1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderData.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/symbolTable.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "gc/cms/cmsCollectorPolicy.hpp"
  32 #include "gc/cms/cmsGCStats.hpp"
  33 #include "gc/cms/cmsHeap.hpp"
  34 #include "gc/cms/cmsOopClosures.inline.hpp"
  35 #include "gc/cms/compactibleFreeListSpace.hpp"
  36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  37 #include "gc/cms/concurrentMarkSweepThread.hpp"
  38 #include "gc/cms/parNewGeneration.hpp"
  39 #include "gc/cms/vmCMSOperations.hpp"
  40 #include "gc/serial/genMarkSweep.hpp"
  41 #include "gc/serial/tenuredGeneration.hpp"
  42 #include "gc/shared/adaptiveSizePolicy.hpp"
  43 #include "gc/shared/cardGeneration.inline.hpp"
  44 #include "gc/shared/cardTableRS.hpp"
  45 #include "gc/shared/collectedHeap.inline.hpp"
  46 #include "gc/shared/collectorCounters.hpp"
  47 #include "gc/shared/collectorPolicy.hpp"
  48 #include "gc/shared/gcLocker.hpp"
  49 #include "gc/shared/gcPolicyCounters.hpp"
  50 #include "gc/shared/gcTimer.hpp"
  51 #include "gc/shared/gcTrace.hpp"
  52 #include "gc/shared/gcTraceTime.inline.hpp"
  53 #include "gc/shared/genCollectedHeap.hpp"
  54 #include "gc/shared/genOopClosures.inline.hpp"
  55 #include "gc/shared/isGCActiveMark.hpp"
  56 #include "gc/shared/referencePolicy.hpp"
  57 #include "gc/shared/strongRootsScope.hpp"
  58 #include "gc/shared/taskqueue.inline.hpp"
  59 #include "gc/shared/weakProcessor.hpp"
  60 #include "logging/log.hpp"
  61 #include "logging/logStream.hpp"
  62 #include "memory/allocation.hpp"
  63 #include "memory/binaryTreeDictionary.inline.hpp"
  64 #include "memory/iterator.inline.hpp"
  65 #include "memory/padded.hpp"
  66 #include "memory/resourceArea.hpp"
  67 #include "oops/access.inline.hpp"
  68 #include "oops/oop.inline.hpp"
  69 #include "prims/jvmtiExport.hpp"
  70 #include "runtime/atomic.hpp"
  71 #include "runtime/flags/flagSetting.hpp"
  72 #include "runtime/globals_extension.hpp"
  73 #include "runtime/handles.inline.hpp"
  74 #include "runtime/java.hpp"
  75 #include "runtime/orderAccess.inline.hpp"
  76 #include "runtime/timer.hpp"
  77 #include "runtime/vmThread.hpp"
  78 #include "services/memoryService.hpp"
  79 #include "services/runtimeService.hpp"
  80 #include "utilities/align.hpp"
  81 #include "utilities/stack.inline.hpp"
  82 
  83 // statics
  84 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  85 bool CMSCollector::_full_gc_requested = false;
  86 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  87 
  88 //////////////////////////////////////////////////////////////////
  89 // In support of CMS/VM thread synchronization
  90 //////////////////////////////////////////////////////////////////
  91 // We split use of the CGC_lock into 2 "levels".
  92 // The low-level locking is of the usual CGC_lock monitor. We introduce
  93 // a higher level "token" (hereafter "CMS token") built on top of the
  94 // low level monitor (hereafter "CGC lock").
  95 // The token-passing protocol gives priority to the VM thread. The
  96 // CMS-lock doesn't provide any fairness guarantees, but clients
  97 // should ensure that it is only held for very short, bounded
  98 // durations.
  99 //
 100 // When either of the CMS thread or the VM thread is involved in
 101 // collection operations during which it does not want the other
 102 // thread to interfere, it obtains the CMS token.
 103 //
 104 // If either thread tries to get the token while the other has
 105 // it, that thread waits. However, if the VM thread and CMS thread
 106 // both want the token, then the VM thread gets priority while the
 107 // CMS thread waits. This ensures, for instance, that the "concurrent"
 108 // phases of the CMS thread's work do not block out the VM thread
 109 // for long periods of time as the CMS thread continues to hog
 110 // the token. (See bug 4616232).
 111 //
 112 // The baton-passing functions are, however, controlled by the
 113 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 114 // and here the low-level CMS lock, not the high level token,
 115 // ensures mutual exclusion.
 116 //
 117 // Two important conditions that we have to satisfy:
 118 // 1. if a thread does a low-level wait on the CMS lock, then it
 119 //    relinquishes the CMS token if it were holding that token
 120 //    when it acquired the low-level CMS lock.
 121 // 2. any low-level notifications on the low-level lock
 122 //    should only be sent when a thread has relinquished the token.
 123 //
 124 // In the absence of either property, we'd have potential deadlock.
 125 //
 126 // We protect each of the CMS (concurrent and sequential) phases
 127 // with the CMS _token_, not the CMS _lock_.
 128 //
 129 // The only code protected by CMS lock is the token acquisition code
 130 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 131 // baton-passing code.
 132 //
 133 // Unfortunately, i couldn't come up with a good abstraction to factor and
 134 // hide the naked CGC_lock manipulation in the baton-passing code
 135 // further below. That's something we should try to do. Also, the proof
 136 // of correctness of this 2-level locking scheme is far from obvious,
 137 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 138 // that there may be a theoretical possibility of delay/starvation in the
 139 // low-level lock/wait/notify scheme used for the baton-passing because of
 140 // potential interference with the priority scheme embodied in the
 141 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 142 // invocation further below and marked with "XXX 20011219YSR".
 143 // Indeed, as we note elsewhere, this may become yet more slippery
 144 // in the presence of multiple CMS and/or multiple VM threads. XXX
 145 
 146 class CMSTokenSync: public StackObj {
 147  private:
 148   bool _is_cms_thread;
 149  public:
 150   CMSTokenSync(bool is_cms_thread):
 151     _is_cms_thread(is_cms_thread) {
 152     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 153            "Incorrect argument to constructor");
 154     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 155   }
 156 
 157   ~CMSTokenSync() {
 158     assert(_is_cms_thread ?
 159              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 160              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 161           "Incorrect state");
 162     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 163   }
 164 };
 165 
 166 // Convenience class that does a CMSTokenSync, and then acquires
 167 // upto three locks.
 168 class CMSTokenSyncWithLocks: public CMSTokenSync {
 169  private:
 170   // Note: locks are acquired in textual declaration order
 171   // and released in the opposite order
 172   MutexLockerEx _locker1, _locker2, _locker3;
 173  public:
 174   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 175                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 176     CMSTokenSync(is_cms_thread),
 177     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 178     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 179     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 180   { }
 181 };
 182 
 183 
 184 //////////////////////////////////////////////////////////////////
 185 //  Concurrent Mark-Sweep Generation /////////////////////////////
 186 //////////////////////////////////////////////////////////////////
 187 
 188 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 189 
 190 // This struct contains per-thread things necessary to support parallel
 191 // young-gen collection.
 192 class CMSParGCThreadState: public CHeapObj<mtGC> {
 193  public:
 194   CompactibleFreeListSpaceLAB lab;
 195   PromotionInfo promo;
 196 
 197   // Constructor.
 198   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 199     promo.setSpace(cfls);
 200   }
 201 };
 202 
 203 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 204      ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
 205   CardGeneration(rs, initial_byte_size, ct),
 206   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 207   _did_compact(false)
 208 {
 209   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 210   HeapWord* end    = (HeapWord*) _virtual_space.high();
 211 
 212   _direct_allocated_words = 0;
 213   NOT_PRODUCT(
 214     _numObjectsPromoted = 0;
 215     _numWordsPromoted = 0;
 216     _numObjectsAllocated = 0;
 217     _numWordsAllocated = 0;
 218   )
 219 
 220   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
 221   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 222   _cmsSpace->_old_gen = this;
 223 
 224   _gc_stats = new CMSGCStats();
 225 
 226   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 227   // offsets match. The ability to tell free chunks from objects
 228   // depends on this property.
 229   debug_only(
 230     FreeChunk* junk = NULL;
 231     assert(UseCompressedClassPointers ||
 232            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 233            "Offset of FreeChunk::_prev within FreeChunk must match"
 234            "  that of OopDesc::_klass within OopDesc");
 235   )
 236 
 237   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 238   for (uint i = 0; i < ParallelGCThreads; i++) {
 239     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 240   }
 241 
 242   _incremental_collection_failed = false;
 243   // The "dilatation_factor" is the expansion that can occur on
 244   // account of the fact that the minimum object size in the CMS
 245   // generation may be larger than that in, say, a contiguous young
 246   //  generation.
 247   // Ideally, in the calculation below, we'd compute the dilatation
 248   // factor as: MinChunkSize/(promoting_gen's min object size)
 249   // Since we do not have such a general query interface for the
 250   // promoting generation, we'll instead just use the minimum
 251   // object size (which today is a header's worth of space);
 252   // note that all arithmetic is in units of HeapWords.
 253   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 254   assert(_dilatation_factor >= 1.0, "from previous assert");
 255 }
 256 
 257 
 258 // The field "_initiating_occupancy" represents the occupancy percentage
 259 // at which we trigger a new collection cycle.  Unless explicitly specified
 260 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 261 // is calculated by:
 262 //
 263 //   Let "f" be MinHeapFreeRatio in
 264 //
 265 //    _initiating_occupancy = 100-f +
 266 //                           f * (CMSTriggerRatio/100)
 267 //   where CMSTriggerRatio is the argument "tr" below.
 268 //
 269 // That is, if we assume the heap is at its desired maximum occupancy at the
 270 // end of a collection, we let CMSTriggerRatio of the (purported) free
 271 // space be allocated before initiating a new collection cycle.
 272 //
 273 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 274   assert(io <= 100 && tr <= 100, "Check the arguments");
 275   if (io >= 0) {
 276     _initiating_occupancy = (double)io / 100.0;
 277   } else {
 278     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 279                              (double)(tr * MinHeapFreeRatio) / 100.0)
 280                             / 100.0;
 281   }
 282 }
 283 
 284 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 285   assert(collector() != NULL, "no collector");
 286   collector()->ref_processor_init();
 287 }
 288 
 289 void CMSCollector::ref_processor_init() {
 290   if (_ref_processor == NULL) {
 291     // Allocate and initialize a reference processor
 292     _ref_processor =
 293       new SpanReferenceProcessor(_span,                               // span
 294                                  (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 295                                  ParallelGCThreads,                   // mt processing degree
 296                                  _cmsGen->refs_discovery_is_mt(),     // mt discovery
 297                                  MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 298                                  _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 299                                  &_is_alive_closure);                 // closure for liveness info
 300     // Initialize the _ref_processor field of CMSGen
 301     _cmsGen->set_ref_processor(_ref_processor);
 302 
 303   }
 304 }
 305 
 306 AdaptiveSizePolicy* CMSCollector::size_policy() {
 307   return CMSHeap::heap()->size_policy();
 308 }
 309 
 310 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 311 
 312   const char* gen_name = "old";
 313   GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
 314   // Generation Counters - generation 1, 1 subspace
 315   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 316       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 317 
 318   _space_counters = new GSpaceCounters(gen_name, 0,
 319                                        _virtual_space.reserved_size(),
 320                                        this, _gen_counters);
 321 }
 322 
 323 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 324   _cms_gen(cms_gen)
 325 {
 326   assert(alpha <= 100, "bad value");
 327   _saved_alpha = alpha;
 328 
 329   // Initialize the alphas to the bootstrap value of 100.
 330   _gc0_alpha = _cms_alpha = 100;
 331 
 332   _cms_begin_time.update();
 333   _cms_end_time.update();
 334 
 335   _gc0_duration = 0.0;
 336   _gc0_period = 0.0;
 337   _gc0_promoted = 0;
 338 
 339   _cms_duration = 0.0;
 340   _cms_period = 0.0;
 341   _cms_allocated = 0;
 342 
 343   _cms_used_at_gc0_begin = 0;
 344   _cms_used_at_gc0_end = 0;
 345   _allow_duty_cycle_reduction = false;
 346   _valid_bits = 0;
 347 }
 348 
 349 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 350   // TBD: CR 6909490
 351   return 1.0;
 352 }
 353 
 354 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 355 }
 356 
 357 // If promotion failure handling is on use
 358 // the padded average size of the promotion for each
 359 // young generation collection.
 360 double CMSStats::time_until_cms_gen_full() const {
 361   size_t cms_free = _cms_gen->cmsSpace()->free();
 362   CMSHeap* heap = CMSHeap::heap();
 363   size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
 364                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 365   if (cms_free > expected_promotion) {
 366     // Start a cms collection if there isn't enough space to promote
 367     // for the next young collection.  Use the padded average as
 368     // a safety factor.
 369     cms_free -= expected_promotion;
 370 
 371     // Adjust by the safety factor.
 372     double cms_free_dbl = (double)cms_free;
 373     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 374     // Apply a further correction factor which tries to adjust
 375     // for recent occurance of concurrent mode failures.
 376     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 377     cms_free_dbl = cms_free_dbl * cms_adjustment;
 378 
 379     log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 380                   cms_free, expected_promotion);
 381     log_trace(gc)("  cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
 382     // Add 1 in case the consumption rate goes to zero.
 383     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 384   }
 385   return 0.0;
 386 }
 387 
 388 // Compare the duration of the cms collection to the
 389 // time remaining before the cms generation is empty.
 390 // Note that the time from the start of the cms collection
 391 // to the start of the cms sweep (less than the total
 392 // duration of the cms collection) can be used.  This
 393 // has been tried and some applications experienced
 394 // promotion failures early in execution.  This was
 395 // possibly because the averages were not accurate
 396 // enough at the beginning.
 397 double CMSStats::time_until_cms_start() const {
 398   // We add "gc0_period" to the "work" calculation
 399   // below because this query is done (mostly) at the
 400   // end of a scavenge, so we need to conservatively
 401   // account for that much possible delay
 402   // in the query so as to avoid concurrent mode failures
 403   // due to starting the collection just a wee bit too
 404   // late.
 405   double work = cms_duration() + gc0_period();
 406   double deadline = time_until_cms_gen_full();
 407   // If a concurrent mode failure occurred recently, we want to be
 408   // more conservative and halve our expected time_until_cms_gen_full()
 409   if (work > deadline) {
 410     log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
 411                           cms_duration(), gc0_period(), time_until_cms_gen_full());
 412     return 0.0;
 413   }
 414   return work - deadline;
 415 }
 416 
 417 #ifndef PRODUCT
 418 void CMSStats::print_on(outputStream *st) const {
 419   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 420   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 421                gc0_duration(), gc0_period(), gc0_promoted());
 422   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 423             cms_duration(), cms_period(), cms_allocated());
 424   st->print(",cms_since_beg=%g,cms_since_end=%g",
 425             cms_time_since_begin(), cms_time_since_end());
 426   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 427             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 428 
 429   if (valid()) {
 430     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 431               promotion_rate(), cms_allocation_rate());
 432     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 433               cms_consumption_rate(), time_until_cms_gen_full());
 434   }
 435   st->cr();
 436 }
 437 #endif // #ifndef PRODUCT
 438 
 439 CMSCollector::CollectorState CMSCollector::_collectorState =
 440                              CMSCollector::Idling;
 441 bool CMSCollector::_foregroundGCIsActive = false;
 442 bool CMSCollector::_foregroundGCShouldWait = false;
 443 
 444 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 445                            CardTableRS*                   ct,
 446                            ConcurrentMarkSweepPolicy*     cp):
 447   _cmsGen(cmsGen),
 448   _ct(ct),
 449   _ref_processor(NULL),    // will be set later
 450   _conc_workers(NULL),     // may be set later
 451   _abort_preclean(false),
 452   _start_sampling(false),
 453   _between_prologue_and_epilogue(false),
 454   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 455   _modUnionTable((CardTable::card_shift - LogHeapWordSize),
 456                  -1 /* lock-free */, "No_lock" /* dummy */),
 457   _modUnionClosurePar(&_modUnionTable),
 458   // Adjust my span to cover old (cms) gen
 459   _span(cmsGen->reserved()),
 460   // Construct the is_alive_closure with _span & markBitMap
 461   _is_alive_closure(_span, &_markBitMap),
 462   _restart_addr(NULL),
 463   _overflow_list(NULL),
 464   _stats(cmsGen),
 465   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 466                              //verify that this lock should be acquired with safepoint check.
 467                              Monitor::_safepoint_check_sometimes)),
 468   _eden_chunk_array(NULL),     // may be set in ctor body
 469   _eden_chunk_capacity(0),     // -- ditto --
 470   _eden_chunk_index(0),        // -- ditto --
 471   _survivor_plab_array(NULL),  // -- ditto --
 472   _survivor_chunk_array(NULL), // -- ditto --
 473   _survivor_chunk_capacity(0), // -- ditto --
 474   _survivor_chunk_index(0),    // -- ditto --
 475   _ser_pmc_preclean_ovflw(0),
 476   _ser_kac_preclean_ovflw(0),
 477   _ser_pmc_remark_ovflw(0),
 478   _par_pmc_remark_ovflw(0),
 479   _ser_kac_ovflw(0),
 480   _par_kac_ovflw(0),
 481 #ifndef PRODUCT
 482   _num_par_pushes(0),
 483 #endif
 484   _collection_count_start(0),
 485   _verifying(false),
 486   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 487   _completed_initialization(false),
 488   _collector_policy(cp),
 489   _should_unload_classes(CMSClassUnloadingEnabled),
 490   _concurrent_cycles_since_last_unload(0),
 491   _roots_scanning_options(GenCollectedHeap::SO_None),
 492   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 493   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 494   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 495   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 496   _cms_start_registered(false)
 497 {
 498   // Now expand the span and allocate the collection support structures
 499   // (MUT, marking bit map etc.) to cover both generations subject to
 500   // collection.
 501 
 502   // For use by dirty card to oop closures.
 503   _cmsGen->cmsSpace()->set_collector(this);
 504 
 505   // Allocate MUT and marking bit map
 506   {
 507     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 508     if (!_markBitMap.allocate(_span)) {
 509       log_warning(gc)("Failed to allocate CMS Bit Map");
 510       return;
 511     }
 512     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 513   }
 514   {
 515     _modUnionTable.allocate(_span);
 516     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 517   }
 518 
 519   if (!_markStack.allocate(MarkStackSize)) {
 520     log_warning(gc)("Failed to allocate CMS Marking Stack");
 521     return;
 522   }
 523 
 524   // Support for multi-threaded concurrent phases
 525   if (CMSConcurrentMTEnabled) {
 526     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 527       // just for now
 528       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
 529     }
 530     if (ConcGCThreads > 1) {
 531       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 532                                  ConcGCThreads, true);
 533       if (_conc_workers == NULL) {
 534         log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
 535         CMSConcurrentMTEnabled = false;
 536       } else {
 537         _conc_workers->initialize_workers();
 538       }
 539     } else {
 540       CMSConcurrentMTEnabled = false;
 541     }
 542   }
 543   if (!CMSConcurrentMTEnabled) {
 544     ConcGCThreads = 0;
 545   } else {
 546     // Turn off CMSCleanOnEnter optimization temporarily for
 547     // the MT case where it's not fixed yet; see 6178663.
 548     CMSCleanOnEnter = false;
 549   }
 550   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 551          "Inconsistency");
 552   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 553   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 554 
 555   // Parallel task queues; these are shared for the
 556   // concurrent and stop-world phases of CMS, but
 557   // are not shared with parallel scavenge (ParNew).
 558   {
 559     uint i;
 560     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 561 
 562     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 563          || ParallelRefProcEnabled)
 564         && num_queues > 0) {
 565       _task_queues = new OopTaskQueueSet(num_queues);
 566       if (_task_queues == NULL) {
 567         log_warning(gc)("task_queues allocation failure.");
 568         return;
 569       }
 570       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 571       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 572       for (i = 0; i < num_queues; i++) {
 573         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 574         if (q == NULL) {
 575           log_warning(gc)("work_queue allocation failure.");
 576           return;
 577         }
 578         _task_queues->register_queue(i, q);
 579       }
 580       for (i = 0; i < num_queues; i++) {
 581         _task_queues->queue(i)->initialize();
 582         _hash_seed[i] = 17;  // copied from ParNew
 583       }
 584     }
 585   }
 586 
 587   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 588 
 589   // Clip CMSBootstrapOccupancy between 0 and 100.
 590   _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
 591 
 592   // Now tell CMS generations the identity of their collector
 593   ConcurrentMarkSweepGeneration::set_collector(this);
 594 
 595   // Create & start a CMS thread for this CMS collector
 596   _cmsThread = ConcurrentMarkSweepThread::start(this);
 597   assert(cmsThread() != NULL, "CMS Thread should have been created");
 598   assert(cmsThread()->collector() == this,
 599          "CMS Thread should refer to this gen");
 600   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 601 
 602   // Support for parallelizing young gen rescan
 603   CMSHeap* heap = CMSHeap::heap();
 604   assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 605   _young_gen = (ParNewGeneration*)heap->young_gen();
 606   if (heap->supports_inline_contig_alloc()) {
 607     _top_addr = heap->top_addr();
 608     _end_addr = heap->end_addr();
 609     assert(_young_gen != NULL, "no _young_gen");
 610     _eden_chunk_index = 0;
 611     _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
 612     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 613   }
 614 
 615   // Support for parallelizing survivor space rescan
 616   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 617     const size_t max_plab_samples =
 618       _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
 619 
 620     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 621     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 622     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 623     _survivor_chunk_capacity = max_plab_samples;
 624     for (uint i = 0; i < ParallelGCThreads; i++) {
 625       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 626       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 627       assert(cur->end() == 0, "Should be 0");
 628       assert(cur->array() == vec, "Should be vec");
 629       assert(cur->capacity() == max_plab_samples, "Error");
 630     }
 631   }
 632 
 633   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 634   _gc_counters = new CollectorCounters("CMS", 1);
 635   _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
 636   _completed_initialization = true;
 637   _inter_sweep_timer.start();  // start of time
 638 }
 639 
 640 const char* ConcurrentMarkSweepGeneration::name() const {
 641   return "concurrent mark-sweep generation";
 642 }
 643 void ConcurrentMarkSweepGeneration::update_counters() {
 644   if (UsePerfData) {
 645     _space_counters->update_all();
 646     _gen_counters->update_all();
 647   }
 648 }
 649 
 650 // this is an optimized version of update_counters(). it takes the
 651 // used value as a parameter rather than computing it.
 652 //
 653 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 654   if (UsePerfData) {
 655     _space_counters->update_used(used);
 656     _space_counters->update_capacity();
 657     _gen_counters->update_all();
 658   }
 659 }
 660 
 661 void ConcurrentMarkSweepGeneration::print() const {
 662   Generation::print();
 663   cmsSpace()->print();
 664 }
 665 
 666 #ifndef PRODUCT
 667 void ConcurrentMarkSweepGeneration::print_statistics() {
 668   cmsSpace()->printFLCensus(0);
 669 }
 670 #endif
 671 
 672 size_t
 673 ConcurrentMarkSweepGeneration::contiguous_available() const {
 674   // dld proposes an improvement in precision here. If the committed
 675   // part of the space ends in a free block we should add that to
 676   // uncommitted size in the calculation below. Will make this
 677   // change later, staying with the approximation below for the
 678   // time being. -- ysr.
 679   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 680 }
 681 
 682 size_t
 683 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 684   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 685 }
 686 
 687 size_t ConcurrentMarkSweepGeneration::max_available() const {
 688   return free() + _virtual_space.uncommitted_size();
 689 }
 690 
 691 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 692   size_t available = max_available();
 693   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 694   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 695   log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
 696                            res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
 697   return res;
 698 }
 699 
 700 // At a promotion failure dump information on block layout in heap
 701 // (cms old generation).
 702 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 703   Log(gc, promotion) log;
 704   if (log.is_trace()) {
 705     LogStream ls(log.trace());
 706     cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
 707   }
 708 }
 709 
 710 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 711   // Clear the promotion information.  These pointers can be adjusted
 712   // along with all the other pointers into the heap but
 713   // compaction is expected to be a rare event with
 714   // a heap using cms so don't do it without seeing the need.
 715   for (uint i = 0; i < ParallelGCThreads; i++) {
 716     _par_gc_thread_states[i]->promo.reset();
 717   }
 718 }
 719 
 720 void ConcurrentMarkSweepGeneration::compute_new_size() {
 721   assert_locked_or_safepoint(Heap_lock);
 722 
 723   // If incremental collection failed, we just want to expand
 724   // to the limit.
 725   if (incremental_collection_failed()) {
 726     clear_incremental_collection_failed();
 727     grow_to_reserved();
 728     return;
 729   }
 730 
 731   // The heap has been compacted but not reset yet.
 732   // Any metric such as free() or used() will be incorrect.
 733 
 734   CardGeneration::compute_new_size();
 735 
 736   // Reset again after a possible resizing
 737   if (did_compact()) {
 738     cmsSpace()->reset_after_compaction();
 739   }
 740 }
 741 
 742 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
 743   assert_locked_or_safepoint(Heap_lock);
 744 
 745   // If incremental collection failed, we just want to expand
 746   // to the limit.
 747   if (incremental_collection_failed()) {
 748     clear_incremental_collection_failed();
 749     grow_to_reserved();
 750     return;
 751   }
 752 
 753   double free_percentage = ((double) free()) / capacity();
 754   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 755   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 756 
 757   // compute expansion delta needed for reaching desired free percentage
 758   if (free_percentage < desired_free_percentage) {
 759     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 760     assert(desired_capacity >= capacity(), "invalid expansion size");
 761     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 762     Log(gc) log;
 763     if (log.is_trace()) {
 764       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 765       log.trace("From compute_new_size: ");
 766       log.trace("  Free fraction %f", free_percentage);
 767       log.trace("  Desired free fraction %f", desired_free_percentage);
 768       log.trace("  Maximum free fraction %f", maximum_free_percentage);
 769       log.trace("  Capacity " SIZE_FORMAT, capacity() / 1000);
 770       log.trace("  Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
 771       CMSHeap* heap = CMSHeap::heap();
 772       assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
 773       size_t young_size = heap->young_gen()->capacity();
 774       log.trace("  Young gen size " SIZE_FORMAT, young_size / 1000);
 775       log.trace("  unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
 776       log.trace("  contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
 777       log.trace("  Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
 778     }
 779     // safe if expansion fails
 780     expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 781     log.trace("  Expanded free fraction %f", ((double) free()) / capacity());
 782   } else {
 783     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 784     assert(desired_capacity <= capacity(), "invalid expansion size");
 785     size_t shrink_bytes = capacity() - desired_capacity;
 786     // Don't shrink unless the delta is greater than the minimum shrink we want
 787     if (shrink_bytes >= MinHeapDeltaBytes) {
 788       shrink_free_list_by(shrink_bytes);
 789     }
 790   }
 791 }
 792 
 793 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 794   return cmsSpace()->freelistLock();
 795 }
 796 
 797 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
 798   CMSSynchronousYieldRequest yr;
 799   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
 800   return have_lock_and_allocate(size, tlab);
 801 }
 802 
 803 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 804                                                                 bool   tlab /* ignored */) {
 805   assert_lock_strong(freelistLock());
 806   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 807   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 808   // Allocate the object live (grey) if the background collector has
 809   // started marking. This is necessary because the marker may
 810   // have passed this address and consequently this object will
 811   // not otherwise be greyed and would be incorrectly swept up.
 812   // Note that if this object contains references, the writing
 813   // of those references will dirty the card containing this object
 814   // allowing the object to be blackened (and its references scanned)
 815   // either during a preclean phase or at the final checkpoint.
 816   if (res != NULL) {
 817     // We may block here with an uninitialized object with
 818     // its mark-bit or P-bits not yet set. Such objects need
 819     // to be safely navigable by block_start().
 820     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
 821     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
 822     collector()->direct_allocated(res, adjustedSize);
 823     _direct_allocated_words += adjustedSize;
 824     // allocation counters
 825     NOT_PRODUCT(
 826       _numObjectsAllocated++;
 827       _numWordsAllocated += (int)adjustedSize;
 828     )
 829   }
 830   return res;
 831 }
 832 
 833 // In the case of direct allocation by mutators in a generation that
 834 // is being concurrently collected, the object must be allocated
 835 // live (grey) if the background collector has started marking.
 836 // This is necessary because the marker may
 837 // have passed this address and consequently this object will
 838 // not otherwise be greyed and would be incorrectly swept up.
 839 // Note that if this object contains references, the writing
 840 // of those references will dirty the card containing this object
 841 // allowing the object to be blackened (and its references scanned)
 842 // either during a preclean phase or at the final checkpoint.
 843 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
 844   assert(_markBitMap.covers(start, size), "Out of bounds");
 845   if (_collectorState >= Marking) {
 846     MutexLockerEx y(_markBitMap.lock(),
 847                     Mutex::_no_safepoint_check_flag);
 848     // [see comments preceding SweepClosure::do_blk() below for details]
 849     //
 850     // Can the P-bits be deleted now?  JJJ
 851     //
 852     // 1. need to mark the object as live so it isn't collected
 853     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
 854     // 3. need to mark the end of the object so marking, precleaning or sweeping
 855     //    can skip over uninitialized or unparsable objects. An allocated
 856     //    object is considered uninitialized for our purposes as long as
 857     //    its klass word is NULL.  All old gen objects are parsable
 858     //    as soon as they are initialized.)
 859     _markBitMap.mark(start);          // object is live
 860     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
 861     _markBitMap.mark(start + size - 1);
 862                                       // mark end of object
 863   }
 864   // check that oop looks uninitialized
 865   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
 866 }
 867 
 868 void CMSCollector::promoted(bool par, HeapWord* start,
 869                             bool is_obj_array, size_t obj_size) {
 870   assert(_markBitMap.covers(start), "Out of bounds");
 871   // See comment in direct_allocated() about when objects should
 872   // be allocated live.
 873   if (_collectorState >= Marking) {
 874     // we already hold the marking bit map lock, taken in
 875     // the prologue
 876     if (par) {
 877       _markBitMap.par_mark(start);
 878     } else {
 879       _markBitMap.mark(start);
 880     }
 881     // We don't need to mark the object as uninitialized (as
 882     // in direct_allocated above) because this is being done with the
 883     // world stopped and the object will be initialized by the
 884     // time the marking, precleaning or sweeping get to look at it.
