1 /*
   2  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/symbolTable.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
  30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
  31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
  32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
  33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
  34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
  35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
  36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
  37 #include "gc_implementation/parNew/parNewGeneration.hpp"
  38 #include "gc_implementation/shared/collectorCounters.hpp"
  39 #include "gc_implementation/shared/isGCActiveMark.hpp"
  40 #include "gc_interface/collectedHeap.inline.hpp"
  41 #include "memory/cardTableRS.hpp"
  42 #include "memory/collectorPolicy.hpp"
  43 #include "memory/gcLocker.inline.hpp"
  44 #include "memory/genCollectedHeap.hpp"
  45 #include "memory/genMarkSweep.hpp"
  46 #include "memory/genOopClosures.inline.hpp"
  47 #include "memory/iterator.hpp"
  48 #include "memory/referencePolicy.hpp"
  49 #include "memory/resourceArea.hpp"
  50 #include "oops/oop.inline.hpp"
  51 #include "prims/jvmtiExport.hpp"
  52 #include "runtime/globals_extension.hpp"
  53 #include "runtime/handles.inline.hpp"
  54 #include "runtime/java.hpp"
  55 #include "runtime/vmThread.hpp"
  56 #include "services/memoryService.hpp"
  57 #include "services/runtimeService.hpp"
  58 
  59 // statics
  60 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  61 bool          CMSCollector::_full_gc_requested          = false;
  62 
  63 //////////////////////////////////////////////////////////////////
  64 // In support of CMS/VM thread synchronization
  65 //////////////////////////////////////////////////////////////////
  66 // We split use of the CGC_lock into 2 "levels".
  67 // The low-level locking is of the usual CGC_lock monitor. We introduce
  68 // a higher level "token" (hereafter "CMS token") built on top of the
  69 // low level monitor (hereafter "CGC lock").
  70 // The token-passing protocol gives priority to the VM thread. The
  71 // CMS-lock doesn't provide any fairness guarantees, but clients
  72 // should ensure that it is only held for very short, bounded
  73 // durations.
  74 //
  75 // When either of the CMS thread or the VM thread is involved in
  76 // collection operations during which it does not want the other
  77 // thread to interfere, it obtains the CMS token.
  78 //
  79 // If either thread tries to get the token while the other has
  80 // it, that thread waits. However, if the VM thread and CMS thread
  81 // both want the token, then the VM thread gets priority while the
  82 // CMS thread waits. This ensures, for instance, that the "concurrent"
  83 // phases of the CMS thread's work do not block out the VM thread
  84 // for long periods of time as the CMS thread continues to hog
  85 // the token. (See bug 4616232).
  86 //
  87 // The baton-passing functions are, however, controlled by the
  88 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
  89 // and here the low-level CMS lock, not the high level token,
  90 // ensures mutual exclusion.
  91 //
  92 // Two important conditions that we have to satisfy:
  93 // 1. if a thread does a low-level wait on the CMS lock, then it
  94 //    relinquishes the CMS token if it were holding that token
  95 //    when it acquired the low-level CMS lock.
  96 // 2. any low-level notifications on the low-level lock
  97 //    should only be sent when a thread has relinquished the token.
  98 //
  99 // In the absence of either property, we'd have potential deadlock.
 100 //
 101 // We protect each of the CMS (concurrent and sequential) phases
 102 // with the CMS _token_, not the CMS _lock_.
 103 //
 104 // The only code protected by CMS lock is the token acquisition code
 105 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 106 // baton-passing code.
 107 //
 108 // Unfortunately, i couldn't come up with a good abstraction to factor and
 109 // hide the naked CGC_lock manipulation in the baton-passing code
 110 // further below. That's something we should try to do. Also, the proof
 111 // of correctness of this 2-level locking scheme is far from obvious,
 112 // and potentially quite slippery. We have an uneasy supsicion, for instance,
 113 // that there may be a theoretical possibility of delay/starvation in the
 114 // low-level lock/wait/notify scheme used for the baton-passing because of
 115 // potential intereference with the priority scheme embodied in the
 116 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 117 // invocation further below and marked with "XXX 20011219YSR".
 118 // Indeed, as we note elsewhere, this may become yet more slippery
 119 // in the presence of multiple CMS and/or multiple VM threads. XXX
 120 
 121 class CMSTokenSync: public StackObj {
 122  private:
 123   bool _is_cms_thread;
 124  public:
 125   CMSTokenSync(bool is_cms_thread):
 126     _is_cms_thread(is_cms_thread) {
 127     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 128            "Incorrect argument to constructor");
 129     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 130   }
 131 
 132   ~CMSTokenSync() {
 133     assert(_is_cms_thread ?
 134              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 135              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 136           "Incorrect state");
 137     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 138   }
 139 };
 140 
 141 // Convenience class that does a CMSTokenSync, and then acquires
 142 // upto three locks.
 143 class CMSTokenSyncWithLocks: public CMSTokenSync {
 144  private:
 145   // Note: locks are acquired in textual declaration order
 146   // and released in the opposite order
 147   MutexLockerEx _locker1, _locker2, _locker3;
 148  public:
 149   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 150                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 151     CMSTokenSync(is_cms_thread),
 152     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 153     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 154     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 155   { }
 156 };
 157 
 158 
 159 // Wrapper class to temporarily disable icms during a foreground cms collection.
 160 class ICMSDisabler: public StackObj {
 161  public:
 162   // The ctor disables icms and wakes up the thread so it notices the change;
 163   // the dtor re-enables icms.  Note that the CMSCollector methods will check
 164   // CMSIncrementalMode.
 165   ICMSDisabler()  { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
 166   ~ICMSDisabler() { CMSCollector::enable_icms(); }
 167 };
 168 
 169 //////////////////////////////////////////////////////////////////
 170 //  Concurrent Mark-Sweep Generation /////////////////////////////
 171 //////////////////////////////////////////////////////////////////
 172 
 173 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 174 
 175 // This struct contains per-thread things necessary to support parallel
 176 // young-gen collection.
 177 class CMSParGCThreadState: public CHeapObj {
 178  public:
 179   CFLS_LAB lab;
 180   PromotionInfo promo;
 181 
 182   // Constructor.
 183   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 184     promo.setSpace(cfls);
 185   }
 186 };
 187 
 188 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 189      ReservedSpace rs, size_t initial_byte_size, int level,
 190      CardTableRS* ct, bool use_adaptive_freelists,
 191      FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
 192   CardGeneration(rs, initial_byte_size, level, ct),
 193   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 194   _debug_collection_type(Concurrent_collection_type)
 195 {
 196   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 197   HeapWord* end    = (HeapWord*) _virtual_space.high();
 198 
 199   _direct_allocated_words = 0;
 200   NOT_PRODUCT(
 201     _numObjectsPromoted = 0;
 202     _numWordsPromoted = 0;
 203     _numObjectsAllocated = 0;
 204     _numWordsAllocated = 0;
 205   )
 206 
 207   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
 208                                            use_adaptive_freelists,
 209                                            dictionaryChoice);
 210   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 211   if (_cmsSpace == NULL) {
 212     vm_exit_during_initialization(
 213       "CompactibleFreeListSpace allocation failure");
 214   }
 215   _cmsSpace->_gen = this;
 216 
 217   _gc_stats = new CMSGCStats();
 218 
 219   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 220   // offsets match. The ability to tell free chunks from objects
 221   // depends on this property.
 222   debug_only(
 223     FreeChunk* junk = NULL;
 224     assert(UseCompressedOops ||
 225            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 226            "Offset of FreeChunk::_prev within FreeChunk must match"
 227            "  that of OopDesc::_klass within OopDesc");
 228   )
 229   if (CollectedHeap::use_parallel_gc_threads()) {
 230     typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
 231     _par_gc_thread_states =
 232       NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
 233     if (_par_gc_thread_states == NULL) {
 234       vm_exit_during_initialization("Could not allocate par gc structs");
 235     }
 236     for (uint i = 0; i < ParallelGCThreads; i++) {
 237       _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 238       if (_par_gc_thread_states[i] == NULL) {
 239         vm_exit_during_initialization("Could not allocate par gc structs");
 240       }
 241     }
 242   } else {
 243     _par_gc_thread_states = NULL;
 244   }
 245   _incremental_collection_failed = false;
 246   // The "dilatation_factor" is the expansion that can occur on
 247   // account of the fact that the minimum object size in the CMS
 248   // generation may be larger than that in, say, a contiguous young
 249   //  generation.
 250   // Ideally, in the calculation below, we'd compute the dilatation
 251   // factor as: MinChunkSize/(promoting_gen's min object size)
 252   // Since we do not have such a general query interface for the
 253   // promoting generation, we'll instead just use the mimimum
 254   // object size (which today is a header's worth of space);
 255   // note that all arithmetic is in units of HeapWords.
 256   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 257   assert(_dilatation_factor >= 1.0, "from previous assert");
 258 }
 259 
 260 
 261 // The field "_initiating_occupancy" represents the occupancy percentage
 262 // at which we trigger a new collection cycle.  Unless explicitly specified
 263 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
 264 // is calculated by:
 265 //
 266 //   Let "f" be MinHeapFreeRatio in
 267 //
 268 //    _intiating_occupancy = 100-f +
 269 //                           f * (CMSTrigger[Perm]Ratio/100)
 270 //   where CMSTrigger[Perm]Ratio is the argument "tr" below.
 271 //
 272 // That is, if we assume the heap is at its desired maximum occupancy at the
 273 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
 274 // space be allocated before initiating a new collection cycle.
 275 //
 276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
 277   assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
 278   if (io >= 0) {
 279     _initiating_occupancy = (double)io / 100.0;
 280   } else {
 281     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 282                              (double)(tr * MinHeapFreeRatio) / 100.0)
 283                             / 100.0;
 284   }
 285 }
 286 
 287 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 288   assert(collector() != NULL, "no collector");
 289   collector()->ref_processor_init();
 290 }
 291 
 292 void CMSCollector::ref_processor_init() {
 293   if (_ref_processor == NULL) {
 294     // Allocate and initialize a reference processor
 295     _ref_processor =
 296       new ReferenceProcessor(_span,                               // span
 297                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 298                              ParallelGCThreads,                   // mt processing degree
 299                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 300                              ConcGCThreads,                       // mt discovery degree
 301                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 302                              &_is_alive_closure,                  // closure for liveness info
 303                              false);                              // next field updates do not need write barrier
 304     // Initialize the _ref_processor field of CMSGen
 305     _cmsGen->set_ref_processor(_ref_processor);
 306 
 307     // Allocate a dummy ref processor for perm gen.
 308     ReferenceProcessor* rp2 = new ReferenceProcessor();
 309     if (rp2 == NULL) {
 310       vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
 311     }
 312     _permGen->set_ref_processor(rp2);
 313   }
 314 }
 315 
 316 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
 317   GenCollectedHeap* gch = GenCollectedHeap::heap();
 318   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
 319     "Wrong type of heap");
 320   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
 321     gch->gen_policy()->size_policy();
 322   assert(sp->is_gc_cms_adaptive_size_policy(),
 323     "Wrong type of size policy");
 324   return sp;
 325 }
 326 
 327 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
 328   CMSGCAdaptivePolicyCounters* results =
 329     (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
 330   assert(
 331     results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
 332     "Wrong gc policy counter kind");
 333   return results;
 334 }
 335 
 336 
 337 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 338 
 339   const char* gen_name = "old";
 340 
 341   // Generation Counters - generation 1, 1 subspace
 342   _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
 343 
 344   _space_counters = new GSpaceCounters(gen_name, 0,
 345                                        _virtual_space.reserved_size(),
 346                                        this, _gen_counters);
 347 }
 348 
 349 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 350   _cms_gen(cms_gen)
 351 {
 352   assert(alpha <= 100, "bad value");
 353   _saved_alpha = alpha;
 354 
 355   // Initialize the alphas to the bootstrap value of 100.
 356   _gc0_alpha = _cms_alpha = 100;
 357 
 358   _cms_begin_time.update();
 359   _cms_end_time.update();
 360 
 361   _gc0_duration = 0.0;
 362   _gc0_period = 0.0;
 363   _gc0_promoted = 0;
 364 
 365   _cms_duration = 0.0;
 366   _cms_period = 0.0;
 367   _cms_allocated = 0;
 368 
 369   _cms_used_at_gc0_begin = 0;
 370   _cms_used_at_gc0_end = 0;
 371   _allow_duty_cycle_reduction = false;
 372   _valid_bits = 0;
 373   _icms_duty_cycle = CMSIncrementalDutyCycle;
 374 }
 375 
 376 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 377   // TBD: CR 6909490
 378   return 1.0;
 379 }
 380 
 381 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 382 }
 383 
 384 // If promotion failure handling is on use
 385 // the padded average size of the promotion for each
 386 // young generation collection.
 387 double CMSStats::time_until_cms_gen_full() const {
 388   size_t cms_free = _cms_gen->cmsSpace()->free();
 389   GenCollectedHeap* gch = GenCollectedHeap::heap();
 390   size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
 391                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 392   if (cms_free > expected_promotion) {
 393     // Start a cms collection if there isn't enough space to promote
 394     // for the next minor collection.  Use the padded average as
 395     // a safety factor.
 396     cms_free -= expected_promotion;
 397 
 398     // Adjust by the safety factor.
 399     double cms_free_dbl = (double)cms_free;
 400     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
 401     // Apply a further correction factor which tries to adjust
 402     // for recent occurance of concurrent mode failures.
 403     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 404     cms_free_dbl = cms_free_dbl * cms_adjustment;
 405 
 406     if (PrintGCDetails && Verbose) {
 407       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
 408         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 409         cms_free, expected_promotion);
 410       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
 411         cms_free_dbl, cms_consumption_rate() + 1.0);
 412     }
 413     // Add 1 in case the consumption rate goes to zero.
 414     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 415   }
 416   return 0.0;
 417 }
 418 
 419 // Compare the duration of the cms collection to the
 420 // time remaining before the cms generation is empty.
 421 // Note that the time from the start of the cms collection
 422 // to the start of the cms sweep (less than the total
 423 // duration of the cms collection) can be used.  This
 424 // has been tried and some applications experienced
 425 // promotion failures early in execution.  This was
 426 // possibly because the averages were not accurate
 427 // enough at the beginning.
 428 double CMSStats::time_until_cms_start() const {
 429   // We add "gc0_period" to the "work" calculation
 430   // below because this query is done (mostly) at the
 431   // end of a scavenge, so we need to conservatively
 432   // account for that much possible delay
 433   // in the query so as to avoid concurrent mode failures
 434   // due to starting the collection just a wee bit too
 435   // late.
 436   double work = cms_duration() + gc0_period();
 437   double deadline = time_until_cms_gen_full();
 438   // If a concurrent mode failure occurred recently, we want to be
 439   // more conservative and halve our expected time_until_cms_gen_full()
 440   if (work > deadline) {
 441     if (Verbose && PrintGCDetails) {
 442       gclog_or_tty->print(
 443         " CMSCollector: collect because of anticipated promotion "
 444         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
 445         gc0_period(), time_until_cms_gen_full());
 446     }
 447     return 0.0;
 448   }
 449   return work - deadline;
 450 }
 451 
 452 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
 453 // amount of change to prevent wild oscillation.
 454 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
 455                                               unsigned int new_duty_cycle) {
 456   assert(old_duty_cycle <= 100, "bad input value");
 457   assert(new_duty_cycle <= 100, "bad input value");
 458 
 459   // Note:  use subtraction with caution since it may underflow (values are
 460   // unsigned).  Addition is safe since we're in the range 0-100.
 461   unsigned int damped_duty_cycle = new_duty_cycle;
 462   if (new_duty_cycle < old_duty_cycle) {
 463     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
 464     if (new_duty_cycle + largest_delta < old_duty_cycle) {
 465       damped_duty_cycle = old_duty_cycle - largest_delta;
 466     }
 467   } else if (new_duty_cycle > old_duty_cycle) {
 468     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
 469     if (new_duty_cycle > old_duty_cycle + largest_delta) {
 470       damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
 471     }
 472   }
 473   assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
 474 
 475   if (CMSTraceIncrementalPacing) {
 476     gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
 477                            old_duty_cycle, new_duty_cycle, damped_duty_cycle);
 478   }
 479   return damped_duty_cycle;
 480 }
 481 
 482 unsigned int CMSStats::icms_update_duty_cycle_impl() {
 483   assert(CMSIncrementalPacing && valid(),
 484          "should be handled in icms_update_duty_cycle()");
 485 
 486   double cms_time_so_far = cms_timer().seconds();
 487   double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
 488   double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
 489 
 490   // Avoid division by 0.
 491   double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
 492   double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
 493 
 494   unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
 495   if (new_duty_cycle > _icms_duty_cycle) {
 496     // Avoid very small duty cycles (1 or 2); 0 is allowed.
 497     if (new_duty_cycle > 2) {
 498       _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
 499                                                 new_duty_cycle);
 500     }
 501   } else if (_allow_duty_cycle_reduction) {
 502     // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
 503     new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
 504     // Respect the minimum duty cycle.
 505     unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
 506     _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
 507   }
 508 
 509   if (PrintGCDetails || CMSTraceIncrementalPacing) {
 510     gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
 511   }
 512 
 513   _allow_duty_cycle_reduction = false;
 514   return _icms_duty_cycle;
 515 }
 516 
 517 #ifndef PRODUCT
 518 void CMSStats::print_on(outputStream *st) const {
 519   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 520   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 521                gc0_duration(), gc0_period(), gc0_promoted());
 522   st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 523             cms_duration(), cms_duration_per_mb(),
 524             cms_period(), cms_allocated());
 525   st->print(",cms_since_beg=%g,cms_since_end=%g",
 526             cms_time_since_begin(), cms_time_since_end());
 527   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 528             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 529   if (CMSIncrementalMode) {
 530     st->print(",dc=%d", icms_duty_cycle());
 531   }
 532 
 533   if (valid()) {
 534     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 535               promotion_rate(), cms_allocation_rate());
 536     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 537               cms_consumption_rate(), time_until_cms_gen_full());
 538   }
 539   st->print(" ");
 540 }
 541 #endif // #ifndef PRODUCT
 542 
 543 CMSCollector::CollectorState CMSCollector::_collectorState =
 544                              CMSCollector::Idling;
 545 bool CMSCollector::_foregroundGCIsActive = false;
 546 bool CMSCollector::_foregroundGCShouldWait = false;
 547 
 548 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 549                            ConcurrentMarkSweepGeneration* permGen,
 550                            CardTableRS*                   ct,
 551                            ConcurrentMarkSweepPolicy*     cp):
 552   _cmsGen(cmsGen),
 553   _permGen(permGen),
 554   _ct(ct),
 555   _ref_processor(NULL),    // will be set later
 556   _conc_workers(NULL),     // may be set later
 557   _abort_preclean(false),
 558   _start_sampling(false),
 559   _between_prologue_and_epilogue(false),
 560   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 561   _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
 562   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 563                  -1 /* lock-free */, "No_lock" /* dummy */),
 564   _modUnionClosure(&_modUnionTable),
 565   _modUnionClosurePar(&_modUnionTable),
 566   // Adjust my span to cover old (cms) gen and perm gen
 567   _span(cmsGen->reserved()._union(permGen->reserved())),
 568   // Construct the is_alive_closure with _span & markBitMap
 569   _is_alive_closure(_span, &_markBitMap),
 570   _restart_addr(NULL),
 571   _overflow_list(NULL),
 572   _stats(cmsGen),
 573   _eden_chunk_array(NULL),     // may be set in ctor body
 574   _eden_chunk_capacity(0),     // -- ditto --
 575   _eden_chunk_index(0),        // -- ditto --
 576   _survivor_plab_array(NULL),  // -- ditto --
 577   _survivor_chunk_array(NULL), // -- ditto --
 578   _survivor_chunk_capacity(0), // -- ditto --
 579   _survivor_chunk_index(0),    // -- ditto --
 580   _ser_pmc_preclean_ovflw(0),
 581   _ser_kac_preclean_ovflw(0),
 582   _ser_pmc_remark_ovflw(0),
 583   _par_pmc_remark_ovflw(0),
 584   _ser_kac_ovflw(0),
 585   _par_kac_ovflw(0),
 586 #ifndef PRODUCT
 587   _num_par_pushes(0),
 588 #endif
 589   _collection_count_start(0),
 590   _verifying(false),
 591   _icms_start_limit(NULL),
 592   _icms_stop_limit(NULL),
 593   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 594   _completed_initialization(false),
 595   _collector_policy(cp),
 596   _should_unload_classes(false),
 597   _concurrent_cycles_since_last_unload(0),
 598   _roots_scanning_options(0),
 599   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 600   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
 601 {
 602   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
 603     ExplicitGCInvokesConcurrent = true;
 604   }
 605   // Now expand the span and allocate the collection support structures
 606   // (MUT, marking bit map etc.) to cover both generations subject to
 607   // collection.
 608 
 609   // First check that _permGen is adjacent to _cmsGen and above it.
 610   assert(   _cmsGen->reserved().word_size()  > 0
 611          && _permGen->reserved().word_size() > 0,
 612          "generations should not be of zero size");
 613   assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
 614          "_cmsGen and _permGen should not overlap");
 615   assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
 616          "_cmsGen->end() different from _permGen->start()");
 617 
 618   // For use by dirty card to oop closures.
 619   _cmsGen->cmsSpace()->set_collector(this);
 620   _permGen->cmsSpace()->set_collector(this);
 621 
 622   // Allocate MUT and marking bit map
 623   {
 624     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 625     if (!_markBitMap.allocate(_span)) {
 626       warning("Failed to allocate CMS Bit Map");
 627       return;
 628     }
 629     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 630   }
 631   {
 632     _modUnionTable.allocate(_span);
 633     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 634   }
 635 
 636   if (!_markStack.allocate(MarkStackSize)) {
 637     warning("Failed to allocate CMS Marking Stack");
 638     return;
 639   }
 640   if (!_revisitStack.allocate(CMSRevisitStackSize)) {
 641     warning("Failed to allocate CMS Revisit Stack");
 642     return;
 643   }
 644 
 645   // Support for multi-threaded concurrent phases
 646   if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) {
 647     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 648       // just for now
 649       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
 650     }
 651     if (ConcGCThreads > 1) {
 652       _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
 653                                  ConcGCThreads, true);
 654       if (_conc_workers == NULL) {
 655         warning("GC/CMS: _conc_workers allocation failure: "
 656               "forcing -CMSConcurrentMTEnabled");
 657         CMSConcurrentMTEnabled = false;
 658       } else {
 659         _conc_workers->initialize_workers();
 660       }
 661     } else {
 662       CMSConcurrentMTEnabled = false;
 663     }
 664   }
 665   if (!CMSConcurrentMTEnabled) {
 666     ConcGCThreads = 0;
 667   } else {
 668     // Turn off CMSCleanOnEnter optimization temporarily for
 669     // the MT case where it's not fixed yet; see 6178663.
 670     CMSCleanOnEnter = false;
 671   }
 672   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 673          "Inconsistency");
 674 
 675   // Parallel task queues; these are shared for the
 676   // concurrent and stop-world phases of CMS, but
 677   // are not shared with parallel scavenge (ParNew).
 678   {
 679     uint i;
 680     uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
 681 
 682     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 683          || ParallelRefProcEnabled)
 684         && num_queues > 0) {
 685       _task_queues = new OopTaskQueueSet(num_queues);
 686       if (_task_queues == NULL) {
 687         warning("task_queues allocation failure.");
 688         return;
 689       }
 690       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
 691       if (_hash_seed == NULL) {
 692         warning("_hash_seed array allocation failure");
 693         return;
 694       }
 695 
 696       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 697       for (i = 0; i < num_queues; i++) {
 698         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 699         if (q == NULL) {
 700           warning("work_queue allocation failure.");
 701           return;
 702         }
 703         _task_queues->register_queue(i, q);
 704       }
 705       for (i = 0; i < num_queues; i++) {
 706         _task_queues->queue(i)->initialize();
 707         _hash_seed[i] = 17;  // copied from ParNew
 708       }
 709     }
 710   }
 711 
 712   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 713   _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
 714 
 715   // Clip CMSBootstrapOccupancy between 0 and 100.
 716   _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
 717                          /(double)100;
 718 
 719   _full_gcs_since_conc_gc = 0;
 720 
 721   // Now tell CMS generations the identity of their collector
 722   ConcurrentMarkSweepGeneration::set_collector(this);
 723 
 724   // Create & start a CMS thread for this CMS collector
 725   _cmsThread = ConcurrentMarkSweepThread::start(this);
 726   assert(cmsThread() != NULL, "CMS Thread should have been created");
 727   assert(cmsThread()->collector() == this,
 728          "CMS Thread should refer to this gen");
 729   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 730 
 731   // Support for parallelizing young gen rescan
 732   GenCollectedHeap* gch = GenCollectedHeap::heap();
 733   _young_gen = gch->prev_gen(_cmsGen);
 734   if (gch->supports_inline_contig_alloc()) {
 735     _top_addr = gch->top_addr();
 736     _end_addr = gch->end_addr();
 737     assert(_young_gen != NULL, "no _young_gen");
 738     _eden_chunk_index = 0;
 739     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
 740     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
 741     if (_eden_chunk_array == NULL) {
 742       _eden_chunk_capacity = 0;
 743       warning("GC/CMS: _eden_chunk_array allocation failure");
 744     }
 745   }
 746   assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
 747 
 748   // Support for parallelizing survivor space rescan
 749   if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
 750     const size_t max_plab_samples =
 751       ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
 752 
 753     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
 754     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
 755     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
 756     if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
 757         || _cursor == NULL) {
 758       warning("Failed to allocate survivor plab/chunk array");
 759       if (_survivor_plab_array  != NULL) {
 760         FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
 761         _survivor_plab_array = NULL;
 762       }
 763       if (_survivor_chunk_array != NULL) {
 764         FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
 765         _survivor_chunk_array = NULL;
 766       }
 767       if (_cursor != NULL) {
 768         FREE_C_HEAP_ARRAY(size_t, _cursor);
 769         _cursor = NULL;
 770       }
 771     } else {
 772       _survivor_chunk_capacity = 2*max_plab_samples;
 773       for (uint i = 0; i < ParallelGCThreads; i++) {
 774         HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
 775         if (vec == NULL) {
 776           warning("Failed to allocate survivor plab array");
 777           for (int j = i; j > 0; j--) {
 778             FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
 779           }
 780           FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
 781           FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
 782           _survivor_plab_array = NULL;
 783           _survivor_chunk_array = NULL;
 784           _survivor_chunk_capacity = 0;
 785           break;
 786         } else {
 787           ChunkArray* cur =
 788             ::new (&_survivor_plab_array[i]) ChunkArray(vec,
 789                                                         max_plab_samples);
 790           assert(cur->end() == 0, "Should be 0");
 791           assert(cur->array() == vec, "Should be vec");
 792           assert(cur->capacity() == max_plab_samples, "Error");
 793         }
 794       }
 795     }
 796   }
 797   assert(   (   _survivor_plab_array  != NULL
 798              && _survivor_chunk_array != NULL)
 799          || (   _survivor_chunk_capacity == 0
 800              && _survivor_chunk_index == 0),
 801          "Error");
 802 
 803   // Choose what strong roots should be scanned depending on verification options
 804   // and perm gen collection mode.
 805   if (!CMSClassUnloadingEnabled) {
 806     // If class unloading is disabled we want to include all classes into the root set.
 807     add_root_scanning_option(SharedHeap::SO_AllClasses);
 808   } else {
 809     add_root_scanning_option(SharedHeap::SO_SystemClasses);
 810   }
 811 
 812   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 813   _gc_counters = new CollectorCounters("CMS", 1);
 814   _completed_initialization = true;
 815   _inter_sweep_timer.start();  // start of time
 816 #ifdef SPARC
 817   // Issue a stern warning, but allow use for experimentation and debugging.
 818   if (VM_Version::is_sun4v() && UseMemSetInBOT) {
 819     assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
 820     warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
 821             " on sun4v; please understand that you are using at your own risk!");
 822   }
 823 #endif
 824 }
 825 
 826 const char* ConcurrentMarkSweepGeneration::name() const {
 827   return "concurrent mark-sweep generation";
 828 }
 829 void ConcurrentMarkSweepGeneration::update_counters() {
 830   if (UsePerfData) {
 831     _space_counters->update_all();
 832     _gen_counters->update_all();
 833   }
 834 }
 835 
 836 // this is an optimized version of update_counters(). it takes the
 837 // used value as a parameter rather than computing it.
 838 //
 839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 840   if (UsePerfData) {
 841     _space_counters->update_used(used);
 842     _space_counters->update_capacity();
 843     _gen_counters->update_all();
 844   }
 845 }
 846 
 847 void ConcurrentMarkSweepGeneration::print() const {
 848   Generation::print();
 849   cmsSpace()->print();
 850 }
 851 
 852 #ifndef PRODUCT
 853 void ConcurrentMarkSweepGeneration::print_statistics() {
 854   cmsSpace()->printFLCensus(0);
 855 }
 856 #endif
 857 
 858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
 859   GenCollectedHeap* gch = GenCollectedHeap::heap();
 860   if (PrintGCDetails) {
 861     if (Verbose) {
 862       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
 863         level(), short_name(), s, used(), capacity());
 864     } else {
 865       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
 866         level(), short_name(), s, used() / K, capacity() / K);
 867     }
 868   }
 869   if (Verbose) {
 870     gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
 871               gch->used(), gch->capacity());
 872   } else {
 873     gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
 874               gch->used() / K, gch->capacity() / K);
 875   }
 876 }
 877 
 878 size_t
 879 ConcurrentMarkSweepGeneration::contiguous_available() const {
 880   // dld proposes an improvement in precision here. If the committed
 881   // part of the space ends in a free block we should add that to
 882   // uncommitted size in the calculation below. Will make this
 883   // change later, staying with the approximation below for the
 884   // time being. -- ysr.
 885   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 886 }
 887 
 888 size_t
 889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 890   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 891 }
 892 
 893 size_t ConcurrentMarkSweepGeneration::max_available() const {
 894   return free() + _virtual_space.uncommitted_size();
 895 }
 896 
 897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 898   size_t available = max_available();
 899   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 900   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 901   if (Verbose && PrintGCDetails) {
 902     gclog_or_tty->print_cr(
 903       "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
 904       "max_promo("SIZE_FORMAT")",
 905       res? "":" not", available, res? ">=":"<",
 906       av_promo, max_promotion_in_bytes);
 907   }
 908   return res;
 909 }
 910 
 911 // At a promotion failure dump information on block layout in heap
 912 // (cms old generation).
 913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 914   if (CMSDumpAtPromotionFailure) {
 915     cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
 916   }
 917 }
 918 
 919 CompactibleSpace*
 920 ConcurrentMarkSweepGeneration::first_compaction_space() const {
 921   return _cmsSpace;
 922 }
 923 
 924 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 925   // Clear the promotion information.  These pointers can be adjusted
 926   // along with all the other pointers into the heap but
 927   // compaction is expected to be a rare event with
 928   // a heap using cms so don't do it without seeing the need.
 929   if (CollectedHeap::use_parallel_gc_threads()) {
 930     for (uint i = 0; i < ParallelGCThreads; i++) {
 931       _par_gc_thread_states[i]->promo.reset();
 932     }
 933   }
 934 }
 935 
 936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
 937   blk->do_space(_cmsSpace);
 938 }
 939 
 940 void ConcurrentMarkSweepGeneration::compute_new_size() {
 941   assert_locked_or_safepoint(Heap_lock);
 942 
 943   // If incremental collection failed, we just want to expand
 944   // to the limit.
 945   if (incremental_collection_failed()) {
 946     clear_incremental_collection_failed();
 947     grow_to_reserved();
 948     return;
 949   }
 950 
 951   size_t expand_bytes = 0;
 952   double free_percentage = ((double) free()) / capacity();
 953   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 954   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 955 
 956   // compute expansion delta needed for reaching desired free percentage
 957   if (free_percentage < desired_free_percentage) {
 958     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 959     assert(desired_capacity >= capacity(), "invalid expansion size");
 960     expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 961   }
 962   if (expand_bytes > 0) {
 963     if (PrintGCDetails && Verbose) {
 964       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 965       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
 966       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
 967       gclog_or_tty->print_cr("  Desired free fraction %f",
 968         desired_free_percentage);
 969       gclog_or_tty->print_cr("  Maximum free fraction %f",
 970         maximum_free_percentage);
 971       gclog_or_tty->print_cr("  Capactiy "SIZE_FORMAT, capacity()/1000);
 972       gclog_or_tty->print_cr("  Desired capacity "SIZE_FORMAT,
 973         desired_capacity/1000);
 974       int prev_level = level() - 1;
 975       if (prev_level >= 0) {
 976         size_t prev_size = 0;
 977         GenCollectedHeap* gch = GenCollectedHeap::heap();
 978         Generation* prev_gen = gch->_gens[prev_level];
 979         prev_size = prev_gen->capacity();
 980           gclog_or_tty->print_cr("  Younger gen size "SIZE_FORMAT,
 981                                  prev_size/1000);
 982       }
 983       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc "SIZE_FORMAT,
 984         unsafe_max_alloc_nogc()/1000);
 985       gclog_or_tty->print_cr("  contiguous available "SIZE_FORMAT,
 986         contiguous_available()/1000);
 987       gclog_or_tty->print_cr("  Expand by "SIZE_FORMAT" (bytes)",
 988         expand_bytes);
 989     }
 990     // safe if expansion fails
 991     expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 992     if (PrintGCDetails && Verbose) {
 993       gclog_or_tty->print_cr("  Expanded free fraction %f",
 994         ((double) free()) / capacity());
 995     }
 996   }
 997 }
 998 
 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1000   return cmsSpace()->freelistLock();
1001 }
1002 
1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1004                                                   bool   tlab) {
1005   CMSSynchronousYieldRequest yr;
1006   MutexLockerEx x(freelistLock(),
1007                   Mutex::_no_safepoint_check_flag);
1008   return have_lock_and_allocate(size, tlab);
1009 }
1010 
1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1012                                                   bool   tlab /* ignored */) {
1013   assert_lock_strong(freelistLock());
1014   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1015   HeapWord* res = cmsSpace()->allocate(adjustedSize);
1016   // Allocate the object live (grey) if the background collector has
1017   // started marking. This is necessary because the marker may
1018   // have passed this address and consequently this object will
1019   // not otherwise be greyed and would be incorrectly swept up.
1020   // Note that if this object contains references, the writing
1021   // of those references will dirty the card containing this object
1022   // allowing the object to be blackened (and its references scanned)
1023   // either during a preclean phase or at the final checkpoint.
1024   if (res != NULL) {
1025     // We may block here with an uninitialized object with
1026     // its mark-bit or P-bits not yet set. Such objects need
1027     // to be safely navigable by block_start().
1028     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1029     assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size");
1030     collector()->direct_allocated(res, adjustedSize);
1031     _direct_allocated_words += adjustedSize;
1032     // allocation counters
1033     NOT_PRODUCT(
1034       _numObjectsAllocated++;
1035       _numWordsAllocated += (int)adjustedSize;
1036     )
1037   }
1038   return res;
1039 }
1040 
1041 // In the case of direct allocation by mutators in a generation that
1042 // is being concurrently collected, the object must be allocated
1043 // live (grey) if the background collector has started marking.
1044 // This is necessary because the marker may
1045 // have passed this address and consequently this object will
1046 // not otherwise be greyed and would be incorrectly swept up.
1047 // Note that if this object contains references, the writing
1048 // of those references will dirty the card containing this object
1049 // allowing the object to be blackened (and its references scanned)
1050 // either during a preclean phase or at the final checkpoint.
1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1052   assert(_markBitMap.covers(start, size), "Out of bounds");
1053   if (_collectorState >= Marking) {
1054     MutexLockerEx y(_markBitMap.lock(),
1055                     Mutex::_no_safepoint_check_flag);
1056     // [see comments preceding SweepClosure::do_blk() below for details]
1057     // 1. need to mark the object as live so it isn't collected
1058     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1059     // 3. need to mark the end of the object so marking, precleaning or sweeping
1060     //    can skip over uninitialized or unparsable objects. An allocated
1061     //    object is considered uninitialized for our purposes as long as
1062     //    its klass word is NULL. (Unparsable objects are those which are
1063     //    initialized in the sense just described, but whose sizes can still
1064     //    not be correctly determined. Note that the class of unparsable objects
1065     //    can only occur in the perm gen. All old gen objects are parsable
1066     //    as soon as they are initialized.)
1067     _markBitMap.mark(start);          // object is live
1068     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
1069     _markBitMap.mark(start + size - 1);
1070                                       // mark end of object
1071   }
1072   // check that oop looks uninitialized
1073   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1074 }
1075 
1076 void CMSCollector::promoted(bool par, HeapWord* start,
1077                             bool is_obj_array, size_t obj_size) {
1078   assert(_markBitMap.covers(start), "Out of bounds");
1079   // See comment in direct_allocated() about when objects should
1080   // be allocated live.
1081   if (_collectorState >= Marking) {
1082     // we already hold the marking bit map lock, taken in
1083     // the prologue
1084     if (par) {
1085       _markBitMap.par_mark(start);
1086     } else {
1087       _markBitMap.mark(start);
1088     }
1089     // We don't need to mark the object as uninitialized (as
1090     // in direct_allocated above) because this is being done with the
1091     // world stopped and the object will be initialized by the
1092     // time the marking, precleaning or sweeping get to look at it.
1093     // But see the code for copying objects into the CMS generation,
1094     // where we need to ensure that concurrent readers of the
1095     // block offset table are able to safely navigate a block that
1096     // is in flux from being free to being allocated (and in
1097     // transition while being copied into) and subsequently
1098     // becoming a bona-fide object when the copy/promotion is complete.
1099     assert(SafepointSynchronize::is_at_safepoint(),
1100            "expect promotion only at safepoints");
1101 
1102     if (_collectorState < Sweeping) {
1103       // Mark the appropriate cards in the modUnionTable, so that
1104       // this object gets scanned before the sweep. If this is
1105       // not done, CMS generation references in the object might
1106       // not get marked.
1107       // For the case of arrays, which are otherwise precisely
1108       // marked, we need to dirty the entire array, not just its head.
1109       if (is_obj_array) {
1110         // The [par_]mark_range() method expects mr.end() below to
1111         // be aligned to the granularity of a bit's representation
1112         // in the heap. In the case of the MUT below, that's a
1113         // card size.
1114         MemRegion mr(start,
1115                      (HeapWord*)round_to((intptr_t)(start + obj_size),
1116                         CardTableModRefBS::card_size /* bytes */));
1117         if (par) {
1118           _modUnionTable.par_mark_range(mr);
1119         } else {
1120           _modUnionTable.mark_range(mr);
1121         }
1122       } else {  // not an obj array; we can just mark the head
1123         if (par) {
1124           _modUnionTable.par_mark(start);
1125         } else {
1126           _modUnionTable.mark(start);
1127         }
1128       }
1129     }
1130   }
1131 }
1132 
1133 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1134 {
1135   size_t delta = pointer_delta(addr, space->bottom());
1136   return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1137 }
1138 
1139 void CMSCollector::icms_update_allocation_limits()
1140 {
1141   Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1142   EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1143 
1144   const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1145   if (CMSTraceIncrementalPacing) {
1146     stats().print();
1147   }
1148 
1149   assert(duty_cycle <= 100, "invalid duty cycle");
1150   if (duty_cycle != 0) {
1151     // The duty_cycle is a percentage between 0 and 100; convert to words and
1152     // then compute the offset from the endpoints of the space.
1153     size_t free_words = eden->free() / HeapWordSize;
1154     double free_words_dbl = (double)free_words;
1155     size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1156     size_t offset_words = (free_words - duty_cycle_words) / 2;
1157 
1158     _icms_start_limit = eden->top() + offset_words;
1159     _icms_stop_limit = eden->end() - offset_words;
1160 
1161     // The limits may be adjusted (shifted to the right) by
1162     // CMSIncrementalOffset, to allow the application more mutator time after a
1163     // young gen gc (when all mutators were stopped) and before CMS starts and
1164     // takes away one or more cpus.
1165     if (CMSIncrementalOffset != 0) {
1166       double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1167       size_t adjustment = (size_t)adjustment_dbl;
1168       HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1169       if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1170         _icms_start_limit += adjustment;
1171         _icms_stop_limit = tmp_stop;
1172       }
1173     }
1174   }
1175   if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1176     _icms_start_limit = _icms_stop_limit = eden->end();
1177   }
1178 
1179   // Install the new start limit.
1180   eden->set_soft_end(_icms_start_limit);
1181 
1182   if (CMSTraceIncrementalMode) {
1183     gclog_or_tty->print(" icms alloc limits:  "
1184                            PTR_FORMAT "," PTR_FORMAT
1185                            " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1186                            _icms_start_limit, _icms_stop_limit,
1187                            percent_of_space(eden, _icms_start_limit),
1188                            percent_of_space(eden, _icms_stop_limit));
1189     if (Verbose) {
1190       gclog_or_tty->print("eden:  ");
1191       eden->print_on(gclog_or_tty);
1192     }
1193   }
1194 }
1195 
1196 // Any changes here should try to maintain the invariant
1197 // that if this method is called with _icms_start_limit
1198 // and _icms_stop_limit both NULL, then it should return NULL
1199 // and not notify the icms thread.
1200 HeapWord*
1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1202                                        size_t word_size)
1203 {
1204   // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1205   // nop.
1206   if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1207     if (top <= _icms_start_limit) {
1208       if (CMSTraceIncrementalMode) {
1209         space->print_on(gclog_or_tty);
1210         gclog_or_tty->stamp();
1211         gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1212                                ", new limit=" PTR_FORMAT
1213                                " (" SIZE_FORMAT "%%)",
1214                                top, _icms_stop_limit,
1215                                percent_of_space(space, _icms_stop_limit));
1216       }
1217       ConcurrentMarkSweepThread::start_icms();
1218       assert(top < _icms_stop_limit, "Tautology");
1219       if (word_size < pointer_delta(_icms_stop_limit, top)) {
1220         return _icms_stop_limit;
1221       }
1222 
1223       // The allocation will cross both the _start and _stop limits, so do the
1224       // stop notification also and return end().
