1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderData.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/systemDictionary.hpp"
  29 #include "code/codeCache.hpp"
  30 #include "gc/cms/cmsCollectorPolicy.hpp"
  31 #include "gc/cms/cmsOopClosures.inline.hpp"
  32 #include "gc/cms/compactibleFreeListSpace.hpp"
  33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  34 #include "gc/cms/concurrentMarkSweepThread.hpp"
  35 #include "gc/cms/parNewGeneration.hpp"
  36 #include "gc/cms/vmCMSOperations.hpp"
  37 #include "gc/serial/genMarkSweep.hpp"
  38 #include "gc/serial/tenuredGeneration.hpp"
  39 #include "gc/shared/adaptiveSizePolicy.hpp"
  40 #include "gc/shared/cardGeneration.inline.hpp"
  41 #include "gc/shared/cardTableRS.hpp"
  42 #include "gc/shared/collectedHeap.inline.hpp"
  43 #include "gc/shared/collectorCounters.hpp"
  44 #include "gc/shared/collectorPolicy.hpp"
  45 #include "gc/shared/gcLocker.inline.hpp"
  46 #include "gc/shared/gcPolicyCounters.hpp"
  47 #include "gc/shared/gcTimer.hpp"
  48 #include "gc/shared/gcTrace.hpp"
  49 #include "gc/shared/gcTraceTime.hpp"
  50 #include "gc/shared/genCollectedHeap.hpp"
  51 #include "gc/shared/genOopClosures.inline.hpp"
  52 #include "gc/shared/isGCActiveMark.hpp"
  53 #include "gc/shared/referencePolicy.hpp"
  54 #include "gc/shared/strongRootsScope.hpp"
  55 #include "gc/shared/taskqueue.inline.hpp"
  56 #include "memory/allocation.hpp"
  57 #include "memory/iterator.inline.hpp"
  58 #include "memory/padded.hpp"
  59 #include "memory/resourceArea.hpp"
  60 #include "oops/oop.inline.hpp"
  61 #include "prims/jvmtiExport.hpp"
  62 #include "runtime/atomic.inline.hpp"
  63 #include "runtime/globals_extension.hpp"
  64 #include "runtime/handles.inline.hpp"
  65 #include "runtime/java.hpp"
  66 #include "runtime/orderAccess.inline.hpp"
  67 #include "runtime/vmThread.hpp"
  68 #include "services/memoryService.hpp"
  69 #include "services/runtimeService.hpp"
  70 #include "utilities/stack.inline.hpp"
  71 
  72 // statics
  73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  74 bool CMSCollector::_full_gc_requested = false;
  75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  76 
  77 //////////////////////////////////////////////////////////////////
  78 // In support of CMS/VM thread synchronization
  79 //////////////////////////////////////////////////////////////////
  80 // We split use of the CGC_lock into 2 "levels".
  81 // The low-level locking is of the usual CGC_lock monitor. We introduce
  82 // a higher level "token" (hereafter "CMS token") built on top of the
  83 // low level monitor (hereafter "CGC lock").
  84 // The token-passing protocol gives priority to the VM thread. The
  85 // CMS-lock doesn't provide any fairness guarantees, but clients
  86 // should ensure that it is only held for very short, bounded
  87 // durations.
  88 //
  89 // When either of the CMS thread or the VM thread is involved in
  90 // collection operations during which it does not want the other
  91 // thread to interfere, it obtains the CMS token.
  92 //
  93 // If either thread tries to get the token while the other has
  94 // it, that thread waits. However, if the VM thread and CMS thread
  95 // both want the token, then the VM thread gets priority while the
  96 // CMS thread waits. This ensures, for instance, that the "concurrent"
  97 // phases of the CMS thread's work do not block out the VM thread
  98 // for long periods of time as the CMS thread continues to hog
  99 // the token. (See bug 4616232).
 100 //
 101 // The baton-passing functions are, however, controlled by the
 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 103 // and here the low-level CMS lock, not the high level token,
 104 // ensures mutual exclusion.
 105 //
 106 // Two important conditions that we have to satisfy:
 107 // 1. if a thread does a low-level wait on the CMS lock, then it
 108 //    relinquishes the CMS token if it were holding that token
 109 //    when it acquired the low-level CMS lock.
 110 // 2. any low-level notifications on the low-level lock
 111 //    should only be sent when a thread has relinquished the token.
 112 //
 113 // In the absence of either property, we'd have potential deadlock.
 114 //
 115 // We protect each of the CMS (concurrent and sequential) phases
 116 // with the CMS _token_, not the CMS _lock_.
 117 //
 118 // The only code protected by CMS lock is the token acquisition code
 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 120 // baton-passing code.
 121 //
 122 // Unfortunately, i couldn't come up with a good abstraction to factor and
 123 // hide the naked CGC_lock manipulation in the baton-passing code
 124 // further below. That's something we should try to do. Also, the proof
 125 // of correctness of this 2-level locking scheme is far from obvious,
 126 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 127 // that there may be a theoretical possibility of delay/starvation in the
 128 // low-level lock/wait/notify scheme used for the baton-passing because of
 129 // potential interference with the priority scheme embodied in the
 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 131 // invocation further below and marked with "XXX 20011219YSR".
 132 // Indeed, as we note elsewhere, this may become yet more slippery
 133 // in the presence of multiple CMS and/or multiple VM threads. XXX
 134 
 135 class CMSTokenSync: public StackObj {
 136  private:
 137   bool _is_cms_thread;
 138  public:
 139   CMSTokenSync(bool is_cms_thread):
 140     _is_cms_thread(is_cms_thread) {
 141     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 142            "Incorrect argument to constructor");
 143     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 144   }
 145 
 146   ~CMSTokenSync() {
 147     assert(_is_cms_thread ?
 148              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 149              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 150           "Incorrect state");
 151     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 152   }
 153 };
 154 
 155 // Convenience class that does a CMSTokenSync, and then acquires
 156 // upto three locks.
 157 class CMSTokenSyncWithLocks: public CMSTokenSync {
 158  private:
 159   // Note: locks are acquired in textual declaration order
 160   // and released in the opposite order
 161   MutexLockerEx _locker1, _locker2, _locker3;
 162  public:
 163   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 164                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 165     CMSTokenSync(is_cms_thread),
 166     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 167     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 168     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 169   { }
 170 };
 171 
 172 
 173 //////////////////////////////////////////////////////////////////
 174 //  Concurrent Mark-Sweep Generation /////////////////////////////
 175 //////////////////////////////////////////////////////////////////
 176 
 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 178 
 179 // This struct contains per-thread things necessary to support parallel
 180 // young-gen collection.
 181 class CMSParGCThreadState: public CHeapObj<mtGC> {
 182  public:
 183   CFLS_LAB lab;
 184   PromotionInfo promo;
 185 
 186   // Constructor.
 187   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 188     promo.setSpace(cfls);
 189   }
 190 };
 191 
 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 193      ReservedSpace rs, size_t initial_byte_size,
 194      CardTableRS* ct, bool use_adaptive_freelists,
 195      FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
 196   CardGeneration(rs, initial_byte_size, ct),
 197   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 198   _did_compact(false)
 199 {
 200   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 201   HeapWord* end    = (HeapWord*) _virtual_space.high();
 202 
 203   _direct_allocated_words = 0;
 204   NOT_PRODUCT(
 205     _numObjectsPromoted = 0;
 206     _numWordsPromoted = 0;
 207     _numObjectsAllocated = 0;
 208     _numWordsAllocated = 0;
 209   )
 210 
 211   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
 212                                            use_adaptive_freelists,
 213                                            dictionaryChoice);
 214   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 215   _cmsSpace->_old_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(UseCompressedClassPointers ||
 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 
 230   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 231   for (uint i = 0; i < ParallelGCThreads; i++) {
 232     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 233   }
 234 
 235   _incremental_collection_failed = false;
 236   // The "dilatation_factor" is the expansion that can occur on
 237   // account of the fact that the minimum object size in the CMS
 238   // generation may be larger than that in, say, a contiguous young
 239   //  generation.
 240   // Ideally, in the calculation below, we'd compute the dilatation
 241   // factor as: MinChunkSize/(promoting_gen's min object size)
 242   // Since we do not have such a general query interface for the
 243   // promoting generation, we'll instead just use the minimum
 244   // object size (which today is a header's worth of space);
 245   // note that all arithmetic is in units of HeapWords.
 246   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 247   assert(_dilatation_factor >= 1.0, "from previous assert");
 248 }
 249 
 250 
 251 // The field "_initiating_occupancy" represents the occupancy percentage
 252 // at which we trigger a new collection cycle.  Unless explicitly specified
 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 254 // is calculated by:
 255 //
 256 //   Let "f" be MinHeapFreeRatio in
 257 //
 258 //    _initiating_occupancy = 100-f +
 259 //                           f * (CMSTriggerRatio/100)
 260 //   where CMSTriggerRatio is the argument "tr" below.
 261 //
 262 // That is, if we assume the heap is at its desired maximum occupancy at the
 263 // end of a collection, we let CMSTriggerRatio of the (purported) free
 264 // space be allocated before initiating a new collection cycle.
 265 //
 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 267   assert(io <= 100 && tr <= 100, "Check the arguments");
 268   if (io >= 0) {
 269     _initiating_occupancy = (double)io / 100.0;
 270   } else {
 271     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 272                              (double)(tr * MinHeapFreeRatio) / 100.0)
 273                             / 100.0;
 274   }
 275 }
 276 
 277 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 278   assert(collector() != NULL, "no collector");
 279   collector()->ref_processor_init();
 280 }
 281 
 282 void CMSCollector::ref_processor_init() {
 283   if (_ref_processor == NULL) {
 284     // Allocate and initialize a reference processor
 285     _ref_processor =
 286       new ReferenceProcessor(_span,                               // span
 287                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 288                              ParallelGCThreads,                   // mt processing degree
 289                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 290                              MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 291                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 292                              &_is_alive_closure);                 // closure for liveness info
 293     // Initialize the _ref_processor field of CMSGen
 294     _cmsGen->set_ref_processor(_ref_processor);
 295 
 296   }
 297 }
 298 
 299 AdaptiveSizePolicy* CMSCollector::size_policy() {
 300   GenCollectedHeap* gch = GenCollectedHeap::heap();
 301   return gch->gen_policy()->size_policy();
 302 }
 303 
 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 305 
 306   const char* gen_name = "old";
 307   GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
 308   // Generation Counters - generation 1, 1 subspace
 309   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 310       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 311 
 312   _space_counters = new GSpaceCounters(gen_name, 0,
 313                                        _virtual_space.reserved_size(),
 314                                        this, _gen_counters);
 315 }
 316 
 317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 318   _cms_gen(cms_gen)
 319 {
 320   assert(alpha <= 100, "bad value");
 321   _saved_alpha = alpha;
 322 
 323   // Initialize the alphas to the bootstrap value of 100.
 324   _gc0_alpha = _cms_alpha = 100;
 325 
 326   _cms_begin_time.update();
 327   _cms_end_time.update();
 328 
 329   _gc0_duration = 0.0;
 330   _gc0_period = 0.0;
 331   _gc0_promoted = 0;
 332 
 333   _cms_duration = 0.0;
 334   _cms_period = 0.0;
 335   _cms_allocated = 0;
 336 
 337   _cms_used_at_gc0_begin = 0;
 338   _cms_used_at_gc0_end = 0;
 339   _allow_duty_cycle_reduction = false;
 340   _valid_bits = 0;
 341 }
 342 
 343 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 344   // TBD: CR 6909490
 345   return 1.0;
 346 }
 347 
 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 349 }
 350 
 351 // If promotion failure handling is on use
 352 // the padded average size of the promotion for each
 353 // young generation collection.
 354 double CMSStats::time_until_cms_gen_full() const {
 355   size_t cms_free = _cms_gen->cmsSpace()->free();
 356   GenCollectedHeap* gch = GenCollectedHeap::heap();
 357   size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
 358                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 359   if (cms_free > expected_promotion) {
 360     // Start a cms collection if there isn't enough space to promote
 361     // for the next young collection.  Use the padded average as
 362     // a safety factor.
 363     cms_free -= expected_promotion;
 364 
 365     // Adjust by the safety factor.
 366     double cms_free_dbl = (double)cms_free;
 367     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 368     // Apply a further correction factor which tries to adjust
 369     // for recent occurance of concurrent mode failures.
 370     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 371     cms_free_dbl = cms_free_dbl * cms_adjustment;
 372 
 373     if (PrintGCDetails && Verbose) {
 374       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
 375         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 376         cms_free, expected_promotion);
 377       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
 378         cms_free_dbl, cms_consumption_rate() + 1.0);
 379     }
 380     // Add 1 in case the consumption rate goes to zero.
 381     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 382   }
 383   return 0.0;
 384 }
 385 
 386 // Compare the duration of the cms collection to the
 387 // time remaining before the cms generation is empty.
 388 // Note that the time from the start of the cms collection
 389 // to the start of the cms sweep (less than the total
 390 // duration of the cms collection) can be used.  This
 391 // has been tried and some applications experienced
 392 // promotion failures early in execution.  This was
 393 // possibly because the averages were not accurate
 394 // enough at the beginning.
 395 double CMSStats::time_until_cms_start() const {
 396   // We add "gc0_period" to the "work" calculation
 397   // below because this query is done (mostly) at the
 398   // end of a scavenge, so we need to conservatively
 399   // account for that much possible delay
 400   // in the query so as to avoid concurrent mode failures
 401   // due to starting the collection just a wee bit too
 402   // late.
 403   double work = cms_duration() + gc0_period();
 404   double deadline = time_until_cms_gen_full();
 405   // If a concurrent mode failure occurred recently, we want to be
 406   // more conservative and halve our expected time_until_cms_gen_full()
 407   if (work > deadline) {
 408     if (Verbose && PrintGCDetails) {
 409       gclog_or_tty->print(
 410         " CMSCollector: collect because of anticipated promotion "
 411         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
 412         gc0_period(), time_until_cms_gen_full());
 413     }
 414     return 0.0;
 415   }
 416   return work - deadline;
 417 }
 418 
 419 #ifndef PRODUCT
 420 void CMSStats::print_on(outputStream *st) const {
 421   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 422   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 423                gc0_duration(), gc0_period(), gc0_promoted());
 424   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 425             cms_duration(), cms_period(), cms_allocated());
 426   st->print(",cms_since_beg=%g,cms_since_end=%g",
 427             cms_time_since_begin(), cms_time_since_end());
 428   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 429             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 430 
 431   if (valid()) {
 432     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 433               promotion_rate(), cms_allocation_rate());
 434     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 435               cms_consumption_rate(), time_until_cms_gen_full());
 436   }
 437   st->print(" ");
 438 }
 439 #endif // #ifndef PRODUCT
 440 
 441 CMSCollector::CollectorState CMSCollector::_collectorState =
 442                              CMSCollector::Idling;
 443 bool CMSCollector::_foregroundGCIsActive = false;
 444 bool CMSCollector::_foregroundGCShouldWait = false;
 445 
 446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 447                            CardTableRS*                   ct,
 448                            ConcurrentMarkSweepPolicy*     cp):
 449   _cmsGen(cmsGen),
 450   _ct(ct),
 451   _ref_processor(NULL),    // will be set later
 452   _conc_workers(NULL),     // may be set later
 453   _abort_preclean(false),
 454   _start_sampling(false),
 455   _between_prologue_and_epilogue(false),
 456   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 457   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 458                  -1 /* lock-free */, "No_lock" /* dummy */),
 459   _modUnionClosurePar(&_modUnionTable),
 460   // Adjust my span to cover old (cms) gen
 461   _span(cmsGen->reserved()),
 462   // Construct the is_alive_closure with _span & markBitMap
 463   _is_alive_closure(_span, &_markBitMap),
 464   _restart_addr(NULL),
 465   _overflow_list(NULL),
 466   _stats(cmsGen),
 467   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 468                              //verify that this lock should be acquired with safepoint check.
 469                              Monitor::_safepoint_check_sometimes)),
 470   _eden_chunk_array(NULL),     // may be set in ctor body
 471   _eden_chunk_capacity(0),     // -- ditto --
 472   _eden_chunk_index(0),        // -- ditto --
 473   _survivor_plab_array(NULL),  // -- ditto --
 474   _survivor_chunk_array(NULL), // -- ditto --
 475   _survivor_chunk_capacity(0), // -- ditto --
 476   _survivor_chunk_index(0),    // -- ditto --
 477   _ser_pmc_preclean_ovflw(0),
 478   _ser_kac_preclean_ovflw(0),
 479   _ser_pmc_remark_ovflw(0),
 480   _par_pmc_remark_ovflw(0),
 481   _ser_kac_ovflw(0),
 482   _par_kac_ovflw(0),
 483 #ifndef PRODUCT
 484   _num_par_pushes(0),
 485 #endif
 486   _collection_count_start(0),
 487   _verifying(false),
 488   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 489   _completed_initialization(false),
 490   _collector_policy(cp),
 491   _should_unload_classes(CMSClassUnloadingEnabled),
 492   _concurrent_cycles_since_last_unload(0),
 493   _roots_scanning_options(GenCollectedHeap::SO_None),
 494   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 495   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 496   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 497   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 498   _cms_start_registered(false)
 499 {
 500   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
 501     ExplicitGCInvokesConcurrent = true;
 502   }
 503   // Now expand the span and allocate the collection support structures
 504   // (MUT, marking bit map etc.) to cover both generations subject to
 505   // collection.
 506 
 507   // For use by dirty card to oop closures.
 508   _cmsGen->cmsSpace()->set_collector(this);
 509 
 510   // Allocate MUT and marking bit map
 511   {
 512     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 513     if (!_markBitMap.allocate(_span)) {
 514       warning("Failed to allocate CMS Bit Map");
 515       return;
 516     }
 517     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 518   }
 519   {
 520     _modUnionTable.allocate(_span);
 521     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 522   }
 523 
 524   if (!_markStack.allocate(MarkStackSize)) {
 525     warning("Failed to allocate CMS Marking Stack");
 526     return;
 527   }
 528 
 529   // Support for multi-threaded concurrent phases
 530   if (CMSConcurrentMTEnabled) {
 531     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 532       // just for now
 533       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
 534     }
 535     if (ConcGCThreads > 1) {
 536       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 537                                  ConcGCThreads, true);
 538       if (_conc_workers == NULL) {
 539         warning("GC/CMS: _conc_workers allocation failure: "
 540               "forcing -CMSConcurrentMTEnabled");
 541         CMSConcurrentMTEnabled = false;
 542       } else {
 543         _conc_workers->initialize_workers();
 544       }
 545     } else {
 546       CMSConcurrentMTEnabled = false;
 547     }
 548   }
 549   if (!CMSConcurrentMTEnabled) {
 550     ConcGCThreads = 0;
 551   } else {
 552     // Turn off CMSCleanOnEnter optimization temporarily for
 553     // the MT case where it's not fixed yet; see 6178663.
 554     CMSCleanOnEnter = false;
 555   }
 556   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 557          "Inconsistency");
 558 
 559   // Parallel task queues; these are shared for the
 560   // concurrent and stop-world phases of CMS, but
 561   // are not shared with parallel scavenge (ParNew).
 562   {
 563     uint i;
 564     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 565 
 566     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 567          || ParallelRefProcEnabled)
 568         && num_queues > 0) {
 569       _task_queues = new OopTaskQueueSet(num_queues);
 570       if (_task_queues == NULL) {
 571         warning("task_queues allocation failure.");
 572         return;
 573       }
 574       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 575       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 576       for (i = 0; i < num_queues; i++) {
 577         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 578         if (q == NULL) {
 579           warning("work_queue allocation failure.");
 580           return;
 581         }
 582         _task_queues->register_queue(i, q);
 583       }
 584       for (i = 0; i < num_queues; i++) {
 585         _task_queues->queue(i)->initialize();
 586         _hash_seed[i] = 17;  // copied from ParNew
 587       }
 588     }
 589   }
 590 
 591   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 592 
 593   // Clip CMSBootstrapOccupancy between 0 and 100.
 594   _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
 595 
 596   // Now tell CMS generations the identity of their collector
 597   ConcurrentMarkSweepGeneration::set_collector(this);
 598 
 599   // Create & start a CMS thread for this CMS collector
 600   _cmsThread = ConcurrentMarkSweepThread::start(this);
 601   assert(cmsThread() != NULL, "CMS Thread should have been created");
 602   assert(cmsThread()->collector() == this,
 603          "CMS Thread should refer to this gen");
 604   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 605 
 606   // Support for parallelizing young gen rescan
 607   GenCollectedHeap* gch = GenCollectedHeap::heap();
 608   assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 609   _young_gen = (ParNewGeneration*)gch->young_gen();
 610   if (gch->supports_inline_contig_alloc()) {
 611     _top_addr = gch->top_addr();
 612     _end_addr = gch->end_addr();
 613     assert(_young_gen != NULL, "no _young_gen");
 614     _eden_chunk_index = 0;
 615     _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
 616     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 617   }
 618 
 619   // Support for parallelizing survivor space rescan
 620   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 621     const size_t max_plab_samples =
 622       _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
 623 
 624     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 625     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 626     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 627     _survivor_chunk_capacity = max_plab_samples;
 628     for (uint i = 0; i < ParallelGCThreads; i++) {
 629       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 630       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 631       assert(cur->end() == 0, "Should be 0");
 632       assert(cur->array() == vec, "Should be vec");
 633       assert(cur->capacity() == max_plab_samples, "Error");
 634     }
 635   }
 636 
 637   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 638   _gc_counters = new CollectorCounters("CMS", 1);
 639   _completed_initialization = true;
 640   _inter_sweep_timer.start();  // start of time
 641 }
 642 
 643 const char* ConcurrentMarkSweepGeneration::name() const {
 644   return "concurrent mark-sweep generation";
 645 }
 646 void ConcurrentMarkSweepGeneration::update_counters() {
 647   if (UsePerfData) {
 648     _space_counters->update_all();
 649     _gen_counters->update_all();
 650   }
 651 }
 652 
 653 // this is an optimized version of update_counters(). it takes the
 654 // used value as a parameter rather than computing it.
 655 //
 656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 657   if (UsePerfData) {
 658     _space_counters->update_used(used);
 659     _space_counters->update_capacity();
 660     _gen_counters->update_all();
 661   }
 662 }
 663 
 664 void ConcurrentMarkSweepGeneration::print() const {
 665   Generation::print();
 666   cmsSpace()->print();
 667 }
 668 
 669 #ifndef PRODUCT
 670 void ConcurrentMarkSweepGeneration::print_statistics() {
 671   cmsSpace()->printFLCensus(0);
 672 }
 673 #endif
 674 
 675 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
 676   GenCollectedHeap* gch = GenCollectedHeap::heap();
 677   if (PrintGCDetails) {
 678     // I didn't want to change the logging when removing the level concept,
 679     // but I guess this logging could say "old" or something instead of "1".
 680     assert(gch->is_old_gen(this),
 681            "The CMS generation should be the old generation");
 682     uint level = 1;
 683     if (Verbose) {
 684       gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
 685         level, short_name(), s, used(), capacity());
 686     } else {
 687       gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
 688         level, short_name(), s, used() / K, capacity() / K);
 689     }
 690   }
 691   if (Verbose) {
 692     gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
 693               gch->used(), gch->capacity());
 694   } else {
 695     gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
 696               gch->used() / K, gch->capacity() / K);
 697   }
 698 }
 699 
 700 size_t
 701 ConcurrentMarkSweepGeneration::contiguous_available() const {
 702   // dld proposes an improvement in precision here. If the committed
 703   // part of the space ends in a free block we should add that to
 704   // uncommitted size in the calculation below. Will make this
 705   // change later, staying with the approximation below for the
 706   // time being. -- ysr.
 707   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 708 }
 709 
 710 size_t
 711 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 712   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 713 }
 714 
 715 size_t ConcurrentMarkSweepGeneration::max_available() const {
 716   return free() + _virtual_space.uncommitted_size();
 717 }
 718 
 719 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 720   size_t available = max_available();
 721   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 722   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 723   if (Verbose && PrintGCDetails) {
 724     gclog_or_tty->print_cr(
 725       "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
 726       "max_promo(" SIZE_FORMAT ")",
 727       res? "":" not", available, res? ">=":"<",
 728       av_promo, max_promotion_in_bytes);
 729   }
 730   return res;
 731 }
 732 
 733 // At a promotion failure dump information on block layout in heap
 734 // (cms old generation).
 735 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 736   if (CMSDumpAtPromotionFailure) {
 737     cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
 738   }
 739 }
 740 
 741 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 742   // Clear the promotion information.  These pointers can be adjusted
 743   // along with all the other pointers into the heap but
 744   // compaction is expected to be a rare event with
 745   // a heap using cms so don't do it without seeing the need.
 746   for (uint i = 0; i < ParallelGCThreads; i++) {
 747     _par_gc_thread_states[i]->promo.reset();
 748   }
 749 }
 750 
 751 void ConcurrentMarkSweepGeneration::compute_new_size() {
 752   assert_locked_or_safepoint(Heap_lock);
 753 
 754   // If incremental collection failed, we just want to expand
 755   // to the limit.
 756   if (incremental_collection_failed()) {
 757     clear_incremental_collection_failed();
 758     grow_to_reserved();
 759     return;
 760   }
 761 
 762   // The heap has been compacted but not reset yet.
 763   // Any metric such as free() or used() will be incorrect.
 764 
 765   CardGeneration::compute_new_size();
 766 
 767   // Reset again after a possible resizing
 768   if (did_compact()) {
 769     cmsSpace()->reset_after_compaction();
 770   }
 771 }
 772 
 773 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
 774   assert_locked_or_safepoint(Heap_lock);
 775 
 776   // If incremental collection failed, we just want to expand
 777   // to the limit.
 778   if (incremental_collection_failed()) {
 779     clear_incremental_collection_failed();
 780     grow_to_reserved();
 781     return;
 782   }
 783 
 784   double free_percentage = ((double) free()) / capacity();
 785   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 786   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 787 
 788   // compute expansion delta needed for reaching desired free percentage
 789   if (free_percentage < desired_free_percentage) {
 790     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 791     assert(desired_capacity >= capacity(), "invalid expansion size");
 792     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 793     if (PrintGCDetails && Verbose) {
 794       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 795       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
 796       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
 797       gclog_or_tty->print_cr("  Desired free fraction %f", desired_free_percentage);
 798       gclog_or_tty->print_cr("  Maximum free fraction %f", maximum_free_percentage);
 799       gclog_or_tty->print_cr("  Capacity " SIZE_FORMAT, capacity() / 1000);
 800       gclog_or_tty->print_cr("  Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
 801       GenCollectedHeap* gch = GenCollectedHeap::heap();
 802       assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
 803       size_t young_size = gch->young_gen()->capacity();
 804       gclog_or_tty->print_cr("  Young gen size " SIZE_FORMAT, young_size / 1000);
 805       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
 806       gclog_or_tty->print_cr("  contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
 807       gclog_or_tty->print_cr("  Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
 808     }
 809     // safe if expansion fails
 810     expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 811     if (PrintGCDetails && Verbose) {
 812       gclog_or_tty->print_cr("  Expanded free fraction %f", ((double) free()) / capacity());
 813     }
 814   } else {
 815     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 816     assert(desired_capacity <= capacity(), "invalid expansion size");
 817     size_t shrink_bytes = capacity() - desired_capacity;
 818     // Don't shrink unless the delta is greater than the minimum shrink we want
 819     if (shrink_bytes >= MinHeapDeltaBytes) {
 820       shrink_free_list_by(shrink_bytes);
 821     }
 822   }
 823 }
 824 
 825 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 826   return cmsSpace()->freelistLock();
 827 }
 828 
 829 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
 830   CMSSynchronousYieldRequest yr;
 831   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
 832   return have_lock_and_allocate(size, tlab);
 833 }
 834 
 835 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 836                                                                 bool   tlab /* ignored */) {
 837   assert_lock_strong(freelistLock());
 838   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 839   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 840   // Allocate the object live (grey) if the background collector has
 841   // started marking. This is necessary because the marker may
 842   // have passed this address and consequently this object will
 843   // not otherwise be greyed and would be incorrectly swept up.
 844   // Note that if this object contains references, the writing
 845   // of those references will dirty the card containing this object
 846   // allowing the object to be blackened (and its references scanned)
 847   // either during a preclean phase or at the final checkpoint.
 848   if (res != NULL) {
 849     // We may block here with an uninitialized object with
 850     // its mark-bit or P-bits not yet set. Such objects need
 851     // to be safely navigable by block_start().
 852     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
 853     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
 854     collector()->direct_allocated(res, adjustedSize);
 855     _direct_allocated_words += adjustedSize;
 856     // allocation counters
 857     NOT_PRODUCT(
 858       _numObjectsAllocated++;
 859       _numWordsAllocated += (int)adjustedSize;
 860     )
 861   }
 862   return res;
 863 }
 864 
 865 // In the case of direct allocation by mutators in a generation that
 866 // is being concurrently collected, the object must be allocated
 867 // live (grey) if the background collector has started marking.
 868 // This is necessary because the marker may
 869 // have passed this address and consequently this object will
 870 // not otherwise be greyed and would be incorrectly swept up.
 871 // Note that if this object contains references, the writing
 872 // of those references will dirty the card containing this object
 873 // allowing the object to be blackened (and its references scanned)
 874 // either during a preclean phase or at the final checkpoint.
 875 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
 876   assert(_markBitMap.covers(start, size), "Out of bounds");
 877   if (_collectorState >= Marking) {
 878     MutexLockerEx y(_markBitMap.lock(),
 879                     Mutex::_no_safepoint_check_flag);
 880     // [see comments preceding SweepClosure::do_blk() below for details]
 881     //
 882     // Can the P-bits be deleted now?  JJJ
 883     //
 884     // 1. need to mark the object as live so it isn't collected
 885     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
 886     // 3. need to mark the end of the object so marking, precleaning or sweeping
 887     //    can skip over uninitialized or unparsable objects. An allocated
 888     //    object is considered uninitialized for our purposes as long as
 889     //    its klass word is NULL.  All old gen objects are parsable
 890     //    as soon as they are initialized.)
 891     _markBitMap.mark(start);          // object is live
 892     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
 893     _markBitMap.mark(start + size - 1);
 894                                       // mark end of object
 895   }
 896   // check that oop looks uninitialized
 897   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
 898 }
 899 
 900 void CMSCollector::promoted(bool par, HeapWord* start,
 901                             bool is_obj_array, size_t obj_size) {
 902   assert(_markBitMap.covers(start), "Out of bounds");
 903   // See comment in direct_allocated() about when objects should
 904   // be allocated live.
 905   if (_collectorState >= Marking) {
 906     // we already hold the marking bit map lock, taken in
 907     // the prologue
 908     if (par) {
 909       _markBitMap.par_mark(start);
 910     } else {
 911       _markBitMap.mark(start);
 912     }
 913     // We don't need to mark the object as uninitialized (as
 914     // in direct_allocated above) because this is being done with the
 915     // world stopped and the object will be initialized by the
 916     // time the marking, precleaning or sweeping get to look at it.
 917     // But see the code for copying objects into the CMS generation,
 918     // where we need to ensure that concurrent readers of the
 919     // block offset table are able to safely navigate a block that
 920     // is in flux from being free to being allocated (and in
 921     // transition while being copied into) and subsequently
 922     // becoming a bona-fide object when the copy/promotion is complete.
 923     assert(SafepointSynchronize::is_at_safepoint(),
 924            "expect promotion only at safepoints");
 925 
 926     if (_collectorState < Sweeping) {
 927       // Mark the appropriate cards in the modUnionTable, so that
 928       // this object gets scanned before the sweep. If this is
 929       // not done, CMS generation references in the object might
 930       // not get marked.
 931       // For the case of arrays, which are otherwise precisely
 932       // marked, we need to dirty the entire array, not just its head.
 933       if (is_obj_array) {
 934         // The [par_]mark_range() method expects mr.end() below to
 935         // be aligned to the granularity of a bit's representation
 936         // in the heap. In the case of the MUT below, that's a
 937         // card size.
 938         MemRegion mr(start,
 939                      (HeapWord*)round_to((intptr_t)(start + obj_size),
 940                         CardTableModRefBS::card_size /* bytes */));
 941         if (par) {
 942           _modUnionTable.par_mark_range(mr);
 943         } else {
 944           _modUnionTable.mark_range(mr);
 945         }
 946       } else {  // not an obj array; we can just mark the head
 947         if (par) {
 948           _modUnionTable.par_mark(start);
 949         } else {
 950           _modUnionTable.mark(start);
 951         }
 952       }
 953     }
 954   }
 955 }
 956 
 957 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
 958   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
 959   // allocate, copy and if necessary update promoinfo --
 960   // delegate to underlying space.
 961   assert_lock_strong(freelistLock());
 962 
 963 #ifndef PRODUCT
 964   if (GenCollectedHeap::heap()->promotion_should_fail()) {
 965     return NULL;
 966   }
 967 #endif  // #ifndef PRODUCT
 968 
 969   oop res = _cmsSpace->promote(obj, obj_size);
 970   if (res == NULL) {
 971     // expand and retry
 972     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
 973     expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
 974     // Since this is the old generation, we don't try to promote
 975     // into a more senior generation.
 976     res = _cmsSpace->promote(obj, obj_size);
 977   }
 978   if (res != NULL) {
 979     // See comment in allocate() about when objects should
 980     // be allocated live.
 981     assert(obj->is_oop(), "Will dereference klass pointer below");
 982     collector()->promoted(false,           // Not parallel
 983                           (HeapWord*)res, obj->is_objArray(), obj_size);
 984     // promotion counters
 985     NOT_PRODUCT(
 986       _numObjectsPromoted++;
 987       _numWordsPromoted +=
 988         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
 989     )
 990   }
 991   return res;
 992 }
 993 
 994 
 995 // IMPORTANT: Notes on object size recognition in CMS.
 996 // ---------------------------------------------------
 997 // A block of storage in the CMS generation is always in
 998 // one of three states. A free block (FREE), an allocated
 999 // object (OBJECT) whose size() method reports the correct size,
1000 // and an intermediate state (TRANSIENT) in which its size cannot
1001 // be accurately determined.
1002 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
1003 // -----------------------------------------------------
1004 // FREE:      klass_word & 1 == 1; mark_word holds block size
1005 //
1006 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1007 //            obj->size() computes correct size
1008 //
1009 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1010 //
1011 // STATE IDENTIFICATION: (64 bit+COOPS)
1012 // ------------------------------------
1013 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1014 //
1015 // OBJECT:    klass_word installed; klass_word != 0;
1016 //            obj->size() computes correct size
1017 //
1018 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1019 //
1020 //
1021 // STATE TRANSITION DIAGRAM
1022 //
1023 //        mut / parnew                     mut  /  parnew
1024 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1025 //  ^                                                                   |
1026 //  |------------------------ DEAD <------------------------------------|
1027 //         sweep                            mut
1028 //
1029 // While a block is in TRANSIENT state its size cannot be determined
1030 // so readers will either need to come back later or stall until
1031 // the size can be determined. Note that for the case of direct
1032 // allocation, P-bits, when available, may be used to determine the
1033 // size of an object that may not yet have been initialized.
1034 
1035 // Things to support parallel young-gen collection.
1036 oop
1037 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1038                                            oop old, markOop m,
1039                                            size_t word_sz) {
1040 #ifndef PRODUCT
1041   if (GenCollectedHeap::heap()->promotion_should_fail()) {
1042     return NULL;
1043   }
1044 #endif  // #ifndef PRODUCT
1045 
1046   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1047   PromotionInfo* promoInfo = &ps->promo;
1048   // if we are tracking promotions, then first ensure space for
1049   // promotion (including spooling space for saving header if necessary).
1050   // then allocate and copy, then track promoted info if needed.
1051   // When tracking (see PromotionInfo::track()), the mark word may
1052   // be displaced and in this case restoration of the mark word
1053   // occurs in the (oop_since_save_marks_)iterate phase.
1054   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1055     // Out of space for allocating spooling buffers;
1056     // try expanding and allocating spooling buffers.
1057     if (!expand_and_ensure_spooling_space(promoInfo)) {
1058       return NULL;
1059     }
1060   }
1061   assert(promoInfo->has_spooling_space(), "Control point invariant");
1062   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1063   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1064   if (obj_ptr == NULL) {
1065      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1066      if (obj_ptr == NULL) {
1067        return NULL;
1068      }
1069   }
1070   oop obj = oop(obj_ptr);
1071   OrderAccess::storestore();
1072   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1073   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1074   // IMPORTANT: See note on object initialization for CMS above.
1075   // Otherwise, copy the object.  Here we must be careful to insert the
1076   // klass pointer last, since this marks the block as an allocated object.
1077   // Except with compressed oops it's the mark word.
1078   HeapWord* old_ptr = (HeapWord*)old;
1079   // Restore the mark word copied above.
1080   obj->set_mark(m);
1081   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1082   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1083   OrderAccess::storestore();
1084 
1085   if (UseCompressedClassPointers) {
1086     // Copy gap missed by (aligned) header size calculation below
1087     obj->set_klass_gap(old->klass_gap());
1088   }
1089   if (word_sz > (size_t)oopDesc::header_size()) {
1090     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1091                                  obj_ptr + oopDesc::header_size(),
1092                                  word_sz - oopDesc::header_size());
1093   }
1094 
1095   // Now we can track the promoted object, if necessary.  We take care
1096   // to delay the transition from uninitialized to full object
1097   // (i.e., insertion of klass pointer) until after, so that it
1098   // atomically becomes a promoted object.