 885     // But see the code for copying objects into the CMS generation,
 886     // where we need to ensure that concurrent readers of the
 887     // block offset table are able to safely navigate a block that
 888     // is in flux from being free to being allocated (and in
 889     // transition while being copied into) and subsequently
 890     // becoming a bona-fide object when the copy/promotion is complete.
 891     assert(SafepointSynchronize::is_at_safepoint(),
 892            "expect promotion only at safepoints");
 893 
 894     if (_collectorState < Sweeping) {
 895       // Mark the appropriate cards in the modUnionTable, so that
 896       // this object gets scanned before the sweep. If this is
 897       // not done, CMS generation references in the object might
 898       // not get marked.
 899       // For the case of arrays, which are otherwise precisely
 900       // marked, we need to dirty the entire array, not just its head.
 901       if (is_obj_array) {
 902         // The [par_]mark_range() method expects mr.end() below to
 903         // be aligned to the granularity of a bit's representation
 904         // in the heap. In the case of the MUT below, that's a
 905         // card size.
 906         MemRegion mr(start,
 907                      align_up(start + obj_size,
 908                               CardTable::card_size /* bytes */));
 909         if (par) {
 910           _modUnionTable.par_mark_range(mr);
 911         } else {
 912           _modUnionTable.mark_range(mr);
 913         }
 914       } else {  // not an obj array; we can just mark the head
 915         if (par) {
 916           _modUnionTable.par_mark(start);
 917         } else {
 918           _modUnionTable.mark(start);
 919         }
 920       }
 921     }
 922   }
 923 }
 924 
 925 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
 926   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
 927   // allocate, copy and if necessary update promoinfo --
 928   // delegate to underlying space.
 929   assert_lock_strong(freelistLock());
 930 
 931 #ifndef PRODUCT
 932   if (CMSHeap::heap()->promotion_should_fail()) {
 933     return NULL;
 934   }
 935 #endif  // #ifndef PRODUCT
 936 
 937   oop res = _cmsSpace->promote(obj, obj_size);
 938   if (res == NULL) {
 939     // expand and retry
 940     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
 941     expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
 942     // Since this is the old generation, we don't try to promote
 943     // into a more senior generation.
 944     res = _cmsSpace->promote(obj, obj_size);
 945   }
 946   if (res != NULL) {
 947     // See comment in allocate() about when objects should
 948     // be allocated live.
 949     assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
 950     collector()->promoted(false,           // Not parallel
 951                           (HeapWord*)res, obj->is_objArray(), obj_size);
 952     // promotion counters
 953     NOT_PRODUCT(
 954       _numObjectsPromoted++;
 955       _numWordsPromoted +=
 956         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
 957     )
 958   }
 959   return res;
 960 }
 961 
 962 
 963 // IMPORTANT: Notes on object size recognition in CMS.
 964 // ---------------------------------------------------
 965 // A block of storage in the CMS generation is always in
 966 // one of three states. A free block (FREE), an allocated
 967 // object (OBJECT) whose size() method reports the correct size,
 968 // and an intermediate state (TRANSIENT) in which its size cannot
 969 // be accurately determined.
 970 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
 971 // -----------------------------------------------------
 972 // FREE:      klass_word & 1 == 1; mark_word holds block size
 973 //
 974 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
 975 //            obj->size() computes correct size
 976 //
 977 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 978 //
 979 // STATE IDENTIFICATION: (64 bit+COOPS)
 980 // ------------------------------------
 981 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
 982 //
 983 // OBJECT:    klass_word installed; klass_word != 0;
 984 //            obj->size() computes correct size
 985 //
 986 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 987 //
 988 //
 989 // STATE TRANSITION DIAGRAM
 990 //
 991 //        mut / parnew                     mut  /  parnew
 992 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
 993 //  ^                                                                   |
 994 //  |------------------------ DEAD <------------------------------------|
 995 //         sweep                            mut
 996 //
 997 // While a block is in TRANSIENT state its size cannot be determined
 998 // so readers will either need to come back later or stall until
 999 // the size can be determined. Note that for the case of direct
1000 // allocation, P-bits, when available, may be used to determine the
1001 // size of an object that may not yet have been initialized.
1002 
1003 // Things to support parallel young-gen collection.
1004 oop
1005 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1006                                            oop old, markOop m,
1007                                            size_t word_sz) {
1008 #ifndef PRODUCT
1009   if (CMSHeap::heap()->promotion_should_fail()) {
1010     return NULL;
1011   }
1012 #endif  // #ifndef PRODUCT
1013 
1014   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1015   PromotionInfo* promoInfo = &ps->promo;
1016   // if we are tracking promotions, then first ensure space for
1017   // promotion (including spooling space for saving header if necessary).
1018   // then allocate and copy, then track promoted info if needed.
1019   // When tracking (see PromotionInfo::track()), the mark word may
1020   // be displaced and in this case restoration of the mark word
1021   // occurs in the (oop_since_save_marks_)iterate phase.
1022   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1023     // Out of space for allocating spooling buffers;
1024     // try expanding and allocating spooling buffers.
1025     if (!expand_and_ensure_spooling_space(promoInfo)) {
1026       return NULL;
1027     }
1028   }
1029   assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1030   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1031   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1032   if (obj_ptr == NULL) {
1033      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1034      if (obj_ptr == NULL) {
1035        return NULL;
1036      }
1037   }
1038   oop obj = oop(obj_ptr);
1039   OrderAccess::storestore();
1040   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1041   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1042   // IMPORTANT: See note on object initialization for CMS above.
1043   // Otherwise, copy the object.  Here we must be careful to insert the
1044   // klass pointer last, since this marks the block as an allocated object.
1045   // Except with compressed oops it's the mark word.
1046   HeapWord* old_ptr = (HeapWord*)old;
1047   // Restore the mark word copied above.
1048   obj->set_mark_raw(m);
1049   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1050   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1051   OrderAccess::storestore();
1052 
1053   if (UseCompressedClassPointers) {
1054     // Copy gap missed by (aligned) header size calculation below
1055     obj->set_klass_gap(old->klass_gap());
1056   }
1057   if (word_sz > (size_t)oopDesc::header_size()) {
1058     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1059                                  obj_ptr + oopDesc::header_size(),
1060                                  word_sz - oopDesc::header_size());
1061   }
1062 
1063   // Now we can track the promoted object, if necessary.  We take care
1064   // to delay the transition from uninitialized to full object
1065   // (i.e., insertion of klass pointer) until after, so that it
1066   // atomically becomes a promoted object.
1067   if (promoInfo->tracking()) {
1068     promoInfo->track((PromotedObject*)obj, old->klass());
1069   }
1070   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1071   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1072   assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1073 
1074   // Finally, install the klass pointer (this should be volatile).
1075   OrderAccess::storestore();
1076   obj->set_klass(old->klass());
1077   // We should now be able to calculate the right size for this object
1078   assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1079 
1080   collector()->promoted(true,          // parallel
1081                         obj_ptr, old->is_objArray(), word_sz);
1082 
1083   NOT_PRODUCT(
1084     Atomic::inc(&_numObjectsPromoted);
1085     Atomic::add(alloc_sz, &_numWordsPromoted);
1086   )
1087 
1088   return obj;
1089 }
1090 
1091 void
1092 ConcurrentMarkSweepGeneration::
1093 par_promote_alloc_done(int thread_num) {
1094   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1095   ps->lab.retire(thread_num);
1096 }
1097 
1098 void
1099 ConcurrentMarkSweepGeneration::
1100 par_oop_since_save_marks_iterate_done(int thread_num) {
1101   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1102   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1103   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1104 
1105   // Because card-scanning has been completed, subsequent phases
1106   // (e.g., reference processing) will not need to recognize which
1107   // objects have been promoted during this GC. So, we can now disable
1108   // promotion tracking.
1109   ps->promo.stopTrackingPromotions();
1110 }
1111 
1112 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1113                                                    size_t size,
1114                                                    bool   tlab)
1115 {
1116   // We allow a STW collection only if a full
1117   // collection was requested.
1118   return full || should_allocate(size, tlab); // FIX ME !!!
1119   // This and promotion failure handling are connected at the
1120   // hip and should be fixed by untying them.
1121 }
1122 
1123 bool CMSCollector::shouldConcurrentCollect() {
1124   LogTarget(Trace, gc) log;
1125 
1126   if (_full_gc_requested) {
1127     log.print("CMSCollector: collect because of explicit  gc request (or GCLocker)");
1128     return true;
1129   }
1130 
1131   FreelistLocker x(this);
1132   // ------------------------------------------------------------------
1133   // Print out lots of information which affects the initiation of
1134   // a collection.
1135   if (log.is_enabled() && stats().valid()) {
1136     log.print("CMSCollector shouldConcurrentCollect: ");
1137 
1138     LogStream out(log);
1139     stats().print_on(&out);
1140 
1141     log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1142     log.print("free=" SIZE_FORMAT, _cmsGen->free());
1143     log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1144     log.print("promotion_rate=%g", stats().promotion_rate());
1145     log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1146     log.print("occupancy=%3.7f", _cmsGen->occupancy());
1147     log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1148     log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1149     log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1150     log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1151   }
1152   // ------------------------------------------------------------------
1153 
1154   // If the estimated time to complete a cms collection (cms_duration())
1155   // is less than the estimated time remaining until the cms generation
1156   // is full, start a collection.
1157   if (!UseCMSInitiatingOccupancyOnly) {
1158     if (stats().valid()) {
1159       if (stats().time_until_cms_start() == 0.0) {
1160         return true;
1161       }
1162     } else {
1163       // We want to conservatively collect somewhat early in order
1164       // to try and "bootstrap" our CMS/promotion statistics;
1165       // this branch will not fire after the first successful CMS
1166       // collection because the stats should then be valid.
1167       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1168         log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1169                   _cmsGen->occupancy(), _bootstrap_occupancy);
1170         return true;
1171       }
1172     }
1173   }
1174 
1175   // Otherwise, we start a collection cycle if
1176   // old gen want a collection cycle started. Each may use
1177   // an appropriate criterion for making this decision.
1178   // XXX We need to make sure that the gen expansion
1179   // criterion dovetails well with this. XXX NEED TO FIX THIS
1180   if (_cmsGen->should_concurrent_collect()) {
1181     log.print("CMS old gen initiated");
1182     return true;
1183   }
1184 
1185   // We start a collection if we believe an incremental collection may fail;
1186   // this is not likely to be productive in practice because it's probably too
1187   // late anyway.
1188   CMSHeap* heap = CMSHeap::heap();
1189   if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1190     log.print("CMSCollector: collect because incremental collection will fail ");
1191     return true;
1192   }
1193 
1194   if (MetaspaceGC::should_concurrent_collect()) {
1195     log.print("CMSCollector: collect for metadata allocation ");
1196     return true;
1197   }
1198 
1199   // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1200   if (CMSTriggerInterval >= 0) {
1201     if (CMSTriggerInterval == 0) {
1202       // Trigger always
1203       return true;
1204     }
1205 
1206     // Check the CMS time since begin (we do not check the stats validity
1207     // as we want to be able to trigger the first CMS cycle as well)
1208     if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1209       if (stats().valid()) {
1210         log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1211                   stats().cms_time_since_begin());
1212       } else {
1213         log.print("CMSCollector: collect because of trigger interval (first collection)");
1214       }
1215       return true;
1216     }
1217   }
1218 
1219   return false;
1220 }
1221 
1222 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1223 
1224 // Clear _expansion_cause fields of constituent generations
1225 void CMSCollector::clear_expansion_cause() {
1226   _cmsGen->clear_expansion_cause();
1227 }
1228 
1229 // We should be conservative in starting a collection cycle.  To
1230 // start too eagerly runs the risk of collecting too often in the
1231 // extreme.  To collect too rarely falls back on full collections,
1232 // which works, even if not optimum in terms of concurrent work.
1233 // As a work around for too eagerly collecting, use the flag
1234 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1235 // giving the user an easily understandable way of controlling the
1236 // collections.
1237 // We want to start a new collection cycle if any of the following
1238 // conditions hold:
1239 // . our current occupancy exceeds the configured initiating occupancy
1240 //   for this generation, or
1241 // . we recently needed to expand this space and have not, since that
1242 //   expansion, done a collection of this generation, or
1243 // . the underlying space believes that it may be a good idea to initiate
1244 //   a concurrent collection (this may be based on criteria such as the
1245 //   following: the space uses linear allocation and linear allocation is
1246 //   going to fail, or there is believed to be excessive fragmentation in
1247 //   the generation, etc... or ...
1248 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1249 //   the case of the old generation; see CR 6543076):
1250 //   we may be approaching a point at which allocation requests may fail because
1251 //   we will be out of sufficient free space given allocation rate estimates.]
1252 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1253 
1254   assert_lock_strong(freelistLock());
1255   if (occupancy() > initiating_occupancy()) {
1256     log_trace(gc)(" %s: collect because of occupancy %f / %f  ",
1257                   short_name(), occupancy(), initiating_occupancy());
1258     return true;
1259   }
1260   if (UseCMSInitiatingOccupancyOnly) {
1261     return false;
1262   }
1263   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1264     log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1265     return true;
1266   }
1267   return false;
1268 }
1269 
1270 void ConcurrentMarkSweepGeneration::collect(bool   full,
1271                                             bool   clear_all_soft_refs,
1272                                             size_t size,
1273                                             bool   tlab)
1274 {
1275   collector()->collect(full, clear_all_soft_refs, size, tlab);
1276 }
1277 
1278 void CMSCollector::collect(bool   full,
1279                            bool   clear_all_soft_refs,
1280                            size_t size,
1281                            bool   tlab)
1282 {
1283   // The following "if" branch is present for defensive reasons.
1284   // In the current uses of this interface, it can be replaced with:
1285   // assert(!GCLocker.is_active(), "Can't be called otherwise");
1286   // But I am not placing that assert here to allow future
1287   // generality in invoking this interface.
1288   if (GCLocker::is_active()) {
1289     // A consistency test for GCLocker
1290     assert(GCLocker::needs_gc(), "Should have been set already");
1291     // Skip this foreground collection, instead
1292     // expanding the heap if necessary.
1293     // Need the free list locks for the call to free() in compute_new_size()
1294     compute_new_size();
1295     return;
1296   }
1297   acquire_control_and_collect(full, clear_all_soft_refs);
1298 }
1299 
1300 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1301   CMSHeap* heap = CMSHeap::heap();
1302   unsigned int gc_count = heap->total_full_collections();
1303   if (gc_count == full_gc_count) {
1304     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1305     _full_gc_requested = true;
1306     _full_gc_cause = cause;
1307     CGC_lock->notify();   // nudge CMS thread
1308   } else {
1309     assert(gc_count > full_gc_count, "Error: causal loop");
1310   }
1311 }
1312 
1313 bool CMSCollector::is_external_interruption() {
1314   GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1315   return GCCause::is_user_requested_gc(cause) ||
1316          GCCause::is_serviceability_requested_gc(cause);
1317 }
1318 
1319 void CMSCollector::report_concurrent_mode_interruption() {
1320   if (is_external_interruption()) {
1321     log_debug(gc)("Concurrent mode interrupted");
1322   } else {
1323     log_debug(gc)("Concurrent mode failure");
1324     _gc_tracer_cm->report_concurrent_mode_failure();
1325   }
1326 }
1327 
1328 
1329 // The foreground and background collectors need to coordinate in order
1330 // to make sure that they do not mutually interfere with CMS collections.
1331 // When a background collection is active,
1332 // the foreground collector may need to take over (preempt) and
1333 // synchronously complete an ongoing collection. Depending on the
1334 // frequency of the background collections and the heap usage
1335 // of the application, this preemption can be seldom or frequent.
1336 // There are only certain
1337 // points in the background collection that the "collection-baton"
1338 // can be passed to the foreground collector.
1339 //
1340 // The foreground collector will wait for the baton before
1341 // starting any part of the collection.  The foreground collector
1342 // will only wait at one location.
1343 //
1344 // The background collector will yield the baton before starting a new
1345 // phase of the collection (e.g., before initial marking, marking from roots,
1346 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1347 // of the loop which switches the phases. The background collector does some
1348 // of the phases (initial mark, final re-mark) with the world stopped.
1349 // Because of locking involved in stopping the world,
1350 // the foreground collector should not block waiting for the background
1351 // collector when it is doing a stop-the-world phase.  The background
1352 // collector will yield the baton at an additional point just before
1353 // it enters a stop-the-world phase.  Once the world is stopped, the
1354 // background collector checks the phase of the collection.  If the
1355 // phase has not changed, it proceeds with the collection.  If the
1356 // phase has changed, it skips that phase of the collection.  See
1357 // the comments on the use of the Heap_lock in collect_in_background().
1358 //
1359 // Variable used in baton passing.
1360 //   _foregroundGCIsActive - Set to true by the foreground collector when
1361 //      it wants the baton.  The foreground clears it when it has finished
1362 //      the collection.
1363 //   _foregroundGCShouldWait - Set to true by the background collector
1364 //        when it is running.  The foreground collector waits while
1365 //      _foregroundGCShouldWait is true.
1366 //  CGC_lock - monitor used to protect access to the above variables
1367 //      and to notify the foreground and background collectors.
1368 //  _collectorState - current state of the CMS collection.
1369 //
1370 // The foreground collector
1371 //   acquires the CGC_lock
1372 //   sets _foregroundGCIsActive
1373 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1374 //     various locks acquired in preparation for the collection
1375 //     are released so as not to block the background collector
1376 //     that is in the midst of a collection
1377 //   proceeds with the collection
1378 //   clears _foregroundGCIsActive
1379 //   returns
1380 //
1381 // The background collector in a loop iterating on the phases of the
1382 //      collection
1383 //   acquires the CGC_lock
1384 //   sets _foregroundGCShouldWait
1385 //   if _foregroundGCIsActive is set
1386 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1387 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1388 //     and exits the loop.
1389 //   otherwise
1390 //     proceed with that phase of the collection
1391 //     if the phase is a stop-the-world phase,
1392 //       yield the baton once more just before enqueueing
1393 //       the stop-world CMS operation (executed by the VM thread).
1394 //   returns after all phases of the collection are done
1395 //
1396 
1397 void CMSCollector::acquire_control_and_collect(bool full,
1398         bool clear_all_soft_refs) {
1399   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1400   assert(!Thread::current()->is_ConcurrentGC_thread(),
1401          "shouldn't try to acquire control from self!");
1402 
1403   // Start the protocol for acquiring control of the
1404   // collection from the background collector (aka CMS thread).
1405   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1406          "VM thread should have CMS token");
1407   // Remember the possibly interrupted state of an ongoing
1408   // concurrent collection
1409   CollectorState first_state = _collectorState;
1410 
1411   // Signal to a possibly ongoing concurrent collection that
1412   // we want to do a foreground collection.
1413   _foregroundGCIsActive = true;
1414 
1415   // release locks and wait for a notify from the background collector
1416   // releasing the locks in only necessary for phases which
1417   // do yields to improve the granularity of the collection.
1418   assert_lock_strong(bitMapLock());
1419   // We need to lock the Free list lock for the space that we are
1420   // currently collecting.
1421   assert(haveFreelistLocks(), "Must be holding free list locks");
1422   bitMapLock()->unlock();
1423   releaseFreelistLocks();
1424   {
1425     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1426     if (_foregroundGCShouldWait) {
1427       // We are going to be waiting for action for the CMS thread;
1428       // it had better not be gone (for instance at shutdown)!
1429       assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1430              "CMS thread must be running");
1431       // Wait here until the background collector gives us the go-ahead
1432       ConcurrentMarkSweepThread::clear_CMS_flag(
1433         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1434       // Get a possibly blocked CMS thread going:
1435       //   Note that we set _foregroundGCIsActive true above,
1436       //   without protection of the CGC_lock.
1437       CGC_lock->notify();
1438       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1439              "Possible deadlock");
1440       while (_foregroundGCShouldWait) {
1441         // wait for notification
1442         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1443         // Possibility of delay/starvation here, since CMS token does
1444         // not know to give priority to VM thread? Actually, i think
1445         // there wouldn't be any delay/starvation, but the proof of
1446         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1447       }
1448       ConcurrentMarkSweepThread::set_CMS_flag(
1449         ConcurrentMarkSweepThread::CMS_vm_has_token);
1450     }
1451   }
1452   // The CMS_token is already held.  Get back the other locks.
1453   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1454          "VM thread should have CMS token");
1455   getFreelistLocks();
1456   bitMapLock()->lock_without_safepoint_check();
1457   log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1458                        p2i(Thread::current()), first_state);
1459   log_debug(gc, state)("    gets control with state %d", _collectorState);
1460 
1461   // Inform cms gen if this was due to partial collection failing.
1462   // The CMS gen may use this fact to determine its expansion policy.
1463   CMSHeap* heap = CMSHeap::heap();
1464   if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1465     assert(!_cmsGen->incremental_collection_failed(),
1466            "Should have been noticed, reacted to and cleared");
1467     _cmsGen->set_incremental_collection_failed();
1468   }
1469 
1470   if (first_state > Idling) {
1471     report_concurrent_mode_interruption();
1472   }
1473 
1474   set_did_compact(true);
1475 
1476   // If the collection is being acquired from the background
1477   // collector, there may be references on the discovered
1478   // references lists.  Abandon those references, since some
1479   // of them may have become unreachable after concurrent
1480   // discovery; the STW compacting collector will redo discovery
1481   // more precisely, without being subject to floating garbage.
1482   // Leaving otherwise unreachable references in the discovered
1483   // lists would require special handling.
1484   ref_processor()->disable_discovery();
1485   ref_processor()->abandon_partial_discovery();
1486   ref_processor()->verify_no_references_recorded();
1487 
1488   if (first_state > Idling) {
1489     save_heap_summary();
1490   }
1491 
1492   do_compaction_work(clear_all_soft_refs);
1493 
1494   // Has the GC time limit been exceeded?
1495   size_t max_eden_size = _young_gen->max_eden_size();
1496   GCCause::Cause gc_cause = heap->gc_cause();
1497   size_policy()->check_gc_overhead_limit(_young_gen->used(),
1498                                          _young_gen->eden()->used(),
1499                                          _cmsGen->max_capacity(),
1500                                          max_eden_size,
1501                                          full,
1502                                          gc_cause,
1503                                          heap->soft_ref_policy());
1504 
1505   // Reset the expansion cause, now that we just completed
1506   // a collection cycle.
1507   clear_expansion_cause();
1508   _foregroundGCIsActive = false;
1509   return;
1510 }
1511 
1512 // Resize the tenured generation
1513 // after obtaining the free list locks for the
1514 // two generations.
1515 void CMSCollector::compute_new_size() {
1516   assert_locked_or_safepoint(Heap_lock);
1517   FreelistLocker z(this);
1518   MetaspaceGC::compute_new_size();
1519   _cmsGen->compute_new_size_free_list();
1520 }
1521 
1522 // A work method used by the foreground collector to do
1523 // a mark-sweep-compact.
1524 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1525   CMSHeap* heap = CMSHeap::heap();
1526 
1527   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1528   gc_timer->register_gc_start();
1529 
1530   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1531   gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1532 
1533   heap->pre_full_gc_dump(gc_timer);
1534 
1535   GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1536 
1537   // Temporarily widen the span of the weak reference processing to
1538   // the entire heap.
1539   MemRegion new_span(CMSHeap::heap()->reserved_region());
1540   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1541   // Temporarily, clear the "is_alive_non_header" field of the
1542   // reference processor.
1543   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1544   // Temporarily make reference _processing_ single threaded (non-MT).
1545   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1546   // Temporarily make refs discovery atomic
1547   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1548   // Temporarily make reference _discovery_ single threaded (non-MT)
1549   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1550 
1551   ref_processor()->set_enqueuing_is_done(false);
1552   ref_processor()->enable_discovery();
1553   ref_processor()->setup_policy(clear_all_soft_refs);
1554   // If an asynchronous collection finishes, the _modUnionTable is
1555   // all clear.  If we are assuming the collection from an asynchronous
1556   // collection, clear the _modUnionTable.
1557   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1558     "_modUnionTable should be clear if the baton was not passed");
1559   _modUnionTable.clear_all();
1560   assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1561     "mod union for klasses should be clear if the baton was passed");
1562   _ct->cld_rem_set()->clear_mod_union();
1563 
1564 
1565   // We must adjust the allocation statistics being maintained
1566   // in the free list space. We do so by reading and clearing
1567   // the sweep timer and updating the block flux rate estimates below.
1568   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1569   if (_inter_sweep_timer.is_active()) {
1570     _inter_sweep_timer.stop();
1571     // Note that we do not use this sample to update the _inter_sweep_estimate.
1572     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1573                                             _inter_sweep_estimate.padded_average(),
1574                                             _intra_sweep_estimate.padded_average());
1575   }
1576 
1577   GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1578   #ifdef ASSERT
1579     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1580     size_t free_size = cms_space->free();
1581     assert(free_size ==
1582            pointer_delta(cms_space->end(), cms_space->compaction_top())
1583            * HeapWordSize,
1584       "All the free space should be compacted into one chunk at top");
1585     assert(cms_space->dictionary()->total_chunk_size(
1586                                       debug_only(cms_space->freelistLock())) == 0 ||
1587            cms_space->totalSizeInIndexedFreeLists() == 0,
1588       "All the free space should be in a single chunk");
1589     size_t num = cms_space->totalCount();
1590     assert((free_size == 0 && num == 0) ||
1591            (free_size > 0  && (num == 1 || num == 2)),
1592          "There should be at most 2 free chunks after compaction");
1593   #endif // ASSERT
1594   _collectorState = Resetting;
1595   assert(_restart_addr == NULL,
1596          "Should have been NULL'd before baton was passed");
1597   reset_stw();
1598   _cmsGen->reset_after_compaction();
1599   _concurrent_cycles_since_last_unload = 0;
1600 
1601   // Clear any data recorded in the PLAB chunk arrays.
1602   if (_survivor_plab_array != NULL) {
1603     reset_survivor_plab_arrays();
1604   }
1605 
1606   // Adjust the per-size allocation stats for the next epoch.
1607   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1608   // Restart the "inter sweep timer" for the next epoch.
1609   _inter_sweep_timer.reset();
1610   _inter_sweep_timer.start();
1611 
1612   // No longer a need to do a concurrent collection for Metaspace.
1613   MetaspaceGC::set_should_concurrent_collect(false);
1614 
1615   heap->post_full_gc_dump(gc_timer);
1616 
1617   gc_timer->register_gc_end();
1618 
1619   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1620 
1621   // For a mark-sweep-compact, compute_new_size() will be called
1622   // in the heap's do_collection() method.
1623 }
1624 
1625 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1626   Log(gc, heap) log;
1627   if (!log.is_trace()) {
1628     return;
1629   }
1630 
1631   ContiguousSpace* eden_space = _young_gen->eden();
1632   ContiguousSpace* from_space = _young_gen->from();
1633   ContiguousSpace* to_space   = _young_gen->to();
1634   // Eden
1635   if (_eden_chunk_array != NULL) {
1636     log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1637               p2i(eden_space->bottom()), p2i(eden_space->top()),
1638               p2i(eden_space->end()), eden_space->capacity());
1639     log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1640               _eden_chunk_index, _eden_chunk_capacity);
1641     for (size_t i = 0; i < _eden_chunk_index; i++) {
1642       log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1643     }
1644   }
1645   // Survivor
1646   if (_survivor_chunk_array != NULL) {
1647     log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1648               p2i(from_space->bottom()), p2i(from_space->top()),
1649               p2i(from_space->end()), from_space->capacity());
1650     log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1651               _survivor_chunk_index, _survivor_chunk_capacity);
1652     for (size_t i = 0; i < _survivor_chunk_index; i++) {
1653       log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1654     }
1655   }
1656 }
1657 
1658 void CMSCollector::getFreelistLocks() const {
1659   // Get locks for all free lists in all generations that this
1660   // collector is responsible for
1661   _cmsGen->freelistLock()->lock_without_safepoint_check();
1662 }
1663 
1664 void CMSCollector::releaseFreelistLocks() const {
1665   // Release locks for all free lists in all generations that this
1666   // collector is responsible for
1667   _cmsGen->freelistLock()->unlock();
1668 }
1669 
1670 bool CMSCollector::haveFreelistLocks() const {
1671   // Check locks for all free lists in all generations that this
1672   // collector is responsible for
1673   assert_lock_strong(_cmsGen->freelistLock());
1674   PRODUCT_ONLY(ShouldNotReachHere());
1675   return true;
1676 }
1677 
1678 // A utility class that is used by the CMS collector to
1679 // temporarily "release" the foreground collector from its
1680 // usual obligation to wait for the background collector to
1681 // complete an ongoing phase before proceeding.
1682 class ReleaseForegroundGC: public StackObj {
1683  private:
1684   CMSCollector* _c;
1685  public:
1686   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1687     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1688     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1689     // allow a potentially blocked foreground collector to proceed
1690     _c->_foregroundGCShouldWait = false;
1691     if (_c->_foregroundGCIsActive) {
1692       CGC_lock->notify();
1693     }
1694     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1695            "Possible deadlock");
1696   }
1697 
1698   ~ReleaseForegroundGC() {
1699     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1700     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1701     _c->_foregroundGCShouldWait = true;
1702   }
1703 };
1704 
1705 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1706   assert(Thread::current()->is_ConcurrentGC_thread(),
1707     "A CMS asynchronous collection is only allowed on a CMS thread.");
1708 
1709   CMSHeap* heap = CMSHeap::heap();
1710   {
1711     bool safepoint_check = Mutex::_no_safepoint_check_flag;
1712     MutexLockerEx hl(Heap_lock, safepoint_check);
1713     FreelistLocker fll(this);
1714     MutexLockerEx x(CGC_lock, safepoint_check);
1715     if (_foregroundGCIsActive) {
1716       // The foreground collector is. Skip this
1717       // background collection.
1718       assert(!_foregroundGCShouldWait, "Should be clear");
1719       return;
1720     } else {
1721       assert(_collectorState == Idling, "Should be idling before start.");
1722       _collectorState = InitialMarking;
1723       register_gc_start(cause);
1724       // Reset the expansion cause, now that we are about to begin
1725       // a new cycle.
1726       clear_expansion_cause();
1727 
1728       // Clear the MetaspaceGC flag since a concurrent collection
1729       // is starting but also clear it after the collection.