1225       if (CMSTraceIncrementalMode) {
1226         space->print_on(gclog_or_tty);
1227         gclog_or_tty->stamp();
1228         gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1229                                ", new limit=" PTR_FORMAT
1230                                " (" SIZE_FORMAT "%%)",
1231                                top, space->end(),
1232                                percent_of_space(space, space->end()));
1233       }
1234       ConcurrentMarkSweepThread::stop_icms();
1235       return space->end();
1236     }
1237 
1238     if (top <= _icms_stop_limit) {
1239       if (CMSTraceIncrementalMode) {
1240         space->print_on(gclog_or_tty);
1241         gclog_or_tty->stamp();
1242         gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1243                                ", new limit=" PTR_FORMAT
1244                                " (" SIZE_FORMAT "%%)",
1245                                top, space->end(),
1246                                percent_of_space(space, space->end()));
1247       }
1248       ConcurrentMarkSweepThread::stop_icms();
1249       return space->end();
1250     }
1251 
1252     if (CMSTraceIncrementalMode) {
1253       space->print_on(gclog_or_tty);
1254       gclog_or_tty->stamp();
1255       gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1256                              ", new limit=" PTR_FORMAT,
1257                              top, NULL);
1258     }
1259   }
1260 
1261   return NULL;
1262 }
1263 
1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1265   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1266   // allocate, copy and if necessary update promoinfo --
1267   // delegate to underlying space.
1268   assert_lock_strong(freelistLock());
1269 
1270 #ifndef PRODUCT
1271   if (Universe::heap()->promotion_should_fail()) {
1272     return NULL;
1273   }
1274 #endif  // #ifndef PRODUCT
1275 
1276   oop res = _cmsSpace->promote(obj, obj_size);
1277   if (res == NULL) {
1278     // expand and retry
1279     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
1280     expand(s*HeapWordSize, MinHeapDeltaBytes,
1281       CMSExpansionCause::_satisfy_promotion);
1282     // Since there's currently no next generation, we don't try to promote
1283     // into a more senior generation.
1284     assert(next_gen() == NULL, "assumption, based upon which no attempt "
1285                                "is made to pass on a possibly failing "
1286                                "promotion to next generation");
1287     res = _cmsSpace->promote(obj, obj_size);
1288   }
1289   if (res != NULL) {
1290     // See comment in allocate() about when objects should
1291     // be allocated live.
1292     assert(obj->is_oop(), "Will dereference klass pointer below");
1293     collector()->promoted(false,           // Not parallel
1294                           (HeapWord*)res, obj->is_objArray(), obj_size);
1295     // promotion counters
1296     NOT_PRODUCT(
1297       _numObjectsPromoted++;
1298       _numWordsPromoted +=
1299         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1300     )
1301   }
1302   return res;
1303 }
1304 
1305 
1306 HeapWord*
1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1308                                              HeapWord* top,
1309                                              size_t word_sz)
1310 {
1311   return collector()->allocation_limit_reached(space, top, word_sz);
1312 }
1313 
1314 // IMPORTANT: Notes on object size recognition in CMS.
1315 // ---------------------------------------------------
1316 // A block of storage in the CMS generation is always in
1317 // one of three states. A free block (FREE), an allocated
1318 // object (OBJECT) whose size() method reports the correct size,
1319 // and an intermediate state (TRANSIENT) in which its size cannot
1320 // be accurately determined.
1321 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
1322 // -----------------------------------------------------
1323 // FREE:      klass_word & 1 == 1; mark_word holds block size
1324 //
1325 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1326 //            obj->size() computes correct size
1327 //            [Perm Gen objects needs to be "parsable" before they can be navigated]
1328 //
1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1330 //
1331 // STATE IDENTIFICATION: (64 bit+COOPS)
1332 // ------------------------------------
1333 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1334 //
1335 // OBJECT:    klass_word installed; klass_word != 0;
1336 //            obj->size() computes correct size
1337 //            [Perm Gen comment above continues to hold]
1338 //
1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1340 //
1341 //
1342 // STATE TRANSITION DIAGRAM
1343 //
1344 //        mut / parnew                     mut  /  parnew
1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1346 //  ^                                                                   |
1347 //  |------------------------ DEAD <------------------------------------|
1348 //         sweep                            mut
1349 //
1350 // While a block is in TRANSIENT state its size cannot be determined
1351 // so readers will either need to come back later or stall until
1352 // the size can be determined. Note that for the case of direct
1353 // allocation, P-bits, when available, may be used to determine the
1354 // size of an object that may not yet have been initialized.
1355 
1356 // Things to support parallel young-gen collection.
1357 oop
1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1359                                            oop old, markOop m,
1360                                            size_t word_sz) {
1361 #ifndef PRODUCT
1362   if (Universe::heap()->promotion_should_fail()) {
1363     return NULL;
1364   }
1365 #endif  // #ifndef PRODUCT
1366 
1367   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1368   PromotionInfo* promoInfo = &ps->promo;
1369   // if we are tracking promotions, then first ensure space for
1370   // promotion (including spooling space for saving header if necessary).
1371   // then allocate and copy, then track promoted info if needed.
1372   // When tracking (see PromotionInfo::track()), the mark word may
1373   // be displaced and in this case restoration of the mark word
1374   // occurs in the (oop_since_save_marks_)iterate phase.
1375   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1376     // Out of space for allocating spooling buffers;
1377     // try expanding and allocating spooling buffers.
1378     if (!expand_and_ensure_spooling_space(promoInfo)) {
1379       return NULL;
1380     }
1381   }
1382   assert(promoInfo->has_spooling_space(), "Control point invariant");
1383   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1384   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1385   if (obj_ptr == NULL) {
1386      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1387      if (obj_ptr == NULL) {
1388        return NULL;
1389      }
1390   }
1391   oop obj = oop(obj_ptr);
1392   OrderAccess::storestore();
1393   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1394   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1395   // IMPORTANT: See note on object initialization for CMS above.
1396   // Otherwise, copy the object.  Here we must be careful to insert the
1397   // klass pointer last, since this marks the block as an allocated object.
1398   // Except with compressed oops it's the mark word.
1399   HeapWord* old_ptr = (HeapWord*)old;
1400   // Restore the mark word copied above.
1401   obj->set_mark(m);
1402   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1403   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1404   OrderAccess::storestore();
1405 
1406   if (UseCompressedOops) {
1407     // Copy gap missed by (aligned) header size calculation below
1408     obj->set_klass_gap(old->klass_gap());
1409   }
1410   if (word_sz > (size_t)oopDesc::header_size()) {
1411     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1412                                  obj_ptr + oopDesc::header_size(),
1413                                  word_sz - oopDesc::header_size());
1414   }
1415 
1416   // Now we can track the promoted object, if necessary.  We take care
1417   // to delay the transition from uninitialized to full object
1418   // (i.e., insertion of klass pointer) until after, so that it
1419   // atomically becomes a promoted object.
1420   if (promoInfo->tracking()) {
1421     promoInfo->track((PromotedObject*)obj, old->klass());
1422   }
1423   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1424   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
1425   assert(old->is_oop(), "Will use and dereference old klass ptr below");
1426 
1427   // Finally, install the klass pointer (this should be volatile).
1428   OrderAccess::storestore();
1429   obj->set_klass(old->klass());
1430   // We should now be able to calculate the right size for this object
1431   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1432 
1433   collector()->promoted(true,          // parallel
1434                         obj_ptr, old->is_objArray(), word_sz);
1435 
1436   NOT_PRODUCT(
1437     Atomic::inc_ptr(&_numObjectsPromoted);
1438     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1439   )
1440 
1441   return obj;
1442 }
1443 
1444 void
1445 ConcurrentMarkSweepGeneration::
1446 par_promote_alloc_undo(int thread_num,
1447                        HeapWord* obj, size_t word_sz) {
1448   // CMS does not support promotion undo.
1449   ShouldNotReachHere();
1450 }
1451 
1452 void
1453 ConcurrentMarkSweepGeneration::
1454 par_promote_alloc_done(int thread_num) {
1455   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1456   ps->lab.retire(thread_num);
1457 }
1458 
1459 void
1460 ConcurrentMarkSweepGeneration::
1461 par_oop_since_save_marks_iterate_done(int thread_num) {
1462   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1463   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1464   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1465 }
1466 
1467 // XXXPERM
1468 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1469                                                    size_t size,
1470                                                    bool   tlab)
1471 {
1472   // We allow a STW collection only if a full
1473   // collection was requested.
1474   return full || should_allocate(size, tlab); // FIX ME !!!
1475   // This and promotion failure handling are connected at the
1476   // hip and should be fixed by untying them.
1477 }
1478 
1479 bool CMSCollector::shouldConcurrentCollect() {
1480   if (_full_gc_requested) {
1481     if (Verbose && PrintGCDetails) {
1482       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1483                              " gc request (or gc_locker)");
1484     }
1485     return true;
1486   }
1487 
1488   // For debugging purposes, change the type of collection.
1489   // If the rotation is not on the concurrent collection
1490   // type, don't start a concurrent collection.
1491   NOT_PRODUCT(
1492     if (RotateCMSCollectionTypes &&
1493         (_cmsGen->debug_collection_type() !=
1494           ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1495       assert(_cmsGen->debug_collection_type() !=
1496         ConcurrentMarkSweepGeneration::Unknown_collection_type,
1497         "Bad cms collection type");
1498       return false;
1499     }
1500   )
1501 
1502   FreelistLocker x(this);
1503   // ------------------------------------------------------------------
1504   // Print out lots of information which affects the initiation of
1505   // a collection.
1506   if (PrintCMSInitiationStatistics && stats().valid()) {
1507     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1508     gclog_or_tty->stamp();
1509     gclog_or_tty->print_cr("");
1510     stats().print_on(gclog_or_tty);
1511     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1512       stats().time_until_cms_gen_full());
1513     gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1514     gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1515                            _cmsGen->contiguous_available());
1516     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1517     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1518     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1519     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1520     gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1521   }
1522   // ------------------------------------------------------------------
1523 
1524   // If the estimated time to complete a cms collection (cms_duration())
1525   // is less than the estimated time remaining until the cms generation
1526   // is full, start a collection.
1527   if (!UseCMSInitiatingOccupancyOnly) {
1528     if (stats().valid()) {
1529       if (stats().time_until_cms_start() == 0.0) {
1530         return true;
1531       }
1532     } else {
1533       // We want to conservatively collect somewhat early in order
1534       // to try and "bootstrap" our CMS/promotion statistics;
1535       // this branch will not fire after the first successful CMS
1536       // collection because the stats should then be valid.
1537       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1538         if (Verbose && PrintGCDetails) {
1539           gclog_or_tty->print_cr(
1540             " CMSCollector: collect for bootstrapping statistics:"
1541             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1542             _bootstrap_occupancy);
1543         }
1544         return true;
1545       }
1546     }
1547   }
1548 
1549   // Otherwise, we start a collection cycle if either the perm gen or
1550   // old gen want a collection cycle started. Each may use
1551   // an appropriate criterion for making this decision.
1552   // XXX We need to make sure that the gen expansion
1553   // criterion dovetails well with this. XXX NEED TO FIX THIS
1554   if (_cmsGen->should_concurrent_collect()) {
1555     if (Verbose && PrintGCDetails) {
1556       gclog_or_tty->print_cr("CMS old gen initiated");
1557     }
1558     return true;
1559   }
1560 
1561   // We start a collection if we believe an incremental collection may fail;
1562   // this is not likely to be productive in practice because it's probably too
1563   // late anyway.
1564   GenCollectedHeap* gch = GenCollectedHeap::heap();
1565   assert(gch->collector_policy()->is_two_generation_policy(),
1566          "You may want to check the correctness of the following");
1567   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1568     if (Verbose && PrintGCDetails) {
1569       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1570     }
1571     return true;
1572   }
1573 
1574   if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1575     bool res = update_should_unload_classes();
1576     if (res) {
1577       if (Verbose && PrintGCDetails) {
1578         gclog_or_tty->print_cr("CMS perm gen initiated");
1579       }
1580       return true;
1581     }
1582   }
1583   return false;
1584 }
1585 
1586 // Clear _expansion_cause fields of constituent generations
1587 void CMSCollector::clear_expansion_cause() {
1588   _cmsGen->clear_expansion_cause();
1589   _permGen->clear_expansion_cause();
1590 }
1591 
1592 // We should be conservative in starting a collection cycle.  To
1593 // start too eagerly runs the risk of collecting too often in the
1594 // extreme.  To collect too rarely falls back on full collections,
1595 // which works, even if not optimum in terms of concurrent work.
1596 // As a work around for too eagerly collecting, use the flag
1597 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1598 // giving the user an easily understandable way of controlling the
1599 // collections.
1600 // We want to start a new collection cycle if any of the following
1601 // conditions hold:
1602 // . our current occupancy exceeds the configured initiating occupancy
1603 //   for this generation, or
1604 // . we recently needed to expand this space and have not, since that
1605 //   expansion, done a collection of this generation, or
1606 // . the underlying space believes that it may be a good idea to initiate
1607 //   a concurrent collection (this may be based on criteria such as the
1608 //   following: the space uses linear allocation and linear allocation is
1609 //   going to fail, or there is believed to be excessive fragmentation in
1610 //   the generation, etc... or ...
1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1612 //   the case of the old generation, not the perm generation; see CR 6543076):
1613 //   we may be approaching a point at which allocation requests may fail because
1614 //   we will be out of sufficient free space given allocation rate estimates.]
1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1616 
1617   assert_lock_strong(freelistLock());
1618   if (occupancy() > initiating_occupancy()) {
1619     if (PrintGCDetails && Verbose) {
1620       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
1621         short_name(), occupancy(), initiating_occupancy());
1622     }
1623     return true;
1624   }
1625   if (UseCMSInitiatingOccupancyOnly) {
1626     return false;
1627   }
1628   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1629     if (PrintGCDetails && Verbose) {
1630       gclog_or_tty->print(" %s: collect because expanded for allocation ",
1631         short_name());
1632     }
1633     return true;
1634   }
1635   if (_cmsSpace->should_concurrent_collect()) {
1636     if (PrintGCDetails && Verbose) {
1637       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1638         short_name());
1639     }
1640     return true;
1641   }
1642   return false;
1643 }
1644 
1645 void ConcurrentMarkSweepGeneration::collect(bool   full,
1646                                             bool   clear_all_soft_refs,
1647                                             size_t size,
1648                                             bool   tlab)
1649 {
1650   collector()->collect(full, clear_all_soft_refs, size, tlab);
1651 }
1652 
1653 void CMSCollector::collect(bool   full,
1654                            bool   clear_all_soft_refs,
1655                            size_t size,
1656                            bool   tlab)
1657 {
1658   if (!UseCMSCollectionPassing && _collectorState > Idling) {
1659     // For debugging purposes skip the collection if the state
1660     // is not currently idle
1661     if (TraceCMSState) {
1662       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1663         Thread::current(), full, _collectorState);
1664     }
1665     return;
1666   }
1667 
1668   // The following "if" branch is present for defensive reasons.
1669   // In the current uses of this interface, it can be replaced with:
1670   // assert(!GC_locker.is_active(), "Can't be called otherwise");
1671   // But I am not placing that assert here to allow future
1672   // generality in invoking this interface.
1673   if (GC_locker::is_active()) {
1674     // A consistency test for GC_locker
1675     assert(GC_locker::needs_gc(), "Should have been set already");
1676     // Skip this foreground collection, instead
1677     // expanding the heap if necessary.
1678     // Need the free list locks for the call to free() in compute_new_size()
1679     compute_new_size();
1680     return;
1681   }
1682   acquire_control_and_collect(full, clear_all_soft_refs);
1683   _full_gcs_since_conc_gc++;
1684 
1685 }
1686 
1687 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1688   GenCollectedHeap* gch = GenCollectedHeap::heap();
1689   unsigned int gc_count = gch->total_full_collections();
1690   if (gc_count == full_gc_count) {
1691     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1692     _full_gc_requested = true;
1693     CGC_lock->notify();   // nudge CMS thread
1694   }
1695 }
1696 
1697 
1698 // The foreground and background collectors need to coordinate in order
1699 // to make sure that they do not mutually interfere with CMS collections.
1700 // When a background collection is active,
1701 // the foreground collector may need to take over (preempt) and
1702 // synchronously complete an ongoing collection. Depending on the
1703 // frequency of the background collections and the heap usage
1704 // of the application, this preemption can be seldom or frequent.
1705 // There are only certain
1706 // points in the background collection that the "collection-baton"
1707 // can be passed to the foreground collector.
1708 //
1709 // The foreground collector will wait for the baton before
1710 // starting any part of the collection.  The foreground collector
1711 // will only wait at one location.
1712 //
1713 // The background collector will yield the baton before starting a new
1714 // phase of the collection (e.g., before initial marking, marking from roots,
1715 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1716 // of the loop which switches the phases. The background collector does some
1717 // of the phases (initial mark, final re-mark) with the world stopped.
1718 // Because of locking involved in stopping the world,
1719 // the foreground collector should not block waiting for the background
1720 // collector when it is doing a stop-the-world phase.  The background
1721 // collector will yield the baton at an additional point just before
1722 // it enters a stop-the-world phase.  Once the world is stopped, the
1723 // background collector checks the phase of the collection.  If the
1724 // phase has not changed, it proceeds with the collection.  If the
1725 // phase has changed, it skips that phase of the collection.  See
1726 // the comments on the use of the Heap_lock in collect_in_background().
1727 //
1728 // Variable used in baton passing.
1729 //   _foregroundGCIsActive - Set to true by the foreground collector when
1730 //      it wants the baton.  The foreground clears it when it has finished
1731 //      the collection.
1732 //   _foregroundGCShouldWait - Set to true by the background collector
1733 //        when it is running.  The foreground collector waits while
1734 //      _foregroundGCShouldWait is true.
1735 //  CGC_lock - monitor used to protect access to the above variables
1736 //      and to notify the foreground and background collectors.
1737 //  _collectorState - current state of the CMS collection.
1738 //
1739 // The foreground collector
1740 //   acquires the CGC_lock
1741 //   sets _foregroundGCIsActive
1742 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1743 //     various locks acquired in preparation for the collection
1744 //     are released so as not to block the background collector
1745 //     that is in the midst of a collection
1746 //   proceeds with the collection
1747 //   clears _foregroundGCIsActive
1748 //   returns
1749 //
1750 // The background collector in a loop iterating on the phases of the
1751 //      collection
1752 //   acquires the CGC_lock
1753 //   sets _foregroundGCShouldWait
1754 //   if _foregroundGCIsActive is set
1755 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1756 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1757 //     and exits the loop.
1758 //   otherwise
1759 //     proceed with that phase of the collection
1760 //     if the phase is a stop-the-world phase,
1761 //       yield the baton once more just before enqueueing
1762 //       the stop-world CMS operation (executed by the VM thread).
1763 //   returns after all phases of the collection are done
1764 //
1765 
1766 void CMSCollector::acquire_control_and_collect(bool full,
1767         bool clear_all_soft_refs) {
1768   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1769   assert(!Thread::current()->is_ConcurrentGC_thread(),
1770          "shouldn't try to acquire control from self!");
1771 
1772   // Start the protocol for acquiring control of the
1773   // collection from the background collector (aka CMS thread).
1774   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1775          "VM thread should have CMS token");
1776   // Remember the possibly interrupted state of an ongoing
1777   // concurrent collection
1778   CollectorState first_state = _collectorState;
1779 
1780   // Signal to a possibly ongoing concurrent collection that
1781   // we want to do a foreground collection.
1782   _foregroundGCIsActive = true;
1783 
1784   // Disable incremental mode during a foreground collection.
1785   ICMSDisabler icms_disabler;
1786 
1787   // release locks and wait for a notify from the background collector
1788   // releasing the locks in only necessary for phases which
1789   // do yields to improve the granularity of the collection.
1790   assert_lock_strong(bitMapLock());
1791   // We need to lock the Free list lock for the space that we are
1792   // currently collecting.
1793   assert(haveFreelistLocks(), "Must be holding free list locks");
1794   bitMapLock()->unlock();
1795   releaseFreelistLocks();
1796   {
1797     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1798     if (_foregroundGCShouldWait) {
1799       // We are going to be waiting for action for the CMS thread;
1800       // it had better not be gone (for instance at shutdown)!
1801       assert(ConcurrentMarkSweepThread::cmst() != NULL,
1802              "CMS thread must be running");
1803       // Wait here until the background collector gives us the go-ahead
1804       ConcurrentMarkSweepThread::clear_CMS_flag(
1805         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1806       // Get a possibly blocked CMS thread going:
1807       //   Note that we set _foregroundGCIsActive true above,
1808       //   without protection of the CGC_lock.
1809       CGC_lock->notify();
1810       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1811              "Possible deadlock");
1812       while (_foregroundGCShouldWait) {
1813         // wait for notification
1814         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1815         // Possibility of delay/starvation here, since CMS token does
1816         // not know to give priority to VM thread? Actually, i think
1817         // there wouldn't be any delay/starvation, but the proof of
1818         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1819       }
1820       ConcurrentMarkSweepThread::set_CMS_flag(
1821         ConcurrentMarkSweepThread::CMS_vm_has_token);
1822     }
1823   }
1824   // The CMS_token is already held.  Get back the other locks.
1825   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1826          "VM thread should have CMS token");
1827   getFreelistLocks();
1828   bitMapLock()->lock_without_safepoint_check();
1829   if (TraceCMSState) {
1830     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1831       INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1832     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
1833   }
1834 
1835   // Check if we need to do a compaction, or if not, whether
1836   // we need to start the mark-sweep from scratch.
1837   bool should_compact    = false;
1838   bool should_start_over = false;
1839   decide_foreground_collection_type(clear_all_soft_refs,
1840     &should_compact, &should_start_over);
1841 
1842 NOT_PRODUCT(
1843   if (RotateCMSCollectionTypes) {
1844     if (_cmsGen->debug_collection_type() ==
1845         ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1846       should_compact = true;
1847     } else if (_cmsGen->debug_collection_type() ==
1848                ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1849       should_compact = false;
1850     }
1851   }
1852 )
1853 
1854   if (PrintGCDetails && first_state > Idling) {
1855     GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1856     if (GCCause::is_user_requested_gc(cause) ||
1857         GCCause::is_serviceability_requested_gc(cause)) {
1858       gclog_or_tty->print(" (concurrent mode interrupted)");
1859     } else {
1860       gclog_or_tty->print(" (concurrent mode failure)");
1861     }
1862   }
1863 
1864   if (should_compact) {
1865     // If the collection is being acquired from the background
1866     // collector, there may be references on the discovered
1867     // references lists that have NULL referents (being those
1868     // that were concurrently cleared by a mutator) or
1869     // that are no longer active (having been enqueued concurrently
1870     // by the mutator).
1871     // Scrub the list of those references because Mark-Sweep-Compact
1872     // code assumes referents are not NULL and that all discovered
1873     // Reference objects are active.
1874     ref_processor()->clean_up_discovered_references();
1875 
1876     do_compaction_work(clear_all_soft_refs);
1877 
1878     // Has the GC time limit been exceeded?
1879     DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1880     size_t max_eden_size = young_gen->max_capacity() -
1881                            young_gen->to()->capacity() -
1882                            young_gen->from()->capacity();
1883     GenCollectedHeap* gch = GenCollectedHeap::heap();
1884     GCCause::Cause gc_cause = gch->gc_cause();
1885     size_policy()->check_gc_overhead_limit(_young_gen->used(),
1886                                            young_gen->eden()->used(),
1887                                            _cmsGen->max_capacity(),
1888                                            max_eden_size,
1889                                            full,
1890                                            gc_cause,
1891                                            gch->collector_policy());
1892   } else {
1893     do_mark_sweep_work(clear_all_soft_refs, first_state,
1894       should_start_over);
1895   }
1896   // Reset the expansion cause, now that we just completed
1897   // a collection cycle.
1898   clear_expansion_cause();
1899   _foregroundGCIsActive = false;
1900   return;
1901 }
1902 
1903 // Resize the perm generation and the tenured generation
1904 // after obtaining the free list locks for the
1905 // two generations.
1906 void CMSCollector::compute_new_size() {
1907   assert_locked_or_safepoint(Heap_lock);
1908   FreelistLocker z(this);
1909   _permGen->compute_new_size();
1910   _cmsGen->compute_new_size();
1911 }
1912 
1913 // A work method used by foreground collection to determine
1914 // what type of collection (compacting or not, continuing or fresh)
1915 // it should do.
1916 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1917 // and CMSCompactWhenClearAllSoftRefs the default in the future
1918 // and do away with the flags after a suitable period.
1919 void CMSCollector::decide_foreground_collection_type(
1920   bool clear_all_soft_refs, bool* should_compact,
1921   bool* should_start_over) {
1922   // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1923   // flag is set, and we have either requested a System.gc() or
1924   // the number of full gc's since the last concurrent cycle
1925   // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1926   // or if an incremental collection has failed
1927   GenCollectedHeap* gch = GenCollectedHeap::heap();
1928   assert(gch->collector_policy()->is_two_generation_policy(),
1929          "You may want to check the correctness of the following");
1930   // Inform cms gen if this was due to partial collection failing.
1931   // The CMS gen may use this fact to determine its expansion policy.
1932   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1933     assert(!_cmsGen->incremental_collection_failed(),
1934            "Should have been noticed, reacted to and cleared");
1935     _cmsGen->set_incremental_collection_failed();
1936   }
1937   *should_compact =
1938     UseCMSCompactAtFullCollection &&
1939     ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1940      GCCause::is_user_requested_gc(gch->gc_cause()) ||
1941      gch->incremental_collection_will_fail(true /* consult_young */));
1942   *should_start_over = false;
1943   if (clear_all_soft_refs && !*should_compact) {
1944     // We are about to do a last ditch collection attempt
1945     // so it would normally make sense to do a compaction
1946     // to reclaim as much space as possible.
1947     if (CMSCompactWhenClearAllSoftRefs) {
1948       // Default: The rationale is that in this case either
1949       // we are past the final marking phase, in which case
1950       // we'd have to start over, or so little has been done
1951       // that there's little point in saving that work. Compaction
1952       // appears to be the sensible choice in either case.
1953       *should_compact = true;
1954     } else {
1955       // We have been asked to clear all soft refs, but not to
1956       // compact. Make sure that we aren't past the final checkpoint
1957       // phase, for that is where we process soft refs. If we are already
1958       // past that phase, we'll need to redo the refs discovery phase and
1959       // if necessary clear soft refs that weren't previously
1960       // cleared. We do so by remembering the phase in which
1961       // we came in, and if we are past the refs processing
1962       // phase, we'll choose to just redo the mark-sweep
1963       // collection from scratch.
1964       if (_collectorState > FinalMarking) {
1965         // We are past the refs processing phase;
1966         // start over and do a fresh synchronous CMS cycle
1967         _collectorState = Resetting; // skip to reset to start new cycle
1968         reset(false /* == !asynch */);
1969         *should_start_over = true;
1970       } // else we can continue a possibly ongoing current cycle
1971     }
1972   }
1973 }
1974 
1975 // A work method used by the foreground collector to do
1976 // a mark-sweep-compact.
1977 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1978   GenCollectedHeap* gch = GenCollectedHeap::heap();
1979   TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1980   if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1981     gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1982       "collections passed to foreground collector", _full_gcs_since_conc_gc);
1983   }
1984 
1985   // Sample collection interval time and reset for collection pause.
1986   if (UseAdaptiveSizePolicy) {
1987     size_policy()->msc_collection_begin();
1988   }
1989 
1990   // Temporarily widen the span of the weak reference processing to
1991   // the entire heap.
1992   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1993   ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1994 
1995   // Temporarily, clear the "is_alive_non_header" field of the
1996   // reference processor.
1997   ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1998 
1999   // Temporarily make reference _processing_ single threaded (non-MT).
2000   ReferenceProcessorMTProcMutator z(ref_processor(), false);
2001 
2002   // Temporarily make refs discovery atomic
2003   ReferenceProcessorAtomicMutator w(ref_processor(), true);
2004 
2005   ref_processor()->set_enqueuing_is_done(false);
2006   ref_processor()->enable_discovery();
2007   ref_processor()->setup_policy(clear_all_soft_refs);
2008   // If an asynchronous collection finishes, the _modUnionTable is
2009   // all clear.  If we are assuming the collection from an asynchronous
2010   // collection, clear the _modUnionTable.
2011   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2012     "_modUnionTable should be clear if the baton was not passed");
2013   _modUnionTable.clear_all();
2014 
2015   // We must adjust the allocation statistics being maintained
2016   // in the free list space. We do so by reading and clearing
2017   // the sweep timer and updating the block flux rate estimates below.
2018   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2019   if (_inter_sweep_timer.is_active()) {
2020     _inter_sweep_timer.stop();
2021     // Note that we do not use this sample to update the _inter_sweep_estimate.
2022     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2023                                             _inter_sweep_estimate.padded_average(),
2024                                             _intra_sweep_estimate.padded_average());
2025   }
2026 
2027   {
2028     TraceCMSMemoryManagerStats();
2029   }
2030   GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2031     ref_processor(), clear_all_soft_refs);
2032   #ifdef ASSERT
2033     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2034     size_t free_size = cms_space->free();
2035     assert(free_size ==
2036            pointer_delta(cms_space->end(), cms_space->compaction_top())
2037            * HeapWordSize,
2038       "All the free space should be compacted into one chunk at top");
2039     assert(cms_space->dictionary()->totalChunkSize(
2040                                       debug_only(cms_space->freelistLock())) == 0 ||
2041            cms_space->totalSizeInIndexedFreeLists() == 0,
2042       "All the free space should be in a single chunk");
2043     size_t num = cms_space->totalCount();
2044     assert((free_size == 0 && num == 0) ||
2045            (free_size > 0  && (num == 1 || num == 2)),
2046          "There should be at most 2 free chunks after compaction");
2047   #endif // ASSERT
2048   _collectorState = Resetting;
2049   assert(_restart_addr == NULL,
2050          "Should have been NULL'd before baton was passed");
2051   reset(false /* == !asynch */);
2052   _cmsGen->reset_after_compaction();
2053   _concurrent_cycles_since_last_unload = 0;
2054 
2055   if (verifying() && !should_unload_classes()) {
2056     perm_gen_verify_bit_map()->clear_all();
2057   }
2058 
2059   // Clear any data recorded in the PLAB chunk arrays.
2060   if (_survivor_plab_array != NULL) {
2061     reset_survivor_plab_arrays();
2062   }
2063 
2064   // Adjust the per-size allocation stats for the next epoch.
2065   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2066   // Restart the "inter sweep timer" for the next epoch.
2067   _inter_sweep_timer.reset();
2068   _inter_sweep_timer.start();
2069 
2070   // Sample collection pause time and reset for collection interval.
2071   if (UseAdaptiveSizePolicy) {
2072     size_policy()->msc_collection_end(gch->gc_cause());
2073   }
2074 
2075   // For a mark-sweep-compact, compute_new_size() will be called
2076   // in the heap's do_collection() method.
2077 }
2078 
2079 // A work method used by the foreground collector to do
2080 // a mark-sweep, after taking over from a possibly on-going
2081 // concurrent mark-sweep collection.
2082 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2083   CollectorState first_state, bool should_start_over) {
2084   if (PrintGC && Verbose) {
2085     gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2086       "collector with count %d",
2087       _full_gcs_since_conc_gc);
2088   }
2089   switch (_collectorState) {
2090     case Idling:
2091       if (first_state == Idling || should_start_over) {
2092         // The background GC was not active, or should
2093         // restarted from scratch;  start the cycle.
2094         _collectorState = InitialMarking;
2095       }
2096       // If first_state was not Idling, then a background GC
2097       // was in progress and has now finished.  No need to do it
2098       // again.  Leave the state as Idling.
2099       break;
2100     case Precleaning:
2101       // In the foreground case don't do the precleaning since
2102       // it is not done concurrently and there is extra work
2103       // required.
2104       _collectorState = FinalMarking;
2105   }
2106   if (PrintGCDetails &&
2107       (_collectorState > Idling ||
2108        !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2109     gclog_or_tty->print(" (concurrent mode failure)");
2110   }
2111   collect_in_foreground(clear_all_soft_refs);
2112 
2113   // For a mark-sweep, compute_new_size() will be called
2114   // in the heap's do_collection() method.
2115 }
2116 
2117 
2118 void CMSCollector::getFreelistLocks() const {
2119   // Get locks for all free lists in all generations that this
2120   // collector is responsible for
2121   _cmsGen->freelistLock()->lock_without_safepoint_check();
2122   _permGen->freelistLock()->lock_without_safepoint_check();
2123 }
2124 
2125 void CMSCollector::releaseFreelistLocks() const {
2126   // Release locks for all free lists in all generations that this
2127   // collector is responsible for
2128   _cmsGen->freelistLock()->unlock();
2129   _permGen->freelistLock()->unlock();
2130 }
2131 
2132 bool CMSCollector::haveFreelistLocks() const {
2133   // Check locks for all free lists in all generations that this
2134   // collector is responsible for
2135   assert_lock_strong(_cmsGen->freelistLock());
2136   assert_lock_strong(_permGen->freelistLock());
2137   PRODUCT_ONLY(ShouldNotReachHere());
2138   return true;
2139 }
2140 
2141 // A utility class that is used by the CMS collector to
2142 // temporarily "release" the foreground collector from its
2143 // usual obligation to wait for the background collector to
2144 // complete an ongoing phase before proceeding.
2145 class ReleaseForegroundGC: public StackObj {
2146  private:
2147   CMSCollector* _c;
2148  public:
2149   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2150     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2151     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2152     // allow a potentially blocked foreground collector to proceed
2153     _c->_foregroundGCShouldWait = false;
2154     if (_c->_foregroundGCIsActive) {
2155       CGC_lock->notify();
2156     }
2157     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2158            "Possible deadlock");
2159   }
2160 
2161   ~ReleaseForegroundGC() {
2162     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2163     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2164     _c->_foregroundGCShouldWait = true;
2165   }
2166 };
2167 
2168 // There are separate collect_in_background and collect_in_foreground because of
2169 // the different locking requirements of the background collector and the
2170 // foreground collector.  There was originally an attempt to share
2171 // one "collect" method between the background collector and the foreground
2172 // collector but the if-then-else required made it cleaner to have
2173 // separate methods.
2174 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2175   assert(Thread::current()->is_ConcurrentGC_thread(),
2176     "A CMS asynchronous collection is only allowed on a CMS thread.");
2177 
2178   GenCollectedHeap* gch = GenCollectedHeap::heap();
2179   {
2180     bool safepoint_check = Mutex::_no_safepoint_check_flag;
2181     MutexLockerEx hl(Heap_lock, safepoint_check);
2182     FreelistLocker fll(this);
2183     MutexLockerEx x(CGC_lock, safepoint_check);
2184     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2185       // The foreground collector is active or we're
2186       // not using asynchronous collections.  Skip this
2187       // background collection.
2188       assert(!_foregroundGCShouldWait, "Should be clear");
2189       return;
2190     } else {
2191       assert(_collectorState == Idling, "Should be idling before start.");
2192       _collectorState = InitialMarking;
2193       // Reset the expansion cause, now that we are about to begin
2194       // a new cycle.
2195       clear_expansion_cause();
2196     }
2197     // Decide if we want to enable class unloading as part of the
2198     // ensuing concurrent GC cycle.
2199     update_should_unload_classes();
2200     _full_gc_requested = false;           // acks all outstanding full gc requests
2201     // Signal that we are about to start a collection
2202     gch->increment_total_full_collections();  // ... starting a collection cycle
2203     _collection_count_start = gch->total_full_collections();
2204   }
2205 
2206   // Used for PrintGC
2207   size_t prev_used;
2208   if (PrintGC && Verbose) {
2209     prev_used = _cmsGen->used(); // XXXPERM
2210   }
2211 
2212   // The change of the collection state is normally done at this level;
2213   // the exceptions are phases that are executed while the world is
2214   // stopped.  For those phases the change of state is done while the
2215   // world is stopped.  For baton passing purposes this allows the
2216   // background collector to finish the phase and change state atomically.
2217   // The foreground collector cannot wait on a phase that is done
2218   // while the world is stopped because the foreground collector already
2219   // has the world stopped and would deadlock.
2220   while (_collectorState != Idling) {
2221     if (TraceCMSState) {
2222       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2223         Thread::current(), _collectorState);
2224     }
2225     // The foreground collector
2226     //   holds the Heap_lock throughout its collection.
2227     //   holds the CMS token (but not the lock)
2228     //     except while it is waiting for the background collector to yield.
2229     //
2230     // The foreground collector should be blocked (not for long)
2231     //   if the background collector is about to start a phase
2232     //   executed with world stopped.  If the background
2233     //   collector has already started such a phase, the
2234     //   foreground collector is blocked waiting for the
2235     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
2236     //   are executed in the VM thread.
2237     //
2238     // The locking order is
2239     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
2240     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
2241     //   CMS token  (claimed in
2242     //                stop_world_and_do() -->
2243     //                  safepoint_synchronize() -->
2244     //                    CMSThread::synchronize())
2245 
2246     {
2247       // Check if the FG collector wants us to yield.
2248       CMSTokenSync x(true); // is cms thread
2249       if (waitForForegroundGC()) {
2250         // We yielded to a foreground GC, nothing more to be
2251         // done this round.
2252         assert(_foregroundGCShouldWait == false, "We set it to false in "
2253                "waitForForegroundGC()");
2254         if (TraceCMSState) {
2255           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2256             " exiting collection CMS state %d",
2257             Thread::current(), _collectorState);
2258         }
2259         return;
2260       } else {
2261         // The background collector can run but check to see if the
2262         // foreground collector has done a collection while the
2263         // background collector was waiting to get the CGC_lock
2264         // above.  If yes, break so that _foregroundGCShouldWait
2265         // is cleared before returning.
2266         if (_collectorState == Idling) {
2267           break;
2268         }
2269       }
2270     }
2271 
2272     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2273       "should be waiting");
2274 
2275     switch (_collectorState) {
2276       case InitialMarking:
2277         {
2278           ReleaseForegroundGC x(this);
2279           stats().record_cms_begin();
2280 
2281           VM_CMS_Initial_Mark initial_mark_op(this);
2282           VMThread::execute(&initial_mark_op);
2283         }
2284         // The collector state may be any legal state at this point
2285         // since the background collector may have yielded to the
2286         // foreground collector.
2287         break;
2288       case Marking:
2289         // initial marking in checkpointRootsInitialWork has been completed
2290         if (markFromRoots(true)) { // we were successful
2291           assert(_collectorState == Precleaning, "Collector state should "
2292             "have changed");
2293         } else {
2294           assert(_foregroundGCIsActive, "Internal state inconsistency");
2295         }
2296         break;
2297       case Precleaning:
2298         if (UseAdaptiveSizePolicy) {
2299           size_policy()->concurrent_precleaning_begin();
2300         }
2301         // marking from roots in markFromRoots has been completed
2302         preclean();
2303         if (UseAdaptiveSizePolicy) {
2304           size_policy()->concurrent_precleaning_end();
2305         }
2306         assert(_collectorState == AbortablePreclean ||
2307                _collectorState == FinalMarking,
2308                "Collector state should have changed");
2309         break;
2310       case AbortablePreclean:
2311         if (UseAdaptiveSizePolicy) {
2312         size_policy()->concurrent_phases_resume();
2313         }
2314         abortable_preclean();
2315         if (UseAdaptiveSizePolicy) {
2316           size_policy()->concurrent_precleaning_end();
2317         }
2318         assert(_collectorState == FinalMarking, "Collector state should "
2319           "have changed");
2320         break;
2321       case FinalMarking:
2322         {
2323           ReleaseForegroundGC x(this);
2324 
2325           VM_CMS_Final_Remark final_remark_op(this);
2326           VMThread::execute(&final_remark_op);
2327         }
2328         assert(_foregroundGCShouldWait, "block post-condition");
2329         break;
2330       case Sweeping:
2331         if (UseAdaptiveSizePolicy) {
2332           size_policy()->concurrent_sweeping_begin();
2333         }
2334         // final marking in checkpointRootsFinal has been completed
2335         sweep(true);
2336         assert(_collectorState == Resizing, "Collector state change "
2337           "to Resizing must be done under the free_list_lock");
2338         _full_gcs_since_conc_gc = 0;
2339 
2340         // Stop the timers for adaptive size policy for the concurrent phases
2341         if (UseAdaptiveSizePolicy) {
2342           size_policy()->concurrent_sweeping_end();
2343           size_policy()->concurrent_phases_end(gch->gc_cause(),
2344                                              gch->prev_gen(_cmsGen)->capacity(),
2345                                              _cmsGen->free());
2346         }
2347 
2348       case Resizing: {
2349         // Sweeping has been completed...
2350         // At this point the background collection has completed.
2351         // Don't move the call to compute_new_size() down
2352         // into code that might be executed if the background
2353         // collection was preempted.
2354         {
2355           ReleaseForegroundGC x(this);   // unblock FG collection
2356           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
2357           CMSTokenSync        z(true);   // not strictly needed.
2358           if (_collectorState == Resizing) {
2359             compute_new_size();
2360             _collectorState = Resetting;
2361           } else {
2362             assert(_collectorState == Idling, "The state should only change"
2363                    " because the foreground collector has finished the collection");
2364           }
2365         }
2366         break;
2367       }
2368       case Resetting:
2369         // CMS heap resizing has been completed
2370         reset(true);
2371         assert(_collectorState == Idling, "Collector state should "
2372           "have changed");
2373         stats().record_cms_end();
2374         // Don't move the concurrent_phases_end() and compute_new_size()
2375         // calls to here because a preempted background collection
2376         // has it's state set to "Resetting".