1099   if (promoInfo->tracking()) {
1100     promoInfo->track((PromotedObject*)obj, old->klass());
1101   }
1102   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1103   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1104   assert(old->is_oop(), "Will use and dereference old klass ptr below");
1105 
1106   // Finally, install the klass pointer (this should be volatile).
1107   OrderAccess::storestore();
1108   obj->set_klass(old->klass());
1109   // We should now be able to calculate the right size for this object
1110   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1111 
1112   collector()->promoted(true,          // parallel
1113                         obj_ptr, old->is_objArray(), word_sz);
1114 
1115   NOT_PRODUCT(
1116     Atomic::inc_ptr(&_numObjectsPromoted);
1117     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1118   )
1119 
1120   return obj;
1121 }
1122 
1123 void
1124 ConcurrentMarkSweepGeneration::
1125 par_promote_alloc_done(int thread_num) {
1126   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1127   ps->lab.retire(thread_num);
1128 }
1129 
1130 void
1131 ConcurrentMarkSweepGeneration::
1132 par_oop_since_save_marks_iterate_done(int thread_num) {
1133   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1134   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1135   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1136 }
1137 
1138 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1139                                                    size_t size,
1140                                                    bool   tlab)
1141 {
1142   // We allow a STW collection only if a full
1143   // collection was requested.
1144   return full || should_allocate(size, tlab); // FIX ME !!!
1145   // This and promotion failure handling are connected at the
1146   // hip and should be fixed by untying them.
1147 }
1148 
1149 bool CMSCollector::shouldConcurrentCollect() {
1150   if (_full_gc_requested) {
1151     if (Verbose && PrintGCDetails) {
1152       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1153                              " gc request (or gc_locker)");
1154     }
1155     return true;
1156   }
1157 
1158   FreelistLocker x(this);
1159   // ------------------------------------------------------------------
1160   // Print out lots of information which affects the initiation of
1161   // a collection.
1162   if (PrintCMSInitiationStatistics && stats().valid()) {
1163     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1164     gclog_or_tty->stamp();
1165     gclog_or_tty->cr();
1166     stats().print_on(gclog_or_tty);
1167     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1168       stats().time_until_cms_gen_full());
1169     gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1170     gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1171                            _cmsGen->contiguous_available());
1172     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1173     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1174     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1175     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1176     gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1177     gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1178     gclog_or_tty->print_cr("metadata initialized %d",
1179       MetaspaceGC::should_concurrent_collect());
1180   }
1181   // ------------------------------------------------------------------
1182 
1183   // If the estimated time to complete a cms collection (cms_duration())
1184   // is less than the estimated time remaining until the cms generation
1185   // is full, start a collection.
1186   if (!UseCMSInitiatingOccupancyOnly) {
1187     if (stats().valid()) {
1188       if (stats().time_until_cms_start() == 0.0) {
1189         return true;
1190       }
1191     } else {
1192       // We want to conservatively collect somewhat early in order
1193       // to try and "bootstrap" our CMS/promotion statistics;
1194       // this branch will not fire after the first successful CMS
1195       // collection because the stats should then be valid.
1196       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1197         if (Verbose && PrintGCDetails) {
1198           gclog_or_tty->print_cr(
1199             " CMSCollector: collect for bootstrapping statistics:"
1200             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1201             _bootstrap_occupancy);
1202         }
1203         return true;
1204       }
1205     }
1206   }
1207 
1208   // Otherwise, we start a collection cycle if
1209   // old gen want a collection cycle started. Each may use
1210   // an appropriate criterion for making this decision.
1211   // XXX We need to make sure that the gen expansion
1212   // criterion dovetails well with this. XXX NEED TO FIX THIS
1213   if (_cmsGen->should_concurrent_collect()) {
1214     if (Verbose && PrintGCDetails) {
1215       gclog_or_tty->print_cr("CMS old gen initiated");
1216     }
1217     return true;
1218   }
1219 
1220   // We start a collection if we believe an incremental collection may fail;
1221   // this is not likely to be productive in practice because it's probably too
1222   // late anyway.
1223   GenCollectedHeap* gch = GenCollectedHeap::heap();
1224   assert(gch->collector_policy()->is_generation_policy(),
1225          "You may want to check the correctness of the following");
1226   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1227     if (Verbose && PrintGCDetails) {
1228       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1229     }
1230     return true;
1231   }
1232 
1233   if (MetaspaceGC::should_concurrent_collect()) {
1234     if (Verbose && PrintGCDetails) {
1235       gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1236     }
1237     return true;
1238   }
1239 
1240   // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1241   if (CMSTriggerInterval >= 0) {
1242     if (CMSTriggerInterval == 0) {
1243       // Trigger always
1244       return true;
1245     }
1246 
1247     // Check the CMS time since begin (we do not check the stats validity
1248     // as we want to be able to trigger the first CMS cycle as well)
1249     if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1250       if (Verbose && PrintGCDetails) {
1251         if (stats().valid()) {
1252           gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1253                                  stats().cms_time_since_begin());
1254         } else {
1255           gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1256         }
1257       }
1258       return true;
1259     }
1260   }
1261 
1262   return false;
1263 }
1264 
1265 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1266 
1267 // Clear _expansion_cause fields of constituent generations
1268 void CMSCollector::clear_expansion_cause() {
1269   _cmsGen->clear_expansion_cause();
1270 }
1271 
1272 // We should be conservative in starting a collection cycle.  To
1273 // start too eagerly runs the risk of collecting too often in the
1274 // extreme.  To collect too rarely falls back on full collections,
1275 // which works, even if not optimum in terms of concurrent work.
1276 // As a work around for too eagerly collecting, use the flag
1277 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1278 // giving the user an easily understandable way of controlling the
1279 // collections.
1280 // We want to start a new collection cycle if any of the following
1281 // conditions hold:
1282 // . our current occupancy exceeds the configured initiating occupancy
1283 //   for this generation, or
1284 // . we recently needed to expand this space and have not, since that
1285 //   expansion, done a collection of this generation, or
1286 // . the underlying space believes that it may be a good idea to initiate
1287 //   a concurrent collection (this may be based on criteria such as the
1288 //   following: the space uses linear allocation and linear allocation is
1289 //   going to fail, or there is believed to be excessive fragmentation in
1290 //   the generation, etc... or ...
1291 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1292 //   the case of the old generation; see CR 6543076):
1293 //   we may be approaching a point at which allocation requests may fail because
1294 //   we will be out of sufficient free space given allocation rate estimates.]
1295 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1296 
1297   assert_lock_strong(freelistLock());
1298   if (occupancy() > initiating_occupancy()) {
1299     if (PrintGCDetails && Verbose) {
1300       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
1301         short_name(), occupancy(), initiating_occupancy());
1302     }
1303     return true;
1304   }
1305   if (UseCMSInitiatingOccupancyOnly) {
1306     return false;
1307   }
1308   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1309     if (PrintGCDetails && Verbose) {
1310       gclog_or_tty->print(" %s: collect because expanded for allocation ",
1311         short_name());
1312     }
1313     return true;
1314   }
1315   if (_cmsSpace->should_concurrent_collect()) {
1316     if (PrintGCDetails && Verbose) {
1317       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1318         short_name());
1319     }
1320     return true;
1321   }
1322   return false;
1323 }
1324 
1325 void ConcurrentMarkSweepGeneration::collect(bool   full,
1326                                             bool   clear_all_soft_refs,
1327                                             size_t size,
1328                                             bool   tlab)
1329 {
1330   collector()->collect(full, clear_all_soft_refs, size, tlab);
1331 }
1332 
1333 void CMSCollector::collect(bool   full,
1334                            bool   clear_all_soft_refs,
1335                            size_t size,
1336                            bool   tlab)
1337 {
1338   // The following "if" branch is present for defensive reasons.
1339   // In the current uses of this interface, it can be replaced with:
1340   // assert(!GC_locker.is_active(), "Can't be called otherwise");
1341   // But I am not placing that assert here to allow future
1342   // generality in invoking this interface.
1343   if (GC_locker::is_active()) {
1344     // A consistency test for GC_locker
1345     assert(GC_locker::needs_gc(), "Should have been set already");
1346     // Skip this foreground collection, instead
1347     // expanding the heap if necessary.
1348     // Need the free list locks for the call to free() in compute_new_size()
1349     compute_new_size();
1350     return;
1351   }
1352   acquire_control_and_collect(full, clear_all_soft_refs);
1353 }
1354 
1355 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1356   GenCollectedHeap* gch = GenCollectedHeap::heap();
1357   unsigned int gc_count = gch->total_full_collections();
1358   if (gc_count == full_gc_count) {
1359     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1360     _full_gc_requested = true;
1361     _full_gc_cause = cause;
1362     CGC_lock->notify();   // nudge CMS thread
1363   } else {
1364     assert(gc_count > full_gc_count, "Error: causal loop");
1365   }
1366 }
1367 
1368 bool CMSCollector::is_external_interruption() {
1369   GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1370   return GCCause::is_user_requested_gc(cause) ||
1371          GCCause::is_serviceability_requested_gc(cause);
1372 }
1373 
1374 void CMSCollector::report_concurrent_mode_interruption() {
1375   if (is_external_interruption()) {
1376     if (PrintGCDetails) {
1377       gclog_or_tty->print(" (concurrent mode interrupted)");
1378     }
1379   } else {
1380     if (PrintGCDetails) {
1381       gclog_or_tty->print(" (concurrent mode failure)");
1382     }
1383     _gc_tracer_cm->report_concurrent_mode_failure();
1384   }
1385 }
1386 
1387 
1388 // The foreground and background collectors need to coordinate in order
1389 // to make sure that they do not mutually interfere with CMS collections.
1390 // When a background collection is active,
1391 // the foreground collector may need to take over (preempt) and
1392 // synchronously complete an ongoing collection. Depending on the
1393 // frequency of the background collections and the heap usage
1394 // of the application, this preemption can be seldom or frequent.
1395 // There are only certain
1396 // points in the background collection that the "collection-baton"
1397 // can be passed to the foreground collector.
1398 //
1399 // The foreground collector will wait for the baton before
1400 // starting any part of the collection.  The foreground collector
1401 // will only wait at one location.
1402 //
1403 // The background collector will yield the baton before starting a new
1404 // phase of the collection (e.g., before initial marking, marking from roots,
1405 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1406 // of the loop which switches the phases. The background collector does some
1407 // of the phases (initial mark, final re-mark) with the world stopped.
1408 // Because of locking involved in stopping the world,
1409 // the foreground collector should not block waiting for the background
1410 // collector when it is doing a stop-the-world phase.  The background
1411 // collector will yield the baton at an additional point just before
1412 // it enters a stop-the-world phase.  Once the world is stopped, the
1413 // background collector checks the phase of the collection.  If the
1414 // phase has not changed, it proceeds with the collection.  If the
1415 // phase has changed, it skips that phase of the collection.  See
1416 // the comments on the use of the Heap_lock in collect_in_background().
1417 //
1418 // Variable used in baton passing.
1419 //   _foregroundGCIsActive - Set to true by the foreground collector when
1420 //      it wants the baton.  The foreground clears it when it has finished
1421 //      the collection.
1422 //   _foregroundGCShouldWait - Set to true by the background collector
1423 //        when it is running.  The foreground collector waits while
1424 //      _foregroundGCShouldWait is true.
1425 //  CGC_lock - monitor used to protect access to the above variables
1426 //      and to notify the foreground and background collectors.
1427 //  _collectorState - current state of the CMS collection.
1428 //
1429 // The foreground collector
1430 //   acquires the CGC_lock
1431 //   sets _foregroundGCIsActive
1432 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1433 //     various locks acquired in preparation for the collection
1434 //     are released so as not to block the background collector
1435 //     that is in the midst of a collection
1436 //   proceeds with the collection
1437 //   clears _foregroundGCIsActive
1438 //   returns
1439 //
1440 // The background collector in a loop iterating on the phases of the
1441 //      collection
1442 //   acquires the CGC_lock
1443 //   sets _foregroundGCShouldWait
1444 //   if _foregroundGCIsActive is set
1445 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1446 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1447 //     and exits the loop.
1448 //   otherwise
1449 //     proceed with that phase of the collection
1450 //     if the phase is a stop-the-world phase,
1451 //       yield the baton once more just before enqueueing
1452 //       the stop-world CMS operation (executed by the VM thread).
1453 //   returns after all phases of the collection are done
1454 //
1455 
1456 void CMSCollector::acquire_control_and_collect(bool full,
1457         bool clear_all_soft_refs) {
1458   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1459   assert(!Thread::current()->is_ConcurrentGC_thread(),
1460          "shouldn't try to acquire control from self!");
1461 
1462   // Start the protocol for acquiring control of the
1463   // collection from the background collector (aka CMS thread).
1464   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1465          "VM thread should have CMS token");
1466   // Remember the possibly interrupted state of an ongoing
1467   // concurrent collection
1468   CollectorState first_state = _collectorState;
1469 
1470   // Signal to a possibly ongoing concurrent collection that
1471   // we want to do a foreground collection.
1472   _foregroundGCIsActive = true;
1473 
1474   // release locks and wait for a notify from the background collector
1475   // releasing the locks in only necessary for phases which
1476   // do yields to improve the granularity of the collection.
1477   assert_lock_strong(bitMapLock());
1478   // We need to lock the Free list lock for the space that we are
1479   // currently collecting.
1480   assert(haveFreelistLocks(), "Must be holding free list locks");
1481   bitMapLock()->unlock();
1482   releaseFreelistLocks();
1483   {
1484     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1485     if (_foregroundGCShouldWait) {
1486       // We are going to be waiting for action for the CMS thread;
1487       // it had better not be gone (for instance at shutdown)!
1488       assert(ConcurrentMarkSweepThread::cmst() != NULL,
1489              "CMS thread must be running");
1490       // Wait here until the background collector gives us the go-ahead
1491       ConcurrentMarkSweepThread::clear_CMS_flag(
1492         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1493       // Get a possibly blocked CMS thread going:
1494       //   Note that we set _foregroundGCIsActive true above,
1495       //   without protection of the CGC_lock.
1496       CGC_lock->notify();
1497       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1498              "Possible deadlock");
1499       while (_foregroundGCShouldWait) {
1500         // wait for notification
1501         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1502         // Possibility of delay/starvation here, since CMS token does
1503         // not know to give priority to VM thread? Actually, i think
1504         // there wouldn't be any delay/starvation, but the proof of
1505         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1506       }
1507       ConcurrentMarkSweepThread::set_CMS_flag(
1508         ConcurrentMarkSweepThread::CMS_vm_has_token);
1509     }
1510   }
1511   // The CMS_token is already held.  Get back the other locks.
1512   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1513          "VM thread should have CMS token");
1514   getFreelistLocks();
1515   bitMapLock()->lock_without_safepoint_check();
1516   if (TraceCMSState) {
1517     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1518       INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1519     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
1520   }
1521 
1522   // Inform cms gen if this was due to partial collection failing.
1523   // The CMS gen may use this fact to determine its expansion policy.
1524   GenCollectedHeap* gch = GenCollectedHeap::heap();
1525   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1526     assert(!_cmsGen->incremental_collection_failed(),
1527            "Should have been noticed, reacted to and cleared");
1528     _cmsGen->set_incremental_collection_failed();
1529   }
1530 
1531   if (first_state > Idling) {
1532     report_concurrent_mode_interruption();
1533   }
1534 
1535   set_did_compact(true);
1536 
1537   // If the collection is being acquired from the background
1538   // collector, there may be references on the discovered
1539   // references lists.  Abandon those references, since some
1540   // of them may have become unreachable after concurrent
1541   // discovery; the STW compacting collector will redo discovery
1542   // more precisely, without being subject to floating garbage.
1543   // Leaving otherwise unreachable references in the discovered
1544   // lists would require special handling.
1545   ref_processor()->disable_discovery();
1546   ref_processor()->abandon_partial_discovery();
1547   ref_processor()->verify_no_references_recorded();
1548 
1549   if (first_state > Idling) {
1550     save_heap_summary();
1551   }
1552 
1553   do_compaction_work(clear_all_soft_refs);
1554 
1555   // Has the GC time limit been exceeded?
1556   size_t max_eden_size = _young_gen->max_capacity() -
1557                          _young_gen->to()->capacity() -
1558                          _young_gen->from()->capacity();
1559   GCCause::Cause gc_cause = gch->gc_cause();
1560   size_policy()->check_gc_overhead_limit(_young_gen->used(),
1561                                          _young_gen->eden()->used(),
1562                                          _cmsGen->max_capacity(),
1563                                          max_eden_size,
1564                                          full,
1565                                          gc_cause,
1566                                          gch->collector_policy());
1567 
1568   // Reset the expansion cause, now that we just completed
1569   // a collection cycle.
1570   clear_expansion_cause();
1571   _foregroundGCIsActive = false;
1572   return;
1573 }
1574 
1575 // Resize the tenured generation
1576 // after obtaining the free list locks for the
1577 // two generations.
1578 void CMSCollector::compute_new_size() {
1579   assert_locked_or_safepoint(Heap_lock);
1580   FreelistLocker z(this);
1581   MetaspaceGC::compute_new_size();
1582   _cmsGen->compute_new_size_free_list();
1583 }
1584 
1585 // A work method used by the foreground collector to do
1586 // a mark-sweep-compact.
1587 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1588   GenCollectedHeap* gch = GenCollectedHeap::heap();
1589 
1590   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1591   gc_timer->register_gc_start();
1592 
1593   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1594   gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1595 
1596   GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
1597 
1598   // Temporarily widen the span of the weak reference processing to
1599   // the entire heap.
1600   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1601   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1602   // Temporarily, clear the "is_alive_non_header" field of the
1603   // reference processor.
1604   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1605   // Temporarily make reference _processing_ single threaded (non-MT).
1606   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1607   // Temporarily make refs discovery atomic
1608   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1609   // Temporarily make reference _discovery_ single threaded (non-MT)
1610   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1611 
1612   ref_processor()->set_enqueuing_is_done(false);
1613   ref_processor()->enable_discovery();
1614   ref_processor()->setup_policy(clear_all_soft_refs);
1615   // If an asynchronous collection finishes, the _modUnionTable is
1616   // all clear.  If we are assuming the collection from an asynchronous
1617   // collection, clear the _modUnionTable.
1618   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1619     "_modUnionTable should be clear if the baton was not passed");
1620   _modUnionTable.clear_all();
1621   assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1622     "mod union for klasses should be clear if the baton was passed");
1623   _ct->klass_rem_set()->clear_mod_union();
1624 
1625   // We must adjust the allocation statistics being maintained
1626   // in the free list space. We do so by reading and clearing
1627   // the sweep timer and updating the block flux rate estimates below.
1628   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1629   if (_inter_sweep_timer.is_active()) {
1630     _inter_sweep_timer.stop();
1631     // Note that we do not use this sample to update the _inter_sweep_estimate.
1632     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1633                                             _inter_sweep_estimate.padded_average(),
1634                                             _intra_sweep_estimate.padded_average());
1635   }
1636 
1637   GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1638   #ifdef ASSERT
1639     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1640     size_t free_size = cms_space->free();
1641     assert(free_size ==
1642            pointer_delta(cms_space->end(), cms_space->compaction_top())
1643            * HeapWordSize,
1644       "All the free space should be compacted into one chunk at top");
1645     assert(cms_space->dictionary()->total_chunk_size(
1646                                       debug_only(cms_space->freelistLock())) == 0 ||
1647            cms_space->totalSizeInIndexedFreeLists() == 0,
1648       "All the free space should be in a single chunk");
1649     size_t num = cms_space->totalCount();
1650     assert((free_size == 0 && num == 0) ||
1651            (free_size > 0  && (num == 1 || num == 2)),
1652          "There should be at most 2 free chunks after compaction");
1653   #endif // ASSERT
1654   _collectorState = Resetting;
1655   assert(_restart_addr == NULL,
1656          "Should have been NULL'd before baton was passed");
1657   reset_stw();
1658   _cmsGen->reset_after_compaction();
1659   _concurrent_cycles_since_last_unload = 0;
1660 
1661   // Clear any data recorded in the PLAB chunk arrays.
1662   if (_survivor_plab_array != NULL) {
1663     reset_survivor_plab_arrays();
1664   }
1665 
1666   // Adjust the per-size allocation stats for the next epoch.
1667   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1668   // Restart the "inter sweep timer" for the next epoch.
1669   _inter_sweep_timer.reset();
1670   _inter_sweep_timer.start();
1671 
1672   gc_timer->register_gc_end();
1673 
1674   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1675 
1676   // For a mark-sweep-compact, compute_new_size() will be called
1677   // in the heap's do_collection() method.
1678 }
1679 
1680 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1681   ContiguousSpace* eden_space = _young_gen->eden();
1682   ContiguousSpace* from_space = _young_gen->from();
1683   ContiguousSpace* to_space   = _young_gen->to();
1684   // Eden
1685   if (_eden_chunk_array != NULL) {
1686     gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1687                            p2i(eden_space->bottom()), p2i(eden_space->top()),
1688                            p2i(eden_space->end()), eden_space->capacity());
1689     gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1690                            "_eden_chunk_capacity=" SIZE_FORMAT,
1691                            _eden_chunk_index, _eden_chunk_capacity);
1692     for (size_t i = 0; i < _eden_chunk_index; i++) {
1693       gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1694                              i, p2i(_eden_chunk_array[i]));
1695     }
1696   }
1697   // Survivor
1698   if (_survivor_chunk_array != NULL) {
1699     gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1700                            p2i(from_space->bottom()), p2i(from_space->top()),
1701                            p2i(from_space->end()), from_space->capacity());
1702     gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1703                            "_survivor_chunk_capacity=" SIZE_FORMAT,
1704                            _survivor_chunk_index, _survivor_chunk_capacity);
1705     for (size_t i = 0; i < _survivor_chunk_index; i++) {
1706       gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1707                              i, p2i(_survivor_chunk_array[i]));
1708     }
1709   }
1710 }
1711 
1712 void CMSCollector::getFreelistLocks() const {
1713   // Get locks for all free lists in all generations that this
1714   // collector is responsible for
1715   _cmsGen->freelistLock()->lock_without_safepoint_check();
1716 }
1717 
1718 void CMSCollector::releaseFreelistLocks() const {
1719   // Release locks for all free lists in all generations that this
1720   // collector is responsible for
1721   _cmsGen->freelistLock()->unlock();
1722 }
1723 
1724 bool CMSCollector::haveFreelistLocks() const {
1725   // Check locks for all free lists in all generations that this
1726   // collector is responsible for
1727   assert_lock_strong(_cmsGen->freelistLock());
1728   PRODUCT_ONLY(ShouldNotReachHere());
1729   return true;
1730 }
1731 
1732 // A utility class that is used by the CMS collector to
1733 // temporarily "release" the foreground collector from its
1734 // usual obligation to wait for the background collector to
1735 // complete an ongoing phase before proceeding.
1736 class ReleaseForegroundGC: public StackObj {
1737  private:
1738   CMSCollector* _c;
1739  public:
1740   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1741     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1742     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1743     // allow a potentially blocked foreground collector to proceed
1744     _c->_foregroundGCShouldWait = false;
1745     if (_c->_foregroundGCIsActive) {
1746       CGC_lock->notify();
1747     }
1748     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1749            "Possible deadlock");
1750   }
1751 
1752   ~ReleaseForegroundGC() {
1753     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1754     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1755     _c->_foregroundGCShouldWait = true;
1756   }
1757 };
1758 
1759 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1760   assert(Thread::current()->is_ConcurrentGC_thread(),
1761     "A CMS asynchronous collection is only allowed on a CMS thread.");
1762 
1763   GenCollectedHeap* gch = GenCollectedHeap::heap();
1764   {
1765     bool safepoint_check = Mutex::_no_safepoint_check_flag;
1766     MutexLockerEx hl(Heap_lock, safepoint_check);
1767     FreelistLocker fll(this);
1768     MutexLockerEx x(CGC_lock, safepoint_check);
1769     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1770       // The foreground collector is active or we're
1771       // not using asynchronous collections.  Skip this
1772       // background collection.
1773       assert(!_foregroundGCShouldWait, "Should be clear");
1774       return;
1775     } else {
1776       assert(_collectorState == Idling, "Should be idling before start.");
1777       _collectorState = InitialMarking;
1778       register_gc_start(cause);
1779       // Reset the expansion cause, now that we are about to begin
1780       // a new cycle.
1781       clear_expansion_cause();
1782 
1783       // Clear the MetaspaceGC flag since a concurrent collection
1784       // is starting but also clear it after the collection.
1785       MetaspaceGC::set_should_concurrent_collect(false);
1786     }
1787     // Decide if we want to enable class unloading as part of the
1788     // ensuing concurrent GC cycle.
1789     update_should_unload_classes();
1790     _full_gc_requested = false;           // acks all outstanding full gc requests
1791     _full_gc_cause = GCCause::_no_gc;
1792     // Signal that we are about to start a collection
1793     gch->increment_total_full_collections();  // ... starting a collection cycle
1794     _collection_count_start = gch->total_full_collections();
1795   }
1796 
1797   // Used for PrintGC
1798   size_t prev_used;
1799   if (PrintGC && Verbose) {
1800     prev_used = _cmsGen->used();
1801   }
1802 
1803   // The change of the collection state is normally done at this level;
1804   // the exceptions are phases that are executed while the world is
1805   // stopped.  For those phases the change of state is done while the
1806   // world is stopped.  For baton passing purposes this allows the
1807   // background collector to finish the phase and change state atomically.
1808   // The foreground collector cannot wait on a phase that is done
1809   // while the world is stopped because the foreground collector already
1810   // has the world stopped and would deadlock.
1811   while (_collectorState != Idling) {
1812     if (TraceCMSState) {
1813       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1814         p2i(Thread::current()), _collectorState);
1815     }
1816     // The foreground collector
1817     //   holds the Heap_lock throughout its collection.
1818     //   holds the CMS token (but not the lock)
1819     //     except while it is waiting for the background collector to yield.
1820     //
1821     // The foreground collector should be blocked (not for long)
1822     //   if the background collector is about to start a phase
1823     //   executed with world stopped.  If the background
1824     //   collector has already started such a phase, the
1825     //   foreground collector is blocked waiting for the
1826     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
1827     //   are executed in the VM thread.
1828     //
1829     // The locking order is
1830     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
1831     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
1832     //   CMS token  (claimed in
1833     //                stop_world_and_do() -->
1834     //                  safepoint_synchronize() -->
1835     //                    CMSThread::synchronize())
1836 
1837     {
1838       // Check if the FG collector wants us to yield.
1839       CMSTokenSync x(true); // is cms thread
1840       if (waitForForegroundGC()) {
1841         // We yielded to a foreground GC, nothing more to be
1842         // done this round.
1843         assert(_foregroundGCShouldWait == false, "We set it to false in "
1844                "waitForForegroundGC()");
1845         if (TraceCMSState) {
1846           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1847             " exiting collection CMS state %d",
1848             p2i(Thread::current()), _collectorState);
1849         }
1850         return;
1851       } else {
1852         // The background collector can run but check to see if the
1853         // foreground collector has done a collection while the
1854         // background collector was waiting to get the CGC_lock
1855         // above.  If yes, break so that _foregroundGCShouldWait
1856         // is cleared before returning.
1857         if (_collectorState == Idling) {
1858           break;
1859         }
1860       }
1861     }
1862 
1863     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1864       "should be waiting");
1865 
1866     switch (_collectorState) {
1867       case InitialMarking:
1868         {
1869           ReleaseForegroundGC x(this);
1870           stats().record_cms_begin();
1871           VM_CMS_Initial_Mark initial_mark_op(this);
1872           VMThread::execute(&initial_mark_op);
1873         }
1874         // The collector state may be any legal state at this point
1875         // since the background collector may have yielded to the
1876         // foreground collector.
1877         break;
1878       case Marking:
1879         // initial marking in checkpointRootsInitialWork has been completed
1880         if (markFromRoots()) { // we were successful
1881           assert(_collectorState == Precleaning, "Collector state should "
1882             "have changed");
1883         } else {
1884           assert(_foregroundGCIsActive, "Internal state inconsistency");
1885         }
1886         break;
1887       case Precleaning:
1888         // marking from roots in markFromRoots has been completed
1889         preclean();
1890         assert(_collectorState == AbortablePreclean ||
1891                _collectorState == FinalMarking,
1892                "Collector state should have changed");
1893         break;
1894       case AbortablePreclean:
1895         abortable_preclean();
1896         assert(_collectorState == FinalMarking, "Collector state should "
1897           "have changed");
1898         break;
1899       case FinalMarking:
1900         {
1901           ReleaseForegroundGC x(this);
1902 
1903           VM_CMS_Final_Remark final_remark_op(this);
1904           VMThread::execute(&final_remark_op);
1905         }
1906         assert(_foregroundGCShouldWait, "block post-condition");
1907         break;
1908       case Sweeping:
1909         // final marking in checkpointRootsFinal has been completed
1910         sweep();
1911         assert(_collectorState == Resizing, "Collector state change "
1912           "to Resizing must be done under the free_list_lock");
1913 
1914       case Resizing: {
1915         // Sweeping has been completed...
1916         // At this point the background collection has completed.
1917         // Don't move the call to compute_new_size() down
1918         // into code that might be executed if the background
1919         // collection was preempted.
1920         {
1921           ReleaseForegroundGC x(this);   // unblock FG collection
1922           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
1923           CMSTokenSync        z(true);   // not strictly needed.
1924           if (_collectorState == Resizing) {
1925             compute_new_size();
1926             save_heap_summary();
1927             _collectorState = Resetting;
1928           } else {
1929             assert(_collectorState == Idling, "The state should only change"
1930                    " because the foreground collector has finished the collection");
1931           }
1932         }
1933         break;
1934       }
1935       case Resetting:
1936         // CMS heap resizing has been completed
1937         reset_concurrent();
1938         assert(_collectorState == Idling, "Collector state should "
1939           "have changed");
1940 
1941         MetaspaceGC::set_should_concurrent_collect(false);
1942 
1943         stats().record_cms_end();
1944         // Don't move the concurrent_phases_end() and compute_new_size()
1945         // calls to here because a preempted background collection
1946         // has it's state set to "Resetting".
1947         break;
1948       case Idling:
1949       default:
1950         ShouldNotReachHere();
1951         break;
1952     }
1953     if (TraceCMSState) {
1954       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
1955         p2i(Thread::current()), _collectorState);
1956     }
1957     assert(_foregroundGCShouldWait, "block post-condition");
1958   }
1959 
1960   // Should this be in gc_epilogue?
1961   collector_policy()->counters()->update_counters();
1962 
1963   {
1964     // Clear _foregroundGCShouldWait and, in the event that the
1965     // foreground collector is waiting, notify it, before
1966     // returning.
1967     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1968     _foregroundGCShouldWait = false;
1969     if (_foregroundGCIsActive) {
1970       CGC_lock->notify();
1971     }
1972     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1973            "Possible deadlock");
1974   }
1975   if (TraceCMSState) {
1976     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1977       " exiting collection CMS state %d",
1978       p2i(Thread::current()), _collectorState);
1979   }
1980   if (PrintGC && Verbose) {
1981     _cmsGen->print_heap_change(prev_used);
1982   }
1983 }
1984 
1985 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1986   _cms_start_registered = true;
1987   _gc_timer_cm->register_gc_start();
1988   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1989 }
1990 
1991 void CMSCollector::register_gc_end() {
1992   if (_cms_start_registered) {
1993     report_heap_summary(GCWhen::AfterGC);
1994 
1995     _gc_timer_cm->register_gc_end();
1996     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1997     _cms_start_registered = false;
1998   }
1999 }
2000 
2001 void CMSCollector::save_heap_summary() {
2002   GenCollectedHeap* gch = GenCollectedHeap::heap();
2003   _last_heap_summary = gch->create_heap_summary();
2004   _last_metaspace_summary = gch->create_metaspace_summary();
2005 }
2006 
2007 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2008   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2009   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2010 }
2011 
2012 bool CMSCollector::waitForForegroundGC() {
2013   bool res = false;
2014   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2015          "CMS thread should have CMS token");
2016   // Block the foreground collector until the
2017   // background collectors decides whether to
2018   // yield.
2019   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2020   _foregroundGCShouldWait = true;
2021   if (_foregroundGCIsActive) {
2022     // The background collector yields to the
2023     // foreground collector and returns a value
2024     // indicating that it has yielded.  The foreground
2025     // collector can proceed.
2026     res = true;
2027     _foregroundGCShouldWait = false;
2028     ConcurrentMarkSweepThread::clear_CMS_flag(
2029       ConcurrentMarkSweepThread::CMS_cms_has_token);
2030     ConcurrentMarkSweepThread::set_CMS_flag(
2031       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2032     // Get a possibly blocked foreground thread going
2033     CGC_lock->notify();
2034     if (TraceCMSState) {
2035       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2036         p2i(Thread::current()), _collectorState);
2037     }
2038     while (_foregroundGCIsActive) {
2039       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2040     }
2041     ConcurrentMarkSweepThread::set_CMS_flag(
2042       ConcurrentMarkSweepThread::CMS_cms_has_token);
2043     ConcurrentMarkSweepThread::clear_CMS_flag(
2044       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2045   }
2046   if (TraceCMSState) {
2047     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2048       p2i(Thread::current()), _collectorState);
2049   }
2050   return res;
2051 }
2052 
2053 // Because of the need to lock the free lists and other structures in
2054 // the collector, common to all the generations that the collector is
2055 // collecting, we need the gc_prologues of individual CMS generations
2056 // delegate to their collector. It may have been simpler had the
2057 // current infrastructure allowed one to call a prologue on a
2058 // collector. In the absence of that we have the generation's
2059 // prologue delegate to the collector, which delegates back
2060 // some "local" work to a worker method in the individual generations
2061 // that it's responsible for collecting, while itself doing any
2062 // work common to all generations it's responsible for. A similar
2063 // comment applies to the  gc_epilogue()'s.
2064 // The role of the variable _between_prologue_and_epilogue is to
2065 // enforce the invocation protocol.
2066 void CMSCollector::gc_prologue(bool full) {
2067   // Call gc_prologue_work() for the CMSGen
2068   // we are responsible for.
2069 
2070   // The following locking discipline assumes that we are only called
2071   // when the world is stopped.
2072   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2073 
2074   // The CMSCollector prologue must call the gc_prologues for the
2075   // "generations" that it's responsible
2076   // for.
2077 
2078   assert(   Thread::current()->is_VM_thread()
2079          || (   CMSScavengeBeforeRemark
2080              && Thread::current()->is_ConcurrentGC_thread()),
2081          "Incorrect thread type for prologue execution");
2082 
2083   if (_between_prologue_and_epilogue) {
2084     // We have already been invoked; this is a gc_prologue delegation
2085     // from yet another CMS generation that we are responsible for, just
2086     // ignore it since all relevant work has already been done.
2087     return;
2088   }
2089 
2090   // set a bit saying prologue has been called; cleared in epilogue
2091   _between_prologue_and_epilogue = true;
2092   // Claim locks for common data structures, then call gc_prologue_work()
2093   // for each CMSGen.
2094 
2095   getFreelistLocks();   // gets free list locks on constituent spaces
2096   bitMapLock()->lock_without_safepoint_check();
2097 
2098   // Should call gc_prologue_work() for all cms gens we are responsible for
2099   bool duringMarking =    _collectorState >= Marking
2100                          && _collectorState < Sweeping;
2101 
2102   // The young collections clear the modified oops state, which tells if
2103   // there are any modified oops in the class. The remark phase also needs
2104   // that information. Tell the young collection to save the union of all
2105   // modified klasses.
2106   if (duringMarking) {
2107     _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2108   }
2109 
2110   bool registerClosure = duringMarking;
2111 
2112   _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2113 
2114   if (!full) {
2115     stats().record_gc0_begin();
2116   }
2117 }
2118 
2119 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2120 
2121   _capacity_at_prologue = capacity();
2122   _used_at_prologue = used();
2123 
2124   // Delegate to CMScollector which knows how to coordinate between
2125   // this and any other CMS generations that it is responsible for
2126   // collecting.
2127   collector()->gc_prologue(full);
2128 }
2129 
2130 // This is a "private" interface for use by this generation's CMSCollector.
2131 // Not to be called directly by any other entity (for instance,
2132 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2133 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2134   bool registerClosure, ModUnionClosure* modUnionClosure) {
2135   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2136   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2137     "Should be NULL");
2138   if (registerClosure) {
2139     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2140   }
2141   cmsSpace()->gc_prologue();
2142   // Clear stat counters
2143   NOT_PRODUCT(
2144     assert(_numObjectsPromoted == 0, "check");
2145     assert(_numWordsPromoted   == 0, "check");
2146     if (Verbose && PrintGC) {
2147       gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2148                           SIZE_FORMAT " bytes concurrently",
2149       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2150     }
2151     _numObjectsAllocated = 0;
2152     _numWordsAllocated   = 0;
2153   )
2154 }
2155 
2156 void CMSCollector::gc_epilogue(bool full) {
2157   // The following locking discipline assumes that we are only called
2158   // when the world is stopped.
2159   assert(SafepointSynchronize::is_at_safepoint(),
2160          "world is stopped assumption");
2161 
2162   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2163   // if linear allocation blocks need to be appropriately marked to allow the
2164   // the blocks to be parsable. We also check here whether we need to nudge the
2165   // CMS collector thread to start a new cycle (if it's not already active).