1730       MetaspaceGC::set_should_concurrent_collect(false);
1731     }
1732     // Decide if we want to enable class unloading as part of the
1733     // ensuing concurrent GC cycle.
1734     update_should_unload_classes();
1735     _full_gc_requested = false;           // acks all outstanding full gc requests
1736     _full_gc_cause = GCCause::_no_gc;
1737     // Signal that we are about to start a collection
1738     heap->increment_total_full_collections();  // ... starting a collection cycle
1739     _collection_count_start = heap->total_full_collections();
1740   }
1741 
1742   size_t prev_used = _cmsGen->used();
1743 
1744   // The change of the collection state is normally done at this level;
1745   // the exceptions are phases that are executed while the world is
1746   // stopped.  For those phases the change of state is done while the
1747   // world is stopped.  For baton passing purposes this allows the
1748   // background collector to finish the phase and change state atomically.
1749   // The foreground collector cannot wait on a phase that is done
1750   // while the world is stopped because the foreground collector already
1751   // has the world stopped and would deadlock.
1752   while (_collectorState != Idling) {
1753     log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1754                          p2i(Thread::current()), _collectorState);
1755     // The foreground collector
1756     //   holds the Heap_lock throughout its collection.
1757     //   holds the CMS token (but not the lock)
1758     //     except while it is waiting for the background collector to yield.
1759     //
1760     // The foreground collector should be blocked (not for long)
1761     //   if the background collector is about to start a phase
1762     //   executed with world stopped.  If the background
1763     //   collector has already started such a phase, the
1764     //   foreground collector is blocked waiting for the
1765     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
1766     //   are executed in the VM thread.
1767     //
1768     // The locking order is
1769     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
1770     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
1771     //   CMS token  (claimed in
1772     //                stop_world_and_do() -->
1773     //                  safepoint_synchronize() -->
1774     //                    CMSThread::synchronize())
1775 
1776     {
1777       // Check if the FG collector wants us to yield.
1778       CMSTokenSync x(true); // is cms thread
1779       if (waitForForegroundGC()) {
1780         // We yielded to a foreground GC, nothing more to be
1781         // done this round.
1782         assert(_foregroundGCShouldWait == false, "We set it to false in "
1783                "waitForForegroundGC()");
1784         log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1785                              p2i(Thread::current()), _collectorState);
1786         return;
1787       } else {
1788         // The background collector can run but check to see if the
1789         // foreground collector has done a collection while the
1790         // background collector was waiting to get the CGC_lock
1791         // above.  If yes, break so that _foregroundGCShouldWait
1792         // is cleared before returning.
1793         if (_collectorState == Idling) {
1794           break;
1795         }
1796       }
1797     }
1798 
1799     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1800       "should be waiting");
1801 
1802     switch (_collectorState) {
1803       case InitialMarking:
1804         {
1805           ReleaseForegroundGC x(this);
1806           stats().record_cms_begin();
1807           VM_CMS_Initial_Mark initial_mark_op(this);
1808           VMThread::execute(&initial_mark_op);
1809         }
1810         // The collector state may be any legal state at this point
1811         // since the background collector may have yielded to the
1812         // foreground collector.
1813         break;
1814       case Marking:
1815         // initial marking in checkpointRootsInitialWork has been completed
1816         if (markFromRoots()) { // we were successful
1817           assert(_collectorState == Precleaning, "Collector state should "
1818             "have changed");
1819         } else {
1820           assert(_foregroundGCIsActive, "Internal state inconsistency");
1821         }
1822         break;
1823       case Precleaning:
1824         // marking from roots in markFromRoots has been completed
1825         preclean();
1826         assert(_collectorState == AbortablePreclean ||
1827                _collectorState == FinalMarking,
1828                "Collector state should have changed");
1829         break;
1830       case AbortablePreclean:
1831         abortable_preclean();
1832         assert(_collectorState == FinalMarking, "Collector state should "
1833           "have changed");
1834         break;
1835       case FinalMarking:
1836         {
1837           ReleaseForegroundGC x(this);
1838 
1839           VM_CMS_Final_Remark final_remark_op(this);
1840           VMThread::execute(&final_remark_op);
1841         }
1842         assert(_foregroundGCShouldWait, "block post-condition");
1843         break;
1844       case Sweeping:
1845         // final marking in checkpointRootsFinal has been completed
1846         sweep();
1847         assert(_collectorState == Resizing, "Collector state change "
1848           "to Resizing must be done under the free_list_lock");
1849 
1850       case Resizing: {
1851         // Sweeping has been completed...
1852         // At this point the background collection has completed.
1853         // Don't move the call to compute_new_size() down
1854         // into code that might be executed if the background
1855         // collection was preempted.
1856         {
1857           ReleaseForegroundGC x(this);   // unblock FG collection
1858           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
1859           CMSTokenSync        z(true);   // not strictly needed.
1860           if (_collectorState == Resizing) {
1861             compute_new_size();
1862             save_heap_summary();
1863             _collectorState = Resetting;
1864           } else {
1865             assert(_collectorState == Idling, "The state should only change"
1866                    " because the foreground collector has finished the collection");
1867           }
1868         }
1869         break;
1870       }
1871       case Resetting:
1872         // CMS heap resizing has been completed
1873         reset_concurrent();
1874         assert(_collectorState == Idling, "Collector state should "
1875           "have changed");
1876 
1877         MetaspaceGC::set_should_concurrent_collect(false);
1878 
1879         stats().record_cms_end();
1880         // Don't move the concurrent_phases_end() and compute_new_size()
1881         // calls to here because a preempted background collection
1882         // has it's state set to "Resetting".
1883         break;
1884       case Idling:
1885       default:
1886         ShouldNotReachHere();
1887         break;
1888     }
1889     log_debug(gc, state)("  Thread " INTPTR_FORMAT " done - next CMS state %d",
1890                          p2i(Thread::current()), _collectorState);
1891     assert(_foregroundGCShouldWait, "block post-condition");
1892   }
1893 
1894   // Should this be in gc_epilogue?
1895   heap->counters()->update_counters();
1896 
1897   {
1898     // Clear _foregroundGCShouldWait and, in the event that the
1899     // foreground collector is waiting, notify it, before
1900     // returning.
1901     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1902     _foregroundGCShouldWait = false;
1903     if (_foregroundGCIsActive) {
1904       CGC_lock->notify();
1905     }
1906     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1907            "Possible deadlock");
1908   }
1909   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1910                        p2i(Thread::current()), _collectorState);
1911   log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K("  SIZE_FORMAT "K)",
1912                      prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1913 }
1914 
1915 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1916   _cms_start_registered = true;
1917   _gc_timer_cm->register_gc_start();
1918   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1919 }
1920 
1921 void CMSCollector::register_gc_end() {
1922   if (_cms_start_registered) {
1923     report_heap_summary(GCWhen::AfterGC);
1924 
1925     _gc_timer_cm->register_gc_end();
1926     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1927     _cms_start_registered = false;
1928   }
1929 }
1930 
1931 void CMSCollector::save_heap_summary() {
1932   CMSHeap* heap = CMSHeap::heap();
1933   _last_heap_summary = heap->create_heap_summary();
1934   _last_metaspace_summary = heap->create_metaspace_summary();
1935 }
1936 
1937 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1938   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1939   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1940 }
1941 
1942 bool CMSCollector::waitForForegroundGC() {
1943   bool res = false;
1944   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1945          "CMS thread should have CMS token");
1946   // Block the foreground collector until the
1947   // background collectors decides whether to
1948   // yield.
1949   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1950   _foregroundGCShouldWait = true;
1951   if (_foregroundGCIsActive) {
1952     // The background collector yields to the
1953     // foreground collector and returns a value
1954     // indicating that it has yielded.  The foreground
1955     // collector can proceed.
1956     res = true;
1957     _foregroundGCShouldWait = false;
1958     ConcurrentMarkSweepThread::clear_CMS_flag(
1959       ConcurrentMarkSweepThread::CMS_cms_has_token);
1960     ConcurrentMarkSweepThread::set_CMS_flag(
1961       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1962     // Get a possibly blocked foreground thread going
1963     CGC_lock->notify();
1964     log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1965                          p2i(Thread::current()), _collectorState);
1966     while (_foregroundGCIsActive) {
1967       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1968     }
1969     ConcurrentMarkSweepThread::set_CMS_flag(
1970       ConcurrentMarkSweepThread::CMS_cms_has_token);
1971     ConcurrentMarkSweepThread::clear_CMS_flag(
1972       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1973   }
1974   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1975                        p2i(Thread::current()), _collectorState);
1976   return res;
1977 }
1978 
1979 // Because of the need to lock the free lists and other structures in
1980 // the collector, common to all the generations that the collector is
1981 // collecting, we need the gc_prologues of individual CMS generations
1982 // delegate to their collector. It may have been simpler had the
1983 // current infrastructure allowed one to call a prologue on a
1984 // collector. In the absence of that we have the generation's
1985 // prologue delegate to the collector, which delegates back
1986 // some "local" work to a worker method in the individual generations
1987 // that it's responsible for collecting, while itself doing any
1988 // work common to all generations it's responsible for. A similar
1989 // comment applies to the  gc_epilogue()'s.
1990 // The role of the variable _between_prologue_and_epilogue is to
1991 // enforce the invocation protocol.
1992 void CMSCollector::gc_prologue(bool full) {
1993   // Call gc_prologue_work() for the CMSGen
1994   // we are responsible for.
1995 
1996   // The following locking discipline assumes that we are only called
1997   // when the world is stopped.
1998   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1999 
2000   // The CMSCollector prologue must call the gc_prologues for the
2001   // "generations" that it's responsible
2002   // for.
2003 
2004   assert(   Thread::current()->is_VM_thread()
2005          || (   CMSScavengeBeforeRemark
2006              && Thread::current()->is_ConcurrentGC_thread()),
2007          "Incorrect thread type for prologue execution");
2008 
2009   if (_between_prologue_and_epilogue) {
2010     // We have already been invoked; this is a gc_prologue delegation
2011     // from yet another CMS generation that we are responsible for, just
2012     // ignore it since all relevant work has already been done.
2013     return;
2014   }
2015 
2016   // set a bit saying prologue has been called; cleared in epilogue
2017   _between_prologue_and_epilogue = true;
2018   // Claim locks for common data structures, then call gc_prologue_work()
2019   // for each CMSGen.
2020 
2021   getFreelistLocks();   // gets free list locks on constituent spaces
2022   bitMapLock()->lock_without_safepoint_check();
2023 
2024   // Should call gc_prologue_work() for all cms gens we are responsible for
2025   bool duringMarking =    _collectorState >= Marking
2026                          && _collectorState < Sweeping;
2027 
2028   // The young collections clear the modified oops state, which tells if
2029   // there are any modified oops in the class. The remark phase also needs
2030   // that information. Tell the young collection to save the union of all
2031   // modified klasses.
2032   if (duringMarking) {
2033     _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2034   }
2035 
2036   bool registerClosure = duringMarking;
2037 
2038   _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2039 
2040   if (!full) {
2041     stats().record_gc0_begin();
2042   }
2043 }
2044 
2045 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2046 
2047   _capacity_at_prologue = capacity();
2048   _used_at_prologue = used();
2049 
2050   // We enable promotion tracking so that card-scanning can recognize
2051   // which objects have been promoted during this GC and skip them.
2052   for (uint i = 0; i < ParallelGCThreads; i++) {
2053     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2054   }
2055 
2056   // Delegate to CMScollector which knows how to coordinate between
2057   // this and any other CMS generations that it is responsible for
2058   // collecting.
2059   collector()->gc_prologue(full);
2060 }
2061 
2062 // This is a "private" interface for use by this generation's CMSCollector.
2063 // Not to be called directly by any other entity (for instance,
2064 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2065 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2066   bool registerClosure, ModUnionClosure* modUnionClosure) {
2067   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2068   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2069     "Should be NULL");
2070   if (registerClosure) {
2071     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2072   }
2073   cmsSpace()->gc_prologue();
2074   // Clear stat counters
2075   NOT_PRODUCT(
2076     assert(_numObjectsPromoted == 0, "check");
2077     assert(_numWordsPromoted   == 0, "check");
2078     log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2079                                  _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2080     _numObjectsAllocated = 0;
2081     _numWordsAllocated   = 0;
2082   )
2083 }
2084 
2085 void CMSCollector::gc_epilogue(bool full) {
2086   // The following locking discipline assumes that we are only called
2087   // when the world is stopped.
2088   assert(SafepointSynchronize::is_at_safepoint(),
2089          "world is stopped assumption");
2090 
2091   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2092   // if linear allocation blocks need to be appropriately marked to allow the
2093   // the blocks to be parsable. We also check here whether we need to nudge the
2094   // CMS collector thread to start a new cycle (if it's not already active).
2095   assert(   Thread::current()->is_VM_thread()
2096          || (   CMSScavengeBeforeRemark
2097              && Thread::current()->is_ConcurrentGC_thread()),
2098          "Incorrect thread type for epilogue execution");
2099 
2100   if (!_between_prologue_and_epilogue) {
2101     // We have already been invoked; this is a gc_epilogue delegation
2102     // from yet another CMS generation that we are responsible for, just
2103     // ignore it since all relevant work has already been done.
2104     return;
2105   }
2106   assert(haveFreelistLocks(), "must have freelist locks");
2107   assert_lock_strong(bitMapLock());
2108 
2109   _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2110 
2111   _cmsGen->gc_epilogue_work(full);
2112 
2113   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2114     // in case sampling was not already enabled, enable it
2115     _start_sampling = true;
2116   }
2117   // reset _eden_chunk_array so sampling starts afresh
2118   _eden_chunk_index = 0;
2119 
2120   size_t cms_used   = _cmsGen->cmsSpace()->used();
2121 
2122   // update performance counters - this uses a special version of
2123   // update_counters() that allows the utilization to be passed as a
2124   // parameter, avoiding multiple calls to used().
2125   //
2126   _cmsGen->update_counters(cms_used);
2127 
2128   bitMapLock()->unlock();
2129   releaseFreelistLocks();
2130 
2131   if (!CleanChunkPoolAsync) {
2132     Chunk::clean_chunk_pool();
2133   }
2134 
2135   set_did_compact(false);
2136   _between_prologue_and_epilogue = false;  // ready for next cycle
2137 }
2138 
2139 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2140   collector()->gc_epilogue(full);
2141 
2142   // When using ParNew, promotion tracking should have already been
2143   // disabled. However, the prologue (which enables promotion
2144   // tracking) and epilogue are called irrespective of the type of
2145   // GC. So they will also be called before and after Full GCs, during
2146   // which promotion tracking will not be explicitly disabled. So,
2147   // it's safer to also disable it here too (to be symmetric with
2148   // enabling it in the prologue).
2149   for (uint i = 0; i < ParallelGCThreads; i++) {
2150     _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2151   }
2152 }
2153 
2154 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2155   assert(!incremental_collection_failed(), "Should have been cleared");
2156   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2157   cmsSpace()->gc_epilogue();
2158     // Print stat counters
2159   NOT_PRODUCT(
2160     assert(_numObjectsAllocated == 0, "check");
2161     assert(_numWordsAllocated == 0, "check");
2162     log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2163                                      _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2164     _numObjectsPromoted = 0;
2165     _numWordsPromoted   = 0;
2166   )
2167 
2168   // Call down the chain in contiguous_available needs the freelistLock
2169   // so print this out before releasing the freeListLock.
2170   log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2171 }
2172 
2173 #ifndef PRODUCT
2174 bool CMSCollector::have_cms_token() {
2175   Thread* thr = Thread::current();
2176   if (thr->is_VM_thread()) {
2177     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2178   } else if (thr->is_ConcurrentGC_thread()) {
2179     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2180   } else if (thr->is_GC_task_thread()) {
2181     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2182            ParGCRareEvent_lock->owned_by_self();
2183   }
2184   return false;
2185 }
2186 
2187 // Check reachability of the given heap address in CMS generation,
2188 // treating all other generations as roots.
2189 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2190   // We could "guarantee" below, rather than assert, but I'll
2191   // leave these as "asserts" so that an adventurous debugger
2192   // could try this in the product build provided some subset of
2193   // the conditions were met, provided they were interested in the
2194   // results and knew that the computation below wouldn't interfere
2195   // with other concurrent computations mutating the structures
2196   // being read or written.
2197   assert(SafepointSynchronize::is_at_safepoint(),
2198          "Else mutations in object graph will make answer suspect");
2199   assert(have_cms_token(), "Should hold cms token");
2200   assert(haveFreelistLocks(), "must hold free list locks");
2201   assert_lock_strong(bitMapLock());
2202 
2203   // Clear the marking bit map array before starting, but, just
2204   // for kicks, first report if the given address is already marked
2205   tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2206                 _markBitMap.isMarked(addr) ? "" : " not");
2207 
2208   if (verify_after_remark()) {
2209     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2210     bool result = verification_mark_bm()->isMarked(addr);
2211     tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2212                   result ? "IS" : "is NOT");
2213     return result;
2214   } else {
2215     tty->print_cr("Could not compute result");
2216     return false;
2217   }
2218 }
2219 #endif
2220 
2221 void
2222 CMSCollector::print_on_error(outputStream* st) {
2223   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2224   if (collector != NULL) {
2225     CMSBitMap* bitmap = &collector->_markBitMap;
2226     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2227     bitmap->print_on_error(st, " Bits: ");
2228 
2229     st->cr();
2230 
2231     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2232     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2233     mut_bitmap->print_on_error(st, " Bits: ");
2234   }
2235 }
2236 
2237 ////////////////////////////////////////////////////////
2238 // CMS Verification Support
2239 ////////////////////////////////////////////////////////
2240 // Following the remark phase, the following invariant
2241 // should hold -- each object in the CMS heap which is
2242 // marked in markBitMap() should be marked in the verification_mark_bm().
2243 
2244 class VerifyMarkedClosure: public BitMapClosure {
2245   CMSBitMap* _marks;
2246   bool       _failed;
2247 
2248  public:
2249   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2250 
2251   bool do_bit(size_t offset) {
2252     HeapWord* addr = _marks->offsetToHeapWord(offset);
2253     if (!_marks->isMarked(addr)) {
2254       Log(gc, verify) log;
2255       ResourceMark rm;
2256       LogStream ls(log.error());
2257       oop(addr)->print_on(&ls);
2258       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2259       _failed = true;
2260     }
2261     return true;
2262   }
2263 
2264   bool failed() { return _failed; }
2265 };
2266 
2267 bool CMSCollector::verify_after_remark() {
2268   GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2269   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2270   static bool init = false;
2271 
2272   assert(SafepointSynchronize::is_at_safepoint(),
2273          "Else mutations in object graph will make answer suspect");
2274   assert(have_cms_token(),
2275          "Else there may be mutual interference in use of "
2276          " verification data structures");
2277   assert(_collectorState > Marking && _collectorState <= Sweeping,
2278          "Else marking info checked here may be obsolete");
2279   assert(haveFreelistLocks(), "must hold free list locks");
2280   assert_lock_strong(bitMapLock());
2281 
2282 
2283   // Allocate marking bit map if not already allocated
2284   if (!init) { // first time
2285     if (!verification_mark_bm()->allocate(_span)) {
2286       return false;
2287     }
2288     init = true;
2289   }
2290 
2291   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2292 
2293   // Turn off refs discovery -- so we will be tracing through refs.
2294   // This is as intended, because by this time
2295   // GC must already have cleared any refs that need to be cleared,
2296   // and traced those that need to be marked; moreover,
2297   // the marking done here is not going to interfere in any
2298   // way with the marking information used by GC.
2299   NoRefDiscovery no_discovery(ref_processor());
2300 
2301 #if COMPILER2_OR_JVMCI
2302   DerivedPointerTableDeactivate dpt_deact;
2303 #endif
2304 
2305   // Clear any marks from a previous round
2306   verification_mark_bm()->clear_all();
2307   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2308   verify_work_stacks_empty();
2309 
2310   CMSHeap* heap = CMSHeap::heap();
2311   heap->ensure_parsability(false);  // fill TLABs, but no need to retire them
2312   // Update the saved marks which may affect the root scans.
2313   heap->save_marks();
2314 
2315   if (CMSRemarkVerifyVariant == 1) {
2316     // In this first variant of verification, we complete
2317     // all marking, then check if the new marks-vector is
2318     // a subset of the CMS marks-vector.
2319     verify_after_remark_work_1();
2320   } else {
2321     guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2322     // In this second variant of verification, we flag an error
2323     // (i.e. an object reachable in the new marks-vector not reachable
2324     // in the CMS marks-vector) immediately, also indicating the
2325     // identify of an object (A) that references the unmarked object (B) --
2326     // presumably, a mutation to A failed to be picked up by preclean/remark?
2327     verify_after_remark_work_2();
2328   }
2329 
2330   return true;
2331 }
2332 
2333 void CMSCollector::verify_after_remark_work_1() {
2334   ResourceMark rm;
2335   HandleMark  hm;
2336   CMSHeap* heap = CMSHeap::heap();
2337 
2338   // Get a clear set of claim bits for the roots processing to work with.
2339   ClassLoaderDataGraph::clear_claimed_marks();
2340 
2341   // Mark from roots one level into CMS
2342   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2343   heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2344 
2345   {
2346     StrongRootsScope srs(1);
2347 
2348     heap->cms_process_roots(&srs,
2349                            true,   // young gen as roots
2350                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2351                            should_unload_classes(),
2352                            &notOlder,
2353                            NULL);
2354   }
2355 
2356   // Now mark from the roots
2357   MarkFromRootsClosure markFromRootsClosure(this, _span,
2358     verification_mark_bm(), verification_mark_stack(),
2359     false /* don't yield */, true /* verifying */);
2360   assert(_restart_addr == NULL, "Expected pre-condition");
2361   verification_mark_bm()->iterate(&markFromRootsClosure);
2362   while (_restart_addr != NULL) {
2363     // Deal with stack overflow: by restarting at the indicated
2364     // address.
2365     HeapWord* ra = _restart_addr;
2366     markFromRootsClosure.reset(ra);
2367     _restart_addr = NULL;
2368     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2369   }
2370   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2371   verify_work_stacks_empty();
2372 
2373   // Marking completed -- now verify that each bit marked in
2374   // verification_mark_bm() is also marked in markBitMap(); flag all
2375   // errors by printing corresponding objects.
2376   VerifyMarkedClosure vcl(markBitMap());
2377   verification_mark_bm()->iterate(&vcl);
2378   if (vcl.failed()) {
2379     Log(gc, verify) log;
2380     log.error("Failed marking verification after remark");
2381     ResourceMark rm;
2382     LogStream ls(log.error());
2383     heap->print_on(&ls);
2384     fatal("CMS: failed marking verification after remark");
2385   }
2386 }
2387 
2388 class VerifyCLDOopsCLDClosure : public CLDClosure {
2389   class VerifyCLDOopsClosure : public OopClosure {
2390     CMSBitMap* _bitmap;
2391    public:
2392     VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2393     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2394     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2395   } _oop_closure;
2396  public:
2397   VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2398   void do_cld(ClassLoaderData* cld) {
2399     cld->oops_do(&_oop_closure, false, false);
2400   }
2401 };
2402 
2403 void CMSCollector::verify_after_remark_work_2() {
2404   ResourceMark rm;
2405   HandleMark  hm;
2406   CMSHeap* heap = CMSHeap::heap();
2407 
2408   // Get a clear set of claim bits for the roots processing to work with.
2409   ClassLoaderDataGraph::clear_claimed_marks();
2410 
2411   // Mark from roots one level into CMS
2412   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2413                                      markBitMap());
2414   CLDToOopClosure cld_closure(&notOlder, true);
2415 
2416   heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2417 
2418   {
2419     StrongRootsScope srs(1);
2420 
2421     heap->cms_process_roots(&srs,
2422                            true,   // young gen as roots
2423                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2424                            should_unload_classes(),
2425                            &notOlder,
2426                            &cld_closure);
2427   }
2428 
2429   // Now mark from the roots
2430   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2431     verification_mark_bm(), markBitMap(), verification_mark_stack());
2432   assert(_restart_addr == NULL, "Expected pre-condition");
2433   verification_mark_bm()->iterate(&markFromRootsClosure);
2434   while (_restart_addr != NULL) {
2435     // Deal with stack overflow: by restarting at the indicated
2436     // address.
2437     HeapWord* ra = _restart_addr;
2438     markFromRootsClosure.reset(ra);
2439     _restart_addr = NULL;
2440     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2441   }
2442   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2443   verify_work_stacks_empty();
2444 
2445   VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2446   ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2447 
2448   // Marking completed -- now verify that each bit marked in
2449   // verification_mark_bm() is also marked in markBitMap(); flag all
2450   // errors by printing corresponding objects.
2451   VerifyMarkedClosure vcl(markBitMap());
2452   verification_mark_bm()->iterate(&vcl);
2453   assert(!vcl.failed(), "Else verification above should not have succeeded");
2454 }
2455 
2456 void ConcurrentMarkSweepGeneration::save_marks() {
2457   // delegate to CMS space
2458   cmsSpace()->save_marks();
2459 }
2460 
2461 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2462   return cmsSpace()->no_allocs_since_save_marks();
2463 }
2464 
2465 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2466                                                                 \
2467 void ConcurrentMarkSweepGeneration::                            \
2468 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2469   cl->set_generation(this);                                     \
2470   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2471   cl->reset_generation();                                       \
2472   save_marks();                                                 \
2473 }
2474 
2475 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2476 
2477 void
2478 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2479   if (freelistLock()->owned_by_self()) {
2480     Generation::oop_iterate(cl);
2481   } else {
2482     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2483     Generation::oop_iterate(cl);
2484   }
2485 }
2486 
2487 void
2488 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2489   if (freelistLock()->owned_by_self()) {
2490     Generation::object_iterate(cl);
2491   } else {
2492     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2493     Generation::object_iterate(cl);
2494   }
2495 }
2496 
2497 void
2498 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2499   if (freelistLock()->owned_by_self()) {
2500     Generation::safe_object_iterate(cl);
2501   } else {
2502     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2503     Generation::safe_object_iterate(cl);
2504   }
2505 }
2506 
2507 void
2508 ConcurrentMarkSweepGeneration::post_compact() {
2509 }
2510 
2511 void
2512 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2513   // Fix the linear allocation blocks to look like free blocks.
2514 
2515   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2516   // are not called when the heap is verified during universe initialization and
2517   // at vm shutdown.
2518   if (freelistLock()->owned_by_self()) {
2519     cmsSpace()->prepare_for_verify();
2520   } else {
2521     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2522     cmsSpace()->prepare_for_verify();
2523   }
2524 }
2525 
2526 void
2527 ConcurrentMarkSweepGeneration::verify() {
2528   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2529   // are not called when the heap is verified during universe initialization and
2530   // at vm shutdown.
2531   if (freelistLock()->owned_by_self()) {
2532     cmsSpace()->verify();
2533   } else {
2534     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2535     cmsSpace()->verify();
2536   }
2537 }
2538 
2539 void CMSCollector::verify() {
2540   _cmsGen->verify();
2541 }
2542 
2543 #ifndef PRODUCT
2544 bool CMSCollector::overflow_list_is_empty() const {
2545   assert(_num_par_pushes >= 0, "Inconsistency");
2546   if (_overflow_list == NULL) {
2547     assert(_num_par_pushes == 0, "Inconsistency");
2548   }
2549   return _overflow_list == NULL;
2550 }
2551 
2552 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2553 // merely consolidate assertion checks that appear to occur together frequently.
2554 void CMSCollector::verify_work_stacks_empty() const {
2555   assert(_markStack.isEmpty(), "Marking stack should be empty");
2556   assert(overflow_list_is_empty(), "Overflow list should be empty");
2557 }
2558 
2559 void CMSCollector::verify_overflow_empty() const {
2560   assert(overflow_list_is_empty(), "Overflow list should be empty");
2561   assert(no_preserved_marks(), "No preserved marks");
2562 }
2563 #endif // PRODUCT
2564 
2565 // Decide if we want to enable class unloading as part of the
2566 // ensuing concurrent GC cycle. We will collect and
2567 // unload classes if it's the case that:
2568 //  (a) class unloading is enabled at the command line, and
2569 //  (b) old gen is getting really full
2570 // NOTE: Provided there is no change in the state of the heap between
2571 // calls to this method, it should have idempotent results. Moreover,
2572 // its results should be monotonically increasing (i.e. going from 0 to 1,
2573 // but not 1 to 0) between successive calls between which the heap was
2574 // not collected. For the implementation below, it must thus rely on
2575 // the property that concurrent_cycles_since_last_unload()
2576 // will not decrease unless a collection cycle happened and that
2577 // _cmsGen->is_too_full() are
2578 // themselves also monotonic in that sense. See check_monotonicity()
2579 // below.
2580 void CMSCollector::update_should_unload_classes() {
2581   _should_unload_classes = false;
2582   if (CMSClassUnloadingEnabled) {
2583     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2584                               CMSClassUnloadingMaxInterval)
2585                            || _cmsGen->is_too_full();
2586   }
2587 }
2588 
2589 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2590   bool res = should_concurrent_collect();
2591   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2592   return res;
2593 }
2594 
2595 void CMSCollector::setup_cms_unloading_and_verification_state() {
2596   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2597                              || VerifyBeforeExit;
2598   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2599 
2600   // We set the proper root for this CMS cycle here.
2601   if (should_unload_classes()) {   // Should unload classes this cycle
2602     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2603     set_verifying(should_verify);    // Set verification state for this cycle
2604     return;                            // Nothing else needs to be done at this time
2605   }
2606 
2607   // Not unloading classes this cycle
2608   assert(!should_unload_classes(), "Inconsistency!");
2609 
2610   // If we are not unloading classes then add SO_AllCodeCache to root
2611   // scanning options.
2612   add_root_scanning_option(rso);
2613 
2614   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2615     set_verifying(true);
2616   } else if (verifying() && !should_verify) {
2617     // We were verifying, but some verification flags got disabled.
2618     set_verifying(false);
2619     // Exclude symbols, strings and code cache elements from root scanning to
2620     // reduce IM and RM pauses.