2377         break;
2378       case Idling:
2379       default:
2380         ShouldNotReachHere();
2381         break;
2382     }
2383     if (TraceCMSState) {
2384       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
2385         Thread::current(), _collectorState);
2386     }
2387     assert(_foregroundGCShouldWait, "block post-condition");
2388   }
2389 
2390   // Should this be in gc_epilogue?
2391   collector_policy()->counters()->update_counters();
2392 
2393   {
2394     // Clear _foregroundGCShouldWait and, in the event that the
2395     // foreground collector is waiting, notify it, before
2396     // returning.
2397     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2398     _foregroundGCShouldWait = false;
2399     if (_foregroundGCIsActive) {
2400       CGC_lock->notify();
2401     }
2402     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2403            "Possible deadlock");
2404   }
2405   if (TraceCMSState) {
2406     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2407       " exiting collection CMS state %d",
2408       Thread::current(), _collectorState);
2409   }
2410   if (PrintGC && Verbose) {
2411     _cmsGen->print_heap_change(prev_used);
2412   }
2413 }
2414 
2415 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2416   assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2417          "Foreground collector should be waiting, not executing");
2418   assert(Thread::current()->is_VM_thread(), "A foreground collection"
2419     "may only be done by the VM Thread with the world stopped");
2420   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2421          "VM thread should have CMS token");
2422 
2423   NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2424     true, gclog_or_tty);)
2425   if (UseAdaptiveSizePolicy) {
2426     size_policy()->ms_collection_begin();
2427   }
2428   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2429 
2430   HandleMark hm;  // Discard invalid handles created during verification
2431 
2432   if (VerifyBeforeGC &&
2433       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2434     Universe::verify(true);
2435   }
2436 
2437   // Snapshot the soft reference policy to be used in this collection cycle.
2438   ref_processor()->setup_policy(clear_all_soft_refs);
2439 
2440   bool init_mark_was_synchronous = false; // until proven otherwise
2441   while (_collectorState != Idling) {
2442     if (TraceCMSState) {
2443       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2444         Thread::current(), _collectorState);
2445     }
2446     switch (_collectorState) {
2447       case InitialMarking:
2448         init_mark_was_synchronous = true;  // fact to be exploited in re-mark
2449         checkpointRootsInitial(false);
2450         assert(_collectorState == Marking, "Collector state should have changed"
2451           " within checkpointRootsInitial()");
2452         break;
2453       case Marking:
2454         // initial marking in checkpointRootsInitialWork has been completed
2455         if (VerifyDuringGC &&
2456             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2457           gclog_or_tty->print("Verify before initial mark: ");
2458           Universe::verify(true);
2459         }
2460         {
2461           bool res = markFromRoots(false);
2462           assert(res && _collectorState == FinalMarking, "Collector state should "
2463             "have changed");
2464           break;
2465         }
2466       case FinalMarking:
2467         if (VerifyDuringGC &&
2468             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2469           gclog_or_tty->print("Verify before re-mark: ");
2470           Universe::verify(true);
2471         }
2472         checkpointRootsFinal(false, clear_all_soft_refs,
2473                              init_mark_was_synchronous);
2474         assert(_collectorState == Sweeping, "Collector state should not "
2475           "have changed within checkpointRootsFinal()");
2476         break;
2477       case Sweeping:
2478         // final marking in checkpointRootsFinal has been completed
2479         if (VerifyDuringGC &&
2480             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2481           gclog_or_tty->print("Verify before sweep: ");
2482           Universe::verify(true);
2483         }
2484         sweep(false);
2485         assert(_collectorState == Resizing, "Incorrect state");
2486         break;
2487       case Resizing: {
2488         // Sweeping has been completed; the actual resize in this case
2489         // is done separately; nothing to be done in this state.
2490         _collectorState = Resetting;
2491         break;
2492       }
2493       case Resetting:
2494         // The heap has been resized.
2495         if (VerifyDuringGC &&
2496             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2497           gclog_or_tty->print("Verify before reset: ");
2498           Universe::verify(true);
2499         }
2500         reset(false);
2501         assert(_collectorState == Idling, "Collector state should "
2502           "have changed");
2503         break;
2504       case Precleaning:
2505       case AbortablePreclean:
2506         // Elide the preclean phase
2507         _collectorState = FinalMarking;
2508         break;
2509       default:
2510         ShouldNotReachHere();
2511     }
2512     if (TraceCMSState) {
2513       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
2514         Thread::current(), _collectorState);
2515     }
2516   }
2517 
2518   if (UseAdaptiveSizePolicy) {
2519     GenCollectedHeap* gch = GenCollectedHeap::heap();
2520     size_policy()->ms_collection_end(gch->gc_cause());
2521   }
2522 
2523   if (VerifyAfterGC &&
2524       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2525     Universe::verify(true);
2526   }
2527   if (TraceCMSState) {
2528     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2529       " exiting collection CMS state %d",
2530       Thread::current(), _collectorState);
2531   }
2532 }
2533 
2534 bool CMSCollector::waitForForegroundGC() {
2535   bool res = false;
2536   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2537          "CMS thread should have CMS token");
2538   // Block the foreground collector until the
2539   // background collectors decides whether to
2540   // yield.
2541   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2542   _foregroundGCShouldWait = true;
2543   if (_foregroundGCIsActive) {
2544     // The background collector yields to the
2545     // foreground collector and returns a value
2546     // indicating that it has yielded.  The foreground
2547     // collector can proceed.
2548     res = true;
2549     _foregroundGCShouldWait = false;
2550     ConcurrentMarkSweepThread::clear_CMS_flag(
2551       ConcurrentMarkSweepThread::CMS_cms_has_token);
2552     ConcurrentMarkSweepThread::set_CMS_flag(
2553       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2554     // Get a possibly blocked foreground thread going
2555     CGC_lock->notify();
2556     if (TraceCMSState) {
2557       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2558         Thread::current(), _collectorState);
2559     }
2560     while (_foregroundGCIsActive) {
2561       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2562     }
2563     ConcurrentMarkSweepThread::set_CMS_flag(
2564       ConcurrentMarkSweepThread::CMS_cms_has_token);
2565     ConcurrentMarkSweepThread::clear_CMS_flag(
2566       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2567   }
2568   if (TraceCMSState) {
2569     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2570       Thread::current(), _collectorState);
2571   }
2572   return res;
2573 }
2574 
2575 // Because of the need to lock the free lists and other structures in
2576 // the collector, common to all the generations that the collector is
2577 // collecting, we need the gc_prologues of individual CMS generations
2578 // delegate to their collector. It may have been simpler had the
2579 // current infrastructure allowed one to call a prologue on a
2580 // collector. In the absence of that we have the generation's
2581 // prologue delegate to the collector, which delegates back
2582 // some "local" work to a worker method in the individual generations
2583 // that it's responsible for collecting, while itself doing any
2584 // work common to all generations it's responsible for. A similar
2585 // comment applies to the  gc_epilogue()'s.
2586 // The role of the varaible _between_prologue_and_epilogue is to
2587 // enforce the invocation protocol.
2588 void CMSCollector::gc_prologue(bool full) {
2589   // Call gc_prologue_work() for each CMSGen and PermGen that
2590   // we are responsible for.
2591 
2592   // The following locking discipline assumes that we are only called
2593   // when the world is stopped.
2594   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2595 
2596   // The CMSCollector prologue must call the gc_prologues for the
2597   // "generations" (including PermGen if any) that it's responsible
2598   // for.
2599 
2600   assert(   Thread::current()->is_VM_thread()
2601          || (   CMSScavengeBeforeRemark
2602              && Thread::current()->is_ConcurrentGC_thread()),
2603          "Incorrect thread type for prologue execution");
2604 
2605   if (_between_prologue_and_epilogue) {
2606     // We have already been invoked; this is a gc_prologue delegation
2607     // from yet another CMS generation that we are responsible for, just
2608     // ignore it since all relevant work has already been done.
2609     return;
2610   }
2611 
2612   // set a bit saying prologue has been called; cleared in epilogue
2613   _between_prologue_and_epilogue = true;
2614   // Claim locks for common data structures, then call gc_prologue_work()
2615   // for each CMSGen and PermGen that we are responsible for.
2616 
2617   getFreelistLocks();   // gets free list locks on constituent spaces
2618   bitMapLock()->lock_without_safepoint_check();
2619 
2620   // Should call gc_prologue_work() for all cms gens we are responsible for
2621   bool registerClosure =    _collectorState >= Marking
2622                          && _collectorState < Sweeping;
2623   ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2624                                                &_modUnionClosurePar
2625                                                : &_modUnionClosure;
2626   _cmsGen->gc_prologue_work(full, registerClosure, muc);
2627   _permGen->gc_prologue_work(full, registerClosure, muc);
2628 
2629   if (!full) {
2630     stats().record_gc0_begin();
2631   }
2632 }
2633 
2634 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2635   // Delegate to CMScollector which knows how to coordinate between
2636   // this and any other CMS generations that it is responsible for
2637   // collecting.
2638   collector()->gc_prologue(full);
2639 }
2640 
2641 // This is a "private" interface for use by this generation's CMSCollector.
2642 // Not to be called directly by any other entity (for instance,
2643 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2644 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2645   bool registerClosure, ModUnionClosure* modUnionClosure) {
2646   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2647   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2648     "Should be NULL");
2649   if (registerClosure) {
2650     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2651   }
2652   cmsSpace()->gc_prologue();
2653   // Clear stat counters
2654   NOT_PRODUCT(
2655     assert(_numObjectsPromoted == 0, "check");
2656     assert(_numWordsPromoted   == 0, "check");
2657     if (Verbose && PrintGC) {
2658       gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2659                           SIZE_FORMAT" bytes concurrently",
2660       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2661     }
2662     _numObjectsAllocated = 0;
2663     _numWordsAllocated   = 0;
2664   )
2665 }
2666 
2667 void CMSCollector::gc_epilogue(bool full) {
2668   // The following locking discipline assumes that we are only called
2669   // when the world is stopped.
2670   assert(SafepointSynchronize::is_at_safepoint(),
2671          "world is stopped assumption");
2672 
2673   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2674   // if linear allocation blocks need to be appropriately marked to allow the
2675   // the blocks to be parsable. We also check here whether we need to nudge the
2676   // CMS collector thread to start a new cycle (if it's not already active).
2677   assert(   Thread::current()->is_VM_thread()
2678          || (   CMSScavengeBeforeRemark
2679              && Thread::current()->is_ConcurrentGC_thread()),
2680          "Incorrect thread type for epilogue execution");
2681 
2682   if (!_between_prologue_and_epilogue) {
2683     // We have already been invoked; this is a gc_epilogue delegation
2684     // from yet another CMS generation that we are responsible for, just
2685     // ignore it since all relevant work has already been done.
2686     return;
2687   }
2688   assert(haveFreelistLocks(), "must have freelist locks");
2689   assert_lock_strong(bitMapLock());
2690 
2691   _cmsGen->gc_epilogue_work(full);
2692   _permGen->gc_epilogue_work(full);
2693 
2694   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2695     // in case sampling was not already enabled, enable it
2696     _start_sampling = true;
2697   }
2698   // reset _eden_chunk_array so sampling starts afresh
2699   _eden_chunk_index = 0;
2700 
2701   size_t cms_used   = _cmsGen->cmsSpace()->used();
2702   size_t perm_used  = _permGen->cmsSpace()->used();
2703 
2704   // update performance counters - this uses a special version of
2705   // update_counters() that allows the utilization to be passed as a
2706   // parameter, avoiding multiple calls to used().
2707   //
2708   _cmsGen->update_counters(cms_used);
2709   _permGen->update_counters(perm_used);
2710 
2711   if (CMSIncrementalMode) {
2712     icms_update_allocation_limits();
2713   }
2714 
2715   bitMapLock()->unlock();
2716   releaseFreelistLocks();
2717 
2718   _between_prologue_and_epilogue = false;  // ready for next cycle
2719 }
2720 
2721 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2722   collector()->gc_epilogue(full);
2723 
2724   // Also reset promotion tracking in par gc thread states.
2725   if (CollectedHeap::use_parallel_gc_threads()) {
2726     for (uint i = 0; i < ParallelGCThreads; i++) {
2727       _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2728     }
2729   }
2730 }
2731 
2732 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2733   assert(!incremental_collection_failed(), "Should have been cleared");
2734   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2735   cmsSpace()->gc_epilogue();
2736     // Print stat counters
2737   NOT_PRODUCT(
2738     assert(_numObjectsAllocated == 0, "check");
2739     assert(_numWordsAllocated == 0, "check");
2740     if (Verbose && PrintGC) {
2741       gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2742                           SIZE_FORMAT" bytes",
2743                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2744     }
2745     _numObjectsPromoted = 0;
2746     _numWordsPromoted   = 0;
2747   )
2748 
2749   if (PrintGC && Verbose) {
2750     // Call down the chain in contiguous_available needs the freelistLock
2751     // so print this out before releasing the freeListLock.
2752     gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2753                         contiguous_available());
2754   }
2755 }
2756 
2757 #ifndef PRODUCT
2758 bool CMSCollector::have_cms_token() {
2759   Thread* thr = Thread::current();
2760   if (thr->is_VM_thread()) {
2761     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2762   } else if (thr->is_ConcurrentGC_thread()) {
2763     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2764   } else if (thr->is_GC_task_thread()) {
2765     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2766            ParGCRareEvent_lock->owned_by_self();
2767   }
2768   return false;
2769 }
2770 #endif
2771 
2772 // Check reachability of the given heap address in CMS generation,
2773 // treating all other generations as roots.
2774 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2775   // We could "guarantee" below, rather than assert, but i'll
2776   // leave these as "asserts" so that an adventurous debugger
2777   // could try this in the product build provided some subset of
2778   // the conditions were met, provided they were intersted in the
2779   // results and knew that the computation below wouldn't interfere
2780   // with other concurrent computations mutating the structures
2781   // being read or written.
2782   assert(SafepointSynchronize::is_at_safepoint(),
2783          "Else mutations in object graph will make answer suspect");
2784   assert(have_cms_token(), "Should hold cms token");
2785   assert(haveFreelistLocks(), "must hold free list locks");
2786   assert_lock_strong(bitMapLock());
2787 
2788   // Clear the marking bit map array before starting, but, just
2789   // for kicks, first report if the given address is already marked
2790   gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2791                 _markBitMap.isMarked(addr) ? "" : " not");
2792 
2793   if (verify_after_remark()) {
2794     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2795     bool result = verification_mark_bm()->isMarked(addr);
2796     gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2797                            result ? "IS" : "is NOT");
2798     return result;
2799   } else {
2800     gclog_or_tty->print_cr("Could not compute result");
2801     return false;
2802   }
2803 }
2804 
2805 ////////////////////////////////////////////////////////
2806 // CMS Verification Support
2807 ////////////////////////////////////////////////////////
2808 // Following the remark phase, the following invariant
2809 // should hold -- each object in the CMS heap which is
2810 // marked in markBitMap() should be marked in the verification_mark_bm().
2811 
2812 class VerifyMarkedClosure: public BitMapClosure {
2813   CMSBitMap* _marks;
2814   bool       _failed;
2815 
2816  public:
2817   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2818 
2819   bool do_bit(size_t offset) {
2820     HeapWord* addr = _marks->offsetToHeapWord(offset);
2821     if (!_marks->isMarked(addr)) {
2822       oop(addr)->print_on(gclog_or_tty);
2823       gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2824       _failed = true;
2825     }
2826     return true;
2827   }
2828 
2829   bool failed() { return _failed; }
2830 };
2831 
2832 bool CMSCollector::verify_after_remark() {
2833   gclog_or_tty->print(" [Verifying CMS Marking... ");
2834   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2835   static bool init = false;
2836 
2837   assert(SafepointSynchronize::is_at_safepoint(),
2838          "Else mutations in object graph will make answer suspect");
2839   assert(have_cms_token(),
2840          "Else there may be mutual interference in use of "
2841          " verification data structures");
2842   assert(_collectorState > Marking && _collectorState <= Sweeping,
2843          "Else marking info checked here may be obsolete");
2844   assert(haveFreelistLocks(), "must hold free list locks");
2845   assert_lock_strong(bitMapLock());
2846 
2847 
2848   // Allocate marking bit map if not already allocated
2849   if (!init) { // first time
2850     if (!verification_mark_bm()->allocate(_span)) {
2851       return false;
2852     }
2853     init = true;
2854   }
2855 
2856   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2857 
2858   // Turn off refs discovery -- so we will be tracing through refs.
2859   // This is as intended, because by this time
2860   // GC must already have cleared any refs that need to be cleared,
2861   // and traced those that need to be marked; moreover,
2862   // the marking done here is not going to intefere in any
2863   // way with the marking information used by GC.
2864   NoRefDiscovery no_discovery(ref_processor());
2865 
2866   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2867 
2868   // Clear any marks from a previous round
2869   verification_mark_bm()->clear_all();
2870   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2871   verify_work_stacks_empty();
2872 
2873   GenCollectedHeap* gch = GenCollectedHeap::heap();
2874   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2875   // Update the saved marks which may affect the root scans.
2876   gch->save_marks();
2877 
2878   if (CMSRemarkVerifyVariant == 1) {
2879     // In this first variant of verification, we complete
2880     // all marking, then check if the new marks-verctor is
2881     // a subset of the CMS marks-vector.
2882     verify_after_remark_work_1();
2883   } else if (CMSRemarkVerifyVariant == 2) {
2884     // In this second variant of verification, we flag an error
2885     // (i.e. an object reachable in the new marks-vector not reachable
2886     // in the CMS marks-vector) immediately, also indicating the
2887     // identify of an object (A) that references the unmarked object (B) --
2888     // presumably, a mutation to A failed to be picked up by preclean/remark?
2889     verify_after_remark_work_2();
2890   } else {
2891     warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2892             CMSRemarkVerifyVariant);
2893   }
2894   gclog_or_tty->print(" done] ");
2895   return true;
2896 }
2897 
2898 void CMSCollector::verify_after_remark_work_1() {
2899   ResourceMark rm;
2900   HandleMark  hm;
2901   GenCollectedHeap* gch = GenCollectedHeap::heap();
2902 
2903   // Mark from roots one level into CMS
2904   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2905   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2906 
2907   gch->gen_process_strong_roots(_cmsGen->level(),
2908                                 true,   // younger gens are roots
2909                                 true,   // activate StrongRootsScope
2910                                 true,   // collecting perm gen
2911                                 SharedHeap::ScanningOption(roots_scanning_options()),
2912                                 &notOlder,
2913                                 true,   // walk code active on stacks
2914                                 NULL);
2915 
2916   // Now mark from the roots
2917   assert(_revisitStack.isEmpty(), "Should be empty");
2918   MarkFromRootsClosure markFromRootsClosure(this, _span,
2919     verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2920     false /* don't yield */, true /* verifying */);
2921   assert(_restart_addr == NULL, "Expected pre-condition");
2922   verification_mark_bm()->iterate(&markFromRootsClosure);
2923   while (_restart_addr != NULL) {
2924     // Deal with stack overflow: by restarting at the indicated
2925     // address.
2926     HeapWord* ra = _restart_addr;
2927     markFromRootsClosure.reset(ra);
2928     _restart_addr = NULL;
2929     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2930   }
2931   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2932   verify_work_stacks_empty();
2933   // Should reset the revisit stack above, since no class tree
2934   // surgery is forthcoming.
2935   _revisitStack.reset(); // throwing away all contents
2936 
2937   // Marking completed -- now verify that each bit marked in
2938   // verification_mark_bm() is also marked in markBitMap(); flag all
2939   // errors by printing corresponding objects.
2940   VerifyMarkedClosure vcl(markBitMap());
2941   verification_mark_bm()->iterate(&vcl);
2942   if (vcl.failed()) {
2943     gclog_or_tty->print("Verification failed");
2944     Universe::heap()->print_on(gclog_or_tty);
2945     fatal("CMS: failed marking verification after remark");
2946   }
2947 }
2948 
2949 void CMSCollector::verify_after_remark_work_2() {
2950   ResourceMark rm;
2951   HandleMark  hm;
2952   GenCollectedHeap* gch = GenCollectedHeap::heap();
2953 
2954   // Mark from roots one level into CMS
2955   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2956                                      markBitMap());
2957   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2958   gch->gen_process_strong_roots(_cmsGen->level(),
2959                                 true,   // younger gens are roots
2960                                 true,   // activate StrongRootsScope
2961                                 true,   // collecting perm gen
2962                                 SharedHeap::ScanningOption(roots_scanning_options()),
2963                                 &notOlder,
2964                                 true,   // walk code active on stacks
2965                                 NULL);
2966 
2967   // Now mark from the roots
2968   assert(_revisitStack.isEmpty(), "Should be empty");
2969   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2970     verification_mark_bm(), markBitMap(), verification_mark_stack());
2971   assert(_restart_addr == NULL, "Expected pre-condition");
2972   verification_mark_bm()->iterate(&markFromRootsClosure);
2973   while (_restart_addr != NULL) {
2974     // Deal with stack overflow: by restarting at the indicated
2975     // address.
2976     HeapWord* ra = _restart_addr;
2977     markFromRootsClosure.reset(ra);
2978     _restart_addr = NULL;
2979     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2980   }
2981   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2982   verify_work_stacks_empty();
2983   // Should reset the revisit stack above, since no class tree
2984   // surgery is forthcoming.
2985   _revisitStack.reset(); // throwing away all contents
2986 
2987   // Marking completed -- now verify that each bit marked in
2988   // verification_mark_bm() is also marked in markBitMap(); flag all
2989   // errors by printing corresponding objects.
2990   VerifyMarkedClosure vcl(markBitMap());
2991   verification_mark_bm()->iterate(&vcl);
2992   assert(!vcl.failed(), "Else verification above should not have succeeded");
2993 }
2994 
2995 void ConcurrentMarkSweepGeneration::save_marks() {
2996   // delegate to CMS space
2997   cmsSpace()->save_marks();
2998   for (uint i = 0; i < ParallelGCThreads; i++) {
2999     _par_gc_thread_states[i]->promo.startTrackingPromotions();
3000   }
3001 }
3002 
3003 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3004   return cmsSpace()->no_allocs_since_save_marks();
3005 }
3006 
3007 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
3008                                                                 \
3009 void ConcurrentMarkSweepGeneration::                            \
3010 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
3011   cl->set_generation(this);                                     \
3012   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
3013   cl->reset_generation();                                       \
3014   save_marks();                                                 \
3015 }
3016 
3017 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3018 
3019 void
3020 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3021 {
3022   // Not currently implemented; need to do the following. -- ysr.
3023   // dld -- I think that is used for some sort of allocation profiler.  So it
3024   // really means the objects allocated by the mutator since the last
3025   // GC.  We could potentially implement this cheaply by recording only
3026   // the direct allocations in a side data structure.
3027   //
3028   // I think we probably ought not to be required to support these
3029   // iterations at any arbitrary point; I think there ought to be some
3030   // call to enable/disable allocation profiling in a generation/space,
3031   // and the iterator ought to return the objects allocated in the
3032   // gen/space since the enable call, or the last iterator call (which
3033   // will probably be at a GC.)  That way, for gens like CM&S that would
3034   // require some extra data structure to support this, we only pay the
3035   // cost when it's in use...
3036   cmsSpace()->object_iterate_since_last_GC(blk);
3037 }
3038 
3039 void
3040 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3041   cl->set_generation(this);
3042   younger_refs_in_space_iterate(_cmsSpace, cl);
3043   cl->reset_generation();
3044 }
3045 
3046 void
3047 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3048   if (freelistLock()->owned_by_self()) {
3049     Generation::oop_iterate(mr, cl);
3050   } else {
3051     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3052     Generation::oop_iterate(mr, cl);
3053   }
3054 }
3055 
3056 void
3057 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3058   if (freelistLock()->owned_by_self()) {
3059     Generation::oop_iterate(cl);
3060   } else {
3061     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3062     Generation::oop_iterate(cl);
3063   }
3064 }
3065 
3066 void
3067 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3068   if (freelistLock()->owned_by_self()) {
3069     Generation::object_iterate(cl);
3070   } else {
3071     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3072     Generation::object_iterate(cl);
3073   }
3074 }
3075 
3076 void
3077 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3078   if (freelistLock()->owned_by_self()) {
3079     Generation::safe_object_iterate(cl);
3080   } else {
3081     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3082     Generation::safe_object_iterate(cl);
3083   }
3084 }
3085 
3086 void
3087 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3088 }
3089 
3090 void
3091 ConcurrentMarkSweepGeneration::post_compact() {
3092 }
3093 
3094 void
3095 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3096   // Fix the linear allocation blocks to look like free blocks.
3097 
3098   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3099   // are not called when the heap is verified during universe initialization and
3100   // at vm shutdown.
3101   if (freelistLock()->owned_by_self()) {
3102     cmsSpace()->prepare_for_verify();
3103   } else {
3104     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3105     cmsSpace()->prepare_for_verify();
3106   }
3107 }
3108 
3109 void
3110 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3111   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3112   // are not called when the heap is verified during universe initialization and
3113   // at vm shutdown.
3114   if (freelistLock()->owned_by_self()) {
3115     cmsSpace()->verify(false /* ignored */);
3116   } else {
3117     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3118     cmsSpace()->verify(false /* ignored */);
3119   }
3120 }
3121 
3122 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3123   _cmsGen->verify(allow_dirty);
3124   _permGen->verify(allow_dirty);
3125 }
3126 
3127 #ifndef PRODUCT
3128 bool CMSCollector::overflow_list_is_empty() const {
3129   assert(_num_par_pushes >= 0, "Inconsistency");
3130   if (_overflow_list == NULL) {
3131     assert(_num_par_pushes == 0, "Inconsistency");
3132   }
3133   return _overflow_list == NULL;
3134 }
3135 
3136 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3137 // merely consolidate assertion checks that appear to occur together frequently.
3138 void CMSCollector::verify_work_stacks_empty() const {
3139   assert(_markStack.isEmpty(), "Marking stack should be empty");
3140   assert(overflow_list_is_empty(), "Overflow list should be empty");
3141 }
3142 
3143 void CMSCollector::verify_overflow_empty() const {
3144   assert(overflow_list_is_empty(), "Overflow list should be empty");
3145   assert(no_preserved_marks(), "No preserved marks");
3146 }
3147 #endif // PRODUCT
3148 
3149 // Decide if we want to enable class unloading as part of the
3150 // ensuing concurrent GC cycle. We will collect the perm gen and
3151 // unload classes if it's the case that:
3152 // (1) an explicit gc request has been made and the flag
3153 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3154 // (2) (a) class unloading is enabled at the command line, and
3155 //     (b) (i)   perm gen threshold has been crossed, or
3156 //         (ii)  old gen is getting really full, or
3157 //         (iii) the previous N CMS collections did not collect the
3158 //               perm gen
3159 // NOTE: Provided there is no change in the state of the heap between
3160 // calls to this method, it should have idempotent results. Moreover,
3161 // its results should be monotonically increasing (i.e. going from 0 to 1,
3162 // but not 1 to 0) between successive calls between which the heap was
3163 // not collected. For the implementation below, it must thus rely on
3164 // the property that concurrent_cycles_since_last_unload()
3165 // will not decrease unless a collection cycle happened and that
3166 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3167 // themselves also monotonic in that sense. See check_monotonicity()
3168 // below.
3169 bool CMSCollector::update_should_unload_classes() {
3170   _should_unload_classes = false;
3171   // Condition 1 above
3172   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3173     _should_unload_classes = true;
3174   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3175     // Disjuncts 2.b.(i,ii,iii) above
3176     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3177                               CMSClassUnloadingMaxInterval)
3178                            || _permGen->should_concurrent_collect()
3179                            || _cmsGen->is_too_full();
3180   }
3181   return _should_unload_classes;
3182 }
3183 
3184 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3185   bool res = should_concurrent_collect();
3186   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3187   return res;
3188 }
3189 
3190 void CMSCollector::setup_cms_unloading_and_verification_state() {
3191   const  bool should_verify =    VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3192                              || VerifyBeforeExit;
3193   const  int  rso           =    SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3194                              |   SharedHeap::SO_CodeCache;
3195 
3196   if (should_unload_classes()) {   // Should unload classes this cycle
3197     remove_root_scanning_option(rso);  // Shrink the root set appropriately
3198     set_verifying(should_verify);    // Set verification state for this cycle
3199     return;                            // Nothing else needs to be done at this time
3200   }
3201 
3202   // Not unloading classes this cycle
3203   assert(!should_unload_classes(), "Inconsitency!");
3204   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3205     // We were not verifying, or we _were_ unloading classes in the last cycle,
3206     // AND some verification options are enabled this cycle; in this case,
3207     // we must make sure that the deadness map is allocated if not already so,
3208     // and cleared (if already allocated previously --
3209     // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3210     if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3211       if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3212         warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3213                 "permanent generation verification disabled");
3214         return;  // Note that we leave verification disabled, so we'll retry this
3215                  // allocation next cycle. We _could_ remember this failure
3216                  // and skip further attempts and permanently disable verification
3217                  // attempts if that is considered more desirable.
3218       }
3219       assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3220               "_perm_gen_ver_bit_map inconsistency?");
3221     } else {
3222       perm_gen_verify_bit_map()->clear_all();
3223     }
3224     // Include symbols, strings and code cache elements to prevent their resurrection.
3225     add_root_scanning_option(rso);
3226     set_verifying(true);
3227   } else if (verifying() && !should_verify) {
3228     // We were verifying, but some verification flags got disabled.
3229     set_verifying(false);
3230     // Exclude symbols, strings and code cache elements from root scanning to
3231     // reduce IM and RM pauses.
3232     remove_root_scanning_option(rso);
3233   }
3234 }
3235 
3236 
3237 #ifndef PRODUCT
3238 HeapWord* CMSCollector::block_start(const void* p) const {
3239   const HeapWord* addr = (HeapWord*)p;
3240   if (_span.contains(p)) {
3241     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3242       return _cmsGen->cmsSpace()->block_start(p);
3243     } else {
3244       assert(_permGen->cmsSpace()->is_in_reserved(addr),
3245              "Inconsistent _span?");
3246       return _permGen->cmsSpace()->block_start(p);
3247     }
3248   }
3249   return NULL;
3250 }
3251 #endif
3252 
3253 HeapWord*
3254 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3255                                                    bool   tlab,
3256                                                    bool   parallel) {
3257   CMSSynchronousYieldRequest yr;
3258   assert(!tlab, "Can't deal with TLAB allocation");
3259   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3260   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3261     CMSExpansionCause::_satisfy_allocation);
3262   if (GCExpandToAllocateDelayMillis > 0) {
3263     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3264   }
3265   return have_lock_and_allocate(word_size, tlab);
3266 }
3267 
3268 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3269 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3270 // to CardGeneration and share it...
3271 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3272   return CardGeneration::expand(bytes, expand_bytes);
3273 }
3274 
3275 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3276   CMSExpansionCause::Cause cause)
3277 {
3278 
3279   bool success = expand(bytes, expand_bytes);
3280 
3281   // remember why we expanded; this information is used
3282   // by shouldConcurrentCollect() when making decisions on whether to start
3283   // a new CMS cycle.
3284   if (success) {
3285     set_expansion_cause(cause);
3286     if (PrintGCDetails && Verbose) {
3287       gclog_or_tty->print_cr("Expanded CMS gen for %s",
3288         CMSExpansionCause::to_string(cause));
3289     }
3290   }
3291 }
3292 
3293 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3294   HeapWord* res = NULL;
3295   MutexLocker x(ParGCRareEvent_lock);
3296   while (true) {
3297     // Expansion by some other thread might make alloc OK now:
3298     res = ps->lab.alloc(word_sz);
3299     if (res != NULL) return res;
3300     // If there's not enough expansion space available, give up.
3301     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3302       return NULL;
3303     }
3304     // Otherwise, we try expansion.
3305     expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3306       CMSExpansionCause::_allocate_par_lab);
3307     // Now go around the loop and try alloc again;
3308     // A competing par_promote might beat us to the expansion space,
3309     // so we may go around the loop again if promotion fails agaion.
3310     if (GCExpandToAllocateDelayMillis > 0) {
3311       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3312     }
3313   }
3314 }
3315 
3316 
3317 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3318   PromotionInfo* promo) {
3319   MutexLocker x(ParGCRareEvent_lock);
3320   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3321   while (true) {
3322     // Expansion by some other thread might make alloc OK now:
3323     if (promo->ensure_spooling_space()) {
3324       assert(promo->has_spooling_space(),
3325              "Post-condition of successful ensure_spooling_space()");
3326       return true;
3327     }
3328     // If there's not enough expansion space available, give up.
3329     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3330       return false;
3331     }
3332     // Otherwise, we try expansion.
3333     expand(refill_size_bytes, MinHeapDeltaBytes,
3334       CMSExpansionCause::_allocate_par_spooling_space);
3335     // Now go around the loop and try alloc again;
3336     // A competing allocation might beat us to the expansion space,
3337     // so we may go around the loop again if allocation fails again.
3338     if (GCExpandToAllocateDelayMillis > 0) {
3339       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3340     }
3341   }
3342 }
3343 
3344 
3345 
3346 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3347   assert_locked_or_safepoint(Heap_lock);
3348   size_t size = ReservedSpace::page_align_size_down(bytes);
3349   if (size > 0) {
3350     shrink_by(size);
3351   }
3352 }
3353 
3354 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3355   assert_locked_or_safepoint(Heap_lock);
3356   bool result = _virtual_space.expand_by(bytes);
3357   if (result) {
3358     HeapWord* old_end = _cmsSpace->end();
3359     size_t new_word_size =
3360       heap_word_size(_virtual_space.committed_size());
3361     MemRegion mr(_cmsSpace->bottom(), new_word_size);
3362     _bts->resize(new_word_size);  // resize the block offset shared array
3363     Universe::heap()->barrier_set()->resize_covered_region(mr);
3364     // Hmmmm... why doesn't CFLS::set_end verify locking?
3365     // This is quite ugly; FIX ME XXX
3366     _cmsSpace->assert_locked(freelistLock());
3367     _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3368 
3369     // update the space and generation capacity counters
3370     if (UsePerfData) {
3371       _space_counters->update_capacity();
3372       _gen_counters->update_all();
3373     }
3374 
3375     if (Verbose && PrintGC) {
3376       size_t new_mem_size = _virtual_space.committed_size();
3377       size_t old_mem_size = new_mem_size - bytes;
3378       gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3379                     name(), old_mem_size/K, bytes/K, new_mem_size/K);
3380     }
3381   }
3382   return result;
3383 }
3384 
3385 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3386   assert_locked_or_safepoint(Heap_lock);
3387   bool success = true;
3388   const size_t remaining_bytes = _virtual_space.uncommitted_size();
3389   if (remaining_bytes > 0) {
3390     success = grow_by(remaining_bytes);
3391     DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3392   }
3393   return success;
3394 }
3395 
3396 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3397   assert_locked_or_safepoint(Heap_lock);
3398   assert_lock_strong(freelistLock());
3399   // XXX Fix when compaction is implemented.
3400   warning("Shrinking of CMS not yet implemented");
3401   return;
3402 }
3403 
3404 
3405 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3406 // phases.
3407 class CMSPhaseAccounting: public StackObj {
3408  public:
3409   CMSPhaseAccounting(CMSCollector *collector,
3410                      const char *phase,
3411                      bool print_cr = true);
3412   ~CMSPhaseAccounting();
3413 
3414  private:
3415   CMSCollector *_collector;
3416   const char *_phase;
3417   elapsedTimer _wallclock;
3418   bool _print_cr;
3419 
3420  public:
3421   // Not MT-safe; so do not pass around these StackObj's
3422   // where they may be accessed by other threads.
3423   jlong wallclock_millis() {
3424     assert(_wallclock.is_active(), "Wall clock should not stop");
3425     _wallclock.stop();  // to record time
3426     jlong ret = _wallclock.milliseconds();
3427     _wallclock.start(); // restart
3428     return ret;
3429   }
3430 };
3431 
3432 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3433                                        const char *phase,
3434                                        bool print_cr) :
3435   _collector(collector), _phase(phase), _print_cr(print_cr) {
3436 
3437   if (PrintCMSStatistics != 0) {
3438     _collector->resetYields();
3439   }
3440   if (PrintGCDetails && PrintGCTimeStamps) {
3441     gclog_or_tty->date_stamp(PrintGCDateStamps);
3442     gclog_or_tty->stamp();
3443     gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3444       _collector->cmsGen()->short_name(), _phase);
3445   }
3446   _collector->resetTimer();
3447   _wallclock.start();
3448   _collector->startTimer();
3449 }
3450 
3451 CMSPhaseAccounting::~CMSPhaseAccounting() {
3452   assert(_wallclock.is_active(), "Wall clock should not have stopped");
3453   _collector->stopTimer();
3454   _wallclock.stop();
3455   if (PrintGCDetails) {
3456     gclog_or_tty->date_stamp(PrintGCDateStamps);
3457     if (PrintGCTimeStamps) {
3458       gclog_or_tty->stamp();
3459       gclog_or_tty->print(": ");
3460     }
3461     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3462                  _collector->cmsGen()->short_name(),
3463                  _phase, _collector->timerValue(), _wallclock.seconds());
3464     if (_print_cr) {
3465       gclog_or_tty->print_cr("");
3466     }
3467     if (PrintCMSStatistics != 0) {
3468       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3469                     _collector->yields());
3470     }
3471   }
3472 }
3473 
3474 // CMS work
3475 
3476 // Checkpoint the roots into this generation from outside
3477 // this generation. [Note this initial checkpoint need only
3478 // be approximate -- we'll do a catch up phase subsequently.]
3479 void CMSCollector::checkpointRootsInitial(bool asynch) {
3480   assert(_collectorState == InitialMarking, "Wrong collector state");
3481   check_correct_thread_executing();
3482   TraceCMSMemoryManagerStats tms(_collectorState);
3483 
3484   ReferenceProcessor* rp = ref_processor();
3485   SpecializationStats::clear();
3486   assert(_restart_addr == NULL, "Control point invariant");
3487   if (asynch) {
3488     // acquire locks for subsequent manipulations
3489     MutexLockerEx x(bitMapLock(),
3490                     Mutex::_no_safepoint_check_flag);
3491     checkpointRootsInitialWork(asynch);
3492     rp->verify_no_references_recorded();
3493     rp->enable_discovery(); // enable ("weak") refs discovery
3494     _collectorState = Marking;
3495   } else {
3496     // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3497     // which recognizes if we are a CMS generation, and doesn't try to turn on
3498     // discovery; verify that they aren't meddling.
3499     assert(!rp->discovery_is_atomic(),
3500            "incorrect setting of discovery predicate");
3501     assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3502            "ref discovery for this generation kind");
3503     // already have locks
3504     checkpointRootsInitialWork(asynch);
3505     rp->enable_discovery(); // now enable ("weak") refs discovery
3506     _collectorState = Marking;
3507   }
3508   SpecializationStats::print();
3509 }
3510 
3511 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3512   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3513   assert(_collectorState == InitialMarking, "just checking");
3514 
3515   // If there has not been a GC[n-1] since last GC[n] cycle completed,
3516   // precede our marking with a collection of all
3517   // younger generations to keep floating garbage to a minimum.
3518   // XXX: we won't do this for now -- it's an optimization to be done later.
3519 
3520   // already have locks
3521   assert_lock_strong(bitMapLock());
3522   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3523 
3524   // Setup the verification and class unloading state for this
3525   // CMS collection cycle.
3526   setup_cms_unloading_and_verification_state();
3527 
3528   NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3529     PrintGCDetails && Verbose, true, gclog_or_tty);)
3530   if (UseAdaptiveSizePolicy) {
3531     size_policy()->checkpoint_roots_initial_begin();
3532   }
3533 
3534   // Reset all the PLAB chunk arrays if necessary.
3535   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3536     reset_survivor_plab_arrays();
3537   }
3538 
3539   ResourceMark rm;
3540   HandleMark  hm;
3541 
3542   FalseClosure falseClosure;
3543   // In the case of a synchronous collection, we will elide the
3544   // remark step, so it's important to catch all the nmethod oops
3545   // in this step.
3546   // The final 'true' flag to gen_process_strong_roots will ensure this.
3547   // If 'async' is true, we can relax the nmethod tracing.
3548   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3549   GenCollectedHeap* gch = GenCollectedHeap::heap();
3550 
3551   verify_work_stacks_empty();
3552   verify_overflow_empty();
3553 
3554   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
3555   // Update the saved marks which may affect the root scans.
3556   gch->save_marks();
3557 
3558   // weak reference processing has not started yet.
3559   ref_processor()->set_enqueuing_is_done(false);
3560 
3561   {
3562     // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3563     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3564     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3565     gch->gen_process_strong_roots(_cmsGen->level(),
3566                                   true,   // younger gens are roots
3567                                   true,   // activate StrongRootsScope
3568                                   true,   // collecting perm gen
3569                                   SharedHeap::ScanningOption(roots_scanning_options()),
3570                                   &notOlder,
3571                                   true,   // walk all of code cache if (so & SO_CodeCache)
3572                                   NULL);
3573   }
3574 
3575   // Clear mod-union table; it will be dirtied in the prologue of
3576   // CMS generation per each younger generation collection.
3577 
3578   assert(_modUnionTable.isAllClear(),
3579        "Was cleared in most recent final checkpoint phase"
3580        " or no bits are set in the gc_prologue before the start of the next "
3581        "subsequent marking phase.");
3582 
3583   // Temporarily disabled, since pre/post-consumption closures don't
3584   // care about precleaned cards
3585   #if 0
3586   {
3587     MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3588                              (HeapWord*)_virtual_space.high());
3589     _ct->ct_bs()->preclean_dirty_cards(mr);
3590   }
3591   #endif
3592 
3593   // Save the end of the used_region of the constituent generations
3594   // to be used to limit the extent of sweep in each generation.