2166   assert(   Thread::current()->is_VM_thread()
2167          || (   CMSScavengeBeforeRemark
2168              && Thread::current()->is_ConcurrentGC_thread()),
2169          "Incorrect thread type for epilogue execution");
2170 
2171   if (!_between_prologue_and_epilogue) {
2172     // We have already been invoked; this is a gc_epilogue delegation
2173     // from yet another CMS generation that we are responsible for, just
2174     // ignore it since all relevant work has already been done.
2175     return;
2176   }
2177   assert(haveFreelistLocks(), "must have freelist locks");
2178   assert_lock_strong(bitMapLock());
2179 
2180   _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2181 
2182   _cmsGen->gc_epilogue_work(full);
2183 
2184   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2185     // in case sampling was not already enabled, enable it
2186     _start_sampling = true;
2187   }
2188   // reset _eden_chunk_array so sampling starts afresh
2189   _eden_chunk_index = 0;
2190 
2191   size_t cms_used   = _cmsGen->cmsSpace()->used();
2192 
2193   // update performance counters - this uses a special version of
2194   // update_counters() that allows the utilization to be passed as a
2195   // parameter, avoiding multiple calls to used().
2196   //
2197   _cmsGen->update_counters(cms_used);
2198 
2199   bitMapLock()->unlock();
2200   releaseFreelistLocks();
2201 
2202   if (!CleanChunkPoolAsync) {
2203     Chunk::clean_chunk_pool();
2204   }
2205 
2206   set_did_compact(false);
2207   _between_prologue_and_epilogue = false;  // ready for next cycle
2208 }
2209 
2210 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2211   collector()->gc_epilogue(full);
2212 
2213   // Also reset promotion tracking in par gc thread states.
2214   for (uint i = 0; i < ParallelGCThreads; i++) {
2215     _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2216   }
2217 }
2218 
2219 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2220   assert(!incremental_collection_failed(), "Should have been cleared");
2221   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2222   cmsSpace()->gc_epilogue();
2223     // Print stat counters
2224   NOT_PRODUCT(
2225     assert(_numObjectsAllocated == 0, "check");
2226     assert(_numWordsAllocated == 0, "check");
2227     if (Verbose && PrintGC) {
2228       gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2229                           SIZE_FORMAT " bytes",
2230                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2231     }
2232     _numObjectsPromoted = 0;
2233     _numWordsPromoted   = 0;
2234   )
2235 
2236   if (PrintGC && Verbose) {
2237     // Call down the chain in contiguous_available needs the freelistLock
2238     // so print this out before releasing the freeListLock.
2239     gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2240                         contiguous_available());
2241   }
2242 }
2243 
2244 #ifndef PRODUCT
2245 bool CMSCollector::have_cms_token() {
2246   Thread* thr = Thread::current();
2247   if (thr->is_VM_thread()) {
2248     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2249   } else if (thr->is_ConcurrentGC_thread()) {
2250     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2251   } else if (thr->is_GC_task_thread()) {
2252     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2253            ParGCRareEvent_lock->owned_by_self();
2254   }
2255   return false;
2256 }
2257 #endif
2258 
2259 // Check reachability of the given heap address in CMS generation,
2260 // treating all other generations as roots.
2261 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2262   // We could "guarantee" below, rather than assert, but I'll
2263   // leave these as "asserts" so that an adventurous debugger
2264   // could try this in the product build provided some subset of
2265   // the conditions were met, provided they were interested in the
2266   // results and knew that the computation below wouldn't interfere
2267   // with other concurrent computations mutating the structures
2268   // being read or written.
2269   assert(SafepointSynchronize::is_at_safepoint(),
2270          "Else mutations in object graph will make answer suspect");
2271   assert(have_cms_token(), "Should hold cms token");
2272   assert(haveFreelistLocks(), "must hold free list locks");
2273   assert_lock_strong(bitMapLock());
2274 
2275   // Clear the marking bit map array before starting, but, just
2276   // for kicks, first report if the given address is already marked
2277   gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2278                 _markBitMap.isMarked(addr) ? "" : " not");
2279 
2280   if (verify_after_remark()) {
2281     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2282     bool result = verification_mark_bm()->isMarked(addr);
2283     gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2284                            result ? "IS" : "is NOT");
2285     return result;
2286   } else {
2287     gclog_or_tty->print_cr("Could not compute result");
2288     return false;
2289   }
2290 }
2291 
2292 
2293 void
2294 CMSCollector::print_on_error(outputStream* st) {
2295   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2296   if (collector != NULL) {
2297     CMSBitMap* bitmap = &collector->_markBitMap;
2298     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2299     bitmap->print_on_error(st, " Bits: ");
2300 
2301     st->cr();
2302 
2303     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2304     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2305     mut_bitmap->print_on_error(st, " Bits: ");
2306   }
2307 }
2308 
2309 ////////////////////////////////////////////////////////
2310 // CMS Verification Support
2311 ////////////////////////////////////////////////////////
2312 // Following the remark phase, the following invariant
2313 // should hold -- each object in the CMS heap which is
2314 // marked in markBitMap() should be marked in the verification_mark_bm().
2315 
2316 class VerifyMarkedClosure: public BitMapClosure {
2317   CMSBitMap* _marks;
2318   bool       _failed;
2319 
2320  public:
2321   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2322 
2323   bool do_bit(size_t offset) {
2324     HeapWord* addr = _marks->offsetToHeapWord(offset);
2325     if (!_marks->isMarked(addr)) {
2326       oop(addr)->print_on(gclog_or_tty);
2327       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2328       _failed = true;
2329     }
2330     return true;
2331   }
2332 
2333   bool failed() { return _failed; }
2334 };
2335 
2336 bool CMSCollector::verify_after_remark(bool silent) {
2337   if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2338   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2339   static bool init = false;
2340 
2341   assert(SafepointSynchronize::is_at_safepoint(),
2342          "Else mutations in object graph will make answer suspect");
2343   assert(have_cms_token(),
2344          "Else there may be mutual interference in use of "
2345          " verification data structures");
2346   assert(_collectorState > Marking && _collectorState <= Sweeping,
2347          "Else marking info checked here may be obsolete");
2348   assert(haveFreelistLocks(), "must hold free list locks");
2349   assert_lock_strong(bitMapLock());
2350 
2351 
2352   // Allocate marking bit map if not already allocated
2353   if (!init) { // first time
2354     if (!verification_mark_bm()->allocate(_span)) {
2355       return false;
2356     }
2357     init = true;
2358   }
2359 
2360   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2361 
2362   // Turn off refs discovery -- so we will be tracing through refs.
2363   // This is as intended, because by this time
2364   // GC must already have cleared any refs that need to be cleared,
2365   // and traced those that need to be marked; moreover,
2366   // the marking done here is not going to interfere in any
2367   // way with the marking information used by GC.
2368   NoRefDiscovery no_discovery(ref_processor());
2369 
2370 #if defined(COMPILER2) || INCLUDE_JVMCI
2371   DerivedPointerTableDeactivate dpt_deact;
2372 #endif
2373 
2374   // Clear any marks from a previous round
2375   verification_mark_bm()->clear_all();
2376   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2377   verify_work_stacks_empty();
2378 
2379   GenCollectedHeap* gch = GenCollectedHeap::heap();
2380   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2381   // Update the saved marks which may affect the root scans.
2382   gch->save_marks();
2383 
2384   if (CMSRemarkVerifyVariant == 1) {
2385     // In this first variant of verification, we complete
2386     // all marking, then check if the new marks-vector is
2387     // a subset of the CMS marks-vector.
2388     verify_after_remark_work_1();
2389   } else if (CMSRemarkVerifyVariant == 2) {
2390     // In this second variant of verification, we flag an error
2391     // (i.e. an object reachable in the new marks-vector not reachable
2392     // in the CMS marks-vector) immediately, also indicating the
2393     // identify of an object (A) that references the unmarked object (B) --
2394     // presumably, a mutation to A failed to be picked up by preclean/remark?
2395     verify_after_remark_work_2();
2396   } else {
2397     warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2398             CMSRemarkVerifyVariant);
2399   }
2400   if (!silent) gclog_or_tty->print(" done] ");
2401   return true;
2402 }
2403 
2404 void CMSCollector::verify_after_remark_work_1() {
2405   ResourceMark rm;
2406   HandleMark  hm;
2407   GenCollectedHeap* gch = GenCollectedHeap::heap();
2408 
2409   // Get a clear set of claim bits for the roots processing to work with.
2410   ClassLoaderDataGraph::clear_claimed_marks();
2411 
2412   // Mark from roots one level into CMS
2413   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2414   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2415 
2416   {
2417     StrongRootsScope srs(1);
2418 
2419     gch->old_process_roots(&srs,
2420                            true,   // young gen as roots
2421                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2422                            should_unload_classes(),
2423                            &notOlder,
2424                            NULL);
2425   }
2426 
2427   // Now mark from the roots
2428   MarkFromRootsClosure markFromRootsClosure(this, _span,
2429     verification_mark_bm(), verification_mark_stack(),
2430     false /* don't yield */, true /* verifying */);
2431   assert(_restart_addr == NULL, "Expected pre-condition");
2432   verification_mark_bm()->iterate(&markFromRootsClosure);
2433   while (_restart_addr != NULL) {
2434     // Deal with stack overflow: by restarting at the indicated
2435     // address.
2436     HeapWord* ra = _restart_addr;
2437     markFromRootsClosure.reset(ra);
2438     _restart_addr = NULL;
2439     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2440   }
2441   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2442   verify_work_stacks_empty();
2443 
2444   // Marking completed -- now verify that each bit marked in
2445   // verification_mark_bm() is also marked in markBitMap(); flag all
2446   // errors by printing corresponding objects.
2447   VerifyMarkedClosure vcl(markBitMap());
2448   verification_mark_bm()->iterate(&vcl);
2449   if (vcl.failed()) {
2450     gclog_or_tty->print("Verification failed");
2451     gch->print_on(gclog_or_tty);
2452     fatal("CMS: failed marking verification after remark");
2453   }
2454 }
2455 
2456 class VerifyKlassOopsKlassClosure : public KlassClosure {
2457   class VerifyKlassOopsClosure : public OopClosure {
2458     CMSBitMap* _bitmap;
2459    public:
2460     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2461     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2462     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2463   } _oop_closure;
2464  public:
2465   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2466   void do_klass(Klass* k) {
2467     k->oops_do(&_oop_closure);
2468   }
2469 };
2470 
2471 void CMSCollector::verify_after_remark_work_2() {
2472   ResourceMark rm;
2473   HandleMark  hm;
2474   GenCollectedHeap* gch = GenCollectedHeap::heap();
2475 
2476   // Get a clear set of claim bits for the roots processing to work with.
2477   ClassLoaderDataGraph::clear_claimed_marks();
2478 
2479   // Mark from roots one level into CMS
2480   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2481                                      markBitMap());
2482   CLDToOopClosure cld_closure(&notOlder, true);
2483 
2484   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2485 
2486   {
2487     StrongRootsScope srs(1);
2488 
2489     gch->old_process_roots(&srs,
2490                            true,   // young gen as roots
2491                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2492                            should_unload_classes(),
2493                            &notOlder,
2494                            &cld_closure);
2495   }
2496 
2497   // Now mark from the roots
2498   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2499     verification_mark_bm(), markBitMap(), verification_mark_stack());
2500   assert(_restart_addr == NULL, "Expected pre-condition");
2501   verification_mark_bm()->iterate(&markFromRootsClosure);
2502   while (_restart_addr != NULL) {
2503     // Deal with stack overflow: by restarting at the indicated
2504     // address.
2505     HeapWord* ra = _restart_addr;
2506     markFromRootsClosure.reset(ra);
2507     _restart_addr = NULL;
2508     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2509   }
2510   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2511   verify_work_stacks_empty();
2512 
2513   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2514   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2515 
2516   // Marking completed -- now verify that each bit marked in
2517   // verification_mark_bm() is also marked in markBitMap(); flag all
2518   // errors by printing corresponding objects.
2519   VerifyMarkedClosure vcl(markBitMap());
2520   verification_mark_bm()->iterate(&vcl);
2521   assert(!vcl.failed(), "Else verification above should not have succeeded");
2522 }
2523 
2524 void ConcurrentMarkSweepGeneration::save_marks() {
2525   // delegate to CMS space
2526   cmsSpace()->save_marks();
2527   for (uint i = 0; i < ParallelGCThreads; i++) {
2528     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2529   }
2530 }
2531 
2532 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2533   return cmsSpace()->no_allocs_since_save_marks();
2534 }
2535 
2536 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2537                                                                 \
2538 void ConcurrentMarkSweepGeneration::                            \
2539 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2540   cl->assert_generation(this);                                  \
2541   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2542   save_marks();                                                 \
2543 }
2544 
2545 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2546 
2547 void
2548 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2549   if (freelistLock()->owned_by_self()) {
2550     Generation::oop_iterate(cl);
2551   } else {
2552     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2553     Generation::oop_iterate(cl);
2554   }
2555 }
2556 
2557 void
2558 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2559   if (freelistLock()->owned_by_self()) {
2560     Generation::object_iterate(cl);
2561   } else {
2562     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2563     Generation::object_iterate(cl);
2564   }
2565 }
2566 
2567 void
2568 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2569   if (freelistLock()->owned_by_self()) {
2570     Generation::safe_object_iterate(cl);
2571   } else {
2572     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2573     Generation::safe_object_iterate(cl);
2574   }
2575 }
2576 
2577 void
2578 ConcurrentMarkSweepGeneration::post_compact() {
2579 }
2580 
2581 void
2582 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2583   // Fix the linear allocation blocks to look like free blocks.
2584 
2585   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2586   // are not called when the heap is verified during universe initialization and
2587   // at vm shutdown.
2588   if (freelistLock()->owned_by_self()) {
2589     cmsSpace()->prepare_for_verify();
2590   } else {
2591     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2592     cmsSpace()->prepare_for_verify();
2593   }
2594 }
2595 
2596 void
2597 ConcurrentMarkSweepGeneration::verify() {
2598   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2599   // are not called when the heap is verified during universe initialization and
2600   // at vm shutdown.
2601   if (freelistLock()->owned_by_self()) {
2602     cmsSpace()->verify();
2603   } else {
2604     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2605     cmsSpace()->verify();
2606   }
2607 }
2608 
2609 void CMSCollector::verify() {
2610   _cmsGen->verify();
2611 }
2612 
2613 #ifndef PRODUCT
2614 bool CMSCollector::overflow_list_is_empty() const {
2615   assert(_num_par_pushes >= 0, "Inconsistency");
2616   if (_overflow_list == NULL) {
2617     assert(_num_par_pushes == 0, "Inconsistency");
2618   }
2619   return _overflow_list == NULL;
2620 }
2621 
2622 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2623 // merely consolidate assertion checks that appear to occur together frequently.
2624 void CMSCollector::verify_work_stacks_empty() const {
2625   assert(_markStack.isEmpty(), "Marking stack should be empty");
2626   assert(overflow_list_is_empty(), "Overflow list should be empty");
2627 }
2628 
2629 void CMSCollector::verify_overflow_empty() const {
2630   assert(overflow_list_is_empty(), "Overflow list should be empty");
2631   assert(no_preserved_marks(), "No preserved marks");
2632 }
2633 #endif // PRODUCT
2634 
2635 // Decide if we want to enable class unloading as part of the
2636 // ensuing concurrent GC cycle. We will collect and
2637 // unload classes if it's the case that:
2638 // (1) an explicit gc request has been made and the flag
2639 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2640 // (2) (a) class unloading is enabled at the command line, and
2641 //     (b) old gen is getting really full
2642 // NOTE: Provided there is no change in the state of the heap between
2643 // calls to this method, it should have idempotent results. Moreover,
2644 // its results should be monotonically increasing (i.e. going from 0 to 1,
2645 // but not 1 to 0) between successive calls between which the heap was
2646 // not collected. For the implementation below, it must thus rely on
2647 // the property that concurrent_cycles_since_last_unload()
2648 // will not decrease unless a collection cycle happened and that
2649 // _cmsGen->is_too_full() are
2650 // themselves also monotonic in that sense. See check_monotonicity()
2651 // below.
2652 void CMSCollector::update_should_unload_classes() {
2653   _should_unload_classes = false;
2654   // Condition 1 above
2655   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2656     _should_unload_classes = true;
2657   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2658     // Disjuncts 2.b.(i,ii,iii) above
2659     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2660                               CMSClassUnloadingMaxInterval)
2661                            || _cmsGen->is_too_full();
2662   }
2663 }
2664 
2665 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2666   bool res = should_concurrent_collect();
2667   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2668   return res;
2669 }
2670 
2671 void CMSCollector::setup_cms_unloading_and_verification_state() {
2672   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2673                              || VerifyBeforeExit;
2674   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2675 
2676   // We set the proper root for this CMS cycle here.
2677   if (should_unload_classes()) {   // Should unload classes this cycle
2678     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2679     set_verifying(should_verify);    // Set verification state for this cycle
2680     return;                            // Nothing else needs to be done at this time
2681   }
2682 
2683   // Not unloading classes this cycle
2684   assert(!should_unload_classes(), "Inconsistency!");
2685 
2686   // If we are not unloading classes then add SO_AllCodeCache to root
2687   // scanning options.
2688   add_root_scanning_option(rso);
2689 
2690   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2691     set_verifying(true);
2692   } else if (verifying() && !should_verify) {
2693     // We were verifying, but some verification flags got disabled.
2694     set_verifying(false);
2695     // Exclude symbols, strings and code cache elements from root scanning to
2696     // reduce IM and RM pauses.
2697     remove_root_scanning_option(rso);
2698   }
2699 }
2700 
2701 
2702 #ifndef PRODUCT
2703 HeapWord* CMSCollector::block_start(const void* p) const {
2704   const HeapWord* addr = (HeapWord*)p;
2705   if (_span.contains(p)) {
2706     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2707       return _cmsGen->cmsSpace()->block_start(p);
2708     }
2709   }
2710   return NULL;
2711 }
2712 #endif
2713 
2714 HeapWord*
2715 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2716                                                    bool   tlab,
2717                                                    bool   parallel) {
2718   CMSSynchronousYieldRequest yr;
2719   assert(!tlab, "Can't deal with TLAB allocation");
2720   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2721   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2722   if (GCExpandToAllocateDelayMillis > 0) {
2723     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2724   }
2725   return have_lock_and_allocate(word_size, tlab);
2726 }
2727 
2728 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2729     size_t bytes,
2730     size_t expand_bytes,
2731     CMSExpansionCause::Cause cause)
2732 {
2733 
2734   bool success = expand(bytes, expand_bytes);
2735 
2736   // remember why we expanded; this information is used
2737   // by shouldConcurrentCollect() when making decisions on whether to start
2738   // a new CMS cycle.
2739   if (success) {
2740     set_expansion_cause(cause);
2741     if (PrintGCDetails && Verbose) {
2742       gclog_or_tty->print_cr("Expanded CMS gen for %s",
2743         CMSExpansionCause::to_string(cause));
2744     }
2745   }
2746 }
2747 
2748 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2749   HeapWord* res = NULL;
2750   MutexLocker x(ParGCRareEvent_lock);
2751   while (true) {
2752     // Expansion by some other thread might make alloc OK now:
2753     res = ps->lab.alloc(word_sz);
2754     if (res != NULL) return res;
2755     // If there's not enough expansion space available, give up.
2756     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2757       return NULL;
2758     }
2759     // Otherwise, we try expansion.
2760     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2761     // Now go around the loop and try alloc again;
2762     // A competing par_promote might beat us to the expansion space,
2763     // so we may go around the loop again if promotion fails again.
2764     if (GCExpandToAllocateDelayMillis > 0) {
2765       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2766     }
2767   }
2768 }
2769 
2770 
2771 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2772   PromotionInfo* promo) {
2773   MutexLocker x(ParGCRareEvent_lock);
2774   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2775   while (true) {
2776     // Expansion by some other thread might make alloc OK now:
2777     if (promo->ensure_spooling_space()) {
2778       assert(promo->has_spooling_space(),
2779              "Post-condition of successful ensure_spooling_space()");
2780       return true;
2781     }
2782     // If there's not enough expansion space available, give up.
2783     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2784       return false;
2785     }
2786     // Otherwise, we try expansion.
2787     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2788     // Now go around the loop and try alloc again;
2789     // A competing allocation might beat us to the expansion space,
2790     // so we may go around the loop again if allocation fails again.
2791     if (GCExpandToAllocateDelayMillis > 0) {
2792       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2793     }
2794   }
2795 }
2796 
2797 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2798   // Only shrink if a compaction was done so that all the free space
2799   // in the generation is in a contiguous block at the end.
2800   if (did_compact()) {
2801     CardGeneration::shrink(bytes);
2802   }
2803 }
2804 
2805 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2806   assert_locked_or_safepoint(Heap_lock);
2807 }
2808 
2809 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2810   assert_locked_or_safepoint(Heap_lock);
2811   assert_lock_strong(freelistLock());
2812   if (PrintGCDetails && Verbose) {
2813     warning("Shrinking of CMS not yet implemented");
2814   }
2815   return;
2816 }
2817 
2818 
2819 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2820 // phases.
2821 class CMSPhaseAccounting: public StackObj {
2822  public:
2823   CMSPhaseAccounting(CMSCollector *collector,
2824                      const char *phase,
2825                      bool print_cr = true);
2826   ~CMSPhaseAccounting();
2827 
2828  private:
2829   CMSCollector *_collector;
2830   const char *_phase;
2831   elapsedTimer _wallclock;
2832   bool _print_cr;
2833 
2834  public:
2835   // Not MT-safe; so do not pass around these StackObj's
2836   // where they may be accessed by other threads.
2837   jlong wallclock_millis() {
2838     assert(_wallclock.is_active(), "Wall clock should not stop");
2839     _wallclock.stop();  // to record time
2840     jlong ret = _wallclock.milliseconds();
2841     _wallclock.start(); // restart
2842     return ret;
2843   }
2844 };
2845 
2846 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2847                                        const char *phase,
2848                                        bool print_cr) :
2849   _collector(collector), _phase(phase), _print_cr(print_cr) {
2850 
2851   if (PrintCMSStatistics != 0) {
2852     _collector->resetYields();
2853   }
2854   if (PrintGCDetails) {
2855     gclog_or_tty->gclog_stamp();
2856     gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2857       _collector->cmsGen()->short_name(), _phase);
2858   }
2859   _collector->resetTimer();
2860   _wallclock.start();
2861   _collector->startTimer();
2862 }
2863 
2864 CMSPhaseAccounting::~CMSPhaseAccounting() {
2865   assert(_wallclock.is_active(), "Wall clock should not have stopped");
2866   _collector->stopTimer();
2867   _wallclock.stop();
2868   if (PrintGCDetails) {
2869     gclog_or_tty->gclog_stamp();
2870     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2871                  _collector->cmsGen()->short_name(),
2872                  _phase, _collector->timerValue(), _wallclock.seconds());
2873     if (_print_cr) {
2874       gclog_or_tty->cr();
2875     }
2876     if (PrintCMSStatistics != 0) {
2877       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2878                     _collector->yields());
2879     }
2880   }
2881 }
2882 
2883 // CMS work
2884 
2885 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2886 class CMSParMarkTask : public AbstractGangTask {
2887  protected:
2888   CMSCollector*     _collector;
2889   uint              _n_workers;
2890   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2891       AbstractGangTask(name),
2892       _collector(collector),
2893       _n_workers(n_workers) {}
2894   // Work method in support of parallel rescan ... of young gen spaces
2895   void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2896                              ContiguousSpace* space,
2897                              HeapWord** chunk_array, size_t chunk_top);
2898   void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2899 };
2900 
2901 // Parallel initial mark task
2902 class CMSParInitialMarkTask: public CMSParMarkTask {
2903   StrongRootsScope* _strong_roots_scope;
2904  public:
2905   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2906       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2907       _strong_roots_scope(strong_roots_scope) {}
2908   void work(uint worker_id);
2909 };
2910 
2911 // Checkpoint the roots into this generation from outside
2912 // this generation. [Note this initial checkpoint need only
2913 // be approximate -- we'll do a catch up phase subsequently.]
2914 void CMSCollector::checkpointRootsInitial() {
2915   assert(_collectorState == InitialMarking, "Wrong collector state");
2916   check_correct_thread_executing();
2917   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2918 
2919   save_heap_summary();
2920   report_heap_summary(GCWhen::BeforeGC);
2921 
2922   ReferenceProcessor* rp = ref_processor();
2923   assert(_restart_addr == NULL, "Control point invariant");
2924   {
2925     // acquire locks for subsequent manipulations
2926     MutexLockerEx x(bitMapLock(),
2927                     Mutex::_no_safepoint_check_flag);
2928     checkpointRootsInitialWork();
2929     // enable ("weak") refs discovery
2930     rp->enable_discovery();
2931     _collectorState = Marking;
2932   }
2933 }
2934 
2935 void CMSCollector::checkpointRootsInitialWork() {
2936   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2937   assert(_collectorState == InitialMarking, "just checking");
2938 
2939   // Already have locks.
2940   assert_lock_strong(bitMapLock());
2941   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2942 
2943   // Setup the verification and class unloading state for this
2944   // CMS collection cycle.
2945   setup_cms_unloading_and_verification_state();
2946 
2947   NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2948     PrintGCDetails && Verbose, true, _gc_timer_cm);)
2949 
2950   // Reset all the PLAB chunk arrays if necessary.
2951   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2952     reset_survivor_plab_arrays();
2953   }
2954 
2955   ResourceMark rm;
2956   HandleMark  hm;
2957 
2958   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2959   GenCollectedHeap* gch = GenCollectedHeap::heap();
2960 
2961   verify_work_stacks_empty();
2962   verify_overflow_empty();
2963 
2964   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2965   // Update the saved marks which may affect the root scans.
2966   gch->save_marks();
2967 
2968   // weak reference processing has not started yet.
2969   ref_processor()->set_enqueuing_is_done(false);
2970 
2971   // Need to remember all newly created CLDs,
2972   // so that we can guarantee that the remark finds them.
2973   ClassLoaderDataGraph::remember_new_clds(true);
2974 
2975   // Whenever a CLD is found, it will be claimed before proceeding to mark
2976   // the klasses. The claimed marks need to be cleared before marking starts.
2977   ClassLoaderDataGraph::clear_claimed_marks();
2978 
2979   if (CMSPrintEdenSurvivorChunks) {
2980     print_eden_and_survivor_chunk_arrays();
2981   }
2982 
2983   {
2984 #if defined(COMPILER2) || INCLUDE_JVMCI
2985     DerivedPointerTableDeactivate dpt_deact;
2986 #endif
2987     if (CMSParallelInitialMarkEnabled) {
2988       // The parallel version.
2989       WorkGang* workers = gch->workers();
2990       assert(workers != NULL, "Need parallel worker threads.");
2991       uint n_workers = workers->active_workers();
2992 
2993       StrongRootsScope srs(n_workers);
2994 
2995       CMSParInitialMarkTask tsk(this, &srs, n_workers);
2996       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2997       if (n_workers > 1) {
2998         workers->run_task(&tsk);
2999       } else {
3000         tsk.work(0);
3001       }
3002     } else {
3003       // The serial version.
3004       CLDToOopClosure cld_closure(&notOlder, true);
3005       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3006 
3007       StrongRootsScope srs(1);
3008 
3009       gch->old_process_roots(&srs,
3010                              true,   // young gen as roots
3011                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
3012                              should_unload_classes(),
3013                              &notOlder,
3014                              &cld_closure);
3015     }
3016   }
3017 
3018   // Clear mod-union table; it will be dirtied in the prologue of
3019   // CMS generation per each young generation collection.
3020 
3021   assert(_modUnionTable.isAllClear(),
3022        "Was cleared in most recent final checkpoint phase"
3023        " or no bits are set in the gc_prologue before the start of the next "
3024        "subsequent marking phase.");
3025 
3026   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3027 
3028   // Save the end of the used_region of the constituent generations
3029   // to be used to limit the extent of sweep in each generation.
3030   save_sweep_limits();
3031   verify_overflow_empty();
3032 }
3033 
3034 bool CMSCollector::markFromRoots() {
3035   // we might be tempted to assert that:
3036   // assert(!SafepointSynchronize::is_at_safepoint(),
3037   //        "inconsistent argument?");
3038   // However that wouldn't be right, because it's possible that
3039   // a safepoint is indeed in progress as a young generation
3040   // stop-the-world GC happens even as we mark in this generation.
3041   assert(_collectorState == Marking, "inconsistent state?");
3042   check_correct_thread_executing();
3043   verify_overflow_empty();
3044 
3045   // Weak ref discovery note: We may be discovering weak
3046   // refs in this generation concurrent (but interleaved) with
3047   // weak ref discovery by the young generation collector.
3048 
3049   CMSTokenSyncWithLocks ts(true, bitMapLock());
3050   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3051   CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3052   bool res = markFromRootsWork();
3053   if (res) {
3054     _collectorState = Precleaning;
3055   } else { // We failed and a foreground collection wants to take over
3056     assert(_foregroundGCIsActive, "internal state inconsistency");
3057     assert(_restart_addr == NULL,  "foreground will restart from scratch");
3058     if (PrintGCDetails) {
3059       gclog_or_tty->print_cr("bailing out to foreground collection");
3060     }
3061   }
3062   verify_overflow_empty();
3063   return res;
3064 }
3065 
3066 bool CMSCollector::markFromRootsWork() {
3067   // iterate over marked bits in bit map, doing a full scan and mark
3068   // from these roots using the following algorithm:
3069   // . if oop is to the right of the current scan pointer,
3070   //   mark corresponding bit (we'll process it later)
3071   // . else (oop is to left of current scan pointer)
3072   //   push oop on marking stack
3073   // . drain the marking stack
3074 
3075   // Note that when we do a marking step we need to hold the
3076   // bit map lock -- recall that direct allocation (by mutators)
3077   // and promotion (by the young generation collector) is also
3078   // marking the bit map. [the so-called allocate live policy.]
3079   // Because the implementation of bit map marking is not
3080   // robust wrt simultaneous marking of bits in the same word,
3081   // we need to make sure that there is no such interference
3082   // between concurrent such updates.
3083 
3084   // already have locks
3085   assert_lock_strong(bitMapLock());
3086 
3087   verify_work_stacks_empty();
3088   verify_overflow_empty();
3089   bool result = false;
3090   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3091     result = do_marking_mt();
3092   } else {
3093     result = do_marking_st();
3094   }
3095   return result;
3096 }
3097 
3098 // Forward decl
3099 class CMSConcMarkingTask;
3100 
3101 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3102   CMSCollector*       _collector;
3103   CMSConcMarkingTask* _task;
3104  public:
3105   virtual void yield();
3106 
3107   // "n_threads" is the number of threads to be terminated.
3108   // "queue_set" is a set of work queues of other threads.
3109   // "collector" is the CMS collector associated with this task terminator.
3110   // "yield" indicates whether we need the gang as a whole to yield.
3111   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3112     ParallelTaskTerminator(n_threads, queue_set),
3113     _collector(collector) { }
3114 
3115   void set_task(CMSConcMarkingTask* task) {
3116     _task = task;
3117   }
3118 };
3119 
3120 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3121   CMSConcMarkingTask* _task;
3122  public:
3123   bool should_exit_termination();
3124   void set_task(CMSConcMarkingTask* task) {
3125     _task = task;
3126   }
3127 };
3128 
3129 // MT Concurrent Marking Task
3130 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3131   CMSCollector* _collector;
3132   uint          _n_workers;       // requested/desired # workers
3133   bool          _result;
3134   CompactibleFreeListSpace*  _cms_space;
3135   char          _pad_front[64];   // padding to ...
3136   HeapWord*     _global_finger;   // ... avoid sharing cache line
3137   char          _pad_back[64];
3138   HeapWord*     _restart_addr;
3139 
3140   //  Exposed here for yielding support
3141   Mutex* const _bit_map_lock;
3142 
3143   // The per thread work queues, available here for stealing
3144   OopTaskQueueSet*  _task_queues;
3145 
3146   // Termination (and yielding) support
3147   CMSConcMarkingTerminator _term;
3148   CMSConcMarkingTerminatorTerminator _term_term;
3149 
3150  public:
3151   CMSConcMarkingTask(CMSCollector* collector,
3152                  CompactibleFreeListSpace* cms_space,
3153                  YieldingFlexibleWorkGang* workers,
3154                  OopTaskQueueSet* task_queues):
3155     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3156     _collector(collector),
3157     _cms_space(cms_space),
3158     _n_workers(0), _result(true),
3159     _task_queues(task_queues),
3160     _term(_n_workers, task_queues, _collector),
3161     _bit_map_lock(collector->bitMapLock())
3162   {
3163     _requested_size = _n_workers;
3164     _term.set_task(this);
3165     _term_term.set_task(this);
3166     _restart_addr = _global_finger = _cms_space->bottom();
3167   }
3168 
3169 
3170   OopTaskQueueSet* task_queues()  { return _task_queues; }
3171 
3172   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3173 
3174   HeapWord** global_finger_addr() { return &_global_finger; }
3175 
3176   CMSConcMarkingTerminator* terminator() { return &_term; }
3177 
3178   virtual void set_for_termination(uint active_workers) {
3179     terminator()->reset_for_reuse(active_workers);
3180   }
3181 
3182   void work(uint worker_id);
3183   bool should_yield() {
3184     return    ConcurrentMarkSweepThread::should_yield()
3185            && !_collector->foregroundGCIsActive();
3186   }
3187 
3188   virtual void coordinator_yield();  // stuff done by coordinator
3189   bool result() { return _result; }
3190 
3191   void reset(HeapWord* ra) {
3192     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3193     _restart_addr = _global_finger = ra;
3194     _term.reset_for_reuse();
3195   }
3196 
3197   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3198                                            OopTaskQueue* work_q);
3199 
3200  private:
3201   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3202   void do_work_steal(int i);
3203   void bump_global_finger(HeapWord* f);
3204 };
3205 
3206 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3207   assert(_task != NULL, "Error");
3208   return _task->yielding();
3209   // Note that we do not need the disjunct || _task->should_yield() above
3210   // because we want terminating threads to yield only if the task
3211   // is already in the midst of yielding, which happens only after at least one
3212   // thread has yielded.
3213 }
3214 
3215 void CMSConcMarkingTerminator::yield() {
3216   if (_task->should_yield()) {
3217     _task->yield();
3218   } else {
3219     ParallelTaskTerminator::yield();
3220   }
3221 }
3222 
3223 ////////////////////////////////////////////////////////////////
3224 // Concurrent Marking Algorithm Sketch
3225 ////////////////////////////////////////////////////////////////
3226 // Until all tasks exhausted (both spaces):
3227 // -- claim next available chunk
3228 // -- bump global finger via CAS
3229 // -- find first object that starts in this chunk
3230 //    and start scanning bitmap from that position
3231 // -- scan marked objects for oops
3232 // -- CAS-mark target, and if successful:
3233 //    . if target oop is above global finger (volatile read)
3234 //      nothing to do
3235 //    . if target oop is in chunk and above local finger
3236 //        then nothing to do
3237 //    . else push on work-queue
3238 // -- Deal with possible overflow issues:
3239 //    . local work-queue overflow causes stuff to be pushed on
3240 //      global (common) overflow queue
3241 //    . always first empty local work queue
3242 //    . then get a batch of oops from global work queue if any
3243 //    . then do work stealing
3244 // -- When all tasks claimed (both spaces)
3245 //    and local work queue empty,
3246 //    then in a loop do:
3247 //    . check global overflow stack; steal a batch of oops and trace
3248 //    . try to steal from other threads oif GOS is empty
3249 //    . if neither is available, offer termination
3250 // -- Terminate and return result
3251 //
3252 void CMSConcMarkingTask::work(uint worker_id) {
3253   elapsedTimer _timer;
3254   ResourceMark rm;
3255   HandleMark hm;
3256 
3257   DEBUG_ONLY(_collector->verify_overflow_empty();)
3258 
3259   // Before we begin work, our work queue should be empty
3260   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3261   // Scan the bitmap covering _cms_space, tracing through grey objects.
3262   _timer.start();
3263   do_scan_and_mark(worker_id, _cms_space);
3264   _timer.stop();
3265   if (PrintCMSStatistics != 0) {
3266     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3267       worker_id, _timer.seconds());
3268       // XXX: need xxx/xxx type of notation, two timers
3269   }
3270 
3271   // ... do work stealing
3272   _timer.reset();
3273   _timer.start();
3274   do_work_steal(worker_id);
3275   _timer.stop();
3276   if (PrintCMSStatistics != 0) {
3277     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3278       worker_id, _timer.seconds());
3279       // XXX: need xxx/xxx type of notation, two timers
3280   }
3281   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3282   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3283   // Note that under the current task protocol, the
3284   // following assertion is true even of the spaces
3285   // expanded since the completion of the concurrent
3286   // marking. XXX This will likely change under a strict
3287   // ABORT semantics.
3288   // After perm removal the comparison was changed to
3289   // greater than or equal to from strictly greater than.
3290   // Before perm removal the highest address sweep would
3291   // have been at the end of perm gen but now is at the
3292   // end of the tenured gen.