2621     remove_root_scanning_option(rso);
2622   }
2623 }
2624 
2625 
2626 #ifndef PRODUCT
2627 HeapWord* CMSCollector::block_start(const void* p) const {
2628   const HeapWord* addr = (HeapWord*)p;
2629   if (_span.contains(p)) {
2630     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2631       return _cmsGen->cmsSpace()->block_start(p);
2632     }
2633   }
2634   return NULL;
2635 }
2636 #endif
2637 
2638 HeapWord*
2639 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2640                                                    bool   tlab,
2641                                                    bool   parallel) {
2642   CMSSynchronousYieldRequest yr;
2643   assert(!tlab, "Can't deal with TLAB allocation");
2644   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2645   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2646   if (GCExpandToAllocateDelayMillis > 0) {
2647     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2648   }
2649   return have_lock_and_allocate(word_size, tlab);
2650 }
2651 
2652 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2653     size_t bytes,
2654     size_t expand_bytes,
2655     CMSExpansionCause::Cause cause)
2656 {
2657 
2658   bool success = expand(bytes, expand_bytes);
2659 
2660   // remember why we expanded; this information is used
2661   // by shouldConcurrentCollect() when making decisions on whether to start
2662   // a new CMS cycle.
2663   if (success) {
2664     set_expansion_cause(cause);
2665     log_trace(gc)("Expanded CMS gen for %s",  CMSExpansionCause::to_string(cause));
2666   }
2667 }
2668 
2669 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2670   HeapWord* res = NULL;
2671   MutexLocker x(ParGCRareEvent_lock);
2672   while (true) {
2673     // Expansion by some other thread might make alloc OK now:
2674     res = ps->lab.alloc(word_sz);
2675     if (res != NULL) return res;
2676     // If there's not enough expansion space available, give up.
2677     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2678       return NULL;
2679     }
2680     // Otherwise, we try expansion.
2681     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2682     // Now go around the loop and try alloc again;
2683     // A competing par_promote might beat us to the expansion space,
2684     // so we may go around the loop again if promotion fails again.
2685     if (GCExpandToAllocateDelayMillis > 0) {
2686       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2687     }
2688   }
2689 }
2690 
2691 
2692 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2693   PromotionInfo* promo) {
2694   MutexLocker x(ParGCRareEvent_lock);
2695   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2696   while (true) {
2697     // Expansion by some other thread might make alloc OK now:
2698     if (promo->ensure_spooling_space()) {
2699       assert(promo->has_spooling_space(),
2700              "Post-condition of successful ensure_spooling_space()");
2701       return true;
2702     }
2703     // If there's not enough expansion space available, give up.
2704     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2705       return false;
2706     }
2707     // Otherwise, we try expansion.
2708     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2709     // Now go around the loop and try alloc again;
2710     // A competing allocation might beat us to the expansion space,
2711     // so we may go around the loop again if allocation fails again.
2712     if (GCExpandToAllocateDelayMillis > 0) {
2713       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2714     }
2715   }
2716 }
2717 
2718 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2719   // Only shrink if a compaction was done so that all the free space
2720   // in the generation is in a contiguous block at the end.
2721   if (did_compact()) {
2722     CardGeneration::shrink(bytes);
2723   }
2724 }
2725 
2726 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2727   assert_locked_or_safepoint(Heap_lock);
2728 }
2729 
2730 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2731   assert_locked_or_safepoint(Heap_lock);
2732   assert_lock_strong(freelistLock());
2733   log_trace(gc)("Shrinking of CMS not yet implemented");
2734   return;
2735 }
2736 
2737 
2738 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2739 // phases.
2740 class CMSPhaseAccounting: public StackObj {
2741  public:
2742   CMSPhaseAccounting(CMSCollector *collector,
2743                      const char *title);
2744   ~CMSPhaseAccounting();
2745 
2746  private:
2747   CMSCollector *_collector;
2748   const char *_title;
2749   GCTraceConcTime(Info, gc) _trace_time;
2750 
2751  public:
2752   // Not MT-safe; so do not pass around these StackObj's
2753   // where they may be accessed by other threads.
2754   double wallclock_millis() {
2755     return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2756   }
2757 };
2758 
2759 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2760                                        const char *title) :
2761   _collector(collector), _title(title), _trace_time(title) {
2762 
2763   _collector->resetYields();
2764   _collector->resetTimer();
2765   _collector->startTimer();
2766   _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2767 }
2768 
2769 CMSPhaseAccounting::~CMSPhaseAccounting() {
2770   _collector->gc_timer_cm()->register_gc_concurrent_end();
2771   _collector->stopTimer();
2772   log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2773   log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2774 }
2775 
2776 // CMS work
2777 
2778 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2779 class CMSParMarkTask : public AbstractGangTask {
2780  protected:
2781   CMSCollector*     _collector;
2782   uint              _n_workers;
2783   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2784       AbstractGangTask(name),
2785       _collector(collector),
2786       _n_workers(n_workers) {}
2787   // Work method in support of parallel rescan ... of young gen spaces
2788   void do_young_space_rescan(OopsInGenClosure* cl,
2789                              ContiguousSpace* space,
2790                              HeapWord** chunk_array, size_t chunk_top);
2791   void work_on_young_gen_roots(OopsInGenClosure* cl);
2792 };
2793 
2794 // Parallel initial mark task
2795 class CMSParInitialMarkTask: public CMSParMarkTask {
2796   StrongRootsScope* _strong_roots_scope;
2797  public:
2798   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2799       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2800       _strong_roots_scope(strong_roots_scope) {}
2801   void work(uint worker_id);
2802 };
2803 
2804 // Checkpoint the roots into this generation from outside
2805 // this generation. [Note this initial checkpoint need only
2806 // be approximate -- we'll do a catch up phase subsequently.]
2807 void CMSCollector::checkpointRootsInitial() {
2808   assert(_collectorState == InitialMarking, "Wrong collector state");
2809   check_correct_thread_executing();
2810   TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2811 
2812   save_heap_summary();
2813   report_heap_summary(GCWhen::BeforeGC);
2814 
2815   ReferenceProcessor* rp = ref_processor();
2816   assert(_restart_addr == NULL, "Control point invariant");
2817   {
2818     // acquire locks for subsequent manipulations
2819     MutexLockerEx x(bitMapLock(),
2820                     Mutex::_no_safepoint_check_flag);
2821     checkpointRootsInitialWork();
2822     // enable ("weak") refs discovery
2823     rp->enable_discovery();
2824     _collectorState = Marking;
2825   }
2826 }
2827 
2828 void CMSCollector::checkpointRootsInitialWork() {
2829   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2830   assert(_collectorState == InitialMarking, "just checking");
2831 
2832   // Already have locks.
2833   assert_lock_strong(bitMapLock());
2834   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2835 
2836   // Setup the verification and class unloading state for this
2837   // CMS collection cycle.
2838   setup_cms_unloading_and_verification_state();
2839 
2840   GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2841 
2842   // Reset all the PLAB chunk arrays if necessary.
2843   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2844     reset_survivor_plab_arrays();
2845   }
2846 
2847   ResourceMark rm;
2848   HandleMark  hm;
2849 
2850   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2851   CMSHeap* heap = CMSHeap::heap();
2852 
2853   verify_work_stacks_empty();
2854   verify_overflow_empty();
2855 
2856   heap->ensure_parsability(false);  // fill TLABs, but no need to retire them
2857   // Update the saved marks which may affect the root scans.
2858   heap->save_marks();
2859 
2860   // weak reference processing has not started yet.
2861   ref_processor()->set_enqueuing_is_done(false);
2862 
2863   // Need to remember all newly created CLDs,
2864   // so that we can guarantee that the remark finds them.
2865   ClassLoaderDataGraph::remember_new_clds(true);
2866 
2867   // Whenever a CLD is found, it will be claimed before proceeding to mark
2868   // the klasses. The claimed marks need to be cleared before marking starts.
2869   ClassLoaderDataGraph::clear_claimed_marks();
2870 
2871   print_eden_and_survivor_chunk_arrays();
2872 
2873   {
2874 #if COMPILER2_OR_JVMCI
2875     DerivedPointerTableDeactivate dpt_deact;
2876 #endif
2877     if (CMSParallelInitialMarkEnabled) {
2878       // The parallel version.
2879       WorkGang* workers = heap->workers();
2880       assert(workers != NULL, "Need parallel worker threads.");
2881       uint n_workers = workers->active_workers();
2882 
2883       StrongRootsScope srs(n_workers);
2884 
2885       CMSParInitialMarkTask tsk(this, &srs, n_workers);
2886       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2887       // If the total workers is greater than 1, then multiple workers
2888       // may be used at some time and the initialization has been set
2889       // such that the single threaded path cannot be used.
2890       if (workers->total_workers() > 1) {
2891         workers->run_task(&tsk);
2892       } else {
2893         tsk.work(0);
2894       }
2895     } else {
2896       // The serial version.
2897       CLDToOopClosure cld_closure(&notOlder, true);
2898       heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2899 
2900       StrongRootsScope srs(1);
2901 
2902       heap->cms_process_roots(&srs,
2903                              true,   // young gen as roots
2904                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
2905                              should_unload_classes(),
2906                              &notOlder,
2907                              &cld_closure);
2908     }
2909   }
2910 
2911   // Clear mod-union table; it will be dirtied in the prologue of
2912   // CMS generation per each young generation collection.
2913 
2914   assert(_modUnionTable.isAllClear(),
2915        "Was cleared in most recent final checkpoint phase"
2916        " or no bits are set in the gc_prologue before the start of the next "
2917        "subsequent marking phase.");
2918 
2919   assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2920 
2921   // Save the end of the used_region of the constituent generations
2922   // to be used to limit the extent of sweep in each generation.
2923   save_sweep_limits();
2924   verify_overflow_empty();
2925 }
2926 
2927 bool CMSCollector::markFromRoots() {
2928   // we might be tempted to assert that:
2929   // assert(!SafepointSynchronize::is_at_safepoint(),
2930   //        "inconsistent argument?");
2931   // However that wouldn't be right, because it's possible that
2932   // a safepoint is indeed in progress as a young generation
2933   // stop-the-world GC happens even as we mark in this generation.
2934   assert(_collectorState == Marking, "inconsistent state?");
2935   check_correct_thread_executing();
2936   verify_overflow_empty();
2937 
2938   // Weak ref discovery note: We may be discovering weak
2939   // refs in this generation concurrent (but interleaved) with
2940   // weak ref discovery by the young generation collector.
2941 
2942   CMSTokenSyncWithLocks ts(true, bitMapLock());
2943   GCTraceCPUTime tcpu;
2944   CMSPhaseAccounting pa(this, "Concurrent Mark");
2945   bool res = markFromRootsWork();
2946   if (res) {
2947     _collectorState = Precleaning;
2948   } else { // We failed and a foreground collection wants to take over
2949     assert(_foregroundGCIsActive, "internal state inconsistency");
2950     assert(_restart_addr == NULL,  "foreground will restart from scratch");
2951     log_debug(gc)("bailing out to foreground collection");
2952   }
2953   verify_overflow_empty();
2954   return res;
2955 }
2956 
2957 bool CMSCollector::markFromRootsWork() {
2958   // iterate over marked bits in bit map, doing a full scan and mark
2959   // from these roots using the following algorithm:
2960   // . if oop is to the right of the current scan pointer,
2961   //   mark corresponding bit (we'll process it later)
2962   // . else (oop is to left of current scan pointer)
2963   //   push oop on marking stack
2964   // . drain the marking stack
2965 
2966   // Note that when we do a marking step we need to hold the
2967   // bit map lock -- recall that direct allocation (by mutators)
2968   // and promotion (by the young generation collector) is also
2969   // marking the bit map. [the so-called allocate live policy.]
2970   // Because the implementation of bit map marking is not
2971   // robust wrt simultaneous marking of bits in the same word,
2972   // we need to make sure that there is no such interference
2973   // between concurrent such updates.
2974 
2975   // already have locks
2976   assert_lock_strong(bitMapLock());
2977 
2978   verify_work_stacks_empty();
2979   verify_overflow_empty();
2980   bool result = false;
2981   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2982     result = do_marking_mt();
2983   } else {
2984     result = do_marking_st();
2985   }
2986   return result;
2987 }
2988 
2989 // Forward decl
2990 class CMSConcMarkingTask;
2991 
2992 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2993   CMSCollector*       _collector;
2994   CMSConcMarkingTask* _task;
2995  public:
2996   virtual void yield();
2997 
2998   // "n_threads" is the number of threads to be terminated.
2999   // "queue_set" is a set of work queues of other threads.
3000   // "collector" is the CMS collector associated with this task terminator.
3001   // "yield" indicates whether we need the gang as a whole to yield.
3002   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3003     ParallelTaskTerminator(n_threads, queue_set),
3004     _collector(collector) { }
3005 
3006   void set_task(CMSConcMarkingTask* task) {
3007     _task = task;
3008   }
3009 };
3010 
3011 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3012   CMSConcMarkingTask* _task;
3013  public:
3014   bool should_exit_termination();
3015   void set_task(CMSConcMarkingTask* task) {
3016     _task = task;
3017   }
3018 };
3019 
3020 // MT Concurrent Marking Task
3021 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3022   CMSCollector*             _collector;
3023   uint                      _n_workers;      // requested/desired # workers
3024   bool                      _result;
3025   CompactibleFreeListSpace* _cms_space;
3026   char                      _pad_front[64];   // padding to ...
3027   HeapWord* volatile        _global_finger;   // ... avoid sharing cache line
3028   char                      _pad_back[64];
3029   HeapWord*                 _restart_addr;
3030 
3031   //  Exposed here for yielding support
3032   Mutex* const _bit_map_lock;
3033 
3034   // The per thread work queues, available here for stealing
3035   OopTaskQueueSet*  _task_queues;
3036 
3037   // Termination (and yielding) support
3038   CMSConcMarkingTerminator _term;
3039   CMSConcMarkingTerminatorTerminator _term_term;
3040 
3041  public:
3042   CMSConcMarkingTask(CMSCollector* collector,
3043                  CompactibleFreeListSpace* cms_space,
3044                  YieldingFlexibleWorkGang* workers,
3045                  OopTaskQueueSet* task_queues):
3046     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3047     _collector(collector),
3048     _cms_space(cms_space),
3049     _n_workers(0), _result(true),
3050     _task_queues(task_queues),
3051     _term(_n_workers, task_queues, _collector),
3052     _bit_map_lock(collector->bitMapLock())
3053   {
3054     _requested_size = _n_workers;
3055     _term.set_task(this);
3056     _term_term.set_task(this);
3057     _restart_addr = _global_finger = _cms_space->bottom();
3058   }
3059 
3060 
3061   OopTaskQueueSet* task_queues()  { return _task_queues; }
3062 
3063   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3064 
3065   HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3066 
3067   CMSConcMarkingTerminator* terminator() { return &_term; }
3068 
3069   virtual void set_for_termination(uint active_workers) {
3070     terminator()->reset_for_reuse(active_workers);
3071   }
3072 
3073   void work(uint worker_id);
3074   bool should_yield() {
3075     return    ConcurrentMarkSweepThread::should_yield()
3076            && !_collector->foregroundGCIsActive();
3077   }
3078 
3079   virtual void coordinator_yield();  // stuff done by coordinator
3080   bool result() { return _result; }
3081 
3082   void reset(HeapWord* ra) {
3083     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3084     _restart_addr = _global_finger = ra;
3085     _term.reset_for_reuse();
3086   }
3087 
3088   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3089                                            OopTaskQueue* work_q);
3090 
3091  private:
3092   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3093   void do_work_steal(int i);
3094   void bump_global_finger(HeapWord* f);
3095 };
3096 
3097 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3098   assert(_task != NULL, "Error");
3099   return _task->yielding();
3100   // Note that we do not need the disjunct || _task->should_yield() above
3101   // because we want terminating threads to yield only if the task
3102   // is already in the midst of yielding, which happens only after at least one
3103   // thread has yielded.
3104 }
3105 
3106 void CMSConcMarkingTerminator::yield() {
3107   if (_task->should_yield()) {
3108     _task->yield();
3109   } else {
3110     ParallelTaskTerminator::yield();
3111   }
3112 }
3113 
3114 ////////////////////////////////////////////////////////////////
3115 // Concurrent Marking Algorithm Sketch
3116 ////////////////////////////////////////////////////////////////
3117 // Until all tasks exhausted (both spaces):
3118 // -- claim next available chunk
3119 // -- bump global finger via CAS
3120 // -- find first object that starts in this chunk
3121 //    and start scanning bitmap from that position
3122 // -- scan marked objects for oops
3123 // -- CAS-mark target, and if successful:
3124 //    . if target oop is above global finger (volatile read)
3125 //      nothing to do
3126 //    . if target oop is in chunk and above local finger
3127 //        then nothing to do
3128 //    . else push on work-queue
3129 // -- Deal with possible overflow issues:
3130 //    . local work-queue overflow causes stuff to be pushed on
3131 //      global (common) overflow queue
3132 //    . always first empty local work queue
3133 //    . then get a batch of oops from global work queue if any
3134 //    . then do work stealing
3135 // -- When all tasks claimed (both spaces)
3136 //    and local work queue empty,
3137 //    then in a loop do:
3138 //    . check global overflow stack; steal a batch of oops and trace
3139 //    . try to steal from other threads oif GOS is empty
3140 //    . if neither is available, offer termination
3141 // -- Terminate and return result
3142 //
3143 void CMSConcMarkingTask::work(uint worker_id) {
3144   elapsedTimer _timer;
3145   ResourceMark rm;
3146   HandleMark hm;
3147 
3148   DEBUG_ONLY(_collector->verify_overflow_empty();)
3149 
3150   // Before we begin work, our work queue should be empty
3151   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3152   // Scan the bitmap covering _cms_space, tracing through grey objects.
3153   _timer.start();
3154   do_scan_and_mark(worker_id, _cms_space);
3155   _timer.stop();
3156   log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3157 
3158   // ... do work stealing
3159   _timer.reset();
3160   _timer.start();
3161   do_work_steal(worker_id);
3162   _timer.stop();
3163   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3164   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3165   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3166   // Note that under the current task protocol, the
3167   // following assertion is true even of the spaces
3168   // expanded since the completion of the concurrent
3169   // marking. XXX This will likely change under a strict
3170   // ABORT semantics.
3171   // After perm removal the comparison was changed to
3172   // greater than or equal to from strictly greater than.
3173   // Before perm removal the highest address sweep would
3174   // have been at the end of perm gen but now is at the
3175   // end of the tenured gen.
3176   assert(_global_finger >=  _cms_space->end(),
3177          "All tasks have been completed");
3178   DEBUG_ONLY(_collector->verify_overflow_empty();)
3179 }
3180 
3181 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3182   HeapWord* read = _global_finger;
3183   HeapWord* cur  = read;
3184   while (f > read) {
3185     cur = read;
3186     read = Atomic::cmpxchg(f, &_global_finger, cur);
3187     if (cur == read) {
3188       // our cas succeeded
3189       assert(_global_finger >= f, "protocol consistency");
3190       break;
3191     }
3192   }
3193 }
3194 
3195 // This is really inefficient, and should be redone by
3196 // using (not yet available) block-read and -write interfaces to the
3197 // stack and the work_queue. XXX FIX ME !!!
3198 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3199                                                       OopTaskQueue* work_q) {
3200   // Fast lock-free check
3201   if (ovflw_stk->length() == 0) {
3202     return false;
3203   }
3204   assert(work_q->size() == 0, "Shouldn't steal");
3205   MutexLockerEx ml(ovflw_stk->par_lock(),
3206                    Mutex::_no_safepoint_check_flag);
3207   // Grab up to 1/4 the size of the work queue
3208   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3209                     (size_t)ParGCDesiredObjsFromOverflowList);
3210   num = MIN2(num, ovflw_stk->length());
3211   for (int i = (int) num; i > 0; i--) {
3212     oop cur = ovflw_stk->pop();
3213     assert(cur != NULL, "Counted wrong?");
3214     work_q->push(cur);
3215   }
3216   return num > 0;
3217 }
3218 
3219 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3220   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3221   int n_tasks = pst->n_tasks();
3222   // We allow that there may be no tasks to do here because
3223   // we are restarting after a stack overflow.
3224   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3225   uint nth_task = 0;
3226 
3227   HeapWord* aligned_start = sp->bottom();
3228   if (sp->used_region().contains(_restart_addr)) {
3229     // Align down to a card boundary for the start of 0th task
3230     // for this space.
3231     aligned_start = align_down(_restart_addr, CardTable::card_size);
3232   }
3233 
3234   size_t chunk_size = sp->marking_task_size();
3235   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3236     // Having claimed the nth task in this space,
3237     // compute the chunk that it corresponds to:
3238     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3239                                aligned_start + (nth_task+1)*chunk_size);
3240     // Try and bump the global finger via a CAS;
3241     // note that we need to do the global finger bump
3242     // _before_ taking the intersection below, because
3243     // the task corresponding to that region will be
3244     // deemed done even if the used_region() expands
3245     // because of allocation -- as it almost certainly will
3246     // during start-up while the threads yield in the
3247     // closure below.
3248     HeapWord* finger = span.end();
3249     bump_global_finger(finger);   // atomically
3250     // There are null tasks here corresponding to chunks
3251     // beyond the "top" address of the space.
3252     span = span.intersection(sp->used_region());
3253     if (!span.is_empty()) {  // Non-null task
3254       HeapWord* prev_obj;
3255       assert(!span.contains(_restart_addr) || nth_task == 0,
3256              "Inconsistency");
3257       if (nth_task == 0) {
3258         // For the 0th task, we'll not need to compute a block_start.
3259         if (span.contains(_restart_addr)) {
3260           // In the case of a restart because of stack overflow,
3261           // we might additionally skip a chunk prefix.
3262           prev_obj = _restart_addr;
3263         } else {
3264           prev_obj = span.start();
3265         }
3266       } else {
3267         // We want to skip the first object because
3268         // the protocol is to scan any object in its entirety
3269         // that _starts_ in this span; a fortiori, any
3270         // object starting in an earlier span is scanned
3271         // as part of an earlier claimed task.
3272         // Below we use the "careful" version of block_start
3273         // so we do not try to navigate uninitialized objects.
3274         prev_obj = sp->block_start_careful(span.start());
3275         // Below we use a variant of block_size that uses the
3276         // Printezis bits to avoid waiting for allocated
3277         // objects to become initialized/parsable.
3278         while (prev_obj < span.start()) {
3279           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3280           if (sz > 0) {
3281             prev_obj += sz;
3282           } else {
3283             // In this case we may end up doing a bit of redundant
3284             // scanning, but that appears unavoidable, short of
3285             // locking the free list locks; see bug 6324141.
3286             break;
3287           }
3288         }
3289       }
3290       if (prev_obj < span.end()) {
3291         MemRegion my_span = MemRegion(prev_obj, span.end());
3292         // Do the marking work within a non-empty span --
3293         // the last argument to the constructor indicates whether the
3294         // iteration should be incremental with periodic yields.
3295         ParMarkFromRootsClosure cl(this, _collector, my_span,
3296                                    &_collector->_markBitMap,
3297                                    work_queue(i),
3298                                    &_collector->_markStack);
3299         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3300       } // else nothing to do for this task
3301     }   // else nothing to do for this task
3302   }
3303   // We'd be tempted to assert here that since there are no
3304   // more tasks left to claim in this space, the global_finger
3305   // must exceed space->top() and a fortiori space->end(). However,
3306   // that would not quite be correct because the bumping of
3307   // global_finger occurs strictly after the claiming of a task,
3308   // so by the time we reach here the global finger may not yet
3309   // have been bumped up by the thread that claimed the last
3310   // task.
3311   pst->all_tasks_completed();
3312 }
3313 
3314 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3315  private:
3316   CMSCollector* _collector;
3317   CMSConcMarkingTask* _task;
3318   MemRegion     _span;
3319   CMSBitMap*    _bit_map;
3320   CMSMarkStack* _overflow_stack;
3321   OopTaskQueue* _work_queue;
3322  protected:
3323   DO_OOP_WORK_DEFN
3324  public:
3325   ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3326                         CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3327     MetadataAwareOopClosure(collector->ref_processor()),
3328     _collector(collector),
3329     _task(task),
3330     _span(collector->_span),
3331     _work_queue(work_queue),
3332     _bit_map(bit_map),
3333     _overflow_stack(overflow_stack)
3334   { }
3335   virtual void do_oop(oop* p);
3336   virtual void do_oop(narrowOop* p);
3337 
3338   void trim_queue(size_t max);
3339   void handle_stack_overflow(HeapWord* lost);
3340   void do_yield_check() {
3341     if (_task->should_yield()) {
3342       _task->yield();
3343     }
3344   }
3345 };
3346 
3347 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3348 
3349 // Grey object scanning during work stealing phase --
3350 // the salient assumption here is that any references
3351 // that are in these stolen objects being scanned must
3352 // already have been initialized (else they would not have
3353 // been published), so we do not need to check for
3354 // uninitialized objects before pushing here.
3355 void ParConcMarkingClosure::do_oop(oop obj) {
3356   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3357   HeapWord* addr = (HeapWord*)obj;
3358   // Check if oop points into the CMS generation
3359   // and is not marked
3360   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3361     // a white object ...
3362     // If we manage to "claim" the object, by being the
3363     // first thread to mark it, then we push it on our
3364     // marking stack
3365     if (_bit_map->par_mark(addr)) {     // ... now grey
3366       // push on work queue (grey set)
3367       bool simulate_overflow = false;
3368       NOT_PRODUCT(
3369         if (CMSMarkStackOverflowALot &&
3370             _collector->simulate_overflow()) {
3371           // simulate a stack overflow
3372           simulate_overflow = true;
3373         }
3374       )
3375       if (simulate_overflow ||
3376           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3377         // stack overflow
3378         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3379         // We cannot assert that the overflow stack is full because
3380         // it may have been emptied since.
3381         assert(simulate_overflow ||
3382                _work_queue->size() == _work_queue->max_elems(),
3383               "Else push should have succeeded");
3384         handle_stack_overflow(addr);
3385       }
3386     } // Else, some other thread got there first
3387     do_yield_check();
3388   }
3389 }
3390 
3391 void ParConcMarkingClosure::do_oop(oop* p)       { ParConcMarkingClosure::do_oop_work(p); }
3392 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3393 
3394 void ParConcMarkingClosure::trim_queue(size_t max) {
3395   while (_work_queue->size() > max) {
3396     oop new_oop;
3397     if (_work_queue->pop_local(new_oop)) {
3398       assert(oopDesc::is_oop(new_oop), "Should be an oop");
3399       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3400       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3401       new_oop->oop_iterate(this);  // do_oop() above
3402       do_yield_check();
3403     }
3404   }
3405 }
3406 
3407 // Upon stack overflow, we discard (part of) the stack,
3408 // remembering the least address amongst those discarded
3409 // in CMSCollector's _restart_address.
3410 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3411   // We need to do this under a mutex to prevent other
3412   // workers from interfering with the work done below.
3413   MutexLockerEx ml(_overflow_stack->par_lock(),
3414                    Mutex::_no_safepoint_check_flag);
3415   // Remember the least grey address discarded
3416   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3417   _collector->lower_restart_addr(ra);
3418   _overflow_stack->reset();  // discard stack contents
3419   _overflow_stack->expand(); // expand the stack if possible
3420 }
3421 
3422 
3423 void CMSConcMarkingTask::do_work_steal(int i) {
3424   OopTaskQueue* work_q = work_queue(i);
3425   oop obj_to_scan;
3426   CMSBitMap* bm = &(_collector->_markBitMap);
3427   CMSMarkStack* ovflw = &(_collector->_markStack);
3428   int* seed = _collector->hash_seed(i);
3429   ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3430   while (true) {
3431     cl.trim_queue(0);
3432     assert(work_q->size() == 0, "Should have been emptied above");
3433     if (get_work_from_overflow_stack(ovflw, work_q)) {
3434       // Can't assert below because the work obtained from the
3435       // overflow stack may already have been stolen from us.
3436       // assert(work_q->size() > 0, "Work from overflow stack");
3437       continue;
3438     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3439       assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3440       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3441       obj_to_scan->oop_iterate(&cl);
3442     } else if (terminator()->offer_termination(&_term_term)) {
3443       assert(work_q->size() == 0, "Impossible!");
3444       break;
3445     } else if (yielding() || should_yield()) {
3446       yield();
3447     }
3448   }
3449 }
3450 
3451 // This is run by the CMS (coordinator) thread.
3452 void CMSConcMarkingTask::coordinator_yield() {
3453   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3454          "CMS thread should hold CMS token");
3455   // First give up the locks, then yield, then re-lock
3456   // We should probably use a constructor/destructor idiom to
3457   // do this unlock/lock or modify the MutexUnlocker class to
3458   // serve our purpose. XXX
3459   assert_lock_strong(_bit_map_lock);
3460   _bit_map_lock->unlock();
3461   ConcurrentMarkSweepThread::desynchronize(true);
3462   _collector->stopTimer();
3463   _collector->incrementYields();
3464 
3465   // It is possible for whichever thread initiated the yield request
3466   // not to get a chance to wake up and take the bitmap lock between
3467   // this thread releasing it and reacquiring it. So, while the
3468   // should_yield() flag is on, let's sleep for a bit to give the
3469   // other thread a chance to wake up. The limit imposed on the number
3470   // of iterations is defensive, to avoid any unforseen circumstances
3471   // putting us into an infinite loop. Since it's always been this
3472   // (coordinator_yield()) method that was observed to cause the
3473   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3474   // which is by default non-zero. For the other seven methods that
3475   // also perform the yield operation, as are using a different
3476   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3477   // can enable the sleeping for those methods too, if necessary.
3478   // See 6442774.
3479   //
3480   // We really need to reconsider the synchronization between the GC
3481   // thread and the yield-requesting threads in the future and we
3482   // should really use wait/notify, which is the recommended
3483   // way of doing this type of interaction. Additionally, we should
3484   // consolidate the eight methods that do the yield operation and they
3485   // are almost identical into one for better maintainability and
3486   // readability. See 6445193.