3595   save_sweep_limits();
3596   if (UseAdaptiveSizePolicy) {
3597     size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3598   }
3599   verify_overflow_empty();
3600 }
3601 
3602 bool CMSCollector::markFromRoots(bool asynch) {
3603   // we might be tempted to assert that:
3604   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3605   //        "inconsistent argument?");
3606   // However that wouldn't be right, because it's possible that
3607   // a safepoint is indeed in progress as a younger generation
3608   // stop-the-world GC happens even as we mark in this generation.
3609   assert(_collectorState == Marking, "inconsistent state?");
3610   check_correct_thread_executing();
3611   verify_overflow_empty();
3612 
3613   bool res;
3614   if (asynch) {
3615 
3616     // Start the timers for adaptive size policy for the concurrent phases
3617     // Do it here so that the foreground MS can use the concurrent
3618     // timer since a foreground MS might has the sweep done concurrently
3619     // or STW.
3620     if (UseAdaptiveSizePolicy) {
3621       size_policy()->concurrent_marking_begin();
3622     }
3623 
3624     // Weak ref discovery note: We may be discovering weak
3625     // refs in this generation concurrent (but interleaved) with
3626     // weak ref discovery by a younger generation collector.
3627 
3628     CMSTokenSyncWithLocks ts(true, bitMapLock());
3629     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3630     CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3631     res = markFromRootsWork(asynch);
3632     if (res) {
3633       _collectorState = Precleaning;
3634     } else { // We failed and a foreground collection wants to take over
3635       assert(_foregroundGCIsActive, "internal state inconsistency");
3636       assert(_restart_addr == NULL,  "foreground will restart from scratch");
3637       if (PrintGCDetails) {
3638         gclog_or_tty->print_cr("bailing out to foreground collection");
3639       }
3640     }
3641     if (UseAdaptiveSizePolicy) {
3642       size_policy()->concurrent_marking_end();
3643     }
3644   } else {
3645     assert(SafepointSynchronize::is_at_safepoint(),
3646            "inconsistent with asynch == false");
3647     if (UseAdaptiveSizePolicy) {
3648       size_policy()->ms_collection_marking_begin();
3649     }
3650     // already have locks
3651     res = markFromRootsWork(asynch);
3652     _collectorState = FinalMarking;
3653     if (UseAdaptiveSizePolicy) {
3654       GenCollectedHeap* gch = GenCollectedHeap::heap();
3655       size_policy()->ms_collection_marking_end(gch->gc_cause());
3656     }
3657   }
3658   verify_overflow_empty();
3659   return res;
3660 }
3661 
3662 bool CMSCollector::markFromRootsWork(bool asynch) {
3663   // iterate over marked bits in bit map, doing a full scan and mark
3664   // from these roots using the following algorithm:
3665   // . if oop is to the right of the current scan pointer,
3666   //   mark corresponding bit (we'll process it later)
3667   // . else (oop is to left of current scan pointer)
3668   //   push oop on marking stack
3669   // . drain the marking stack
3670 
3671   // Note that when we do a marking step we need to hold the
3672   // bit map lock -- recall that direct allocation (by mutators)
3673   // and promotion (by younger generation collectors) is also
3674   // marking the bit map. [the so-called allocate live policy.]
3675   // Because the implementation of bit map marking is not
3676   // robust wrt simultaneous marking of bits in the same word,
3677   // we need to make sure that there is no such interference
3678   // between concurrent such updates.
3679 
3680   // already have locks
3681   assert_lock_strong(bitMapLock());
3682 
3683   // Clear the revisit stack, just in case there are any
3684   // obsolete contents from a short-circuited previous CMS cycle.
3685   _revisitStack.reset();
3686   verify_work_stacks_empty();
3687   verify_overflow_empty();
3688   assert(_revisitStack.isEmpty(), "tabula rasa");
3689   DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3690   bool result = false;
3691   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3692     result = do_marking_mt(asynch);
3693   } else {
3694     result = do_marking_st(asynch);
3695   }
3696   return result;
3697 }
3698 
3699 // Forward decl
3700 class CMSConcMarkingTask;
3701 
3702 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3703   CMSCollector*       _collector;
3704   CMSConcMarkingTask* _task;
3705  public:
3706   virtual void yield();
3707 
3708   // "n_threads" is the number of threads to be terminated.
3709   // "queue_set" is a set of work queues of other threads.
3710   // "collector" is the CMS collector associated with this task terminator.
3711   // "yield" indicates whether we need the gang as a whole to yield.
3712   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3713     ParallelTaskTerminator(n_threads, queue_set),
3714     _collector(collector) { }
3715 
3716   void set_task(CMSConcMarkingTask* task) {
3717     _task = task;
3718   }
3719 };
3720 
3721 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3722   CMSConcMarkingTask* _task;
3723  public:
3724   bool should_exit_termination();
3725   void set_task(CMSConcMarkingTask* task) {
3726     _task = task;
3727   }
3728 };
3729 
3730 // MT Concurrent Marking Task
3731 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3732   CMSCollector* _collector;
3733   int           _n_workers;                  // requested/desired # workers
3734   bool          _asynch;
3735   bool          _result;
3736   CompactibleFreeListSpace*  _cms_space;
3737   CompactibleFreeListSpace* _perm_space;
3738   char          _pad_front[64];   // padding to ...
3739   HeapWord*     _global_finger;   // ... avoid sharing cache line
3740   char          _pad_back[64];
3741   HeapWord*     _restart_addr;
3742 
3743   //  Exposed here for yielding support
3744   Mutex* const _bit_map_lock;
3745 
3746   // The per thread work queues, available here for stealing
3747   OopTaskQueueSet*  _task_queues;
3748 
3749   // Termination (and yielding) support
3750   CMSConcMarkingTerminator _term;
3751   CMSConcMarkingTerminatorTerminator _term_term;
3752 
3753  public:
3754   CMSConcMarkingTask(CMSCollector* collector,
3755                  CompactibleFreeListSpace* cms_space,
3756                  CompactibleFreeListSpace* perm_space,
3757                  bool asynch,
3758                  YieldingFlexibleWorkGang* workers,
3759                  OopTaskQueueSet* task_queues):
3760     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3761     _collector(collector),
3762     _cms_space(cms_space),
3763     _perm_space(perm_space),
3764     _asynch(asynch), _n_workers(0), _result(true),
3765     _task_queues(task_queues),
3766     _term(_n_workers, task_queues, _collector),
3767     _bit_map_lock(collector->bitMapLock())
3768   {
3769     _requested_size = _n_workers;
3770     _term.set_task(this);
3771     _term_term.set_task(this);
3772     assert(_cms_space->bottom() < _perm_space->bottom(),
3773            "Finger incorrectly initialized below");
3774     _restart_addr = _global_finger = _cms_space->bottom();
3775   }
3776 
3777 
3778   OopTaskQueueSet* task_queues()  { return _task_queues; }
3779 
3780   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3781 
3782   HeapWord** global_finger_addr() { return &_global_finger; }
3783 
3784   CMSConcMarkingTerminator* terminator() { return &_term; }
3785 
3786   virtual void set_for_termination(int active_workers) {
3787     terminator()->reset_for_reuse(active_workers);
3788   }
3789 
3790   void work(int i);
3791   bool should_yield() {
3792     return    ConcurrentMarkSweepThread::should_yield()
3793            && !_collector->foregroundGCIsActive()
3794            && _asynch;
3795   }
3796 
3797   virtual void coordinator_yield();  // stuff done by coordinator
3798   bool result() { return _result; }
3799 
3800   void reset(HeapWord* ra) {
3801     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3802     assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3803     assert(ra             <  _perm_space->end(), "ra too large");
3804     _restart_addr = _global_finger = ra;
3805     _term.reset_for_reuse();
3806   }
3807 
3808   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3809                                            OopTaskQueue* work_q);
3810 
3811  private:
3812   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3813   void do_work_steal(int i);
3814   void bump_global_finger(HeapWord* f);
3815 };
3816 
3817 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3818   assert(_task != NULL, "Error");
3819   return _task->yielding();
3820   // Note that we do not need the disjunct || _task->should_yield() above
3821   // because we want terminating threads to yield only if the task
3822   // is already in the midst of yielding, which happens only after at least one
3823   // thread has yielded.
3824 }
3825 
3826 void CMSConcMarkingTerminator::yield() {
3827   if (_task->should_yield()) {
3828     _task->yield();
3829   } else {
3830     ParallelTaskTerminator::yield();
3831   }
3832 }
3833 
3834 ////////////////////////////////////////////////////////////////
3835 // Concurrent Marking Algorithm Sketch
3836 ////////////////////////////////////////////////////////////////
3837 // Until all tasks exhausted (both spaces):
3838 // -- claim next available chunk
3839 // -- bump global finger via CAS
3840 // -- find first object that starts in this chunk
3841 //    and start scanning bitmap from that position
3842 // -- scan marked objects for oops
3843 // -- CAS-mark target, and if successful:
3844 //    . if target oop is above global finger (volatile read)
3845 //      nothing to do
3846 //    . if target oop is in chunk and above local finger
3847 //        then nothing to do
3848 //    . else push on work-queue
3849 // -- Deal with possible overflow issues:
3850 //    . local work-queue overflow causes stuff to be pushed on
3851 //      global (common) overflow queue
3852 //    . always first empty local work queue
3853 //    . then get a batch of oops from global work queue if any
3854 //    . then do work stealing
3855 // -- When all tasks claimed (both spaces)
3856 //    and local work queue empty,
3857 //    then in a loop do:
3858 //    . check global overflow stack; steal a batch of oops and trace
3859 //    . try to steal from other threads oif GOS is empty
3860 //    . if neither is available, offer termination
3861 // -- Terminate and return result
3862 //
3863 void CMSConcMarkingTask::work(int i) {
3864   elapsedTimer _timer;
3865   ResourceMark rm;
3866   HandleMark hm;
3867 
3868   DEBUG_ONLY(_collector->verify_overflow_empty();)
3869 
3870   // Before we begin work, our work queue should be empty
3871   assert(work_queue(i)->size() == 0, "Expected to be empty");
3872   // Scan the bitmap covering _cms_space, tracing through grey objects.
3873   _timer.start();
3874   do_scan_and_mark(i, _cms_space);
3875   _timer.stop();
3876   if (PrintCMSStatistics != 0) {
3877     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3878       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3879   }
3880 
3881   // ... do the same for the _perm_space
3882   _timer.reset();
3883   _timer.start();
3884   do_scan_and_mark(i, _perm_space);
3885   _timer.stop();
3886   if (PrintCMSStatistics != 0) {
3887     gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3888       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3889   }
3890 
3891   // ... do work stealing
3892   _timer.reset();
3893   _timer.start();
3894   do_work_steal(i);
3895   _timer.stop();
3896   if (PrintCMSStatistics != 0) {
3897     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3898       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3899   }
3900   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3901   assert(work_queue(i)->size() == 0, "Should have been emptied");
3902   // Note that under the current task protocol, the
3903   // following assertion is true even of the spaces
3904   // expanded since the completion of the concurrent
3905   // marking. XXX This will likely change under a strict
3906   // ABORT semantics.
3907   assert(_global_finger >  _cms_space->end() &&
3908          _global_finger >= _perm_space->end(),
3909          "All tasks have been completed");
3910   DEBUG_ONLY(_collector->verify_overflow_empty();)
3911 }
3912 
3913 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3914   HeapWord* read = _global_finger;
3915   HeapWord* cur  = read;
3916   while (f > read) {
3917     cur = read;
3918     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3919     if (cur == read) {
3920       // our cas succeeded
3921       assert(_global_finger >= f, "protocol consistency");
3922       break;
3923     }
3924   }
3925 }
3926 
3927 // This is really inefficient, and should be redone by
3928 // using (not yet available) block-read and -write interfaces to the
3929 // stack and the work_queue. XXX FIX ME !!!
3930 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3931                                                       OopTaskQueue* work_q) {
3932   // Fast lock-free check
3933   if (ovflw_stk->length() == 0) {
3934     return false;
3935   }
3936   assert(work_q->size() == 0, "Shouldn't steal");
3937   MutexLockerEx ml(ovflw_stk->par_lock(),
3938                    Mutex::_no_safepoint_check_flag);
3939   // Grab up to 1/4 the size of the work queue
3940   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3941                     (size_t)ParGCDesiredObjsFromOverflowList);
3942   num = MIN2(num, ovflw_stk->length());
3943   for (int i = (int) num; i > 0; i--) {
3944     oop cur = ovflw_stk->pop();
3945     assert(cur != NULL, "Counted wrong?");
3946     work_q->push(cur);
3947   }
3948   return num > 0;
3949 }
3950 
3951 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3952   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3953   int n_tasks = pst->n_tasks();
3954   // We allow that there may be no tasks to do here because
3955   // we are restarting after a stack overflow.
3956   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3957   int nth_task = 0;
3958 
3959   HeapWord* aligned_start = sp->bottom();
3960   if (sp->used_region().contains(_restart_addr)) {
3961     // Align down to a card boundary for the start of 0th task
3962     // for this space.
3963     aligned_start =
3964       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3965                                  CardTableModRefBS::card_size);
3966   }
3967 
3968   size_t chunk_size = sp->marking_task_size();
3969   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3970     // Having claimed the nth task in this space,
3971     // compute the chunk that it corresponds to:
3972     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3973                                aligned_start + (nth_task+1)*chunk_size);
3974     // Try and bump the global finger via a CAS;
3975     // note that we need to do the global finger bump
3976     // _before_ taking the intersection below, because
3977     // the task corresponding to that region will be
3978     // deemed done even if the used_region() expands
3979     // because of allocation -- as it almost certainly will
3980     // during start-up while the threads yield in the
3981     // closure below.
3982     HeapWord* finger = span.end();
3983     bump_global_finger(finger);   // atomically
3984     // There are null tasks here corresponding to chunks
3985     // beyond the "top" address of the space.
3986     span = span.intersection(sp->used_region());
3987     if (!span.is_empty()) {  // Non-null task
3988       HeapWord* prev_obj;
3989       assert(!span.contains(_restart_addr) || nth_task == 0,
3990              "Inconsistency");
3991       if (nth_task == 0) {
3992         // For the 0th task, we'll not need to compute a block_start.
3993         if (span.contains(_restart_addr)) {
3994           // In the case of a restart because of stack overflow,
3995           // we might additionally skip a chunk prefix.
3996           prev_obj = _restart_addr;
3997         } else {
3998           prev_obj = span.start();
3999         }
4000       } else {
4001         // We want to skip the first object because
4002         // the protocol is to scan any object in its entirety
4003         // that _starts_ in this span; a fortiori, any
4004         // object starting in an earlier span is scanned
4005         // as part of an earlier claimed task.
4006         // Below we use the "careful" version of block_start
4007         // so we do not try to navigate uninitialized objects.
4008         prev_obj = sp->block_start_careful(span.start());
4009         // Below we use a variant of block_size that uses the
4010         // Printezis bits to avoid waiting for allocated
4011         // objects to become initialized/parsable.
4012         while (prev_obj < span.start()) {
4013           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4014           if (sz > 0) {
4015             prev_obj += sz;
4016           } else {
4017             // In this case we may end up doing a bit of redundant
4018             // scanning, but that appears unavoidable, short of
4019             // locking the free list locks; see bug 6324141.
4020             break;
4021           }
4022         }
4023       }
4024       if (prev_obj < span.end()) {
4025         MemRegion my_span = MemRegion(prev_obj, span.end());
4026         // Do the marking work within a non-empty span --
4027         // the last argument to the constructor indicates whether the
4028         // iteration should be incremental with periodic yields.
4029         Par_MarkFromRootsClosure cl(this, _collector, my_span,
4030                                     &_collector->_markBitMap,
4031                                     work_queue(i),
4032                                     &_collector->_markStack,
4033                                     &_collector->_revisitStack,
4034                                     _asynch);
4035         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4036       } // else nothing to do for this task
4037     }   // else nothing to do for this task
4038   }
4039   // We'd be tempted to assert here that since there are no
4040   // more tasks left to claim in this space, the global_finger
4041   // must exceed space->top() and a fortiori space->end(). However,
4042   // that would not quite be correct because the bumping of
4043   // global_finger occurs strictly after the claiming of a task,
4044   // so by the time we reach here the global finger may not yet
4045   // have been bumped up by the thread that claimed the last
4046   // task.
4047   pst->all_tasks_completed();
4048 }
4049 
4050 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4051  private:
4052   CMSConcMarkingTask* _task;
4053   MemRegion     _span;
4054   CMSBitMap*    _bit_map;
4055   CMSMarkStack* _overflow_stack;
4056   OopTaskQueue* _work_queue;
4057  protected:
4058   DO_OOP_WORK_DEFN
4059  public:
4060   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4061                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4062                          CMSMarkStack* revisit_stack):
4063     Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
4064     _task(task),
4065     _span(collector->_span),
4066     _work_queue(work_queue),
4067     _bit_map(bit_map),
4068     _overflow_stack(overflow_stack)
4069   { }
4070   virtual void do_oop(oop* p);
4071   virtual void do_oop(narrowOop* p);
4072   void trim_queue(size_t max);
4073   void handle_stack_overflow(HeapWord* lost);
4074   void do_yield_check() {
4075     if (_task->should_yield()) {
4076       _task->yield();
4077     }
4078   }
4079 };
4080 
4081 // Grey object scanning during work stealing phase --
4082 // the salient assumption here is that any references
4083 // that are in these stolen objects being scanned must
4084 // already have been initialized (else they would not have
4085 // been published), so we do not need to check for
4086 // uninitialized objects before pushing here.
4087 void Par_ConcMarkingClosure::do_oop(oop obj) {
4088   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4089   HeapWord* addr = (HeapWord*)obj;
4090   // Check if oop points into the CMS generation
4091   // and is not marked
4092   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4093     // a white object ...
4094     // If we manage to "claim" the object, by being the
4095     // first thread to mark it, then we push it on our
4096     // marking stack
4097     if (_bit_map->par_mark(addr)) {     // ... now grey
4098       // push on work queue (grey set)
4099       bool simulate_overflow = false;
4100       NOT_PRODUCT(
4101         if (CMSMarkStackOverflowALot &&
4102             _collector->simulate_overflow()) {
4103           // simulate a stack overflow
4104           simulate_overflow = true;
4105         }
4106       )
4107       if (simulate_overflow ||
4108           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4109         // stack overflow
4110         if (PrintCMSStatistics != 0) {
4111           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4112                                  SIZE_FORMAT, _overflow_stack->capacity());
4113         }
4114         // We cannot assert that the overflow stack is full because
4115         // it may have been emptied since.
4116         assert(simulate_overflow ||
4117                _work_queue->size() == _work_queue->max_elems(),
4118               "Else push should have succeeded");
4119         handle_stack_overflow(addr);
4120       }
4121     } // Else, some other thread got there first
4122     do_yield_check();
4123   }
4124 }
4125 
4126 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
4127 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4128 
4129 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4130   while (_work_queue->size() > max) {
4131     oop new_oop;
4132     if (_work_queue->pop_local(new_oop)) {
4133       assert(new_oop->is_oop(), "Should be an oop");
4134       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4135       assert(_span.contains((HeapWord*)new_oop), "Not in span");
4136       assert(new_oop->is_parsable(), "Should be parsable");
4137       new_oop->oop_iterate(this);  // do_oop() above
4138       do_yield_check();
4139     }
4140   }
4141 }
4142 
4143 // Upon stack overflow, we discard (part of) the stack,
4144 // remembering the least address amongst those discarded
4145 // in CMSCollector's _restart_address.
4146 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4147   // We need to do this under a mutex to prevent other
4148   // workers from interfering with the work done below.
4149   MutexLockerEx ml(_overflow_stack->par_lock(),
4150                    Mutex::_no_safepoint_check_flag);
4151   // Remember the least grey address discarded
4152   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4153   _collector->lower_restart_addr(ra);
4154   _overflow_stack->reset();  // discard stack contents
4155   _overflow_stack->expand(); // expand the stack if possible
4156 }
4157 
4158 
4159 void CMSConcMarkingTask::do_work_steal(int i) {
4160   OopTaskQueue* work_q = work_queue(i);
4161   oop obj_to_scan;
4162   CMSBitMap* bm = &(_collector->_markBitMap);
4163   CMSMarkStack* ovflw = &(_collector->_markStack);
4164   CMSMarkStack* revisit = &(_collector->_revisitStack);
4165   int* seed = _collector->hash_seed(i);
4166   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4167   while (true) {
4168     cl.trim_queue(0);
4169     assert(work_q->size() == 0, "Should have been emptied above");
4170     if (get_work_from_overflow_stack(ovflw, work_q)) {
4171       // Can't assert below because the work obtained from the
4172       // overflow stack may already have been stolen from us.
4173       // assert(work_q->size() > 0, "Work from overflow stack");
4174       continue;
4175     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4176       assert(obj_to_scan->is_oop(), "Should be an oop");
4177       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4178       obj_to_scan->oop_iterate(&cl);
4179     } else if (terminator()->offer_termination(&_term_term)) {
4180       assert(work_q->size() == 0, "Impossible!");
4181       break;
4182     } else if (yielding() || should_yield()) {
4183       yield();
4184     }
4185   }
4186 }
4187 
4188 // This is run by the CMS (coordinator) thread.
4189 void CMSConcMarkingTask::coordinator_yield() {
4190   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4191          "CMS thread should hold CMS token");
4192   DEBUG_ONLY(RememberKlassesChecker mux(false);)
4193   // First give up the locks, then yield, then re-lock
4194   // We should probably use a constructor/destructor idiom to
4195   // do this unlock/lock or modify the MutexUnlocker class to
4196   // serve our purpose. XXX
4197   assert_lock_strong(_bit_map_lock);
4198   _bit_map_lock->unlock();
4199   ConcurrentMarkSweepThread::desynchronize(true);
4200   ConcurrentMarkSweepThread::acknowledge_yield_request();
4201   _collector->stopTimer();
4202   if (PrintCMSStatistics != 0) {
4203     _collector->incrementYields();
4204   }
4205   _collector->icms_wait();
4206 
4207   // It is possible for whichever thread initiated the yield request
4208   // not to get a chance to wake up and take the bitmap lock between
4209   // this thread releasing it and reacquiring it. So, while the
4210   // should_yield() flag is on, let's sleep for a bit to give the
4211   // other thread a chance to wake up. The limit imposed on the number
4212   // of iterations is defensive, to avoid any unforseen circumstances
4213   // putting us into an infinite loop. Since it's always been this
4214   // (coordinator_yield()) method that was observed to cause the
4215   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4216   // which is by default non-zero. For the other seven methods that
4217   // also perform the yield operation, as are using a different
4218   // parameter (CMSYieldSleepCount) which is by default zero. This way we
4219   // can enable the sleeping for those methods too, if necessary.
4220   // See 6442774.
4221   //
4222   // We really need to reconsider the synchronization between the GC
4223   // thread and the yield-requesting threads in the future and we
4224   // should really use wait/notify, which is the recommended
4225   // way of doing this type of interaction. Additionally, we should
4226   // consolidate the eight methods that do the yield operation and they
4227   // are almost identical into one for better maintenability and
4228   // readability. See 6445193.
4229   //
4230   // Tony 2006.06.29
4231   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4232                    ConcurrentMarkSweepThread::should_yield() &&
4233                    !CMSCollector::foregroundGCIsActive(); ++i) {
4234     os::sleep(Thread::current(), 1, false);
4235     ConcurrentMarkSweepThread::acknowledge_yield_request();
4236   }
4237 
4238   ConcurrentMarkSweepThread::synchronize(true);
4239   _bit_map_lock->lock_without_safepoint_check();
4240   _collector->startTimer();
4241 }
4242 
4243 bool CMSCollector::do_marking_mt(bool asynch) {
4244   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4245   // In the future this would be determined ergonomically, based
4246   // on #cpu's, # active mutator threads (and load), and mutation rate.
4247   int num_workers = ConcGCThreads;
4248 
4249   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4250   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4251 
4252   CMSConcMarkingTask tsk(this,
4253                          cms_space,
4254                          perm_space,
4255                          asynch,
4256                          conc_workers(),
4257                          task_queues());
4258 
4259   // Since the actual number of workers we get may be different
4260   // from the number we requested above, do we need to do anything different
4261   // below? In particular, may be we need to subclass the SequantialSubTasksDone
4262   // class?? XXX
4263   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4264   perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4265 
4266   // Refs discovery is already non-atomic.
4267   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4268   // Mutate the Refs discovery so it is MT during the
4269   // multi-threaded marking phase.
4270   ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4271   DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4272   conc_workers()->start_task(&tsk);
4273   while (tsk.yielded()) {
4274     tsk.coordinator_yield();
4275     conc_workers()->continue_task(&tsk);
4276   }
4277   // If the task was aborted, _restart_addr will be non-NULL
4278   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4279   while (_restart_addr != NULL) {
4280     // XXX For now we do not make use of ABORTED state and have not
4281     // yet implemented the right abort semantics (even in the original
4282     // single-threaded CMS case). That needs some more investigation
4283     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4284     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4285     // If _restart_addr is non-NULL, a marking stack overflow
4286     // occurred; we need to do a fresh marking iteration from the
4287     // indicated restart address.
4288     if (_foregroundGCIsActive && asynch) {
4289       // We may be running into repeated stack overflows, having
4290       // reached the limit of the stack size, while making very
4291       // slow forward progress. It may be best to bail out and
4292       // let the foreground collector do its job.
4293       // Clear _restart_addr, so that foreground GC
4294       // works from scratch. This avoids the headache of
4295       // a "rescan" which would otherwise be needed because
4296       // of the dirty mod union table & card table.
4297       _restart_addr = NULL;
4298       return false;
4299     }
4300     // Adjust the task to restart from _restart_addr
4301     tsk.reset(_restart_addr);
4302     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4303                   _restart_addr);
4304     perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4305                   _restart_addr);
4306     _restart_addr = NULL;
4307     // Get the workers going again
4308     conc_workers()->start_task(&tsk);
4309     while (tsk.yielded()) {
4310       tsk.coordinator_yield();
4311       conc_workers()->continue_task(&tsk);
4312     }
4313   }
4314   assert(tsk.completed(), "Inconsistency");
4315   assert(tsk.result() == true, "Inconsistency");
4316   return true;
4317 }
4318 
4319 bool CMSCollector::do_marking_st(bool asynch) {
4320   ResourceMark rm;
4321   HandleMark   hm;
4322 
4323   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4324     &_markStack, &_revisitStack, CMSYield && asynch);
4325   // the last argument to iterate indicates whether the iteration
4326   // should be incremental with periodic yields.
4327   _markBitMap.iterate(&markFromRootsClosure);
4328   // If _restart_addr is non-NULL, a marking stack overflow
4329   // occurred; we need to do a fresh iteration from the
4330   // indicated restart address.
4331   while (_restart_addr != NULL) {
4332     if (_foregroundGCIsActive && asynch) {
4333       // We may be running into repeated stack overflows, having
4334       // reached the limit of the stack size, while making very
4335       // slow forward progress. It may be best to bail out and
4336       // let the foreground collector do its job.
4337       // Clear _restart_addr, so that foreground GC
4338       // works from scratch. This avoids the headache of
4339       // a "rescan" which would otherwise be needed because
4340       // of the dirty mod union table & card table.
4341       _restart_addr = NULL;
4342       return false;  // indicating failure to complete marking
4343     }
4344     // Deal with stack overflow:
4345     // we restart marking from _restart_addr
4346     HeapWord* ra = _restart_addr;
4347     markFromRootsClosure.reset(ra);
4348     _restart_addr = NULL;
4349     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4350   }
4351   return true;
4352 }
4353 
4354 void CMSCollector::preclean() {
4355   check_correct_thread_executing();
4356   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4357   verify_work_stacks_empty();
4358   verify_overflow_empty();
4359   _abort_preclean = false;
4360   if (CMSPrecleaningEnabled) {
4361     // Precleaning is currently not MT but the reference processor
4362     // may be set for MT.  Disable it temporarily here.
4363     ReferenceProcessor* rp = ref_processor();
4364     ReferenceProcessorMTProcMutator z(rp, false);
4365     _eden_chunk_index = 0;
4366     size_t used = get_eden_used();
4367     size_t capacity = get_eden_capacity();
4368     // Don't start sampling unless we will get sufficiently
4369     // many samples.
4370     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4371                 * CMSScheduleRemarkEdenPenetration)) {
4372       _start_sampling = true;
4373     } else {
4374       _start_sampling = false;
4375     }
4376     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4377     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4378     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4379   }
4380   CMSTokenSync x(true); // is cms thread
4381   if (CMSPrecleaningEnabled) {
4382     sample_eden();
4383     _collectorState = AbortablePreclean;
4384   } else {
4385     _collectorState = FinalMarking;
4386   }
4387   verify_work_stacks_empty();
4388   verify_overflow_empty();
4389 }
4390 
4391 // Try and schedule the remark such that young gen
4392 // occupancy is CMSScheduleRemarkEdenPenetration %.
4393 void CMSCollector::abortable_preclean() {
4394   check_correct_thread_executing();
4395   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
4396   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4397 
4398   // If Eden's current occupancy is below this threshold,
4399   // immediately schedule the remark; else preclean
4400   // past the next scavenge in an effort to
4401   // schedule the pause as described avove. By choosing
4402   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4403   // we will never do an actual abortable preclean cycle.
4404   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4405     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4406     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4407     // We need more smarts in the abortable preclean
4408     // loop below to deal with cases where allocation
4409     // in young gen is very very slow, and our precleaning
4410     // is running a losing race against a horde of
4411     // mutators intent on flooding us with CMS updates
4412     // (dirty cards).
4413     // One, admittedly dumb, strategy is to give up
4414     // after a certain number of abortable precleaning loops
4415     // or after a certain maximum time. We want to make
4416     // this smarter in the next iteration.
4417     // XXX FIX ME!!! YSR
4418     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4419     while (!(should_abort_preclean() ||
4420              ConcurrentMarkSweepThread::should_terminate())) {
4421       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4422       cumworkdone += workdone;
4423       loops++;
4424       // Voluntarily terminate abortable preclean phase if we have
4425       // been at it for too long.
4426       if ((CMSMaxAbortablePrecleanLoops != 0) &&
4427           loops >= CMSMaxAbortablePrecleanLoops) {
4428         if (PrintGCDetails) {
4429           gclog_or_tty->print(" CMS: abort preclean due to loops ");
4430         }
4431         break;
4432       }
4433       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4434         if (PrintGCDetails) {
4435           gclog_or_tty->print(" CMS: abort preclean due to time ");
4436         }
4437         break;
4438       }
4439       // If we are doing little work each iteration, we should
4440       // take a short break.
4441       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4442         // Sleep for some time, waiting for work to accumulate
4443         stopTimer();
4444         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4445         startTimer();
4446         waited++;
4447       }
4448     }
4449     if (PrintCMSStatistics > 0) {
4450       gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4451                           loops, waited, cumworkdone);
4452     }
4453   }
4454   CMSTokenSync x(true); // is cms thread
4455   if (_collectorState != Idling) {
4456     assert(_collectorState == AbortablePreclean,
4457            "Spontaneous state transition?");
4458     _collectorState = FinalMarking;
4459   } // Else, a foreground collection completed this CMS cycle.
4460   return;
4461 }
4462 
4463 // Respond to an Eden sampling opportunity
4464 void CMSCollector::sample_eden() {
4465   // Make sure a young gc cannot sneak in between our
4466   // reading and recording of a sample.
4467   assert(Thread::current()->is_ConcurrentGC_thread(),
4468          "Only the cms thread may collect Eden samples");
4469   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4470          "Should collect samples while holding CMS token");
4471   if (!_start_sampling) {
4472     return;
4473   }
4474   if (_eden_chunk_array) {
4475     if (_eden_chunk_index < _eden_chunk_capacity) {
4476       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
4477       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4478              "Unexpected state of Eden");
4479       // We'd like to check that what we just sampled is an oop-start address;
4480       // however, we cannot do that here since the object may not yet have been
4481       // initialized. So we'll instead do the check when we _use_ this sample
4482       // later.
4483       if (_eden_chunk_index == 0 ||
4484           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4485                          _eden_chunk_array[_eden_chunk_index-1])
4486            >= CMSSamplingGrain)) {
4487         _eden_chunk_index++;  // commit sample
4488       }
4489     }
4490   }
4491   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4492     size_t used = get_eden_used();
4493     size_t capacity = get_eden_capacity();
4494     assert(used <= capacity, "Unexpected state of Eden");
4495     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4496       _abort_preclean = true;
4497     }
4498   }
4499 }
4500 
4501 
4502 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4503   assert(_collectorState == Precleaning ||
4504          _collectorState == AbortablePreclean, "incorrect state");
4505   ResourceMark rm;
4506   HandleMark   hm;
4507   // Do one pass of scrubbing the discovered reference lists
4508   // to remove any reference objects with strongly-reachable
4509   // referents.
4510   if (clean_refs) {
4511     ReferenceProcessor* rp = ref_processor();
4512     CMSPrecleanRefsYieldClosure yield_cl(this);
4513     assert(rp->span().equals(_span), "Spans should be equal");
4514     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4515                                    &_markStack, &_revisitStack,
4516                                    true /* preclean */);
4517     CMSDrainMarkingStackClosure complete_trace(this,
4518                                    _span, &_markBitMap, &_markStack,
4519                                    &keep_alive, true /* preclean */);
4520 
4521     // We don't want this step to interfere with a young
4522     // collection because we don't want to take CPU
4523     // or memory bandwidth away from the young GC threads
4524     // (which may be as many as there are CPUs).
4525     // Note that we don't need to protect ourselves from
4526     // interference with mutators because they can't
4527     // manipulate the discovered reference lists nor affect
4528     // the computed reachability of the referents, the
4529     // only properties manipulated by the precleaning
4530     // of these reference lists.
4531     stopTimer();
4532     CMSTokenSyncWithLocks x(true /* is cms thread */,
4533                             bitMapLock());
4534     startTimer();
4535     sample_eden();
4536 
4537     // The following will yield to allow foreground
4538     // collection to proceed promptly. XXX YSR:
4539     // The code in this method may need further
4540     // tweaking for better performance and some restructuring
4541     // for cleaner interfaces.
4542     rp->preclean_discovered_references(
4543           rp->is_alive_non_header(), &keep_alive, &complete_trace,
4544           &yield_cl, should_unload_classes());
4545   }
4546 
4547   if (clean_survivor) {  // preclean the active survivor space(s)
4548     assert(_young_gen->kind() == Generation::DefNew ||
4549            _young_gen->kind() == Generation::ParNew ||
4550            _young_gen->kind() == Generation::ASParNew,
4551          "incorrect type for cast");
4552     DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4553     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4554                              &_markBitMap, &_modUnionTable,
4555                              &_markStack, &_revisitStack,
4556                              true /* precleaning phase */);
4557     stopTimer();
4558     CMSTokenSyncWithLocks ts(true /* is cms thread */,
4559                              bitMapLock());
4560     startTimer();
4561     unsigned int before_count =
4562       GenCollectedHeap::heap()->total_collections();
4563     SurvivorSpacePrecleanClosure
4564       sss_cl(this, _span, &_markBitMap, &_markStack,
4565              &pam_cl, before_count, CMSYield);
4566     DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4567     dng->from()->object_iterate_careful(&sss_cl);
4568     dng->to()->object_iterate_careful(&sss_cl);
4569   }
4570   MarkRefsIntoAndScanClosure
4571     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4572              &_markStack, &_revisitStack, this, CMSYield,
4573              true /* precleaning phase */);
4574   // CAUTION: The following closure has persistent state that may need to
4575   // be reset upon a decrease in the sequence of addresses it
4576   // processes.
4577   ScanMarkedObjectsAgainCarefullyClosure
4578     smoac_cl(this, _span,
4579       &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4580 
4581   // Preclean dirty cards in ModUnionTable and CardTable using
4582   // appropriate convergence criterion;
4583   // repeat CMSPrecleanIter times unless we find that
4584   // we are losing.
4585   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4586   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4587          "Bad convergence multiplier");
4588   assert(CMSPrecleanThreshold >= 100,
4589          "Unreasonably low CMSPrecleanThreshold");
4590 
4591   size_t numIter, cumNumCards, lastNumCards, curNumCards;
4592   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4593        numIter < CMSPrecleanIter;
4594        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4595     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
4596     if (CMSPermGenPrecleaningEnabled) {
4597       curNumCards  += preclean_mod_union_table(_permGen, &smoac_cl);
4598     }
4599     if (Verbose && PrintGCDetails) {
4600       gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4601     }
4602     // Either there are very few dirty cards, so re-mark
4603     // pause will be small anyway, or our pre-cleaning isn't
4604     // that much faster than the rate at which cards are being
4605     // dirtied, so we might as well stop and re-mark since
4606     // precleaning won't improve our re-mark time by much.
4607     if (curNumCards <= CMSPrecleanThreshold ||
4608         (numIter > 0 &&
4609          (curNumCards * CMSPrecleanDenominator >
4610          lastNumCards * CMSPrecleanNumerator))) {
4611       numIter++;
4612       cumNumCards += curNumCards;
4613       break;
4614     }
4615   }
4616   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4617   if (CMSPermGenPrecleaningEnabled) {
4618     curNumCards += preclean_card_table(_permGen, &smoac_cl);
4619   }
4620   cumNumCards += curNumCards;
4621   if (PrintGCDetails && PrintCMSStatistics != 0) {
4622     gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4623                   curNumCards, cumNumCards, numIter);
4624   }
4625   return cumNumCards;   // as a measure of useful work done
4626 }
4627 
4628 // PRECLEANING NOTES:
4629 // Precleaning involves:
4630 // . reading the bits of the modUnionTable and clearing the set bits.
4631 // . For the cards corresponding to the set bits, we scan the
4632 //   objects on those cards. This means we need the free_list_lock
4633 //   so that we can safely iterate over the CMS space when scanning
4634 //   for oops.
4635 // . When we scan the objects, we'll be both reading and setting
4636 //   marks in the marking bit map, so we'll need the marking bit map.
4637 // . For protecting _collector_state transitions, we take the CGC_lock.
4638 //   Note that any races in the reading of of card table entries by the
4639 //   CMS thread on the one hand and the clearing of those entries by the
4640 //   VM thread or the setting of those entries by the mutator threads on the
4641 //   other are quite benign. However, for efficiency it makes sense to keep
4642 //   the VM thread from racing with the CMS thread while the latter is
4643 //   dirty card info to the modUnionTable. We therefore also use the
4644 //   CGC_lock to protect the reading of the card table and the mod union
4645 //   table by the CM thread.
4646 // . We run concurrently with mutator updates, so scanning
4647 //   needs to be done carefully  -- we should not try to scan
4648 //   potentially uninitialized objects.
4649 //
4650 // Locking strategy: While holding the CGC_lock, we scan over and
4651 // reset a maximal dirty range of the mod union / card tables, then lock
4652 // the free_list_lock and bitmap lock to do a full marking, then
4653 // release these locks; and repeat the cycle. This allows for a
4654 // certain amount of fairness in the sharing of these locks between
4655 // the CMS collector on the one hand, and the VM thread and the
4656 // mutators on the other.
4657 
4658 // NOTE: preclean_mod_union_table() and preclean_card_table()
4659 // further below are largely identical; if you need to modify
4660 // one of these methods, please check the other method too.
4661 
4662 size_t CMSCollector::preclean_mod_union_table(
4663   ConcurrentMarkSweepGeneration* gen,
4664   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4665   verify_work_stacks_empty();
4666   verify_overflow_empty();
4667 
4668   // Turn off checking for this method but turn it back on
4669   // selectively.  There are yield points in this method
4670   // but it is difficult to turn the checking off just around
4671   // the yield points.  It is simpler to selectively turn
4672   // it on.
4673   DEBUG_ONLY(RememberKlassesChecker mux(false);)
4674 
4675   // strategy: starting with the first card, accumulate contiguous
4676   // ranges of dirty cards; clear these cards, then scan the region
4677   // covered by these cards.
4678 
4679   // Since all of the MUT is committed ahead, we can just use
4680   // that, in case the generations expand while we are precleaning.
4681   // It might also be fine to just use the committed part of the
4682   // generation, but we might potentially miss cards when the
4683   // generation is rapidly expanding while we are in the midst
4684   // of precleaning.
4685   HeapWord* startAddr = gen->reserved().start();
4686   HeapWord* endAddr   = gen->reserved().end();
4687 
4688   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4689 
4690   size_t numDirtyCards, cumNumDirtyCards;
4691   HeapWord *nextAddr, *lastAddr;
4692   for (cumNumDirtyCards = numDirtyCards = 0,
4693        nextAddr = lastAddr = startAddr;
4694        nextAddr < endAddr;
4695        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4696 
4697     ResourceMark rm;
4698     HandleMark   hm;
4699 
4700     MemRegion dirtyRegion;
4701     {
4702       stopTimer();
4703       // Potential yield point
4704       CMSTokenSync ts(true);
4705       startTimer();
4706       sample_eden();
4707       // Get dirty region starting at nextOffset (inclusive),
4708       // simultaneously clearing it.
4709       dirtyRegion =
4710         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4711       assert(dirtyRegion.start() >= nextAddr,
4712              "returned region inconsistent?");
4713     }
4714     // Remember where the next search should begin.
4715     // The returned region (if non-empty) is a right open interval,
4716     // so lastOffset is obtained from the right end of that
4717     // interval.
4718     lastAddr = dirtyRegion.end();
4719     // Should do something more transparent and less hacky XXX
4720     numDirtyCards =
4721       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4722 
4723     // We'll scan the cards in the dirty region (with periodic
4724     // yields for foreground GC as needed).