3293   assert(_global_finger >=  _cms_space->end(),
3294          "All tasks have been completed");
3295   DEBUG_ONLY(_collector->verify_overflow_empty();)
3296 }
3297 
3298 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3299   HeapWord* read = _global_finger;
3300   HeapWord* cur  = read;
3301   while (f > read) {
3302     cur = read;
3303     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3304     if (cur == read) {
3305       // our cas succeeded
3306       assert(_global_finger >= f, "protocol consistency");
3307       break;
3308     }
3309   }
3310 }
3311 
3312 // This is really inefficient, and should be redone by
3313 // using (not yet available) block-read and -write interfaces to the
3314 // stack and the work_queue. XXX FIX ME !!!
3315 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3316                                                       OopTaskQueue* work_q) {
3317   // Fast lock-free check
3318   if (ovflw_stk->length() == 0) {
3319     return false;
3320   }
3321   assert(work_q->size() == 0, "Shouldn't steal");
3322   MutexLockerEx ml(ovflw_stk->par_lock(),
3323                    Mutex::_no_safepoint_check_flag);
3324   // Grab up to 1/4 the size of the work queue
3325   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3326                     (size_t)ParGCDesiredObjsFromOverflowList);
3327   num = MIN2(num, ovflw_stk->length());
3328   for (int i = (int) num; i > 0; i--) {
3329     oop cur = ovflw_stk->pop();
3330     assert(cur != NULL, "Counted wrong?");
3331     work_q->push(cur);
3332   }
3333   return num > 0;
3334 }
3335 
3336 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3337   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3338   int n_tasks = pst->n_tasks();
3339   // We allow that there may be no tasks to do here because
3340   // we are restarting after a stack overflow.
3341   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3342   uint nth_task = 0;
3343 
3344   HeapWord* aligned_start = sp->bottom();
3345   if (sp->used_region().contains(_restart_addr)) {
3346     // Align down to a card boundary for the start of 0th task
3347     // for this space.
3348     aligned_start =
3349       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3350                                  CardTableModRefBS::card_size);
3351   }
3352 
3353   size_t chunk_size = sp->marking_task_size();
3354   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3355     // Having claimed the nth task in this space,
3356     // compute the chunk that it corresponds to:
3357     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3358                                aligned_start + (nth_task+1)*chunk_size);
3359     // Try and bump the global finger via a CAS;
3360     // note that we need to do the global finger bump
3361     // _before_ taking the intersection below, because
3362     // the task corresponding to that region will be
3363     // deemed done even if the used_region() expands
3364     // because of allocation -- as it almost certainly will
3365     // during start-up while the threads yield in the
3366     // closure below.
3367     HeapWord* finger = span.end();
3368     bump_global_finger(finger);   // atomically
3369     // There are null tasks here corresponding to chunks
3370     // beyond the "top" address of the space.
3371     span = span.intersection(sp->used_region());
3372     if (!span.is_empty()) {  // Non-null task
3373       HeapWord* prev_obj;
3374       assert(!span.contains(_restart_addr) || nth_task == 0,
3375              "Inconsistency");
3376       if (nth_task == 0) {
3377         // For the 0th task, we'll not need to compute a block_start.
3378         if (span.contains(_restart_addr)) {
3379           // In the case of a restart because of stack overflow,
3380           // we might additionally skip a chunk prefix.
3381           prev_obj = _restart_addr;
3382         } else {
3383           prev_obj = span.start();
3384         }
3385       } else {
3386         // We want to skip the first object because
3387         // the protocol is to scan any object in its entirety
3388         // that _starts_ in this span; a fortiori, any
3389         // object starting in an earlier span is scanned
3390         // as part of an earlier claimed task.
3391         // Below we use the "careful" version of block_start
3392         // so we do not try to navigate uninitialized objects.
3393         prev_obj = sp->block_start_careful(span.start());
3394         // Below we use a variant of block_size that uses the
3395         // Printezis bits to avoid waiting for allocated
3396         // objects to become initialized/parsable.
3397         while (prev_obj < span.start()) {
3398           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3399           if (sz > 0) {
3400             prev_obj += sz;
3401           } else {
3402             // In this case we may end up doing a bit of redundant
3403             // scanning, but that appears unavoidable, short of
3404             // locking the free list locks; see bug 6324141.
3405             break;
3406           }
3407         }
3408       }
3409       if (prev_obj < span.end()) {
3410         MemRegion my_span = MemRegion(prev_obj, span.end());
3411         // Do the marking work within a non-empty span --
3412         // the last argument to the constructor indicates whether the
3413         // iteration should be incremental with periodic yields.
3414         Par_MarkFromRootsClosure cl(this, _collector, my_span,
3415                                     &_collector->_markBitMap,
3416                                     work_queue(i),
3417                                     &_collector->_markStack);
3418         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3419       } // else nothing to do for this task
3420     }   // else nothing to do for this task
3421   }
3422   // We'd be tempted to assert here that since there are no
3423   // more tasks left to claim in this space, the global_finger
3424   // must exceed space->top() and a fortiori space->end(). However,
3425   // that would not quite be correct because the bumping of
3426   // global_finger occurs strictly after the claiming of a task,
3427   // so by the time we reach here the global finger may not yet
3428   // have been bumped up by the thread that claimed the last
3429   // task.
3430   pst->all_tasks_completed();
3431 }
3432 
3433 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3434  private:
3435   CMSCollector* _collector;
3436   CMSConcMarkingTask* _task;
3437   MemRegion     _span;
3438   CMSBitMap*    _bit_map;
3439   CMSMarkStack* _overflow_stack;
3440   OopTaskQueue* _work_queue;
3441  protected:
3442   DO_OOP_WORK_DEFN
3443  public:
3444   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3445                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3446     MetadataAwareOopClosure(collector->ref_processor()),
3447     _collector(collector),
3448     _task(task),
3449     _span(collector->_span),
3450     _work_queue(work_queue),
3451     _bit_map(bit_map),
3452     _overflow_stack(overflow_stack)
3453   { }
3454   virtual void do_oop(oop* p);
3455   virtual void do_oop(narrowOop* p);
3456 
3457   void trim_queue(size_t max);
3458   void handle_stack_overflow(HeapWord* lost);
3459   void do_yield_check() {
3460     if (_task->should_yield()) {
3461       _task->yield();
3462     }
3463   }
3464 };
3465 
3466 // Grey object scanning during work stealing phase --
3467 // the salient assumption here is that any references
3468 // that are in these stolen objects being scanned must
3469 // already have been initialized (else they would not have
3470 // been published), so we do not need to check for
3471 // uninitialized objects before pushing here.
3472 void Par_ConcMarkingClosure::do_oop(oop obj) {
3473   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3474   HeapWord* addr = (HeapWord*)obj;
3475   // Check if oop points into the CMS generation
3476   // and is not marked
3477   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3478     // a white object ...
3479     // If we manage to "claim" the object, by being the
3480     // first thread to mark it, then we push it on our
3481     // marking stack
3482     if (_bit_map->par_mark(addr)) {     // ... now grey
3483       // push on work queue (grey set)
3484       bool simulate_overflow = false;
3485       NOT_PRODUCT(
3486         if (CMSMarkStackOverflowALot &&
3487             _collector->simulate_overflow()) {
3488           // simulate a stack overflow
3489           simulate_overflow = true;
3490         }
3491       )
3492       if (simulate_overflow ||
3493           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3494         // stack overflow
3495         if (PrintCMSStatistics != 0) {
3496           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3497                                  SIZE_FORMAT, _overflow_stack->capacity());
3498         }
3499         // We cannot assert that the overflow stack is full because
3500         // it may have been emptied since.
3501         assert(simulate_overflow ||
3502                _work_queue->size() == _work_queue->max_elems(),
3503               "Else push should have succeeded");
3504         handle_stack_overflow(addr);
3505       }
3506     } // Else, some other thread got there first
3507     do_yield_check();
3508   }
3509 }
3510 
3511 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
3512 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3513 
3514 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3515   while (_work_queue->size() > max) {
3516     oop new_oop;
3517     if (_work_queue->pop_local(new_oop)) {
3518       assert(new_oop->is_oop(), "Should be an oop");
3519       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3520       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3521       new_oop->oop_iterate(this);  // do_oop() above
3522       do_yield_check();
3523     }
3524   }
3525 }
3526 
3527 // Upon stack overflow, we discard (part of) the stack,
3528 // remembering the least address amongst those discarded
3529 // in CMSCollector's _restart_address.
3530 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3531   // We need to do this under a mutex to prevent other
3532   // workers from interfering with the work done below.
3533   MutexLockerEx ml(_overflow_stack->par_lock(),
3534                    Mutex::_no_safepoint_check_flag);
3535   // Remember the least grey address discarded
3536   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3537   _collector->lower_restart_addr(ra);
3538   _overflow_stack->reset();  // discard stack contents
3539   _overflow_stack->expand(); // expand the stack if possible
3540 }
3541 
3542 
3543 void CMSConcMarkingTask::do_work_steal(int i) {
3544   OopTaskQueue* work_q = work_queue(i);
3545   oop obj_to_scan;
3546   CMSBitMap* bm = &(_collector->_markBitMap);
3547   CMSMarkStack* ovflw = &(_collector->_markStack);
3548   int* seed = _collector->hash_seed(i);
3549   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3550   while (true) {
3551     cl.trim_queue(0);
3552     assert(work_q->size() == 0, "Should have been emptied above");
3553     if (get_work_from_overflow_stack(ovflw, work_q)) {
3554       // Can't assert below because the work obtained from the
3555       // overflow stack may already have been stolen from us.
3556       // assert(work_q->size() > 0, "Work from overflow stack");
3557       continue;
3558     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3559       assert(obj_to_scan->is_oop(), "Should be an oop");
3560       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3561       obj_to_scan->oop_iterate(&cl);
3562     } else if (terminator()->offer_termination(&_term_term)) {
3563       assert(work_q->size() == 0, "Impossible!");
3564       break;
3565     } else if (yielding() || should_yield()) {
3566       yield();
3567     }
3568   }
3569 }
3570 
3571 // This is run by the CMS (coordinator) thread.
3572 void CMSConcMarkingTask::coordinator_yield() {
3573   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3574          "CMS thread should hold CMS token");
3575   // First give up the locks, then yield, then re-lock
3576   // We should probably use a constructor/destructor idiom to
3577   // do this unlock/lock or modify the MutexUnlocker class to
3578   // serve our purpose. XXX
3579   assert_lock_strong(_bit_map_lock);
3580   _bit_map_lock->unlock();
3581   ConcurrentMarkSweepThread::desynchronize(true);
3582   _collector->stopTimer();
3583   if (PrintCMSStatistics != 0) {
3584     _collector->incrementYields();
3585   }
3586 
3587   // It is possible for whichever thread initiated the yield request
3588   // not to get a chance to wake up and take the bitmap lock between
3589   // this thread releasing it and reacquiring it. So, while the
3590   // should_yield() flag is on, let's sleep for a bit to give the
3591   // other thread a chance to wake up. The limit imposed on the number
3592   // of iterations is defensive, to avoid any unforseen circumstances
3593   // putting us into an infinite loop. Since it's always been this
3594   // (coordinator_yield()) method that was observed to cause the
3595   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3596   // which is by default non-zero. For the other seven methods that
3597   // also perform the yield operation, as are using a different
3598   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3599   // can enable the sleeping for those methods too, if necessary.
3600   // See 6442774.
3601   //
3602   // We really need to reconsider the synchronization between the GC
3603   // thread and the yield-requesting threads in the future and we
3604   // should really use wait/notify, which is the recommended
3605   // way of doing this type of interaction. Additionally, we should
3606   // consolidate the eight methods that do the yield operation and they
3607   // are almost identical into one for better maintainability and
3608   // readability. See 6445193.
3609   //
3610   // Tony 2006.06.29
3611   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3612                    ConcurrentMarkSweepThread::should_yield() &&
3613                    !CMSCollector::foregroundGCIsActive(); ++i) {
3614     os::sleep(Thread::current(), 1, false);
3615   }
3616 
3617   ConcurrentMarkSweepThread::synchronize(true);
3618   _bit_map_lock->lock_without_safepoint_check();
3619   _collector->startTimer();
3620 }
3621 
3622 bool CMSCollector::do_marking_mt() {
3623   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3624   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3625                                                                   conc_workers()->active_workers(),
3626                                                                   Threads::number_of_non_daemon_threads());
3627   conc_workers()->set_active_workers(num_workers);
3628 
3629   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3630 
3631   CMSConcMarkingTask tsk(this,
3632                          cms_space,
3633                          conc_workers(),
3634                          task_queues());
3635 
3636   // Since the actual number of workers we get may be different
3637   // from the number we requested above, do we need to do anything different
3638   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3639   // class?? XXX
3640   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3641 
3642   // Refs discovery is already non-atomic.
3643   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3644   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3645   conc_workers()->start_task(&tsk);
3646   while (tsk.yielded()) {
3647     tsk.coordinator_yield();
3648     conc_workers()->continue_task(&tsk);
3649   }
3650   // If the task was aborted, _restart_addr will be non-NULL
3651   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3652   while (_restart_addr != NULL) {
3653     // XXX For now we do not make use of ABORTED state and have not
3654     // yet implemented the right abort semantics (even in the original
3655     // single-threaded CMS case). That needs some more investigation
3656     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3657     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3658     // If _restart_addr is non-NULL, a marking stack overflow
3659     // occurred; we need to do a fresh marking iteration from the
3660     // indicated restart address.
3661     if (_foregroundGCIsActive) {
3662       // We may be running into repeated stack overflows, having
3663       // reached the limit of the stack size, while making very
3664       // slow forward progress. It may be best to bail out and
3665       // let the foreground collector do its job.
3666       // Clear _restart_addr, so that foreground GC
3667       // works from scratch. This avoids the headache of
3668       // a "rescan" which would otherwise be needed because
3669       // of the dirty mod union table & card table.
3670       _restart_addr = NULL;
3671       return false;
3672     }
3673     // Adjust the task to restart from _restart_addr
3674     tsk.reset(_restart_addr);
3675     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3676                   _restart_addr);
3677     _restart_addr = NULL;
3678     // Get the workers going again
3679     conc_workers()->start_task(&tsk);
3680     while (tsk.yielded()) {
3681       tsk.coordinator_yield();
3682       conc_workers()->continue_task(&tsk);
3683     }
3684   }
3685   assert(tsk.completed(), "Inconsistency");
3686   assert(tsk.result() == true, "Inconsistency");
3687   return true;
3688 }
3689 
3690 bool CMSCollector::do_marking_st() {
3691   ResourceMark rm;
3692   HandleMark   hm;
3693 
3694   // Temporarily make refs discovery single threaded (non-MT)
3695   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3696   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3697     &_markStack, CMSYield);
3698   // the last argument to iterate indicates whether the iteration
3699   // should be incremental with periodic yields.
3700   _markBitMap.iterate(&markFromRootsClosure);
3701   // If _restart_addr is non-NULL, a marking stack overflow
3702   // occurred; we need to do a fresh iteration from the
3703   // indicated restart address.
3704   while (_restart_addr != NULL) {
3705     if (_foregroundGCIsActive) {
3706       // We may be running into repeated stack overflows, having
3707       // reached the limit of the stack size, while making very
3708       // slow forward progress. It may be best to bail out and
3709       // let the foreground collector do its job.
3710       // Clear _restart_addr, so that foreground GC
3711       // works from scratch. This avoids the headache of
3712       // a "rescan" which would otherwise be needed because
3713       // of the dirty mod union table & card table.
3714       _restart_addr = NULL;
3715       return false;  // indicating failure to complete marking
3716     }
3717     // Deal with stack overflow:
3718     // we restart marking from _restart_addr
3719     HeapWord* ra = _restart_addr;
3720     markFromRootsClosure.reset(ra);
3721     _restart_addr = NULL;
3722     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3723   }
3724   return true;
3725 }
3726 
3727 void CMSCollector::preclean() {
3728   check_correct_thread_executing();
3729   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3730   verify_work_stacks_empty();
3731   verify_overflow_empty();
3732   _abort_preclean = false;
3733   if (CMSPrecleaningEnabled) {
3734     if (!CMSEdenChunksRecordAlways) {
3735       _eden_chunk_index = 0;
3736     }
3737     size_t used = get_eden_used();
3738     size_t capacity = get_eden_capacity();
3739     // Don't start sampling unless we will get sufficiently
3740     // many samples.
3741     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3742                 * CMSScheduleRemarkEdenPenetration)) {
3743       _start_sampling = true;
3744     } else {
3745       _start_sampling = false;
3746     }
3747     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3748     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
3749     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3750   }
3751   CMSTokenSync x(true); // is cms thread
3752   if (CMSPrecleaningEnabled) {
3753     sample_eden();
3754     _collectorState = AbortablePreclean;
3755   } else {
3756     _collectorState = FinalMarking;
3757   }
3758   verify_work_stacks_empty();
3759   verify_overflow_empty();
3760 }
3761 
3762 // Try and schedule the remark such that young gen
3763 // occupancy is CMSScheduleRemarkEdenPenetration %.
3764 void CMSCollector::abortable_preclean() {
3765   check_correct_thread_executing();
3766   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3767   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3768 
3769   // If Eden's current occupancy is below this threshold,
3770   // immediately schedule the remark; else preclean
3771   // past the next scavenge in an effort to
3772   // schedule the pause as described above. By choosing
3773   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3774   // we will never do an actual abortable preclean cycle.
3775   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3776     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3777     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
3778     // We need more smarts in the abortable preclean
3779     // loop below to deal with cases where allocation
3780     // in young gen is very very slow, and our precleaning
3781     // is running a losing race against a horde of
3782     // mutators intent on flooding us with CMS updates
3783     // (dirty cards).
3784     // One, admittedly dumb, strategy is to give up
3785     // after a certain number of abortable precleaning loops
3786     // or after a certain maximum time. We want to make
3787     // this smarter in the next iteration.
3788     // XXX FIX ME!!! YSR
3789     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3790     while (!(should_abort_preclean() ||
3791              ConcurrentMarkSweepThread::should_terminate())) {
3792       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3793       cumworkdone += workdone;
3794       loops++;
3795       // Voluntarily terminate abortable preclean phase if we have
3796       // been at it for too long.
3797       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3798           loops >= CMSMaxAbortablePrecleanLoops) {
3799         if (PrintGCDetails) {
3800           gclog_or_tty->print(" CMS: abort preclean due to loops ");
3801         }
3802         break;
3803       }
3804       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3805         if (PrintGCDetails) {
3806           gclog_or_tty->print(" CMS: abort preclean due to time ");
3807         }
3808         break;
3809       }
3810       // If we are doing little work each iteration, we should
3811       // take a short break.
3812       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3813         // Sleep for some time, waiting for work to accumulate
3814         stopTimer();
3815         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3816         startTimer();
3817         waited++;
3818       }
3819     }
3820     if (PrintCMSStatistics > 0) {
3821       gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3822                           loops, waited, cumworkdone);
3823     }
3824   }
3825   CMSTokenSync x(true); // is cms thread
3826   if (_collectorState != Idling) {
3827     assert(_collectorState == AbortablePreclean,
3828            "Spontaneous state transition?");
3829     _collectorState = FinalMarking;
3830   } // Else, a foreground collection completed this CMS cycle.
3831   return;
3832 }
3833 
3834 // Respond to an Eden sampling opportunity
3835 void CMSCollector::sample_eden() {
3836   // Make sure a young gc cannot sneak in between our
3837   // reading and recording of a sample.
3838   assert(Thread::current()->is_ConcurrentGC_thread(),
3839          "Only the cms thread may collect Eden samples");
3840   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3841          "Should collect samples while holding CMS token");
3842   if (!_start_sampling) {
3843     return;
3844   }
3845   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3846   // is populated by the young generation.
3847   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3848     if (_eden_chunk_index < _eden_chunk_capacity) {
3849       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3850       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3851              "Unexpected state of Eden");
3852       // We'd like to check that what we just sampled is an oop-start address;
3853       // however, we cannot do that here since the object may not yet have been
3854       // initialized. So we'll instead do the check when we _use_ this sample
3855       // later.
3856       if (_eden_chunk_index == 0 ||
3857           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3858                          _eden_chunk_array[_eden_chunk_index-1])
3859            >= CMSSamplingGrain)) {
3860         _eden_chunk_index++;  // commit sample
3861       }
3862     }
3863   }
3864   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3865     size_t used = get_eden_used();
3866     size_t capacity = get_eden_capacity();
3867     assert(used <= capacity, "Unexpected state of Eden");
3868     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3869       _abort_preclean = true;
3870     }
3871   }
3872 }
3873 
3874 
3875 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3876   assert(_collectorState == Precleaning ||
3877          _collectorState == AbortablePreclean, "incorrect state");
3878   ResourceMark rm;
3879   HandleMark   hm;
3880 
3881   // Precleaning is currently not MT but the reference processor
3882   // may be set for MT.  Disable it temporarily here.
3883   ReferenceProcessor* rp = ref_processor();
3884   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3885 
3886   // Do one pass of scrubbing the discovered reference lists
3887   // to remove any reference objects with strongly-reachable
3888   // referents.
3889   if (clean_refs) {
3890     CMSPrecleanRefsYieldClosure yield_cl(this);
3891     assert(rp->span().equals(_span), "Spans should be equal");
3892     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3893                                    &_markStack, true /* preclean */);
3894     CMSDrainMarkingStackClosure complete_trace(this,
3895                                    _span, &_markBitMap, &_markStack,
3896                                    &keep_alive, true /* preclean */);
3897 
3898     // We don't want this step to interfere with a young
3899     // collection because we don't want to take CPU
3900     // or memory bandwidth away from the young GC threads
3901     // (which may be as many as there are CPUs).
3902     // Note that we don't need to protect ourselves from
3903     // interference with mutators because they can't
3904     // manipulate the discovered reference lists nor affect
3905     // the computed reachability of the referents, the
3906     // only properties manipulated by the precleaning
3907     // of these reference lists.
3908     stopTimer();
3909     CMSTokenSyncWithLocks x(true /* is cms thread */,
3910                             bitMapLock());
3911     startTimer();
3912     sample_eden();
3913 
3914     // The following will yield to allow foreground
3915     // collection to proceed promptly. XXX YSR:
3916     // The code in this method may need further
3917     // tweaking for better performance and some restructuring
3918     // for cleaner interfaces.
3919     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3920     rp->preclean_discovered_references(
3921           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3922           gc_timer);
3923   }
3924 
3925   if (clean_survivor) {  // preclean the active survivor space(s)
3926     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3927                              &_markBitMap, &_modUnionTable,
3928                              &_markStack, true /* precleaning phase */);
3929     stopTimer();
3930     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3931                              bitMapLock());
3932     startTimer();
3933     unsigned int before_count =
3934       GenCollectedHeap::heap()->total_collections();
3935     SurvivorSpacePrecleanClosure
3936       sss_cl(this, _span, &_markBitMap, &_markStack,
3937              &pam_cl, before_count, CMSYield);
3938     _young_gen->from()->object_iterate_careful(&sss_cl);
3939     _young_gen->to()->object_iterate_careful(&sss_cl);
3940   }
3941   MarkRefsIntoAndScanClosure
3942     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3943              &_markStack, this, CMSYield,
3944              true /* precleaning phase */);
3945   // CAUTION: The following closure has persistent state that may need to
3946   // be reset upon a decrease in the sequence of addresses it
3947   // processes.
3948   ScanMarkedObjectsAgainCarefullyClosure
3949     smoac_cl(this, _span,
3950       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3951 
3952   // Preclean dirty cards in ModUnionTable and CardTable using
3953   // appropriate convergence criterion;
3954   // repeat CMSPrecleanIter times unless we find that
3955   // we are losing.
3956   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3957   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3958          "Bad convergence multiplier");
3959   assert(CMSPrecleanThreshold >= 100,
3960          "Unreasonably low CMSPrecleanThreshold");
3961 
3962   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3963   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3964        numIter < CMSPrecleanIter;
3965        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3966     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3967     if (Verbose && PrintGCDetails) {
3968       gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3969     }
3970     // Either there are very few dirty cards, so re-mark
3971     // pause will be small anyway, or our pre-cleaning isn't
3972     // that much faster than the rate at which cards are being
3973     // dirtied, so we might as well stop and re-mark since
3974     // precleaning won't improve our re-mark time by much.
3975     if (curNumCards <= CMSPrecleanThreshold ||
3976         (numIter > 0 &&
3977          (curNumCards * CMSPrecleanDenominator >
3978          lastNumCards * CMSPrecleanNumerator))) {
3979       numIter++;
3980       cumNumCards += curNumCards;
3981       break;
3982     }
3983   }
3984 
3985   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3986 
3987   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3988   cumNumCards += curNumCards;
3989   if (PrintGCDetails && PrintCMSStatistics != 0) {
3990     gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3991                   curNumCards, cumNumCards, numIter);
3992   }
3993   return cumNumCards;   // as a measure of useful work done
3994 }
3995 
3996 // PRECLEANING NOTES:
3997 // Precleaning involves:
3998 // . reading the bits of the modUnionTable and clearing the set bits.
3999 // . For the cards corresponding to the set bits, we scan the
4000 //   objects on those cards. This means we need the free_list_lock
4001 //   so that we can safely iterate over the CMS space when scanning
4002 //   for oops.
4003 // . When we scan the objects, we'll be both reading and setting
4004 //   marks in the marking bit map, so we'll need the marking bit map.
4005 // . For protecting _collector_state transitions, we take the CGC_lock.
4006 //   Note that any races in the reading of of card table entries by the
4007 //   CMS thread on the one hand and the clearing of those entries by the
4008 //   VM thread or the setting of those entries by the mutator threads on the
4009 //   other are quite benign. However, for efficiency it makes sense to keep
4010 //   the VM thread from racing with the CMS thread while the latter is
4011 //   dirty card info to the modUnionTable. We therefore also use the
4012 //   CGC_lock to protect the reading of the card table and the mod union
4013 //   table by the CM thread.
4014 // . We run concurrently with mutator updates, so scanning
4015 //   needs to be done carefully  -- we should not try to scan
4016 //   potentially uninitialized objects.
4017 //
4018 // Locking strategy: While holding the CGC_lock, we scan over and
4019 // reset a maximal dirty range of the mod union / card tables, then lock
4020 // the free_list_lock and bitmap lock to do a full marking, then
4021 // release these locks; and repeat the cycle. This allows for a
4022 // certain amount of fairness in the sharing of these locks between
4023 // the CMS collector on the one hand, and the VM thread and the
4024 // mutators on the other.
4025 
4026 // NOTE: preclean_mod_union_table() and preclean_card_table()
4027 // further below are largely identical; if you need to modify
4028 // one of these methods, please check the other method too.
4029 
4030 size_t CMSCollector::preclean_mod_union_table(
4031   ConcurrentMarkSweepGeneration* old_gen,
4032   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4033   verify_work_stacks_empty();
4034   verify_overflow_empty();
4035 
4036   // strategy: starting with the first card, accumulate contiguous
4037   // ranges of dirty cards; clear these cards, then scan the region
4038   // covered by these cards.
4039 
4040   // Since all of the MUT is committed ahead, we can just use
4041   // that, in case the generations expand while we are precleaning.
4042   // It might also be fine to just use the committed part of the
4043   // generation, but we might potentially miss cards when the
4044   // generation is rapidly expanding while we are in the midst
4045   // of precleaning.
4046   HeapWord* startAddr = old_gen->reserved().start();
4047   HeapWord* endAddr   = old_gen->reserved().end();
4048 
4049   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
4050 
4051   size_t numDirtyCards, cumNumDirtyCards;
4052   HeapWord *nextAddr, *lastAddr;
4053   for (cumNumDirtyCards = numDirtyCards = 0,
4054        nextAddr = lastAddr = startAddr;
4055        nextAddr < endAddr;
4056        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4057 
4058     ResourceMark rm;
4059     HandleMark   hm;
4060 
4061     MemRegion dirtyRegion;
4062     {
4063       stopTimer();
4064       // Potential yield point
4065       CMSTokenSync ts(true);
4066       startTimer();
4067       sample_eden();
4068       // Get dirty region starting at nextOffset (inclusive),
4069       // simultaneously clearing it.
4070       dirtyRegion =
4071         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4072       assert(dirtyRegion.start() >= nextAddr,
4073              "returned region inconsistent?");
4074     }
4075     // Remember where the next search should begin.
4076     // The returned region (if non-empty) is a right open interval,
4077     // so lastOffset is obtained from the right end of that
4078     // interval.
4079     lastAddr = dirtyRegion.end();
4080     // Should do something more transparent and less hacky XXX
4081     numDirtyCards =
4082       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4083 
4084     // We'll scan the cards in the dirty region (with periodic
4085     // yields for foreground GC as needed).
4086     if (!dirtyRegion.is_empty()) {
4087       assert(numDirtyCards > 0, "consistency check");
4088       HeapWord* stop_point = NULL;
4089       stopTimer();
4090       // Potential yield point
4091       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
4092                                bitMapLock());
4093       startTimer();
4094       {
4095         verify_work_stacks_empty();
4096         verify_overflow_empty();
4097         sample_eden();
4098         stop_point =
4099           old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4100       }
4101       if (stop_point != NULL) {
4102         // The careful iteration stopped early either because it found an
4103         // uninitialized object, or because we were in the midst of an
4104         // "abortable preclean", which should now be aborted. Redirty
4105         // the bits corresponding to the partially-scanned or unscanned
4106         // cards. We'll either restart at the next block boundary or
4107         // abort the preclean.
4108         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4109                "Should only be AbortablePreclean.");
4110         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4111         if (should_abort_preclean()) {
4112           break; // out of preclean loop
4113         } else {
4114           // Compute the next address at which preclean should pick up;
4115           // might need bitMapLock in order to read P-bits.
4116           lastAddr = next_card_start_after_block(stop_point);
4117         }
4118       }
4119     } else {
4120       assert(lastAddr == endAddr, "consistency check");
4121       assert(numDirtyCards == 0, "consistency check");
4122       break;
4123     }
4124   }
4125   verify_work_stacks_empty();
4126   verify_overflow_empty();
4127   return cumNumDirtyCards;
4128 }
4129 
4130 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4131 // below are largely identical; if you need to modify
4132 // one of these methods, please check the other method too.
4133 
4134 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4135   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4136   // strategy: it's similar to precleamModUnionTable above, in that
4137   // we accumulate contiguous ranges of dirty cards, mark these cards
4138   // precleaned, then scan the region covered by these cards.
4139   HeapWord* endAddr   = (HeapWord*)(old_gen->_virtual_space.high());
4140   HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4141 
4142   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
4143 
4144   size_t numDirtyCards, cumNumDirtyCards;
4145   HeapWord *lastAddr, *nextAddr;
4146 
4147   for (cumNumDirtyCards = numDirtyCards = 0,
4148        nextAddr = lastAddr = startAddr;
4149        nextAddr < endAddr;
4150        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4151 
4152     ResourceMark rm;
4153     HandleMark   hm;
4154 
4155     MemRegion dirtyRegion;
4156     {
4157       // See comments in "Precleaning notes" above on why we
4158       // do this locking. XXX Could the locking overheads be
4159       // too high when dirty cards are sparse? [I don't think so.]
4160       stopTimer();
4161       CMSTokenSync x(true); // is cms thread
4162       startTimer();
4163       sample_eden();
4164       // Get and clear dirty region from card table
4165       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4166                                     MemRegion(nextAddr, endAddr),
4167                                     true,
4168                                     CardTableModRefBS::precleaned_card_val());
4169 
4170       assert(dirtyRegion.start() >= nextAddr,
4171              "returned region inconsistent?");
4172     }
4173     lastAddr = dirtyRegion.end();
4174     numDirtyCards =
4175       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4176 
4177     if (!dirtyRegion.is_empty()) {
4178       stopTimer();
4179       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4180       startTimer();
4181       sample_eden();
4182       verify_work_stacks_empty();
4183       verify_overflow_empty();
4184       HeapWord* stop_point =
4185         old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4186       if (stop_point != NULL) {
4187         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4188                "Should only be AbortablePreclean.");
4189         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4190         if (should_abort_preclean()) {
4191           break; // out of preclean loop
4192         } else {
4193           // Compute the next address at which preclean should pick up.
4194           lastAddr = next_card_start_after_block(stop_point);
4195         }
4196       }
4197     } else {
4198       break;
4199     }
4200   }
4201   verify_work_stacks_empty();
4202   verify_overflow_empty();
4203   return cumNumDirtyCards;
4204 }
4205 
4206 class PrecleanKlassClosure : public KlassClosure {
4207   KlassToOopClosure _cm_klass_closure;
4208  public:
4209   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4210   void do_klass(Klass* k) {
4211     if (k->has_accumulated_modified_oops()) {
4212       k->clear_accumulated_modified_oops();
4213 
4214       _cm_klass_closure.do_klass(k);
4215     }
4216   }
4217 };
4218 
4219 // The freelist lock is needed to prevent asserts, is it really needed?
4220 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4221 
4222   cl->set_freelistLock(freelistLock);
4223 
4224   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4225 
4226   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4227   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4228   PrecleanKlassClosure preclean_klass_closure(cl);
4229   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4230 
4231   verify_work_stacks_empty();
4232   verify_overflow_empty();
4233 }
4234 
4235 void CMSCollector::checkpointRootsFinal() {
4236   assert(_collectorState == FinalMarking, "incorrect state transition?");
4237   check_correct_thread_executing();
4238   // world is stopped at this checkpoint
4239   assert(SafepointSynchronize::is_at_safepoint(),
4240          "world should be stopped");
4241   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4242 
4243   verify_work_stacks_empty();
4244   verify_overflow_empty();
4245 
4246   if (PrintGCDetails) {
4247     gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
4248                         _young_gen->used() / K,
4249                         _young_gen->capacity() / K);
4250   }
4251   {
4252     if (CMSScavengeBeforeRemark) {
4253       GenCollectedHeap* gch = GenCollectedHeap::heap();
4254       // Temporarily set flag to false, GCH->do_collection will
4255       // expect it to be false and set to true
4256       FlagSetting fl(gch->_is_gc_active, false);
4257       NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4258         PrintGCDetails && Verbose, true, _gc_timer_cm);)
4259       gch->do_collection(true,                      // full (i.e. force, see below)
4260                          false,                     // !clear_all_soft_refs
4261                          0,                         // size
4262                          false,                     // is_tlab
4263                          GenCollectedHeap::YoungGen // type
4264         );
4265     }
4266     FreelistLocker x(this);
4267     MutexLockerEx y(bitMapLock(),
4268                     Mutex::_no_safepoint_check_flag);
4269     checkpointRootsFinalWork();
4270   }
4271   verify_work_stacks_empty();
4272   verify_overflow_empty();
4273 }
4274 
4275 void CMSCollector::checkpointRootsFinalWork() {
4276   NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
4277 
4278   assert(haveFreelistLocks(), "must have free list locks");
4279   assert_lock_strong(bitMapLock());
4280 
4281   ResourceMark rm;
4282   HandleMark   hm;
4283 
4284   GenCollectedHeap* gch = GenCollectedHeap::heap();
4285 
4286   if (should_unload_classes()) {
4287     CodeCache::gc_prologue();
4288   }
4289   assert(haveFreelistLocks(), "must have free list locks");
4290   assert_lock_strong(bitMapLock());
4291 
4292   // We might assume that we need not fill TLAB's when
4293   // CMSScavengeBeforeRemark is set, because we may have just done
4294   // a scavenge which would have filled all TLAB's -- and besides
4295   // Eden would be empty. This however may not always be the case --
4296   // for instance although we asked for a scavenge, it may not have
4297   // happened because of a JNI critical section. We probably need
4298   // a policy for deciding whether we can in that case wait until
4299   // the critical section releases and then do the remark following
4300   // the scavenge, and skip it here. In the absence of that policy,
4301   // or of an indication of whether the scavenge did indeed occur,
4302   // we cannot rely on TLAB's having been filled and must do
4303   // so here just in case a scavenge did not happen.
4304   gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4305   // Update the saved marks which may affect the root scans.
4306   gch->save_marks();
4307 
4308   if (CMSPrintEdenSurvivorChunks) {
4309     print_eden_and_survivor_chunk_arrays();
4310   }
4311 
4312   {
4313 #if defined(COMPILER2) || INCLUDE_JVMCI
4314     DerivedPointerTableDeactivate dpt_deact;
4315 #endif
4316 
4317     // Note on the role of the mod union table:
4318     // Since the marker in "markFromRoots" marks concurrently with
4319     // mutators, it is possible for some reachable objects not to have been
4320     // scanned. For instance, an only reference to an object A was
4321     // placed in object B after the marker scanned B. Unless B is rescanned,
4322     // A would be collected. Such updates to references in marked objects
4323     // are detected via the mod union table which is the set of all cards
4324     // dirtied since the first checkpoint in this GC cycle and prior to
4325     // the most recent young generation GC, minus those cleaned up by the
4326     // concurrent precleaning.
4327     if (CMSParallelRemarkEnabled) {
4328       GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
4329       do_remark_parallel();
4330     } else {
4331       GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, _gc_timer_cm);
4332       do_remark_non_parallel();
4333     }
4334   }
4335   verify_work_stacks_empty();
4336   verify_overflow_empty();
4337 
4338   {
4339     NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
4340     refProcessingWork();
4341   }
4342   verify_work_stacks_empty();
4343   verify_overflow_empty();
4344 
4345   if (should_unload_classes()) {
4346     CodeCache::gc_epilogue();
4347   }
4348   JvmtiExport::gc_epilogue();
4349 
4350   // If we encountered any (marking stack / work queue) overflow
4351   // events during the current CMS cycle, take appropriate
4352   // remedial measures, where possible, so as to try and avoid
4353   // recurrence of that condition.