3487   //
3488   // Tony 2006.06.29
3489   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3490                    ConcurrentMarkSweepThread::should_yield() &&
3491                    !CMSCollector::foregroundGCIsActive(); ++i) {
3492     os::sleep(Thread::current(), 1, false);
3493   }
3494 
3495   ConcurrentMarkSweepThread::synchronize(true);
3496   _bit_map_lock->lock_without_safepoint_check();
3497   _collector->startTimer();
3498 }
3499 
3500 bool CMSCollector::do_marking_mt() {
3501   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3502   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3503                                                                   conc_workers()->active_workers(),
3504                                                                   Threads::number_of_non_daemon_threads());
3505   num_workers = conc_workers()->update_active_workers(num_workers);
3506   log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3507 
3508   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3509 
3510   CMSConcMarkingTask tsk(this,
3511                          cms_space,
3512                          conc_workers(),
3513                          task_queues());
3514 
3515   // Since the actual number of workers we get may be different
3516   // from the number we requested above, do we need to do anything different
3517   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3518   // class?? XXX
3519   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3520 
3521   // Refs discovery is already non-atomic.
3522   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3523   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3524   conc_workers()->start_task(&tsk);
3525   while (tsk.yielded()) {
3526     tsk.coordinator_yield();
3527     conc_workers()->continue_task(&tsk);
3528   }
3529   // If the task was aborted, _restart_addr will be non-NULL
3530   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3531   while (_restart_addr != NULL) {
3532     // XXX For now we do not make use of ABORTED state and have not
3533     // yet implemented the right abort semantics (even in the original
3534     // single-threaded CMS case). That needs some more investigation
3535     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3536     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3537     // If _restart_addr is non-NULL, a marking stack overflow
3538     // occurred; we need to do a fresh marking iteration from the
3539     // indicated restart address.
3540     if (_foregroundGCIsActive) {
3541       // We may be running into repeated stack overflows, having
3542       // reached the limit of the stack size, while making very
3543       // slow forward progress. It may be best to bail out and
3544       // let the foreground collector do its job.
3545       // Clear _restart_addr, so that foreground GC
3546       // works from scratch. This avoids the headache of
3547       // a "rescan" which would otherwise be needed because
3548       // of the dirty mod union table & card table.
3549       _restart_addr = NULL;
3550       return false;
3551     }
3552     // Adjust the task to restart from _restart_addr
3553     tsk.reset(_restart_addr);
3554     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3555                   _restart_addr);
3556     _restart_addr = NULL;
3557     // Get the workers going again
3558     conc_workers()->start_task(&tsk);
3559     while (tsk.yielded()) {
3560       tsk.coordinator_yield();
3561       conc_workers()->continue_task(&tsk);
3562     }
3563   }
3564   assert(tsk.completed(), "Inconsistency");
3565   assert(tsk.result() == true, "Inconsistency");
3566   return true;
3567 }
3568 
3569 bool CMSCollector::do_marking_st() {
3570   ResourceMark rm;
3571   HandleMark   hm;
3572 
3573   // Temporarily make refs discovery single threaded (non-MT)
3574   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3575   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3576     &_markStack, CMSYield);
3577   // the last argument to iterate indicates whether the iteration
3578   // should be incremental with periodic yields.
3579   _markBitMap.iterate(&markFromRootsClosure);
3580   // If _restart_addr is non-NULL, a marking stack overflow
3581   // occurred; we need to do a fresh iteration from the
3582   // indicated restart address.
3583   while (_restart_addr != NULL) {
3584     if (_foregroundGCIsActive) {
3585       // We may be running into repeated stack overflows, having
3586       // reached the limit of the stack size, while making very
3587       // slow forward progress. It may be best to bail out and
3588       // let the foreground collector do its job.
3589       // Clear _restart_addr, so that foreground GC
3590       // works from scratch. This avoids the headache of
3591       // a "rescan" which would otherwise be needed because
3592       // of the dirty mod union table & card table.
3593       _restart_addr = NULL;
3594       return false;  // indicating failure to complete marking
3595     }
3596     // Deal with stack overflow:
3597     // we restart marking from _restart_addr
3598     HeapWord* ra = _restart_addr;
3599     markFromRootsClosure.reset(ra);
3600     _restart_addr = NULL;
3601     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3602   }
3603   return true;
3604 }
3605 
3606 void CMSCollector::preclean() {
3607   check_correct_thread_executing();
3608   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3609   verify_work_stacks_empty();
3610   verify_overflow_empty();
3611   _abort_preclean = false;
3612   if (CMSPrecleaningEnabled) {
3613     if (!CMSEdenChunksRecordAlways) {
3614       _eden_chunk_index = 0;
3615     }
3616     size_t used = get_eden_used();
3617     size_t capacity = get_eden_capacity();
3618     // Don't start sampling unless we will get sufficiently
3619     // many samples.
3620     if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3621                 * CMSScheduleRemarkEdenPenetration)) {
3622       _start_sampling = true;
3623     } else {
3624       _start_sampling = false;
3625     }
3626     GCTraceCPUTime tcpu;
3627     CMSPhaseAccounting pa(this, "Concurrent Preclean");
3628     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3629   }
3630   CMSTokenSync x(true); // is cms thread
3631   if (CMSPrecleaningEnabled) {
3632     sample_eden();
3633     _collectorState = AbortablePreclean;
3634   } else {
3635     _collectorState = FinalMarking;
3636   }
3637   verify_work_stacks_empty();
3638   verify_overflow_empty();
3639 }
3640 
3641 // Try and schedule the remark such that young gen
3642 // occupancy is CMSScheduleRemarkEdenPenetration %.
3643 void CMSCollector::abortable_preclean() {
3644   check_correct_thread_executing();
3645   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3646   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3647 
3648   // If Eden's current occupancy is below this threshold,
3649   // immediately schedule the remark; else preclean
3650   // past the next scavenge in an effort to
3651   // schedule the pause as described above. By choosing
3652   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3653   // we will never do an actual abortable preclean cycle.
3654   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3655     GCTraceCPUTime tcpu;
3656     CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3657     // We need more smarts in the abortable preclean
3658     // loop below to deal with cases where allocation
3659     // in young gen is very very slow, and our precleaning
3660     // is running a losing race against a horde of
3661     // mutators intent on flooding us with CMS updates
3662     // (dirty cards).
3663     // One, admittedly dumb, strategy is to give up
3664     // after a certain number of abortable precleaning loops
3665     // or after a certain maximum time. We want to make
3666     // this smarter in the next iteration.
3667     // XXX FIX ME!!! YSR
3668     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3669     while (!(should_abort_preclean() ||
3670              ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3671       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3672       cumworkdone += workdone;
3673       loops++;
3674       // Voluntarily terminate abortable preclean phase if we have
3675       // been at it for too long.
3676       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3677           loops >= CMSMaxAbortablePrecleanLoops) {
3678         log_debug(gc)(" CMS: abort preclean due to loops ");
3679         break;
3680       }
3681       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3682         log_debug(gc)(" CMS: abort preclean due to time ");
3683         break;
3684       }
3685       // If we are doing little work each iteration, we should
3686       // take a short break.
3687       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3688         // Sleep for some time, waiting for work to accumulate
3689         stopTimer();
3690         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3691         startTimer();
3692         waited++;
3693       }
3694     }
3695     log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3696                                loops, waited, cumworkdone);
3697   }
3698   CMSTokenSync x(true); // is cms thread
3699   if (_collectorState != Idling) {
3700     assert(_collectorState == AbortablePreclean,
3701            "Spontaneous state transition?");
3702     _collectorState = FinalMarking;
3703   } // Else, a foreground collection completed this CMS cycle.
3704   return;
3705 }
3706 
3707 // Respond to an Eden sampling opportunity
3708 void CMSCollector::sample_eden() {
3709   // Make sure a young gc cannot sneak in between our
3710   // reading and recording of a sample.
3711   assert(Thread::current()->is_ConcurrentGC_thread(),
3712          "Only the cms thread may collect Eden samples");
3713   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3714          "Should collect samples while holding CMS token");
3715   if (!_start_sampling) {
3716     return;
3717   }
3718   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3719   // is populated by the young generation.
3720   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3721     if (_eden_chunk_index < _eden_chunk_capacity) {
3722       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3723       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3724              "Unexpected state of Eden");
3725       // We'd like to check that what we just sampled is an oop-start address;
3726       // however, we cannot do that here since the object may not yet have been
3727       // initialized. So we'll instead do the check when we _use_ this sample
3728       // later.
3729       if (_eden_chunk_index == 0 ||
3730           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3731                          _eden_chunk_array[_eden_chunk_index-1])
3732            >= CMSSamplingGrain)) {
3733         _eden_chunk_index++;  // commit sample
3734       }
3735     }
3736   }
3737   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3738     size_t used = get_eden_used();
3739     size_t capacity = get_eden_capacity();
3740     assert(used <= capacity, "Unexpected state of Eden");
3741     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3742       _abort_preclean = true;
3743     }
3744   }
3745 }
3746 
3747 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3748   assert(_collectorState == Precleaning ||
3749          _collectorState == AbortablePreclean, "incorrect state");
3750   ResourceMark rm;
3751   HandleMark   hm;
3752 
3753   // Precleaning is currently not MT but the reference processor
3754   // may be set for MT.  Disable it temporarily here.
3755   SpanReferenceProcessor* rp = ref_processor();
3756   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3757 
3758   // Do one pass of scrubbing the discovered reference lists
3759   // to remove any reference objects with strongly-reachable
3760   // referents.
3761   if (clean_refs) {
3762     CMSPrecleanRefsYieldClosure yield_cl(this);
3763     assert(rp->span().equals(_span), "Spans should be equal");
3764     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3765                                    &_markStack, true /* preclean */);
3766     CMSDrainMarkingStackClosure complete_trace(this,
3767                                    _span, &_markBitMap, &_markStack,
3768                                    &keep_alive, true /* preclean */);
3769 
3770     // We don't want this step to interfere with a young
3771     // collection because we don't want to take CPU
3772     // or memory bandwidth away from the young GC threads
3773     // (which may be as many as there are CPUs).
3774     // Note that we don't need to protect ourselves from
3775     // interference with mutators because they can't
3776     // manipulate the discovered reference lists nor affect
3777     // the computed reachability of the referents, the
3778     // only properties manipulated by the precleaning
3779     // of these reference lists.
3780     stopTimer();
3781     CMSTokenSyncWithLocks x(true /* is cms thread */,
3782                             bitMapLock());
3783     startTimer();
3784     sample_eden();
3785 
3786     // The following will yield to allow foreground
3787     // collection to proceed promptly. XXX YSR:
3788     // The code in this method may need further
3789     // tweaking for better performance and some restructuring
3790     // for cleaner interfaces.
3791     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3792     rp->preclean_discovered_references(
3793           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3794           gc_timer);
3795   }
3796 
3797   if (clean_survivor) {  // preclean the active survivor space(s)
3798     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3799                              &_markBitMap, &_modUnionTable,
3800                              &_markStack, true /* precleaning phase */);
3801     stopTimer();
3802     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3803                              bitMapLock());
3804     startTimer();
3805     unsigned int before_count =
3806       CMSHeap::heap()->total_collections();
3807     SurvivorSpacePrecleanClosure
3808       sss_cl(this, _span, &_markBitMap, &_markStack,
3809              &pam_cl, before_count, CMSYield);
3810     _young_gen->from()->object_iterate_careful(&sss_cl);
3811     _young_gen->to()->object_iterate_careful(&sss_cl);
3812   }
3813   MarkRefsIntoAndScanClosure
3814     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3815              &_markStack, this, CMSYield,
3816              true /* precleaning phase */);
3817   // CAUTION: The following closure has persistent state that may need to
3818   // be reset upon a decrease in the sequence of addresses it
3819   // processes.
3820   ScanMarkedObjectsAgainCarefullyClosure
3821     smoac_cl(this, _span,
3822       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3823 
3824   // Preclean dirty cards in ModUnionTable and CardTable using
3825   // appropriate convergence criterion;
3826   // repeat CMSPrecleanIter times unless we find that
3827   // we are losing.
3828   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3829   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3830          "Bad convergence multiplier");
3831   assert(CMSPrecleanThreshold >= 100,
3832          "Unreasonably low CMSPrecleanThreshold");
3833 
3834   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3835   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3836        numIter < CMSPrecleanIter;
3837        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3838     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3839     log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3840     // Either there are very few dirty cards, so re-mark
3841     // pause will be small anyway, or our pre-cleaning isn't
3842     // that much faster than the rate at which cards are being
3843     // dirtied, so we might as well stop and re-mark since
3844     // precleaning won't improve our re-mark time by much.
3845     if (curNumCards <= CMSPrecleanThreshold ||
3846         (numIter > 0 &&
3847          (curNumCards * CMSPrecleanDenominator >
3848          lastNumCards * CMSPrecleanNumerator))) {
3849       numIter++;
3850       cumNumCards += curNumCards;
3851       break;
3852     }
3853   }
3854 
3855   preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3856 
3857   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3858   cumNumCards += curNumCards;
3859   log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3860                              curNumCards, cumNumCards, numIter);
3861   return cumNumCards;   // as a measure of useful work done
3862 }
3863 
3864 // PRECLEANING NOTES:
3865 // Precleaning involves:
3866 // . reading the bits of the modUnionTable and clearing the set bits.
3867 // . For the cards corresponding to the set bits, we scan the
3868 //   objects on those cards. This means we need the free_list_lock
3869 //   so that we can safely iterate over the CMS space when scanning
3870 //   for oops.
3871 // . When we scan the objects, we'll be both reading and setting
3872 //   marks in the marking bit map, so we'll need the marking bit map.
3873 // . For protecting _collector_state transitions, we take the CGC_lock.
3874 //   Note that any races in the reading of of card table entries by the
3875 //   CMS thread on the one hand and the clearing of those entries by the
3876 //   VM thread or the setting of those entries by the mutator threads on the
3877 //   other are quite benign. However, for efficiency it makes sense to keep
3878 //   the VM thread from racing with the CMS thread while the latter is
3879 //   dirty card info to the modUnionTable. We therefore also use the
3880 //   CGC_lock to protect the reading of the card table and the mod union
3881 //   table by the CM thread.
3882 // . We run concurrently with mutator updates, so scanning
3883 //   needs to be done carefully  -- we should not try to scan
3884 //   potentially uninitialized objects.
3885 //
3886 // Locking strategy: While holding the CGC_lock, we scan over and
3887 // reset a maximal dirty range of the mod union / card tables, then lock
3888 // the free_list_lock and bitmap lock to do a full marking, then
3889 // release these locks; and repeat the cycle. This allows for a
3890 // certain amount of fairness in the sharing of these locks between
3891 // the CMS collector on the one hand, and the VM thread and the
3892 // mutators on the other.
3893 
3894 // NOTE: preclean_mod_union_table() and preclean_card_table()
3895 // further below are largely identical; if you need to modify
3896 // one of these methods, please check the other method too.
3897 
3898 size_t CMSCollector::preclean_mod_union_table(
3899   ConcurrentMarkSweepGeneration* old_gen,
3900   ScanMarkedObjectsAgainCarefullyClosure* cl) {
3901   verify_work_stacks_empty();
3902   verify_overflow_empty();
3903 
3904   // strategy: starting with the first card, accumulate contiguous
3905   // ranges of dirty cards; clear these cards, then scan the region
3906   // covered by these cards.
3907 
3908   // Since all of the MUT is committed ahead, we can just use
3909   // that, in case the generations expand while we are precleaning.
3910   // It might also be fine to just use the committed part of the
3911   // generation, but we might potentially miss cards when the
3912   // generation is rapidly expanding while we are in the midst
3913   // of precleaning.
3914   HeapWord* startAddr = old_gen->reserved().start();
3915   HeapWord* endAddr   = old_gen->reserved().end();
3916 
3917   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
3918 
3919   size_t numDirtyCards, cumNumDirtyCards;
3920   HeapWord *nextAddr, *lastAddr;
3921   for (cumNumDirtyCards = numDirtyCards = 0,
3922        nextAddr = lastAddr = startAddr;
3923        nextAddr < endAddr;
3924        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3925 
3926     ResourceMark rm;
3927     HandleMark   hm;
3928 
3929     MemRegion dirtyRegion;
3930     {
3931       stopTimer();
3932       // Potential yield point
3933       CMSTokenSync ts(true);
3934       startTimer();
3935       sample_eden();
3936       // Get dirty region starting at nextOffset (inclusive),
3937       // simultaneously clearing it.
3938       dirtyRegion =
3939         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3940       assert(dirtyRegion.start() >= nextAddr,
3941              "returned region inconsistent?");
3942     }
3943     // Remember where the next search should begin.
3944     // The returned region (if non-empty) is a right open interval,
3945     // so lastOffset is obtained from the right end of that
3946     // interval.
3947     lastAddr = dirtyRegion.end();
3948     // Should do something more transparent and less hacky XXX
3949     numDirtyCards =
3950       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3951 
3952     // We'll scan the cards in the dirty region (with periodic
3953     // yields for foreground GC as needed).
3954     if (!dirtyRegion.is_empty()) {
3955       assert(numDirtyCards > 0, "consistency check");
3956       HeapWord* stop_point = NULL;
3957       stopTimer();
3958       // Potential yield point
3959       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3960                                bitMapLock());
3961       startTimer();
3962       {
3963         verify_work_stacks_empty();
3964         verify_overflow_empty();
3965         sample_eden();
3966         stop_point =
3967           old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3968       }
3969       if (stop_point != NULL) {
3970         // The careful iteration stopped early either because it found an
3971         // uninitialized object, or because we were in the midst of an
3972         // "abortable preclean", which should now be aborted. Redirty
3973         // the bits corresponding to the partially-scanned or unscanned
3974         // cards. We'll either restart at the next block boundary or
3975         // abort the preclean.
3976         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3977                "Should only be AbortablePreclean.");
3978         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3979         if (should_abort_preclean()) {
3980           break; // out of preclean loop
3981         } else {
3982           // Compute the next address at which preclean should pick up;
3983           // might need bitMapLock in order to read P-bits.
3984           lastAddr = next_card_start_after_block(stop_point);
3985         }
3986       }
3987     } else {
3988       assert(lastAddr == endAddr, "consistency check");
3989       assert(numDirtyCards == 0, "consistency check");
3990       break;
3991     }
3992   }
3993   verify_work_stacks_empty();
3994   verify_overflow_empty();
3995   return cumNumDirtyCards;
3996 }
3997 
3998 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3999 // below are largely identical; if you need to modify
4000 // one of these methods, please check the other method too.
4001 
4002 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4003   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4004   // strategy: it's similar to precleamModUnionTable above, in that
4005   // we accumulate contiguous ranges of dirty cards, mark these cards
4006   // precleaned, then scan the region covered by these cards.
4007   HeapWord* endAddr   = (HeapWord*)(old_gen->_virtual_space.high());
4008   HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4009 
4010   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
4011 
4012   size_t numDirtyCards, cumNumDirtyCards;
4013   HeapWord *lastAddr, *nextAddr;
4014 
4015   for (cumNumDirtyCards = numDirtyCards = 0,
4016        nextAddr = lastAddr = startAddr;
4017        nextAddr < endAddr;
4018        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4019 
4020     ResourceMark rm;
4021     HandleMark   hm;
4022 
4023     MemRegion dirtyRegion;
4024     {
4025       // See comments in "Precleaning notes" above on why we
4026       // do this locking. XXX Could the locking overheads be
4027       // too high when dirty cards are sparse? [I don't think so.]
4028       stopTimer();
4029       CMSTokenSync x(true); // is cms thread
4030       startTimer();
4031       sample_eden();
4032       // Get and clear dirty region from card table
4033       dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr),
4034                                                       true,
4035                                                       CardTable::precleaned_card_val());
4036 
4037       assert(dirtyRegion.start() >= nextAddr,
4038              "returned region inconsistent?");
4039     }
4040     lastAddr = dirtyRegion.end();
4041     numDirtyCards =
4042       dirtyRegion.word_size()/CardTable::card_size_in_words;
4043 
4044     if (!dirtyRegion.is_empty()) {
4045       stopTimer();
4046       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4047       startTimer();
4048       sample_eden();
4049       verify_work_stacks_empty();
4050       verify_overflow_empty();
4051       HeapWord* stop_point =
4052         old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4053       if (stop_point != NULL) {
4054         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4055                "Should only be AbortablePreclean.");
4056         _ct->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4057         if (should_abort_preclean()) {
4058           break; // out of preclean loop
4059         } else {
4060           // Compute the next address at which preclean should pick up.
4061           lastAddr = next_card_start_after_block(stop_point);
4062         }
4063       }
4064     } else {
4065       break;
4066     }
4067   }
4068   verify_work_stacks_empty();
4069   verify_overflow_empty();
4070   return cumNumDirtyCards;
4071 }
4072 
4073 class PrecleanCLDClosure : public CLDClosure {
4074   MetadataAwareOopsInGenClosure* _cm_closure;
4075  public:
4076   PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4077   void do_cld(ClassLoaderData* cld) {
4078     if (cld->has_accumulated_modified_oops()) {
4079       cld->clear_accumulated_modified_oops();
4080 
4081       _cm_closure->do_cld(cld);
4082     }
4083   }
4084 };
4085 
4086 // The freelist lock is needed to prevent asserts, is it really needed?
4087 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4088 
4089   cl->set_freelistLock(freelistLock);
4090 
4091   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4092 
4093   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4094   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4095   PrecleanCLDClosure preclean_closure(cl);
4096   ClassLoaderDataGraph::cld_do(&preclean_closure);
4097 
4098   verify_work_stacks_empty();
4099   verify_overflow_empty();
4100 }
4101 
4102 void CMSCollector::checkpointRootsFinal() {
4103   assert(_collectorState == FinalMarking, "incorrect state transition?");
4104   check_correct_thread_executing();
4105   // world is stopped at this checkpoint
4106   assert(SafepointSynchronize::is_at_safepoint(),
4107          "world should be stopped");
4108   TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4109 
4110   verify_work_stacks_empty();
4111   verify_overflow_empty();
4112 
4113   log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4114                 _young_gen->used() / K, _young_gen->capacity() / K);
4115   {
4116     if (CMSScavengeBeforeRemark) {
4117       CMSHeap* heap = CMSHeap::heap();
4118       // Temporarily set flag to false, GCH->do_collection will
4119       // expect it to be false and set to true
4120       FlagSetting fl(heap->_is_gc_active, false);
4121 
4122       heap->do_collection(true,                      // full (i.e. force, see below)
4123                           false,                     // !clear_all_soft_refs
4124                           0,                         // size
4125                           false,                     // is_tlab
4126                           GenCollectedHeap::YoungGen // type
4127         );
4128     }
4129     FreelistLocker x(this);
4130     MutexLockerEx y(bitMapLock(),
4131                     Mutex::_no_safepoint_check_flag);
4132     checkpointRootsFinalWork();
4133   }
4134   verify_work_stacks_empty();
4135   verify_overflow_empty();
4136 }
4137 
4138 void CMSCollector::checkpointRootsFinalWork() {
4139   GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4140 
4141   assert(haveFreelistLocks(), "must have free list locks");
4142   assert_lock_strong(bitMapLock());
4143 
4144   ResourceMark rm;
4145   HandleMark   hm;
4146 
4147   CMSHeap* heap = CMSHeap::heap();
4148 
4149   if (should_unload_classes()) {
4150     CodeCache::gc_prologue();
4151   }
4152   assert(haveFreelistLocks(), "must have free list locks");
4153   assert_lock_strong(bitMapLock());
4154 
4155   // We might assume that we need not fill TLAB's when
4156   // CMSScavengeBeforeRemark is set, because we may have just done
4157   // a scavenge which would have filled all TLAB's -- and besides
4158   // Eden would be empty. This however may not always be the case --
4159   // for instance although we asked for a scavenge, it may not have
4160   // happened because of a JNI critical section. We probably need
4161   // a policy for deciding whether we can in that case wait until
4162   // the critical section releases and then do the remark following
4163   // the scavenge, and skip it here. In the absence of that policy,
4164   // or of an indication of whether the scavenge did indeed occur,
4165   // we cannot rely on TLAB's having been filled and must do
4166   // so here just in case a scavenge did not happen.
4167   heap->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4168   // Update the saved marks which may affect the root scans.
4169   heap->save_marks();
4170 
4171   print_eden_and_survivor_chunk_arrays();
4172 
4173   {
4174 #if COMPILER2_OR_JVMCI
4175     DerivedPointerTableDeactivate dpt_deact;
4176 #endif
4177 
4178     // Note on the role of the mod union table:
4179     // Since the marker in "markFromRoots" marks concurrently with
4180     // mutators, it is possible for some reachable objects not to have been
4181     // scanned. For instance, an only reference to an object A was
4182     // placed in object B after the marker scanned B. Unless B is rescanned,
4183     // A would be collected. Such updates to references in marked objects
4184     // are detected via the mod union table which is the set of all cards
4185     // dirtied since the first checkpoint in this GC cycle and prior to
4186     // the most recent young generation GC, minus those cleaned up by the
4187     // concurrent precleaning.
4188     if (CMSParallelRemarkEnabled) {
4189       GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4190       do_remark_parallel();
4191     } else {
4192       GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4193       do_remark_non_parallel();
4194     }
4195   }
4196   verify_work_stacks_empty();
4197   verify_overflow_empty();
4198 
4199   {
4200     GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4201     refProcessingWork();
4202   }
4203   verify_work_stacks_empty();
4204   verify_overflow_empty();
4205 
4206   if (should_unload_classes()) {
4207     CodeCache::gc_epilogue();
4208   }
4209   JvmtiExport::gc_epilogue();
4210 
4211   // If we encountered any (marking stack / work queue) overflow
4212   // events during the current CMS cycle, take appropriate
4213   // remedial measures, where possible, so as to try and avoid
4214   // recurrence of that condition.
4215   assert(_markStack.isEmpty(), "No grey objects");
4216   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4217                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4218   if (ser_ovflw > 0) {
4219     log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4220                          _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4221     _markStack.expand();
4222     _ser_pmc_remark_ovflw = 0;
4223     _ser_pmc_preclean_ovflw = 0;
4224     _ser_kac_preclean_ovflw = 0;
4225     _ser_kac_ovflw = 0;
4226   }
4227   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4228      log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4229                           _par_pmc_remark_ovflw, _par_kac_ovflw);
4230      _par_pmc_remark_ovflw = 0;
4231     _par_kac_ovflw = 0;
4232   }
4233    if (_markStack._hit_limit > 0) {
4234      log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4235                           _markStack._hit_limit);
4236    }
4237    if (_markStack._failed_double > 0) {
4238      log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4239                           _markStack._failed_double, _markStack.capacity());
4240    }
4241   _markStack._hit_limit = 0;
4242   _markStack._failed_double = 0;
4243 
4244   if ((VerifyAfterGC || VerifyDuringGC) &&
4245       CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4246     verify_after_remark();
4247   }
4248 
4249   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4250 
4251   // Change under the freelistLocks.
4252   _collectorState = Sweeping;
4253   // Call isAllClear() under bitMapLock
4254   assert(_modUnionTable.isAllClear(),
4255       "Should be clear by end of the final marking");
4256   assert(_ct->cld_rem_set()->mod_union_is_clear(),
4257       "Should be clear by end of the final marking");
4258 }
4259 
4260 void CMSParInitialMarkTask::work(uint worker_id) {
4261   elapsedTimer _timer;
4262   ResourceMark rm;
4263   HandleMark   hm;
4264 
4265   // ---------- scan from roots --------------
4266   _timer.start();
4267   CMSHeap* heap = CMSHeap::heap();
4268   ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4269 
4270   // ---------- young gen roots --------------
4271   {
4272     work_on_young_gen_roots(&par_mri_cl);
4273     _timer.stop();
4274     log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4275   }
4276 
4277   // ---------- remaining roots --------------
4278   _timer.reset();
4279   _timer.start();
4280 
4281   CLDToOopClosure cld_closure(&par_mri_cl, true);
4282 
4283   heap->cms_process_roots(_strong_roots_scope,
4284                           false,     // yg was scanned above
4285                           GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4286                           _collector->should_unload_classes(),
4287                           &par_mri_cl,
4288                           &cld_closure);
4289   assert(_collector->should_unload_classes()
4290          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4291          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4292   _timer.stop();
4293   log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4294 }
4295 
4296 // Parallel remark task
4297 class CMSParRemarkTask: public CMSParMarkTask {
4298   CompactibleFreeListSpace* _cms_space;
4299 
4300   // The per-thread work queues, available here for stealing.
4301   OopTaskQueueSet*       _task_queues;
4302   ParallelTaskTerminator _term;
4303   StrongRootsScope*      _strong_roots_scope;
4304 
4305  public:
4306   // A value of 0 passed to n_workers will cause the number of
4307   // workers to be taken from the active workers in the work gang.
4308   CMSParRemarkTask(CMSCollector* collector,
4309                    CompactibleFreeListSpace* cms_space,
4310                    uint n_workers, WorkGang* workers,
4311                    OopTaskQueueSet* task_queues,
4312                    StrongRootsScope* strong_roots_scope):
4313     CMSParMarkTask("Rescan roots and grey objects in parallel",
4314                    collector, n_workers),
4315     _cms_space(cms_space),
4316     _task_queues(task_queues),
4317     _term(n_workers, task_queues),
4318     _strong_roots_scope(strong_roots_scope) { }
4319 
4320   OopTaskQueueSet* task_queues() { return _task_queues; }
4321 
4322   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4323 
4324   ParallelTaskTerminator* terminator() { return &_term; }
4325   uint n_workers() { return _n_workers; }
4326 
4327   void work(uint worker_id);
4328 
4329  private:
4330   // ... of  dirty cards in old space
4331   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4332                                   ParMarkRefsIntoAndScanClosure* cl);
4333 
4334   // ... work stealing for the above
4335   void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4336 };
4337 
4338 class RemarkCLDClosure : public CLDClosure {
4339   CLDToOopClosure _cm_closure;
4340  public:
4341   RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4342   void do_cld(ClassLoaderData* cld) {
4343     // Check if we have modified any oops in the CLD during the concurrent marking.
4344     if (cld->has_accumulated_modified_oops()) {
4345       cld->clear_accumulated_modified_oops();
4346 
4347       // We could have transfered the current modified marks to the accumulated marks,
4348       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4349     } else if (cld->has_modified_oops()) {
4350       // Don't clear anything, this info is needed by the next young collection.
4351     } else {
4352       // No modified oops in the ClassLoaderData.
4353       return;
4354     }
4355 
4356     // The klass has modified fields, need to scan the klass.