4725     if (!dirtyRegion.is_empty()) {
4726       assert(numDirtyCards > 0, "consistency check");
4727       HeapWord* stop_point = NULL;
4728       stopTimer();
4729       // Potential yield point
4730       CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4731                                bitMapLock());
4732       startTimer();
4733       {
4734         verify_work_stacks_empty();
4735         verify_overflow_empty();
4736         sample_eden();
4737         DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4738         stop_point =
4739           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4740       }
4741       if (stop_point != NULL) {
4742         // The careful iteration stopped early either because it found an
4743         // uninitialized object, or because we were in the midst of an
4744         // "abortable preclean", which should now be aborted. Redirty
4745         // the bits corresponding to the partially-scanned or unscanned
4746         // cards. We'll either restart at the next block boundary or
4747         // abort the preclean.
4748         assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4749                (_collectorState == AbortablePreclean && should_abort_preclean()),
4750                "Unparsable objects should only be in perm gen.");
4751         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4752         if (should_abort_preclean()) {
4753           break; // out of preclean loop
4754         } else {
4755           // Compute the next address at which preclean should pick up;
4756           // might need bitMapLock in order to read P-bits.
4757           lastAddr = next_card_start_after_block(stop_point);
4758         }
4759       }
4760     } else {
4761       assert(lastAddr == endAddr, "consistency check");
4762       assert(numDirtyCards == 0, "consistency check");
4763       break;
4764     }
4765   }
4766   verify_work_stacks_empty();
4767   verify_overflow_empty();
4768   return cumNumDirtyCards;
4769 }
4770 
4771 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4772 // below are largely identical; if you need to modify
4773 // one of these methods, please check the other method too.
4774 
4775 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4776   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4777   // strategy: it's similar to precleamModUnionTable above, in that
4778   // we accumulate contiguous ranges of dirty cards, mark these cards
4779   // precleaned, then scan the region covered by these cards.
4780   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
4781   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4782 
4783   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4784 
4785   size_t numDirtyCards, cumNumDirtyCards;
4786   HeapWord *lastAddr, *nextAddr;
4787 
4788   for (cumNumDirtyCards = numDirtyCards = 0,
4789        nextAddr = lastAddr = startAddr;
4790        nextAddr < endAddr;
4791        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4792 
4793     ResourceMark rm;
4794     HandleMark   hm;
4795 
4796     MemRegion dirtyRegion;
4797     {
4798       // See comments in "Precleaning notes" above on why we
4799       // do this locking. XXX Could the locking overheads be
4800       // too high when dirty cards are sparse? [I don't think so.]
4801       stopTimer();
4802       CMSTokenSync x(true); // is cms thread
4803       startTimer();
4804       sample_eden();
4805       // Get and clear dirty region from card table
4806       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4807                                     MemRegion(nextAddr, endAddr),
4808                                     true,
4809                                     CardTableModRefBS::precleaned_card_val());
4810 
4811       assert(dirtyRegion.start() >= nextAddr,
4812              "returned region inconsistent?");
4813     }
4814     lastAddr = dirtyRegion.end();
4815     numDirtyCards =
4816       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4817 
4818     if (!dirtyRegion.is_empty()) {
4819       stopTimer();
4820       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4821       startTimer();
4822       sample_eden();
4823       verify_work_stacks_empty();
4824       verify_overflow_empty();
4825       DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4826       HeapWord* stop_point =
4827         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4828       if (stop_point != NULL) {
4829         // The careful iteration stopped early because it found an
4830         // uninitialized object.  Redirty the bits corresponding to the
4831         // partially-scanned or unscanned cards, and start again at the
4832         // next block boundary.
4833         assert(CMSPermGenPrecleaningEnabled ||
4834                (_collectorState == AbortablePreclean && should_abort_preclean()),
4835                "Unparsable objects should only be in perm gen.");
4836         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4837         if (should_abort_preclean()) {
4838           break; // out of preclean loop
4839         } else {
4840           // Compute the next address at which preclean should pick up.
4841           lastAddr = next_card_start_after_block(stop_point);
4842         }
4843       }
4844     } else {
4845       break;
4846     }
4847   }
4848   verify_work_stacks_empty();
4849   verify_overflow_empty();
4850   return cumNumDirtyCards;
4851 }
4852 
4853 void CMSCollector::checkpointRootsFinal(bool asynch,
4854   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4855   assert(_collectorState == FinalMarking, "incorrect state transition?");
4856   check_correct_thread_executing();
4857   // world is stopped at this checkpoint
4858   assert(SafepointSynchronize::is_at_safepoint(),
4859          "world should be stopped");
4860   TraceCMSMemoryManagerStats tms(_collectorState);
4861   verify_work_stacks_empty();
4862   verify_overflow_empty();
4863 
4864   SpecializationStats::clear();
4865   if (PrintGCDetails) {
4866     gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4867                         _young_gen->used() / K,
4868                         _young_gen->capacity() / K);
4869   }
4870   if (asynch) {
4871     if (CMSScavengeBeforeRemark) {
4872       GenCollectedHeap* gch = GenCollectedHeap::heap();
4873       // Temporarily set flag to false, GCH->do_collection will
4874       // expect it to be false and set to true
4875       FlagSetting fl(gch->_is_gc_active, false);
4876       NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4877         PrintGCDetails && Verbose, true, gclog_or_tty);)
4878       int level = _cmsGen->level() - 1;
4879       if (level >= 0) {
4880         gch->do_collection(true,        // full (i.e. force, see below)
4881                            false,       // !clear_all_soft_refs
4882                            0,           // size
4883                            false,       // is_tlab
4884                            level        // max_level
4885                           );
4886       }
4887     }
4888     FreelistLocker x(this);
4889     MutexLockerEx y(bitMapLock(),
4890                     Mutex::_no_safepoint_check_flag);
4891     assert(!init_mark_was_synchronous, "but that's impossible!");
4892     checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4893   } else {
4894     // already have all the locks
4895     checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4896                              init_mark_was_synchronous);
4897   }
4898   verify_work_stacks_empty();
4899   verify_overflow_empty();
4900   SpecializationStats::print();
4901 }
4902 
4903 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4904   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4905 
4906   NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4907 
4908   assert(haveFreelistLocks(), "must have free list locks");
4909   assert_lock_strong(bitMapLock());
4910 
4911   if (UseAdaptiveSizePolicy) {
4912     size_policy()->checkpoint_roots_final_begin();
4913   }
4914 
4915   ResourceMark rm;
4916   HandleMark   hm;
4917 
4918   GenCollectedHeap* gch = GenCollectedHeap::heap();
4919 
4920   if (should_unload_classes()) {
4921     CodeCache::gc_prologue();
4922   }
4923   assert(haveFreelistLocks(), "must have free list locks");
4924   assert_lock_strong(bitMapLock());
4925 
4926   DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4927   if (!init_mark_was_synchronous) {
4928     // We might assume that we need not fill TLAB's when
4929     // CMSScavengeBeforeRemark is set, because we may have just done
4930     // a scavenge which would have filled all TLAB's -- and besides
4931     // Eden would be empty. This however may not always be the case --
4932     // for instance although we asked for a scavenge, it may not have
4933     // happened because of a JNI critical section. We probably need
4934     // a policy for deciding whether we can in that case wait until
4935     // the critical section releases and then do the remark following
4936     // the scavenge, and skip it here. In the absence of that policy,
4937     // or of an indication of whether the scavenge did indeed occur,
4938     // we cannot rely on TLAB's having been filled and must do
4939     // so here just in case a scavenge did not happen.
4940     gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4941     // Update the saved marks which may affect the root scans.
4942     gch->save_marks();
4943 
4944     {
4945       COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4946 
4947       // Note on the role of the mod union table:
4948       // Since the marker in "markFromRoots" marks concurrently with
4949       // mutators, it is possible for some reachable objects not to have been
4950       // scanned. For instance, an only reference to an object A was
4951       // placed in object B after the marker scanned B. Unless B is rescanned,
4952       // A would be collected. Such updates to references in marked objects
4953       // are detected via the mod union table which is the set of all cards
4954       // dirtied since the first checkpoint in this GC cycle and prior to
4955       // the most recent young generation GC, minus those cleaned up by the
4956       // concurrent precleaning.
4957       if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4958         TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4959         do_remark_parallel();
4960       } else {
4961         TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4962                     gclog_or_tty);
4963         do_remark_non_parallel();
4964       }
4965     }
4966   } else {
4967     assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4968     // The initial mark was stop-world, so there's no rescanning to
4969     // do; go straight on to the next step below.
4970   }
4971   verify_work_stacks_empty();
4972   verify_overflow_empty();
4973 
4974   {
4975     NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4976     refProcessingWork(asynch, clear_all_soft_refs);
4977   }
4978   verify_work_stacks_empty();
4979   verify_overflow_empty();
4980 
4981   if (should_unload_classes()) {
4982     CodeCache::gc_epilogue();
4983   }
4984   JvmtiExport::gc_epilogue();
4985 
4986   // If we encountered any (marking stack / work queue) overflow
4987   // events during the current CMS cycle, take appropriate
4988   // remedial measures, where possible, so as to try and avoid
4989   // recurrence of that condition.
4990   assert(_markStack.isEmpty(), "No grey objects");
4991   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4992                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4993   if (ser_ovflw > 0) {
4994     if (PrintCMSStatistics != 0) {
4995       gclog_or_tty->print_cr("Marking stack overflow (benign) "
4996         "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4997         ", kac_preclean="SIZE_FORMAT")",
4998         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4999         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5000     }
5001     _markStack.expand();
5002     _ser_pmc_remark_ovflw = 0;
5003     _ser_pmc_preclean_ovflw = 0;
5004     _ser_kac_preclean_ovflw = 0;
5005     _ser_kac_ovflw = 0;
5006   }
5007   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5008     if (PrintCMSStatistics != 0) {
5009       gclog_or_tty->print_cr("Work queue overflow (benign) "
5010         "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5011         _par_pmc_remark_ovflw, _par_kac_ovflw);
5012     }
5013     _par_pmc_remark_ovflw = 0;
5014     _par_kac_ovflw = 0;
5015   }
5016   if (PrintCMSStatistics != 0) {
5017      if (_markStack._hit_limit > 0) {
5018        gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5019                               _markStack._hit_limit);
5020      }
5021      if (_markStack._failed_double > 0) {
5022        gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5023                               " current capacity "SIZE_FORMAT,
5024                               _markStack._failed_double,
5025                               _markStack.capacity());
5026      }
5027   }
5028   _markStack._hit_limit = 0;
5029   _markStack._failed_double = 0;
5030 
5031   // Check that all the klasses have been checked
5032   assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5033 
5034   if ((VerifyAfterGC || VerifyDuringGC) &&
5035       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5036     verify_after_remark();
5037   }
5038 
5039   // Change under the freelistLocks.
5040   _collectorState = Sweeping;
5041   // Call isAllClear() under bitMapLock
5042   assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5043     " final marking");
5044   if (UseAdaptiveSizePolicy) {
5045     size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5046   }
5047 }
5048 
5049 // Parallel remark task
5050 class CMSParRemarkTask: public AbstractGangTask {
5051   CMSCollector* _collector;
5052   int           _n_workers;
5053   CompactibleFreeListSpace* _cms_space;
5054   CompactibleFreeListSpace* _perm_space;
5055 
5056   // The per-thread work queues, available here for stealing.
5057   OopTaskQueueSet*       _task_queues;
5058   ParallelTaskTerminator _term;
5059 
5060  public:
5061   CMSParRemarkTask(CMSCollector* collector,
5062                    CompactibleFreeListSpace* cms_space,
5063                    CompactibleFreeListSpace* perm_space,
5064                    int n_workers, FlexibleWorkGang* workers,
5065                    OopTaskQueueSet* task_queues):
5066     AbstractGangTask("Rescan roots and grey objects in parallel"),
5067     _collector(collector),
5068     _cms_space(cms_space), _perm_space(perm_space),
5069     _n_workers(n_workers),
5070     _task_queues(task_queues),
5071     _term(n_workers, task_queues) { }
5072 
5073   OopTaskQueueSet* task_queues() { return _task_queues; }
5074 
5075   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5076 
5077   ParallelTaskTerminator* terminator() { return &_term; }
5078   int n_workers() { return _n_workers; }
5079 
5080   void work(int i);
5081 
5082  private:
5083   // Work method in support of parallel rescan ... of young gen spaces
5084   void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5085                              ContiguousSpace* space,
5086                              HeapWord** chunk_array, size_t chunk_top);
5087 
5088   // ... of  dirty cards in old space
5089   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5090                                   Par_MarkRefsIntoAndScanClosure* cl);
5091 
5092   // ... work stealing for the above
5093   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5094 };
5095 
5096 // work_queue(i) is passed to the closure
5097 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
5098 // also is passed to do_dirty_card_rescan_tasks() and to
5099 // do_work_steal() to select the i-th task_queue.
5100 
5101 void CMSParRemarkTask::work(int i) {
5102   elapsedTimer _timer;
5103   ResourceMark rm;
5104   HandleMark   hm;
5105 
5106   // ---------- rescan from roots --------------
5107   _timer.start();
5108   GenCollectedHeap* gch = GenCollectedHeap::heap();
5109   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5110     _collector->_span, _collector->ref_processor(),
5111     &(_collector->_markBitMap),
5112     work_queue(i), &(_collector->_revisitStack));
5113 
5114   // Rescan young gen roots first since these are likely
5115   // coarsely partitioned and may, on that account, constitute
5116   // the critical path; thus, it's best to start off that
5117   // work first.
5118   // ---------- young gen roots --------------
5119   {
5120     DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5121     EdenSpace* eden_space = dng->eden();
5122     ContiguousSpace* from_space = dng->from();
5123     ContiguousSpace* to_space   = dng->to();
5124 
5125     HeapWord** eca = _collector->_eden_chunk_array;
5126     size_t     ect = _collector->_eden_chunk_index;
5127     HeapWord** sca = _collector->_survivor_chunk_array;
5128     size_t     sct = _collector->_survivor_chunk_index;
5129 
5130     assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5131     assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5132 
5133     do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
5134     do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
5135     do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
5136 
5137     _timer.stop();
5138     if (PrintCMSStatistics != 0) {
5139       gclog_or_tty->print_cr(
5140         "Finished young gen rescan work in %dth thread: %3.3f sec",
5141         i, _timer.seconds());
5142     }
5143   }
5144 
5145   // ---------- remaining roots --------------
5146   _timer.reset();
5147   _timer.start();
5148   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5149                                 false,     // yg was scanned above
5150                                 false,     // this is parallel code
5151                                 true,      // collecting perm gen
5152                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5153                                 &par_mrias_cl,
5154                                 true,   // walk all of code cache if (so & SO_CodeCache)
5155                                 NULL);
5156   assert(_collector->should_unload_classes()
5157          || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5158          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5159   _timer.stop();
5160   if (PrintCMSStatistics != 0) {
5161     gclog_or_tty->print_cr(
5162       "Finished remaining root rescan work in %dth thread: %3.3f sec",
5163       i, _timer.seconds());
5164   }
5165 
5166   // ---------- rescan dirty cards ------------
5167   _timer.reset();
5168   _timer.start();
5169 
5170   // Do the rescan tasks for each of the two spaces
5171   // (cms_space and perm_space) in turn.
5172   // "i" is passed to select the "i-th" task_queue
5173   do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5174   do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5175   _timer.stop();
5176   if (PrintCMSStatistics != 0) {
5177     gclog_or_tty->print_cr(
5178       "Finished dirty card rescan work in %dth thread: %3.3f sec",
5179       i, _timer.seconds());
5180   }
5181 
5182   // ---------- steal work from other threads ...
5183   // ---------- ... and drain overflow list.
5184   _timer.reset();
5185   _timer.start();
5186   do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5187   _timer.stop();
5188   if (PrintCMSStatistics != 0) {
5189     gclog_or_tty->print_cr(
5190       "Finished work stealing in %dth thread: %3.3f sec",
5191       i, _timer.seconds());
5192   }
5193 }
5194 
5195 // Note that parameter "i" is not used.
5196 void
5197 CMSParRemarkTask::do_young_space_rescan(int i,
5198   Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5199   HeapWord** chunk_array, size_t chunk_top) {
5200   // Until all tasks completed:
5201   // . claim an unclaimed task
5202   // . compute region boundaries corresponding to task claimed
5203   //   using chunk_array
5204   // . par_oop_iterate(cl) over that region
5205 
5206   ResourceMark rm;
5207   HandleMark   hm;
5208 
5209   SequentialSubTasksDone* pst = space->par_seq_tasks();
5210   assert(pst->valid(), "Uninitialized use?");
5211 
5212   int nth_task = 0;
5213   int n_tasks  = pst->n_tasks();
5214 
5215   HeapWord *start, *end;
5216   while (!pst->is_task_claimed(/* reference */ nth_task)) {
5217     // We claimed task # nth_task; compute its boundaries.
5218     if (chunk_top == 0) {  // no samples were taken
5219       assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5220       start = space->bottom();
5221       end   = space->top();
5222     } else if (nth_task == 0) {
5223       start = space->bottom();
5224       end   = chunk_array[nth_task];
5225     } else if (nth_task < (jint)chunk_top) {
5226       assert(nth_task >= 1, "Control point invariant");
5227       start = chunk_array[nth_task - 1];
5228       end   = chunk_array[nth_task];
5229     } else {
5230       assert(nth_task == (jint)chunk_top, "Control point invariant");
5231       start = chunk_array[chunk_top - 1];
5232       end   = space->top();
5233     }
5234     MemRegion mr(start, end);
5235     // Verify that mr is in space
5236     assert(mr.is_empty() || space->used_region().contains(mr),
5237            "Should be in space");
5238     // Verify that "start" is an object boundary
5239     assert(mr.is_empty() || oop(mr.start())->is_oop(),
5240            "Should be an oop");
5241     space->par_oop_iterate(mr, cl);
5242   }
5243   pst->all_tasks_completed();
5244 }
5245 
5246 void
5247 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5248   CompactibleFreeListSpace* sp, int i,
5249   Par_MarkRefsIntoAndScanClosure* cl) {
5250   // Until all tasks completed:
5251   // . claim an unclaimed task
5252   // . compute region boundaries corresponding to task claimed
5253   // . transfer dirty bits ct->mut for that region
5254   // . apply rescanclosure to dirty mut bits for that region
5255 
5256   ResourceMark rm;
5257   HandleMark   hm;
5258 
5259   OopTaskQueue* work_q = work_queue(i);
5260   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5261   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5262   // CAUTION: This closure has state that persists across calls to
5263   // the work method dirty_range_iterate_clear() in that it has
5264   // imbedded in it a (subtype of) UpwardsObjectClosure. The
5265   // use of that state in the imbedded UpwardsObjectClosure instance
5266   // assumes that the cards are always iterated (even if in parallel
5267   // by several threads) in monotonically increasing order per each
5268   // thread. This is true of the implementation below which picks
5269   // card ranges (chunks) in monotonically increasing order globally
5270   // and, a-fortiori, in monotonically increasing order per thread
5271   // (the latter order being a subsequence of the former).
5272   // If the work code below is ever reorganized into a more chaotic
5273   // work-partitioning form than the current "sequential tasks"
5274   // paradigm, the use of that persistent state will have to be
5275   // revisited and modified appropriately. See also related
5276   // bug 4756801 work on which should examine this code to make
5277   // sure that the changes there do not run counter to the
5278   // assumptions made here and necessary for correctness and
5279   // efficiency. Note also that this code might yield inefficient
5280   // behaviour in the case of very large objects that span one or
5281   // more work chunks. Such objects would potentially be scanned
5282   // several times redundantly. Work on 4756801 should try and
5283   // address that performance anomaly if at all possible. XXX
5284   MemRegion  full_span  = _collector->_span;
5285   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
5286   CMSMarkStack* rs = &(_collector->_revisitStack);   // shared
5287   MarkFromDirtyCardsClosure
5288     greyRescanClosure(_collector, full_span, // entire span of interest
5289                       sp, bm, work_q, rs, cl);
5290 
5291   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5292   assert(pst->valid(), "Uninitialized use?");
5293   int nth_task = 0;
5294   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5295   MemRegion span = sp->used_region();
5296   HeapWord* start_addr = span.start();
5297   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5298                                            alignment);
5299   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5300   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5301          start_addr, "Check alignment");
5302   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5303          chunk_size, "Check alignment");
5304 
5305   while (!pst->is_task_claimed(/* reference */ nth_task)) {
5306     // Having claimed the nth_task, compute corresponding mem-region,
5307     // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5308     // The alignment restriction ensures that we do not need any
5309     // synchronization with other gang-workers while setting or
5310     // clearing bits in thus chunk of the MUT.
5311     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5312                                     start_addr + (nth_task+1)*chunk_size);
5313     // The last chunk's end might be way beyond end of the
5314     // used region. In that case pull back appropriately.
5315     if (this_span.end() > end_addr) {
5316       this_span.set_end(end_addr);
5317       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5318     }
5319     // Iterate over the dirty cards covering this chunk, marking them
5320     // precleaned, and setting the corresponding bits in the mod union
5321     // table. Since we have been careful to partition at Card and MUT-word
5322     // boundaries no synchronization is needed between parallel threads.
5323     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5324                                                  &modUnionClosure);
5325 
5326     // Having transferred these marks into the modUnionTable,
5327     // rescan the marked objects on the dirty cards in the modUnionTable.
5328     // Even if this is at a synchronous collection, the initial marking
5329     // may have been done during an asynchronous collection so there
5330     // may be dirty bits in the mod-union table.
5331     _collector->_modUnionTable.dirty_range_iterate_clear(
5332                   this_span, &greyRescanClosure);
5333     _collector->_modUnionTable.verifyNoOneBitsInRange(
5334                                  this_span.start(),
5335                                  this_span.end());
5336   }
5337   pst->all_tasks_completed();  // declare that i am done
5338 }
5339 
5340 // . see if we can share work_queues with ParNew? XXX
5341 void
5342 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5343                                 int* seed) {
5344   OopTaskQueue* work_q = work_queue(i);
5345   NOT_PRODUCT(int num_steals = 0;)
5346   oop obj_to_scan;
5347   CMSBitMap* bm = &(_collector->_markBitMap);
5348 
5349   while (true) {
5350     // Completely finish any left over work from (an) earlier round(s)
5351     cl->trim_queue(0);
5352     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5353                                          (size_t)ParGCDesiredObjsFromOverflowList);
5354     // Now check if there's any work in the overflow list
5355     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5356     // only affects the number of attempts made to get work from the
5357     // overflow list and does not affect the number of workers.  Just
5358     // pass ParallelGCThreads so this behavior is unchanged.
5359     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5360                                                 work_q,
5361                                                 ParallelGCThreads)) {
5362       // found something in global overflow list;
5363       // not yet ready to go stealing work from others.
5364       // We'd like to assert(work_q->size() != 0, ...)
5365       // because we just took work from the overflow list,
5366       // but of course we can't since all of that could have
5367       // been already stolen from us.
5368       // "He giveth and He taketh away."
5369       continue;
5370     }
5371     // Verify that we have no work before we resort to stealing
5372     assert(work_q->size() == 0, "Have work, shouldn't steal");
5373     // Try to steal from other queues that have work
5374     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5375       NOT_PRODUCT(num_steals++;)
5376       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5377       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5378       // Do scanning work
5379       obj_to_scan->oop_iterate(cl);
5380       // Loop around, finish this work, and try to steal some more
5381     } else if (terminator()->offer_termination()) {
5382         break;  // nirvana from the infinite cycle
5383     }
5384   }
5385   NOT_PRODUCT(
5386     if (PrintCMSStatistics != 0) {
5387       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5388     }
5389   )
5390   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5391          "Else our work is not yet done");
5392 }
5393 
5394 // Return a thread-local PLAB recording array, as appropriate.
5395 void* CMSCollector::get_data_recorder(int thr_num) {
5396   if (_survivor_plab_array != NULL &&
5397       (CMSPLABRecordAlways ||
5398        (_collectorState > Marking && _collectorState < FinalMarking))) {
5399     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5400     ChunkArray* ca = &_survivor_plab_array[thr_num];
5401     ca->reset();   // clear it so that fresh data is recorded
5402     return (void*) ca;
5403   } else {
5404     return NULL;
5405   }
5406 }
5407 
5408 // Reset all the thread-local PLAB recording arrays
5409 void CMSCollector::reset_survivor_plab_arrays() {
5410   for (uint i = 0; i < ParallelGCThreads; i++) {
5411     _survivor_plab_array[i].reset();
5412   }
5413 }
5414 
5415 // Merge the per-thread plab arrays into the global survivor chunk
5416 // array which will provide the partitioning of the survivor space
5417 // for CMS rescan.
5418 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5419                                               int no_of_gc_threads) {
5420   assert(_survivor_plab_array  != NULL, "Error");
5421   assert(_survivor_chunk_array != NULL, "Error");
5422   assert(_collectorState == FinalMarking, "Error");
5423   for (int j = 0; j < no_of_gc_threads; j++) {
5424     _cursor[j] = 0;
5425   }
5426   HeapWord* top = surv->top();
5427   size_t i;
5428   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
5429     HeapWord* min_val = top;          // Higher than any PLAB address
5430     uint      min_tid = 0;            // position of min_val this round
5431     for (int j = 0; j < no_of_gc_threads; j++) {
5432       ChunkArray* cur_sca = &_survivor_plab_array[j];
5433       if (_cursor[j] == cur_sca->end()) {
5434         continue;
5435       }
5436       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5437       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5438       assert(surv->used_region().contains(cur_val), "Out of bounds value");
5439       if (cur_val < min_val) {
5440         min_tid = j;
5441         min_val = cur_val;
5442       } else {
5443         assert(cur_val < top, "All recorded addresses should be less");
5444       }
5445     }
5446     // At this point min_val and min_tid are respectively
5447     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5448     // and the thread (j) that witnesses that address.
5449     // We record this address in the _survivor_chunk_array[i]
5450     // and increment _cursor[min_tid] prior to the next round i.
5451     if (min_val == top) {
5452       break;
5453     }
5454     _survivor_chunk_array[i] = min_val;
5455     _cursor[min_tid]++;
5456   }
5457   // We are all done; record the size of the _survivor_chunk_array
5458   _survivor_chunk_index = i; // exclusive: [0, i)
5459   if (PrintCMSStatistics > 0) {
5460     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5461   }
5462   // Verify that we used up all the recorded entries
5463   #ifdef ASSERT
5464     size_t total = 0;
5465     for (int j = 0; j < no_of_gc_threads; j++) {
5466       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5467       total += _cursor[j];
5468     }
5469     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5470     // Check that the merged array is in sorted order
5471     if (total > 0) {
5472       for (size_t i = 0; i < total - 1; i++) {
5473         if (PrintCMSStatistics > 0) {
5474           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5475                               i, _survivor_chunk_array[i]);
5476         }
5477         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5478                "Not sorted");
5479       }
5480     }
5481   #endif // ASSERT
5482 }
5483 
5484 // Set up the space's par_seq_tasks structure for work claiming
5485 // for parallel rescan of young gen.
5486 // See ParRescanTask where this is currently used.
5487 void
5488 CMSCollector::
5489 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5490   assert(n_threads > 0, "Unexpected n_threads argument");
5491   DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5492 
5493   // Eden space
5494   {
5495     SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5496     assert(!pst->valid(), "Clobbering existing data?");
5497     // Each valid entry in [0, _eden_chunk_index) represents a task.
5498     size_t n_tasks = _eden_chunk_index + 1;
5499     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5500     // Sets the condition for completion of the subtask (how many threads
5501     // need to finish in order to be done).
5502     pst->set_n_threads(n_threads);
5503     pst->set_n_tasks((int)n_tasks);
5504   }
5505 
5506   // Merge the survivor plab arrays into _survivor_chunk_array
5507   if (_survivor_plab_array != NULL) {
5508     merge_survivor_plab_arrays(dng->from(), n_threads);
5509   } else {
5510     assert(_survivor_chunk_index == 0, "Error");
5511   }
5512 
5513   // To space
5514   {
5515     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5516     assert(!pst->valid(), "Clobbering existing data?");
5517     // Sets the condition for completion of the subtask (how many threads
5518     // need to finish in order to be done).
5519     pst->set_n_threads(n_threads);
5520     pst->set_n_tasks(1);
5521     assert(pst->valid(), "Error");
5522   }
5523 
5524   // From space
5525   {
5526     SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5527     assert(!pst->valid(), "Clobbering existing data?");
5528     size_t n_tasks = _survivor_chunk_index + 1;
5529     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5530     // Sets the condition for completion of the subtask (how many threads
5531     // need to finish in order to be done).
5532     pst->set_n_threads(n_threads);
5533     pst->set_n_tasks((int)n_tasks);
5534     assert(pst->valid(), "Error");
5535   }
5536 }
5537 
5538 // Parallel version of remark
5539 void CMSCollector::do_remark_parallel() {
5540   GenCollectedHeap* gch = GenCollectedHeap::heap();
5541   FlexibleWorkGang* workers = gch->workers();
5542   assert(workers != NULL, "Need parallel worker threads.");
5543   int n_workers = workers->total_workers();
5544   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
5545   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5546 
5547   CMSParRemarkTask tsk(this,
5548     cms_space, perm_space,
5549     n_workers, workers, task_queues());
5550 
5551   // Set up for parallel process_strong_roots work.
5552   gch->set_par_threads(n_workers);
5553   // We won't be iterating over the cards in the card table updating
5554   // the younger_gen cards, so we shouldn't call the following else
5555   // the verification code as well as subsequent younger_refs_iterate
5556   // code would get confused. XXX
5557   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5558 
5559   // The young gen rescan work will not be done as part of
5560   // process_strong_roots (which currently doesn't knw how to
5561   // parallelize such a scan), but rather will be broken up into
5562   // a set of parallel tasks (via the sampling that the [abortable]
5563   // preclean phase did of EdenSpace, plus the [two] tasks of
5564   // scanning the [two] survivor spaces. Further fine-grain
5565   // parallelization of the scanning of the survivor spaces
5566   // themselves, and of precleaning of the younger gen itself
5567   // is deferred to the future.
5568   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5569 
5570   // The dirty card rescan work is broken up into a "sequence"
5571   // of parallel tasks (per constituent space) that are dynamically
5572   // claimed by the parallel threads.
5573   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5574   perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5575 
5576   // It turns out that even when we're using 1 thread, doing the work in a
5577   // separate thread causes wide variance in run times.  We can't help this
5578   // in the multi-threaded case, but we special-case n=1 here to get
5579   // repeatable measurements of the 1-thread overhead of the parallel code.
5580   if (n_workers > 1) {
5581     // Make refs discovery MT-safe
5582     ReferenceProcessorMTMutator mt(ref_processor(), true);
5583     GenCollectedHeap::StrongRootsScope srs(gch);
5584     workers->run_task(&tsk);
5585   } else {
5586     GenCollectedHeap::StrongRootsScope srs(gch);
5587     tsk.work(0);
5588   }
5589   gch->set_par_threads(0);  // 0 ==> non-parallel.
5590   // restore, single-threaded for now, any preserved marks
5591   // as a result of work_q overflow
5592   restore_preserved_marks_if_any();
5593 }
5594 
5595 // Non-parallel version of remark
5596 void CMSCollector::do_remark_non_parallel() {
5597   ResourceMark rm;
5598   HandleMark   hm;
5599   GenCollectedHeap* gch = GenCollectedHeap::heap();
5600   MarkRefsIntoAndScanClosure
5601     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5602              &_markStack, &_revisitStack, this,
5603              false /* should_yield */, false /* not precleaning */);
5604   MarkFromDirtyCardsClosure
5605     markFromDirtyCardsClosure(this, _span,
5606                               NULL,  // space is set further below
5607                               &_markBitMap, &_markStack, &_revisitStack,
5608                               &mrias_cl);
5609   {
5610     TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5611     // Iterate over the dirty cards, setting the corresponding bits in the
5612     // mod union table.
5613     {
5614       ModUnionClosure modUnionClosure(&_modUnionTable);
5615       _ct->ct_bs()->dirty_card_iterate(
5616                       _cmsGen->used_region(),
5617                       &modUnionClosure);
5618       _ct->ct_bs()->dirty_card_iterate(
5619                       _permGen->used_region(),
5620                       &modUnionClosure);
5621     }
5622     // Having transferred these marks into the modUnionTable, we just need
5623     // to rescan the marked objects on the dirty cards in the modUnionTable.
5624     // The initial marking may have been done during an asynchronous
5625     // collection so there may be dirty bits in the mod-union table.
5626     const int alignment =
5627       CardTableModRefBS::card_size * BitsPerWord;
5628     {
5629       // ... First handle dirty cards in CMS gen
5630       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5631       MemRegion ur = _cmsGen->used_region();
5632       HeapWord* lb = ur.start();
5633       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5634       MemRegion cms_span(lb, ub);
5635       _modUnionTable.dirty_range_iterate_clear(cms_span,
5636                                                &markFromDirtyCardsClosure);
5637       verify_work_stacks_empty();
5638       if (PrintCMSStatistics != 0) {
5639         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5640           markFromDirtyCardsClosure.num_dirty_cards());
5641       }
5642     }
5643     {
5644       // .. and then repeat for dirty cards in perm gen
5645       markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5646       MemRegion ur = _permGen->used_region();
5647       HeapWord* lb = ur.start();
5648       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5649       MemRegion perm_span(lb, ub);
5650       _modUnionTable.dirty_range_iterate_clear(perm_span,
5651                                                &markFromDirtyCardsClosure);
5652       verify_work_stacks_empty();
5653       if (PrintCMSStatistics != 0) {
5654         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5655           markFromDirtyCardsClosure.num_dirty_cards());
5656       }
5657     }
5658   }
5659   if (VerifyDuringGC &&
5660       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5661     HandleMark hm;  // Discard invalid handles created during verification
5662     Universe::verify(true);
5663   }
5664   {
5665     TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5666 
5667     verify_work_stacks_empty();
5668 
5669     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5670     GenCollectedHeap::StrongRootsScope srs(gch);
5671     gch->gen_process_strong_roots(_cmsGen->level(),
5672                                   true,  // younger gens as roots
5673                                   false, // use the local StrongRootsScope
5674                                   true,  // collecting perm gen
5675                                   SharedHeap::ScanningOption(roots_scanning_options()),
5676                                   &mrias_cl,
5677                                   true,   // walk code active on stacks
5678                                   NULL);
5679     assert(should_unload_classes()
5680            || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5681            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5682   }
5683   verify_work_stacks_empty();
5684   // Restore evacuated mark words, if any, used for overflow list links
5685   if (!CMSOverflowEarlyRestoration) {
5686     restore_preserved_marks_if_any();
5687   }
5688   verify_overflow_empty();
5689 }
5690 
5691 ////////////////////////////////////////////////////////
5692 // Parallel Reference Processing Task Proxy Class
5693 ////////////////////////////////////////////////////////
5694 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5695   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5696   CMSCollector*          _collector;
5697   CMSBitMap*             _mark_bit_map;
5698   const MemRegion        _span;
5699   ProcessTask&           _task;
5700 
5701 public:
5702   CMSRefProcTaskProxy(ProcessTask&     task,
5703                       CMSCollector*    collector,
5704                       const MemRegion& span,
5705                       CMSBitMap*       mark_bit_map,
5706                       AbstractWorkGang* workers,
5707                       OopTaskQueueSet* task_queues):
5708     // XXX Should superclass AGTWOQ also know about AWG since it knows
5709     // about the task_queues used by the AWG? Then it could initialize
5710     // the terminator() object. See 6984287. The set_for_termination()
5711     // below is a temporary band-aid for the regression in 6984287.
5712     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5713       task_queues),
5714     _task(task),
5715     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5716   {
5717     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5718            "Inconsistency in _span");
5719     set_for_termination(workers->active_workers());
5720   }
5721 
5722   OopTaskQueueSet* task_queues() { return queues(); }
5723 
5724   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5725 
5726   void do_work_steal(int i,
5727                      CMSParDrainMarkingStackClosure* drain,
5728                      CMSParKeepAliveClosure* keep_alive,
5729                      int* seed);
5730 
5731   virtual void work(int i);
5732 };
5733 
5734 void CMSRefProcTaskProxy::work(int i) {
5735   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5736   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5737                                         _mark_bit_map,
5738                                         &_collector->_revisitStack,
5739                                         work_queue(i));
5740   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5741                                                  _mark_bit_map,
5742                                                  &_collector->_revisitStack,
5743                                                  work_queue(i));
5744   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5745   _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5746   if (_task.marks_oops_alive()) {
5747     do_work_steal(i, &par_drain_stack, &par_keep_alive,
5748                   _collector->hash_seed(i));
5749   }
5750   assert(work_queue(i)->size() == 0, "work_queue should be empty");
5751   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5752 }
5753 
5754 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5755   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5756   EnqueueTask& _task;
5757 
5758 public:
5759   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5760     : AbstractGangTask("Enqueue reference objects in parallel"),
5761       _task(task)
5762   { }
5763 
5764   virtual void work(int i)
5765   {
5766     _task.work(i);
5767   }
5768 };
5769 
5770 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5771   MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5772   OopTaskQueue* work_queue):
5773    Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5774    _span(span),
5775    _bit_map(bit_map),
5776    _work_queue(work_queue),
5777    _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5778    _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5779                         (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5780 { }
5781 
5782 // . see if we can share work_queues with ParNew? XXX
5783 void CMSRefProcTaskProxy::do_work_steal(int i,
5784   CMSParDrainMarkingStackClosure* drain,
5785   CMSParKeepAliveClosure* keep_alive,
5786   int* seed) {
5787   OopTaskQueue* work_q = work_queue(i);
5788   NOT_PRODUCT(int num_steals = 0;)
5789   oop obj_to_scan;
5790 
5791   while (true) {
5792     // Completely finish any left over work from (an) earlier round(s)
5793     drain->trim_queue(0);
5794     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5795                                          (size_t)ParGCDesiredObjsFromOverflowList);
5796     // Now check if there's any work in the overflow list
5797     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5798     // only affects the number of attempts made to get work from the
5799     // overflow list and does not affect the number of workers.  Just
5800     // pass ParallelGCThreads so this behavior is unchanged.
5801     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5802                                                 work_q,
5803                                                 ParallelGCThreads)) {
5804       // Found something in global overflow list;
5805       // not yet ready to go stealing work from others.
5806       // We'd like to assert(work_q->size() != 0, ...)
5807       // because we just took work from the overflow list,
5808       // but of course we can't, since all of that might have
5809       // been already stolen from us.
5810       continue;
5811     }
5812     // Verify that we have no work before we resort to stealing
5813     assert(work_q->size() == 0, "Have work, shouldn't steal");
5814     // Try to steal from other queues that have work
5815     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5816       NOT_PRODUCT(num_steals++;)
5817       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5818       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5819       // Do scanning work
5820       obj_to_scan->oop_iterate(keep_alive);
5821       // Loop around, finish this work, and try to steal some more
5822     } else if (terminator()->offer_termination()) {
5823       break;  // nirvana from the infinite cycle
5824     }
5825   }
5826   NOT_PRODUCT(
5827     if (PrintCMSStatistics != 0) {
5828       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5829     }
5830   )
5831 }
5832 
5833 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5834 {
5835   GenCollectedHeap* gch = GenCollectedHeap::heap();
5836   FlexibleWorkGang* workers = gch->workers();
5837   assert(workers != NULL, "Need parallel worker threads.");
5838   CMSRefProcTaskProxy rp_task(task, &_collector,
5839                               _collector.ref_processor()->span(),
5840                               _collector.markBitMap(),
5841                               workers, _collector.task_queues());
5842   workers->run_task(&rp_task);
5843 }
5844 
5845 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5846 {
5847 
5848   GenCollectedHeap* gch = GenCollectedHeap::heap();
5849   FlexibleWorkGang* workers = gch->workers();
5850   assert(workers != NULL, "Need parallel worker threads.");
5851   CMSRefEnqueueTaskProxy enq_task(task);
5852   workers->run_task(&enq_task);
5853 }
5854 
5855 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5856 
5857   ResourceMark rm;
5858   HandleMark   hm;
5859 
5860   ReferenceProcessor* rp = ref_processor();
5861   assert(rp->span().equals(_span), "Spans should be equal");
5862   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5863   // Process weak references.
5864   rp->setup_policy(clear_all_soft_refs);
5865   verify_work_stacks_empty();
5866 
5867   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5868                                           &_markStack, &_revisitStack,
5869                                           false /* !preclean */);
5870   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5871                                 _span, &_markBitMap, &_markStack,
5872                                 &cmsKeepAliveClosure, false /* !preclean */);
5873   {
5874     TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5875     if (rp->processing_is_mt()) {
5876       // Set the degree of MT here.  If the discovery is done MT, there
5877       // may have been a different number of threads doing the discovery
5878       // and a different number of discovered lists may have Ref objects.
5879       // That is OK as long as the Reference lists are balanced (see
5880       // balance_all_queues() and balance_queues()).
5881 
5882 
5883       rp->set_mt_degree(ParallelGCThreads);
5884       CMSRefProcTaskExecutor task_executor(*this);
5885       rp->process_discovered_references(&_is_alive_closure,
5886                                         &cmsKeepAliveClosure,
5887                                         &cmsDrainMarkingStackClosure,
5888                                         &task_executor);
5889     } else {
5890       rp->process_discovered_references(&_is_alive_closure,
5891                                         &cmsKeepAliveClosure,
5892                                         &cmsDrainMarkingStackClosure,
5893                                         NULL);
5894     }
5895     verify_work_stacks_empty();
5896   }
5897 
5898   if (should_unload_classes()) {
5899     {
5900       TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5901 
5902       // Follow SystemDictionary roots and unload classes
5903       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5904 
5905       // Follow CodeCache roots and unload any methods marked for unloading
5906       CodeCache::do_unloading(&_is_alive_closure,
5907                               &cmsKeepAliveClosure,
5908                               purged_class);
5909 
5910       cmsDrainMarkingStackClosure.do_void();
5911       verify_work_stacks_empty();
5912 
5913       // Update subklass/sibling/implementor links in KlassKlass descendants
5914       assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5915       oop k;
5916       while ((k = _revisitStack.pop()) != NULL) {
5917         ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5918                        &_is_alive_closure,
5919                        &cmsKeepAliveClosure);
5920       }
5921       assert(!ClassUnloading ||
5922              (_markStack.isEmpty() && overflow_list_is_empty()),
5923              "Should not have found new reachable objects");
5924       assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5925       cmsDrainMarkingStackClosure.do_void();
5926       verify_work_stacks_empty();
5927     }
5928 
5929     {
5930       TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5931       // Now clean up stale oops in StringTable
5932       StringTable::unlink(&_is_alive_closure);
5933       // Clean up unreferenced symbols in symbol table.