4354   assert(_markStack.isEmpty(), "No grey objects");
4355   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4356                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4357   if (ser_ovflw > 0) {
4358     if (PrintCMSStatistics != 0) {
4359       gclog_or_tty->print_cr("Marking stack overflow (benign) "
4360         "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
4361         ", kac_preclean=" SIZE_FORMAT ")",
4362         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4363         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4364     }
4365     _markStack.expand();
4366     _ser_pmc_remark_ovflw = 0;
4367     _ser_pmc_preclean_ovflw = 0;
4368     _ser_kac_preclean_ovflw = 0;
4369     _ser_kac_ovflw = 0;
4370   }
4371   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4372     if (PrintCMSStatistics != 0) {
4373       gclog_or_tty->print_cr("Work queue overflow (benign) "
4374         "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4375         _par_pmc_remark_ovflw, _par_kac_ovflw);
4376     }
4377     _par_pmc_remark_ovflw = 0;
4378     _par_kac_ovflw = 0;
4379   }
4380   if (PrintCMSStatistics != 0) {
4381      if (_markStack._hit_limit > 0) {
4382        gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4383                               _markStack._hit_limit);
4384      }
4385      if (_markStack._failed_double > 0) {
4386        gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
4387                               " current capacity " SIZE_FORMAT,
4388                               _markStack._failed_double,
4389                               _markStack.capacity());
4390      }
4391   }
4392   _markStack._hit_limit = 0;
4393   _markStack._failed_double = 0;
4394 
4395   if ((VerifyAfterGC || VerifyDuringGC) &&
4396       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4397     verify_after_remark();
4398   }
4399 
4400   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4401 
4402   // Change under the freelistLocks.
4403   _collectorState = Sweeping;
4404   // Call isAllClear() under bitMapLock
4405   assert(_modUnionTable.isAllClear(),
4406       "Should be clear by end of the final marking");
4407   assert(_ct->klass_rem_set()->mod_union_is_clear(),
4408       "Should be clear by end of the final marking");
4409 }
4410 
4411 void CMSParInitialMarkTask::work(uint worker_id) {
4412   elapsedTimer _timer;
4413   ResourceMark rm;
4414   HandleMark   hm;
4415 
4416   // ---------- scan from roots --------------
4417   _timer.start();
4418   GenCollectedHeap* gch = GenCollectedHeap::heap();
4419   Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4420 
4421   // ---------- young gen roots --------------
4422   {
4423     work_on_young_gen_roots(worker_id, &par_mri_cl);
4424     _timer.stop();
4425     if (PrintCMSStatistics != 0) {
4426       gclog_or_tty->print_cr(
4427         "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4428         worker_id, _timer.seconds());
4429     }
4430   }
4431 
4432   // ---------- remaining roots --------------
4433   _timer.reset();
4434   _timer.start();
4435 
4436   CLDToOopClosure cld_closure(&par_mri_cl, true);
4437 
4438   gch->old_process_roots(_strong_roots_scope,
4439                          false,     // yg was scanned above
4440                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4441                          _collector->should_unload_classes(),
4442                          &par_mri_cl,
4443                          &cld_closure);
4444   assert(_collector->should_unload_classes()
4445          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4446          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4447   _timer.stop();
4448   if (PrintCMSStatistics != 0) {
4449     gclog_or_tty->print_cr(
4450       "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4451       worker_id, _timer.seconds());
4452   }
4453 }
4454 
4455 // Parallel remark task
4456 class CMSParRemarkTask: public CMSParMarkTask {
4457   CompactibleFreeListSpace* _cms_space;
4458 
4459   // The per-thread work queues, available here for stealing.
4460   OopTaskQueueSet*       _task_queues;
4461   ParallelTaskTerminator _term;
4462   StrongRootsScope*      _strong_roots_scope;
4463 
4464  public:
4465   // A value of 0 passed to n_workers will cause the number of
4466   // workers to be taken from the active workers in the work gang.
4467   CMSParRemarkTask(CMSCollector* collector,
4468                    CompactibleFreeListSpace* cms_space,
4469                    uint n_workers, WorkGang* workers,
4470                    OopTaskQueueSet* task_queues,
4471                    StrongRootsScope* strong_roots_scope):
4472     CMSParMarkTask("Rescan roots and grey objects in parallel",
4473                    collector, n_workers),
4474     _cms_space(cms_space),
4475     _task_queues(task_queues),
4476     _term(n_workers, task_queues),
4477     _strong_roots_scope(strong_roots_scope) { }
4478 
4479   OopTaskQueueSet* task_queues() { return _task_queues; }
4480 
4481   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4482 
4483   ParallelTaskTerminator* terminator() { return &_term; }
4484   uint n_workers() { return _n_workers; }
4485 
4486   void work(uint worker_id);
4487 
4488  private:
4489   // ... of  dirty cards in old space
4490   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4491                                   Par_MarkRefsIntoAndScanClosure* cl);
4492 
4493   // ... work stealing for the above
4494   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4495 };
4496 
4497 class RemarkKlassClosure : public KlassClosure {
4498   KlassToOopClosure _cm_klass_closure;
4499  public:
4500   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4501   void do_klass(Klass* k) {
4502     // Check if we have modified any oops in the Klass during the concurrent marking.
4503     if (k->has_accumulated_modified_oops()) {
4504       k->clear_accumulated_modified_oops();
4505 
4506       // We could have transfered the current modified marks to the accumulated marks,
4507       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4508     } else if (k->has_modified_oops()) {
4509       // Don't clear anything, this info is needed by the next young collection.
4510     } else {
4511       // No modified oops in the Klass.
4512       return;
4513     }
4514 
4515     // The klass has modified fields, need to scan the klass.
4516     _cm_klass_closure.do_klass(k);
4517   }
4518 };
4519 
4520 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4521   ParNewGeneration* young_gen = _collector->_young_gen;
4522   ContiguousSpace* eden_space = young_gen->eden();
4523   ContiguousSpace* from_space = young_gen->from();
4524   ContiguousSpace* to_space   = young_gen->to();
4525 
4526   HeapWord** eca = _collector->_eden_chunk_array;
4527   size_t     ect = _collector->_eden_chunk_index;
4528   HeapWord** sca = _collector->_survivor_chunk_array;
4529   size_t     sct = _collector->_survivor_chunk_index;
4530 
4531   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4532   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4533 
4534   do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4535   do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4536   do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4537 }
4538 
4539 // work_queue(i) is passed to the closure
4540 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
4541 // also is passed to do_dirty_card_rescan_tasks() and to
4542 // do_work_steal() to select the i-th task_queue.
4543 
4544 void CMSParRemarkTask::work(uint worker_id) {
4545   elapsedTimer _timer;
4546   ResourceMark rm;
4547   HandleMark   hm;
4548 
4549   // ---------- rescan from roots --------------
4550   _timer.start();
4551   GenCollectedHeap* gch = GenCollectedHeap::heap();
4552   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4553     _collector->_span, _collector->ref_processor(),
4554     &(_collector->_markBitMap),
4555     work_queue(worker_id));
4556 
4557   // Rescan young gen roots first since these are likely
4558   // coarsely partitioned and may, on that account, constitute
4559   // the critical path; thus, it's best to start off that
4560   // work first.
4561   // ---------- young gen roots --------------
4562   {
4563     work_on_young_gen_roots(worker_id, &par_mrias_cl);
4564     _timer.stop();
4565     if (PrintCMSStatistics != 0) {
4566       gclog_or_tty->print_cr(
4567         "Finished young gen rescan work in %dth thread: %3.3f sec",
4568         worker_id, _timer.seconds());
4569     }
4570   }
4571 
4572   // ---------- remaining roots --------------
4573   _timer.reset();
4574   _timer.start();
4575   gch->old_process_roots(_strong_roots_scope,
4576                          false,     // yg was scanned above
4577                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4578                          _collector->should_unload_classes(),
4579                          &par_mrias_cl,
4580                          NULL);     // The dirty klasses will be handled below
4581 
4582   assert(_collector->should_unload_classes()
4583          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4584          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4585   _timer.stop();
4586   if (PrintCMSStatistics != 0) {
4587     gclog_or_tty->print_cr(
4588       "Finished remaining root rescan work in %dth thread: %3.3f sec",
4589       worker_id, _timer.seconds());
4590   }
4591 
4592   // ---------- unhandled CLD scanning ----------
4593   if (worker_id == 0) { // Single threaded at the moment.
4594     _timer.reset();
4595     _timer.start();
4596 
4597     // Scan all new class loader data objects and new dependencies that were
4598     // introduced during concurrent marking.
4599     ResourceMark rm;
4600     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4601     for (int i = 0; i < array->length(); i++) {
4602       par_mrias_cl.do_cld_nv(array->at(i));
4603     }
4604 
4605     // We don't need to keep track of new CLDs anymore.
4606     ClassLoaderDataGraph::remember_new_clds(false);
4607 
4608     _timer.stop();
4609     if (PrintCMSStatistics != 0) {
4610       gclog_or_tty->print_cr(
4611           "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4612           worker_id, _timer.seconds());
4613     }
4614   }
4615 
4616   // ---------- dirty klass scanning ----------
4617   if (worker_id == 0) { // Single threaded at the moment.
4618     _timer.reset();
4619     _timer.start();
4620 
4621     // Scan all classes that was dirtied during the concurrent marking phase.
4622     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4623     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4624 
4625     _timer.stop();
4626     if (PrintCMSStatistics != 0) {
4627       gclog_or_tty->print_cr(
4628           "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4629           worker_id, _timer.seconds());
4630     }
4631   }
4632 
4633   // We might have added oops to ClassLoaderData::_handles during the
4634   // concurrent marking phase. These oops point to newly allocated objects
4635   // that are guaranteed to be kept alive. Either by the direct allocation
4636   // code, or when the young collector processes the roots. Hence,
4637   // we don't have to revisit the _handles block during the remark phase.
4638 
4639   // ---------- rescan dirty cards ------------
4640   _timer.reset();
4641   _timer.start();
4642 
4643   // Do the rescan tasks for each of the two spaces
4644   // (cms_space) in turn.
4645   // "worker_id" is passed to select the task_queue for "worker_id"
4646   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4647   _timer.stop();
4648   if (PrintCMSStatistics != 0) {
4649     gclog_or_tty->print_cr(
4650       "Finished dirty card rescan work in %dth thread: %3.3f sec",
4651       worker_id, _timer.seconds());
4652   }
4653 
4654   // ---------- steal work from other threads ...
4655   // ---------- ... and drain overflow list.
4656   _timer.reset();
4657   _timer.start();
4658   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4659   _timer.stop();
4660   if (PrintCMSStatistics != 0) {
4661     gclog_or_tty->print_cr(
4662       "Finished work stealing in %dth thread: %3.3f sec",
4663       worker_id, _timer.seconds());
4664   }
4665 }
4666 
4667 // Note that parameter "i" is not used.
4668 void
4669 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4670   OopsInGenClosure* cl, ContiguousSpace* space,
4671   HeapWord** chunk_array, size_t chunk_top) {
4672   // Until all tasks completed:
4673   // . claim an unclaimed task
4674   // . compute region boundaries corresponding to task claimed
4675   //   using chunk_array
4676   // . par_oop_iterate(cl) over that region
4677 
4678   ResourceMark rm;
4679   HandleMark   hm;
4680 
4681   SequentialSubTasksDone* pst = space->par_seq_tasks();
4682 
4683   uint nth_task = 0;
4684   uint n_tasks  = pst->n_tasks();
4685 
4686   if (n_tasks > 0) {
4687     assert(pst->valid(), "Uninitialized use?");
4688     HeapWord *start, *end;
4689     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4690       // We claimed task # nth_task; compute its boundaries.
4691       if (chunk_top == 0) {  // no samples were taken
4692         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4693         start = space->bottom();
4694         end   = space->top();
4695       } else if (nth_task == 0) {
4696         start = space->bottom();
4697         end   = chunk_array[nth_task];
4698       } else if (nth_task < (uint)chunk_top) {
4699         assert(nth_task >= 1, "Control point invariant");
4700         start = chunk_array[nth_task - 1];
4701         end   = chunk_array[nth_task];
4702       } else {
4703         assert(nth_task == (uint)chunk_top, "Control point invariant");
4704         start = chunk_array[chunk_top - 1];
4705         end   = space->top();
4706       }
4707       MemRegion mr(start, end);
4708       // Verify that mr is in space
4709       assert(mr.is_empty() || space->used_region().contains(mr),
4710              "Should be in space");
4711       // Verify that "start" is an object boundary
4712       assert(mr.is_empty() || oop(mr.start())->is_oop(),
4713              "Should be an oop");
4714       space->par_oop_iterate(mr, cl);
4715     }
4716     pst->all_tasks_completed();
4717   }
4718 }
4719 
4720 void
4721 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4722   CompactibleFreeListSpace* sp, int i,
4723   Par_MarkRefsIntoAndScanClosure* cl) {
4724   // Until all tasks completed:
4725   // . claim an unclaimed task
4726   // . compute region boundaries corresponding to task claimed
4727   // . transfer dirty bits ct->mut for that region
4728   // . apply rescanclosure to dirty mut bits for that region
4729 
4730   ResourceMark rm;
4731   HandleMark   hm;
4732 
4733   OopTaskQueue* work_q = work_queue(i);
4734   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4735   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4736   // CAUTION: This closure has state that persists across calls to
4737   // the work method dirty_range_iterate_clear() in that it has
4738   // embedded in it a (subtype of) UpwardsObjectClosure. The
4739   // use of that state in the embedded UpwardsObjectClosure instance
4740   // assumes that the cards are always iterated (even if in parallel
4741   // by several threads) in monotonically increasing order per each
4742   // thread. This is true of the implementation below which picks
4743   // card ranges (chunks) in monotonically increasing order globally
4744   // and, a-fortiori, in monotonically increasing order per thread
4745   // (the latter order being a subsequence of the former).
4746   // If the work code below is ever reorganized into a more chaotic
4747   // work-partitioning form than the current "sequential tasks"
4748   // paradigm, the use of that persistent state will have to be
4749   // revisited and modified appropriately. See also related
4750   // bug 4756801 work on which should examine this code to make
4751   // sure that the changes there do not run counter to the
4752   // assumptions made here and necessary for correctness and
4753   // efficiency. Note also that this code might yield inefficient
4754   // behavior in the case of very large objects that span one or
4755   // more work chunks. Such objects would potentially be scanned
4756   // several times redundantly. Work on 4756801 should try and
4757   // address that performance anomaly if at all possible. XXX
4758   MemRegion  full_span  = _collector->_span;
4759   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4760   MarkFromDirtyCardsClosure
4761     greyRescanClosure(_collector, full_span, // entire span of interest
4762                       sp, bm, work_q, cl);
4763 
4764   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4765   assert(pst->valid(), "Uninitialized use?");
4766   uint nth_task = 0;
4767   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4768   MemRegion span = sp->used_region();
4769   HeapWord* start_addr = span.start();
4770   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4771                                            alignment);
4772   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4773   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4774          start_addr, "Check alignment");
4775   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4776          chunk_size, "Check alignment");
4777 
4778   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4779     // Having claimed the nth_task, compute corresponding mem-region,
4780     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4781     // The alignment restriction ensures that we do not need any
4782     // synchronization with other gang-workers while setting or
4783     // clearing bits in thus chunk of the MUT.
4784     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4785                                     start_addr + (nth_task+1)*chunk_size);
4786     // The last chunk's end might be way beyond end of the
4787     // used region. In that case pull back appropriately.
4788     if (this_span.end() > end_addr) {
4789       this_span.set_end(end_addr);
4790       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4791     }
4792     // Iterate over the dirty cards covering this chunk, marking them
4793     // precleaned, and setting the corresponding bits in the mod union
4794     // table. Since we have been careful to partition at Card and MUT-word
4795     // boundaries no synchronization is needed between parallel threads.
4796     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4797                                                  &modUnionClosure);
4798 
4799     // Having transferred these marks into the modUnionTable,
4800     // rescan the marked objects on the dirty cards in the modUnionTable.
4801     // Even if this is at a synchronous collection, the initial marking
4802     // may have been done during an asynchronous collection so there
4803     // may be dirty bits in the mod-union table.
4804     _collector->_modUnionTable.dirty_range_iterate_clear(
4805                   this_span, &greyRescanClosure);
4806     _collector->_modUnionTable.verifyNoOneBitsInRange(
4807                                  this_span.start(),
4808                                  this_span.end());
4809   }
4810   pst->all_tasks_completed();  // declare that i am done
4811 }
4812 
4813 // . see if we can share work_queues with ParNew? XXX
4814 void
4815 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4816                                 int* seed) {
4817   OopTaskQueue* work_q = work_queue(i);
4818   NOT_PRODUCT(int num_steals = 0;)
4819   oop obj_to_scan;
4820   CMSBitMap* bm = &(_collector->_markBitMap);
4821 
4822   while (true) {
4823     // Completely finish any left over work from (an) earlier round(s)
4824     cl->trim_queue(0);
4825     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4826                                          (size_t)ParGCDesiredObjsFromOverflowList);
4827     // Now check if there's any work in the overflow list
4828     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4829     // only affects the number of attempts made to get work from the
4830     // overflow list and does not affect the number of workers.  Just
4831     // pass ParallelGCThreads so this behavior is unchanged.
4832     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4833                                                 work_q,
4834                                                 ParallelGCThreads)) {
4835       // found something in global overflow list;
4836       // not yet ready to go stealing work from others.
4837       // We'd like to assert(work_q->size() != 0, ...)
4838       // because we just took work from the overflow list,
4839       // but of course we can't since all of that could have
4840       // been already stolen from us.
4841       // "He giveth and He taketh away."
4842       continue;
4843     }
4844     // Verify that we have no work before we resort to stealing
4845     assert(work_q->size() == 0, "Have work, shouldn't steal");
4846     // Try to steal from other queues that have work
4847     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4848       NOT_PRODUCT(num_steals++;)
4849       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4850       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4851       // Do scanning work
4852       obj_to_scan->oop_iterate(cl);
4853       // Loop around, finish this work, and try to steal some more
4854     } else if (terminator()->offer_termination()) {
4855         break;  // nirvana from the infinite cycle
4856     }
4857   }
4858   NOT_PRODUCT(
4859     if (PrintCMSStatistics != 0) {
4860       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4861     }
4862   )
4863   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4864          "Else our work is not yet done");
4865 }
4866 
4867 // Record object boundaries in _eden_chunk_array by sampling the eden
4868 // top in the slow-path eden object allocation code path and record
4869 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4870 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4871 // sampling in sample_eden() that activates during the part of the
4872 // preclean phase.
4873 void CMSCollector::sample_eden_chunk() {
4874   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4875     if (_eden_chunk_lock->try_lock()) {
4876       // Record a sample. This is the critical section. The contents
4877       // of the _eden_chunk_array have to be non-decreasing in the
4878       // address order.
4879       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4880       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4881              "Unexpected state of Eden");
4882       if (_eden_chunk_index == 0 ||
4883           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4884            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4885                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4886         _eden_chunk_index++;  // commit sample
4887       }
4888       _eden_chunk_lock->unlock();
4889     }
4890   }
4891 }
4892 
4893 // Return a thread-local PLAB recording array, as appropriate.
4894 void* CMSCollector::get_data_recorder(int thr_num) {
4895   if (_survivor_plab_array != NULL &&
4896       (CMSPLABRecordAlways ||
4897        (_collectorState > Marking && _collectorState < FinalMarking))) {
4898     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4899     ChunkArray* ca = &_survivor_plab_array[thr_num];
4900     ca->reset();   // clear it so that fresh data is recorded
4901     return (void*) ca;
4902   } else {
4903     return NULL;
4904   }
4905 }
4906 
4907 // Reset all the thread-local PLAB recording arrays
4908 void CMSCollector::reset_survivor_plab_arrays() {
4909   for (uint i = 0; i < ParallelGCThreads; i++) {
4910     _survivor_plab_array[i].reset();
4911   }
4912 }
4913 
4914 // Merge the per-thread plab arrays into the global survivor chunk
4915 // array which will provide the partitioning of the survivor space
4916 // for CMS initial scan and rescan.
4917 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4918                                               int no_of_gc_threads) {
4919   assert(_survivor_plab_array  != NULL, "Error");
4920   assert(_survivor_chunk_array != NULL, "Error");
4921   assert(_collectorState == FinalMarking ||
4922          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4923   for (int j = 0; j < no_of_gc_threads; j++) {
4924     _cursor[j] = 0;
4925   }
4926   HeapWord* top = surv->top();
4927   size_t i;
4928   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4929     HeapWord* min_val = top;          // Higher than any PLAB address
4930     uint      min_tid = 0;            // position of min_val this round
4931     for (int j = 0; j < no_of_gc_threads; j++) {
4932       ChunkArray* cur_sca = &_survivor_plab_array[j];
4933       if (_cursor[j] == cur_sca->end()) {
4934         continue;
4935       }
4936       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4937       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4938       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4939       if (cur_val < min_val) {
4940         min_tid = j;
4941         min_val = cur_val;
4942       } else {
4943         assert(cur_val < top, "All recorded addresses should be less");
4944       }
4945     }
4946     // At this point min_val and min_tid are respectively
4947     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4948     // and the thread (j) that witnesses that address.
4949     // We record this address in the _survivor_chunk_array[i]
4950     // and increment _cursor[min_tid] prior to the next round i.
4951     if (min_val == top) {
4952       break;
4953     }
4954     _survivor_chunk_array[i] = min_val;
4955     _cursor[min_tid]++;
4956   }
4957   // We are all done; record the size of the _survivor_chunk_array
4958   _survivor_chunk_index = i; // exclusive: [0, i)
4959   if (PrintCMSStatistics > 0) {
4960     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4961   }
4962   // Verify that we used up all the recorded entries
4963   #ifdef ASSERT
4964     size_t total = 0;
4965     for (int j = 0; j < no_of_gc_threads; j++) {
4966       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4967       total += _cursor[j];
4968     }
4969     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4970     // Check that the merged array is in sorted order
4971     if (total > 0) {
4972       for (size_t i = 0; i < total - 1; i++) {
4973         if (PrintCMSStatistics > 0) {
4974           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4975                               i, p2i(_survivor_chunk_array[i]));
4976         }
4977         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4978                "Not sorted");
4979       }
4980     }
4981   #endif // ASSERT
4982 }
4983 
4984 // Set up the space's par_seq_tasks structure for work claiming
4985 // for parallel initial scan and rescan of young gen.
4986 // See ParRescanTask where this is currently used.
4987 void
4988 CMSCollector::
4989 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4990   assert(n_threads > 0, "Unexpected n_threads argument");
4991 
4992   // Eden space
4993   if (!_young_gen->eden()->is_empty()) {
4994     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4995     assert(!pst->valid(), "Clobbering existing data?");
4996     // Each valid entry in [0, _eden_chunk_index) represents a task.
4997     size_t n_tasks = _eden_chunk_index + 1;
4998     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4999     // Sets the condition for completion of the subtask (how many threads
5000     // need to finish in order to be done).
5001     pst->set_n_threads(n_threads);
5002     pst->set_n_tasks((int)n_tasks);
5003   }
5004 
5005   // Merge the survivor plab arrays into _survivor_chunk_array
5006   if (_survivor_plab_array != NULL) {
5007     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5008   } else {
5009     assert(_survivor_chunk_index == 0, "Error");
5010   }
5011 
5012   // To space
5013   {
5014     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5015     assert(!pst->valid(), "Clobbering existing data?");
5016     // Sets the condition for completion of the subtask (how many threads
5017     // need to finish in order to be done).
5018     pst->set_n_threads(n_threads);
5019     pst->set_n_tasks(1);
5020     assert(pst->valid(), "Error");
5021   }
5022 
5023   // From space
5024   {
5025     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5026     assert(!pst->valid(), "Clobbering existing data?");
5027     size_t n_tasks = _survivor_chunk_index + 1;
5028     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5029     // Sets the condition for completion of the subtask (how many threads
5030     // need to finish in order to be done).
5031     pst->set_n_threads(n_threads);
5032     pst->set_n_tasks((int)n_tasks);
5033     assert(pst->valid(), "Error");
5034   }
5035 }
5036 
5037 // Parallel version of remark
5038 void CMSCollector::do_remark_parallel() {
5039   GenCollectedHeap* gch = GenCollectedHeap::heap();
5040   WorkGang* workers = gch->workers();
5041   assert(workers != NULL, "Need parallel worker threads.");
5042   // Choose to use the number of GC workers most recently set
5043   // into "active_workers".
5044   uint n_workers = workers->active_workers();
5045 
5046   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
5047 
5048   StrongRootsScope srs(n_workers);
5049 
5050   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5051 
5052   // We won't be iterating over the cards in the card table updating
5053   // the younger_gen cards, so we shouldn't call the following else
5054   // the verification code as well as subsequent younger_refs_iterate
5055   // code would get confused. XXX
5056   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5057 
5058   // The young gen rescan work will not be done as part of
5059   // process_roots (which currently doesn't know how to
5060   // parallelize such a scan), but rather will be broken up into
5061   // a set of parallel tasks (via the sampling that the [abortable]
5062   // preclean phase did of eden, plus the [two] tasks of
5063   // scanning the [two] survivor spaces. Further fine-grain
5064   // parallelization of the scanning of the survivor spaces
5065   // themselves, and of precleaning of the young gen itself
5066   // is deferred to the future.
5067   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5068 
5069   // The dirty card rescan work is broken up into a "sequence"
5070   // of parallel tasks (per constituent space) that are dynamically
5071   // claimed by the parallel threads.
5072   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5073 
5074   // It turns out that even when we're using 1 thread, doing the work in a
5075   // separate thread causes wide variance in run times.  We can't help this
5076   // in the multi-threaded case, but we special-case n=1 here to get
5077   // repeatable measurements of the 1-thread overhead of the parallel code.
5078   if (n_workers > 1) {
5079     // Make refs discovery MT-safe, if it isn't already: it may not
5080     // necessarily be so, since it's possible that we are doing
5081     // ST marking.
5082     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5083     workers->run_task(&tsk);
5084   } else {
5085     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5086     tsk.work(0);
5087   }
5088 
5089   // restore, single-threaded for now, any preserved marks
5090   // as a result of work_q overflow
5091   restore_preserved_marks_if_any();
5092 }
5093 
5094 // Non-parallel version of remark
5095 void CMSCollector::do_remark_non_parallel() {
5096   ResourceMark rm;
5097   HandleMark   hm;
5098   GenCollectedHeap* gch = GenCollectedHeap::heap();
5099   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5100 
5101   MarkRefsIntoAndScanClosure
5102     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5103              &_markStack, this,
5104              false /* should_yield */, false /* not precleaning */);
5105   MarkFromDirtyCardsClosure
5106     markFromDirtyCardsClosure(this, _span,
5107                               NULL,  // space is set further below
5108                               &_markBitMap, &_markStack, &mrias_cl);
5109   {
5110     GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5111     // Iterate over the dirty cards, setting the corresponding bits in the
5112     // mod union table.
5113     {
5114       ModUnionClosure modUnionClosure(&_modUnionTable);
5115       _ct->ct_bs()->dirty_card_iterate(
5116                       _cmsGen->used_region(),
5117                       &modUnionClosure);
5118     }
5119     // Having transferred these marks into the modUnionTable, we just need
5120     // to rescan the marked objects on the dirty cards in the modUnionTable.
5121     // The initial marking may have been done during an asynchronous
5122     // collection so there may be dirty bits in the mod-union table.
5123     const int alignment =
5124       CardTableModRefBS::card_size * BitsPerWord;
5125     {
5126       // ... First handle dirty cards in CMS gen
5127       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5128       MemRegion ur = _cmsGen->used_region();
5129       HeapWord* lb = ur.start();
5130       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5131       MemRegion cms_span(lb, ub);
5132       _modUnionTable.dirty_range_iterate_clear(cms_span,
5133                                                &markFromDirtyCardsClosure);
5134       verify_work_stacks_empty();
5135       if (PrintCMSStatistics != 0) {
5136         gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5137           markFromDirtyCardsClosure.num_dirty_cards());
5138       }
5139     }
5140   }
5141   if (VerifyDuringGC &&
5142       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5143     HandleMark hm;  // Discard invalid handles created during verification
5144     Universe::verify();
5145   }
5146   {
5147     GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5148 
5149     verify_work_stacks_empty();
5150 
5151     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5152     StrongRootsScope srs(1);
5153 
5154     gch->old_process_roots(&srs,
5155                            true,  // young gen as roots
5156                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
5157                            should_unload_classes(),
5158                            &mrias_cl,
5159                            NULL); // The dirty klasses will be handled below
5160 
5161     assert(should_unload_classes()
5162            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5163            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5164   }
5165 
5166   {
5167     GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5168 
5169     verify_work_stacks_empty();
5170 
5171     // Scan all class loader data objects that might have been introduced
5172     // during concurrent marking.
5173     ResourceMark rm;
5174     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5175     for (int i = 0; i < array->length(); i++) {
5176       mrias_cl.do_cld_nv(array->at(i));
5177     }
5178 
5179     // We don't need to keep track of new CLDs anymore.
5180     ClassLoaderDataGraph::remember_new_clds(false);
5181 
5182     verify_work_stacks_empty();
5183   }
5184 
5185   {
5186     GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
5187 
5188     verify_work_stacks_empty();
5189 
5190     RemarkKlassClosure remark_klass_closure(&mrias_cl);
5191     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5192 
5193     verify_work_stacks_empty();
5194   }
5195 
5196   // We might have added oops to ClassLoaderData::_handles during the
5197   // concurrent marking phase. These oops point to newly allocated objects
5198   // that are guaranteed to be kept alive. Either by the direct allocation
5199   // code, or when the young collector processes the roots. Hence,
5200   // we don't have to revisit the _handles block during the remark phase.
5201 
5202   verify_work_stacks_empty();
5203   // Restore evacuated mark words, if any, used for overflow list links
5204   if (!CMSOverflowEarlyRestoration) {
5205     restore_preserved_marks_if_any();
5206   }
5207   verify_overflow_empty();
5208 }
5209 
5210 ////////////////////////////////////////////////////////
5211 // Parallel Reference Processing Task Proxy Class
5212 ////////////////////////////////////////////////////////
5213 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5214   OopTaskQueueSet*       _queues;
5215   ParallelTaskTerminator _terminator;
5216  public:
5217   AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5218     AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5219   ParallelTaskTerminator* terminator() { return &_terminator; }
5220   OopTaskQueueSet* queues() { return _queues; }
5221 };
5222 
5223 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5224   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5225   CMSCollector*          _collector;
5226   CMSBitMap*             _mark_bit_map;
5227   const MemRegion        _span;
5228   ProcessTask&           _task;
5229 
5230 public:
5231   CMSRefProcTaskProxy(ProcessTask&     task,
5232                       CMSCollector*    collector,
5233                       const MemRegion& span,
5234                       CMSBitMap*       mark_bit_map,
5235                       AbstractWorkGang* workers,
5236                       OopTaskQueueSet* task_queues):
5237     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5238       task_queues,
5239       workers->active_workers()),
5240     _task(task),
5241     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5242   {
5243     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5244            "Inconsistency in _span");
5245   }
5246 
5247   OopTaskQueueSet* task_queues() { return queues(); }
5248 
5249   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5250 
5251   void do_work_steal(int i,
5252                      CMSParDrainMarkingStackClosure* drain,
5253                      CMSParKeepAliveClosure* keep_alive,
5254                      int* seed);
5255 
5256   virtual void work(uint worker_id);
5257 };
5258 
5259 void CMSRefProcTaskProxy::work(uint worker_id) {
5260   ResourceMark rm;
5261   HandleMark hm;
5262   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5263   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5264                                         _mark_bit_map,
5265                                         work_queue(worker_id));
5266   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5267                                                  _mark_bit_map,
5268                                                  work_queue(worker_id));
5269   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5270   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5271   if (_task.marks_oops_alive()) {
5272     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5273                   _collector->hash_seed(worker_id));
5274   }
5275   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5276   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5277 }
5278 
5279 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5280   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5281   EnqueueTask& _task;
5282 
5283 public:
5284   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5285     : AbstractGangTask("Enqueue reference objects in parallel"),
5286       _task(task)
5287   { }
5288 
5289   virtual void work(uint worker_id)
5290   {
5291     _task.work(worker_id);
5292   }
5293 };
5294 
5295 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5296   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5297    _span(span),
5298    _bit_map(bit_map),
5299    _work_queue(work_queue),
5300    _mark_and_push(collector, span, bit_map, work_queue),
5301    _low_water_mark(MIN2((work_queue->max_elems()/4),
5302                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5303 { }
5304 
5305 // . see if we can share work_queues with ParNew? XXX
5306 void CMSRefProcTaskProxy::do_work_steal(int i,
5307   CMSParDrainMarkingStackClosure* drain,
5308   CMSParKeepAliveClosure* keep_alive,
5309   int* seed) {
5310   OopTaskQueue* work_q = work_queue(i);
5311   NOT_PRODUCT(int num_steals = 0;)
5312   oop obj_to_scan;
5313 
5314   while (true) {
5315     // Completely finish any left over work from (an) earlier round(s)
5316     drain->trim_queue(0);
5317     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5318                                          (size_t)ParGCDesiredObjsFromOverflowList);
5319     // Now check if there's any work in the overflow list
5320     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5321     // only affects the number of attempts made to get work from the
5322     // overflow list and does not affect the number of workers.  Just
5323     // pass ParallelGCThreads so this behavior is unchanged.
5324     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5325                                                 work_q,
5326                                                 ParallelGCThreads)) {
5327       // Found something in global overflow list;
5328       // not yet ready to go stealing work from others.
5329       // We'd like to assert(work_q->size() != 0, ...)
5330       // because we just took work from the overflow list,
5331       // but of course we can't, since all of that might have
5332       // been already stolen from us.
5333       continue;
5334     }
5335     // Verify that we have no work before we resort to stealing
5336     assert(work_q->size() == 0, "Have work, shouldn't steal");
5337     // Try to steal from other queues that have work
5338     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5339       NOT_PRODUCT(num_steals++;)
5340       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5341       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5342       // Do scanning work
5343       obj_to_scan->oop_iterate(keep_alive);
5344       // Loop around, finish this work, and try to steal some more
5345     } else if (terminator()->offer_termination()) {
5346       break;  // nirvana from the infinite cycle
5347     }
5348   }
5349   NOT_PRODUCT(
5350     if (PrintCMSStatistics != 0) {
5351       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5352     }
5353   )
5354 }
5355 
5356 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5357 {
5358   GenCollectedHeap* gch = GenCollectedHeap::heap();
5359   WorkGang* workers = gch->workers();
5360   assert(workers != NULL, "Need parallel worker threads.");
5361   CMSRefProcTaskProxy rp_task(task, &_collector,
5362                               _collector.ref_processor()->span(),
5363                               _collector.markBitMap(),
5364                               workers, _collector.task_queues());
5365   workers->run_task(&rp_task);
5366 }
5367 
5368 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5369 {
5370 
5371   GenCollectedHeap* gch = GenCollectedHeap::heap();
5372   WorkGang* workers = gch->workers();
5373   assert(workers != NULL, "Need parallel worker threads.");
5374   CMSRefEnqueueTaskProxy enq_task(task);
5375   workers->run_task(&enq_task);
5376 }
5377 
5378 void CMSCollector::refProcessingWork() {
5379   ResourceMark rm;
5380   HandleMark   hm;
5381 
5382   ReferenceProcessor* rp = ref_processor();
5383   assert(rp->span().equals(_span), "Spans should be equal");
5384   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5385   // Process weak references.
5386   rp->setup_policy(false);
5387   verify_work_stacks_empty();
5388 
5389   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5390                                           &_markStack, false /* !preclean */);
5391   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5392                                 _span, &_markBitMap, &_markStack,
5393                                 &cmsKeepAliveClosure, false /* !preclean */);
5394   {
5395     GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
5396 
5397     ReferenceProcessorStats stats;
5398     if (rp->processing_is_mt()) {
5399       // Set the degree of MT here.  If the discovery is done MT, there
5400       // may have been a different number of threads doing the discovery
5401       // and a different number of discovered lists may have Ref objects.
5402       // That is OK as long as the Reference lists are balanced (see
5403       // balance_all_queues() and balance_queues()).
5404       GenCollectedHeap* gch = GenCollectedHeap::heap();
5405       uint active_workers = ParallelGCThreads;
5406       WorkGang* workers = gch->workers();
5407       if (workers != NULL) {
5408         active_workers = workers->active_workers();
5409         // The expectation is that active_workers will have already
5410         // been set to a reasonable value.  If it has not been set,
5411         // investigate.
5412         assert(active_workers > 0, "Should have been set during scavenge");
5413       }
5414       rp->set_active_mt_degree(active_workers);
5415       CMSRefProcTaskExecutor task_executor(*this);
5416       stats = rp->process_discovered_references(&_is_alive_closure,
5417                                         &cmsKeepAliveClosure,
5418                                         &cmsDrainMarkingStackClosure,
5419                                         &task_executor,
5420                                         _gc_timer_cm);
5421     } else {
5422       stats = rp->process_discovered_references(&_is_alive_closure,
5423                                         &cmsKeepAliveClosure,
5424                                         &cmsDrainMarkingStackClosure,
5425                                         NULL,
5426                                         _gc_timer_cm);
5427     }
5428     _gc_tracer_cm->report_gc_reference_stats(stats);
5429 
5430   }
5431 
5432   // This is the point where the entire marking should have completed.