4357     _cm_closure.do_cld(cld);
4358   }
4359 };
4360 
4361 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4362   ParNewGeneration* young_gen = _collector->_young_gen;
4363   ContiguousSpace* eden_space = young_gen->eden();
4364   ContiguousSpace* from_space = young_gen->from();
4365   ContiguousSpace* to_space   = young_gen->to();
4366 
4367   HeapWord** eca = _collector->_eden_chunk_array;
4368   size_t     ect = _collector->_eden_chunk_index;
4369   HeapWord** sca = _collector->_survivor_chunk_array;
4370   size_t     sct = _collector->_survivor_chunk_index;
4371 
4372   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4373   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4374 
4375   do_young_space_rescan(cl, to_space, NULL, 0);
4376   do_young_space_rescan(cl, from_space, sca, sct);
4377   do_young_space_rescan(cl, eden_space, eca, ect);
4378 }
4379 
4380 // work_queue(i) is passed to the closure
4381 // ParMarkRefsIntoAndScanClosure.  The "i" parameter
4382 // also is passed to do_dirty_card_rescan_tasks() and to
4383 // do_work_steal() to select the i-th task_queue.
4384 
4385 void CMSParRemarkTask::work(uint worker_id) {
4386   elapsedTimer _timer;
4387   ResourceMark rm;
4388   HandleMark   hm;
4389 
4390   // ---------- rescan from roots --------------
4391   _timer.start();
4392   CMSHeap* heap = CMSHeap::heap();
4393   ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4394     _collector->_span, _collector->ref_processor(),
4395     &(_collector->_markBitMap),
4396     work_queue(worker_id));
4397 
4398   // Rescan young gen roots first since these are likely
4399   // coarsely partitioned and may, on that account, constitute
4400   // the critical path; thus, it's best to start off that
4401   // work first.
4402   // ---------- young gen roots --------------
4403   {
4404     work_on_young_gen_roots(&par_mrias_cl);
4405     _timer.stop();
4406     log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4407   }
4408 
4409   // ---------- remaining roots --------------
4410   _timer.reset();
4411   _timer.start();
4412   heap->cms_process_roots(_strong_roots_scope,
4413                           false,     // yg was scanned above
4414                           GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4415                           _collector->should_unload_classes(),
4416                           &par_mrias_cl,
4417                           NULL);     // The dirty klasses will be handled below
4418 
4419   assert(_collector->should_unload_classes()
4420          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4421          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4422   _timer.stop();
4423   log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec",  worker_id, _timer.seconds());
4424 
4425   // ---------- unhandled CLD scanning ----------
4426   if (worker_id == 0) { // Single threaded at the moment.
4427     _timer.reset();
4428     _timer.start();
4429 
4430     // Scan all new class loader data objects and new dependencies that were
4431     // introduced during concurrent marking.
4432     ResourceMark rm;
4433     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4434     for (int i = 0; i < array->length(); i++) {
4435       par_mrias_cl.do_cld_nv(array->at(i));
4436     }
4437 
4438     // We don't need to keep track of new CLDs anymore.
4439     ClassLoaderDataGraph::remember_new_clds(false);
4440 
4441     _timer.stop();
4442     log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4443   }
4444 
4445   // We might have added oops to ClassLoaderData::_handles during the
4446   // concurrent marking phase. These oops do not always point to newly allocated objects
4447   // that are guaranteed to be kept alive.  Hence,
4448   // we do have to revisit the _handles block during the remark phase.
4449 
4450   // ---------- dirty CLD scanning ----------
4451   if (worker_id == 0) { // Single threaded at the moment.
4452     _timer.reset();
4453     _timer.start();
4454 
4455     // Scan all classes that was dirtied during the concurrent marking phase.
4456     RemarkCLDClosure remark_closure(&par_mrias_cl);
4457     ClassLoaderDataGraph::cld_do(&remark_closure);
4458 
4459     _timer.stop();
4460     log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4461   }
4462 
4463 
4464   // ---------- rescan dirty cards ------------
4465   _timer.reset();
4466   _timer.start();
4467 
4468   // Do the rescan tasks for each of the two spaces
4469   // (cms_space) in turn.
4470   // "worker_id" is passed to select the task_queue for "worker_id"
4471   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4472   _timer.stop();
4473   log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4474 
4475   // ---------- steal work from other threads ...
4476   // ---------- ... and drain overflow list.
4477   _timer.reset();
4478   _timer.start();
4479   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4480   _timer.stop();
4481   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4482 }
4483 
4484 void
4485 CMSParMarkTask::do_young_space_rescan(
4486   OopsInGenClosure* cl, ContiguousSpace* space,
4487   HeapWord** chunk_array, size_t chunk_top) {
4488   // Until all tasks completed:
4489   // . claim an unclaimed task
4490   // . compute region boundaries corresponding to task claimed
4491   //   using chunk_array
4492   // . par_oop_iterate(cl) over that region
4493 
4494   ResourceMark rm;
4495   HandleMark   hm;
4496 
4497   SequentialSubTasksDone* pst = space->par_seq_tasks();
4498 
4499   uint nth_task = 0;
4500   uint n_tasks  = pst->n_tasks();
4501 
4502   if (n_tasks > 0) {
4503     assert(pst->valid(), "Uninitialized use?");
4504     HeapWord *start, *end;
4505     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4506       // We claimed task # nth_task; compute its boundaries.
4507       if (chunk_top == 0) {  // no samples were taken
4508         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4509         start = space->bottom();
4510         end   = space->top();
4511       } else if (nth_task == 0) {
4512         start = space->bottom();
4513         end   = chunk_array[nth_task];
4514       } else if (nth_task < (uint)chunk_top) {
4515         assert(nth_task >= 1, "Control point invariant");
4516         start = chunk_array[nth_task - 1];
4517         end   = chunk_array[nth_task];
4518       } else {
4519         assert(nth_task == (uint)chunk_top, "Control point invariant");
4520         start = chunk_array[chunk_top - 1];
4521         end   = space->top();
4522       }
4523       MemRegion mr(start, end);
4524       // Verify that mr is in space
4525       assert(mr.is_empty() || space->used_region().contains(mr),
4526              "Should be in space");
4527       // Verify that "start" is an object boundary
4528       assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4529              "Should be an oop");
4530       space->par_oop_iterate(mr, cl);
4531     }
4532     pst->all_tasks_completed();
4533   }
4534 }
4535 
4536 void
4537 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4538   CompactibleFreeListSpace* sp, int i,
4539   ParMarkRefsIntoAndScanClosure* cl) {
4540   // Until all tasks completed:
4541   // . claim an unclaimed task
4542   // . compute region boundaries corresponding to task claimed
4543   // . transfer dirty bits ct->mut for that region
4544   // . apply rescanclosure to dirty mut bits for that region
4545 
4546   ResourceMark rm;
4547   HandleMark   hm;
4548 
4549   OopTaskQueue* work_q = work_queue(i);
4550   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4551   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4552   // CAUTION: This closure has state that persists across calls to
4553   // the work method dirty_range_iterate_clear() in that it has
4554   // embedded in it a (subtype of) UpwardsObjectClosure. The
4555   // use of that state in the embedded UpwardsObjectClosure instance
4556   // assumes that the cards are always iterated (even if in parallel
4557   // by several threads) in monotonically increasing order per each
4558   // thread. This is true of the implementation below which picks
4559   // card ranges (chunks) in monotonically increasing order globally
4560   // and, a-fortiori, in monotonically increasing order per thread
4561   // (the latter order being a subsequence of the former).
4562   // If the work code below is ever reorganized into a more chaotic
4563   // work-partitioning form than the current "sequential tasks"
4564   // paradigm, the use of that persistent state will have to be
4565   // revisited and modified appropriately. See also related
4566   // bug 4756801 work on which should examine this code to make
4567   // sure that the changes there do not run counter to the
4568   // assumptions made here and necessary for correctness and
4569   // efficiency. Note also that this code might yield inefficient
4570   // behavior in the case of very large objects that span one or
4571   // more work chunks. Such objects would potentially be scanned
4572   // several times redundantly. Work on 4756801 should try and
4573   // address that performance anomaly if at all possible. XXX
4574   MemRegion  full_span  = _collector->_span;
4575   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4576   MarkFromDirtyCardsClosure
4577     greyRescanClosure(_collector, full_span, // entire span of interest
4578                       sp, bm, work_q, cl);
4579 
4580   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4581   assert(pst->valid(), "Uninitialized use?");
4582   uint nth_task = 0;
4583   const int alignment = CardTable::card_size * BitsPerWord;
4584   MemRegion span = sp->used_region();
4585   HeapWord* start_addr = span.start();
4586   HeapWord* end_addr = align_up(span.end(), alignment);
4587   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4588   assert(is_aligned(start_addr, alignment), "Check alignment");
4589   assert(is_aligned(chunk_size, alignment), "Check alignment");
4590 
4591   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4592     // Having claimed the nth_task, compute corresponding mem-region,
4593     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4594     // The alignment restriction ensures that we do not need any
4595     // synchronization with other gang-workers while setting or
4596     // clearing bits in thus chunk of the MUT.
4597     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4598                                     start_addr + (nth_task+1)*chunk_size);
4599     // The last chunk's end might be way beyond end of the
4600     // used region. In that case pull back appropriately.
4601     if (this_span.end() > end_addr) {
4602       this_span.set_end(end_addr);
4603       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4604     }
4605     // Iterate over the dirty cards covering this chunk, marking them
4606     // precleaned, and setting the corresponding bits in the mod union
4607     // table. Since we have been careful to partition at Card and MUT-word
4608     // boundaries no synchronization is needed between parallel threads.
4609     _collector->_ct->dirty_card_iterate(this_span,
4610                                                  &modUnionClosure);
4611 
4612     // Having transferred these marks into the modUnionTable,
4613     // rescan the marked objects on the dirty cards in the modUnionTable.
4614     // Even if this is at a synchronous collection, the initial marking
4615     // may have been done during an asynchronous collection so there
4616     // may be dirty bits in the mod-union table.
4617     _collector->_modUnionTable.dirty_range_iterate_clear(
4618                   this_span, &greyRescanClosure);
4619     _collector->_modUnionTable.verifyNoOneBitsInRange(
4620                                  this_span.start(),
4621                                  this_span.end());
4622   }
4623   pst->all_tasks_completed();  // declare that i am done
4624 }
4625 
4626 // . see if we can share work_queues with ParNew? XXX
4627 void
4628 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4629                                 int* seed) {
4630   OopTaskQueue* work_q = work_queue(i);
4631   NOT_PRODUCT(int num_steals = 0;)
4632   oop obj_to_scan;
4633   CMSBitMap* bm = &(_collector->_markBitMap);
4634 
4635   while (true) {
4636     // Completely finish any left over work from (an) earlier round(s)
4637     cl->trim_queue(0);
4638     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4639                                          (size_t)ParGCDesiredObjsFromOverflowList);
4640     // Now check if there's any work in the overflow list
4641     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4642     // only affects the number of attempts made to get work from the
4643     // overflow list and does not affect the number of workers.  Just
4644     // pass ParallelGCThreads so this behavior is unchanged.
4645     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4646                                                 work_q,
4647                                                 ParallelGCThreads)) {
4648       // found something in global overflow list;
4649       // not yet ready to go stealing work from others.
4650       // We'd like to assert(work_q->size() != 0, ...)
4651       // because we just took work from the overflow list,
4652       // but of course we can't since all of that could have
4653       // been already stolen from us.
4654       // "He giveth and He taketh away."
4655       continue;
4656     }
4657     // Verify that we have no work before we resort to stealing
4658     assert(work_q->size() == 0, "Have work, shouldn't steal");
4659     // Try to steal from other queues that have work
4660     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4661       NOT_PRODUCT(num_steals++;)
4662       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4663       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4664       // Do scanning work
4665       obj_to_scan->oop_iterate(cl);
4666       // Loop around, finish this work, and try to steal some more
4667     } else if (terminator()->offer_termination()) {
4668         break;  // nirvana from the infinite cycle
4669     }
4670   }
4671   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4672   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4673          "Else our work is not yet done");
4674 }
4675 
4676 // Record object boundaries in _eden_chunk_array by sampling the eden
4677 // top in the slow-path eden object allocation code path and record
4678 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4679 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4680 // sampling in sample_eden() that activates during the part of the
4681 // preclean phase.
4682 void CMSCollector::sample_eden_chunk() {
4683   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4684     if (_eden_chunk_lock->try_lock()) {
4685       // Record a sample. This is the critical section. The contents
4686       // of the _eden_chunk_array have to be non-decreasing in the
4687       // address order.
4688       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4689       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4690              "Unexpected state of Eden");
4691       if (_eden_chunk_index == 0 ||
4692           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4693            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4694                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4695         _eden_chunk_index++;  // commit sample
4696       }
4697       _eden_chunk_lock->unlock();
4698     }
4699   }
4700 }
4701 
4702 // Return a thread-local PLAB recording array, as appropriate.
4703 void* CMSCollector::get_data_recorder(int thr_num) {
4704   if (_survivor_plab_array != NULL &&
4705       (CMSPLABRecordAlways ||
4706        (_collectorState > Marking && _collectorState < FinalMarking))) {
4707     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4708     ChunkArray* ca = &_survivor_plab_array[thr_num];
4709     ca->reset();   // clear it so that fresh data is recorded
4710     return (void*) ca;
4711   } else {
4712     return NULL;
4713   }
4714 }
4715 
4716 // Reset all the thread-local PLAB recording arrays
4717 void CMSCollector::reset_survivor_plab_arrays() {
4718   for (uint i = 0; i < ParallelGCThreads; i++) {
4719     _survivor_plab_array[i].reset();
4720   }
4721 }
4722 
4723 // Merge the per-thread plab arrays into the global survivor chunk
4724 // array which will provide the partitioning of the survivor space
4725 // for CMS initial scan and rescan.
4726 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4727                                               int no_of_gc_threads) {
4728   assert(_survivor_plab_array  != NULL, "Error");
4729   assert(_survivor_chunk_array != NULL, "Error");
4730   assert(_collectorState == FinalMarking ||
4731          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4732   for (int j = 0; j < no_of_gc_threads; j++) {
4733     _cursor[j] = 0;
4734   }
4735   HeapWord* top = surv->top();
4736   size_t i;
4737   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4738     HeapWord* min_val = top;          // Higher than any PLAB address
4739     uint      min_tid = 0;            // position of min_val this round
4740     for (int j = 0; j < no_of_gc_threads; j++) {
4741       ChunkArray* cur_sca = &_survivor_plab_array[j];
4742       if (_cursor[j] == cur_sca->end()) {
4743         continue;
4744       }
4745       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4746       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4747       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4748       if (cur_val < min_val) {
4749         min_tid = j;
4750         min_val = cur_val;
4751       } else {
4752         assert(cur_val < top, "All recorded addresses should be less");
4753       }
4754     }
4755     // At this point min_val and min_tid are respectively
4756     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4757     // and the thread (j) that witnesses that address.
4758     // We record this address in the _survivor_chunk_array[i]
4759     // and increment _cursor[min_tid] prior to the next round i.
4760     if (min_val == top) {
4761       break;
4762     }
4763     _survivor_chunk_array[i] = min_val;
4764     _cursor[min_tid]++;
4765   }
4766   // We are all done; record the size of the _survivor_chunk_array
4767   _survivor_chunk_index = i; // exclusive: [0, i)
4768   log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4769   // Verify that we used up all the recorded entries
4770   #ifdef ASSERT
4771     size_t total = 0;
4772     for (int j = 0; j < no_of_gc_threads; j++) {
4773       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4774       total += _cursor[j];
4775     }
4776     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4777     // Check that the merged array is in sorted order
4778     if (total > 0) {
4779       for (size_t i = 0; i < total - 1; i++) {
4780         log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4781                                      i, p2i(_survivor_chunk_array[i]));
4782         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4783                "Not sorted");
4784       }
4785     }
4786   #endif // ASSERT
4787 }
4788 
4789 // Set up the space's par_seq_tasks structure for work claiming
4790 // for parallel initial scan and rescan of young gen.
4791 // See ParRescanTask where this is currently used.
4792 void
4793 CMSCollector::
4794 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4795   assert(n_threads > 0, "Unexpected n_threads argument");
4796 
4797   // Eden space
4798   if (!_young_gen->eden()->is_empty()) {
4799     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4800     assert(!pst->valid(), "Clobbering existing data?");
4801     // Each valid entry in [0, _eden_chunk_index) represents a task.
4802     size_t n_tasks = _eden_chunk_index + 1;
4803     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4804     // Sets the condition for completion of the subtask (how many threads
4805     // need to finish in order to be done).
4806     pst->set_n_threads(n_threads);
4807     pst->set_n_tasks((int)n_tasks);
4808   }
4809 
4810   // Merge the survivor plab arrays into _survivor_chunk_array
4811   if (_survivor_plab_array != NULL) {
4812     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4813   } else {
4814     assert(_survivor_chunk_index == 0, "Error");
4815   }
4816 
4817   // To space
4818   {
4819     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4820     assert(!pst->valid(), "Clobbering existing data?");
4821     // Sets the condition for completion of the subtask (how many threads
4822     // need to finish in order to be done).
4823     pst->set_n_threads(n_threads);
4824     pst->set_n_tasks(1);
4825     assert(pst->valid(), "Error");
4826   }
4827 
4828   // From space
4829   {
4830     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4831     assert(!pst->valid(), "Clobbering existing data?");
4832     size_t n_tasks = _survivor_chunk_index + 1;
4833     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4834     // Sets the condition for completion of the subtask (how many threads
4835     // need to finish in order to be done).
4836     pst->set_n_threads(n_threads);
4837     pst->set_n_tasks((int)n_tasks);
4838     assert(pst->valid(), "Error");
4839   }
4840 }
4841 
4842 // Parallel version of remark
4843 void CMSCollector::do_remark_parallel() {
4844   CMSHeap* heap = CMSHeap::heap();
4845   WorkGang* workers = heap->workers();
4846   assert(workers != NULL, "Need parallel worker threads.");
4847   // Choose to use the number of GC workers most recently set
4848   // into "active_workers".
4849   uint n_workers = workers->active_workers();
4850 
4851   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4852 
4853   StrongRootsScope srs(n_workers);
4854 
4855   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4856 
4857   // We won't be iterating over the cards in the card table updating
4858   // the younger_gen cards, so we shouldn't call the following else
4859   // the verification code as well as subsequent younger_refs_iterate
4860   // code would get confused. XXX
4861   // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4862 
4863   // The young gen rescan work will not be done as part of
4864   // process_roots (which currently doesn't know how to
4865   // parallelize such a scan), but rather will be broken up into
4866   // a set of parallel tasks (via the sampling that the [abortable]
4867   // preclean phase did of eden, plus the [two] tasks of
4868   // scanning the [two] survivor spaces. Further fine-grain
4869   // parallelization of the scanning of the survivor spaces
4870   // themselves, and of precleaning of the young gen itself
4871   // is deferred to the future.
4872   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4873 
4874   // The dirty card rescan work is broken up into a "sequence"
4875   // of parallel tasks (per constituent space) that are dynamically
4876   // claimed by the parallel threads.
4877   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4878 
4879   // It turns out that even when we're using 1 thread, doing the work in a
4880   // separate thread causes wide variance in run times.  We can't help this
4881   // in the multi-threaded case, but we special-case n=1 here to get
4882   // repeatable measurements of the 1-thread overhead of the parallel code.
4883   if (n_workers > 1) {
4884     // Make refs discovery MT-safe, if it isn't already: it may not
4885     // necessarily be so, since it's possible that we are doing
4886     // ST marking.
4887     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4888     workers->run_task(&tsk);
4889   } else {
4890     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4891     tsk.work(0);
4892   }
4893 
4894   // restore, single-threaded for now, any preserved marks
4895   // as a result of work_q overflow
4896   restore_preserved_marks_if_any();
4897 }
4898 
4899 // Non-parallel version of remark
4900 void CMSCollector::do_remark_non_parallel() {
4901   ResourceMark rm;
4902   HandleMark   hm;
4903   CMSHeap* heap = CMSHeap::heap();
4904   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4905 
4906   MarkRefsIntoAndScanClosure
4907     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4908              &_markStack, this,
4909              false /* should_yield */, false /* not precleaning */);
4910   MarkFromDirtyCardsClosure
4911     markFromDirtyCardsClosure(this, _span,
4912                               NULL,  // space is set further below
4913                               &_markBitMap, &_markStack, &mrias_cl);
4914   {
4915     GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4916     // Iterate over the dirty cards, setting the corresponding bits in the
4917     // mod union table.
4918     {
4919       ModUnionClosure modUnionClosure(&_modUnionTable);
4920       _ct->dirty_card_iterate(_cmsGen->used_region(),
4921                               &modUnionClosure);
4922     }
4923     // Having transferred these marks into the modUnionTable, we just need
4924     // to rescan the marked objects on the dirty cards in the modUnionTable.
4925     // The initial marking may have been done during an asynchronous
4926     // collection so there may be dirty bits in the mod-union table.
4927     const int alignment = CardTable::card_size * BitsPerWord;
4928     {
4929       // ... First handle dirty cards in CMS gen
4930       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4931       MemRegion ur = _cmsGen->used_region();
4932       HeapWord* lb = ur.start();
4933       HeapWord* ub = align_up(ur.end(), alignment);
4934       MemRegion cms_span(lb, ub);
4935       _modUnionTable.dirty_range_iterate_clear(cms_span,
4936                                                &markFromDirtyCardsClosure);
4937       verify_work_stacks_empty();
4938       log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4939     }
4940   }
4941   if (VerifyDuringGC &&
4942       CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4943     HandleMark hm;  // Discard invalid handles created during verification
4944     Universe::verify();
4945   }
4946   {
4947     GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4948 
4949     verify_work_stacks_empty();
4950 
4951     heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4952     StrongRootsScope srs(1);
4953 
4954     heap->cms_process_roots(&srs,
4955                             true,  // young gen as roots
4956                             GenCollectedHeap::ScanningOption(roots_scanning_options()),
4957                             should_unload_classes(),
4958                             &mrias_cl,
4959                             NULL); // The dirty klasses will be handled below
4960 
4961     assert(should_unload_classes()
4962            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4963            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4964   }
4965 
4966   {
4967     GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4968 
4969     verify_work_stacks_empty();
4970 
4971     // Scan all class loader data objects that might have been introduced
4972     // during concurrent marking.
4973     ResourceMark rm;
4974     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4975     for (int i = 0; i < array->length(); i++) {
4976       mrias_cl.do_cld_nv(array->at(i));
4977     }
4978 
4979     // We don't need to keep track of new CLDs anymore.
4980     ClassLoaderDataGraph::remember_new_clds(false);
4981 
4982     verify_work_stacks_empty();
4983   }
4984 
4985   // We might have added oops to ClassLoaderData::_handles during the
4986   // concurrent marking phase. These oops do not point to newly allocated objects
4987   // that are guaranteed to be kept alive.  Hence,
4988   // we do have to revisit the _handles block during the remark phase.
4989   {
4990     GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4991 
4992     verify_work_stacks_empty();
4993 
4994     RemarkCLDClosure remark_closure(&mrias_cl);
4995     ClassLoaderDataGraph::cld_do(&remark_closure);
4996 
4997     verify_work_stacks_empty();
4998   }
4999 
5000   verify_work_stacks_empty();
5001   // Restore evacuated mark words, if any, used for overflow list links
5002   restore_preserved_marks_if_any();
5003 
5004   verify_overflow_empty();
5005 }
5006 
5007 ////////////////////////////////////////////////////////
5008 // Parallel Reference Processing Task Proxy Class
5009 ////////////////////////////////////////////////////////
5010 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5011   OopTaskQueueSet*       _queues;
5012   ParallelTaskTerminator _terminator;
5013  public:
5014   AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5015     AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5016   ParallelTaskTerminator* terminator() { return &_terminator; }
5017   OopTaskQueueSet* queues() { return _queues; }
5018 };
5019 
5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5021   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5022   CMSCollector*          _collector;
5023   CMSBitMap*             _mark_bit_map;
5024   const MemRegion        _span;
5025   ProcessTask&           _task;
5026 
5027 public:
5028   CMSRefProcTaskProxy(ProcessTask&     task,
5029                       CMSCollector*    collector,
5030                       const MemRegion& span,
5031                       CMSBitMap*       mark_bit_map,
5032                       AbstractWorkGang* workers,
5033                       OopTaskQueueSet* task_queues):
5034     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5035       task_queues,
5036       workers->active_workers()),
5037     _task(task),
5038     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5039   {
5040     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5041            "Inconsistency in _span");
5042   }
5043 
5044   OopTaskQueueSet* task_queues() { return queues(); }
5045 
5046   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5047 
5048   void do_work_steal(int i,
5049                      CMSParDrainMarkingStackClosure* drain,
5050                      CMSParKeepAliveClosure* keep_alive,
5051                      int* seed);
5052 
5053   virtual void work(uint worker_id);
5054 };
5055 
5056 void CMSRefProcTaskProxy::work(uint worker_id) {
5057   ResourceMark rm;
5058   HandleMark hm;
5059   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5060   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5061                                         _mark_bit_map,
5062                                         work_queue(worker_id));
5063   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5064                                                  _mark_bit_map,
5065                                                  work_queue(worker_id));
5066   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5067   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5068   if (_task.marks_oops_alive()) {
5069     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5070                   _collector->hash_seed(worker_id));
5071   }
5072   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5073   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5074 }
5075 
5076 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5077   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5078   EnqueueTask& _task;
5079 
5080 public:
5081   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5082     : AbstractGangTask("Enqueue reference objects in parallel"),
5083       _task(task)
5084   { }
5085 
5086   virtual void work(uint worker_id)
5087   {
5088     _task.work(worker_id);
5089   }
5090 };
5091 
5092 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5093   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5094    _span(span),
5095    _bit_map(bit_map),
5096    _work_queue(work_queue),
5097    _mark_and_push(collector, span, bit_map, work_queue),
5098    _low_water_mark(MIN2((work_queue->max_elems()/4),
5099                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5100 { }
5101 
5102 // . see if we can share work_queues with ParNew? XXX
5103 void CMSRefProcTaskProxy::do_work_steal(int i,
5104   CMSParDrainMarkingStackClosure* drain,
5105   CMSParKeepAliveClosure* keep_alive,
5106   int* seed) {
5107   OopTaskQueue* work_q = work_queue(i);
5108   NOT_PRODUCT(int num_steals = 0;)
5109   oop obj_to_scan;
5110 
5111   while (true) {
5112     // Completely finish any left over work from (an) earlier round(s)
5113     drain->trim_queue(0);
5114     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5115                                          (size_t)ParGCDesiredObjsFromOverflowList);
5116     // Now check if there's any work in the overflow list
5117     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5118     // only affects the number of attempts made to get work from the
5119     // overflow list and does not affect the number of workers.  Just
5120     // pass ParallelGCThreads so this behavior is unchanged.
5121     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5122                                                 work_q,
5123                                                 ParallelGCThreads)) {
5124       // Found something in global overflow list;
5125       // not yet ready to go stealing work from others.
5126       // We'd like to assert(work_q->size() != 0, ...)
5127       // because we just took work from the overflow list,
5128       // but of course we can't, since all of that might have
5129       // been already stolen from us.
5130       continue;
5131     }
5132     // Verify that we have no work before we resort to stealing
5133     assert(work_q->size() == 0, "Have work, shouldn't steal");
5134     // Try to steal from other queues that have work
5135     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5136       NOT_PRODUCT(num_steals++;)
5137       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5138       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5139       // Do scanning work
5140       obj_to_scan->oop_iterate(keep_alive);
5141       // Loop around, finish this work, and try to steal some more
5142     } else if (terminator()->offer_termination()) {
5143       break;  // nirvana from the infinite cycle
5144     }
5145   }
5146   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5147 }
5148 
5149 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5150 {
5151   CMSHeap* heap = CMSHeap::heap();
5152   WorkGang* workers = heap->workers();
5153   assert(workers != NULL, "Need parallel worker threads.");
5154   CMSRefProcTaskProxy rp_task(task, &_collector,
5155                               _collector.ref_processor()->span(),
5156                               _collector.markBitMap(),
5157                               workers, _collector.task_queues());
5158   workers->run_task(&rp_task);
5159 }
5160 
5161 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5162 {
5163 
5164   CMSHeap* heap = CMSHeap::heap();
5165   WorkGang* workers = heap->workers();
5166   assert(workers != NULL, "Need parallel worker threads.");
5167   CMSRefEnqueueTaskProxy enq_task(task);
5168   workers->run_task(&enq_task);
5169 }
5170 
5171 void CMSCollector::refProcessingWork() {
5172   ResourceMark rm;
5173   HandleMark   hm;
5174 
5175   SpanReferenceProcessor* rp = ref_processor();
5176   assert(rp->span().equals(_span), "Spans should be equal");
5177   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5178   // Process weak references.
5179   rp->setup_policy(false);
5180   verify_work_stacks_empty();
5181 
5182   ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5183   {
5184     GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5185 
5186     // Setup keep_alive and complete closures.
5187     CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5188                                             &_markStack, false /* !preclean */);
5189     CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5190                                   _span, &_markBitMap, &_markStack,
5191                                   &cmsKeepAliveClosure, false /* !preclean */);
5192 
5193     ReferenceProcessorStats stats;
5194     if (rp->processing_is_mt()) {
5195       // Set the degree of MT here.  If the discovery is done MT, there
5196       // may have been a different number of threads doing the discovery
5197       // and a different number of discovered lists may have Ref objects.
5198       // That is OK as long as the Reference lists are balanced (see
5199       // balance_all_queues() and balance_queues()).
5200       CMSHeap* heap = CMSHeap::heap();
5201       uint active_workers = ParallelGCThreads;
5202       WorkGang* workers = heap->workers();
5203       if (workers != NULL) {
5204         active_workers = workers->active_workers();
5205         // The expectation is that active_workers will have already
5206         // been set to a reasonable value.  If it has not been set,
5207         // investigate.
5208         assert(active_workers > 0, "Should have been set during scavenge");
5209       }
5210       rp->set_active_mt_degree(active_workers);
5211       CMSRefProcTaskExecutor task_executor(*this);
5212       stats = rp->process_discovered_references(&_is_alive_closure,
5213                                         &cmsKeepAliveClosure,
5214                                         &cmsDrainMarkingStackClosure,
5215                                         &task_executor,
5216                                         &pt);
5217     } else {
5218       stats = rp->process_discovered_references(&_is_alive_closure,
5219                                         &cmsKeepAliveClosure,
5220                                         &cmsDrainMarkingStackClosure,
5221                                         NULL,
5222                                         &pt);
5223     }
5224     _gc_tracer_cm->report_gc_reference_stats(stats);
5225     pt.print_all_references();
5226   }
5227 
5228   // This is the point where the entire marking should have completed.