5934       SymbolTable::unlink();
5935     }
5936   }
5937 
5938   verify_work_stacks_empty();
5939   // Restore any preserved marks as a result of mark stack or
5940   // work queue overflow
5941   restore_preserved_marks_if_any();  // done single-threaded for now
5942 
5943   rp->set_enqueuing_is_done(true);
5944   if (rp->processing_is_mt()) {
5945     rp->balance_all_queues();
5946     CMSRefProcTaskExecutor task_executor(*this);
5947     rp->enqueue_discovered_references(&task_executor);
5948   } else {
5949     rp->enqueue_discovered_references(NULL);
5950   }
5951   rp->verify_no_references_recorded();
5952   assert(!rp->discovery_enabled(), "should have been disabled");
5953 }
5954 
5955 #ifndef PRODUCT
5956 void CMSCollector::check_correct_thread_executing() {
5957   Thread* t = Thread::current();
5958   // Only the VM thread or the CMS thread should be here.
5959   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5960          "Unexpected thread type");
5961   // If this is the vm thread, the foreground process
5962   // should not be waiting.  Note that _foregroundGCIsActive is
5963   // true while the foreground collector is waiting.
5964   if (_foregroundGCShouldWait) {
5965     // We cannot be the VM thread
5966     assert(t->is_ConcurrentGC_thread(),
5967            "Should be CMS thread");
5968   } else {
5969     // We can be the CMS thread only if we are in a stop-world
5970     // phase of CMS collection.
5971     if (t->is_ConcurrentGC_thread()) {
5972       assert(_collectorState == InitialMarking ||
5973              _collectorState == FinalMarking,
5974              "Should be a stop-world phase");
5975       // The CMS thread should be holding the CMS_token.
5976       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5977              "Potential interference with concurrently "
5978              "executing VM thread");
5979     }
5980   }
5981 }
5982 #endif
5983 
5984 void CMSCollector::sweep(bool asynch) {
5985   assert(_collectorState == Sweeping, "just checking");
5986   check_correct_thread_executing();
5987   verify_work_stacks_empty();
5988   verify_overflow_empty();
5989   increment_sweep_count();
5990   TraceCMSMemoryManagerStats tms(_collectorState);
5991 
5992   _inter_sweep_timer.stop();
5993   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5994   size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5995 
5996   // PermGen verification support: If perm gen sweeping is disabled in
5997   // this cycle, we preserve the perm gen object "deadness" information
5998   // in the perm_gen_verify_bit_map. In order to do that we traverse
5999   // all blocks in perm gen and mark all dead objects.
6000   if (verifying() && !should_unload_classes()) {
6001     assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6002            "Should have already been allocated");
6003     MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6004                                markBitMap(), perm_gen_verify_bit_map());
6005     if (asynch) {
6006       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6007                                bitMapLock());
6008       _permGen->cmsSpace()->blk_iterate(&mdo);
6009     } else {
6010       // In the case of synchronous sweep, we already have
6011       // the requisite locks/tokens.
6012       _permGen->cmsSpace()->blk_iterate(&mdo);
6013     }
6014   }
6015 
6016   assert(!_intra_sweep_timer.is_active(), "Should not be active");
6017   _intra_sweep_timer.reset();
6018   _intra_sweep_timer.start();
6019   if (asynch) {
6020     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6021     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6022     // First sweep the old gen then the perm gen
6023     {
6024       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6025                                bitMapLock());
6026       sweepWork(_cmsGen, asynch);
6027     }
6028 
6029     // Now repeat for perm gen
6030     if (should_unload_classes()) {
6031       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6032                              bitMapLock());
6033       sweepWork(_permGen, asynch);
6034     }
6035 
6036     // Update Universe::_heap_*_at_gc figures.
6037     // We need all the free list locks to make the abstract state
6038     // transition from Sweeping to Resetting. See detailed note
6039     // further below.
6040     {
6041       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6042                                _permGen->freelistLock());
6043       // Update heap occupancy information which is used as
6044       // input to soft ref clearing policy at the next gc.
6045       Universe::update_heap_info_at_gc();
6046       _collectorState = Resizing;
6047     }
6048   } else {
6049     // already have needed locks
6050     sweepWork(_cmsGen,  asynch);
6051 
6052     if (should_unload_classes()) {
6053       sweepWork(_permGen, asynch);
6054     }
6055     // Update heap occupancy information which is used as
6056     // input to soft ref clearing policy at the next gc.
6057     Universe::update_heap_info_at_gc();
6058     _collectorState = Resizing;
6059   }
6060   verify_work_stacks_empty();
6061   verify_overflow_empty();
6062 
6063   _intra_sweep_timer.stop();
6064   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6065 
6066   _inter_sweep_timer.reset();
6067   _inter_sweep_timer.start();
6068 
6069   update_time_of_last_gc(os::javaTimeMillis());
6070 
6071   // NOTE on abstract state transitions:
6072   // Mutators allocate-live and/or mark the mod-union table dirty
6073   // based on the state of the collection.  The former is done in
6074   // the interval [Marking, Sweeping] and the latter in the interval
6075   // [Marking, Sweeping).  Thus the transitions into the Marking state
6076   // and out of the Sweeping state must be synchronously visible
6077   // globally to the mutators.
6078   // The transition into the Marking state happens with the world
6079   // stopped so the mutators will globally see it.  Sweeping is
6080   // done asynchronously by the background collector so the transition
6081   // from the Sweeping state to the Resizing state must be done
6082   // under the freelistLock (as is the check for whether to
6083   // allocate-live and whether to dirty the mod-union table).
6084   assert(_collectorState == Resizing, "Change of collector state to"
6085     " Resizing must be done under the freelistLocks (plural)");
6086 
6087   // Now that sweeping has been completed, we clear
6088   // the incremental_collection_failed flag,
6089   // thus inviting a younger gen collection to promote into
6090   // this generation. If such a promotion may still fail,
6091   // the flag will be set again when a young collection is
6092   // attempted.
6093   GenCollectedHeap* gch = GenCollectedHeap::heap();
6094   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
6095   gch->update_full_collections_completed(_collection_count_start);
6096 }
6097 
6098 // FIX ME!!! Looks like this belongs in CFLSpace, with
6099 // CMSGen merely delegating to it.
6100 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6101   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6102   HeapWord*  minAddr        = _cmsSpace->bottom();
6103   HeapWord*  largestAddr    =
6104     (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
6105   if (largestAddr == NULL) {
6106     // The dictionary appears to be empty.  In this case
6107     // try to coalesce at the end of the heap.
6108     largestAddr = _cmsSpace->end();
6109   }
6110   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
6111   size_t nearLargestOffset =
6112     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6113   if (PrintFLSStatistics != 0) {
6114     gclog_or_tty->print_cr(
6115       "CMS: Large Block: " PTR_FORMAT ";"
6116       " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6117       largestAddr,
6118       _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6119   }
6120   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6121 }
6122 
6123 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6124   return addr >= _cmsSpace->nearLargestChunk();
6125 }
6126 
6127 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6128   return _cmsSpace->find_chunk_at_end();
6129 }
6130 
6131 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6132                                                     bool full) {
6133   // The next lower level has been collected.  Gather any statistics
6134   // that are of interest at this point.
6135   if (!full && (current_level + 1) == level()) {
6136     // Gather statistics on the young generation collection.
6137     collector()->stats().record_gc0_end(used());
6138   }
6139 }
6140 
6141 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6142   GenCollectedHeap* gch = GenCollectedHeap::heap();
6143   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6144     "Wrong type of heap");
6145   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6146     gch->gen_policy()->size_policy();
6147   assert(sp->is_gc_cms_adaptive_size_policy(),
6148     "Wrong type of size policy");
6149   return sp;
6150 }
6151 
6152 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6153   if (PrintGCDetails && Verbose) {
6154     gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6155   }
6156   _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6157   _debug_collection_type =
6158     (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6159   if (PrintGCDetails && Verbose) {
6160     gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6161   }
6162 }
6163 
6164 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6165   bool asynch) {
6166   // We iterate over the space(s) underlying this generation,
6167   // checking the mark bit map to see if the bits corresponding
6168   // to specific blocks are marked or not. Blocks that are
6169   // marked are live and are not swept up. All remaining blocks
6170   // are swept up, with coalescing on-the-fly as we sweep up
6171   // contiguous free and/or garbage blocks:
6172   // We need to ensure that the sweeper synchronizes with allocators
6173   // and stop-the-world collectors. In particular, the following
6174   // locks are used:
6175   // . CMS token: if this is held, a stop the world collection cannot occur
6176   // . freelistLock: if this is held no allocation can occur from this
6177   //                 generation by another thread
6178   // . bitMapLock: if this is held, no other thread can access or update
6179   //
6180 
6181   // Note that we need to hold the freelistLock if we use
6182   // block iterate below; else the iterator might go awry if
6183   // a mutator (or promotion) causes block contents to change
6184   // (for instance if the allocator divvies up a block).
6185   // If we hold the free list lock, for all practical purposes
6186   // young generation GC's can't occur (they'll usually need to
6187   // promote), so we might as well prevent all young generation
6188   // GC's while we do a sweeping step. For the same reason, we might
6189   // as well take the bit map lock for the entire duration
6190 
6191   // check that we hold the requisite locks
6192   assert(have_cms_token(), "Should hold cms token");
6193   assert(   (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6194          || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6195         "Should possess CMS token to sweep");
6196   assert_lock_strong(gen->freelistLock());
6197   assert_lock_strong(bitMapLock());
6198 
6199   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6200   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
6201   gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6202                                       _inter_sweep_estimate.padded_average(),
6203                                       _intra_sweep_estimate.padded_average());
6204   gen->setNearLargestChunk();
6205 
6206   {
6207     SweepClosure sweepClosure(this, gen, &_markBitMap,
6208                             CMSYield && asynch);
6209     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6210     // We need to free-up/coalesce garbage/blocks from a
6211     // co-terminal free run. This is done in the SweepClosure
6212     // destructor; so, do not remove this scope, else the
6213     // end-of-sweep-census below will be off by a little bit.
6214   }
6215   gen->cmsSpace()->sweep_completed();
6216   gen->cmsSpace()->endSweepFLCensus(sweep_count());
6217   if (should_unload_classes()) {                // unloaded classes this cycle,
6218     _concurrent_cycles_since_last_unload = 0;   // ... reset count
6219   } else {                                      // did not unload classes,
6220     _concurrent_cycles_since_last_unload++;     // ... increment count
6221   }
6222 }
6223 
6224 // Reset CMS data structures (for now just the marking bit map)
6225 // preparatory for the next cycle.
6226 void CMSCollector::reset(bool asynch) {
6227   GenCollectedHeap* gch = GenCollectedHeap::heap();
6228   CMSAdaptiveSizePolicy* sp = size_policy();
6229   AdaptiveSizePolicyOutput(sp, gch->total_collections());
6230   if (asynch) {
6231     CMSTokenSyncWithLocks ts(true, bitMapLock());
6232 
6233     // If the state is not "Resetting", the foreground  thread
6234     // has done a collection and the resetting.
6235     if (_collectorState != Resetting) {
6236       assert(_collectorState == Idling, "The state should only change"
6237         " because the foreground collector has finished the collection");
6238       return;
6239     }
6240 
6241     // Clear the mark bitmap (no grey objects to start with)
6242     // for the next cycle.
6243     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6244     CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6245 
6246     HeapWord* curAddr = _markBitMap.startWord();
6247     while (curAddr < _markBitMap.endWord()) {
6248       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
6249       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6250       _markBitMap.clear_large_range(chunk);
6251       if (ConcurrentMarkSweepThread::should_yield() &&
6252           !foregroundGCIsActive() &&
6253           CMSYield) {
6254         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6255                "CMS thread should hold CMS token");
6256         assert_lock_strong(bitMapLock());
6257         bitMapLock()->unlock();
6258         ConcurrentMarkSweepThread::desynchronize(true);
6259         ConcurrentMarkSweepThread::acknowledge_yield_request();
6260         stopTimer();
6261         if (PrintCMSStatistics != 0) {
6262           incrementYields();
6263         }
6264         icms_wait();
6265 
6266         // See the comment in coordinator_yield()
6267         for (unsigned i = 0; i < CMSYieldSleepCount &&
6268                          ConcurrentMarkSweepThread::should_yield() &&
6269                          !CMSCollector::foregroundGCIsActive(); ++i) {
6270           os::sleep(Thread::current(), 1, false);
6271           ConcurrentMarkSweepThread::acknowledge_yield_request();
6272         }
6273 
6274         ConcurrentMarkSweepThread::synchronize(true);
6275         bitMapLock()->lock_without_safepoint_check();
6276         startTimer();
6277       }
6278       curAddr = chunk.end();
6279     }
6280     // A successful mostly concurrent collection has been done.
6281     // Because only the full (i.e., concurrent mode failure) collections
6282     // are being measured for gc overhead limits, clean the "near" flag
6283     // and count.
6284     sp->reset_gc_overhead_limit_count();
6285     _collectorState = Idling;
6286   } else {
6287     // already have the lock
6288     assert(_collectorState == Resetting, "just checking");
6289     assert_lock_strong(bitMapLock());
6290     _markBitMap.clear_all();
6291     _collectorState = Idling;
6292   }
6293 
6294   // Stop incremental mode after a cycle completes, so that any future cycles
6295   // are triggered by allocation.
6296   stop_icms();
6297 
6298   NOT_PRODUCT(
6299     if (RotateCMSCollectionTypes) {
6300       _cmsGen->rotate_debug_collection_type();
6301     }
6302   )
6303 }
6304 
6305 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6306   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6307   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6308   TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6309   TraceCollectorStats tcs(counters());
6310 
6311   switch (op) {
6312     case CMS_op_checkpointRootsInitial: {
6313       SvcGCMarker sgcm(SvcGCMarker::OTHER);
6314       checkpointRootsInitial(true);       // asynch
6315       if (PrintGC) {
6316         _cmsGen->printOccupancy("initial-mark");
6317       }
6318       break;
6319     }
6320     case CMS_op_checkpointRootsFinal: {
6321       SvcGCMarker sgcm(SvcGCMarker::OTHER);
6322       checkpointRootsFinal(true,    // asynch
6323                            false,   // !clear_all_soft_refs
6324                            false);  // !init_mark_was_synchronous
6325       if (PrintGC) {
6326         _cmsGen->printOccupancy("remark");
6327       }
6328       break;
6329     }
6330     default:
6331       fatal("No such CMS_op");
6332   }
6333 }
6334 
6335 #ifndef PRODUCT
6336 size_t const CMSCollector::skip_header_HeapWords() {
6337   return FreeChunk::header_size();
6338 }
6339 
6340 // Try and collect here conditions that should hold when
6341 // CMS thread is exiting. The idea is that the foreground GC
6342 // thread should not be blocked if it wants to terminate
6343 // the CMS thread and yet continue to run the VM for a while
6344 // after that.
6345 void CMSCollector::verify_ok_to_terminate() const {
6346   assert(Thread::current()->is_ConcurrentGC_thread(),
6347          "should be called by CMS thread");
6348   assert(!_foregroundGCShouldWait, "should be false");
6349   // We could check here that all the various low-level locks
6350   // are not held by the CMS thread, but that is overkill; see
6351   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6352   // is checked.
6353 }
6354 #endif
6355 
6356 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6357    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6358           "missing Printezis mark?");
6359   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6360   size_t size = pointer_delta(nextOneAddr + 1, addr);
6361   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6362          "alignment problem");
6363   assert(size >= 3, "Necessary for Printezis marks to work");
6364   return size;
6365 }
6366 
6367 // A variant of the above (block_size_using_printezis_bits()) except
6368 // that we return 0 if the P-bits are not yet set.
6369 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6370   if (_markBitMap.isMarked(addr + 1)) {
6371     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6372     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6373     size_t size = pointer_delta(nextOneAddr + 1, addr);
6374     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6375            "alignment problem");
6376     assert(size >= 3, "Necessary for Printezis marks to work");
6377     return size;
6378   }
6379   return 0;
6380 }
6381 
6382 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6383   size_t sz = 0;
6384   oop p = (oop)addr;
6385   if (p->klass_or_null() != NULL && p->is_parsable()) {
6386     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6387   } else {
6388     sz = block_size_using_printezis_bits(addr);
6389   }
6390   assert(sz > 0, "size must be nonzero");
6391   HeapWord* next_block = addr + sz;
6392   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
6393                                              CardTableModRefBS::card_size);
6394   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
6395          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6396          "must be different cards");
6397   return next_card;
6398 }
6399 
6400 
6401 // CMS Bit Map Wrapper /////////////////////////////////////////
6402 
6403 // Construct a CMS bit map infrastructure, but don't create the
6404 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6405 // further below.
6406 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6407   _bm(),
6408   _shifter(shifter),
6409   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6410 {
6411   _bmStartWord = 0;
6412   _bmWordSize  = 0;
6413 }
6414 
6415 bool CMSBitMap::allocate(MemRegion mr) {
6416   _bmStartWord = mr.start();
6417   _bmWordSize  = mr.word_size();
6418   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6419                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6420   if (!brs.is_reserved()) {
6421     warning("CMS bit map allocation failure");
6422     return false;
6423   }
6424   // For now we'll just commit all of the bit map up fromt.
6425   // Later on we'll try to be more parsimonious with swap.
6426   if (!_virtual_space.initialize(brs, brs.size())) {
6427     warning("CMS bit map backing store failure");
6428     return false;
6429   }
6430   assert(_virtual_space.committed_size() == brs.size(),
6431          "didn't reserve backing store for all of CMS bit map?");
6432   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6433   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6434          _bmWordSize, "inconsistency in bit map sizing");
6435   _bm.set_size(_bmWordSize >> _shifter);
6436 
6437   // bm.clear(); // can we rely on getting zero'd memory? verify below
6438   assert(isAllClear(),
6439          "Expected zero'd memory from ReservedSpace constructor");
6440   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6441          "consistency check");
6442   return true;
6443 }
6444 
6445 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6446   HeapWord *next_addr, *end_addr, *last_addr;
6447   assert_locked();
6448   assert(covers(mr), "out-of-range error");
6449   // XXX assert that start and end are appropriately aligned
6450   for (next_addr = mr.start(), end_addr = mr.end();
6451        next_addr < end_addr; next_addr = last_addr) {
6452     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6453     last_addr = dirty_region.end();
6454     if (!dirty_region.is_empty()) {
6455       cl->do_MemRegion(dirty_region);
6456     } else {
6457       assert(last_addr == end_addr, "program logic");
6458       return;
6459     }
6460   }
6461 }
6462 
6463 #ifndef PRODUCT
6464 void CMSBitMap::assert_locked() const {
6465   CMSLockVerifier::assert_locked(lock());
6466 }
6467 
6468 bool CMSBitMap::covers(MemRegion mr) const {
6469   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6470   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6471          "size inconsistency");
6472   return (mr.start() >= _bmStartWord) &&
6473          (mr.end()   <= endWord());
6474 }
6475 
6476 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6477     return (start >= _bmStartWord && (start + size) <= endWord());
6478 }
6479 
6480 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6481   // verify that there are no 1 bits in the interval [left, right)
6482   FalseBitMapClosure falseBitMapClosure;
6483   iterate(&falseBitMapClosure, left, right);
6484 }
6485 
6486 void CMSBitMap::region_invariant(MemRegion mr)
6487 {
6488   assert_locked();
6489   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6490   assert(!mr.is_empty(), "unexpected empty region");
6491   assert(covers(mr), "mr should be covered by bit map");
6492   // convert address range into offset range
6493   size_t start_ofs = heapWordToOffset(mr.start());
6494   // Make sure that end() is appropriately aligned
6495   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6496                         (1 << (_shifter+LogHeapWordSize))),
6497          "Misaligned mr.end()");
6498   size_t end_ofs   = heapWordToOffset(mr.end());
6499   assert(end_ofs > start_ofs, "Should mark at least one bit");
6500 }
6501 
6502 #endif
6503 
6504 bool CMSMarkStack::allocate(size_t size) {
6505   // allocate a stack of the requisite depth
6506   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6507                    size * sizeof(oop)));
6508   if (!rs.is_reserved()) {
6509     warning("CMSMarkStack allocation failure");
6510     return false;
6511   }
6512   if (!_virtual_space.initialize(rs, rs.size())) {
6513     warning("CMSMarkStack backing store failure");
6514     return false;
6515   }
6516   assert(_virtual_space.committed_size() == rs.size(),
6517          "didn't reserve backing store for all of CMS stack?");
6518   _base = (oop*)(_virtual_space.low());
6519   _index = 0;
6520   _capacity = size;
6521   NOT_PRODUCT(_max_depth = 0);
6522   return true;
6523 }
6524 
6525 // XXX FIX ME !!! In the MT case we come in here holding a
6526 // leaf lock. For printing we need to take a further lock
6527 // which has lower rank. We need to recallibrate the two
6528 // lock-ranks involved in order to be able to rpint the
6529 // messages below. (Or defer the printing to the caller.
6530 // For now we take the expedient path of just disabling the
6531 // messages for the problematic case.)
6532 void CMSMarkStack::expand() {
6533   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6534   if (_capacity == MarkStackSizeMax) {
6535     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6536       // We print a warning message only once per CMS cycle.
6537       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6538     }
6539     return;
6540   }
6541   // Double capacity if possible
6542   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6543   // Do not give up existing stack until we have managed to
6544   // get the double capacity that we desired.
6545   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6546                    new_capacity * sizeof(oop)));
6547   if (rs.is_reserved()) {
6548     // Release the backing store associated with old stack
6549     _virtual_space.release();
6550     // Reinitialize virtual space for new stack
6551     if (!_virtual_space.initialize(rs, rs.size())) {
6552       fatal("Not enough swap for expanded marking stack");
6553     }
6554     _base = (oop*)(_virtual_space.low());
6555     _index = 0;
6556     _capacity = new_capacity;
6557   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6558     // Failed to double capacity, continue;
6559     // we print a detail message only once per CMS cycle.
6560     gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6561             SIZE_FORMAT"K",
6562             _capacity / K, new_capacity / K);
6563   }
6564 }
6565 
6566 
6567 // Closures
6568 // XXX: there seems to be a lot of code  duplication here;
6569 // should refactor and consolidate common code.
6570 
6571 // This closure is used to mark refs into the CMS generation in
6572 // the CMS bit map. Called at the first checkpoint. This closure
6573 // assumes that we do not need to re-mark dirty cards; if the CMS
6574 // generation on which this is used is not an oldest (modulo perm gen)
6575 // generation then this will lose younger_gen cards!
6576 
6577 MarkRefsIntoClosure::MarkRefsIntoClosure(
6578   MemRegion span, CMSBitMap* bitMap):
6579     _span(span),
6580     _bitMap(bitMap)
6581 {
6582     assert(_ref_processor == NULL, "deliberately left NULL");
6583     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6584 }
6585 
6586 void MarkRefsIntoClosure::do_oop(oop obj) {
6587   // if p points into _span, then mark corresponding bit in _markBitMap
6588   assert(obj->is_oop(), "expected an oop");
6589   HeapWord* addr = (HeapWord*)obj;
6590   if (_span.contains(addr)) {
6591     // this should be made more efficient
6592     _bitMap->mark(addr);
6593   }
6594 }
6595 
6596 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
6597 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6598 
6599 // A variant of the above, used for CMS marking verification.
6600 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6601   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6602     _span(span),
6603     _verification_bm(verification_bm),
6604     _cms_bm(cms_bm)
6605 {
6606     assert(_ref_processor == NULL, "deliberately left NULL");
6607     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6608 }
6609 
6610 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6611   // if p points into _span, then mark corresponding bit in _markBitMap
6612   assert(obj->is_oop(), "expected an oop");
6613   HeapWord* addr = (HeapWord*)obj;
6614   if (_span.contains(addr)) {
6615     _verification_bm->mark(addr);
6616     if (!_cms_bm->isMarked(addr)) {
6617       oop(addr)->print();
6618       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6619       fatal("... aborting");
6620     }
6621   }
6622 }
6623 
6624 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6625 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6626 
6627 //////////////////////////////////////////////////
6628 // MarkRefsIntoAndScanClosure
6629 //////////////////////////////////////////////////
6630 
6631 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6632                                                        ReferenceProcessor* rp,
6633                                                        CMSBitMap* bit_map,
6634                                                        CMSBitMap* mod_union_table,
6635                                                        CMSMarkStack*  mark_stack,
6636                                                        CMSMarkStack*  revisit_stack,
6637                                                        CMSCollector* collector,
6638                                                        bool should_yield,
6639                                                        bool concurrent_precleaning):
6640   _collector(collector),
6641   _span(span),
6642   _bit_map(bit_map),
6643   _mark_stack(mark_stack),
6644   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6645                       mark_stack, revisit_stack, concurrent_precleaning),
6646   _yield(should_yield),
6647   _concurrent_precleaning(concurrent_precleaning),
6648   _freelistLock(NULL)
6649 {
6650   _ref_processor = rp;
6651   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6652 }
6653 
6654 // This closure is used to mark refs into the CMS generation at the
6655 // second (final) checkpoint, and to scan and transitively follow
6656 // the unmarked oops. It is also used during the concurrent precleaning
6657 // phase while scanning objects on dirty cards in the CMS generation.
6658 // The marks are made in the marking bit map and the marking stack is
6659 // used for keeping the (newly) grey objects during the scan.
6660 // The parallel version (Par_...) appears further below.
6661 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6662   if (obj != NULL) {
6663     assert(obj->is_oop(), "expected an oop");
6664     HeapWord* addr = (HeapWord*)obj;
6665     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6666     assert(_collector->overflow_list_is_empty(),
6667            "overflow list should be empty");
6668     if (_span.contains(addr) &&
6669         !_bit_map->isMarked(addr)) {
6670       // mark bit map (object is now grey)
6671       _bit_map->mark(addr);
6672       // push on marking stack (stack should be empty), and drain the
6673       // stack by applying this closure to the oops in the oops popped
6674       // from the stack (i.e. blacken the grey objects)
6675       bool res = _mark_stack->push(obj);
6676       assert(res, "Should have space to push on empty stack");
6677       do {
6678         oop new_oop = _mark_stack->pop();
6679         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6680         assert(new_oop->is_parsable(), "Found unparsable oop");
6681         assert(_bit_map->isMarked((HeapWord*)new_oop),
6682                "only grey objects on this stack");
6683         // iterate over the oops in this oop, marking and pushing
6684         // the ones in CMS heap (i.e. in _span).
6685         new_oop->oop_iterate(&_pushAndMarkClosure);
6686         // check if it's time to yield
6687         do_yield_check();
6688       } while (!_mark_stack->isEmpty() ||
6689                (!_concurrent_precleaning && take_from_overflow_list()));
6690         // if marking stack is empty, and we are not doing this
6691         // during precleaning, then check the overflow list
6692     }
6693     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6694     assert(_collector->overflow_list_is_empty(),
6695            "overflow list was drained above");
6696     // We could restore evacuated mark words, if any, used for
6697     // overflow list links here because the overflow list is
6698     // provably empty here. That would reduce the maximum
6699     // size requirements for preserved_{oop,mark}_stack.
6700     // But we'll just postpone it until we are all done
6701     // so we can just stream through.
6702     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6703       _collector->restore_preserved_marks_if_any();
6704       assert(_collector->no_preserved_marks(), "No preserved marks");
6705     }
6706     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6707            "All preserved marks should have been restored above");
6708   }
6709 }
6710 
6711 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6712 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6713 
6714 void MarkRefsIntoAndScanClosure::do_yield_work() {
6715   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6716          "CMS thread should hold CMS token");
6717   assert_lock_strong(_freelistLock);
6718   assert_lock_strong(_bit_map->lock());
6719   // relinquish the free_list_lock and bitMaplock()
6720   DEBUG_ONLY(RememberKlassesChecker mux(false);)
6721   _bit_map->lock()->unlock();
6722   _freelistLock->unlock();
6723   ConcurrentMarkSweepThread::desynchronize(true);
6724   ConcurrentMarkSweepThread::acknowledge_yield_request();
6725   _collector->stopTimer();
6726   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6727   if (PrintCMSStatistics != 0) {
6728     _collector->incrementYields();
6729   }
6730   _collector->icms_wait();
6731 
6732   // See the comment in coordinator_yield()
6733   for (unsigned i = 0;
6734        i < CMSYieldSleepCount &&
6735        ConcurrentMarkSweepThread::should_yield() &&
6736        !CMSCollector::foregroundGCIsActive();
6737        ++i) {
6738     os::sleep(Thread::current(), 1, false);
6739     ConcurrentMarkSweepThread::acknowledge_yield_request();
6740   }
6741 
6742   ConcurrentMarkSweepThread::synchronize(true);
6743   _freelistLock->lock_without_safepoint_check();
6744   _bit_map->lock()->lock_without_safepoint_check();
6745   _collector->startTimer();
6746 }
6747 
6748 ///////////////////////////////////////////////////////////
6749 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6750 //                                 MarkRefsIntoAndScanClosure
6751 ///////////////////////////////////////////////////////////
6752 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6753   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6754   CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack*  revisit_stack):
6755   _span(span),
6756   _bit_map(bit_map),
6757   _work_queue(work_queue),
6758   _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6759                        (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6760   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6761                           revisit_stack)
6762 {
6763   _ref_processor = rp;
6764   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6765 }
6766 
6767 // This closure is used to mark refs into the CMS generation at the
6768 // second (final) checkpoint, and to scan and transitively follow
6769 // the unmarked oops. The marks are made in the marking bit map and
6770 // the work_queue is used for keeping the (newly) grey objects during
6771 // the scan phase whence they are also available for stealing by parallel
6772 // threads. Since the marking bit map is shared, updates are
6773 // synchronized (via CAS).
6774 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6775   if (obj != NULL) {
6776     // Ignore mark word because this could be an already marked oop
6777     // that may be chained at the end of the overflow list.
6778     assert(obj->is_oop(true), "expected an oop");
6779     HeapWord* addr = (HeapWord*)obj;
6780     if (_span.contains(addr) &&
6781         !_bit_map->isMarked(addr)) {
6782       // mark bit map (object will become grey):
6783       // It is possible for several threads to be
6784       // trying to "claim" this object concurrently;
6785       // the unique thread that succeeds in marking the
6786       // object first will do the subsequent push on
6787       // to the work queue (or overflow list).
6788       if (_bit_map->par_mark(addr)) {
6789         // push on work_queue (which may not be empty), and trim the
6790         // queue to an appropriate length by applying this closure to
6791         // the oops in the oops popped from the stack (i.e. blacken the
6792         // grey objects)
6793         bool res = _work_queue->push(obj);
6794         assert(res, "Low water mark should be less than capacity?");
6795         trim_queue(_low_water_mark);
6796       } // Else, another thread claimed the object
6797     }
6798   }
6799 }
6800 
6801 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6802 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
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.
6806 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6807   oop p, MemRegion mr) {
6808 
6809   size_t size = 0;
6810   HeapWord* addr = (HeapWord*)p;
6811   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6812   assert(_span.contains(addr), "we are scanning the CMS generation");
6813   // check if it's time to yield
6814   if (do_yield_check()) {
6815     // We yielded for some foreground stop-world work,
6816     // and we have been asked to abort this ongoing preclean cycle.
6817     return 0;
6818   }
6819   if (_bitMap->isMarked(addr)) {
6820     // it's marked; is it potentially uninitialized?
6821     if (p->klass_or_null() != NULL) {
6822       // If is_conc_safe is false, the object may be undergoing
6823       // change by the VM outside a safepoint.  Don't try to
6824       // scan it, but rather leave it for the remark phase.
6825       if (CMSPermGenPrecleaningEnabled &&
6826           (!p->is_conc_safe() || !p->is_parsable())) {
6827         // Signal precleaning to redirty the card since
6828         // the klass pointer is already installed.
6829         assert(size == 0, "Initial value");
6830       } else {
6831         assert(p->is_parsable(), "must be parsable.");
6832         // an initialized object; ignore mark word in verification below
6833         // since we are running concurrent with mutators
6834         assert(p->is_oop(true), "should be an oop");
6835         if (p->is_objArray()) {
6836           // objArrays are precisely marked; restrict scanning
6837           // to dirty cards only.
6838           size = CompactibleFreeListSpace::adjustObjectSize(
6839                    p->oop_iterate(_scanningClosure, mr));
6840         } else {
6841           // A non-array may have been imprecisely marked; we need
6842           // to scan object in its entirety.
6843           size = CompactibleFreeListSpace::adjustObjectSize(
6844                    p->oop_iterate(_scanningClosure));
6845         }
6846         #ifdef DEBUG
6847           size_t direct_size =
6848             CompactibleFreeListSpace::adjustObjectSize(p->size());
6849           assert(size == direct_size, "Inconsistency in size");
6850           assert(size >= 3, "Necessary for Printezis marks to work");
6851           if (!_bitMap->isMarked(addr+1)) {
6852             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6853           } else {
6854             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6855             assert(_bitMap->isMarked(addr+size-1),
6856                    "inconsistent Printezis mark");
6857           }
6858         #endif // DEBUG
6859       }
6860     } else {
6861       // an unitialized object
6862       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6863       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6864       size = pointer_delta(nextOneAddr + 1, addr);
6865       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6866              "alignment problem");
6867       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6868       // will dirty the card when the klass pointer is installed in the
6869       // object (signalling the completion of initialization).
6870     }
6871   } else {
6872     // Either a not yet marked object or an uninitialized object
6873     if (p->klass_or_null() == NULL || !p->is_parsable()) {
6874       // An uninitialized object, skip to the next card, since
6875       // we may not be able to read its P-bits yet.
6876       assert(size == 0, "Initial value");
6877     } else {
6878       // An object not (yet) reached by marking: we merely need to
6879       // compute its size so as to go look at the next block.
6880       assert(p->is_oop(true), "should be an oop");
6881       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6882     }
6883   }
6884   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6885   return size;
6886 }
6887 
6888 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6889   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6890          "CMS thread should hold CMS token");
6891   assert_lock_strong(_freelistLock);
6892   assert_lock_strong(_bitMap->lock());
6893   DEBUG_ONLY(RememberKlassesChecker mux(false);)
6894   // relinquish the free_list_lock and bitMaplock()
6895   _bitMap->lock()->unlock();
6896   _freelistLock->unlock();
6897   ConcurrentMarkSweepThread::desynchronize(true);
6898   ConcurrentMarkSweepThread::acknowledge_yield_request();
6899   _collector->stopTimer();
6900   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6901   if (PrintCMSStatistics != 0) {
6902     _collector->incrementYields();
6903   }
6904   _collector->icms_wait();
6905 
6906   // See the comment in coordinator_yield()
6907   for (unsigned i = 0; i < CMSYieldSleepCount &&
6908                    ConcurrentMarkSweepThread::should_yield() &&
6909                    !CMSCollector::foregroundGCIsActive(); ++i) {
6910     os::sleep(Thread::current(), 1, false);
6911     ConcurrentMarkSweepThread::acknowledge_yield_request();
6912   }
6913 
6914   ConcurrentMarkSweepThread::synchronize(true);
6915   _freelistLock->lock_without_safepoint_check();
6916   _bitMap->lock()->lock_without_safepoint_check();
6917   _collector->startTimer();
6918 }
6919 
6920 
6921 //////////////////////////////////////////////////////////////////
6922 // SurvivorSpacePrecleanClosure
6923 //////////////////////////////////////////////////////////////////
6924 // This (single-threaded) closure is used to preclean the oops in
6925 // the survivor spaces.
6926 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6927 
6928   HeapWord* addr = (HeapWord*)p;
6929   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6930   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6931   assert(p->klass_or_null() != NULL, "object should be initializd");
6932   assert(p->is_parsable(), "must be parsable.");
6933   // an initialized object; ignore mark word in verification below
6934   // since we are running concurrent with mutators
6935   assert(p->is_oop(true), "should be an oop");
6936   // Note that we do not yield while we iterate over
6937   // the interior oops of p, pushing the relevant ones
6938   // on our marking stack.
6939   size_t size = p->oop_iterate(_scanning_closure);
6940   do_yield_check();
6941   // Observe that below, we do not abandon the preclean
6942   // phase as soon as we should; rather we empty the
6943   // marking stack before returning. This is to satisfy
6944   // some existing assertions. In general, it may be a
6945   // good idea to abort immediately and complete the marking
6946   // from the grey objects at a later time.
6947   while (!_mark_stack->isEmpty()) {
6948     oop new_oop = _mark_stack->pop();
6949     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6950     assert(new_oop->is_parsable(), "Found unparsable oop");
6951     assert(_bit_map->isMarked((HeapWord*)new_oop),
6952            "only grey objects on this stack");
6953     // iterate over the oops in this oop, marking and pushing
6954     // the ones in CMS heap (i.e. in _span).
6955     new_oop->oop_iterate(_scanning_closure);
6956     // check if it's time to yield
6957     do_yield_check();
6958   }
6959   unsigned int after_count =
6960     GenCollectedHeap::heap()->total_collections();
6961   bool abort = (_before_count != after_count) ||
6962                _collector->should_abort_preclean();
6963   return abort ? 0 : size;
6964 }
6965 
6966 void SurvivorSpacePrecleanClosure::do_yield_work() {
6967   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6968          "CMS thread should hold CMS token");
6969   assert_lock_strong(_bit_map->lock());
6970   DEBUG_ONLY(RememberKlassesChecker smx(false);)
6971   // Relinquish the bit map lock
6972   _bit_map->lock()->unlock();
6973   ConcurrentMarkSweepThread::desynchronize(true);
6974   ConcurrentMarkSweepThread::acknowledge_yield_request();
6975   _collector->stopTimer();
6976   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6977   if (PrintCMSStatistics != 0) {
6978     _collector->incrementYields();
6979   }
6980   _collector->icms_wait();
6981 
6982   // See the comment in coordinator_yield()
6983   for (unsigned i = 0; i < CMSYieldSleepCount &&
6984                        ConcurrentMarkSweepThread::should_yield() &&
6985                        !CMSCollector::foregroundGCIsActive(); ++i) {
6986     os::sleep(Thread::current(), 1, false);
6987     ConcurrentMarkSweepThread::acknowledge_yield_request();
6988   }
6989 
6990   ConcurrentMarkSweepThread::synchronize(true);
6991   _bit_map->lock()->lock_without_safepoint_check();
6992   _collector->startTimer();
6993 }
6994 
6995 // This closure is used to rescan the marked objects on the dirty cards
6996 // in the mod union table and the card table proper. In the parallel
6997 // case, although the bitMap is shared, we do a single read so the
6998 // isMarked() query is "safe".
6999 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7000   // Ignore mark word because we are running concurrent with mutators
7001   assert(p->is_oop_or_null(true), "expected an oop or null");
7002   HeapWord* addr = (HeapWord*)p;
7003   assert(_span.contains(addr), "we are scanning the CMS generation");
7004   bool is_obj_array = false;
7005   #ifdef DEBUG
7006     if (!_parallel) {
7007       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7008       assert(_collector->overflow_list_is_empty(),
7009              "overflow list should be empty");
7010 
7011     }
7012   #endif // DEBUG
7013   if (_bit_map->isMarked(addr)) {
7014     // Obj arrays are precisely marked, non-arrays are not;
7015     // so we scan objArrays precisely and non-arrays in their
7016     // entirety.
7017     if (p->is_objArray()) {
7018       is_obj_array = true;
7019       if (_parallel) {
7020         p->oop_iterate(_par_scan_closure, mr);
7021       } else {
7022         p->oop_iterate(_scan_closure, mr);
7023       }
7024     } else {
7025       if (_parallel) {
7026         p->oop_iterate(_par_scan_closure);
7027       } else {
7028         p->oop_iterate(_scan_closure);
7029       }
7030     }
7031   }
7032   #ifdef DEBUG
7033     if (!_parallel) {
7034       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7035       assert(_collector->overflow_list_is_empty(),
7036              "overflow list should be empty");
7037 
7038     }
7039   #endif // DEBUG
7040   return is_obj_array;
7041 }
7042 
7043 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7044                         MemRegion span,
7045                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
7046                         CMSMarkStack*  revisitStack,
7047                         bool should_yield, bool verifying):
7048   _collector(collector),
7049   _span(span),
7050   _bitMap(bitMap),
7051   _mut(&collector->_modUnionTable),
7052   _markStack(markStack),
7053   _revisitStack(revisitStack),
7054   _yield(should_yield),
7055   _skipBits(0)
7056 {
7057   assert(_markStack->isEmpty(), "stack should be empty");
7058   _finger = _bitMap->startWord();
7059   _threshold = _finger;
7060   assert(_collector->_restart_addr == NULL, "Sanity check");
7061   assert(_span.contains(_finger), "Out of bounds _finger?");
7062   DEBUG_ONLY(_verifying = verifying;)
7063 }
7064 
7065 void MarkFromRootsClosure::reset(HeapWord* addr) {
7066   assert(_markStack->isEmpty(), "would cause duplicates on stack");
7067   assert(_span.contains(addr), "Out of bounds _finger?");
7068   _finger = addr;
7069   _threshold = (HeapWord*)round_to(
7070                  (intptr_t)_finger, CardTableModRefBS::card_size);
7071 }
7072 
7073 // Should revisit to see if this should be restructured for
7074 // greater efficiency.