5433   verify_work_stacks_empty();
5434 
5435   if (should_unload_classes()) {
5436     {
5437       GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
5438 
5439       // Unload classes and purge the SystemDictionary.
5440       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5441 
5442       // Unload nmethods.
5443       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5444 
5445       // Prune dead klasses from subklass/sibling/implementor lists.
5446       Klass::clean_weak_klass_links(&_is_alive_closure);
5447     }
5448 
5449     {
5450       GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
5451       // Clean up unreferenced symbols in symbol table.
5452       SymbolTable::unlink();
5453     }
5454 
5455     {
5456       GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
5457       // Delete entries for dead interned strings.
5458       StringTable::unlink(&_is_alive_closure);
5459     }
5460   }
5461 
5462 
5463   // Restore any preserved marks as a result of mark stack or
5464   // work queue overflow
5465   restore_preserved_marks_if_any();  // done single-threaded for now
5466 
5467   rp->set_enqueuing_is_done(true);
5468   if (rp->processing_is_mt()) {
5469     rp->balance_all_queues();
5470     CMSRefProcTaskExecutor task_executor(*this);
5471     rp->enqueue_discovered_references(&task_executor);
5472   } else {
5473     rp->enqueue_discovered_references(NULL);
5474   }
5475   rp->verify_no_references_recorded();
5476   assert(!rp->discovery_enabled(), "should have been disabled");
5477 }
5478 
5479 #ifndef PRODUCT
5480 void CMSCollector::check_correct_thread_executing() {
5481   Thread* t = Thread::current();
5482   // Only the VM thread or the CMS thread should be here.
5483   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5484          "Unexpected thread type");
5485   // If this is the vm thread, the foreground process
5486   // should not be waiting.  Note that _foregroundGCIsActive is
5487   // true while the foreground collector is waiting.
5488   if (_foregroundGCShouldWait) {
5489     // We cannot be the VM thread
5490     assert(t->is_ConcurrentGC_thread(),
5491            "Should be CMS thread");
5492   } else {
5493     // We can be the CMS thread only if we are in a stop-world
5494     // phase of CMS collection.
5495     if (t->is_ConcurrentGC_thread()) {
5496       assert(_collectorState == InitialMarking ||
5497              _collectorState == FinalMarking,
5498              "Should be a stop-world phase");
5499       // The CMS thread should be holding the CMS_token.
5500       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5501              "Potential interference with concurrently "
5502              "executing VM thread");
5503     }
5504   }
5505 }
5506 #endif
5507 
5508 void CMSCollector::sweep() {
5509   assert(_collectorState == Sweeping, "just checking");
5510   check_correct_thread_executing();
5511   verify_work_stacks_empty();
5512   verify_overflow_empty();
5513   increment_sweep_count();
5514   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5515 
5516   _inter_sweep_timer.stop();
5517   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5518 
5519   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5520   _intra_sweep_timer.reset();
5521   _intra_sweep_timer.start();
5522   {
5523     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5524     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5525     // First sweep the old gen
5526     {
5527       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5528                                bitMapLock());
5529       sweepWork(_cmsGen);
5530     }
5531 
5532     // Update Universe::_heap_*_at_gc figures.
5533     // We need all the free list locks to make the abstract state
5534     // transition from Sweeping to Resetting. See detailed note
5535     // further below.
5536     {
5537       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5538       // Update heap occupancy information which is used as
5539       // input to soft ref clearing policy at the next gc.
5540       Universe::update_heap_info_at_gc();
5541       _collectorState = Resizing;
5542     }
5543   }
5544   verify_work_stacks_empty();
5545   verify_overflow_empty();
5546 
5547   if (should_unload_classes()) {
5548     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5549     // requires that the virtual spaces are stable and not deleted.
5550     ClassLoaderDataGraph::set_should_purge(true);
5551   }
5552 
5553   _intra_sweep_timer.stop();
5554   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5555 
5556   _inter_sweep_timer.reset();
5557   _inter_sweep_timer.start();
5558 
5559   // We need to use a monotonically non-decreasing time in ms
5560   // or we will see time-warp warnings and os::javaTimeMillis()
5561   // does not guarantee monotonicity.
5562   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5563   update_time_of_last_gc(now);
5564 
5565   // NOTE on abstract state transitions:
5566   // Mutators allocate-live and/or mark the mod-union table dirty
5567   // based on the state of the collection.  The former is done in
5568   // the interval [Marking, Sweeping] and the latter in the interval
5569   // [Marking, Sweeping).  Thus the transitions into the Marking state
5570   // and out of the Sweeping state must be synchronously visible
5571   // globally to the mutators.
5572   // The transition into the Marking state happens with the world
5573   // stopped so the mutators will globally see it.  Sweeping is
5574   // done asynchronously by the background collector so the transition
5575   // from the Sweeping state to the Resizing state must be done
5576   // under the freelistLock (as is the check for whether to
5577   // allocate-live and whether to dirty the mod-union table).
5578   assert(_collectorState == Resizing, "Change of collector state to"
5579     " Resizing must be done under the freelistLocks (plural)");
5580 
5581   // Now that sweeping has been completed, we clear
5582   // the incremental_collection_failed flag,
5583   // thus inviting a younger gen collection to promote into
5584   // this generation. If such a promotion may still fail,
5585   // the flag will be set again when a young collection is
5586   // attempted.
5587   GenCollectedHeap* gch = GenCollectedHeap::heap();
5588   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5589   gch->update_full_collections_completed(_collection_count_start);
5590 }
5591 
5592 // FIX ME!!! Looks like this belongs in CFLSpace, with
5593 // CMSGen merely delegating to it.
5594 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5595   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5596   HeapWord*  minAddr        = _cmsSpace->bottom();
5597   HeapWord*  largestAddr    =
5598     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5599   if (largestAddr == NULL) {
5600     // The dictionary appears to be empty.  In this case
5601     // try to coalesce at the end of the heap.
5602     largestAddr = _cmsSpace->end();
5603   }
5604   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5605   size_t nearLargestOffset =
5606     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5607   if (PrintFLSStatistics != 0) {
5608     gclog_or_tty->print_cr(
5609       "CMS: Large Block: " PTR_FORMAT ";"
5610       " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5611       p2i(largestAddr),
5612       p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5613   }
5614   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5615 }
5616 
5617 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5618   return addr >= _cmsSpace->nearLargestChunk();
5619 }
5620 
5621 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5622   return _cmsSpace->find_chunk_at_end();
5623 }
5624 
5625 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5626                                                     bool full) {
5627   // If the young generation has been collected, gather any statistics
5628   // that are of interest at this point.
5629   bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5630   if (!full && current_is_young) {
5631     // Gather statistics on the young generation collection.
5632     collector()->stats().record_gc0_end(used());
5633   }
5634 }
5635 
5636 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5637   // We iterate over the space(s) underlying this generation,
5638   // checking the mark bit map to see if the bits corresponding
5639   // to specific blocks are marked or not. Blocks that are
5640   // marked are live and are not swept up. All remaining blocks
5641   // are swept up, with coalescing on-the-fly as we sweep up
5642   // contiguous free and/or garbage blocks:
5643   // We need to ensure that the sweeper synchronizes with allocators
5644   // and stop-the-world collectors. In particular, the following
5645   // locks are used:
5646   // . CMS token: if this is held, a stop the world collection cannot occur
5647   // . freelistLock: if this is held no allocation can occur from this
5648   //                 generation by another thread
5649   // . bitMapLock: if this is held, no other thread can access or update
5650   //
5651 
5652   // Note that we need to hold the freelistLock if we use
5653   // block iterate below; else the iterator might go awry if
5654   // a mutator (or promotion) causes block contents to change
5655   // (for instance if the allocator divvies up a block).
5656   // If we hold the free list lock, for all practical purposes
5657   // young generation GC's can't occur (they'll usually need to
5658   // promote), so we might as well prevent all young generation
5659   // GC's while we do a sweeping step. For the same reason, we might
5660   // as well take the bit map lock for the entire duration
5661 
5662   // check that we hold the requisite locks
5663   assert(have_cms_token(), "Should hold cms token");
5664   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5665   assert_lock_strong(old_gen->freelistLock());
5666   assert_lock_strong(bitMapLock());
5667 
5668   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5669   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5670   old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5671                                           _inter_sweep_estimate.padded_average(),
5672                                           _intra_sweep_estimate.padded_average());
5673   old_gen->setNearLargestChunk();
5674 
5675   {
5676     SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5677     old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5678     // We need to free-up/coalesce garbage/blocks from a
5679     // co-terminal free run. This is done in the SweepClosure
5680     // destructor; so, do not remove this scope, else the
5681     // end-of-sweep-census below will be off by a little bit.
5682   }
5683   old_gen->cmsSpace()->sweep_completed();
5684   old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5685   if (should_unload_classes()) {                // unloaded classes this cycle,
5686     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5687   } else {                                      // did not unload classes,
5688     _concurrent_cycles_since_last_unload++;     // ... increment count
5689   }
5690 }
5691 
5692 // Reset CMS data structures (for now just the marking bit map)
5693 // preparatory for the next cycle.
5694 void CMSCollector::reset_concurrent() {
5695   CMSTokenSyncWithLocks ts(true, bitMapLock());
5696 
5697   // If the state is not "Resetting", the foreground  thread
5698   // has done a collection and the resetting.
5699   if (_collectorState != Resetting) {
5700     assert(_collectorState == Idling, "The state should only change"
5701       " because the foreground collector has finished the collection");
5702     return;
5703   }
5704 
5705   // Clear the mark bitmap (no grey objects to start with)
5706   // for the next cycle.
5707   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5708   CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
5709 
5710   HeapWord* curAddr = _markBitMap.startWord();
5711   while (curAddr < _markBitMap.endWord()) {
5712     size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5713     MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5714     _markBitMap.clear_large_range(chunk);
5715     if (ConcurrentMarkSweepThread::should_yield() &&
5716         !foregroundGCIsActive() &&
5717         CMSYield) {
5718       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5719              "CMS thread should hold CMS token");
5720       assert_lock_strong(bitMapLock());
5721       bitMapLock()->unlock();
5722       ConcurrentMarkSweepThread::desynchronize(true);
5723       stopTimer();
5724       if (PrintCMSStatistics != 0) {
5725         incrementYields();
5726       }
5727 
5728       // See the comment in coordinator_yield()
5729       for (unsigned i = 0; i < CMSYieldSleepCount &&
5730                        ConcurrentMarkSweepThread::should_yield() &&
5731                        !CMSCollector::foregroundGCIsActive(); ++i) {
5732         os::sleep(Thread::current(), 1, false);
5733       }
5734 
5735       ConcurrentMarkSweepThread::synchronize(true);
5736       bitMapLock()->lock_without_safepoint_check();
5737       startTimer();
5738     }
5739     curAddr = chunk.end();
5740   }
5741   // A successful mostly concurrent collection has been done.
5742   // Because only the full (i.e., concurrent mode failure) collections
5743   // are being measured for gc overhead limits, clean the "near" flag
5744   // and count.
5745   size_policy()->reset_gc_overhead_limit_count();
5746   _collectorState = Idling;
5747 
5748   register_gc_end();
5749 }
5750 
5751 // Same as above but for STW paths
5752 void CMSCollector::reset_stw() {
5753   // already have the lock
5754   assert(_collectorState == Resetting, "just checking");
5755   assert_lock_strong(bitMapLock());
5756   GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5757   _markBitMap.clear_all();
5758   _collectorState = Idling;
5759   register_gc_end();
5760 }
5761 
5762 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5763   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5764   GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
5765   TraceCollectorStats tcs(counters());
5766 
5767   switch (op) {
5768     case CMS_op_checkpointRootsInitial: {
5769       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5770       checkpointRootsInitial();
5771       if (PrintGC) {
5772         _cmsGen->printOccupancy("initial-mark");
5773       }
5774       break;
5775     }
5776     case CMS_op_checkpointRootsFinal: {
5777       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5778       checkpointRootsFinal();
5779       if (PrintGC) {
5780         _cmsGen->printOccupancy("remark");
5781       }
5782       break;
5783     }
5784     default:
5785       fatal("No such CMS_op");
5786   }
5787 }
5788 
5789 #ifndef PRODUCT
5790 size_t const CMSCollector::skip_header_HeapWords() {
5791   return FreeChunk::header_size();
5792 }
5793 
5794 // Try and collect here conditions that should hold when
5795 // CMS thread is exiting. The idea is that the foreground GC
5796 // thread should not be blocked if it wants to terminate
5797 // the CMS thread and yet continue to run the VM for a while
5798 // after that.
5799 void CMSCollector::verify_ok_to_terminate() const {
5800   assert(Thread::current()->is_ConcurrentGC_thread(),
5801          "should be called by CMS thread");
5802   assert(!_foregroundGCShouldWait, "should be false");
5803   // We could check here that all the various low-level locks
5804   // are not held by the CMS thread, but that is overkill; see
5805   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5806   // is checked.
5807 }
5808 #endif
5809 
5810 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5811    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5812           "missing Printezis mark?");
5813   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5814   size_t size = pointer_delta(nextOneAddr + 1, addr);
5815   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5816          "alignment problem");
5817   assert(size >= 3, "Necessary for Printezis marks to work");
5818   return size;
5819 }
5820 
5821 // A variant of the above (block_size_using_printezis_bits()) except
5822 // that we return 0 if the P-bits are not yet set.
5823 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5824   if (_markBitMap.isMarked(addr + 1)) {
5825     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5826     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5827     size_t size = pointer_delta(nextOneAddr + 1, addr);
5828     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5829            "alignment problem");
5830     assert(size >= 3, "Necessary for Printezis marks to work");
5831     return size;
5832   }
5833   return 0;
5834 }
5835 
5836 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5837   size_t sz = 0;
5838   oop p = (oop)addr;
5839   if (p->klass_or_null() != NULL) {
5840     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5841   } else {
5842     sz = block_size_using_printezis_bits(addr);
5843   }
5844   assert(sz > 0, "size must be nonzero");
5845   HeapWord* next_block = addr + sz;
5846   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
5847                                              CardTableModRefBS::card_size);
5848   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
5849          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5850          "must be different cards");
5851   return next_card;
5852 }
5853 
5854 
5855 // CMS Bit Map Wrapper /////////////////////////////////////////
5856 
5857 // Construct a CMS bit map infrastructure, but don't create the
5858 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5859 // further below.
5860 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5861   _bm(),
5862   _shifter(shifter),
5863   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5864                                     Monitor::_safepoint_check_sometimes) : NULL)
5865 {
5866   _bmStartWord = 0;
5867   _bmWordSize  = 0;
5868 }
5869 
5870 bool CMSBitMap::allocate(MemRegion mr) {
5871   _bmStartWord = mr.start();
5872   _bmWordSize  = mr.word_size();
5873   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5874                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5875   if (!brs.is_reserved()) {
5876     warning("CMS bit map allocation failure");
5877     return false;
5878   }
5879   // For now we'll just commit all of the bit map up front.
5880   // Later on we'll try to be more parsimonious with swap.
5881   if (!_virtual_space.initialize(brs, brs.size())) {
5882     warning("CMS bit map backing store failure");
5883     return false;
5884   }
5885   assert(_virtual_space.committed_size() == brs.size(),
5886          "didn't reserve backing store for all of CMS bit map?");
5887   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5888   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5889          _bmWordSize, "inconsistency in bit map sizing");
5890   _bm.set_size(_bmWordSize >> _shifter);
5891 
5892   // bm.clear(); // can we rely on getting zero'd memory? verify below
5893   assert(isAllClear(),
5894          "Expected zero'd memory from ReservedSpace constructor");
5895   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5896          "consistency check");
5897   return true;
5898 }
5899 
5900 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5901   HeapWord *next_addr, *end_addr, *last_addr;
5902   assert_locked();
5903   assert(covers(mr), "out-of-range error");
5904   // XXX assert that start and end are appropriately aligned
5905   for (next_addr = mr.start(), end_addr = mr.end();
5906        next_addr < end_addr; next_addr = last_addr) {
5907     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5908     last_addr = dirty_region.end();
5909     if (!dirty_region.is_empty()) {
5910       cl->do_MemRegion(dirty_region);
5911     } else {
5912       assert(last_addr == end_addr, "program logic");
5913       return;
5914     }
5915   }
5916 }
5917 
5918 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5919   _bm.print_on_error(st, prefix);
5920 }
5921 
5922 #ifndef PRODUCT
5923 void CMSBitMap::assert_locked() const {
5924   CMSLockVerifier::assert_locked(lock());
5925 }
5926 
5927 bool CMSBitMap::covers(MemRegion mr) const {
5928   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5929   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5930          "size inconsistency");
5931   return (mr.start() >= _bmStartWord) &&
5932          (mr.end()   <= endWord());
5933 }
5934 
5935 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5936     return (start >= _bmStartWord && (start + size) <= endWord());
5937 }
5938 
5939 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5940   // verify that there are no 1 bits in the interval [left, right)
5941   FalseBitMapClosure falseBitMapClosure;
5942   iterate(&falseBitMapClosure, left, right);
5943 }
5944 
5945 void CMSBitMap::region_invariant(MemRegion mr)
5946 {
5947   assert_locked();
5948   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5949   assert(!mr.is_empty(), "unexpected empty region");
5950   assert(covers(mr), "mr should be covered by bit map");
5951   // convert address range into offset range
5952   size_t start_ofs = heapWordToOffset(mr.start());
5953   // Make sure that end() is appropriately aligned
5954   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5955                         (1 << (_shifter+LogHeapWordSize))),
5956          "Misaligned mr.end()");
5957   size_t end_ofs   = heapWordToOffset(mr.end());
5958   assert(end_ofs > start_ofs, "Should mark at least one bit");
5959 }
5960 
5961 #endif
5962 
5963 bool CMSMarkStack::allocate(size_t size) {
5964   // allocate a stack of the requisite depth
5965   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5966                    size * sizeof(oop)));
5967   if (!rs.is_reserved()) {
5968     warning("CMSMarkStack allocation failure");
5969     return false;
5970   }
5971   if (!_virtual_space.initialize(rs, rs.size())) {
5972     warning("CMSMarkStack backing store failure");
5973     return false;
5974   }
5975   assert(_virtual_space.committed_size() == rs.size(),
5976          "didn't reserve backing store for all of CMS stack?");
5977   _base = (oop*)(_virtual_space.low());
5978   _index = 0;
5979   _capacity = size;
5980   NOT_PRODUCT(_max_depth = 0);
5981   return true;
5982 }
5983 
5984 // XXX FIX ME !!! In the MT case we come in here holding a
5985 // leaf lock. For printing we need to take a further lock
5986 // which has lower rank. We need to recalibrate the two
5987 // lock-ranks involved in order to be able to print the
5988 // messages below. (Or defer the printing to the caller.
5989 // For now we take the expedient path of just disabling the
5990 // messages for the problematic case.)
5991 void CMSMarkStack::expand() {
5992   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5993   if (_capacity == MarkStackSizeMax) {
5994     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
5995       // We print a warning message only once per CMS cycle.
5996       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
5997     }
5998     return;
5999   }
6000   // Double capacity if possible
6001   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6002   // Do not give up existing stack until we have managed to
6003   // get the double capacity that we desired.
6004   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6005                    new_capacity * sizeof(oop)));
6006   if (rs.is_reserved()) {
6007     // Release the backing store associated with old stack
6008     _virtual_space.release();
6009     // Reinitialize virtual space for new stack
6010     if (!_virtual_space.initialize(rs, rs.size())) {
6011       fatal("Not enough swap for expanded marking stack");
6012     }
6013     _base = (oop*)(_virtual_space.low());
6014     _index = 0;
6015     _capacity = new_capacity;
6016   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6017     // Failed to double capacity, continue;
6018     // we print a detail message only once per CMS cycle.
6019     gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6020             SIZE_FORMAT "K",
6021             _capacity / K, new_capacity / K);
6022   }
6023 }
6024 
6025 
6026 // Closures
6027 // XXX: there seems to be a lot of code  duplication here;
6028 // should refactor and consolidate common code.
6029 
6030 // This closure is used to mark refs into the CMS generation in
6031 // the CMS bit map. Called at the first checkpoint. This closure
6032 // assumes that we do not need to re-mark dirty cards; if the CMS
6033 // generation on which this is used is not an oldest
6034 // generation then this will lose younger_gen cards!
6035 
6036 MarkRefsIntoClosure::MarkRefsIntoClosure(
6037   MemRegion span, CMSBitMap* bitMap):
6038     _span(span),
6039     _bitMap(bitMap)
6040 {
6041   assert(ref_processor() == NULL, "deliberately left NULL");
6042   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6043 }
6044 
6045 void MarkRefsIntoClosure::do_oop(oop obj) {
6046   // if p points into _span, then mark corresponding bit in _markBitMap
6047   assert(obj->is_oop(), "expected an oop");
6048   HeapWord* addr = (HeapWord*)obj;
6049   if (_span.contains(addr)) {
6050     // this should be made more efficient
6051     _bitMap->mark(addr);
6052   }
6053 }
6054 
6055 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
6056 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6057 
6058 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6059   MemRegion span, CMSBitMap* bitMap):
6060     _span(span),
6061     _bitMap(bitMap)
6062 {
6063   assert(ref_processor() == NULL, "deliberately left NULL");
6064   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6065 }
6066 
6067 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6068   // if p points into _span, then mark corresponding bit in _markBitMap
6069   assert(obj->is_oop(), "expected an oop");
6070   HeapWord* addr = (HeapWord*)obj;
6071   if (_span.contains(addr)) {
6072     // this should be made more efficient
6073     _bitMap->par_mark(addr);
6074   }
6075 }
6076 
6077 void Par_MarkRefsIntoClosure::do_oop(oop* p)       { Par_MarkRefsIntoClosure::do_oop_work(p); }
6078 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6079 
6080 // A variant of the above, used for CMS marking verification.
6081 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6082   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6083     _span(span),
6084     _verification_bm(verification_bm),
6085     _cms_bm(cms_bm)
6086 {
6087   assert(ref_processor() == NULL, "deliberately left NULL");
6088   assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6089 }
6090 
6091 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6092   // if p points into _span, then mark corresponding bit in _markBitMap
6093   assert(obj->is_oop(), "expected an oop");
6094   HeapWord* addr = (HeapWord*)obj;
6095   if (_span.contains(addr)) {
6096     _verification_bm->mark(addr);
6097     if (!_cms_bm->isMarked(addr)) {
6098       oop(addr)->print();
6099       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6100       fatal("... aborting");
6101     }
6102   }
6103 }
6104 
6105 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6106 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6107 
6108 //////////////////////////////////////////////////
6109 // MarkRefsIntoAndScanClosure
6110 //////////////////////////////////////////////////
6111 
6112 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6113                                                        ReferenceProcessor* rp,
6114                                                        CMSBitMap* bit_map,
6115                                                        CMSBitMap* mod_union_table,
6116                                                        CMSMarkStack*  mark_stack,
6117                                                        CMSCollector* collector,
6118                                                        bool should_yield,
6119                                                        bool concurrent_precleaning):
6120   _collector(collector),
6121   _span(span),
6122   _bit_map(bit_map),
6123   _mark_stack(mark_stack),
6124   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6125                       mark_stack, concurrent_precleaning),
6126   _yield(should_yield),
6127   _concurrent_precleaning(concurrent_precleaning),
6128   _freelistLock(NULL)
6129 {
6130   // FIXME: Should initialize in base class constructor.
6131   assert(rp != NULL, "ref_processor shouldn't be NULL");
6132   set_ref_processor_internal(rp);
6133 }
6134 
6135 // This closure is used to mark refs into the CMS generation at the
6136 // second (final) checkpoint, and to scan and transitively follow
6137 // the unmarked oops. It is also used during the concurrent precleaning
6138 // phase while scanning objects on dirty cards in the CMS generation.
6139 // The marks are made in the marking bit map and the marking stack is
6140 // used for keeping the (newly) grey objects during the scan.
6141 // The parallel version (Par_...) appears further below.
6142 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6143   if (obj != NULL) {
6144     assert(obj->is_oop(), "expected an oop");
6145     HeapWord* addr = (HeapWord*)obj;
6146     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6147     assert(_collector->overflow_list_is_empty(),
6148            "overflow list should be empty");
6149     if (_span.contains(addr) &&
6150         !_bit_map->isMarked(addr)) {
6151       // mark bit map (object is now grey)
6152       _bit_map->mark(addr);
6153       // push on marking stack (stack should be empty), and drain the
6154       // stack by applying this closure to the oops in the oops popped
6155       // from the stack (i.e. blacken the grey objects)
6156       bool res = _mark_stack->push(obj);
6157       assert(res, "Should have space to push on empty stack");
6158       do {
6159         oop new_oop = _mark_stack->pop();
6160         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6161         assert(_bit_map->isMarked((HeapWord*)new_oop),
6162                "only grey objects on this stack");
6163         // iterate over the oops in this oop, marking and pushing
6164         // the ones in CMS heap (i.e. in _span).
6165         new_oop->oop_iterate(&_pushAndMarkClosure);
6166         // check if it's time to yield
6167         do_yield_check();
6168       } while (!_mark_stack->isEmpty() ||
6169                (!_concurrent_precleaning && take_from_overflow_list()));
6170         // if marking stack is empty, and we are not doing this
6171         // during precleaning, then check the overflow list
6172     }
6173     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6174     assert(_collector->overflow_list_is_empty(),
6175            "overflow list was drained above");
6176     // We could restore evacuated mark words, if any, used for
6177     // overflow list links here because the overflow list is
6178     // provably empty here. That would reduce the maximum
6179     // size requirements for preserved_{oop,mark}_stack.
6180     // But we'll just postpone it until we are all done
6181     // so we can just stream through.
6182     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6183       _collector->restore_preserved_marks_if_any();
6184       assert(_collector->no_preserved_marks(), "No preserved marks");
6185     }
6186     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6187            "All preserved marks should have been restored above");
6188   }
6189 }
6190 
6191 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6192 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6193 
6194 void MarkRefsIntoAndScanClosure::do_yield_work() {
6195   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6196          "CMS thread should hold CMS token");
6197   assert_lock_strong(_freelistLock);
6198   assert_lock_strong(_bit_map->lock());
6199   // relinquish the free_list_lock and bitMaplock()
6200   _bit_map->lock()->unlock();
6201   _freelistLock->unlock();
6202   ConcurrentMarkSweepThread::desynchronize(true);
6203   _collector->stopTimer();
6204   if (PrintCMSStatistics != 0) {
6205     _collector->incrementYields();
6206   }
6207 
6208   // See the comment in coordinator_yield()
6209   for (unsigned i = 0;
6210        i < CMSYieldSleepCount &&
6211        ConcurrentMarkSweepThread::should_yield() &&
6212        !CMSCollector::foregroundGCIsActive();
6213        ++i) {
6214     os::sleep(Thread::current(), 1, false);
6215   }
6216 
6217   ConcurrentMarkSweepThread::synchronize(true);
6218   _freelistLock->lock_without_safepoint_check();
6219   _bit_map->lock()->lock_without_safepoint_check();
6220   _collector->startTimer();
6221 }
6222 
6223 ///////////////////////////////////////////////////////////
6224 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6225 //                                 MarkRefsIntoAndScanClosure
6226 ///////////////////////////////////////////////////////////
6227 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6228   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6229   CMSBitMap* bit_map, OopTaskQueue* work_queue):
6230   _span(span),
6231   _bit_map(bit_map),
6232   _work_queue(work_queue),
6233   _low_water_mark(MIN2((work_queue->max_elems()/4),
6234                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6235   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6236 {
6237   // FIXME: Should initialize in base class constructor.
6238   assert(rp != NULL, "ref_processor shouldn't be NULL");
6239   set_ref_processor_internal(rp);
6240 }
6241 
6242 // This closure is used to mark refs into the CMS generation at the
6243 // second (final) checkpoint, and to scan and transitively follow
6244 // the unmarked oops. The marks are made in the marking bit map and
6245 // the work_queue is used for keeping the (newly) grey objects during
6246 // the scan phase whence they are also available for stealing by parallel
6247 // threads. Since the marking bit map is shared, updates are
6248 // synchronized (via CAS).
6249 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6250   if (obj != NULL) {
6251     // Ignore mark word because this could be an already marked oop
6252     // that may be chained at the end of the overflow list.
6253     assert(obj->is_oop(true), "expected an oop");
6254     HeapWord* addr = (HeapWord*)obj;
6255     if (_span.contains(addr) &&
6256         !_bit_map->isMarked(addr)) {
6257       // mark bit map (object will become grey):
6258       // It is possible for several threads to be
6259       // trying to "claim" this object concurrently;
6260       // the unique thread that succeeds in marking the
6261       // object first will do the subsequent push on
6262       // to the work queue (or overflow list).
6263       if (_bit_map->par_mark(addr)) {
6264         // push on work_queue (which may not be empty), and trim the
6265         // queue to an appropriate length by applying this closure to
6266         // the oops in the oops popped from the stack (i.e. blacken the
6267         // grey objects)
6268         bool res = _work_queue->push(obj);
6269         assert(res, "Low water mark should be less than capacity?");
6270         trim_queue(_low_water_mark);
6271       } // Else, another thread claimed the object
6272     }
6273   }
6274 }
6275 
6276 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6277 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6278 
6279 // This closure is used to rescan the marked objects on the dirty cards
6280 // in the mod union table and the card table proper.
6281 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6282   oop p, MemRegion mr) {
6283 
6284   size_t size = 0;
6285   HeapWord* addr = (HeapWord*)p;
6286   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6287   assert(_span.contains(addr), "we are scanning the CMS generation");
6288   // check if it's time to yield
6289   if (do_yield_check()) {
6290     // We yielded for some foreground stop-world work,
6291     // and we have been asked to abort this ongoing preclean cycle.
6292     return 0;
6293   }
6294   if (_bitMap->isMarked(addr)) {
6295     // it's marked; is it potentially uninitialized?
6296     if (p->klass_or_null() != NULL) {
6297         // an initialized object; ignore mark word in verification below
6298         // since we are running concurrent with mutators
6299         assert(p->is_oop(true), "should be an oop");
6300         if (p->is_objArray()) {
6301           // objArrays are precisely marked; restrict scanning
6302           // to dirty cards only.
6303           size = CompactibleFreeListSpace::adjustObjectSize(
6304                    p->oop_iterate_size(_scanningClosure, mr));
6305         } else {
6306           // A non-array may have been imprecisely marked; we need
6307           // to scan object in its entirety.
6308           size = CompactibleFreeListSpace::adjustObjectSize(
6309                    p->oop_iterate_size(_scanningClosure));
6310         }
6311         #ifdef ASSERT
6312           size_t direct_size =
6313             CompactibleFreeListSpace::adjustObjectSize(p->size());
6314           assert(size == direct_size, "Inconsistency in size");
6315           assert(size >= 3, "Necessary for Printezis marks to work");
6316           if (!_bitMap->isMarked(addr+1)) {
6317             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6318           } else {
6319             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6320             assert(_bitMap->isMarked(addr+size-1),
6321                    "inconsistent Printezis mark");
6322           }
6323         #endif // ASSERT
6324     } else {
6325       // An uninitialized object.
6326       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6327       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6328       size = pointer_delta(nextOneAddr + 1, addr);
6329       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6330              "alignment problem");
6331       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6332       // will dirty the card when the klass pointer is installed in the
6333       // object (signaling the completion of initialization).
6334     }
6335   } else {
6336     // Either a not yet marked object or an uninitialized object
6337     if (p->klass_or_null() == NULL) {
6338       // An uninitialized object, skip to the next card, since
6339       // we may not be able to read its P-bits yet.
6340       assert(size == 0, "Initial value");
6341     } else {
6342       // An object not (yet) reached by marking: we merely need to
6343       // compute its size so as to go look at the next block.
6344       assert(p->is_oop(true), "should be an oop");
6345       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6346     }
6347   }
6348   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6349   return size;
6350 }
6351 
6352 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6353   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6354          "CMS thread should hold CMS token");
6355   assert_lock_strong(_freelistLock);
6356   assert_lock_strong(_bitMap->lock());
6357   // relinquish the free_list_lock and bitMaplock()
6358   _bitMap->lock()->unlock();
6359   _freelistLock->unlock();
6360   ConcurrentMarkSweepThread::desynchronize(true);
6361   _collector->stopTimer();
6362   if (PrintCMSStatistics != 0) {
6363     _collector->incrementYields();
6364   }
6365 
6366   // See the comment in coordinator_yield()
6367   for (unsigned i = 0; i < CMSYieldSleepCount &&
6368                    ConcurrentMarkSweepThread::should_yield() &&
6369                    !CMSCollector::foregroundGCIsActive(); ++i) {
6370     os::sleep(Thread::current(), 1, false);
6371   }
6372 
6373   ConcurrentMarkSweepThread::synchronize(true);
6374   _freelistLock->lock_without_safepoint_check();
6375   _bitMap->lock()->lock_without_safepoint_check();
6376   _collector->startTimer();
6377 }
6378 
6379 
6380 //////////////////////////////////////////////////////////////////
6381 // SurvivorSpacePrecleanClosure
6382 //////////////////////////////////////////////////////////////////
6383 // This (single-threaded) closure is used to preclean the oops in
6384 // the survivor spaces.
6385 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6386 
6387   HeapWord* addr = (HeapWord*)p;
6388   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6389   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6390   assert(p->klass_or_null() != NULL, "object should be initialized");
6391   // an initialized object; ignore mark word in verification below
6392   // since we are running concurrent with mutators
6393   assert(p->is_oop(true), "should be an oop");
6394   // Note that we do not yield while we iterate over
6395   // the interior oops of p, pushing the relevant ones
6396   // on our marking stack.
6397   size_t size = p->oop_iterate_size(_scanning_closure);
6398   do_yield_check();
6399   // Observe that below, we do not abandon the preclean
6400   // phase as soon as we should; rather we empty the
6401   // marking stack before returning. This is to satisfy
6402   // some existing assertions. In general, it may be a
6403   // good idea to abort immediately and complete the marking
6404   // from the grey objects at a later time.
6405   while (!_mark_stack->isEmpty()) {
6406     oop new_oop = _mark_stack->pop();
6407     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6408     assert(_bit_map->isMarked((HeapWord*)new_oop),
6409            "only grey objects on this stack");
6410     // iterate over the oops in this oop, marking and pushing
6411     // the ones in CMS heap (i.e. in _span).
6412     new_oop->oop_iterate(_scanning_closure);
6413     // check if it's time to yield
6414     do_yield_check();
6415   }
6416   unsigned int after_count =
6417     GenCollectedHeap::heap()->total_collections();
6418   bool abort = (_before_count != after_count) ||
6419                _collector->should_abort_preclean();
6420   return abort ? 0 : size;
6421 }
6422 
6423 void SurvivorSpacePrecleanClosure::do_yield_work() {
6424   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6425          "CMS thread should hold CMS token");
6426   assert_lock_strong(_bit_map->lock());
6427   // Relinquish the bit map lock
6428   _bit_map->lock()->unlock();
6429   ConcurrentMarkSweepThread::desynchronize(true);
6430   _collector->stopTimer();
6431   if (PrintCMSStatistics != 0) {
6432     _collector->incrementYields();
6433   }
6434 
6435   // See the comment in coordinator_yield()
6436   for (unsigned i = 0; i < CMSYieldSleepCount &&
6437                        ConcurrentMarkSweepThread::should_yield() &&
6438                        !CMSCollector::foregroundGCIsActive(); ++i) {
6439     os::sleep(Thread::current(), 1, false);
6440   }
6441 
6442   ConcurrentMarkSweepThread::synchronize(true);
6443   _bit_map->lock()->lock_without_safepoint_check();
6444   _collector->startTimer();
6445 }
6446 
6447 // This closure is used to rescan the marked objects on the dirty cards
6448 // in the mod union table and the card table proper. In the parallel
6449 // case, although the bitMap is shared, we do a single read so the
6450 // isMarked() query is "safe".
6451 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6452   // Ignore mark word because we are running concurrent with mutators
6453   assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6454   HeapWord* addr = (HeapWord*)p;
6455   assert(_span.contains(addr), "we are scanning the CMS generation");
6456   bool is_obj_array = false;
6457   #ifdef ASSERT
6458     if (!_parallel) {
6459       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6460       assert(_collector->overflow_list_is_empty(),
6461              "overflow list should be empty");
6462 
6463     }
6464   #endif // ASSERT
6465   if (_bit_map->isMarked(addr)) {
6466     // Obj arrays are precisely marked, non-arrays are not;
6467     // so we scan objArrays precisely and non-arrays in their
6468     // entirety.