5229   verify_work_stacks_empty();
5230 
5231   {
5232     GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5233     WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5234   }
5235 
5236   if (should_unload_classes()) {
5237     {
5238       GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5239 
5240       // Unload classes and purge the SystemDictionary.
5241       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5242 
5243       // Unload nmethods.
5244       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5245 
5246       // Prune dead klasses from subklass/sibling/implementor lists.
5247       Klass::clean_weak_klass_links();
5248     }
5249 
5250     {
5251       GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5252       // Clean up unreferenced symbols in symbol table.
5253       SymbolTable::unlink();
5254     }
5255 
5256     {
5257       GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5258       // Delete entries for dead interned strings.
5259       StringTable::unlink(&_is_alive_closure);
5260     }
5261   }
5262 
5263   // Restore any preserved marks as a result of mark stack or
5264   // work queue overflow
5265   restore_preserved_marks_if_any();  // done single-threaded for now
5266 
5267   rp->set_enqueuing_is_done(true);
5268   if (rp->processing_is_mt()) {
5269     rp->balance_all_queues();
5270     CMSRefProcTaskExecutor task_executor(*this);
5271     rp->enqueue_discovered_references(&task_executor, &pt);
5272   } else {
5273     rp->enqueue_discovered_references(NULL, &pt);
5274   }
5275   rp->verify_no_references_recorded();
5276   pt.print_enqueue_phase();
5277   assert(!rp->discovery_enabled(), "should have been disabled");
5278 }
5279 
5280 #ifndef PRODUCT
5281 void CMSCollector::check_correct_thread_executing() {
5282   Thread* t = Thread::current();
5283   // Only the VM thread or the CMS thread should be here.
5284   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5285          "Unexpected thread type");
5286   // If this is the vm thread, the foreground process
5287   // should not be waiting.  Note that _foregroundGCIsActive is
5288   // true while the foreground collector is waiting.
5289   if (_foregroundGCShouldWait) {
5290     // We cannot be the VM thread
5291     assert(t->is_ConcurrentGC_thread(),
5292            "Should be CMS thread");
5293   } else {
5294     // We can be the CMS thread only if we are in a stop-world
5295     // phase of CMS collection.
5296     if (t->is_ConcurrentGC_thread()) {
5297       assert(_collectorState == InitialMarking ||
5298              _collectorState == FinalMarking,
5299              "Should be a stop-world phase");
5300       // The CMS thread should be holding the CMS_token.
5301       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5302              "Potential interference with concurrently "
5303              "executing VM thread");
5304     }
5305   }
5306 }
5307 #endif
5308 
5309 void CMSCollector::sweep() {
5310   assert(_collectorState == Sweeping, "just checking");
5311   check_correct_thread_executing();
5312   verify_work_stacks_empty();
5313   verify_overflow_empty();
5314   increment_sweep_count();
5315   TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5316 
5317   _inter_sweep_timer.stop();
5318   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5319 
5320   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5321   _intra_sweep_timer.reset();
5322   _intra_sweep_timer.start();
5323   {
5324     GCTraceCPUTime tcpu;
5325     CMSPhaseAccounting pa(this, "Concurrent Sweep");
5326     // First sweep the old gen
5327     {
5328       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5329                                bitMapLock());
5330       sweepWork(_cmsGen);
5331     }
5332 
5333     // Update Universe::_heap_*_at_gc figures.
5334     // We need all the free list locks to make the abstract state
5335     // transition from Sweeping to Resetting. See detailed note
5336     // further below.
5337     {
5338       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5339       // Update heap occupancy information which is used as
5340       // input to soft ref clearing policy at the next gc.
5341       Universe::update_heap_info_at_gc();
5342       _collectorState = Resizing;
5343     }
5344   }
5345   verify_work_stacks_empty();
5346   verify_overflow_empty();
5347 
5348   if (should_unload_classes()) {
5349     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5350     // requires that the virtual spaces are stable and not deleted.
5351     ClassLoaderDataGraph::set_should_purge(true);
5352   }
5353 
5354   _intra_sweep_timer.stop();
5355   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5356 
5357   _inter_sweep_timer.reset();
5358   _inter_sweep_timer.start();
5359 
5360   // We need to use a monotonically non-decreasing time in ms
5361   // or we will see time-warp warnings and os::javaTimeMillis()
5362   // does not guarantee monotonicity.
5363   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5364   update_time_of_last_gc(now);
5365 
5366   // NOTE on abstract state transitions:
5367   // Mutators allocate-live and/or mark the mod-union table dirty
5368   // based on the state of the collection.  The former is done in
5369   // the interval [Marking, Sweeping] and the latter in the interval
5370   // [Marking, Sweeping).  Thus the transitions into the Marking state
5371   // and out of the Sweeping state must be synchronously visible
5372   // globally to the mutators.
5373   // The transition into the Marking state happens with the world
5374   // stopped so the mutators will globally see it.  Sweeping is
5375   // done asynchronously by the background collector so the transition
5376   // from the Sweeping state to the Resizing state must be done
5377   // under the freelistLock (as is the check for whether to
5378   // allocate-live and whether to dirty the mod-union table).
5379   assert(_collectorState == Resizing, "Change of collector state to"
5380     " Resizing must be done under the freelistLocks (plural)");
5381 
5382   // Now that sweeping has been completed, we clear
5383   // the incremental_collection_failed flag,
5384   // thus inviting a younger gen collection to promote into
5385   // this generation. If such a promotion may still fail,
5386   // the flag will be set again when a young collection is
5387   // attempted.
5388   CMSHeap* heap = CMSHeap::heap();
5389   heap->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5390   heap->update_full_collections_completed(_collection_count_start);
5391 }
5392 
5393 // FIX ME!!! Looks like this belongs in CFLSpace, with
5394 // CMSGen merely delegating to it.
5395 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5396   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5397   HeapWord*  minAddr        = _cmsSpace->bottom();
5398   HeapWord*  largestAddr    =
5399     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5400   if (largestAddr == NULL) {
5401     // The dictionary appears to be empty.  In this case
5402     // try to coalesce at the end of the heap.
5403     largestAddr = _cmsSpace->end();
5404   }
5405   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5406   size_t nearLargestOffset =
5407     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5408   log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5409                           p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5410   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5411 }
5412 
5413 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5414   return addr >= _cmsSpace->nearLargestChunk();
5415 }
5416 
5417 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5418   return _cmsSpace->find_chunk_at_end();
5419 }
5420 
5421 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5422                                                     bool full) {
5423   // If the young generation has been collected, gather any statistics
5424   // that are of interest at this point.
5425   bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5426   if (!full && current_is_young) {
5427     // Gather statistics on the young generation collection.
5428     collector()->stats().record_gc0_end(used());
5429   }
5430 }
5431 
5432 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5433   // We iterate over the space(s) underlying this generation,
5434   // checking the mark bit map to see if the bits corresponding
5435   // to specific blocks are marked or not. Blocks that are
5436   // marked are live and are not swept up. All remaining blocks
5437   // are swept up, with coalescing on-the-fly as we sweep up
5438   // contiguous free and/or garbage blocks:
5439   // We need to ensure that the sweeper synchronizes with allocators
5440   // and stop-the-world collectors. In particular, the following
5441   // locks are used:
5442   // . CMS token: if this is held, a stop the world collection cannot occur
5443   // . freelistLock: if this is held no allocation can occur from this
5444   //                 generation by another thread
5445   // . bitMapLock: if this is held, no other thread can access or update
5446   //
5447 
5448   // Note that we need to hold the freelistLock if we use
5449   // block iterate below; else the iterator might go awry if
5450   // a mutator (or promotion) causes block contents to change
5451   // (for instance if the allocator divvies up a block).
5452   // If we hold the free list lock, for all practical purposes
5453   // young generation GC's can't occur (they'll usually need to
5454   // promote), so we might as well prevent all young generation
5455   // GC's while we do a sweeping step. For the same reason, we might
5456   // as well take the bit map lock for the entire duration
5457 
5458   // check that we hold the requisite locks
5459   assert(have_cms_token(), "Should hold cms token");
5460   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5461   assert_lock_strong(old_gen->freelistLock());
5462   assert_lock_strong(bitMapLock());
5463 
5464   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5465   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5466   old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5467                                           _inter_sweep_estimate.padded_average(),
5468                                           _intra_sweep_estimate.padded_average());
5469   old_gen->setNearLargestChunk();
5470 
5471   {
5472     SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5473     old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5474     // We need to free-up/coalesce garbage/blocks from a
5475     // co-terminal free run. This is done in the SweepClosure
5476     // destructor; so, do not remove this scope, else the
5477     // end-of-sweep-census below will be off by a little bit.
5478   }
5479   old_gen->cmsSpace()->sweep_completed();
5480   old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5481   if (should_unload_classes()) {                // unloaded classes this cycle,
5482     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5483   } else {                                      // did not unload classes,
5484     _concurrent_cycles_since_last_unload++;     // ... increment count
5485   }
5486 }
5487 
5488 // Reset CMS data structures (for now just the marking bit map)
5489 // preparatory for the next cycle.
5490 void CMSCollector::reset_concurrent() {
5491   CMSTokenSyncWithLocks ts(true, bitMapLock());
5492 
5493   // If the state is not "Resetting", the foreground  thread
5494   // has done a collection and the resetting.
5495   if (_collectorState != Resetting) {
5496     assert(_collectorState == Idling, "The state should only change"
5497       " because the foreground collector has finished the collection");
5498     return;
5499   }
5500 
5501   {
5502     // Clear the mark bitmap (no grey objects to start with)
5503     // for the next cycle.
5504     GCTraceCPUTime tcpu;
5505     CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5506 
5507     HeapWord* curAddr = _markBitMap.startWord();
5508     while (curAddr < _markBitMap.endWord()) {
5509       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5510       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5511       _markBitMap.clear_large_range(chunk);
5512       if (ConcurrentMarkSweepThread::should_yield() &&
5513           !foregroundGCIsActive() &&
5514           CMSYield) {
5515         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5516                "CMS thread should hold CMS token");
5517         assert_lock_strong(bitMapLock());
5518         bitMapLock()->unlock();
5519         ConcurrentMarkSweepThread::desynchronize(true);
5520         stopTimer();
5521         incrementYields();
5522 
5523         // See the comment in coordinator_yield()
5524         for (unsigned i = 0; i < CMSYieldSleepCount &&
5525                          ConcurrentMarkSweepThread::should_yield() &&
5526                          !CMSCollector::foregroundGCIsActive(); ++i) {
5527           os::sleep(Thread::current(), 1, false);
5528         }
5529 
5530         ConcurrentMarkSweepThread::synchronize(true);
5531         bitMapLock()->lock_without_safepoint_check();
5532         startTimer();
5533       }
5534       curAddr = chunk.end();
5535     }
5536     // A successful mostly concurrent collection has been done.
5537     // Because only the full (i.e., concurrent mode failure) collections
5538     // are being measured for gc overhead limits, clean the "near" flag
5539     // and count.
5540     size_policy()->reset_gc_overhead_limit_count();
5541     _collectorState = Idling;
5542   }
5543 
5544   register_gc_end();
5545 }
5546 
5547 // Same as above but for STW paths
5548 void CMSCollector::reset_stw() {
5549   // already have the lock
5550   assert(_collectorState == Resetting, "just checking");
5551   assert_lock_strong(bitMapLock());
5552   GCIdMark gc_id_mark(_cmsThread->gc_id());
5553   _markBitMap.clear_all();
5554   _collectorState = Idling;
5555   register_gc_end();
5556 }
5557 
5558 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5559   GCTraceCPUTime tcpu;
5560   TraceCollectorStats tcs_cgc(cgc_counters());
5561 
5562   switch (op) {
5563     case CMS_op_checkpointRootsInitial: {
5564       GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5565       SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5566       checkpointRootsInitial();
5567       break;
5568     }
5569     case CMS_op_checkpointRootsFinal: {
5570       GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5571       SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5572       checkpointRootsFinal();
5573       break;
5574     }
5575     default:
5576       fatal("No such CMS_op");
5577   }
5578 }
5579 
5580 #ifndef PRODUCT
5581 size_t const CMSCollector::skip_header_HeapWords() {
5582   return FreeChunk::header_size();
5583 }
5584 
5585 // Try and collect here conditions that should hold when
5586 // CMS thread is exiting. The idea is that the foreground GC
5587 // thread should not be blocked if it wants to terminate
5588 // the CMS thread and yet continue to run the VM for a while
5589 // after that.
5590 void CMSCollector::verify_ok_to_terminate() const {
5591   assert(Thread::current()->is_ConcurrentGC_thread(),
5592          "should be called by CMS thread");
5593   assert(!_foregroundGCShouldWait, "should be false");
5594   // We could check here that all the various low-level locks
5595   // are not held by the CMS thread, but that is overkill; see
5596   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5597   // is checked.
5598 }
5599 #endif
5600 
5601 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5602    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5603           "missing Printezis mark?");
5604   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5605   size_t size = pointer_delta(nextOneAddr + 1, addr);
5606   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5607          "alignment problem");
5608   assert(size >= 3, "Necessary for Printezis marks to work");
5609   return size;
5610 }
5611 
5612 // A variant of the above (block_size_using_printezis_bits()) except
5613 // that we return 0 if the P-bits are not yet set.
5614 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5615   if (_markBitMap.isMarked(addr + 1)) {
5616     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5617     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5618     size_t size = pointer_delta(nextOneAddr + 1, addr);
5619     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5620            "alignment problem");
5621     assert(size >= 3, "Necessary for Printezis marks to work");
5622     return size;
5623   }
5624   return 0;
5625 }
5626 
5627 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5628   size_t sz = 0;
5629   oop p = (oop)addr;
5630   if (p->klass_or_null_acquire() != NULL) {
5631     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5632   } else {
5633     sz = block_size_using_printezis_bits(addr);
5634   }
5635   assert(sz > 0, "size must be nonzero");
5636   HeapWord* next_block = addr + sz;
5637   HeapWord* next_card  = align_up(next_block, CardTable::card_size);
5638   assert(align_down((uintptr_t)addr,      CardTable::card_size) <
5639          align_down((uintptr_t)next_card, CardTable::card_size),
5640          "must be different cards");
5641   return next_card;
5642 }
5643 
5644 
5645 // CMS Bit Map Wrapper /////////////////////////////////////////
5646 
5647 // Construct a CMS bit map infrastructure, but don't create the
5648 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5649 // further below.
5650 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5651   _bm(),
5652   _shifter(shifter),
5653   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5654                                     Monitor::_safepoint_check_sometimes) : NULL)
5655 {
5656   _bmStartWord = 0;
5657   _bmWordSize  = 0;
5658 }
5659 
5660 bool CMSBitMap::allocate(MemRegion mr) {
5661   _bmStartWord = mr.start();
5662   _bmWordSize  = mr.word_size();
5663   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5664                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5665   if (!brs.is_reserved()) {
5666     log_warning(gc)("CMS bit map allocation failure");
5667     return false;
5668   }
5669   // For now we'll just commit all of the bit map up front.
5670   // Later on we'll try to be more parsimonious with swap.
5671   if (!_virtual_space.initialize(brs, brs.size())) {
5672     log_warning(gc)("CMS bit map backing store failure");
5673     return false;
5674   }
5675   assert(_virtual_space.committed_size() == brs.size(),
5676          "didn't reserve backing store for all of CMS bit map?");
5677   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5678          _bmWordSize, "inconsistency in bit map sizing");
5679   _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5680 
5681   // bm.clear(); // can we rely on getting zero'd memory? verify below
5682   assert(isAllClear(),
5683          "Expected zero'd memory from ReservedSpace constructor");
5684   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5685          "consistency check");
5686   return true;
5687 }
5688 
5689 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5690   HeapWord *next_addr, *end_addr, *last_addr;
5691   assert_locked();
5692   assert(covers(mr), "out-of-range error");
5693   // XXX assert that start and end are appropriately aligned
5694   for (next_addr = mr.start(), end_addr = mr.end();
5695        next_addr < end_addr; next_addr = last_addr) {
5696     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5697     last_addr = dirty_region.end();
5698     if (!dirty_region.is_empty()) {
5699       cl->do_MemRegion(dirty_region);
5700     } else {
5701       assert(last_addr == end_addr, "program logic");
5702       return;
5703     }
5704   }
5705 }
5706 
5707 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5708   _bm.print_on_error(st, prefix);
5709 }
5710 
5711 #ifndef PRODUCT
5712 void CMSBitMap::assert_locked() const {
5713   CMSLockVerifier::assert_locked(lock());
5714 }
5715 
5716 bool CMSBitMap::covers(MemRegion mr) const {
5717   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5718   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5719          "size inconsistency");
5720   return (mr.start() >= _bmStartWord) &&
5721          (mr.end()   <= endWord());
5722 }
5723 
5724 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5725     return (start >= _bmStartWord && (start + size) <= endWord());
5726 }
5727 
5728 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5729   // verify that there are no 1 bits in the interval [left, right)
5730   FalseBitMapClosure falseBitMapClosure;
5731   iterate(&falseBitMapClosure, left, right);
5732 }
5733 
5734 void CMSBitMap::region_invariant(MemRegion mr)
5735 {
5736   assert_locked();
5737   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5738   assert(!mr.is_empty(), "unexpected empty region");
5739   assert(covers(mr), "mr should be covered by bit map");
5740   // convert address range into offset range
5741   size_t start_ofs = heapWordToOffset(mr.start());
5742   // Make sure that end() is appropriately aligned
5743   assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5744          "Misaligned mr.end()");
5745   size_t end_ofs   = heapWordToOffset(mr.end());
5746   assert(end_ofs > start_ofs, "Should mark at least one bit");
5747 }
5748 
5749 #endif
5750 
5751 bool CMSMarkStack::allocate(size_t size) {
5752   // allocate a stack of the requisite depth
5753   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5754                    size * sizeof(oop)));
5755   if (!rs.is_reserved()) {
5756     log_warning(gc)("CMSMarkStack allocation failure");
5757     return false;
5758   }
5759   if (!_virtual_space.initialize(rs, rs.size())) {
5760     log_warning(gc)("CMSMarkStack backing store failure");
5761     return false;
5762   }
5763   assert(_virtual_space.committed_size() == rs.size(),
5764          "didn't reserve backing store for all of CMS stack?");
5765   _base = (oop*)(_virtual_space.low());
5766   _index = 0;
5767   _capacity = size;
5768   NOT_PRODUCT(_max_depth = 0);
5769   return true;
5770 }
5771 
5772 // XXX FIX ME !!! In the MT case we come in here holding a
5773 // leaf lock. For printing we need to take a further lock
5774 // which has lower rank. We need to recalibrate the two
5775 // lock-ranks involved in order to be able to print the
5776 // messages below. (Or defer the printing to the caller.
5777 // For now we take the expedient path of just disabling the
5778 // messages for the problematic case.)
5779 void CMSMarkStack::expand() {
5780   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5781   if (_capacity == MarkStackSizeMax) {
5782     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5783       // We print a warning message only once per CMS cycle.
5784       log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5785     }
5786     return;
5787   }
5788   // Double capacity if possible
5789   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5790   // Do not give up existing stack until we have managed to
5791   // get the double capacity that we desired.
5792   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5793                    new_capacity * sizeof(oop)));
5794   if (rs.is_reserved()) {
5795     // Release the backing store associated with old stack
5796     _virtual_space.release();
5797     // Reinitialize virtual space for new stack
5798     if (!_virtual_space.initialize(rs, rs.size())) {
5799       fatal("Not enough swap for expanded marking stack");
5800     }
5801     _base = (oop*)(_virtual_space.low());
5802     _index = 0;
5803     _capacity = new_capacity;
5804   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5805     // Failed to double capacity, continue;
5806     // we print a detail message only once per CMS cycle.
5807     log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5808                         _capacity / K, new_capacity / K);
5809   }
5810 }
5811 
5812 
5813 // Closures
5814 // XXX: there seems to be a lot of code  duplication here;
5815 // should refactor and consolidate common code.
5816 
5817 // This closure is used to mark refs into the CMS generation in
5818 // the CMS bit map. Called at the first checkpoint. This closure
5819 // assumes that we do not need to re-mark dirty cards; if the CMS
5820 // generation on which this is used is not an oldest
5821 // generation then this will lose younger_gen cards!
5822 
5823 MarkRefsIntoClosure::MarkRefsIntoClosure(
5824   MemRegion span, CMSBitMap* bitMap):
5825     _span(span),
5826     _bitMap(bitMap)
5827 {
5828   assert(ref_discoverer() == NULL, "deliberately left NULL");
5829   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5830 }
5831 
5832 void MarkRefsIntoClosure::do_oop(oop obj) {
5833   // if p points into _span, then mark corresponding bit in _markBitMap
5834   assert(oopDesc::is_oop(obj), "expected an oop");
5835   HeapWord* addr = (HeapWord*)obj;
5836   if (_span.contains(addr)) {
5837     // this should be made more efficient
5838     _bitMap->mark(addr);
5839   }
5840 }
5841 
5842 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
5843 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5844 
5845 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5846   MemRegion span, CMSBitMap* bitMap):
5847     _span(span),
5848     _bitMap(bitMap)
5849 {
5850   assert(ref_discoverer() == NULL, "deliberately left NULL");
5851   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5852 }
5853 
5854 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5855   // if p points into _span, then mark corresponding bit in _markBitMap
5856   assert(oopDesc::is_oop(obj), "expected an oop");
5857   HeapWord* addr = (HeapWord*)obj;
5858   if (_span.contains(addr)) {
5859     // this should be made more efficient
5860     _bitMap->par_mark(addr);
5861   }
5862 }
5863 
5864 void ParMarkRefsIntoClosure::do_oop(oop* p)       { ParMarkRefsIntoClosure::do_oop_work(p); }
5865 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5866 
5867 // A variant of the above, used for CMS marking verification.
5868 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5869   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5870     _span(span),
5871     _verification_bm(verification_bm),
5872     _cms_bm(cms_bm)
5873 {
5874   assert(ref_discoverer() == NULL, "deliberately left NULL");
5875   assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5876 }
5877 
5878 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5879   // if p points into _span, then mark corresponding bit in _markBitMap
5880   assert(oopDesc::is_oop(obj), "expected an oop");
5881   HeapWord* addr = (HeapWord*)obj;
5882   if (_span.contains(addr)) {
5883     _verification_bm->mark(addr);
5884     if (!_cms_bm->isMarked(addr)) {
5885       Log(gc, verify) log;
5886       ResourceMark rm;
5887       LogStream ls(log.error());
5888       oop(addr)->print_on(&ls);
5889       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5890       fatal("... aborting");
5891     }
5892   }
5893 }
5894 
5895 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5896 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5897 
5898 //////////////////////////////////////////////////
5899 // MarkRefsIntoAndScanClosure
5900 //////////////////////////////////////////////////
5901 
5902 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5903                                                        ReferenceDiscoverer* rd,
5904                                                        CMSBitMap* bit_map,
5905                                                        CMSBitMap* mod_union_table,
5906                                                        CMSMarkStack*  mark_stack,
5907                                                        CMSCollector* collector,
5908                                                        bool should_yield,
5909                                                        bool concurrent_precleaning):
5910   _collector(collector),
5911   _span(span),
5912   _bit_map(bit_map),
5913   _mark_stack(mark_stack),
5914   _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table,
5915                       mark_stack, concurrent_precleaning),
5916   _yield(should_yield),
5917   _concurrent_precleaning(concurrent_precleaning),
5918   _freelistLock(NULL)
5919 {
5920   // FIXME: Should initialize in base class constructor.
5921   assert(rd != NULL, "ref_discoverer shouldn't be NULL");
5922   set_ref_discoverer_internal(rd);
5923 }
5924 
5925 // This closure is used to mark refs into the CMS generation at the
5926 // second (final) checkpoint, and to scan and transitively follow
5927 // the unmarked oops. It is also used during the concurrent precleaning
5928 // phase while scanning objects on dirty cards in the CMS generation.
5929 // The marks are made in the marking bit map and the marking stack is
5930 // used for keeping the (newly) grey objects during the scan.
5931 // The parallel version (Par_...) appears further below.
5932 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5933   if (obj != NULL) {
5934     assert(oopDesc::is_oop(obj), "expected an oop");
5935     HeapWord* addr = (HeapWord*)obj;
5936     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5937     assert(_collector->overflow_list_is_empty(),
5938            "overflow list should be empty");
5939     if (_span.contains(addr) &&
5940         !_bit_map->isMarked(addr)) {
5941       // mark bit map (object is now grey)
5942       _bit_map->mark(addr);
5943       // push on marking stack (stack should be empty), and drain the
5944       // stack by applying this closure to the oops in the oops popped
5945       // from the stack (i.e. blacken the grey objects)
5946       bool res = _mark_stack->push(obj);
5947       assert(res, "Should have space to push on empty stack");
5948       do {
5949         oop new_oop = _mark_stack->pop();
5950         assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5951         assert(_bit_map->isMarked((HeapWord*)new_oop),
5952                "only grey objects on this stack");
5953         // iterate over the oops in this oop, marking and pushing
5954         // the ones in CMS heap (i.e. in _span).
5955         new_oop->oop_iterate(&_pushAndMarkClosure);
5956         // check if it's time to yield
5957         do_yield_check();
5958       } while (!_mark_stack->isEmpty() ||
5959                (!_concurrent_precleaning && take_from_overflow_list()));
5960         // if marking stack is empty, and we are not doing this
5961         // during precleaning, then check the overflow list
5962     }
5963     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5964     assert(_collector->overflow_list_is_empty(),
5965            "overflow list was drained above");
5966 
5967     assert(_collector->no_preserved_marks(),
5968            "All preserved marks should have been restored above");
5969   }
5970 }
5971 
5972 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5973 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5974 
5975 void MarkRefsIntoAndScanClosure::do_yield_work() {
5976   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5977          "CMS thread should hold CMS token");
5978   assert_lock_strong(_freelistLock);
5979   assert_lock_strong(_bit_map->lock());
5980   // relinquish the free_list_lock and bitMaplock()
5981   _bit_map->lock()->unlock();
5982   _freelistLock->unlock();
5983   ConcurrentMarkSweepThread::desynchronize(true);
5984   _collector->stopTimer();
5985   _collector->incrementYields();
5986 
5987   // See the comment in coordinator_yield()
5988   for (unsigned i = 0;
5989        i < CMSYieldSleepCount &&
5990        ConcurrentMarkSweepThread::should_yield() &&
5991        !CMSCollector::foregroundGCIsActive();
5992        ++i) {
5993     os::sleep(Thread::current(), 1, false);
5994   }
5995 
5996   ConcurrentMarkSweepThread::synchronize(true);
5997   _freelistLock->lock_without_safepoint_check();
5998   _bit_map->lock()->lock_without_safepoint_check();
5999   _collector->startTimer();
6000 }
6001 
6002 ///////////////////////////////////////////////////////////
6003 // ParMarkRefsIntoAndScanClosure: a parallel version of
6004 //                                MarkRefsIntoAndScanClosure
6005 ///////////////////////////////////////////////////////////
6006 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6007   CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd,
6008   CMSBitMap* bit_map, OopTaskQueue* work_queue):
6009   _span(span),
6010   _bit_map(bit_map),
6011   _work_queue(work_queue),
6012   _low_water_mark(MIN2((work_queue->max_elems()/4),
6013                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6014   _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue)
6015 {
6016   // FIXME: Should initialize in base class constructor.
6017   assert(rd != NULL, "ref_discoverer shouldn't be NULL");
6018   set_ref_discoverer_internal(rd);
6019 }
6020 
6021 // This closure is used to mark refs into the CMS generation at the
6022 // second (final) checkpoint, and to scan and transitively follow
6023 // the unmarked oops. The marks are made in the marking bit map and
6024 // the work_queue is used for keeping the (newly) grey objects during
6025 // the scan phase whence they are also available for stealing by parallel
6026 // threads. Since the marking bit map is shared, updates are
6027 // synchronized (via CAS).
6028 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6029   if (obj != NULL) {
6030     // Ignore mark word because this could be an already marked oop
6031     // that may be chained at the end of the overflow list.
6032     assert(oopDesc::is_oop(obj, true), "expected an oop");
6033     HeapWord* addr = (HeapWord*)obj;
6034     if (_span.contains(addr) &&
6035         !_bit_map->isMarked(addr)) {
6036       // mark bit map (object will become grey):
6037       // It is possible for several threads to be
6038       // trying to "claim" this object concurrently;
6039       // the unique thread that succeeds in marking the
6040       // object first will do the subsequent push on
6041       // to the work queue (or overflow list).
6042       if (_bit_map->par_mark(addr)) {
6043         // push on work_queue (which may not be empty), and trim the
6044         // queue to an appropriate length by applying this closure to
6045         // the oops in the oops popped from the stack (i.e. blacken the
6046         // grey objects)
6047         bool res = _work_queue->push(obj);
6048         assert(res, "Low water mark should be less than capacity?");
6049         trim_queue(_low_water_mark);
6050       } // Else, another thread claimed the object
6051     }
6052   }
6053 }
6054 
6055 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p)       { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6056 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6057 
6058 // This closure is used to rescan the marked objects on the dirty cards
6059 // in the mod union table and the card table proper.
6060 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6061   oop p, MemRegion mr) {
6062 
6063   size_t size = 0;
6064   HeapWord* addr = (HeapWord*)p;
6065   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6066   assert(_span.contains(addr), "we are scanning the CMS generation");
6067   // check if it's time to yield
6068   if (do_yield_check()) {
6069     // We yielded for some foreground stop-world work,
6070     // and we have been asked to abort this ongoing preclean cycle.
6071     return 0;
6072   }
6073   if (_bitMap->isMarked(addr)) {
6074     // it's marked; is it potentially uninitialized?
6075     if (p->klass_or_null_acquire() != NULL) {
6076         // an initialized object; ignore mark word in verification below
6077         // since we are running concurrent with mutators
6078         assert(oopDesc::is_oop(p, true), "should be an oop");
6079         if (p->is_objArray()) {
6080           // objArrays are precisely marked; restrict scanning
6081           // to dirty cards only.