7075 bool MarkFromRootsClosure::do_bit(size_t offset) {
7076   if (_skipBits > 0) {
7077     _skipBits--;
7078     return true;
7079   }
7080   // convert offset into a HeapWord*
7081   HeapWord* addr = _bitMap->startWord() + offset;
7082   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7083          "address out of range");
7084   assert(_bitMap->isMarked(addr), "tautology");
7085   if (_bitMap->isMarked(addr+1)) {
7086     // this is an allocated but not yet initialized object
7087     assert(_skipBits == 0, "tautology");
7088     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
7089     oop p = oop(addr);
7090     if (p->klass_or_null() == NULL || !p->is_parsable()) {
7091       DEBUG_ONLY(if (!_verifying) {)
7092         // We re-dirty the cards on which this object lies and increase
7093         // the _threshold so that we'll come back to scan this object
7094         // during the preclean or remark phase. (CMSCleanOnEnter)
7095         if (CMSCleanOnEnter) {
7096           size_t sz = _collector->block_size_using_printezis_bits(addr);
7097           HeapWord* end_card_addr   = (HeapWord*)round_to(
7098                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7099           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7100           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7101           // Bump _threshold to end_card_addr; note that
7102           // _threshold cannot possibly exceed end_card_addr, anyhow.
7103           // This prevents future clearing of the card as the scan proceeds
7104           // to the right.
7105           assert(_threshold <= end_card_addr,
7106                  "Because we are just scanning into this object");
7107           if (_threshold < end_card_addr) {
7108             _threshold = end_card_addr;
7109           }
7110           if (p->klass_or_null() != NULL) {
7111             // Redirty the range of cards...
7112             _mut->mark_range(redirty_range);
7113           } // ...else the setting of klass will dirty the card anyway.
7114         }
7115       DEBUG_ONLY(})
7116       return true;
7117     }
7118   }
7119   scanOopsInOop(addr);
7120   return true;
7121 }
7122 
7123 // We take a break if we've been at this for a while,
7124 // so as to avoid monopolizing the locks involved.
7125 void MarkFromRootsClosure::do_yield_work() {
7126   // First give up the locks, then yield, then re-lock
7127   // We should probably use a constructor/destructor idiom to
7128   // do this unlock/lock or modify the MutexUnlocker class to
7129   // serve our purpose. XXX
7130   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7131          "CMS thread should hold CMS token");
7132   assert_lock_strong(_bitMap->lock());
7133   DEBUG_ONLY(RememberKlassesChecker mux(false);)
7134   _bitMap->lock()->unlock();
7135   ConcurrentMarkSweepThread::desynchronize(true);
7136   ConcurrentMarkSweepThread::acknowledge_yield_request();
7137   _collector->stopTimer();
7138   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7139   if (PrintCMSStatistics != 0) {
7140     _collector->incrementYields();
7141   }
7142   _collector->icms_wait();
7143 
7144   // See the comment in coordinator_yield()
7145   for (unsigned i = 0; i < CMSYieldSleepCount &&
7146                        ConcurrentMarkSweepThread::should_yield() &&
7147                        !CMSCollector::foregroundGCIsActive(); ++i) {
7148     os::sleep(Thread::current(), 1, false);
7149     ConcurrentMarkSweepThread::acknowledge_yield_request();
7150   }
7151 
7152   ConcurrentMarkSweepThread::synchronize(true);
7153   _bitMap->lock()->lock_without_safepoint_check();
7154   _collector->startTimer();
7155 }
7156 
7157 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7158   assert(_bitMap->isMarked(ptr), "expected bit to be set");
7159   assert(_markStack->isEmpty(),
7160          "should drain stack to limit stack usage");
7161   // convert ptr to an oop preparatory to scanning
7162   oop obj = oop(ptr);
7163   // Ignore mark word in verification below, since we
7164   // may be running concurrent with mutators.
7165   assert(obj->is_oop(true), "should be an oop");
7166   assert(_finger <= ptr, "_finger runneth ahead");
7167   // advance the finger to right end of this object
7168   _finger = ptr + obj->size();
7169   assert(_finger > ptr, "we just incremented it above");
7170   // On large heaps, it may take us some time to get through
7171   // the marking phase (especially if running iCMS). During
7172   // this time it's possible that a lot of mutations have
7173   // accumulated in the card table and the mod union table --
7174   // these mutation records are redundant until we have
7175   // actually traced into the corresponding card.
7176   // Here, we check whether advancing the finger would make
7177   // us cross into a new card, and if so clear corresponding
7178   // cards in the MUT (preclean them in the card-table in the
7179   // future).
7180 
7181   DEBUG_ONLY(if (!_verifying) {)
7182     // The clean-on-enter optimization is disabled by default,
7183     // until we fix 6178663.
7184     if (CMSCleanOnEnter && (_finger > _threshold)) {
7185       // [_threshold, _finger) represents the interval
7186       // of cards to be cleared  in MUT (or precleaned in card table).
7187       // The set of cards to be cleared is all those that overlap
7188       // with the interval [_threshold, _finger); note that
7189       // _threshold is always kept card-aligned but _finger isn't
7190       // always card-aligned.
7191       HeapWord* old_threshold = _threshold;
7192       assert(old_threshold == (HeapWord*)round_to(
7193               (intptr_t)old_threshold, CardTableModRefBS::card_size),
7194              "_threshold should always be card-aligned");
7195       _threshold = (HeapWord*)round_to(
7196                      (intptr_t)_finger, CardTableModRefBS::card_size);
7197       MemRegion mr(old_threshold, _threshold);
7198       assert(!mr.is_empty(), "Control point invariant");
7199       assert(_span.contains(mr), "Should clear within span");
7200       // XXX When _finger crosses from old gen into perm gen
7201       // we may be doing unnecessary cleaning; do better in the
7202       // future by detecting that condition and clearing fewer
7203       // MUT/CT entries.
7204       _mut->clear_range(mr);
7205     }
7206   DEBUG_ONLY(})
7207   // Note: the finger doesn't advance while we drain
7208   // the stack below.
7209   PushOrMarkClosure pushOrMarkClosure(_collector,
7210                                       _span, _bitMap, _markStack,
7211                                       _revisitStack,
7212                                       _finger, this);
7213   bool res = _markStack->push(obj);
7214   assert(res, "Empty non-zero size stack should have space for single push");
7215   while (!_markStack->isEmpty()) {
7216     oop new_oop = _markStack->pop();
7217     // Skip verifying header mark word below because we are
7218     // running concurrent with mutators.
7219     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7220     // now scan this oop's oops
7221     new_oop->oop_iterate(&pushOrMarkClosure);
7222     do_yield_check();
7223   }
7224   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7225 }
7226 
7227 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7228                        CMSCollector* collector, MemRegion span,
7229                        CMSBitMap* bit_map,
7230                        OopTaskQueue* work_queue,
7231                        CMSMarkStack*  overflow_stack,
7232                        CMSMarkStack*  revisit_stack,
7233                        bool should_yield):
7234   _collector(collector),
7235   _whole_span(collector->_span),
7236   _span(span),
7237   _bit_map(bit_map),
7238   _mut(&collector->_modUnionTable),
7239   _work_queue(work_queue),
7240   _overflow_stack(overflow_stack),
7241   _revisit_stack(revisit_stack),
7242   _yield(should_yield),
7243   _skip_bits(0),
7244   _task(task)
7245 {
7246   assert(_work_queue->size() == 0, "work_queue should be empty");
7247   _finger = span.start();
7248   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
7249   assert(_span.contains(_finger), "Out of bounds _finger?");
7250 }
7251 
7252 // Should revisit to see if this should be restructured for
7253 // greater efficiency.
7254 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7255   if (_skip_bits > 0) {
7256     _skip_bits--;
7257     return true;
7258   }
7259   // convert offset into a HeapWord*
7260   HeapWord* addr = _bit_map->startWord() + offset;
7261   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7262          "address out of range");
7263   assert(_bit_map->isMarked(addr), "tautology");
7264   if (_bit_map->isMarked(addr+1)) {
7265     // this is an allocated object that might not yet be initialized
7266     assert(_skip_bits == 0, "tautology");
7267     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
7268     oop p = oop(addr);
7269     if (p->klass_or_null() == NULL || !p->is_parsable()) {
7270       // in the case of Clean-on-Enter optimization, redirty card
7271       // and avoid clearing card by increasing  the threshold.
7272       return true;
7273     }
7274   }
7275   scan_oops_in_oop(addr);
7276   return true;
7277 }
7278 
7279 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7280   assert(_bit_map->isMarked(ptr), "expected bit to be set");
7281   // Should we assert that our work queue is empty or
7282   // below some drain limit?
7283   assert(_work_queue->size() == 0,
7284          "should drain stack to limit stack usage");
7285   // convert ptr to an oop preparatory to scanning
7286   oop obj = oop(ptr);
7287   // Ignore mark word in verification below, since we
7288   // may be running concurrent with mutators.
7289   assert(obj->is_oop(true), "should be an oop");
7290   assert(_finger <= ptr, "_finger runneth ahead");
7291   // advance the finger to right end of this object
7292   _finger = ptr + obj->size();
7293   assert(_finger > ptr, "we just incremented it above");
7294   // On large heaps, it may take us some time to get through
7295   // the marking phase (especially if running iCMS). During
7296   // this time it's possible that a lot of mutations have
7297   // accumulated in the card table and the mod union table --
7298   // these mutation records are redundant until we have
7299   // actually traced into the corresponding card.
7300   // Here, we check whether advancing the finger would make
7301   // us cross into a new card, and if so clear corresponding
7302   // cards in the MUT (preclean them in the card-table in the
7303   // future).
7304 
7305   // The clean-on-enter optimization is disabled by default,
7306   // until we fix 6178663.
7307   if (CMSCleanOnEnter && (_finger > _threshold)) {
7308     // [_threshold, _finger) represents the interval
7309     // of cards to be cleared  in MUT (or precleaned in card table).
7310     // The set of cards to be cleared is all those that overlap
7311     // with the interval [_threshold, _finger); note that
7312     // _threshold is always kept card-aligned but _finger isn't
7313     // always card-aligned.
7314     HeapWord* old_threshold = _threshold;
7315     assert(old_threshold == (HeapWord*)round_to(
7316             (intptr_t)old_threshold, CardTableModRefBS::card_size),
7317            "_threshold should always be card-aligned");
7318     _threshold = (HeapWord*)round_to(
7319                    (intptr_t)_finger, CardTableModRefBS::card_size);
7320     MemRegion mr(old_threshold, _threshold);
7321     assert(!mr.is_empty(), "Control point invariant");
7322     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7323     // XXX When _finger crosses from old gen into perm gen
7324     // we may be doing unnecessary cleaning; do better in the
7325     // future by detecting that condition and clearing fewer
7326     // MUT/CT entries.
7327     _mut->clear_range(mr);
7328   }
7329 
7330   // Note: the local finger doesn't advance while we drain
7331   // the stack below, but the global finger sure can and will.
7332   HeapWord** gfa = _task->global_finger_addr();
7333   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7334                                       _span, _bit_map,
7335                                       _work_queue,
7336                                       _overflow_stack,
7337                                       _revisit_stack,
7338                                       _finger,
7339                                       gfa, this);
7340   bool res = _work_queue->push(obj);   // overflow could occur here
7341   assert(res, "Will hold once we use workqueues");
7342   while (true) {
7343     oop new_oop;
7344     if (!_work_queue->pop_local(new_oop)) {
7345       // We emptied our work_queue; check if there's stuff that can
7346       // be gotten from the overflow stack.
7347       if (CMSConcMarkingTask::get_work_from_overflow_stack(
7348             _overflow_stack, _work_queue)) {
7349         do_yield_check();
7350         continue;
7351       } else {  // done
7352         break;
7353       }
7354     }
7355     // Skip verifying header mark word below because we are
7356     // running concurrent with mutators.
7357     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7358     // now scan this oop's oops
7359     new_oop->oop_iterate(&pushOrMarkClosure);
7360     do_yield_check();
7361   }
7362   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7363 }
7364 
7365 // Yield in response to a request from VM Thread or
7366 // from mutators.
7367 void Par_MarkFromRootsClosure::do_yield_work() {
7368   assert(_task != NULL, "sanity");
7369   _task->yield();
7370 }
7371 
7372 // A variant of the above used for verifying CMS marking work.
7373 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7374                         MemRegion span,
7375                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7376                         CMSMarkStack*  mark_stack):
7377   _collector(collector),
7378   _span(span),
7379   _verification_bm(verification_bm),
7380   _cms_bm(cms_bm),
7381   _mark_stack(mark_stack),
7382   _pam_verify_closure(collector, span, verification_bm, cms_bm,
7383                       mark_stack)
7384 {
7385   assert(_mark_stack->isEmpty(), "stack should be empty");
7386   _finger = _verification_bm->startWord();
7387   assert(_collector->_restart_addr == NULL, "Sanity check");
7388   assert(_span.contains(_finger), "Out of bounds _finger?");
7389 }
7390 
7391 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7392   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7393   assert(_span.contains(addr), "Out of bounds _finger?");
7394   _finger = addr;
7395 }
7396 
7397 // Should revisit to see if this should be restructured for
7398 // greater efficiency.
7399 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7400   // convert offset into a HeapWord*
7401   HeapWord* addr = _verification_bm->startWord() + offset;
7402   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7403          "address out of range");
7404   assert(_verification_bm->isMarked(addr), "tautology");
7405   assert(_cms_bm->isMarked(addr), "tautology");
7406 
7407   assert(_mark_stack->isEmpty(),
7408          "should drain stack to limit stack usage");
7409   // convert addr to an oop preparatory to scanning
7410   oop obj = oop(addr);
7411   assert(obj->is_oop(), "should be an oop");
7412   assert(_finger <= addr, "_finger runneth ahead");
7413   // advance the finger to right end of this object
7414   _finger = addr + obj->size();
7415   assert(_finger > addr, "we just incremented it above");
7416   // Note: the finger doesn't advance while we drain
7417   // the stack below.
7418   bool res = _mark_stack->push(obj);
7419   assert(res, "Empty non-zero size stack should have space for single push");
7420   while (!_mark_stack->isEmpty()) {
7421     oop new_oop = _mark_stack->pop();
7422     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7423     // now scan this oop's oops
7424     new_oop->oop_iterate(&_pam_verify_closure);
7425   }
7426   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7427   return true;
7428 }
7429 
7430 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7431   CMSCollector* collector, MemRegion span,
7432   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7433   CMSMarkStack*  mark_stack):
7434   OopClosure(collector->ref_processor()),
7435   _collector(collector),
7436   _span(span),
7437   _verification_bm(verification_bm),
7438   _cms_bm(cms_bm),
7439   _mark_stack(mark_stack)
7440 { }
7441 
7442 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
7443 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7444 
7445 // Upon stack overflow, we discard (part of) the stack,
7446 // remembering the least address amongst those discarded
7447 // in CMSCollector's _restart_address.
7448 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7449   // Remember the least grey address discarded
7450   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7451   _collector->lower_restart_addr(ra);
7452   _mark_stack->reset();  // discard stack contents
7453   _mark_stack->expand(); // expand the stack if possible
7454 }
7455 
7456 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7457   assert(obj->is_oop_or_null(), "expected an oop or NULL");
7458   HeapWord* addr = (HeapWord*)obj;
7459   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7460     // Oop lies in _span and isn't yet grey or black
7461     _verification_bm->mark(addr);            // now grey
7462     if (!_cms_bm->isMarked(addr)) {
7463       oop(addr)->print();
7464       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7465                              addr);
7466       fatal("... aborting");
7467     }
7468 
7469     if (!_mark_stack->push(obj)) { // stack overflow
7470       if (PrintCMSStatistics != 0) {
7471         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7472                                SIZE_FORMAT, _mark_stack->capacity());
7473       }
7474       assert(_mark_stack->isFull(), "Else push should have succeeded");
7475       handle_stack_overflow(addr);
7476     }
7477     // anything including and to the right of _finger
7478     // will be scanned as we iterate over the remainder of the
7479     // bit map
7480   }
7481 }
7482 
7483 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7484                      MemRegion span,
7485                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
7486                      CMSMarkStack*  revisitStack,
7487                      HeapWord* finger, MarkFromRootsClosure* parent) :
7488   KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7489   _span(span),
7490   _bitMap(bitMap),
7491   _markStack(markStack),
7492   _finger(finger),
7493   _parent(parent)
7494 { }
7495 
7496 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7497                      MemRegion span,
7498                      CMSBitMap* bit_map,
7499                      OopTaskQueue* work_queue,
7500                      CMSMarkStack*  overflow_stack,
7501                      CMSMarkStack*  revisit_stack,
7502                      HeapWord* finger,
7503                      HeapWord** global_finger_addr,
7504                      Par_MarkFromRootsClosure* parent) :
7505   Par_KlassRememberingOopClosure(collector,
7506                             collector->ref_processor(),
7507                             revisit_stack),
7508   _whole_span(collector->_span),
7509   _span(span),
7510   _bit_map(bit_map),
7511   _work_queue(work_queue),
7512   _overflow_stack(overflow_stack),
7513   _finger(finger),
7514   _global_finger_addr(global_finger_addr),
7515   _parent(parent)
7516 { }
7517 
7518 // Assumes thread-safe access by callers, who are
7519 // responsible for mutual exclusion.
7520 void CMSCollector::lower_restart_addr(HeapWord* low) {
7521   assert(_span.contains(low), "Out of bounds addr");
7522   if (_restart_addr == NULL) {
7523     _restart_addr = low;
7524   } else {
7525     _restart_addr = MIN2(_restart_addr, low);
7526   }
7527 }
7528 
7529 // Upon stack overflow, we discard (part of) the stack,
7530 // remembering the least address amongst those discarded
7531 // in CMSCollector's _restart_address.
7532 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7533   // Remember the least grey address discarded
7534   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7535   _collector->lower_restart_addr(ra);
7536   _markStack->reset();  // discard stack contents
7537   _markStack->expand(); // expand the stack if possible
7538 }
7539 
7540 // Upon stack overflow, we discard (part of) the stack,
7541 // remembering the least address amongst those discarded
7542 // in CMSCollector's _restart_address.
7543 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7544   // We need to do this under a mutex to prevent other
7545   // workers from interfering with the work done below.
7546   MutexLockerEx ml(_overflow_stack->par_lock(),
7547                    Mutex::_no_safepoint_check_flag);
7548   // Remember the least grey address discarded
7549   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7550   _collector->lower_restart_addr(ra);
7551   _overflow_stack->reset();  // discard stack contents
7552   _overflow_stack->expand(); // expand the stack if possible
7553 }
7554 
7555 void PushOrMarkClosure::do_oop(oop obj) {
7556   // Ignore mark word because we are running concurrent with mutators.
7557   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7558   HeapWord* addr = (HeapWord*)obj;
7559   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7560     // Oop lies in _span and isn't yet grey or black
7561     _bitMap->mark(addr);            // now grey
7562     if (addr < _finger) {
7563       // the bit map iteration has already either passed, or
7564       // sampled, this bit in the bit map; we'll need to
7565       // use the marking stack to scan this oop's oops.
7566       bool simulate_overflow = false;
7567       NOT_PRODUCT(
7568         if (CMSMarkStackOverflowALot &&
7569             _collector->simulate_overflow()) {
7570           // simulate a stack overflow
7571           simulate_overflow = true;
7572         }
7573       )
7574       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7575         if (PrintCMSStatistics != 0) {
7576           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7577                                  SIZE_FORMAT, _markStack->capacity());
7578         }
7579         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7580         handle_stack_overflow(addr);
7581       }
7582     }
7583     // anything including and to the right of _finger
7584     // will be scanned as we iterate over the remainder of the
7585     // bit map
7586     do_yield_check();
7587   }
7588 }
7589 
7590 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
7591 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7592 
7593 void Par_PushOrMarkClosure::do_oop(oop obj) {
7594   // Ignore mark word because we are running concurrent with mutators.
7595   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7596   HeapWord* addr = (HeapWord*)obj;
7597   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7598     // Oop lies in _span and isn't yet grey or black
7599     // We read the global_finger (volatile read) strictly after marking oop
7600     bool res = _bit_map->par_mark(addr);    // now grey
7601     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7602     // Should we push this marked oop on our stack?
7603     // -- if someone else marked it, nothing to do
7604     // -- if target oop is above global finger nothing to do
7605     // -- if target oop is in chunk and above local finger
7606     //      then nothing to do
7607     // -- else push on work queue
7608     if (   !res       // someone else marked it, they will deal with it
7609         || (addr >= *gfa)  // will be scanned in a later task
7610         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7611       return;
7612     }
7613     // the bit map iteration has already either passed, or
7614     // sampled, this bit in the bit map; we'll need to
7615     // use the marking stack to scan this oop's oops.
7616     bool simulate_overflow = false;
7617     NOT_PRODUCT(
7618       if (CMSMarkStackOverflowALot &&
7619           _collector->simulate_overflow()) {
7620         // simulate a stack overflow
7621         simulate_overflow = true;
7622       }
7623     )
7624     if (simulate_overflow ||
7625         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7626       // stack overflow
7627       if (PrintCMSStatistics != 0) {
7628         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7629                                SIZE_FORMAT, _overflow_stack->capacity());
7630       }
7631       // We cannot assert that the overflow stack is full because
7632       // it may have been emptied since.
7633       assert(simulate_overflow ||
7634              _work_queue->size() == _work_queue->max_elems(),
7635             "Else push should have succeeded");
7636       handle_stack_overflow(addr);
7637     }
7638     do_yield_check();
7639   }
7640 }
7641 
7642 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
7643 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7644 
7645 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7646                                              ReferenceProcessor* rp,
7647                                              CMSMarkStack* revisit_stack) :
7648   OopClosure(rp),
7649   _collector(collector),
7650   _revisit_stack(revisit_stack),
7651   _should_remember_klasses(collector->should_unload_classes()) {}
7652 
7653 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7654                                        MemRegion span,
7655                                        ReferenceProcessor* rp,
7656                                        CMSBitMap* bit_map,
7657                                        CMSBitMap* mod_union_table,
7658                                        CMSMarkStack*  mark_stack,
7659                                        CMSMarkStack*  revisit_stack,
7660                                        bool           concurrent_precleaning):
7661   KlassRememberingOopClosure(collector, rp, revisit_stack),
7662   _span(span),
7663   _bit_map(bit_map),
7664   _mod_union_table(mod_union_table),
7665   _mark_stack(mark_stack),
7666   _concurrent_precleaning(concurrent_precleaning)
7667 {
7668   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7669 }
7670 
7671 // Grey object rescan during pre-cleaning and second checkpoint phases --
7672 // the non-parallel version (the parallel version appears further below.)
7673 void PushAndMarkClosure::do_oop(oop obj) {
7674   // Ignore mark word verification. If during concurrent precleaning,
7675   // the object monitor may be locked. If during the checkpoint
7676   // phases, the object may already have been reached by a  different
7677   // path and may be at the end of the global overflow list (so
7678   // the mark word may be NULL).
7679   assert(obj->is_oop_or_null(true /* ignore mark word */),
7680          "expected an oop or NULL");
7681   HeapWord* addr = (HeapWord*)obj;
7682   // Check if oop points into the CMS generation
7683   // and is not marked
7684   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7685     // a white object ...
7686     _bit_map->mark(addr);         // ... now grey
7687     // push on the marking stack (grey set)
7688     bool simulate_overflow = false;
7689     NOT_PRODUCT(
7690       if (CMSMarkStackOverflowALot &&
7691           _collector->simulate_overflow()) {
7692         // simulate a stack overflow
7693         simulate_overflow = true;
7694       }
7695     )
7696     if (simulate_overflow || !_mark_stack->push(obj)) {
7697       if (_concurrent_precleaning) {
7698          // During precleaning we can just dirty the appropriate card(s)
7699          // in the mod union table, thus ensuring that the object remains
7700          // in the grey set  and continue. In the case of object arrays
7701          // we need to dirty all of the cards that the object spans,
7702          // since the rescan of object arrays will be limited to the
7703          // dirty cards.
7704          // Note that no one can be intefering with us in this action
7705          // of dirtying the mod union table, so no locking or atomics
7706          // are required.
7707          if (obj->is_objArray()) {
7708            size_t sz = obj->size();
7709            HeapWord* end_card_addr = (HeapWord*)round_to(
7710                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7711            MemRegion redirty_range = MemRegion(addr, end_card_addr);
7712            assert(!redirty_range.is_empty(), "Arithmetical tautology");
7713            _mod_union_table->mark_range(redirty_range);
7714          } else {
7715            _mod_union_table->mark(addr);
7716          }
7717          _collector->_ser_pmc_preclean_ovflw++;
7718       } else {
7719          // During the remark phase, we need to remember this oop
7720          // in the overflow list.
7721          _collector->push_on_overflow_list(obj);
7722          _collector->_ser_pmc_remark_ovflw++;
7723       }
7724     }
7725   }
7726 }
7727 
7728 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7729                                                MemRegion span,
7730                                                ReferenceProcessor* rp,
7731                                                CMSBitMap* bit_map,
7732                                                OopTaskQueue* work_queue,
7733                                                CMSMarkStack* revisit_stack):
7734   Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7735   _span(span),
7736   _bit_map(bit_map),
7737   _work_queue(work_queue)
7738 {
7739   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7740 }
7741 
7742 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
7743 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7744 
7745 // Grey object rescan during second checkpoint phase --
7746 // the parallel version.
7747 void Par_PushAndMarkClosure::do_oop(oop obj) {
7748   // In the assert below, we ignore the mark word because
7749   // this oop may point to an already visited object that is
7750   // on the overflow stack (in which case the mark word has
7751   // been hijacked for chaining into the overflow stack --
7752   // if this is the last object in the overflow stack then
7753   // its mark word will be NULL). Because this object may
7754   // have been subsequently popped off the global overflow
7755   // stack, and the mark word possibly restored to the prototypical
7756   // value, by the time we get to examined this failing assert in
7757   // the debugger, is_oop_or_null(false) may subsequently start
7758   // to hold.
7759   assert(obj->is_oop_or_null(true),
7760          "expected an oop or NULL");
7761   HeapWord* addr = (HeapWord*)obj;
7762   // Check if oop points into the CMS generation
7763   // and is not marked
7764   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7765     // a white object ...
7766     // If we manage to "claim" the object, by being the
7767     // first thread to mark it, then we push it on our
7768     // marking stack
7769     if (_bit_map->par_mark(addr)) {     // ... now grey
7770       // push on work queue (grey set)
7771       bool simulate_overflow = false;
7772       NOT_PRODUCT(
7773         if (CMSMarkStackOverflowALot &&
7774             _collector->par_simulate_overflow()) {
7775           // simulate a stack overflow
7776           simulate_overflow = true;
7777         }
7778       )
7779       if (simulate_overflow || !_work_queue->push(obj)) {
7780         _collector->par_push_on_overflow_list(obj);
7781         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
7782       }
7783     } // Else, some other thread got there first
7784   }
7785 }
7786 
7787 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
7788 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7789 
7790 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7791   // TBD
7792 }
7793 
7794 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7795   // TBD
7796 }
7797 
7798 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7799   DEBUG_ONLY(RememberKlassesChecker mux(false);)
7800   Mutex* bml = _collector->bitMapLock();
7801   assert_lock_strong(bml);
7802   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7803          "CMS thread should hold CMS token");
7804 
7805   bml->unlock();
7806   ConcurrentMarkSweepThread::desynchronize(true);
7807 
7808   ConcurrentMarkSweepThread::acknowledge_yield_request();
7809 
7810   _collector->stopTimer();
7811   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7812   if (PrintCMSStatistics != 0) {
7813     _collector->incrementYields();
7814   }
7815   _collector->icms_wait();
7816 
7817   // See the comment in coordinator_yield()
7818   for (unsigned i = 0; i < CMSYieldSleepCount &&
7819                        ConcurrentMarkSweepThread::should_yield() &&
7820                        !CMSCollector::foregroundGCIsActive(); ++i) {
7821     os::sleep(Thread::current(), 1, false);
7822     ConcurrentMarkSweepThread::acknowledge_yield_request();
7823   }
7824 
7825   ConcurrentMarkSweepThread::synchronize(true);
7826   bml->lock();
7827 
7828   _collector->startTimer();
7829 }
7830 
7831 bool CMSPrecleanRefsYieldClosure::should_return() {
7832   if (ConcurrentMarkSweepThread::should_yield()) {
7833     do_yield_work();
7834   }
7835   return _collector->foregroundGCIsActive();
7836 }
7837 
7838 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7839   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7840          "mr should be aligned to start at a card boundary");
7841   // We'd like to assert:
7842   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7843   //        "mr should be a range of cards");
7844   // However, that would be too strong in one case -- the last
7845   // partition ends at _unallocated_block which, in general, can be
7846   // an arbitrary boundary, not necessarily card aligned.
7847   if (PrintCMSStatistics != 0) {
7848     _num_dirty_cards +=
7849          mr.word_size()/CardTableModRefBS::card_size_in_words;
7850   }
7851   _space->object_iterate_mem(mr, &_scan_cl);
7852 }
7853 
7854 SweepClosure::SweepClosure(CMSCollector* collector,
7855                            ConcurrentMarkSweepGeneration* g,
7856                            CMSBitMap* bitMap, bool should_yield) :
7857   _collector(collector),
7858   _g(g),
7859   _sp(g->cmsSpace()),
7860   _limit(_sp->sweep_limit()),
7861   _freelistLock(_sp->freelistLock()),
7862   _bitMap(bitMap),
7863   _yield(should_yield),
7864   _inFreeRange(false),           // No free range at beginning of sweep
7865   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7866   _lastFreeRangeCoalesced(false),
7867   _freeFinger(g->used_region().start())
7868 {
7869   NOT_PRODUCT(
7870     _numObjectsFreed = 0;
7871     _numWordsFreed   = 0;
7872     _numObjectsLive = 0;
7873     _numWordsLive = 0;
7874     _numObjectsAlreadyFree = 0;
7875     _numWordsAlreadyFree = 0;
7876     _last_fc = NULL;
7877 
7878     _sp->initializeIndexedFreeListArrayReturnedBytes();
7879     _sp->dictionary()->initializeDictReturnedBytes();
7880   )
7881   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7882          "sweep _limit out of bounds");
7883   if (CMSTraceSweeper) {
7884     gclog_or_tty->print("\n====================\nStarting new sweep\n");
7885   }
7886 }
7887 
7888 // We need this destructor to reclaim any space at the end
7889 // of the space, which do_blk below may not yet have added back to
7890 // the free lists.
7891 SweepClosure::~SweepClosure() {
7892   assert_lock_strong(_freelistLock);
7893   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7894          "sweep _limit out of bounds");
7895   // Flush any remaining coterminal free run as a single
7896   // coalesced chunk to the appropriate free list.
7897   if (inFreeRange()) {
7898     assert(freeFinger() < _limit, "freeFinger points too high");
7899     flush_cur_free_chunk(freeFinger(), pointer_delta(_limit, freeFinger()));
7900     if (CMSTraceSweeper) {
7901       gclog_or_tty->print("Sweep: last chunk: ");
7902       gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") [coalesced:"SIZE_FORMAT"]\n",
7903                           freeFinger(), pointer_delta(_limit, freeFinger()), lastFreeRangeCoalesced());
7904     }
7905   } // else nothing to flush
7906   NOT_PRODUCT(
7907     if (Verbose && PrintGC) {
7908       gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7909                           SIZE_FORMAT " bytes",
7910                  _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7911       gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects,  "
7912                              SIZE_FORMAT" bytes  "
7913         "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7914         _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7915         _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7916       size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7917         sizeof(HeapWord);
7918       gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7919 
7920       if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7921         size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7922         size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7923         size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7924         gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7925         gclog_or_tty->print("   Indexed List Returned "SIZE_FORMAT" bytes",
7926           indexListReturnedBytes);
7927         gclog_or_tty->print_cr("        Dictionary Returned "SIZE_FORMAT" bytes",
7928           dictReturnedBytes);
7929       }
7930     }
7931   )
7932   // Now, in debug mode, just null out the sweep_limit
7933   NOT_PRODUCT(_sp->clear_sweep_limit();)
7934   if (CMSTraceSweeper) {
7935     gclog_or_tty->print("end of sweep\n================\n");
7936   }
7937 }
7938 
7939 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7940     bool freeRangeInFreeLists) {
7941   if (CMSTraceSweeper) {
7942     gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
7943                freeFinger, freeRangeInFreeLists);
7944   }
7945   assert(!inFreeRange(), "Trampling existing free range");
7946   set_inFreeRange(true);
7947   set_lastFreeRangeCoalesced(false);
7948 
7949   set_freeFinger(freeFinger);
7950   set_freeRangeInFreeLists(freeRangeInFreeLists);
7951   if (CMSTestInFreeList) {
7952     if (freeRangeInFreeLists) {
7953       FreeChunk* fc = (FreeChunk*) freeFinger;
7954       assert(fc->isFree(), "A chunk on the free list should be free.");
7955       assert(fc->size() > 0, "Free range should have a size");
7956       assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7957     }
7958   }
7959 }
7960 
7961 // Note that the sweeper runs concurrently with mutators. Thus,
7962 // it is possible for direct allocation in this generation to happen
7963 // in the middle of the sweep. Note that the sweeper also coalesces
7964 // contiguous free blocks. Thus, unless the sweeper and the allocator
7965 // synchronize appropriately freshly allocated blocks may get swept up.
7966 // This is accomplished by the sweeper locking the free lists while
7967 // it is sweeping. Thus blocks that are determined to be free are
7968 // indeed free. There is however one additional complication:
7969 // blocks that have been allocated since the final checkpoint and
7970 // mark, will not have been marked and so would be treated as
7971 // unreachable and swept up. To prevent this, the allocator marks
7972 // the bit map when allocating during the sweep phase. This leads,
7973 // however, to a further complication -- objects may have been allocated
7974 // but not yet initialized -- in the sense that the header isn't yet
7975 // installed. The sweeper can not then determine the size of the block
7976 // in order to skip over it. To deal with this case, we use a technique
7977 // (due to Printezis) to encode such uninitialized block sizes in the
7978 // bit map. Since the bit map uses a bit per every HeapWord, but the
7979 // CMS generation has a minimum object size of 3 HeapWords, it follows
7980 // that "normal marks" won't be adjacent in the bit map (there will
7981 // always be at least two 0 bits between successive 1 bits). We make use
7982 // of these "unused" bits to represent uninitialized blocks -- the bit
7983 // corresponding to the start of the uninitialized object and the next
7984 // bit are both set. Finally, a 1 bit marks the end of the object that
7985 // started with the two consecutive 1 bits to indicate its potentially
7986 // uninitialized state.
7987 
7988 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7989   FreeChunk* fc = (FreeChunk*)addr;
7990   size_t res;
7991 
7992   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7993   // than "addr == _limit" because although _limit was a block boundary when
7994   // we started the sweep, it may no longer be one because heap expansion
7995   // may have caused us to coalesce the block ending at the address _limit
7996   // with a newly expanded chunk (this happens when _limit was set to the
7997   // previous _end of the space), so we may have stepped past _limit; see CR 6977970.
7998   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7999     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8000            "sweep _limit out of bounds");
8001     assert(addr < _sp->end(), "addr out of bounds");
8002     // Flush any remaining coterminal free run as a single
8003     // coalesced chunk to the appropriate free list.
8004     if (inFreeRange()) {
8005       assert(freeFinger() < _limit, "finger points too high");
8006       flush_cur_free_chunk(freeFinger(),
8007                            pointer_delta(addr, freeFinger()));
8008       if (CMSTraceSweeper) {
8009         gclog_or_tty->print("Sweep: last chunk: ");
8010         gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8011                    "[coalesced:"SIZE_FORMAT"]\n",
8012                    freeFinger(), pointer_delta(addr, freeFinger()),
8013                    lastFreeRangeCoalesced());
8014       }
8015     }
8016 
8017     // help the iterator loop finish
8018     return pointer_delta(_sp->end(), addr);
8019   }
8020 
8021   assert(addr < _limit, "sweep invariant");
8022   // check if we should yield
8023   do_yield_check(addr);
8024   if (fc->isFree()) {
8025     // Chunk that is already free
8026     res = fc->size();
8027     do_already_free_chunk(fc);
8028     debug_only(_sp->verifyFreeLists());
8029     assert(res == fc->size(), "Don't expect the size to change");
8030     NOT_PRODUCT(
8031       _numObjectsAlreadyFree++;
8032       _numWordsAlreadyFree += res;
8033     )
8034     NOT_PRODUCT(_last_fc = fc;)
8035   } else if (!_bitMap->isMarked(addr)) {
8036     // Chunk is fresh garbage
8037     res = do_garbage_chunk(fc);
8038     debug_only(_sp->verifyFreeLists());
8039     NOT_PRODUCT(
8040       _numObjectsFreed++;
8041       _numWordsFreed += res;
8042     )
8043   } else {
8044     // Chunk that is alive.
8045     res = do_live_chunk(fc);
8046     debug_only(_sp->verifyFreeLists());
8047     NOT_PRODUCT(
8048         _numObjectsLive++;
8049         _numWordsLive += res;
8050     )
8051   }
8052   return res;
8053 }
8054 
8055 // For the smart allocation, record following
8056 //  split deaths - a free chunk is removed from its free list because
8057 //      it is being split into two or more chunks.
8058 //  split birth - a free chunk is being added to its free list because
8059 //      a larger free chunk has been split and resulted in this free chunk.
8060 //  coal death - a free chunk is being removed from its free list because
8061 //      it is being coalesced into a large free chunk.
8062 //  coal birth - a free chunk is being added to its free list because
8063 //      it was created when two or more free chunks where coalesced into
8064 //      this free chunk.
8065 //
8066 // These statistics are used to determine the desired number of free
8067 // chunks of a given size.  The desired number is chosen to be relative
8068 // to the end of a CMS sweep.  The desired number at the end of a sweep
8069 // is the
8070 //      count-at-end-of-previous-sweep (an amount that was enough)
8071 //              - count-at-beginning-of-current-sweep  (the excess)
8072 //              + split-births  (gains in this size during interval)
8073 //              - split-deaths  (demands on this size during interval)
8074 // where the interval is from the end of one sweep to the end of the
8075 // next.
8076 //
8077 // When sweeping the sweeper maintains an accumulated chunk which is
8078 // the chunk that is made up of chunks that have been coalesced.  That
8079 // will be termed the left-hand chunk.  A new chunk of garbage that
8080 // is being considered for coalescing will be referred to as the
8081 // right-hand chunk.
8082 //
8083 // When making a decision on whether to coalesce a right-hand chunk with
8084 // the current left-hand chunk, the current count vs. the desired count
8085 // of the left-hand chunk is considered.  Also if the right-hand chunk
8086 // is near the large chunk at the end of the heap (see
8087 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8088 // left-hand chunk is coalesced.
8089 //
8090 // When making a decision about whether to split a chunk, the desired count
8091 // vs. the current count of the candidate to be split is also considered.
8092 // If the candidate is underpopulated (currently fewer chunks than desired)
8093 // a chunk of an overpopulated (currently more chunks than desired) size may
8094 // be chosen.  The "hint" associated with a free list, if non-null, points
8095 // to a free list which may be overpopulated.
8096 //
8097 
8098 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8099   size_t size = fc->size();
8100   // Chunks that cannot be coalesced are not in the
8101   // free lists.
8102   if (CMSTestInFreeList && !fc->cantCoalesce()) {
8103     assert(_sp->verifyChunkInFreeLists(fc),
8104       "free chunk should be in free lists");
8105   }
8106   // a chunk that is already free, should not have been
8107   // marked in the bit map
8108   HeapWord* addr = (HeapWord*) fc;
8109   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8110   // Verify that the bit map has no bits marked between
8111   // addr and purported end of this block.
8112   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8113 
8114   // Some chunks cannot be coalesced under any circumstances.
8115   // See the definition of cantCoalesce().
8116   if (!fc->cantCoalesce()) {
8117     // This chunk can potentially be coalesced.
8118     if (_sp->adaptive_freelists()) {
8119       // All the work is done in
8120       do_post_free_or_garbage_chunk(fc, size);
8121     } else {  // Not adaptive free lists
8122       // this is a free chunk that can potentially be coalesced by the sweeper;
8123       if (!inFreeRange()) {
8124         // if the next chunk is a free block that can't be coalesced
8125         // it doesn't make sense to remove this chunk from the free lists
8126         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8127         assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8128         if ((HeapWord*)nextChunk < _sp->end() &&     // There is another free chunk to the right ...
8129             nextChunk->isFree()               &&     // ... which is free...
8130             nextChunk->cantCoalesce()) {             // ... but can't be coalesced
8131           // nothing to do
8132         } else {
8133           // Potentially the start of a new free range:
8134           // Don't eagerly remove it from the free lists.
8135           // No need to remove it if it will just be put
8136           // back again.  (Also from a pragmatic point of view
8137           // if it is a free block in a region that is beyond
8138           // any allocated blocks, an assertion will fail)
8139           // Remember the start of a free run.
8140           initialize_free_range(addr, true);
8141           // end - can coalesce with next chunk
8142         }
8143       } else {
8144         // the midst of a free range, we are coalescing
8145         debug_only(record_free_block_coalesced(fc);)
8146         if (CMSTraceSweeper) {
8147           gclog_or_tty->print("  -- pick up free block 0x%x (%d)\n", fc, size);
8148         }
8149         // remove it from the free lists
8150         _sp->removeFreeChunkFromFreeLists(fc);
8151         set_lastFreeRangeCoalesced(true);
8152         // If the chunk is being coalesced and the current free range is
8153         // in the free lists, remove the current free range so that it
8154         // will be returned to the free lists in its entirety - all
8155         // the coalesced pieces included.