6469     if (p->is_objArray()) {
6470       is_obj_array = true;
6471       if (_parallel) {
6472         p->oop_iterate(_par_scan_closure, mr);
6473       } else {
6474         p->oop_iterate(_scan_closure, mr);
6475       }
6476     } else {
6477       if (_parallel) {
6478         p->oop_iterate(_par_scan_closure);
6479       } else {
6480         p->oop_iterate(_scan_closure);
6481       }
6482     }
6483   }
6484   #ifdef ASSERT
6485     if (!_parallel) {
6486       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6487       assert(_collector->overflow_list_is_empty(),
6488              "overflow list should be empty");
6489 
6490     }
6491   #endif // ASSERT
6492   return is_obj_array;
6493 }
6494 
6495 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6496                         MemRegion span,
6497                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6498                         bool should_yield, bool verifying):
6499   _collector(collector),
6500   _span(span),
6501   _bitMap(bitMap),
6502   _mut(&collector->_modUnionTable),
6503   _markStack(markStack),
6504   _yield(should_yield),
6505   _skipBits(0)
6506 {
6507   assert(_markStack->isEmpty(), "stack should be empty");
6508   _finger = _bitMap->startWord();
6509   _threshold = _finger;
6510   assert(_collector->_restart_addr == NULL, "Sanity check");
6511   assert(_span.contains(_finger), "Out of bounds _finger?");
6512   DEBUG_ONLY(_verifying = verifying;)
6513 }
6514 
6515 void MarkFromRootsClosure::reset(HeapWord* addr) {
6516   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6517   assert(_span.contains(addr), "Out of bounds _finger?");
6518   _finger = addr;
6519   _threshold = (HeapWord*)round_to(
6520                  (intptr_t)_finger, CardTableModRefBS::card_size);
6521 }
6522 
6523 // Should revisit to see if this should be restructured for
6524 // greater efficiency.
6525 bool MarkFromRootsClosure::do_bit(size_t offset) {
6526   if (_skipBits > 0) {
6527     _skipBits--;
6528     return true;
6529   }
6530   // convert offset into a HeapWord*
6531   HeapWord* addr = _bitMap->startWord() + offset;
6532   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6533          "address out of range");
6534   assert(_bitMap->isMarked(addr), "tautology");
6535   if (_bitMap->isMarked(addr+1)) {
6536     // this is an allocated but not yet initialized object
6537     assert(_skipBits == 0, "tautology");
6538     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6539     oop p = oop(addr);
6540     if (p->klass_or_null() == NULL) {
6541       DEBUG_ONLY(if (!_verifying) {)
6542         // We re-dirty the cards on which this object lies and increase
6543         // the _threshold so that we'll come back to scan this object
6544         // during the preclean or remark phase. (CMSCleanOnEnter)
6545         if (CMSCleanOnEnter) {
6546           size_t sz = _collector->block_size_using_printezis_bits(addr);
6547           HeapWord* end_card_addr   = (HeapWord*)round_to(
6548                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6549           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6550           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6551           // Bump _threshold to end_card_addr; note that
6552           // _threshold cannot possibly exceed end_card_addr, anyhow.
6553           // This prevents future clearing of the card as the scan proceeds
6554           // to the right.
6555           assert(_threshold <= end_card_addr,
6556                  "Because we are just scanning into this object");
6557           if (_threshold < end_card_addr) {
6558             _threshold = end_card_addr;
6559           }
6560           if (p->klass_or_null() != NULL) {
6561             // Redirty the range of cards...
6562             _mut->mark_range(redirty_range);
6563           } // ...else the setting of klass will dirty the card anyway.
6564         }
6565       DEBUG_ONLY(})
6566       return true;
6567     }
6568   }
6569   scanOopsInOop(addr);
6570   return true;
6571 }
6572 
6573 // We take a break if we've been at this for a while,
6574 // so as to avoid monopolizing the locks involved.
6575 void MarkFromRootsClosure::do_yield_work() {
6576   // First give up the locks, then yield, then re-lock
6577   // We should probably use a constructor/destructor idiom to
6578   // do this unlock/lock or modify the MutexUnlocker class to
6579   // serve our purpose. XXX
6580   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6581          "CMS thread should hold CMS token");
6582   assert_lock_strong(_bitMap->lock());
6583   _bitMap->lock()->unlock();
6584   ConcurrentMarkSweepThread::desynchronize(true);
6585   _collector->stopTimer();
6586   if (PrintCMSStatistics != 0) {
6587     _collector->incrementYields();
6588   }
6589 
6590   // See the comment in coordinator_yield()
6591   for (unsigned i = 0; i < CMSYieldSleepCount &&
6592                        ConcurrentMarkSweepThread::should_yield() &&
6593                        !CMSCollector::foregroundGCIsActive(); ++i) {
6594     os::sleep(Thread::current(), 1, false);
6595   }
6596 
6597   ConcurrentMarkSweepThread::synchronize(true);
6598   _bitMap->lock()->lock_without_safepoint_check();
6599   _collector->startTimer();
6600 }
6601 
6602 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6603   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6604   assert(_markStack->isEmpty(),
6605          "should drain stack to limit stack usage");
6606   // convert ptr to an oop preparatory to scanning
6607   oop obj = oop(ptr);
6608   // Ignore mark word in verification below, since we
6609   // may be running concurrent with mutators.
6610   assert(obj->is_oop(true), "should be an oop");
6611   assert(_finger <= ptr, "_finger runneth ahead");
6612   // advance the finger to right end of this object
6613   _finger = ptr + obj->size();
6614   assert(_finger > ptr, "we just incremented it above");
6615   // On large heaps, it may take us some time to get through
6616   // the marking phase. During
6617   // this time it's possible that a lot of mutations have
6618   // accumulated in the card table and the mod union table --
6619   // these mutation records are redundant until we have
6620   // actually traced into the corresponding card.
6621   // Here, we check whether advancing the finger would make
6622   // us cross into a new card, and if so clear corresponding
6623   // cards in the MUT (preclean them in the card-table in the
6624   // future).
6625 
6626   DEBUG_ONLY(if (!_verifying) {)
6627     // The clean-on-enter optimization is disabled by default,
6628     // until we fix 6178663.
6629     if (CMSCleanOnEnter && (_finger > _threshold)) {
6630       // [_threshold, _finger) represents the interval
6631       // of cards to be cleared  in MUT (or precleaned in card table).
6632       // The set of cards to be cleared is all those that overlap
6633       // with the interval [_threshold, _finger); note that
6634       // _threshold is always kept card-aligned but _finger isn't
6635       // always card-aligned.
6636       HeapWord* old_threshold = _threshold;
6637       assert(old_threshold == (HeapWord*)round_to(
6638               (intptr_t)old_threshold, CardTableModRefBS::card_size),
6639              "_threshold should always be card-aligned");
6640       _threshold = (HeapWord*)round_to(
6641                      (intptr_t)_finger, CardTableModRefBS::card_size);
6642       MemRegion mr(old_threshold, _threshold);
6643       assert(!mr.is_empty(), "Control point invariant");
6644       assert(_span.contains(mr), "Should clear within span");
6645       _mut->clear_range(mr);
6646     }
6647   DEBUG_ONLY(})
6648   // Note: the finger doesn't advance while we drain
6649   // the stack below.
6650   PushOrMarkClosure pushOrMarkClosure(_collector,
6651                                       _span, _bitMap, _markStack,
6652                                       _finger, this);
6653   bool res = _markStack->push(obj);
6654   assert(res, "Empty non-zero size stack should have space for single push");
6655   while (!_markStack->isEmpty()) {
6656     oop new_oop = _markStack->pop();
6657     // Skip verifying header mark word below because we are
6658     // running concurrent with mutators.
6659     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6660     // now scan this oop's oops
6661     new_oop->oop_iterate(&pushOrMarkClosure);
6662     do_yield_check();
6663   }
6664   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6665 }
6666 
6667 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6668                        CMSCollector* collector, MemRegion span,
6669                        CMSBitMap* bit_map,
6670                        OopTaskQueue* work_queue,
6671                        CMSMarkStack*  overflow_stack):
6672   _collector(collector),
6673   _whole_span(collector->_span),
6674   _span(span),
6675   _bit_map(bit_map),
6676   _mut(&collector->_modUnionTable),
6677   _work_queue(work_queue),
6678   _overflow_stack(overflow_stack),
6679   _skip_bits(0),
6680   _task(task)
6681 {
6682   assert(_work_queue->size() == 0, "work_queue should be empty");
6683   _finger = span.start();
6684   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6685   assert(_span.contains(_finger), "Out of bounds _finger?");
6686 }
6687 
6688 // Should revisit to see if this should be restructured for
6689 // greater efficiency.
6690 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6691   if (_skip_bits > 0) {
6692     _skip_bits--;
6693     return true;
6694   }
6695   // convert offset into a HeapWord*
6696   HeapWord* addr = _bit_map->startWord() + offset;
6697   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6698          "address out of range");
6699   assert(_bit_map->isMarked(addr), "tautology");
6700   if (_bit_map->isMarked(addr+1)) {
6701     // this is an allocated object that might not yet be initialized
6702     assert(_skip_bits == 0, "tautology");
6703     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6704     oop p = oop(addr);
6705     if (p->klass_or_null() == NULL) {
6706       // in the case of Clean-on-Enter optimization, redirty card
6707       // and avoid clearing card by increasing  the threshold.
6708       return true;
6709     }
6710   }
6711   scan_oops_in_oop(addr);
6712   return true;
6713 }
6714 
6715 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6716   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6717   // Should we assert that our work queue is empty or
6718   // below some drain limit?
6719   assert(_work_queue->size() == 0,
6720          "should drain stack to limit stack usage");
6721   // convert ptr to an oop preparatory to scanning
6722   oop obj = oop(ptr);
6723   // Ignore mark word in verification below, since we
6724   // may be running concurrent with mutators.
6725   assert(obj->is_oop(true), "should be an oop");
6726   assert(_finger <= ptr, "_finger runneth ahead");
6727   // advance the finger to right end of this object
6728   _finger = ptr + obj->size();
6729   assert(_finger > ptr, "we just incremented it above");
6730   // On large heaps, it may take us some time to get through
6731   // the marking phase. During
6732   // this time it's possible that a lot of mutations have
6733   // accumulated in the card table and the mod union table --
6734   // these mutation records are redundant until we have
6735   // actually traced into the corresponding card.
6736   // Here, we check whether advancing the finger would make
6737   // us cross into a new card, and if so clear corresponding
6738   // cards in the MUT (preclean them in the card-table in the
6739   // future).
6740 
6741   // The clean-on-enter optimization is disabled by default,
6742   // until we fix 6178663.
6743   if (CMSCleanOnEnter && (_finger > _threshold)) {
6744     // [_threshold, _finger) represents the interval
6745     // of cards to be cleared  in MUT (or precleaned in card table).
6746     // The set of cards to be cleared is all those that overlap
6747     // with the interval [_threshold, _finger); note that
6748     // _threshold is always kept card-aligned but _finger isn't
6749     // always card-aligned.
6750     HeapWord* old_threshold = _threshold;
6751     assert(old_threshold == (HeapWord*)round_to(
6752             (intptr_t)old_threshold, CardTableModRefBS::card_size),
6753            "_threshold should always be card-aligned");
6754     _threshold = (HeapWord*)round_to(
6755                    (intptr_t)_finger, CardTableModRefBS::card_size);
6756     MemRegion mr(old_threshold, _threshold);
6757     assert(!mr.is_empty(), "Control point invariant");
6758     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6759     _mut->clear_range(mr);
6760   }
6761 
6762   // Note: the local finger doesn't advance while we drain
6763   // the stack below, but the global finger sure can and will.
6764   HeapWord** gfa = _task->global_finger_addr();
6765   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6766                                       _span, _bit_map,
6767                                       _work_queue,
6768                                       _overflow_stack,
6769                                       _finger,
6770                                       gfa, this);
6771   bool res = _work_queue->push(obj);   // overflow could occur here
6772   assert(res, "Will hold once we use workqueues");
6773   while (true) {
6774     oop new_oop;
6775     if (!_work_queue->pop_local(new_oop)) {
6776       // We emptied our work_queue; check if there's stuff that can
6777       // be gotten from the overflow stack.
6778       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6779             _overflow_stack, _work_queue)) {
6780         do_yield_check();
6781         continue;
6782       } else {  // done
6783         break;
6784       }
6785     }
6786     // Skip verifying header mark word below because we are
6787     // running concurrent with mutators.
6788     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6789     // now scan this oop's oops
6790     new_oop->oop_iterate(&pushOrMarkClosure);
6791     do_yield_check();
6792   }
6793   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6794 }
6795 
6796 // Yield in response to a request from VM Thread or
6797 // from mutators.
6798 void Par_MarkFromRootsClosure::do_yield_work() {
6799   assert(_task != NULL, "sanity");
6800   _task->yield();
6801 }
6802 
6803 // A variant of the above used for verifying CMS marking work.
6804 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6805                         MemRegion span,
6806                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6807                         CMSMarkStack*  mark_stack):
6808   _collector(collector),
6809   _span(span),
6810   _verification_bm(verification_bm),
6811   _cms_bm(cms_bm),
6812   _mark_stack(mark_stack),
6813   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6814                       mark_stack)
6815 {
6816   assert(_mark_stack->isEmpty(), "stack should be empty");
6817   _finger = _verification_bm->startWord();
6818   assert(_collector->_restart_addr == NULL, "Sanity check");
6819   assert(_span.contains(_finger), "Out of bounds _finger?");
6820 }
6821 
6822 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6823   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6824   assert(_span.contains(addr), "Out of bounds _finger?");
6825   _finger = addr;
6826 }
6827 
6828 // Should revisit to see if this should be restructured for
6829 // greater efficiency.
6830 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6831   // convert offset into a HeapWord*
6832   HeapWord* addr = _verification_bm->startWord() + offset;
6833   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6834          "address out of range");
6835   assert(_verification_bm->isMarked(addr), "tautology");
6836   assert(_cms_bm->isMarked(addr), "tautology");
6837 
6838   assert(_mark_stack->isEmpty(),
6839          "should drain stack to limit stack usage");
6840   // convert addr to an oop preparatory to scanning
6841   oop obj = oop(addr);
6842   assert(obj->is_oop(), "should be an oop");
6843   assert(_finger <= addr, "_finger runneth ahead");
6844   // advance the finger to right end of this object
6845   _finger = addr + obj->size();
6846   assert(_finger > addr, "we just incremented it above");
6847   // Note: the finger doesn't advance while we drain
6848   // the stack below.
6849   bool res = _mark_stack->push(obj);
6850   assert(res, "Empty non-zero size stack should have space for single push");
6851   while (!_mark_stack->isEmpty()) {
6852     oop new_oop = _mark_stack->pop();
6853     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6854     // now scan this oop's oops
6855     new_oop->oop_iterate(&_pam_verify_closure);
6856   }
6857   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6858   return true;
6859 }
6860 
6861 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6862   CMSCollector* collector, MemRegion span,
6863   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6864   CMSMarkStack*  mark_stack):
6865   MetadataAwareOopClosure(collector->ref_processor()),
6866   _collector(collector),
6867   _span(span),
6868   _verification_bm(verification_bm),
6869   _cms_bm(cms_bm),
6870   _mark_stack(mark_stack)
6871 { }
6872 
6873 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6874 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6875 
6876 // Upon stack overflow, we discard (part of) the stack,
6877 // remembering the least address amongst those discarded
6878 // in CMSCollector's _restart_address.
6879 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6880   // Remember the least grey address discarded
6881   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6882   _collector->lower_restart_addr(ra);
6883   _mark_stack->reset();  // discard stack contents
6884   _mark_stack->expand(); // expand the stack if possible
6885 }
6886 
6887 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6888   assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6889   HeapWord* addr = (HeapWord*)obj;
6890   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6891     // Oop lies in _span and isn't yet grey or black
6892     _verification_bm->mark(addr);            // now grey
6893     if (!_cms_bm->isMarked(addr)) {
6894       oop(addr)->print();
6895       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6896                              p2i(addr));
6897       fatal("... aborting");
6898     }
6899 
6900     if (!_mark_stack->push(obj)) { // stack overflow
6901       if (PrintCMSStatistics != 0) {
6902         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6903                                SIZE_FORMAT, _mark_stack->capacity());
6904       }
6905       assert(_mark_stack->isFull(), "Else push should have succeeded");
6906       handle_stack_overflow(addr);
6907     }
6908     // anything including and to the right of _finger
6909     // will be scanned as we iterate over the remainder of the
6910     // bit map
6911   }
6912 }
6913 
6914 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6915                      MemRegion span,
6916                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6917                      HeapWord* finger, MarkFromRootsClosure* parent) :
6918   MetadataAwareOopClosure(collector->ref_processor()),
6919   _collector(collector),
6920   _span(span),
6921   _bitMap(bitMap),
6922   _markStack(markStack),
6923   _finger(finger),
6924   _parent(parent)
6925 { }
6926 
6927 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6928                      MemRegion span,
6929                      CMSBitMap* bit_map,
6930                      OopTaskQueue* work_queue,
6931                      CMSMarkStack*  overflow_stack,
6932                      HeapWord* finger,
6933                      HeapWord** global_finger_addr,
6934                      Par_MarkFromRootsClosure* parent) :
6935   MetadataAwareOopClosure(collector->ref_processor()),
6936   _collector(collector),
6937   _whole_span(collector->_span),
6938   _span(span),
6939   _bit_map(bit_map),
6940   _work_queue(work_queue),
6941   _overflow_stack(overflow_stack),
6942   _finger(finger),
6943   _global_finger_addr(global_finger_addr),
6944   _parent(parent)
6945 { }
6946 
6947 // Assumes thread-safe access by callers, who are
6948 // responsible for mutual exclusion.
6949 void CMSCollector::lower_restart_addr(HeapWord* low) {
6950   assert(_span.contains(low), "Out of bounds addr");
6951   if (_restart_addr == NULL) {
6952     _restart_addr = low;
6953   } else {
6954     _restart_addr = MIN2(_restart_addr, low);
6955   }
6956 }
6957 
6958 // Upon stack overflow, we discard (part of) the stack,
6959 // remembering the least address amongst those discarded
6960 // in CMSCollector's _restart_address.
6961 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6962   // Remember the least grey address discarded
6963   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6964   _collector->lower_restart_addr(ra);
6965   _markStack->reset();  // discard stack contents
6966   _markStack->expand(); // expand the stack if possible
6967 }
6968 
6969 // Upon stack overflow, we discard (part of) the stack,
6970 // remembering the least address amongst those discarded
6971 // in CMSCollector's _restart_address.
6972 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6973   // We need to do this under a mutex to prevent other
6974   // workers from interfering with the work done below.
6975   MutexLockerEx ml(_overflow_stack->par_lock(),
6976                    Mutex::_no_safepoint_check_flag);
6977   // Remember the least grey address discarded
6978   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6979   _collector->lower_restart_addr(ra);
6980   _overflow_stack->reset();  // discard stack contents
6981   _overflow_stack->expand(); // expand the stack if possible
6982 }
6983 
6984 void PushOrMarkClosure::do_oop(oop obj) {
6985   // Ignore mark word because we are running concurrent with mutators.
6986   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6987   HeapWord* addr = (HeapWord*)obj;
6988   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6989     // Oop lies in _span and isn't yet grey or black
6990     _bitMap->mark(addr);            // now grey
6991     if (addr < _finger) {
6992       // the bit map iteration has already either passed, or
6993       // sampled, this bit in the bit map; we'll need to
6994       // use the marking stack to scan this oop's oops.
6995       bool simulate_overflow = false;
6996       NOT_PRODUCT(
6997         if (CMSMarkStackOverflowALot &&
6998             _collector->simulate_overflow()) {
6999           // simulate a stack overflow
7000           simulate_overflow = true;
7001         }
7002       )
7003       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7004         if (PrintCMSStatistics != 0) {
7005           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7006                                  SIZE_FORMAT, _markStack->capacity());
7007         }
7008         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7009         handle_stack_overflow(addr);
7010       }
7011     }
7012     // anything including and to the right of _finger
7013     // will be scanned as we iterate over the remainder of the
7014     // bit map
7015     do_yield_check();
7016   }
7017 }
7018 
7019 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
7020 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7021 
7022 void Par_PushOrMarkClosure::do_oop(oop obj) {
7023   // Ignore mark word because we are running concurrent with mutators.
7024   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7025   HeapWord* addr = (HeapWord*)obj;
7026   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7027     // Oop lies in _span and isn't yet grey or black
7028     // We read the global_finger (volatile read) strictly after marking oop
7029     bool res = _bit_map->par_mark(addr);    // now grey
7030     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7031     // Should we push this marked oop on our stack?
7032     // -- if someone else marked it, nothing to do
7033     // -- if target oop is above global finger nothing to do
7034     // -- if target oop is in chunk and above local finger
7035     //      then nothing to do
7036     // -- else push on work queue
7037     if (   !res       // someone else marked it, they will deal with it
7038         || (addr >= *gfa)  // will be scanned in a later task
7039         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7040       return;
7041     }
7042     // the bit map iteration has already either passed, or
7043     // sampled, this bit in the bit map; we'll need to
7044     // use the marking stack to scan this oop's oops.
7045     bool simulate_overflow = false;
7046     NOT_PRODUCT(
7047       if (CMSMarkStackOverflowALot &&
7048           _collector->simulate_overflow()) {
7049         // simulate a stack overflow
7050         simulate_overflow = true;
7051       }
7052     )
7053     if (simulate_overflow ||
7054         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7055       // stack overflow
7056       if (PrintCMSStatistics != 0) {
7057         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7058                                SIZE_FORMAT, _overflow_stack->capacity());
7059       }
7060       // We cannot assert that the overflow stack is full because
7061       // it may have been emptied since.
7062       assert(simulate_overflow ||
7063              _work_queue->size() == _work_queue->max_elems(),
7064             "Else push should have succeeded");
7065       handle_stack_overflow(addr);
7066     }
7067     do_yield_check();
7068   }
7069 }
7070 
7071 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
7072 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7073 
7074 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7075                                        MemRegion span,
7076                                        ReferenceProcessor* rp,
7077                                        CMSBitMap* bit_map,
7078                                        CMSBitMap* mod_union_table,
7079                                        CMSMarkStack*  mark_stack,
7080                                        bool           concurrent_precleaning):
7081   MetadataAwareOopClosure(rp),
7082   _collector(collector),
7083   _span(span),
7084   _bit_map(bit_map),
7085   _mod_union_table(mod_union_table),
7086   _mark_stack(mark_stack),
7087   _concurrent_precleaning(concurrent_precleaning)
7088 {
7089   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
7090 }
7091 
7092 // Grey object rescan during pre-cleaning and second checkpoint phases --
7093 // the non-parallel version (the parallel version appears further below.)
7094 void PushAndMarkClosure::do_oop(oop obj) {
7095   // Ignore mark word verification. If during concurrent precleaning,
7096   // the object monitor may be locked. If during the checkpoint
7097   // phases, the object may already have been reached by a  different
7098   // path and may be at the end of the global overflow list (so
7099   // the mark word may be NULL).
7100   assert(obj->is_oop_or_null(true /* ignore mark word */),
7101          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7102   HeapWord* addr = (HeapWord*)obj;
7103   // Check if oop points into the CMS generation
7104   // and is not marked
7105   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7106     // a white object ...
7107     _bit_map->mark(addr);         // ... now grey
7108     // push on the marking stack (grey set)
7109     bool simulate_overflow = false;
7110     NOT_PRODUCT(
7111       if (CMSMarkStackOverflowALot &&
7112           _collector->simulate_overflow()) {
7113         // simulate a stack overflow
7114         simulate_overflow = true;
7115       }
7116     )
7117     if (simulate_overflow || !_mark_stack->push(obj)) {
7118       if (_concurrent_precleaning) {
7119          // During precleaning we can just dirty the appropriate card(s)
7120          // in the mod union table, thus ensuring that the object remains
7121          // in the grey set  and continue. In the case of object arrays
7122          // we need to dirty all of the cards that the object spans,
7123          // since the rescan of object arrays will be limited to the
7124          // dirty cards.
7125          // Note that no one can be interfering with us in this action
7126          // of dirtying the mod union table, so no locking or atomics
7127          // are required.
7128          if (obj->is_objArray()) {
7129            size_t sz = obj->size();
7130            HeapWord* end_card_addr = (HeapWord*)round_to(
7131                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7132            MemRegion redirty_range = MemRegion(addr, end_card_addr);
7133            assert(!redirty_range.is_empty(), "Arithmetical tautology");
7134            _mod_union_table->mark_range(redirty_range);
7135          } else {
7136            _mod_union_table->mark(addr);
7137          }
7138          _collector->_ser_pmc_preclean_ovflw++;
7139       } else {
7140          // During the remark phase, we need to remember this oop
7141          // in the overflow list.
7142          _collector->push_on_overflow_list(obj);
7143          _collector->_ser_pmc_remark_ovflw++;
7144       }
7145     }
7146   }
7147 }
7148 
7149 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7150                                                MemRegion span,
7151                                                ReferenceProcessor* rp,
7152                                                CMSBitMap* bit_map,
7153                                                OopTaskQueue* work_queue):
7154   MetadataAwareOopClosure(rp),
7155   _collector(collector),
7156   _span(span),
7157   _bit_map(bit_map),
7158   _work_queue(work_queue)
7159 {
7160   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
7161 }
7162 
7163 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
7164 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7165 
7166 // Grey object rescan during second checkpoint phase --
7167 // the parallel version.
7168 void Par_PushAndMarkClosure::do_oop(oop obj) {
7169   // In the assert below, we ignore the mark word because
7170   // this oop may point to an already visited object that is
7171   // on the overflow stack (in which case the mark word has
7172   // been hijacked for chaining into the overflow stack --
7173   // if this is the last object in the overflow stack then
7174   // its mark word will be NULL). Because this object may
7175   // have been subsequently popped off the global overflow
7176   // stack, and the mark word possibly restored to the prototypical
7177   // value, by the time we get to examined this failing assert in
7178   // the debugger, is_oop_or_null(false) may subsequently start
7179   // to hold.
7180   assert(obj->is_oop_or_null(true),
7181          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
7182   HeapWord* addr = (HeapWord*)obj;
7183   // Check if oop points into the CMS generation
7184   // and is not marked
7185   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7186     // a white object ...
7187     // If we manage to "claim" the object, by being the
7188     // first thread to mark it, then we push it on our
7189     // marking stack
7190     if (_bit_map->par_mark(addr)) {     // ... now grey
7191       // push on work queue (grey set)
7192       bool simulate_overflow = false;
7193       NOT_PRODUCT(
7194         if (CMSMarkStackOverflowALot &&
7195             _collector->par_simulate_overflow()) {
7196           // simulate a stack overflow
7197           simulate_overflow = true;
7198         }
7199       )
7200       if (simulate_overflow || !_work_queue->push(obj)) {
7201         _collector->par_push_on_overflow_list(obj);
7202         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
7203       }
7204     } // Else, some other thread got there first
7205   }
7206 }
7207 
7208 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
7209 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7210 
7211 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7212   Mutex* bml = _collector->bitMapLock();
7213   assert_lock_strong(bml);
7214   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7215          "CMS thread should hold CMS token");
7216 
7217   bml->unlock();
7218   ConcurrentMarkSweepThread::desynchronize(true);
7219 
7220   _collector->stopTimer();
7221   if (PrintCMSStatistics != 0) {
7222     _collector->incrementYields();
7223   }
7224 
7225   // See the comment in coordinator_yield()
7226   for (unsigned i = 0; i < CMSYieldSleepCount &&
7227                        ConcurrentMarkSweepThread::should_yield() &&
7228                        !CMSCollector::foregroundGCIsActive(); ++i) {
7229     os::sleep(Thread::current(), 1, false);
7230   }
7231 
7232   ConcurrentMarkSweepThread::synchronize(true);
7233   bml->lock();
7234 
7235   _collector->startTimer();
7236 }
7237 
7238 bool CMSPrecleanRefsYieldClosure::should_return() {
7239   if (ConcurrentMarkSweepThread::should_yield()) {
7240     do_yield_work();
7241   }
7242   return _collector->foregroundGCIsActive();
7243 }
7244 
7245 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7246   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7247          "mr should be aligned to start at a card boundary");
7248   // We'd like to assert:
7249   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7250   //        "mr should be a range of cards");
7251   // However, that would be too strong in one case -- the last
7252   // partition ends at _unallocated_block which, in general, can be
7253   // an arbitrary boundary, not necessarily card aligned.
7254   if (PrintCMSStatistics != 0) {
7255     _num_dirty_cards +=
7256          mr.word_size()/CardTableModRefBS::card_size_in_words;
7257   }
7258   _space->object_iterate_mem(mr, &_scan_cl);
7259 }
7260 
7261 SweepClosure::SweepClosure(CMSCollector* collector,
7262                            ConcurrentMarkSweepGeneration* g,
7263                            CMSBitMap* bitMap, bool should_yield) :
7264   _collector(collector),
7265   _g(g),
7266   _sp(g->cmsSpace()),
7267   _limit(_sp->sweep_limit()),
7268   _freelistLock(_sp->freelistLock()),
7269   _bitMap(bitMap),
7270   _yield(should_yield),
7271   _inFreeRange(false),           // No free range at beginning of sweep
7272   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7273   _lastFreeRangeCoalesced(false),
7274   _freeFinger(g->used_region().start())
7275 {
7276   NOT_PRODUCT(
7277     _numObjectsFreed = 0;
7278     _numWordsFreed   = 0;
7279     _numObjectsLive = 0;
7280     _numWordsLive = 0;
7281     _numObjectsAlreadyFree = 0;
7282     _numWordsAlreadyFree = 0;
7283     _last_fc = NULL;
7284 
7285     _sp->initializeIndexedFreeListArrayReturnedBytes();
7286     _sp->dictionary()->initialize_dict_returned_bytes();
7287   )
7288   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7289          "sweep _limit out of bounds");
7290   if (CMSTraceSweeper) {
7291     gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7292                         p2i(_limit));
7293   }
7294 }
7295 
7296 void SweepClosure::print_on(outputStream* st) const {
7297   tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7298                 p2i(_sp->bottom()), p2i(_sp->end()));
7299   tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7300   tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7301   NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7302   tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7303                 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7304 }
7305 
7306 #ifndef PRODUCT
7307 // Assertion checking only:  no useful work in product mode --
7308 // however, if any of the flags below become product flags,
7309 // you may need to review this code to see if it needs to be
7310 // enabled in product mode.
7311 SweepClosure::~SweepClosure() {
7312   assert_lock_strong(_freelistLock);
7313   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7314          "sweep _limit out of bounds");
7315   if (inFreeRange()) {
7316     warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7317     print();
7318     ShouldNotReachHere();
7319   }
7320   if (Verbose && PrintGC) {
7321     gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7322                         _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7323     gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects,  "
7324                            SIZE_FORMAT " bytes  "
7325       "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7326       _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7327       _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7328     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7329                         * sizeof(HeapWord);
7330     gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7331 
7332     if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7333       size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7334       size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7335       size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7336       gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
7337       gclog_or_tty->print("   Indexed List Returned " SIZE_FORMAT " bytes",
7338         indexListReturnedBytes);
7339       gclog_or_tty->print_cr("        Dictionary Returned " SIZE_FORMAT " bytes",
7340         dict_returned_bytes);
7341     }
7342   }
7343   if (CMSTraceSweeper) {
7344     gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7345                            p2i(_limit));
7346   }
7347 }
7348 #endif  // PRODUCT
7349 
7350 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7351     bool freeRangeInFreeLists) {
7352   if (CMSTraceSweeper) {
7353     gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7354                p2i(freeFinger), freeRangeInFreeLists);
7355   }
7356   assert(!inFreeRange(), "Trampling existing free range");
7357   set_inFreeRange(true);
7358   set_lastFreeRangeCoalesced(false);
7359 
7360   set_freeFinger(freeFinger);
7361   set_freeRangeInFreeLists(freeRangeInFreeLists);
7362   if (CMSTestInFreeList) {
7363     if (freeRangeInFreeLists) {
7364       FreeChunk* fc = (FreeChunk*) freeFinger;
7365       assert(fc->is_free(), "A chunk on the free list should be free.");
7366       assert(fc->size() > 0, "Free range should have a size");
7367       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7368     }
7369   }
7370 }
7371 
7372 // Note that the sweeper runs concurrently with mutators. Thus,
7373 // it is possible for direct allocation in this generation to happen
7374 // in the middle of the sweep. Note that the sweeper also coalesces
7375 // contiguous free blocks. Thus, unless the sweeper and the allocator
7376 // synchronize appropriately freshly allocated blocks may get swept up.
7377 // This is accomplished by the sweeper locking the free lists while
7378 // it is sweeping. Thus blocks that are determined to be free are
7379 // indeed free. There is however one additional complication:
7380 // blocks that have been allocated since the final checkpoint and
7381 // mark, will not have been marked and so would be treated as
7382 // unreachable and swept up. To prevent this, the allocator marks
7383 // the bit map when allocating during the sweep phase. This leads,
7384 // however, to a further complication -- objects may have been allocated
7385 // but not yet initialized -- in the sense that the header isn't yet
7386 // installed. The sweeper can not then determine the size of the block
7387 // in order to skip over it. To deal with this case, we use a technique
7388 // (due to Printezis) to encode such uninitialized block sizes in the
7389 // bit map. Since the bit map uses a bit per every HeapWord, but the
7390 // CMS generation has a minimum object size of 3 HeapWords, it follows
7391 // that "normal marks" won't be adjacent in the bit map (there will
7392 // always be at least two 0 bits between successive 1 bits). We make use
7393 // of these "unused" bits to represent uninitialized blocks -- the bit
7394 // corresponding to the start of the uninitialized object and the next
7395 // bit are both set. Finally, a 1 bit marks the end of the object that
7396 // started with the two consecutive 1 bits to indicate its potentially
7397 // uninitialized state.
7398 
7399 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7400   FreeChunk* fc = (FreeChunk*)addr;
7401   size_t res;
7402 
7403   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7404   // than "addr == _limit" because although _limit was a block boundary when
7405   // we started the sweep, it may no longer be one because heap expansion
7406   // may have caused us to coalesce the block ending at the address _limit
7407   // with a newly expanded chunk (this happens when _limit was set to the
7408   // previous _end of the space), so we may have stepped past _limit:
7409   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7410   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7411     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7412            "sweep _limit out of bounds");
7413     assert(addr < _sp->end(), "addr out of bounds");
7414     // Flush any free range we might be holding as a single
7415     // coalesced chunk to the appropriate free list.
7416     if (inFreeRange()) {
7417       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7418              "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7419       flush_cur_free_chunk(freeFinger(),
7420                            pointer_delta(addr, freeFinger()));
7421       if (CMSTraceSweeper) {
7422         gclog_or_tty->print("Sweep: last chunk: ");
7423         gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") "
7424                    "[coalesced:%d]\n",
7425                    p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7426                    lastFreeRangeCoalesced() ? 1 : 0);
7427       }
7428     }
7429 
7430     // help the iterator loop finish
7431     return pointer_delta(_sp->end(), addr);
7432   }
7433 
7434   assert(addr < _limit, "sweep invariant");
7435   // check if we should yield
7436   do_yield_check(addr);
7437   if (fc->is_free()) {
7438     // Chunk that is already free
7439     res = fc->size();
7440     do_already_free_chunk(fc);
7441     debug_only(_sp->verifyFreeLists());
7442     // If we flush the chunk at hand in lookahead_and_flush()
7443     // and it's coalesced with a preceding chunk, then the
7444     // process of "mangling" the payload of the coalesced block
7445     // will cause erasure of the size information from the
7446     // (erstwhile) header of all the coalesced blocks but the
7447     // first, so the first disjunct in the assert will not hold
7448     // in that specific case (in which case the second disjunct
7449     // will hold).
7450     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7451            "Otherwise the size info doesn't change at this step");
7452     NOT_PRODUCT(
7453       _numObjectsAlreadyFree++;
7454       _numWordsAlreadyFree += res;
7455     )
7456     NOT_PRODUCT(_last_fc = fc;)
7457   } else if (!_bitMap->isMarked(addr)) {
7458     // Chunk is fresh garbage
7459     res = do_garbage_chunk(fc);
7460     debug_only(_sp->verifyFreeLists());
7461     NOT_PRODUCT(
7462       _numObjectsFreed++;
7463       _numWordsFreed += res;
7464     )
7465   } else {
7466     // Chunk that is alive.
7467     res = do_live_chunk(fc);
7468     debug_only(_sp->verifyFreeLists());
7469     NOT_PRODUCT(
7470         _numObjectsLive++;
7471         _numWordsLive += res;
7472     )
7473   }
7474   return res;
7475 }
7476 
7477 // For the smart allocation, record following
7478 //  split deaths - a free chunk is removed from its free list because
7479 //      it is being split into two or more chunks.
7480 //  split birth - a free chunk is being added to its free list because
7481 //      a larger free chunk has been split and resulted in this free chunk.
7482 //  coal death - a free chunk is being removed from its free list because
7483 //      it is being coalesced into a large free chunk.
7484 //  coal birth - a free chunk is being added to its free list because
7485 //      it was created when two or more free chunks where coalesced into
7486 //      this free chunk.
7487 //
7488 // These statistics are used to determine the desired number of free
7489 // chunks of a given size.  The desired number is chosen to be relative
7490 // to the end of a CMS sweep.  The desired number at the end of a sweep
7491 // is the
7492 //      count-at-end-of-previous-sweep (an amount that was enough)
7493 //              - count-at-beginning-of-current-sweep  (the excess)
7494 //              + split-births  (gains in this size during interval)
7495 //              - split-deaths  (demands on this size during interval)
7496 // where the interval is from the end of one sweep to the end of the
7497 // next.
7498 //
7499 // When sweeping the sweeper maintains an accumulated chunk which is
7500 // the chunk that is made up of chunks that have been coalesced.  That
7501 // will be termed the left-hand chunk.  A new chunk of garbage that
7502 // is being considered for coalescing will be referred to as the
7503 // right-hand chunk.