8156         if (freeRangeInFreeLists()) {
8157           FreeChunk* ffc = (FreeChunk*) freeFinger();
8158           assert(ffc->size() == pointer_delta(addr, freeFinger()),
8159             "Size of free range is inconsistent with chunk size.");
8160           if (CMSTestInFreeList) {
8161             assert(_sp->verifyChunkInFreeLists(ffc),
8162               "free range is not in free lists");
8163           }
8164           _sp->removeFreeChunkFromFreeLists(ffc);
8165           set_freeRangeInFreeLists(false);
8166         }
8167       }
8168     }
8169   } else {
8170     // Code path common to both original and adaptive free lists.
8171 
8172     // cant coalesce with previous block; this should be treated
8173     // as the end of a free run if any
8174     if (inFreeRange()) {
8175       // we kicked some butt; time to pick up the garbage
8176       assert(freeFinger() < addr, "freeFinger points too high");
8177       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8178     }
8179     // else, nothing to do, just continue
8180   }
8181 }
8182 
8183 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8184   // This is a chunk of garbage.  It is not in any free list.
8185   // Add it to a free list or let it possibly be coalesced into
8186   // a larger chunk.
8187   HeapWord* addr = (HeapWord*) fc;
8188   size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8189 
8190   if (_sp->adaptive_freelists()) {
8191     // Verify that the bit map has no bits marked between
8192     // addr and purported end of just dead object.
8193     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8194 
8195     do_post_free_or_garbage_chunk(fc, size);
8196   } else {
8197     if (!inFreeRange()) {
8198       // start of a new free range
8199       assert(size > 0, "A free range should have a size");
8200       initialize_free_range(addr, false);
8201 
8202     } else {
8203       // this will be swept up when we hit the end of the
8204       // free range
8205       if (CMSTraceSweeper) {
8206         gclog_or_tty->print("  -- pick up garbage 0x%x (%d) \n", fc, size);
8207       }
8208       // If the chunk is being coalesced and the current free range is
8209       // in the free lists, remove the current free range so that it
8210       // will be returned to the free lists in its entirety - all
8211       // the coalesced pieces included.
8212       if (freeRangeInFreeLists()) {
8213         FreeChunk* ffc = (FreeChunk*)freeFinger();
8214         assert(ffc->size() == pointer_delta(addr, freeFinger()),
8215           "Size of free range is inconsistent with chunk size.");
8216         if (CMSTestInFreeList) {
8217           assert(_sp->verifyChunkInFreeLists(ffc),
8218             "free range is not in free lists");
8219         }
8220         _sp->removeFreeChunkFromFreeLists(ffc);
8221         set_freeRangeInFreeLists(false);
8222       }
8223       set_lastFreeRangeCoalesced(true);
8224     }
8225     // this will be swept up when we hit the end of the free range
8226 
8227     // Verify that the bit map has no bits marked between
8228     // addr and purported end of just dead object.
8229     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8230   }
8231   return size;
8232 }
8233 
8234 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8235   HeapWord* addr = (HeapWord*) fc;
8236   // The sweeper has just found a live object. Return any accumulated
8237   // left hand chunk to the free lists.
8238   if (inFreeRange()) {
8239     assert(freeFinger() < addr, "freeFinger points too high");
8240     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8241   }
8242 
8243   // This object is live: we'd normally expect this to be
8244   // an oop, and like to assert the following:
8245   // assert(oop(addr)->is_oop(), "live block should be an oop");
8246   // However, as we commented above, this may be an object whose
8247   // header hasn't yet been initialized.
8248   size_t size;
8249   assert(_bitMap->isMarked(addr), "Tautology for this control point");
8250   if (_bitMap->isMarked(addr + 1)) {
8251     // Determine the size from the bit map, rather than trying to
8252     // compute it from the object header.
8253     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8254     size = pointer_delta(nextOneAddr + 1, addr);
8255     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8256            "alignment problem");
8257 
8258 #ifdef DEBUG
8259       if (oop(addr)->klass_or_null() != NULL &&
8260           (   !_collector->should_unload_classes()
8261            || (oop(addr)->is_parsable()) &&
8262                oop(addr)->is_conc_safe())) {
8263         // Ignore mark word because we are running concurrent with mutators
8264         assert(oop(addr)->is_oop(true), "live block should be an oop");
8265         // is_conc_safe is checked before performing this assertion
8266         // because an object that is not is_conc_safe may yet have
8267         // the return from size() correct.
8268         assert(size ==
8269                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8270                "P-mark and computed size do not agree");
8271       }
8272 #endif
8273 
8274   } else {
8275     // This should be an initialized object that's alive.
8276     assert(oop(addr)->klass_or_null() != NULL &&
8277            (!_collector->should_unload_classes()
8278             || oop(addr)->is_parsable()),
8279            "Should be an initialized object");
8280     // Note that there are objects used during class redefinition
8281     // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite()
8282     // which are discarded with their is_conc_safe state still
8283     // false.  These object may be floating garbage so may be
8284     // seen here.  If they are floating garbage their size
8285     // should be attainable from their klass.  Do not that
8286     // is_conc_safe() is true for oop(addr).
8287     // Ignore mark word because we are running concurrent with mutators
8288     assert(oop(addr)->is_oop(true), "live block should be an oop");
8289     // Verify that the bit map has no bits marked between
8290     // addr and purported end of this block.
8291     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8292     assert(size >= 3, "Necessary for Printezis marks to work");
8293     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8294     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8295   }
8296   return size;
8297 }
8298 
8299 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8300                                                  size_t chunkSize) {
8301   // do_post_free_or_garbage_chunk() should only be called in the case
8302   // of the adaptive free list allocator.
8303   bool fcInFreeLists = fc->isFree();
8304   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8305   assert((HeapWord*)fc <= _limit, "sweep invariant");
8306   if (CMSTestInFreeList && fcInFreeLists) {
8307     assert(_sp->verifyChunkInFreeLists(fc), "free chunk is not in free lists");
8308   }
8309 
8310   if (CMSTraceSweeper) {
8311     gclog_or_tty->print_cr("  -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8312   }
8313 
8314   HeapWord* addr = (HeapWord*) fc;
8315 
8316   bool coalesce;
8317   size_t left  = pointer_delta(addr, freeFinger());
8318   size_t right = chunkSize;
8319   switch (FLSCoalescePolicy) {
8320     // numeric value forms a coalition aggressiveness metric
8321     case 0:  { // never coalesce
8322       coalesce = false;
8323       break;
8324     }
8325     case 1: { // coalesce if left & right chunks on overpopulated lists
8326       coalesce = _sp->coalOverPopulated(left) &&
8327                  _sp->coalOverPopulated(right);
8328       break;
8329     }
8330     case 2: { // coalesce if left chunk on overpopulated list (default)
8331       coalesce = _sp->coalOverPopulated(left);
8332       break;
8333     }
8334     case 3: { // coalesce if left OR right chunk on overpopulated list
8335       coalesce = _sp->coalOverPopulated(left) ||
8336                  _sp->coalOverPopulated(right);
8337       break;
8338     }
8339     case 4: { // always coalesce
8340       coalesce = true;
8341       break;
8342     }
8343     default:
8344      ShouldNotReachHere();
8345   }
8346 
8347   // Should the current free range be coalesced?
8348   // If the chunk is in a free range and either we decided to coalesce above
8349   // or the chunk is near the large block at the end of the heap
8350   // (isNearLargestChunk() returns true), then coalesce this chunk.
8351   bool doCoalesce = inFreeRange() &&
8352     (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8353   if (doCoalesce) {
8354     // Coalesce the current free range on the left with the new
8355     // chunk on the right.  If either is on a free list,
8356     // it must be removed from the list and stashed in the closure.
8357     if (freeRangeInFreeLists()) {
8358       FreeChunk* ffc = (FreeChunk*)freeFinger();
8359       assert(ffc->size() == pointer_delta(addr, freeFinger()),
8360         "Size of free range is inconsistent with chunk size.");
8361       if (CMSTestInFreeList) {
8362         assert(_sp->verifyChunkInFreeLists(ffc),
8363           "Chunk is not in free lists");
8364       }
8365       _sp->coalDeath(ffc->size());
8366       _sp->removeFreeChunkFromFreeLists(ffc);
8367       set_freeRangeInFreeLists(false);
8368     }
8369     if (fcInFreeLists) {
8370       _sp->coalDeath(chunkSize);
8371       assert(fc->size() == chunkSize,
8372         "The chunk has the wrong size or is not in the free lists");
8373       _sp->removeFreeChunkFromFreeLists(fc);
8374     }
8375     set_lastFreeRangeCoalesced(true);
8376   } else {  // not in a free range and/or should not coalesce
8377     // Return the current free range and start a new one.
8378     if (inFreeRange()) {
8379       // In a free range but cannot coalesce with the right hand chunk.
8380       // Put the current free range into the free lists.
8381       flush_cur_free_chunk(freeFinger(),
8382                            pointer_delta(addr, freeFinger()));
8383     }
8384     // Set up for new free range.  Pass along whether the right hand
8385     // chunk is in the free lists.
8386     initialize_free_range((HeapWord*)fc, fcInFreeLists);
8387   }
8388 }
8389 
8390 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8391   assert(inFreeRange(), "Should only be called if currently in a free range.");
8392   assert(size > 0,
8393     "A zero sized chunk cannot be added to the free lists.");
8394   if (!freeRangeInFreeLists()) {
8395     if (CMSTestInFreeList) {
8396       FreeChunk* fc = (FreeChunk*) chunk;
8397       fc->setSize(size);
8398       assert(!_sp->verifyChunkInFreeLists(fc),
8399         "chunk should not be in free lists yet");
8400     }
8401     if (CMSTraceSweeper) {
8402       gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8403                     chunk, size);
8404     }
8405     // A new free range is going to be starting.  The current
8406     // free range has not been added to the free lists yet or
8407     // was removed so add it back.
8408     // If the current free range was coalesced, then the death
8409     // of the free range was recorded.  Record a birth now.
8410     if (lastFreeRangeCoalesced()) {
8411       _sp->coalBirth(size);
8412     }
8413     _sp->addChunkAndRepairOffsetTable(chunk, size,
8414             lastFreeRangeCoalesced());
8415   }
8416   set_inFreeRange(false);
8417   set_freeRangeInFreeLists(false);
8418 }
8419 
8420 // We take a break if we've been at this for a while,
8421 // so as to avoid monopolizing the locks involved.
8422 void SweepClosure::do_yield_work(HeapWord* addr) {
8423   // Return current free chunk being used for coalescing (if any)
8424   // to the appropriate freelist.  After yielding, the next
8425   // free block encountered will start a coalescing range of
8426   // free blocks.  If the next free block is adjacent to the
8427   // chunk just flushed, they will need to wait for the next
8428   // sweep to be coalesced.
8429   if (inFreeRange()) {
8430     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8431   }
8432 
8433   // First give up the locks, then yield, then re-lock.
8434   // We should probably use a constructor/destructor idiom to
8435   // do this unlock/lock or modify the MutexUnlocker class to
8436   // serve our purpose. XXX
8437   assert_lock_strong(_bitMap->lock());
8438   assert_lock_strong(_freelistLock);
8439   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8440          "CMS thread should hold CMS token");
8441   _bitMap->lock()->unlock();
8442   _freelistLock->unlock();
8443   ConcurrentMarkSweepThread::desynchronize(true);
8444   ConcurrentMarkSweepThread::acknowledge_yield_request();
8445   _collector->stopTimer();
8446   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8447   if (PrintCMSStatistics != 0) {
8448     _collector->incrementYields();
8449   }
8450   _collector->icms_wait();
8451 
8452   // See the comment in coordinator_yield()
8453   for (unsigned i = 0; i < CMSYieldSleepCount &&
8454                        ConcurrentMarkSweepThread::should_yield() &&
8455                        !CMSCollector::foregroundGCIsActive(); ++i) {
8456     os::sleep(Thread::current(), 1, false);
8457     ConcurrentMarkSweepThread::acknowledge_yield_request();
8458   }
8459 
8460   ConcurrentMarkSweepThread::synchronize(true);
8461   _freelistLock->lock();
8462   _bitMap->lock()->lock_without_safepoint_check();
8463   _collector->startTimer();
8464 }
8465 
8466 #ifndef PRODUCT
8467 // This is actually very useful in a product build if it can
8468 // be called from the debugger.  Compile it into the product
8469 // as needed.
8470 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8471   return debug_cms_space->verifyChunkInFreeLists(fc);
8472 }
8473 
8474 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8475   if (CMSTraceSweeper) {
8476     gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8477   }
8478 }
8479 #endif
8480 
8481 // CMSIsAliveClosure
8482 bool CMSIsAliveClosure::do_object_b(oop obj) {
8483   HeapWord* addr = (HeapWord*)obj;
8484   return addr != NULL &&
8485          (!_span.contains(addr) || _bit_map->isMarked(addr));
8486 }
8487 
8488 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8489                       MemRegion span,
8490                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8491                       CMSMarkStack* revisit_stack, bool cpc):
8492   KlassRememberingOopClosure(collector, NULL, revisit_stack),
8493   _span(span),
8494   _bit_map(bit_map),
8495   _mark_stack(mark_stack),
8496   _concurrent_precleaning(cpc) {
8497   assert(!_span.is_empty(), "Empty span could spell trouble");
8498 }
8499 
8500 
8501 // CMSKeepAliveClosure: the serial version
8502 void CMSKeepAliveClosure::do_oop(oop obj) {
8503   HeapWord* addr = (HeapWord*)obj;
8504   if (_span.contains(addr) &&
8505       !_bit_map->isMarked(addr)) {
8506     _bit_map->mark(addr);
8507     bool simulate_overflow = false;
8508     NOT_PRODUCT(
8509       if (CMSMarkStackOverflowALot &&
8510           _collector->simulate_overflow()) {
8511         // simulate a stack overflow
8512         simulate_overflow = true;
8513       }
8514     )
8515     if (simulate_overflow || !_mark_stack->push(obj)) {
8516       if (_concurrent_precleaning) {
8517         // We dirty the overflown object and let the remark
8518         // phase deal with it.
8519         assert(_collector->overflow_list_is_empty(), "Error");
8520         // In the case of object arrays, we need to dirty all of
8521         // the cards that the object spans. No locking or atomics
8522         // are needed since no one else can be mutating the mod union
8523         // table.
8524         if (obj->is_objArray()) {
8525           size_t sz = obj->size();
8526           HeapWord* end_card_addr =
8527             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8528           MemRegion redirty_range = MemRegion(addr, end_card_addr);
8529           assert(!redirty_range.is_empty(), "Arithmetical tautology");
8530           _collector->_modUnionTable.mark_range(redirty_range);
8531         } else {
8532           _collector->_modUnionTable.mark(addr);
8533         }
8534         _collector->_ser_kac_preclean_ovflw++;
8535       } else {
8536         _collector->push_on_overflow_list(obj);
8537         _collector->_ser_kac_ovflw++;
8538       }
8539     }
8540   }
8541 }
8542 
8543 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
8544 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8545 
8546 // CMSParKeepAliveClosure: a parallel version of the above.
8547 // The work queues are private to each closure (thread),
8548 // but (may be) available for stealing by other threads.
8549 void CMSParKeepAliveClosure::do_oop(oop obj) {
8550   HeapWord* addr = (HeapWord*)obj;
8551   if (_span.contains(addr) &&
8552       !_bit_map->isMarked(addr)) {
8553     // In general, during recursive tracing, several threads
8554     // may be concurrently getting here; the first one to
8555     // "tag" it, claims it.
8556     if (_bit_map->par_mark(addr)) {
8557       bool res = _work_queue->push(obj);
8558       assert(res, "Low water mark should be much less than capacity");
8559       // Do a recursive trim in the hope that this will keep
8560       // stack usage lower, but leave some oops for potential stealers
8561       trim_queue(_low_water_mark);
8562     } // Else, another thread got there first
8563   }
8564 }
8565 
8566 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
8567 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8568 
8569 void CMSParKeepAliveClosure::trim_queue(uint max) {
8570   while (_work_queue->size() > max) {
8571     oop new_oop;
8572     if (_work_queue->pop_local(new_oop)) {
8573       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8574       assert(_bit_map->isMarked((HeapWord*)new_oop),
8575              "no white objects on this stack!");
8576       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8577       // iterate over the oops in this oop, marking and pushing
8578       // the ones in CMS heap (i.e. in _span).
8579       new_oop->oop_iterate(&_mark_and_push);
8580     }
8581   }
8582 }
8583 
8584 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8585                                 CMSCollector* collector,
8586                                 MemRegion span, CMSBitMap* bit_map,
8587                                 CMSMarkStack* revisit_stack,
8588                                 OopTaskQueue* work_queue):
8589   Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8590   _span(span),
8591   _bit_map(bit_map),
8592   _work_queue(work_queue) { }
8593 
8594 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8595   HeapWord* addr = (HeapWord*)obj;
8596   if (_span.contains(addr) &&
8597       !_bit_map->isMarked(addr)) {
8598     if (_bit_map->par_mark(addr)) {
8599       bool simulate_overflow = false;
8600       NOT_PRODUCT(
8601         if (CMSMarkStackOverflowALot &&
8602             _collector->par_simulate_overflow()) {
8603           // simulate a stack overflow
8604           simulate_overflow = true;
8605         }
8606       )
8607       if (simulate_overflow || !_work_queue->push(obj)) {
8608         _collector->par_push_on_overflow_list(obj);
8609         _collector->_par_kac_ovflw++;
8610       }
8611     } // Else another thread got there already
8612   }
8613 }
8614 
8615 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8616 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8617 
8618 //////////////////////////////////////////////////////////////////
8619 //  CMSExpansionCause                /////////////////////////////
8620 //////////////////////////////////////////////////////////////////
8621 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8622   switch (cause) {
8623     case _no_expansion:
8624       return "No expansion";
8625     case _satisfy_free_ratio:
8626       return "Free ratio";
8627     case _satisfy_promotion:
8628       return "Satisfy promotion";
8629     case _satisfy_allocation:
8630       return "allocation";
8631     case _allocate_par_lab:
8632       return "Par LAB";
8633     case _allocate_par_spooling_space:
8634       return "Par Spooling Space";
8635     case _adaptive_size_policy:
8636       return "Ergonomics";
8637     default:
8638       return "unknown";
8639   }
8640 }
8641 
8642 void CMSDrainMarkingStackClosure::do_void() {
8643   // the max number to take from overflow list at a time
8644   const size_t num = _mark_stack->capacity()/4;
8645   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8646          "Overflow list should be NULL during concurrent phases");
8647   while (!_mark_stack->isEmpty() ||
8648          // if stack is empty, check the overflow list
8649          _collector->take_from_overflow_list(num, _mark_stack)) {
8650     oop obj = _mark_stack->pop();
8651     HeapWord* addr = (HeapWord*)obj;
8652     assert(_span.contains(addr), "Should be within span");
8653     assert(_bit_map->isMarked(addr), "Should be marked");
8654     assert(obj->is_oop(), "Should be an oop");
8655     obj->oop_iterate(_keep_alive);
8656   }
8657 }
8658 
8659 void CMSParDrainMarkingStackClosure::do_void() {
8660   // drain queue
8661   trim_queue(0);
8662 }
8663 
8664 // Trim our work_queue so its length is below max at return
8665 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8666   while (_work_queue->size() > max) {
8667     oop new_oop;
8668     if (_work_queue->pop_local(new_oop)) {
8669       assert(new_oop->is_oop(), "Expected an oop");
8670       assert(_bit_map->isMarked((HeapWord*)new_oop),
8671              "no white objects on this stack!");
8672       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8673       // iterate over the oops in this oop, marking and pushing
8674       // the ones in CMS heap (i.e. in _span).
8675       new_oop->oop_iterate(&_mark_and_push);
8676     }
8677   }
8678 }
8679 
8680 ////////////////////////////////////////////////////////////////////
8681 // Support for Marking Stack Overflow list handling and related code
8682 ////////////////////////////////////////////////////////////////////
8683 // Much of the following code is similar in shape and spirit to the
8684 // code used in ParNewGC. We should try and share that code
8685 // as much as possible in the future.
8686 
8687 #ifndef PRODUCT
8688 // Debugging support for CMSStackOverflowALot
8689 
8690 // It's OK to call this multi-threaded;  the worst thing
8691 // that can happen is that we'll get a bunch of closely
8692 // spaced simulated oveflows, but that's OK, in fact
8693 // probably good as it would exercise the overflow code
8694 // under contention.
8695 bool CMSCollector::simulate_overflow() {
8696   if (_overflow_counter-- <= 0) { // just being defensive
8697     _overflow_counter = CMSMarkStackOverflowInterval;
8698     return true;
8699   } else {
8700     return false;
8701   }
8702 }
8703 
8704 bool CMSCollector::par_simulate_overflow() {
8705   return simulate_overflow();
8706 }
8707 #endif
8708 
8709 // Single-threaded
8710 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8711   assert(stack->isEmpty(), "Expected precondition");
8712   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8713   size_t i = num;
8714   oop  cur = _overflow_list;
8715   const markOop proto = markOopDesc::prototype();
8716   NOT_PRODUCT(ssize_t n = 0;)
8717   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8718     next = oop(cur->mark());
8719     cur->set_mark(proto);   // until proven otherwise
8720     assert(cur->is_oop(), "Should be an oop");
8721     bool res = stack->push(cur);
8722     assert(res, "Bit off more than can chew?");
8723     NOT_PRODUCT(n++;)
8724   }
8725   _overflow_list = cur;
8726 #ifndef PRODUCT
8727   assert(_num_par_pushes >= n, "Too many pops?");
8728   _num_par_pushes -=n;
8729 #endif
8730   return !stack->isEmpty();
8731 }
8732 
8733 #define BUSY  (oop(0x1aff1aff))
8734 // (MT-safe) Get a prefix of at most "num" from the list.
8735 // The overflow list is chained through the mark word of
8736 // each object in the list. We fetch the entire list,
8737 // break off a prefix of the right size and return the
8738 // remainder. If other threads try to take objects from
8739 // the overflow list at that time, they will wait for
8740 // some time to see if data becomes available. If (and
8741 // only if) another thread places one or more object(s)
8742 // on the global list before we have returned the suffix
8743 // to the global list, we will walk down our local list
8744 // to find its end and append the global list to
8745 // our suffix before returning it. This suffix walk can
8746 // prove to be expensive (quadratic in the amount of traffic)
8747 // when there are many objects in the overflow list and
8748 // there is much producer-consumer contention on the list.
8749 // *NOTE*: The overflow list manipulation code here and
8750 // in ParNewGeneration:: are very similar in shape,
8751 // except that in the ParNew case we use the old (from/eden)
8752 // copy of the object to thread the list via its klass word.
8753 // Because of the common code, if you make any changes in
8754 // the code below, please check the ParNew version to see if
8755 // similar changes might be needed.
8756 // CR 6797058 has been filed to consolidate the common code.
8757 bool CMSCollector::par_take_from_overflow_list(size_t num,
8758                                                OopTaskQueue* work_q,
8759                                                int no_of_gc_threads) {
8760   assert(work_q->size() == 0, "First empty local work queue");
8761   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8762   if (_overflow_list == NULL) {
8763     return false;
8764   }
8765   // Grab the entire list; we'll put back a suffix
8766   oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8767   Thread* tid = Thread::current();
8768   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8769   // set to ParallelGCThreads.
8770   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8771   size_t sleep_time_millis = MAX2((size_t)1, num/100);
8772   // If the list is busy, we spin for a short while,
8773   // sleeping between attempts to get the list.
8774   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8775     os::sleep(tid, sleep_time_millis, false);
8776     if (_overflow_list == NULL) {
8777       // Nothing left to take
8778       return false;
8779     } else if (_overflow_list != BUSY) {
8780       // Try and grab the prefix
8781       prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8782     }
8783   }
8784   // If the list was found to be empty, or we spun long
8785   // enough, we give up and return empty-handed. If we leave
8786   // the list in the BUSY state below, it must be the case that
8787   // some other thread holds the overflow list and will set it
8788   // to a non-BUSY state in the future.
8789   if (prefix == NULL || prefix == BUSY) {
8790      // Nothing to take or waited long enough
8791      if (prefix == NULL) {
8792        // Write back the NULL in case we overwrote it with BUSY above
8793        // and it is still the same value.
8794        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8795      }
8796      return false;
8797   }
8798   assert(prefix != NULL && prefix != BUSY, "Error");
8799   size_t i = num;
8800   oop cur = prefix;
8801   // Walk down the first "num" objects, unless we reach the end.
8802   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8803   if (cur->mark() == NULL) {
8804     // We have "num" or fewer elements in the list, so there
8805     // is nothing to return to the global list.
8806     // Write back the NULL in lieu of the BUSY we wrote
8807     // above, if it is still the same value.
8808     if (_overflow_list == BUSY) {
8809       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8810     }
8811   } else {
8812     // Chop off the suffix and rerturn it to the global list.
8813     assert(cur->mark() != BUSY, "Error");
8814     oop suffix_head = cur->mark(); // suffix will be put back on global list
8815     cur->set_mark(NULL);           // break off suffix
8816     // It's possible that the list is still in the empty(busy) state
8817     // we left it in a short while ago; in that case we may be
8818     // able to place back the suffix without incurring the cost
8819     // of a walk down the list.
8820     oop observed_overflow_list = _overflow_list;
8821     oop cur_overflow_list = observed_overflow_list;
8822     bool attached = false;
8823     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8824       observed_overflow_list =
8825         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8826       if (cur_overflow_list == observed_overflow_list) {
8827         attached = true;
8828         break;
8829       } else cur_overflow_list = observed_overflow_list;
8830     }
8831     if (!attached) {
8832       // Too bad, someone else sneaked in (at least) an element; we'll need
8833       // to do a splice. Find tail of suffix so we can prepend suffix to global
8834       // list.
8835       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8836       oop suffix_tail = cur;
8837       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8838              "Tautology");
8839       observed_overflow_list = _overflow_list;
8840       do {
8841         cur_overflow_list = observed_overflow_list;
8842         if (cur_overflow_list != BUSY) {
8843           // Do the splice ...
8844           suffix_tail->set_mark(markOop(cur_overflow_list));
8845         } else { // cur_overflow_list == BUSY
8846           suffix_tail->set_mark(NULL);
8847         }
8848         // ... and try to place spliced list back on overflow_list ...
8849         observed_overflow_list =
8850           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8851       } while (cur_overflow_list != observed_overflow_list);
8852       // ... until we have succeeded in doing so.
8853     }
8854   }
8855 
8856   // Push the prefix elements on work_q
8857   assert(prefix != NULL, "control point invariant");
8858   const markOop proto = markOopDesc::prototype();
8859   oop next;
8860   NOT_PRODUCT(ssize_t n = 0;)
8861   for (cur = prefix; cur != NULL; cur = next) {
8862     next = oop(cur->mark());
8863     cur->set_mark(proto);   // until proven otherwise
8864     assert(cur->is_oop(), "Should be an oop");
8865     bool res = work_q->push(cur);
8866     assert(res, "Bit off more than we can chew?");
8867     NOT_PRODUCT(n++;)
8868   }
8869 #ifndef PRODUCT
8870   assert(_num_par_pushes >= n, "Too many pops?");
8871   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8872 #endif
8873   return true;
8874 }
8875 
8876 // Single-threaded
8877 void CMSCollector::push_on_overflow_list(oop p) {
8878   NOT_PRODUCT(_num_par_pushes++;)
8879   assert(p->is_oop(), "Not an oop");
8880   preserve_mark_if_necessary(p);
8881   p->set_mark((markOop)_overflow_list);
8882   _overflow_list = p;
8883 }
8884 
8885 // Multi-threaded; use CAS to prepend to overflow list
8886 void CMSCollector::par_push_on_overflow_list(oop p) {
8887   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8888   assert(p->is_oop(), "Not an oop");
8889   par_preserve_mark_if_necessary(p);
8890   oop observed_overflow_list = _overflow_list;
8891   oop cur_overflow_list;
8892   do {
8893     cur_overflow_list = observed_overflow_list;
8894     if (cur_overflow_list != BUSY) {
8895       p->set_mark(markOop(cur_overflow_list));
8896     } else {
8897       p->set_mark(NULL);
8898     }
8899     observed_overflow_list =
8900       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8901   } while (cur_overflow_list != observed_overflow_list);
8902 }
8903 #undef BUSY
8904 
8905 // Single threaded
8906 // General Note on GrowableArray: pushes may silently fail
8907 // because we are (temporarily) out of C-heap for expanding
8908 // the stack. The problem is quite ubiquitous and affects
8909 // a lot of code in the JVM. The prudent thing for GrowableArray
8910 // to do (for now) is to exit with an error. However, that may
8911 // be too draconian in some cases because the caller may be
8912 // able to recover without much harm. For such cases, we
8913 // should probably introduce a "soft_push" method which returns
8914 // an indication of success or failure with the assumption that
8915 // the caller may be able to recover from a failure; code in
8916 // the VM can then be changed, incrementally, to deal with such
8917 // failures where possible, thus, incrementally hardening the VM
8918 // in such low resource situations.
8919 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8920   _preserved_oop_stack.push(p);
8921   _preserved_mark_stack.push(m);
8922   assert(m == p->mark(), "Mark word changed");
8923   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8924          "bijection");
8925 }
8926 
8927 // Single threaded
8928 void CMSCollector::preserve_mark_if_necessary(oop p) {
8929   markOop m = p->mark();
8930   if (m->must_be_preserved(p)) {
8931     preserve_mark_work(p, m);
8932   }
8933 }
8934 
8935 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8936   markOop m = p->mark();
8937   if (m->must_be_preserved(p)) {
8938     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8939     // Even though we read the mark word without holding
8940     // the lock, we are assured that it will not change
8941     // because we "own" this oop, so no other thread can
8942     // be trying to push it on the overflow list; see
8943     // the assertion in preserve_mark_work() that checks
8944     // that m == p->mark().
8945     preserve_mark_work(p, m);
8946   }
8947 }
8948 
8949 // We should be able to do this multi-threaded,
8950 // a chunk of stack being a task (this is
8951 // correct because each oop only ever appears
8952 // once in the overflow list. However, it's
8953 // not very easy to completely overlap this with
8954 // other operations, so will generally not be done
8955 // until all work's been completed. Because we
8956 // expect the preserved oop stack (set) to be small,
8957 // it's probably fine to do this single-threaded.
8958 // We can explore cleverer concurrent/overlapped/parallel
8959 // processing of preserved marks if we feel the
8960 // need for this in the future. Stack overflow should
8961 // be so rare in practice and, when it happens, its
8962 // effect on performance so great that this will
8963 // likely just be in the noise anyway.
8964 void CMSCollector::restore_preserved_marks_if_any() {
8965   assert(SafepointSynchronize::is_at_safepoint(),
8966          "world should be stopped");
8967   assert(Thread::current()->is_ConcurrentGC_thread() ||
8968          Thread::current()->is_VM_thread(),
8969          "should be single-threaded");
8970   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8971          "bijection");
8972 
8973   while (!_preserved_oop_stack.is_empty()) {
8974     oop p = _preserved_oop_stack.pop();
8975     assert(p->is_oop(), "Should be an oop");
8976     assert(_span.contains(p), "oop should be in _span");
8977     assert(p->mark() == markOopDesc::prototype(),
8978            "Set when taken from overflow list");
8979     markOop m = _preserved_mark_stack.pop();
8980     p->set_mark(m);
8981   }
8982   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8983          "stacks were cleared above");
8984 }
8985 
8986 #ifndef PRODUCT
8987 bool CMSCollector::no_preserved_marks() const {
8988   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8989 }
8990 #endif
8991 
8992 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8993 {
8994   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8995   CMSAdaptiveSizePolicy* size_policy =
8996     (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8997   assert(size_policy->is_gc_cms_adaptive_size_policy(),
8998     "Wrong type for size policy");
8999   return size_policy;
9000 }
9001 
9002 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9003                                            size_t desired_promo_size) {
9004   if (cur_promo_size < desired_promo_size) {
9005     size_t expand_bytes = desired_promo_size - cur_promo_size;
9006     if (PrintAdaptiveSizePolicy && Verbose) {
9007       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9008         "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9009         expand_bytes);
9010     }
9011     expand(expand_bytes,
9012            MinHeapDeltaBytes,
9013            CMSExpansionCause::_adaptive_size_policy);
9014   } else if (desired_promo_size < cur_promo_size) {
9015     size_t shrink_bytes = cur_promo_size - desired_promo_size;
9016     if (PrintAdaptiveSizePolicy && Verbose) {
9017       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9018         "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9019         shrink_bytes);
9020     }
9021     shrink(shrink_bytes);
9022   }
9023 }
9024 
9025 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9026   GenCollectedHeap* gch = GenCollectedHeap::heap();
9027   CMSGCAdaptivePolicyCounters* counters =
9028     (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9029   assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9030     "Wrong kind of counters");
9031   return counters;
9032 }
9033 
9034 
9035 void ASConcurrentMarkSweepGeneration::update_counters() {
9036   if (UsePerfData) {
9037     _space_counters->update_all();
9038     _gen_counters->update_all();
9039     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9040     GenCollectedHeap* gch = GenCollectedHeap::heap();
9041     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9042     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9043       "Wrong gc statistics type");
9044     counters->update_counters(gc_stats_l);
9045   }
9046 }
9047 
9048 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9049   if (UsePerfData) {
9050     _space_counters->update_used(used);
9051     _space_counters->update_capacity();
9052     _gen_counters->update_all();
9053 
9054     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9055     GenCollectedHeap* gch = GenCollectedHeap::heap();
9056     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9057     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9058       "Wrong gc statistics type");
9059     counters->update_counters(gc_stats_l);
9060   }
9061 }
9062 
9063 // The desired expansion delta is computed so that:
9064 // . desired free percentage or greater is used
9065 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9066   assert_locked_or_safepoint(Heap_lock);
9067 
9068   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9069 
9070   // If incremental collection failed, we just want to expand
9071   // to the limit.
9072   if (incremental_collection_failed()) {
9073     clear_incremental_collection_failed();
9074     grow_to_reserved();
9075     return;
9076   }
9077 
9078   assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9079 
9080   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9081     "Wrong type of heap");
9082   int prev_level = level() - 1;
9083   assert(prev_level >= 0, "The cms generation is the lowest generation");
9084   Generation* prev_gen = gch->get_gen(prev_level);
9085   assert(prev_gen->kind() == Generation::ASParNew,
9086     "Wrong type of young generation");
9087   ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9088   size_t cur_eden = younger_gen->eden()->capacity();
9089   CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9090   size_t cur_promo = free();
9091   size_policy->compute_tenured_generation_free_space(cur_promo,
9092                                                        max_available(),
9093                                                        cur_eden);
9094   resize(cur_promo, size_policy->promo_size());
9095 
9096   // Record the new size of the space in the cms generation
9097   // that is available for promotions.  This is temporary.
9098   // It should be the desired promo size.
9099   size_policy->avg_cms_promo()->sample(free());
9100   size_policy->avg_old_live()->sample(used());
9101 
9102   if (UsePerfData) {
9103     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9104     counters->update_cms_capacity_counter(capacity());
9105   }
9106 }
9107 
9108 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9109   assert_locked_or_safepoint(Heap_lock);
9110   assert_lock_strong(freelistLock());
9111   HeapWord* old_end = _cmsSpace->end();
9112   HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9113   assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9114   FreeChunk* chunk_at_end = find_chunk_at_end();
9115   if (chunk_at_end == NULL) {
9116     // No room to shrink
9117     if (PrintGCDetails && Verbose) {
9118       gclog_or_tty->print_cr("No room to shrink: old_end  "
9119         PTR_FORMAT "  unallocated_start  " PTR_FORMAT
9120         " chunk_at_end  " PTR_FORMAT,
9121         old_end, unallocated_start, chunk_at_end);
9122     }
9123     return;
9124   } else {
9125 
9126     // Find the chunk at the end of the space and determine
9127     // how much it can be shrunk.
9128     size_t shrinkable_size_in_bytes = chunk_at_end->size();
9129     size_t aligned_shrinkable_size_in_bytes =
9130       align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9131     assert(unallocated_start <= chunk_at_end->end(),
9132       "Inconsistent chunk at end of space");
9133     size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9134     size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9135 
9136     // Shrink the underlying space
9137     _virtual_space.shrink_by(bytes);
9138     if (PrintGCDetails && Verbose) {
9139       gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9140         " desired_bytes " SIZE_FORMAT
9141         " shrinkable_size_in_bytes " SIZE_FORMAT
9142         " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9143         "  bytes  " SIZE_FORMAT,
9144         desired_bytes, shrinkable_size_in_bytes,
9145         aligned_shrinkable_size_in_bytes, bytes);
9146       gclog_or_tty->print_cr("          old_end  " SIZE_FORMAT
9147         "  unallocated_start  " SIZE_FORMAT,
9148         old_end, unallocated_start);
9149     }
9150 
9151     // If the space did shrink (shrinking is not guaranteed),
9152     // shrink the chunk at the end by the appropriate amount.
9153     if (((HeapWord*)_virtual_space.high()) < old_end) {
9154       size_t new_word_size =
9155         heap_word_size(_virtual_space.committed_size());
9156 
9157       // Have to remove the chunk from the dictionary because it is changing
9158       // size and might be someplace elsewhere in the dictionary.
9159 
9160       // Get the chunk at end, shrink it, and put it
9161       // back.
9162       _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9163       size_t word_size_change = word_size_before - new_word_size;
9164       size_t chunk_at_end_old_size = chunk_at_end->size();
9165       assert(chunk_at_end_old_size >= word_size_change,
9166         "Shrink is too large");
9167       chunk_at_end->setSize(chunk_at_end_old_size -
9168                           word_size_change);
9169       _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9170         word_size_change);
9171 
9172       _cmsSpace->returnChunkToDictionary(chunk_at_end);
9173 
9174       MemRegion mr(_cmsSpace->bottom(), new_word_size);
9175       _bts->resize(new_word_size);  // resize the block offset shared array
9176       Universe::heap()->barrier_set()->resize_covered_region(mr);
9177       _cmsSpace->assert_locked();
9178       _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9179 
9180       NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9181 
9182       // update the space and generation capacity counters
9183       if (UsePerfData) {
9184         _space_counters->update_capacity();
9185         _gen_counters->update_all();
9186       }
9187 
9188       if (Verbose && PrintGCDetails) {
9189         size_t new_mem_size = _virtual_space.committed_size();
9190         size_t old_mem_size = new_mem_size + bytes;
9191         gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9192                       name(), old_mem_size/K, bytes/K, new_mem_size/K);
9193       }
9194     }
9195 
9196     assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9197       "Inconsistency at end of space");
9198     assert(chunk_at_end->end() == _cmsSpace->end(),
9199       "Shrinking is inconsistent");
9200     return;
9201   }
9202 }
9203 
9204 // Transfer some number of overflown objects to usual marking
9205 // stack. Return true if some objects were transferred.
9206 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9207   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9208                     (size_t)ParGCDesiredObjsFromOverflowList);
9209 
9210   bool res = _collector->take_from_overflow_list(num, _mark_stack);
9211   assert(_collector->overflow_list_is_empty() || res,
9212          "If list is not empty, we should have taken something");
9213   assert(!res || !_mark_stack->isEmpty(),
9214          "If we took something, it should now be on our stack");
9215   return res;
9216 }
9217 
9218 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9219   size_t res = _sp->block_size_no_stall(addr, _collector);
9220   if (_sp->block_is_obj(addr)) {
9221     if (_live_bit_map->isMarked(addr)) {
9222       // It can't have been dead in a previous cycle
9223       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9224     } else {
9225       _dead_bit_map->mark(addr);      // mark the dead object
9226     }
9227   }
9228   // Could be 0, if the block size could not be computed without stalling.
9229   return res;
9230 }
9231 
9232 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase): TraceMemoryManagerStats() {
9233 
9234   switch (phase) {
9235     case CMSCollector::InitialMarking:
9236       initialize(true  /* fullGC */ ,
9237                  true  /* recordGCBeginTime */,
9238                  true  /* recordPreGCUsage */,
9239                  false /* recordPeakUsage */,
9240                  false /* recordPostGCusage */,
9241                  true  /* recordAccumulatedGCTime */,
9242                  false /* recordGCEndTime */,
9243                  false /* countCollection */  );
9244       break;
9245 
9246     case CMSCollector::FinalMarking:
9247       initialize(true  /* fullGC */ ,
9248                  false /* recordGCBeginTime */,
9249                  false /* recordPreGCUsage */,
9250                  false /* recordPeakUsage */,
9251                  false /* recordPostGCusage */,
9252                  true  /* recordAccumulatedGCTime */,
9253                  false /* recordGCEndTime */,
9254                  false /* countCollection */  );
9255       break;
9256 
9257     case CMSCollector::Sweeping:
9258       initialize(true  /* fullGC */ ,
9259                  false /* recordGCBeginTime */,
9260                  false /* recordPreGCUsage */,
9261                  true  /* recordPeakUsage */,
9262                  true  /* recordPostGCusage */,
9263                  false /* recordAccumulatedGCTime */,
9264                  true  /* recordGCEndTime */,
9265                  true  /* countCollection */  );
9266       break;
9267 
9268     default:
9269       ShouldNotReachHere();
9270   }
9271 }
9272 
9273 // when bailing out of cms in concurrent mode failure
9274 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(): TraceMemoryManagerStats() {
9275   initialize(true /* fullGC */ ,
9276              true /* recordGCBeginTime */,
9277              true /* recordPreGCUsage */,
9278              true /* recordPeakUsage */,
9279              true /* recordPostGCusage */,
9280              true /* recordAccumulatedGCTime */,
9281              true /* recordGCEndTime */,
9282              true /* countCollection */ );
9283 }