7504 //
7505 // When making a decision on whether to coalesce a right-hand chunk with
7506 // the current left-hand chunk, the current count vs. the desired count
7507 // of the left-hand chunk is considered.  Also if the right-hand chunk
7508 // is near the large chunk at the end of the heap (see
7509 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7510 // left-hand chunk is coalesced.
7511 //
7512 // When making a decision about whether to split a chunk, the desired count
7513 // vs. the current count of the candidate to be split is also considered.
7514 // If the candidate is underpopulated (currently fewer chunks than desired)
7515 // a chunk of an overpopulated (currently more chunks than desired) size may
7516 // be chosen.  The "hint" associated with a free list, if non-null, points
7517 // to a free list which may be overpopulated.
7518 //
7519 
7520 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7521   const size_t size = fc->size();
7522   // Chunks that cannot be coalesced are not in the
7523   // free lists.
7524   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7525     assert(_sp->verify_chunk_in_free_list(fc),
7526       "free chunk should be in free lists");
7527   }
7528   // a chunk that is already free, should not have been
7529   // marked in the bit map
7530   HeapWord* const addr = (HeapWord*) fc;
7531   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7532   // Verify that the bit map has no bits marked between
7533   // addr and purported end of this block.
7534   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7535 
7536   // Some chunks cannot be coalesced under any circumstances.
7537   // See the definition of cantCoalesce().
7538   if (!fc->cantCoalesce()) {
7539     // This chunk can potentially be coalesced.
7540     if (_sp->adaptive_freelists()) {
7541       // All the work is done in
7542       do_post_free_or_garbage_chunk(fc, size);
7543     } else {  // Not adaptive free lists
7544       // this is a free chunk that can potentially be coalesced by the sweeper;
7545       if (!inFreeRange()) {
7546         // if the next chunk is a free block that can't be coalesced
7547         // it doesn't make sense to remove this chunk from the free lists
7548         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7549         assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7550         if ((HeapWord*)nextChunk < _sp->end() &&     // There is another free chunk to the right ...
7551             nextChunk->is_free()               &&     // ... which is free...
7552             nextChunk->cantCoalesce()) {             // ... but can't be coalesced
7553           // nothing to do
7554         } else {
7555           // Potentially the start of a new free range:
7556           // Don't eagerly remove it from the free lists.
7557           // No need to remove it if it will just be put
7558           // back again.  (Also from a pragmatic point of view
7559           // if it is a free block in a region that is beyond
7560           // any allocated blocks, an assertion will fail)
7561           // Remember the start of a free run.
7562           initialize_free_range(addr, true);
7563           // end - can coalesce with next chunk
7564         }
7565       } else {
7566         // the midst of a free range, we are coalescing
7567         print_free_block_coalesced(fc);
7568         if (CMSTraceSweeper) {
7569           gclog_or_tty->print("  -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7570         }
7571         // remove it from the free lists
7572         _sp->removeFreeChunkFromFreeLists(fc);
7573         set_lastFreeRangeCoalesced(true);
7574         // If the chunk is being coalesced and the current free range is
7575         // in the free lists, remove the current free range so that it
7576         // will be returned to the free lists in its entirety - all
7577         // the coalesced pieces included.
7578         if (freeRangeInFreeLists()) {
7579           FreeChunk* ffc = (FreeChunk*) freeFinger();
7580           assert(ffc->size() == pointer_delta(addr, freeFinger()),
7581             "Size of free range is inconsistent with chunk size.");
7582           if (CMSTestInFreeList) {
7583             assert(_sp->verify_chunk_in_free_list(ffc),
7584               "free range is not in free lists");
7585           }
7586           _sp->removeFreeChunkFromFreeLists(ffc);
7587           set_freeRangeInFreeLists(false);
7588         }
7589       }
7590     }
7591     // Note that if the chunk is not coalescable (the else arm
7592     // below), we unconditionally flush, without needing to do
7593     // a "lookahead," as we do below.
7594     if (inFreeRange()) lookahead_and_flush(fc, size);
7595   } else {
7596     // Code path common to both original and adaptive free lists.
7597 
7598     // cant coalesce with previous block; this should be treated
7599     // as the end of a free run if any
7600     if (inFreeRange()) {
7601       // we kicked some butt; time to pick up the garbage
7602       assert(freeFinger() < addr, "freeFinger points too high");
7603       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7604     }
7605     // else, nothing to do, just continue
7606   }
7607 }
7608 
7609 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7610   // This is a chunk of garbage.  It is not in any free list.
7611   // Add it to a free list or let it possibly be coalesced into
7612   // a larger chunk.
7613   HeapWord* const addr = (HeapWord*) fc;
7614   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7615 
7616   if (_sp->adaptive_freelists()) {
7617     // Verify that the bit map has no bits marked between
7618     // addr and purported end of just dead object.
7619     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7620 
7621     do_post_free_or_garbage_chunk(fc, size);
7622   } else {
7623     if (!inFreeRange()) {
7624       // start of a new free range
7625       assert(size > 0, "A free range should have a size");
7626       initialize_free_range(addr, false);
7627     } else {
7628       // this will be swept up when we hit the end of the
7629       // free range
7630       if (CMSTraceSweeper) {
7631         gclog_or_tty->print("  -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7632       }
7633       // If the chunk is being coalesced and the current free range is
7634       // in the free lists, remove the current free range so that it
7635       // will be returned to the free lists in its entirety - all
7636       // the coalesced pieces included.
7637       if (freeRangeInFreeLists()) {
7638         FreeChunk* ffc = (FreeChunk*)freeFinger();
7639         assert(ffc->size() == pointer_delta(addr, freeFinger()),
7640           "Size of free range is inconsistent with chunk size.");
7641         if (CMSTestInFreeList) {
7642           assert(_sp->verify_chunk_in_free_list(ffc),
7643             "free range is not in free lists");
7644         }
7645         _sp->removeFreeChunkFromFreeLists(ffc);
7646         set_freeRangeInFreeLists(false);
7647       }
7648       set_lastFreeRangeCoalesced(true);
7649     }
7650     // this will be swept up when we hit the end of the free range
7651 
7652     // Verify that the bit map has no bits marked between
7653     // addr and purported end of just dead object.
7654     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7655   }
7656   assert(_limit >= addr + size,
7657          "A freshly garbage chunk can't possibly straddle over _limit");
7658   if (inFreeRange()) lookahead_and_flush(fc, size);
7659   return size;
7660 }
7661 
7662 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7663   HeapWord* addr = (HeapWord*) fc;
7664   // The sweeper has just found a live object. Return any accumulated
7665   // left hand chunk to the free lists.
7666   if (inFreeRange()) {
7667     assert(freeFinger() < addr, "freeFinger points too high");
7668     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7669   }
7670 
7671   // This object is live: we'd normally expect this to be
7672   // an oop, and like to assert the following:
7673   // assert(oop(addr)->is_oop(), "live block should be an oop");
7674   // However, as we commented above, this may be an object whose
7675   // header hasn't yet been initialized.
7676   size_t size;
7677   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7678   if (_bitMap->isMarked(addr + 1)) {
7679     // Determine the size from the bit map, rather than trying to
7680     // compute it from the object header.
7681     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7682     size = pointer_delta(nextOneAddr + 1, addr);
7683     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7684            "alignment problem");
7685 
7686 #ifdef ASSERT
7687       if (oop(addr)->klass_or_null() != NULL) {
7688         // Ignore mark word because we are running concurrent with mutators
7689         assert(oop(addr)->is_oop(true), "live block should be an oop");
7690         assert(size ==
7691                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7692                "P-mark and computed size do not agree");
7693       }
7694 #endif
7695 
7696   } else {
7697     // This should be an initialized object that's alive.
7698     assert(oop(addr)->klass_or_null() != NULL,
7699            "Should be an initialized object");
7700     // Ignore mark word because we are running concurrent with mutators
7701     assert(oop(addr)->is_oop(true), "live block should be an oop");
7702     // Verify that the bit map has no bits marked between
7703     // addr and purported end of this block.
7704     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7705     assert(size >= 3, "Necessary for Printezis marks to work");
7706     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7707     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7708   }
7709   return size;
7710 }
7711 
7712 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7713                                                  size_t chunkSize) {
7714   // do_post_free_or_garbage_chunk() should only be called in the case
7715   // of the adaptive free list allocator.
7716   const bool fcInFreeLists = fc->is_free();
7717   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7718   assert((HeapWord*)fc <= _limit, "sweep invariant");
7719   if (CMSTestInFreeList && fcInFreeLists) {
7720     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7721   }
7722 
7723   if (CMSTraceSweeper) {
7724     gclog_or_tty->print_cr("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7725   }
7726 
7727   HeapWord* const fc_addr = (HeapWord*) fc;
7728 
7729   bool coalesce;
7730   const size_t left  = pointer_delta(fc_addr, freeFinger());
7731   const size_t right = chunkSize;
7732   switch (FLSCoalescePolicy) {
7733     // numeric value forms a coalition aggressiveness metric
7734     case 0:  { // never coalesce
7735       coalesce = false;
7736       break;
7737     }
7738     case 1: { // coalesce if left & right chunks on overpopulated lists
7739       coalesce = _sp->coalOverPopulated(left) &&
7740                  _sp->coalOverPopulated(right);
7741       break;
7742     }
7743     case 2: { // coalesce if left chunk on overpopulated list (default)
7744       coalesce = _sp->coalOverPopulated(left);
7745       break;
7746     }
7747     case 3: { // coalesce if left OR right chunk on overpopulated list
7748       coalesce = _sp->coalOverPopulated(left) ||
7749                  _sp->coalOverPopulated(right);
7750       break;
7751     }
7752     case 4: { // always coalesce
7753       coalesce = true;
7754       break;
7755     }
7756     default:
7757      ShouldNotReachHere();
7758   }
7759 
7760   // Should the current free range be coalesced?
7761   // If the chunk is in a free range and either we decided to coalesce above
7762   // or the chunk is near the large block at the end of the heap
7763   // (isNearLargestChunk() returns true), then coalesce this chunk.
7764   const bool doCoalesce = inFreeRange()
7765                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7766   if (doCoalesce) {
7767     // Coalesce the current free range on the left with the new
7768     // chunk on the right.  If either is on a free list,
7769     // it must be removed from the list and stashed in the closure.
7770     if (freeRangeInFreeLists()) {
7771       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7772       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7773         "Size of free range is inconsistent with chunk size.");
7774       if (CMSTestInFreeList) {
7775         assert(_sp->verify_chunk_in_free_list(ffc),
7776           "Chunk is not in free lists");
7777       }
7778       _sp->coalDeath(ffc->size());
7779       _sp->removeFreeChunkFromFreeLists(ffc);
7780       set_freeRangeInFreeLists(false);
7781     }
7782     if (fcInFreeLists) {
7783       _sp->coalDeath(chunkSize);
7784       assert(fc->size() == chunkSize,
7785         "The chunk has the wrong size or is not in the free lists");
7786       _sp->removeFreeChunkFromFreeLists(fc);
7787     }
7788     set_lastFreeRangeCoalesced(true);
7789     print_free_block_coalesced(fc);
7790   } else {  // not in a free range and/or should not coalesce
7791     // Return the current free range and start a new one.
7792     if (inFreeRange()) {
7793       // In a free range but cannot coalesce with the right hand chunk.
7794       // Put the current free range into the free lists.
7795       flush_cur_free_chunk(freeFinger(),
7796                            pointer_delta(fc_addr, freeFinger()));
7797     }
7798     // Set up for new free range.  Pass along whether the right hand
7799     // chunk is in the free lists.
7800     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7801   }
7802 }
7803 
7804 // Lookahead flush:
7805 // If we are tracking a free range, and this is the last chunk that
7806 // we'll look at because its end crosses past _limit, we'll preemptively
7807 // flush it along with any free range we may be holding on to. Note that
7808 // this can be the case only for an already free or freshly garbage
7809 // chunk. If this block is an object, it can never straddle
7810 // over _limit. The "straddling" occurs when _limit is set at
7811 // the previous end of the space when this cycle started, and
7812 // a subsequent heap expansion caused the previously co-terminal
7813 // free block to be coalesced with the newly expanded portion,
7814 // thus rendering _limit a non-block-boundary making it dangerous
7815 // for the sweeper to step over and examine.
7816 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7817   assert(inFreeRange(), "Should only be called if currently in a free range.");
7818   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7819   assert(_sp->used_region().contains(eob - 1),
7820          "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7821          " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7822          " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7823          p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7824   if (eob >= _limit) {
7825     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7826     if (CMSTraceSweeper) {
7827       gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7828                              "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7829                              "[" PTR_FORMAT "," PTR_FORMAT ")",
7830                              p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7831     }
7832     // Return the storage we are tracking back into the free lists.
7833     if (CMSTraceSweeper) {
7834       gclog_or_tty->print_cr("Flushing ... ");
7835     }
7836     assert(freeFinger() < eob, "Error");
7837     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7838   }
7839 }
7840 
7841 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7842   assert(inFreeRange(), "Should only be called if currently in a free range.");
7843   assert(size > 0,
7844     "A zero sized chunk cannot be added to the free lists.");
7845   if (!freeRangeInFreeLists()) {
7846     if (CMSTestInFreeList) {
7847       FreeChunk* fc = (FreeChunk*) chunk;
7848       fc->set_size(size);
7849       assert(!_sp->verify_chunk_in_free_list(fc),
7850         "chunk should not be in free lists yet");
7851     }
7852     if (CMSTraceSweeper) {
7853       gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7854                     p2i(chunk), size);
7855     }
7856     // A new free range is going to be starting.  The current
7857     // free range has not been added to the free lists yet or
7858     // was removed so add it back.
7859     // If the current free range was coalesced, then the death
7860     // of the free range was recorded.  Record a birth now.
7861     if (lastFreeRangeCoalesced()) {
7862       _sp->coalBirth(size);
7863     }
7864     _sp->addChunkAndRepairOffsetTable(chunk, size,
7865             lastFreeRangeCoalesced());
7866   } else if (CMSTraceSweeper) {
7867     gclog_or_tty->print_cr("Already in free list: nothing to flush");
7868   }
7869   set_inFreeRange(false);
7870   set_freeRangeInFreeLists(false);
7871 }
7872 
7873 // We take a break if we've been at this for a while,
7874 // so as to avoid monopolizing the locks involved.
7875 void SweepClosure::do_yield_work(HeapWord* addr) {
7876   // Return current free chunk being used for coalescing (if any)
7877   // to the appropriate freelist.  After yielding, the next
7878   // free block encountered will start a coalescing range of
7879   // free blocks.  If the next free block is adjacent to the
7880   // chunk just flushed, they will need to wait for the next
7881   // sweep to be coalesced.
7882   if (inFreeRange()) {
7883     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7884   }
7885 
7886   // First give up the locks, then yield, then re-lock.
7887   // We should probably use a constructor/destructor idiom to
7888   // do this unlock/lock or modify the MutexUnlocker class to
7889   // serve our purpose. XXX
7890   assert_lock_strong(_bitMap->lock());
7891   assert_lock_strong(_freelistLock);
7892   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7893          "CMS thread should hold CMS token");
7894   _bitMap->lock()->unlock();
7895   _freelistLock->unlock();
7896   ConcurrentMarkSweepThread::desynchronize(true);
7897   _collector->stopTimer();
7898   if (PrintCMSStatistics != 0) {
7899     _collector->incrementYields();
7900   }
7901 
7902   // See the comment in coordinator_yield()
7903   for (unsigned i = 0; i < CMSYieldSleepCount &&
7904                        ConcurrentMarkSweepThread::should_yield() &&
7905                        !CMSCollector::foregroundGCIsActive(); ++i) {
7906     os::sleep(Thread::current(), 1, false);
7907   }
7908 
7909   ConcurrentMarkSweepThread::synchronize(true);
7910   _freelistLock->lock();
7911   _bitMap->lock()->lock_without_safepoint_check();
7912   _collector->startTimer();
7913 }
7914 
7915 #ifndef PRODUCT
7916 // This is actually very useful in a product build if it can
7917 // be called from the debugger.  Compile it into the product
7918 // as needed.
7919 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7920   return debug_cms_space->verify_chunk_in_free_list(fc);
7921 }
7922 #endif
7923 
7924 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7925   if (CMSTraceSweeper) {
7926     gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7927                            p2i(fc), fc->size());
7928   }
7929 }
7930 
7931 // CMSIsAliveClosure
7932 bool CMSIsAliveClosure::do_object_b(oop obj) {
7933   HeapWord* addr = (HeapWord*)obj;
7934   return addr != NULL &&
7935          (!_span.contains(addr) || _bit_map->isMarked(addr));
7936 }
7937 
7938 
7939 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7940                       MemRegion span,
7941                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7942                       bool cpc):
7943   _collector(collector),
7944   _span(span),
7945   _bit_map(bit_map),
7946   _mark_stack(mark_stack),
7947   _concurrent_precleaning(cpc) {
7948   assert(!_span.is_empty(), "Empty span could spell trouble");
7949 }
7950 
7951 
7952 // CMSKeepAliveClosure: the serial version
7953 void CMSKeepAliveClosure::do_oop(oop obj) {
7954   HeapWord* addr = (HeapWord*)obj;
7955   if (_span.contains(addr) &&
7956       !_bit_map->isMarked(addr)) {
7957     _bit_map->mark(addr);
7958     bool simulate_overflow = false;
7959     NOT_PRODUCT(
7960       if (CMSMarkStackOverflowALot &&
7961           _collector->simulate_overflow()) {
7962         // simulate a stack overflow
7963         simulate_overflow = true;
7964       }
7965     )
7966     if (simulate_overflow || !_mark_stack->push(obj)) {
7967       if (_concurrent_precleaning) {
7968         // We dirty the overflown object and let the remark
7969         // phase deal with it.
7970         assert(_collector->overflow_list_is_empty(), "Error");
7971         // In the case of object arrays, we need to dirty all of
7972         // the cards that the object spans. No locking or atomics
7973         // are needed since no one else can be mutating the mod union
7974         // table.
7975         if (obj->is_objArray()) {
7976           size_t sz = obj->size();
7977           HeapWord* end_card_addr =
7978             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7979           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7980           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7981           _collector->_modUnionTable.mark_range(redirty_range);
7982         } else {
7983           _collector->_modUnionTable.mark(addr);
7984         }
7985         _collector->_ser_kac_preclean_ovflw++;
7986       } else {
7987         _collector->push_on_overflow_list(obj);
7988         _collector->_ser_kac_ovflw++;
7989       }
7990     }
7991   }
7992 }
7993 
7994 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
7995 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7996 
7997 // CMSParKeepAliveClosure: a parallel version of the above.
7998 // The work queues are private to each closure (thread),
7999 // but (may be) available for stealing by other threads.
8000 void CMSParKeepAliveClosure::do_oop(oop obj) {
8001   HeapWord* addr = (HeapWord*)obj;
8002   if (_span.contains(addr) &&
8003       !_bit_map->isMarked(addr)) {
8004     // In general, during recursive tracing, several threads
8005     // may be concurrently getting here; the first one to
8006     // "tag" it, claims it.
8007     if (_bit_map->par_mark(addr)) {
8008       bool res = _work_queue->push(obj);
8009       assert(res, "Low water mark should be much less than capacity");
8010       // Do a recursive trim in the hope that this will keep
8011       // stack usage lower, but leave some oops for potential stealers
8012       trim_queue(_low_water_mark);
8013     } // Else, another thread got there first
8014   }
8015 }
8016 
8017 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
8018 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8019 
8020 void CMSParKeepAliveClosure::trim_queue(uint max) {
8021   while (_work_queue->size() > max) {
8022     oop new_oop;
8023     if (_work_queue->pop_local(new_oop)) {
8024       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8025       assert(_bit_map->isMarked((HeapWord*)new_oop),
8026              "no white objects on this stack!");
8027       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8028       // iterate over the oops in this oop, marking and pushing
8029       // the ones in CMS heap (i.e. in _span).
8030       new_oop->oop_iterate(&_mark_and_push);
8031     }
8032   }
8033 }
8034 
8035 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8036                                 CMSCollector* collector,
8037                                 MemRegion span, CMSBitMap* bit_map,
8038                                 OopTaskQueue* work_queue):
8039   _collector(collector),
8040   _span(span),
8041   _bit_map(bit_map),
8042   _work_queue(work_queue) { }
8043 
8044 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8045   HeapWord* addr = (HeapWord*)obj;
8046   if (_span.contains(addr) &&
8047       !_bit_map->isMarked(addr)) {
8048     if (_bit_map->par_mark(addr)) {
8049       bool simulate_overflow = false;
8050       NOT_PRODUCT(
8051         if (CMSMarkStackOverflowALot &&
8052             _collector->par_simulate_overflow()) {
8053           // simulate a stack overflow
8054           simulate_overflow = true;
8055         }
8056       )
8057       if (simulate_overflow || !_work_queue->push(obj)) {
8058         _collector->par_push_on_overflow_list(obj);
8059         _collector->_par_kac_ovflw++;
8060       }
8061     } // Else another thread got there already
8062   }
8063 }
8064 
8065 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8066 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8067 
8068 //////////////////////////////////////////////////////////////////
8069 //  CMSExpansionCause                /////////////////////////////
8070 //////////////////////////////////////////////////////////////////
8071 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8072   switch (cause) {
8073     case _no_expansion:
8074       return "No expansion";
8075     case _satisfy_free_ratio:
8076       return "Free ratio";
8077     case _satisfy_promotion:
8078       return "Satisfy promotion";
8079     case _satisfy_allocation:
8080       return "allocation";
8081     case _allocate_par_lab:
8082       return "Par LAB";
8083     case _allocate_par_spooling_space:
8084       return "Par Spooling Space";
8085     case _adaptive_size_policy:
8086       return "Ergonomics";
8087     default:
8088       return "unknown";
8089   }
8090 }
8091 
8092 void CMSDrainMarkingStackClosure::do_void() {
8093   // the max number to take from overflow list at a time
8094   const size_t num = _mark_stack->capacity()/4;
8095   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8096          "Overflow list should be NULL during concurrent phases");
8097   while (!_mark_stack->isEmpty() ||
8098          // if stack is empty, check the overflow list
8099          _collector->take_from_overflow_list(num, _mark_stack)) {
8100     oop obj = _mark_stack->pop();
8101     HeapWord* addr = (HeapWord*)obj;
8102     assert(_span.contains(addr), "Should be within span");
8103     assert(_bit_map->isMarked(addr), "Should be marked");
8104     assert(obj->is_oop(), "Should be an oop");
8105     obj->oop_iterate(_keep_alive);
8106   }
8107 }
8108 
8109 void CMSParDrainMarkingStackClosure::do_void() {
8110   // drain queue
8111   trim_queue(0);
8112 }
8113 
8114 // Trim our work_queue so its length is below max at return
8115 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8116   while (_work_queue->size() > max) {
8117     oop new_oop;
8118     if (_work_queue->pop_local(new_oop)) {
8119       assert(new_oop->is_oop(), "Expected an oop");
8120       assert(_bit_map->isMarked((HeapWord*)new_oop),
8121              "no white objects on this stack!");
8122       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8123       // iterate over the oops in this oop, marking and pushing
8124       // the ones in CMS heap (i.e. in _span).
8125       new_oop->oop_iterate(&_mark_and_push);
8126     }
8127   }
8128 }
8129 
8130 ////////////////////////////////////////////////////////////////////
8131 // Support for Marking Stack Overflow list handling and related code
8132 ////////////////////////////////////////////////////////////////////
8133 // Much of the following code is similar in shape and spirit to the
8134 // code used in ParNewGC. We should try and share that code
8135 // as much as possible in the future.
8136 
8137 #ifndef PRODUCT
8138 // Debugging support for CMSStackOverflowALot
8139 
8140 // It's OK to call this multi-threaded;  the worst thing
8141 // that can happen is that we'll get a bunch of closely
8142 // spaced simulated overflows, but that's OK, in fact
8143 // probably good as it would exercise the overflow code
8144 // under contention.
8145 bool CMSCollector::simulate_overflow() {
8146   if (_overflow_counter-- <= 0) { // just being defensive
8147     _overflow_counter = CMSMarkStackOverflowInterval;
8148     return true;
8149   } else {
8150     return false;
8151   }
8152 }
8153 
8154 bool CMSCollector::par_simulate_overflow() {
8155   return simulate_overflow();
8156 }
8157 #endif
8158 
8159 // Single-threaded
8160 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8161   assert(stack->isEmpty(), "Expected precondition");
8162   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8163   size_t i = num;
8164   oop  cur = _overflow_list;
8165   const markOop proto = markOopDesc::prototype();
8166   NOT_PRODUCT(ssize_t n = 0;)
8167   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8168     next = oop(cur->mark());
8169     cur->set_mark(proto);   // until proven otherwise
8170     assert(cur->is_oop(), "Should be an oop");
8171     bool res = stack->push(cur);
8172     assert(res, "Bit off more than can chew?");
8173     NOT_PRODUCT(n++;)
8174   }
8175   _overflow_list = cur;
8176 #ifndef PRODUCT
8177   assert(_num_par_pushes >= n, "Too many pops?");
8178   _num_par_pushes -=n;
8179 #endif
8180   return !stack->isEmpty();
8181 }
8182 
8183 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
8184 // (MT-safe) Get a prefix of at most "num" from the list.
8185 // The overflow list is chained through the mark word of
8186 // each object in the list. We fetch the entire list,
8187 // break off a prefix of the right size and return the
8188 // remainder. If other threads try to take objects from
8189 // the overflow list at that time, they will wait for
8190 // some time to see if data becomes available. If (and
8191 // only if) another thread places one or more object(s)
8192 // on the global list before we have returned the suffix
8193 // to the global list, we will walk down our local list
8194 // to find its end and append the global list to
8195 // our suffix before returning it. This suffix walk can
8196 // prove to be expensive (quadratic in the amount of traffic)
8197 // when there are many objects in the overflow list and
8198 // there is much producer-consumer contention on the list.
8199 // *NOTE*: The overflow list manipulation code here and
8200 // in ParNewGeneration:: are very similar in shape,
8201 // except that in the ParNew case we use the old (from/eden)
8202 // copy of the object to thread the list via its klass word.
8203 // Because of the common code, if you make any changes in
8204 // the code below, please check the ParNew version to see if
8205 // similar changes might be needed.
8206 // CR 6797058 has been filed to consolidate the common code.
8207 bool CMSCollector::par_take_from_overflow_list(size_t num,
8208                                                OopTaskQueue* work_q,
8209                                                int no_of_gc_threads) {
8210   assert(work_q->size() == 0, "First empty local work queue");
8211   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8212   if (_overflow_list == NULL) {
8213     return false;
8214   }
8215   // Grab the entire list; we'll put back a suffix
8216   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8217   Thread* tid = Thread::current();
8218   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8219   // set to ParallelGCThreads.
8220   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8221   size_t sleep_time_millis = MAX2((size_t)1, num/100);
8222   // If the list is busy, we spin for a short while,
8223   // sleeping between attempts to get the list.
8224   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8225     os::sleep(tid, sleep_time_millis, false);
8226     if (_overflow_list == NULL) {
8227       // Nothing left to take
8228       return false;
8229     } else if (_overflow_list != BUSY) {
8230       // Try and grab the prefix
8231       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8232     }
8233   }
8234   // If the list was found to be empty, or we spun long
8235   // enough, we give up and return empty-handed. If we leave
8236   // the list in the BUSY state below, it must be the case that
8237   // some other thread holds the overflow list and will set it
8238   // to a non-BUSY state in the future.
8239   if (prefix == NULL || prefix == BUSY) {
8240      // Nothing to take or waited long enough
8241      if (prefix == NULL) {
8242        // Write back the NULL in case we overwrote it with BUSY above
8243        // and it is still the same value.
8244        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8245      }
8246      return false;
8247   }
8248   assert(prefix != NULL && prefix != BUSY, "Error");
8249   size_t i = num;
8250   oop cur = prefix;
8251   // Walk down the first "num" objects, unless we reach the end.
8252   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8253   if (cur->mark() == NULL) {
8254     // We have "num" or fewer elements in the list, so there
8255     // is nothing to return to the global list.
8256     // Write back the NULL in lieu of the BUSY we wrote
8257     // above, if it is still the same value.
8258     if (_overflow_list == BUSY) {
8259       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8260     }
8261   } else {
8262     // Chop off the suffix and return it to the global list.
8263     assert(cur->mark() != BUSY, "Error");
8264     oop suffix_head = cur->mark(); // suffix will be put back on global list
8265     cur->set_mark(NULL);           // break off suffix
8266     // It's possible that the list is still in the empty(busy) state
8267     // we left it in a short while ago; in that case we may be
8268     // able to place back the suffix without incurring the cost
8269     // of a walk down the list.
8270     oop observed_overflow_list = _overflow_list;
8271     oop cur_overflow_list = observed_overflow_list;
8272     bool attached = false;
8273     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8274       observed_overflow_list =
8275         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8276       if (cur_overflow_list == observed_overflow_list) {
8277         attached = true;
8278         break;
8279       } else cur_overflow_list = observed_overflow_list;
8280     }
8281     if (!attached) {
8282       // Too bad, someone else sneaked in (at least) an element; we'll need
8283       // to do a splice. Find tail of suffix so we can prepend suffix to global
8284       // list.
8285       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8286       oop suffix_tail = cur;
8287       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8288              "Tautology");
8289       observed_overflow_list = _overflow_list;
8290       do {
8291         cur_overflow_list = observed_overflow_list;
8292         if (cur_overflow_list != BUSY) {
8293           // Do the splice ...
8294           suffix_tail->set_mark(markOop(cur_overflow_list));
8295         } else { // cur_overflow_list == BUSY
8296           suffix_tail->set_mark(NULL);
8297         }
8298         // ... and try to place spliced list back on overflow_list ...
8299         observed_overflow_list =
8300           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8301       } while (cur_overflow_list != observed_overflow_list);
8302       // ... until we have succeeded in doing so.
8303     }
8304   }
8305 
8306   // Push the prefix elements on work_q
8307   assert(prefix != NULL, "control point invariant");
8308   const markOop proto = markOopDesc::prototype();
8309   oop next;
8310   NOT_PRODUCT(ssize_t n = 0;)
8311   for (cur = prefix; cur != NULL; cur = next) {
8312     next = oop(cur->mark());
8313     cur->set_mark(proto);   // until proven otherwise
8314     assert(cur->is_oop(), "Should be an oop");
8315     bool res = work_q->push(cur);
8316     assert(res, "Bit off more than we can chew?");
8317     NOT_PRODUCT(n++;)
8318   }
8319 #ifndef PRODUCT
8320   assert(_num_par_pushes >= n, "Too many pops?");
8321   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8322 #endif
8323   return true;
8324 }
8325 
8326 // Single-threaded
8327 void CMSCollector::push_on_overflow_list(oop p) {
8328   NOT_PRODUCT(_num_par_pushes++;)
8329   assert(p->is_oop(), "Not an oop");
8330   preserve_mark_if_necessary(p);
8331   p->set_mark((markOop)_overflow_list);
8332   _overflow_list = p;
8333 }
8334 
8335 // Multi-threaded; use CAS to prepend to overflow list
8336 void CMSCollector::par_push_on_overflow_list(oop p) {
8337   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8338   assert(p->is_oop(), "Not an oop");
8339   par_preserve_mark_if_necessary(p);
8340   oop observed_overflow_list = _overflow_list;
8341   oop cur_overflow_list;
8342   do {
8343     cur_overflow_list = observed_overflow_list;
8344     if (cur_overflow_list != BUSY) {
8345       p->set_mark(markOop(cur_overflow_list));
8346     } else {
8347       p->set_mark(NULL);
8348     }
8349     observed_overflow_list =
8350       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8351   } while (cur_overflow_list != observed_overflow_list);
8352 }
8353 #undef BUSY
8354 
8355 // Single threaded
8356 // General Note on GrowableArray: pushes may silently fail
8357 // because we are (temporarily) out of C-heap for expanding
8358 // the stack. The problem is quite ubiquitous and affects
8359 // a lot of code in the JVM. The prudent thing for GrowableArray
8360 // to do (for now) is to exit with an error. However, that may
8361 // be too draconian in some cases because the caller may be
8362 // able to recover without much harm. For such cases, we
8363 // should probably introduce a "soft_push" method which returns
8364 // an indication of success or failure with the assumption that
8365 // the caller may be able to recover from a failure; code in
8366 // the VM can then be changed, incrementally, to deal with such
8367 // failures where possible, thus, incrementally hardening the VM
8368 // in such low resource situations.
8369 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8370   _preserved_oop_stack.push(p);
8371   _preserved_mark_stack.push(m);
8372   assert(m == p->mark(), "Mark word changed");
8373   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8374          "bijection");
8375 }
8376 
8377 // Single threaded
8378 void CMSCollector::preserve_mark_if_necessary(oop p) {
8379   markOop m = p->mark();
8380   if (m->must_be_preserved(p)) {
8381     preserve_mark_work(p, m);
8382   }
8383 }
8384 
8385 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8386   markOop m = p->mark();
8387   if (m->must_be_preserved(p)) {
8388     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8389     // Even though we read the mark word without holding
8390     // the lock, we are assured that it will not change
8391     // because we "own" this oop, so no other thread can
8392     // be trying to push it on the overflow list; see
8393     // the assertion in preserve_mark_work() that checks
8394     // that m == p->mark().
8395     preserve_mark_work(p, m);
8396   }
8397 }
8398 
8399 // We should be able to do this multi-threaded,
8400 // a chunk of stack being a task (this is
8401 // correct because each oop only ever appears
8402 // once in the overflow list. However, it's
8403 // not very easy to completely overlap this with
8404 // other operations, so will generally not be done
8405 // until all work's been completed. Because we
8406 // expect the preserved oop stack (set) to be small,
8407 // it's probably fine to do this single-threaded.
8408 // We can explore cleverer concurrent/overlapped/parallel
8409 // processing of preserved marks if we feel the
8410 // need for this in the future. Stack overflow should
8411 // be so rare in practice and, when it happens, its
8412 // effect on performance so great that this will
8413 // likely just be in the noise anyway.
8414 void CMSCollector::restore_preserved_marks_if_any() {
8415   assert(SafepointSynchronize::is_at_safepoint(),
8416          "world should be stopped");
8417   assert(Thread::current()->is_ConcurrentGC_thread() ||
8418          Thread::current()->is_VM_thread(),
8419          "should be single-threaded");
8420   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8421          "bijection");
8422 
8423   while (!_preserved_oop_stack.is_empty()) {
8424     oop p = _preserved_oop_stack.pop();
8425     assert(p->is_oop(), "Should be an oop");
8426     assert(_span.contains(p), "oop should be in _span");
8427     assert(p->mark() == markOopDesc::prototype(),
8428            "Set when taken from overflow list");
8429     markOop m = _preserved_mark_stack.pop();
8430     p->set_mark(m);
8431   }
8432   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8433          "stacks were cleared above");
8434 }
8435 
8436 #ifndef PRODUCT
8437 bool CMSCollector::no_preserved_marks() const {
8438   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8439 }
8440 #endif
8441 
8442 // Transfer some number of overflown objects to usual marking
8443 // stack. Return true if some objects were transferred.
8444 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8445   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8446                     (size_t)ParGCDesiredObjsFromOverflowList);
8447 
8448   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8449   assert(_collector->overflow_list_is_empty() || res,
8450          "If list is not empty, we should have taken something");
8451   assert(!res || !_mark_stack->isEmpty(),
8452          "If we took something, it should now be on our stack");
8453   return res;
8454 }
8455 
8456 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8457   size_t res = _sp->block_size_no_stall(addr, _collector);
8458   if (_sp->block_is_obj(addr)) {
8459     if (_live_bit_map->isMarked(addr)) {
8460       // It can't have been dead in a previous cycle
8461       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8462     } else {
8463       _dead_bit_map->mark(addr);      // mark the dead object
8464     }
8465   }
8466   // Could be 0, if the block size could not be computed without stalling.
8467   return res;
8468 }
8469 
8470 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8471 
8472   switch (phase) {
8473     case CMSCollector::InitialMarking:
8474       initialize(true  /* fullGC */ ,
8475                  cause /* cause of the GC */,
8476                  true  /* recordGCBeginTime */,
8477                  true  /* recordPreGCUsage */,
8478                  false /* recordPeakUsage */,
8479                  false /* recordPostGCusage */,
8480                  true  /* recordAccumulatedGCTime */,
8481                  false /* recordGCEndTime */,
8482                  false /* countCollection */  );
8483       break;
8484 
8485     case CMSCollector::FinalMarking:
8486       initialize(true  /* fullGC */ ,
8487                  cause /* cause of the GC */,
8488                  false /* recordGCBeginTime */,
8489                  false /* recordPreGCUsage */,
8490                  false /* recordPeakUsage */,
8491                  false /* recordPostGCusage */,
8492                  true  /* recordAccumulatedGCTime */,
8493                  false /* recordGCEndTime */,
8494                  false /* countCollection */  );
8495       break;
8496 
8497     case CMSCollector::Sweeping:
8498       initialize(true  /* fullGC */ ,
8499                  cause /* cause of the GC */,
8500                  false /* recordGCBeginTime */,
8501                  false /* recordPreGCUsage */,
8502                  true  /* recordPeakUsage */,
8503                  true  /* recordPostGCusage */,
8504                  false /* recordAccumulatedGCTime */,
8505                  true  /* recordGCEndTime */,
8506                  true  /* countCollection */  );
8507       break;
8508 
8509     default:
8510       ShouldNotReachHere();
8511   }
8512 }