6082           size = CompactibleFreeListSpace::adjustObjectSize(
6083                    p->oop_iterate_size(_scanningClosure, mr));
6084         } else {
6085           // A non-array may have been imprecisely marked; we need
6086           // to scan object in its entirety.
6087           size = CompactibleFreeListSpace::adjustObjectSize(
6088                    p->oop_iterate_size(_scanningClosure));
6089         }
6090       #ifdef ASSERT
6091         size_t direct_size =
6092           CompactibleFreeListSpace::adjustObjectSize(p->size());
6093         assert(size == direct_size, "Inconsistency in size");
6094         assert(size >= 3, "Necessary for Printezis marks to work");
6095         HeapWord* start_pbit = addr + 1;
6096         HeapWord* end_pbit = addr + size - 1;
6097         assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6098                "inconsistent Printezis mark");
6099         // Verify inner mark bits (between Printezis bits) are clear,
6100         // but don't repeat if there are multiple dirty regions for
6101         // the same object, to avoid potential O(N^2) performance.
6102         if (addr != _last_scanned_object) {
6103           _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6104           _last_scanned_object = addr;
6105         }
6106       #endif // ASSERT
6107     } else {
6108       // An uninitialized object.
6109       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6110       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6111       size = pointer_delta(nextOneAddr + 1, addr);
6112       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6113              "alignment problem");
6114       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6115       // will dirty the card when the klass pointer is installed in the
6116       // object (signaling the completion of initialization).
6117     }
6118   } else {
6119     // Either a not yet marked object or an uninitialized object
6120     if (p->klass_or_null_acquire() == NULL) {
6121       // An uninitialized object, skip to the next card, since
6122       // we may not be able to read its P-bits yet.
6123       assert(size == 0, "Initial value");
6124     } else {
6125       // An object not (yet) reached by marking: we merely need to
6126       // compute its size so as to go look at the next block.
6127       assert(oopDesc::is_oop(p, true), "should be an oop");
6128       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6129     }
6130   }
6131   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6132   return size;
6133 }
6134 
6135 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6136   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6137          "CMS thread should hold CMS token");
6138   assert_lock_strong(_freelistLock);
6139   assert_lock_strong(_bitMap->lock());
6140   // relinquish the free_list_lock and bitMaplock()
6141   _bitMap->lock()->unlock();
6142   _freelistLock->unlock();
6143   ConcurrentMarkSweepThread::desynchronize(true);
6144   _collector->stopTimer();
6145   _collector->incrementYields();
6146 
6147   // See the comment in coordinator_yield()
6148   for (unsigned i = 0; i < CMSYieldSleepCount &&
6149                    ConcurrentMarkSweepThread::should_yield() &&
6150                    !CMSCollector::foregroundGCIsActive(); ++i) {
6151     os::sleep(Thread::current(), 1, false);
6152   }
6153 
6154   ConcurrentMarkSweepThread::synchronize(true);
6155   _freelistLock->lock_without_safepoint_check();
6156   _bitMap->lock()->lock_without_safepoint_check();
6157   _collector->startTimer();
6158 }
6159 
6160 
6161 //////////////////////////////////////////////////////////////////
6162 // SurvivorSpacePrecleanClosure
6163 //////////////////////////////////////////////////////////////////
6164 // This (single-threaded) closure is used to preclean the oops in
6165 // the survivor spaces.
6166 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6167 
6168   HeapWord* addr = (HeapWord*)p;
6169   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6170   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6171   assert(p->klass_or_null() != NULL, "object should be initialized");
6172   // an initialized object; ignore mark word in verification below
6173   // since we are running concurrent with mutators
6174   assert(oopDesc::is_oop(p, true), "should be an oop");
6175   // Note that we do not yield while we iterate over
6176   // the interior oops of p, pushing the relevant ones
6177   // on our marking stack.
6178   size_t size = p->oop_iterate_size(_scanning_closure);
6179   do_yield_check();
6180   // Observe that below, we do not abandon the preclean
6181   // phase as soon as we should; rather we empty the
6182   // marking stack before returning. This is to satisfy
6183   // some existing assertions. In general, it may be a
6184   // good idea to abort immediately and complete the marking
6185   // from the grey objects at a later time.
6186   while (!_mark_stack->isEmpty()) {
6187     oop new_oop = _mark_stack->pop();
6188     assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6189     assert(_bit_map->isMarked((HeapWord*)new_oop),
6190            "only grey objects on this stack");
6191     // iterate over the oops in this oop, marking and pushing
6192     // the ones in CMS heap (i.e. in _span).
6193     new_oop->oop_iterate(_scanning_closure);
6194     // check if it's time to yield
6195     do_yield_check();
6196   }
6197   unsigned int after_count =
6198     CMSHeap::heap()->total_collections();
6199   bool abort = (_before_count != after_count) ||
6200                _collector->should_abort_preclean();
6201   return abort ? 0 : size;
6202 }
6203 
6204 void SurvivorSpacePrecleanClosure::do_yield_work() {
6205   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6206          "CMS thread should hold CMS token");
6207   assert_lock_strong(_bit_map->lock());
6208   // Relinquish the bit map lock
6209   _bit_map->lock()->unlock();
6210   ConcurrentMarkSweepThread::desynchronize(true);
6211   _collector->stopTimer();
6212   _collector->incrementYields();
6213 
6214   // See the comment in coordinator_yield()
6215   for (unsigned i = 0; i < CMSYieldSleepCount &&
6216                        ConcurrentMarkSweepThread::should_yield() &&
6217                        !CMSCollector::foregroundGCIsActive(); ++i) {
6218     os::sleep(Thread::current(), 1, false);
6219   }
6220 
6221   ConcurrentMarkSweepThread::synchronize(true);
6222   _bit_map->lock()->lock_without_safepoint_check();
6223   _collector->startTimer();
6224 }
6225 
6226 // This closure is used to rescan the marked objects on the dirty cards
6227 // in the mod union table and the card table proper. In the parallel
6228 // case, although the bitMap is shared, we do a single read so the
6229 // isMarked() query is "safe".
6230 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6231   // Ignore mark word because we are running concurrent with mutators
6232   assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6233   HeapWord* addr = (HeapWord*)p;
6234   assert(_span.contains(addr), "we are scanning the CMS generation");
6235   bool is_obj_array = false;
6236   #ifdef ASSERT
6237     if (!_parallel) {
6238       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6239       assert(_collector->overflow_list_is_empty(),
6240              "overflow list should be empty");
6241 
6242     }
6243   #endif // ASSERT
6244   if (_bit_map->isMarked(addr)) {
6245     // Obj arrays are precisely marked, non-arrays are not;
6246     // so we scan objArrays precisely and non-arrays in their
6247     // entirety.
6248     if (p->is_objArray()) {
6249       is_obj_array = true;
6250       if (_parallel) {
6251         p->oop_iterate(_par_scan_closure, mr);
6252       } else {
6253         p->oop_iterate(_scan_closure, mr);
6254       }
6255     } else {
6256       if (_parallel) {
6257         p->oop_iterate(_par_scan_closure);
6258       } else {
6259         p->oop_iterate(_scan_closure);
6260       }
6261     }
6262   }
6263   #ifdef ASSERT
6264     if (!_parallel) {
6265       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6266       assert(_collector->overflow_list_is_empty(),
6267              "overflow list should be empty");
6268 
6269     }
6270   #endif // ASSERT
6271   return is_obj_array;
6272 }
6273 
6274 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6275                         MemRegion span,
6276                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6277                         bool should_yield, bool verifying):
6278   _collector(collector),
6279   _span(span),
6280   _bitMap(bitMap),
6281   _mut(&collector->_modUnionTable),
6282   _markStack(markStack),
6283   _yield(should_yield),
6284   _skipBits(0)
6285 {
6286   assert(_markStack->isEmpty(), "stack should be empty");
6287   _finger = _bitMap->startWord();
6288   _threshold = _finger;
6289   assert(_collector->_restart_addr == NULL, "Sanity check");
6290   assert(_span.contains(_finger), "Out of bounds _finger?");
6291   DEBUG_ONLY(_verifying = verifying;)
6292 }
6293 
6294 void MarkFromRootsClosure::reset(HeapWord* addr) {
6295   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6296   assert(_span.contains(addr), "Out of bounds _finger?");
6297   _finger = addr;
6298   _threshold = align_up(_finger, CardTable::card_size);
6299 }
6300 
6301 // Should revisit to see if this should be restructured for
6302 // greater efficiency.
6303 bool MarkFromRootsClosure::do_bit(size_t offset) {
6304   if (_skipBits > 0) {
6305     _skipBits--;
6306     return true;
6307   }
6308   // convert offset into a HeapWord*
6309   HeapWord* addr = _bitMap->startWord() + offset;
6310   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6311          "address out of range");
6312   assert(_bitMap->isMarked(addr), "tautology");
6313   if (_bitMap->isMarked(addr+1)) {
6314     // this is an allocated but not yet initialized object
6315     assert(_skipBits == 0, "tautology");
6316     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6317     oop p = oop(addr);
6318     if (p->klass_or_null_acquire() == NULL) {
6319       DEBUG_ONLY(if (!_verifying) {)
6320         // We re-dirty the cards on which this object lies and increase
6321         // the _threshold so that we'll come back to scan this object
6322         // during the preclean or remark phase. (CMSCleanOnEnter)
6323         if (CMSCleanOnEnter) {
6324           size_t sz = _collector->block_size_using_printezis_bits(addr);
6325           HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6326           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6327           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6328           // Bump _threshold to end_card_addr; note that
6329           // _threshold cannot possibly exceed end_card_addr, anyhow.
6330           // This prevents future clearing of the card as the scan proceeds
6331           // to the right.
6332           assert(_threshold <= end_card_addr,
6333                  "Because we are just scanning into this object");
6334           if (_threshold < end_card_addr) {
6335             _threshold = end_card_addr;
6336           }
6337           if (p->klass_or_null_acquire() != NULL) {
6338             // Redirty the range of cards...
6339             _mut->mark_range(redirty_range);
6340           } // ...else the setting of klass will dirty the card anyway.
6341         }
6342       DEBUG_ONLY(})
6343       return true;
6344     }
6345   }
6346   scanOopsInOop(addr);
6347   return true;
6348 }
6349 
6350 // We take a break if we've been at this for a while,
6351 // so as to avoid monopolizing the locks involved.
6352 void MarkFromRootsClosure::do_yield_work() {
6353   // First give up the locks, then yield, then re-lock
6354   // We should probably use a constructor/destructor idiom to
6355   // do this unlock/lock or modify the MutexUnlocker class to
6356   // serve our purpose. XXX
6357   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6358          "CMS thread should hold CMS token");
6359   assert_lock_strong(_bitMap->lock());
6360   _bitMap->lock()->unlock();
6361   ConcurrentMarkSweepThread::desynchronize(true);
6362   _collector->stopTimer();
6363   _collector->incrementYields();
6364 
6365   // See the comment in coordinator_yield()
6366   for (unsigned i = 0; i < CMSYieldSleepCount &&
6367                        ConcurrentMarkSweepThread::should_yield() &&
6368                        !CMSCollector::foregroundGCIsActive(); ++i) {
6369     os::sleep(Thread::current(), 1, false);
6370   }
6371 
6372   ConcurrentMarkSweepThread::synchronize(true);
6373   _bitMap->lock()->lock_without_safepoint_check();
6374   _collector->startTimer();
6375 }
6376 
6377 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6378   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6379   assert(_markStack->isEmpty(),
6380          "should drain stack to limit stack usage");
6381   // convert ptr to an oop preparatory to scanning
6382   oop obj = oop(ptr);
6383   // Ignore mark word in verification below, since we
6384   // may be running concurrent with mutators.
6385   assert(oopDesc::is_oop(obj, true), "should be an oop");
6386   assert(_finger <= ptr, "_finger runneth ahead");
6387   // advance the finger to right end of this object
6388   _finger = ptr + obj->size();
6389   assert(_finger > ptr, "we just incremented it above");
6390   // On large heaps, it may take us some time to get through
6391   // the marking phase. During
6392   // this time it's possible that a lot of mutations have
6393   // accumulated in the card table and the mod union table --
6394   // these mutation records are redundant until we have
6395   // actually traced into the corresponding card.
6396   // Here, we check whether advancing the finger would make
6397   // us cross into a new card, and if so clear corresponding
6398   // cards in the MUT (preclean them in the card-table in the
6399   // future).
6400 
6401   DEBUG_ONLY(if (!_verifying) {)
6402     // The clean-on-enter optimization is disabled by default,
6403     // until we fix 6178663.
6404     if (CMSCleanOnEnter && (_finger > _threshold)) {
6405       // [_threshold, _finger) represents the interval
6406       // of cards to be cleared  in MUT (or precleaned in card table).
6407       // The set of cards to be cleared is all those that overlap
6408       // with the interval [_threshold, _finger); note that
6409       // _threshold is always kept card-aligned but _finger isn't
6410       // always card-aligned.
6411       HeapWord* old_threshold = _threshold;
6412       assert(is_aligned(old_threshold, CardTable::card_size),
6413              "_threshold should always be card-aligned");
6414       _threshold = align_up(_finger, CardTable::card_size);
6415       MemRegion mr(old_threshold, _threshold);
6416       assert(!mr.is_empty(), "Control point invariant");
6417       assert(_span.contains(mr), "Should clear within span");
6418       _mut->clear_range(mr);
6419     }
6420   DEBUG_ONLY(})
6421   // Note: the finger doesn't advance while we drain
6422   // the stack below.
6423   PushOrMarkClosure pushOrMarkClosure(_collector,
6424                                       _span, _bitMap, _markStack,
6425                                       _finger, this);
6426   bool res = _markStack->push(obj);
6427   assert(res, "Empty non-zero size stack should have space for single push");
6428   while (!_markStack->isEmpty()) {
6429     oop new_oop = _markStack->pop();
6430     // Skip verifying header mark word below because we are
6431     // running concurrent with mutators.
6432     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6433     // now scan this oop's oops
6434     new_oop->oop_iterate(&pushOrMarkClosure);
6435     do_yield_check();
6436   }
6437   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6438 }
6439 
6440 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6441                        CMSCollector* collector, MemRegion span,
6442                        CMSBitMap* bit_map,
6443                        OopTaskQueue* work_queue,
6444                        CMSMarkStack*  overflow_stack):
6445   _collector(collector),
6446   _whole_span(collector->_span),
6447   _span(span),
6448   _bit_map(bit_map),
6449   _mut(&collector->_modUnionTable),
6450   _work_queue(work_queue),
6451   _overflow_stack(overflow_stack),
6452   _skip_bits(0),
6453   _task(task)
6454 {
6455   assert(_work_queue->size() == 0, "work_queue should be empty");
6456   _finger = span.start();
6457   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6458   assert(_span.contains(_finger), "Out of bounds _finger?");
6459 }
6460 
6461 // Should revisit to see if this should be restructured for
6462 // greater efficiency.
6463 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6464   if (_skip_bits > 0) {
6465     _skip_bits--;
6466     return true;
6467   }
6468   // convert offset into a HeapWord*
6469   HeapWord* addr = _bit_map->startWord() + offset;
6470   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6471          "address out of range");
6472   assert(_bit_map->isMarked(addr), "tautology");
6473   if (_bit_map->isMarked(addr+1)) {
6474     // this is an allocated object that might not yet be initialized
6475     assert(_skip_bits == 0, "tautology");
6476     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6477     oop p = oop(addr);
6478     if (p->klass_or_null_acquire() == NULL) {
6479       // in the case of Clean-on-Enter optimization, redirty card
6480       // and avoid clearing card by increasing  the threshold.
6481       return true;
6482     }
6483   }
6484   scan_oops_in_oop(addr);
6485   return true;
6486 }
6487 
6488 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6489   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6490   // Should we assert that our work queue is empty or
6491   // below some drain limit?
6492   assert(_work_queue->size() == 0,
6493          "should drain stack to limit stack usage");
6494   // convert ptr to an oop preparatory to scanning
6495   oop obj = oop(ptr);
6496   // Ignore mark word in verification below, since we
6497   // may be running concurrent with mutators.
6498   assert(oopDesc::is_oop(obj, true), "should be an oop");
6499   assert(_finger <= ptr, "_finger runneth ahead");
6500   // advance the finger to right end of this object
6501   _finger = ptr + obj->size();
6502   assert(_finger > ptr, "we just incremented it above");
6503   // On large heaps, it may take us some time to get through
6504   // the marking phase. During
6505   // this time it's possible that a lot of mutations have
6506   // accumulated in the card table and the mod union table --
6507   // these mutation records are redundant until we have
6508   // actually traced into the corresponding card.
6509   // Here, we check whether advancing the finger would make
6510   // us cross into a new card, and if so clear corresponding
6511   // cards in the MUT (preclean them in the card-table in the
6512   // future).
6513 
6514   // The clean-on-enter optimization is disabled by default,
6515   // until we fix 6178663.
6516   if (CMSCleanOnEnter && (_finger > _threshold)) {
6517     // [_threshold, _finger) represents the interval
6518     // of cards to be cleared  in MUT (or precleaned in card table).
6519     // The set of cards to be cleared is all those that overlap
6520     // with the interval [_threshold, _finger); note that
6521     // _threshold is always kept card-aligned but _finger isn't
6522     // always card-aligned.
6523     HeapWord* old_threshold = _threshold;
6524     assert(is_aligned(old_threshold, CardTable::card_size),
6525            "_threshold should always be card-aligned");
6526     _threshold = align_up(_finger, CardTable::card_size);
6527     MemRegion mr(old_threshold, _threshold);
6528     assert(!mr.is_empty(), "Control point invariant");
6529     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6530     _mut->clear_range(mr);
6531   }
6532 
6533   // Note: the local finger doesn't advance while we drain
6534   // the stack below, but the global finger sure can and will.
6535   HeapWord* volatile* gfa = _task->global_finger_addr();
6536   ParPushOrMarkClosure pushOrMarkClosure(_collector,
6537                                          _span, _bit_map,
6538                                          _work_queue,
6539                                          _overflow_stack,
6540                                          _finger,
6541                                          gfa, this);
6542   bool res = _work_queue->push(obj);   // overflow could occur here
6543   assert(res, "Will hold once we use workqueues");
6544   while (true) {
6545     oop new_oop;
6546     if (!_work_queue->pop_local(new_oop)) {
6547       // We emptied our work_queue; check if there's stuff that can
6548       // be gotten from the overflow stack.
6549       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6550             _overflow_stack, _work_queue)) {
6551         do_yield_check();
6552         continue;
6553       } else {  // done
6554         break;
6555       }
6556     }
6557     // Skip verifying header mark word below because we are
6558     // running concurrent with mutators.
6559     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6560     // now scan this oop's oops
6561     new_oop->oop_iterate(&pushOrMarkClosure);
6562     do_yield_check();
6563   }
6564   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6565 }
6566 
6567 // Yield in response to a request from VM Thread or
6568 // from mutators.
6569 void ParMarkFromRootsClosure::do_yield_work() {
6570   assert(_task != NULL, "sanity");
6571   _task->yield();
6572 }
6573 
6574 // A variant of the above used for verifying CMS marking work.
6575 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6576                         MemRegion span,
6577                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6578                         CMSMarkStack*  mark_stack):
6579   _collector(collector),
6580   _span(span),
6581   _verification_bm(verification_bm),
6582   _cms_bm(cms_bm),
6583   _mark_stack(mark_stack),
6584   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6585                       mark_stack)
6586 {
6587   assert(_mark_stack->isEmpty(), "stack should be empty");
6588   _finger = _verification_bm->startWord();
6589   assert(_collector->_restart_addr == NULL, "Sanity check");
6590   assert(_span.contains(_finger), "Out of bounds _finger?");
6591 }
6592 
6593 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6594   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6595   assert(_span.contains(addr), "Out of bounds _finger?");
6596   _finger = addr;
6597 }
6598 
6599 // Should revisit to see if this should be restructured for
6600 // greater efficiency.
6601 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6602   // convert offset into a HeapWord*
6603   HeapWord* addr = _verification_bm->startWord() + offset;
6604   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6605          "address out of range");
6606   assert(_verification_bm->isMarked(addr), "tautology");
6607   assert(_cms_bm->isMarked(addr), "tautology");
6608 
6609   assert(_mark_stack->isEmpty(),
6610          "should drain stack to limit stack usage");
6611   // convert addr to an oop preparatory to scanning
6612   oop obj = oop(addr);
6613   assert(oopDesc::is_oop(obj), "should be an oop");
6614   assert(_finger <= addr, "_finger runneth ahead");
6615   // advance the finger to right end of this object
6616   _finger = addr + obj->size();
6617   assert(_finger > addr, "we just incremented it above");
6618   // Note: the finger doesn't advance while we drain
6619   // the stack below.
6620   bool res = _mark_stack->push(obj);
6621   assert(res, "Empty non-zero size stack should have space for single push");
6622   while (!_mark_stack->isEmpty()) {
6623     oop new_oop = _mark_stack->pop();
6624     assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6625     // now scan this oop's oops
6626     new_oop->oop_iterate(&_pam_verify_closure);
6627   }
6628   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6629   return true;
6630 }
6631 
6632 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6633   CMSCollector* collector, MemRegion span,
6634   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6635   CMSMarkStack*  mark_stack):
6636   MetadataAwareOopClosure(collector->ref_processor()),
6637   _collector(collector),
6638   _span(span),
6639   _verification_bm(verification_bm),
6640   _cms_bm(cms_bm),
6641   _mark_stack(mark_stack)
6642 { }
6643 
6644 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) {
6645   oop obj = RawAccess<>::oop_load(p);
6646   do_oop(obj);
6647 }
6648 
6649 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6650 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6651 
6652 // Upon stack overflow, we discard (part of) the stack,
6653 // remembering the least address amongst those discarded
6654 // in CMSCollector's _restart_address.
6655 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6656   // Remember the least grey address discarded
6657   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6658   _collector->lower_restart_addr(ra);
6659   _mark_stack->reset();  // discard stack contents
6660   _mark_stack->expand(); // expand the stack if possible
6661 }
6662 
6663 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6664   assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6665   HeapWord* addr = (HeapWord*)obj;
6666   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6667     // Oop lies in _span and isn't yet grey or black
6668     _verification_bm->mark(addr);            // now grey
6669     if (!_cms_bm->isMarked(addr)) {
6670       Log(gc, verify) log;
6671       ResourceMark rm;
6672       LogStream ls(log.error());
6673       oop(addr)->print_on(&ls);
6674       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6675       fatal("... aborting");
6676     }
6677 
6678     if (!_mark_stack->push(obj)) { // stack overflow
6679       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6680       assert(_mark_stack->isFull(), "Else push should have succeeded");
6681       handle_stack_overflow(addr);
6682     }
6683     // anything including and to the right of _finger
6684     // will be scanned as we iterate over the remainder of the
6685     // bit map
6686   }
6687 }
6688 
6689 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6690                      MemRegion span,
6691                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6692                      HeapWord* finger, MarkFromRootsClosure* parent) :
6693   MetadataAwareOopClosure(collector->ref_processor()),
6694   _collector(collector),
6695   _span(span),
6696   _bitMap(bitMap),
6697   _markStack(markStack),
6698   _finger(finger),
6699   _parent(parent)
6700 { }
6701 
6702 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6703                                            MemRegion span,
6704                                            CMSBitMap* bit_map,
6705                                            OopTaskQueue* work_queue,
6706                                            CMSMarkStack*  overflow_stack,
6707                                            HeapWord* finger,
6708                                            HeapWord* volatile* global_finger_addr,
6709                                            ParMarkFromRootsClosure* parent) :
6710   MetadataAwareOopClosure(collector->ref_processor()),
6711   _collector(collector),
6712   _whole_span(collector->_span),
6713   _span(span),
6714   _bit_map(bit_map),
6715   _work_queue(work_queue),
6716   _overflow_stack(overflow_stack),
6717   _finger(finger),
6718   _global_finger_addr(global_finger_addr),
6719   _parent(parent)
6720 { }
6721 
6722 // Assumes thread-safe access by callers, who are
6723 // responsible for mutual exclusion.
6724 void CMSCollector::lower_restart_addr(HeapWord* low) {
6725   assert(_span.contains(low), "Out of bounds addr");
6726   if (_restart_addr == NULL) {
6727     _restart_addr = low;
6728   } else {
6729     _restart_addr = MIN2(_restart_addr, low);
6730   }
6731 }
6732 
6733 // Upon stack overflow, we discard (part of) the stack,
6734 // remembering the least address amongst those discarded
6735 // in CMSCollector's _restart_address.
6736 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6737   // Remember the least grey address discarded
6738   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6739   _collector->lower_restart_addr(ra);
6740   _markStack->reset();  // discard stack contents
6741   _markStack->expand(); // expand the stack if possible
6742 }
6743 
6744 // Upon stack overflow, we discard (part of) the stack,
6745 // remembering the least address amongst those discarded
6746 // in CMSCollector's _restart_address.
6747 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6748   // We need to do this under a mutex to prevent other
6749   // workers from interfering with the work done below.
6750   MutexLockerEx ml(_overflow_stack->par_lock(),
6751                    Mutex::_no_safepoint_check_flag);
6752   // Remember the least grey address discarded
6753   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6754   _collector->lower_restart_addr(ra);
6755   _overflow_stack->reset();  // discard stack contents
6756   _overflow_stack->expand(); // expand the stack if possible
6757 }
6758 
6759 void PushOrMarkClosure::do_oop(oop obj) {
6760   // Ignore mark word because we are running concurrent with mutators.
6761   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6762   HeapWord* addr = (HeapWord*)obj;
6763   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6764     // Oop lies in _span and isn't yet grey or black
6765     _bitMap->mark(addr);            // now grey
6766     if (addr < _finger) {
6767       // the bit map iteration has already either passed, or
6768       // sampled, this bit in the bit map; we'll need to
6769       // use the marking stack to scan this oop's oops.
6770       bool simulate_overflow = false;
6771       NOT_PRODUCT(
6772         if (CMSMarkStackOverflowALot &&
6773             _collector->simulate_overflow()) {
6774           // simulate a stack overflow
6775           simulate_overflow = true;
6776         }
6777       )
6778       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6779         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6780         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6781         handle_stack_overflow(addr);
6782       }
6783     }
6784     // anything including and to the right of _finger
6785     // will be scanned as we iterate over the remainder of the
6786     // bit map
6787     do_yield_check();
6788   }
6789 }
6790 
6791 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
6792 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6793 
6794 void ParPushOrMarkClosure::do_oop(oop obj) {
6795   // Ignore mark word because we are running concurrent with mutators.
6796   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6797   HeapWord* addr = (HeapWord*)obj;
6798   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6799     // Oop lies in _span and isn't yet grey or black
6800     // We read the global_finger (volatile read) strictly after marking oop
6801     bool res = _bit_map->par_mark(addr);    // now grey
6802     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6803     // Should we push this marked oop on our stack?
6804     // -- if someone else marked it, nothing to do
6805     // -- if target oop is above global finger nothing to do
6806     // -- if target oop is in chunk and above local finger
6807     //      then nothing to do
6808     // -- else push on work queue
6809     if (   !res       // someone else marked it, they will deal with it
6810         || (addr >= *gfa)  // will be scanned in a later task
6811         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6812       return;
6813     }
6814     // the bit map iteration has already either passed, or
6815     // sampled, this bit in the bit map; we'll need to
6816     // use the marking stack to scan this oop's oops.
6817     bool simulate_overflow = false;
6818     NOT_PRODUCT(
6819       if (CMSMarkStackOverflowALot &&
6820           _collector->simulate_overflow()) {
6821         // simulate a stack overflow
6822         simulate_overflow = true;
6823       }
6824     )
6825     if (simulate_overflow ||
6826         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6827       // stack overflow
6828       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6829       // We cannot assert that the overflow stack is full because
6830       // it may have been emptied since.
6831       assert(simulate_overflow ||
6832              _work_queue->size() == _work_queue->max_elems(),
6833             "Else push should have succeeded");
6834       handle_stack_overflow(addr);
6835     }
6836     do_yield_check();
6837   }
6838 }
6839 
6840 void ParPushOrMarkClosure::do_oop(oop* p)       { ParPushOrMarkClosure::do_oop_work(p); }
6841 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6842 
6843 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6844                                        MemRegion span,
6845                                        ReferenceDiscoverer* rd,
6846                                        CMSBitMap* bit_map,
6847                                        CMSBitMap* mod_union_table,
6848                                        CMSMarkStack*  mark_stack,
6849                                        bool           concurrent_precleaning):
6850   MetadataAwareOopClosure(rd),
6851   _collector(collector),
6852   _span(span),
6853   _bit_map(bit_map),
6854   _mod_union_table(mod_union_table),
6855   _mark_stack(mark_stack),
6856   _concurrent_precleaning(concurrent_precleaning)
6857 {
6858   assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6859 }
6860 
6861 // Grey object rescan during pre-cleaning and second checkpoint phases --
6862 // the non-parallel version (the parallel version appears further below.)
6863 void PushAndMarkClosure::do_oop(oop obj) {
6864   // Ignore mark word verification. If during concurrent precleaning,
6865   // the object monitor may be locked. If during the checkpoint
6866   // phases, the object may already have been reached by a  different
6867   // path and may be at the end of the global overflow list (so
6868   // the mark word may be NULL).
6869   assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6870          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6871   HeapWord* addr = (HeapWord*)obj;
6872   // Check if oop points into the CMS generation
6873   // and is not marked
6874   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6875     // a white object ...
6876     _bit_map->mark(addr);         // ... now grey
6877     // push on the marking stack (grey set)
6878     bool simulate_overflow = false;
6879     NOT_PRODUCT(
6880       if (CMSMarkStackOverflowALot &&
6881           _collector->simulate_overflow()) {
6882         // simulate a stack overflow
6883         simulate_overflow = true;
6884       }
6885     )
6886     if (simulate_overflow || !_mark_stack->push(obj)) {
6887       if (_concurrent_precleaning) {
6888          // During precleaning we can just dirty the appropriate card(s)
6889          // in the mod union table, thus ensuring that the object remains
6890          // in the grey set  and continue. In the case of object arrays
6891          // we need to dirty all of the cards that the object spans,
6892          // since the rescan of object arrays will be limited to the
6893          // dirty cards.
6894          // Note that no one can be interfering with us in this action
6895          // of dirtying the mod union table, so no locking or atomics
6896          // are required.
6897          if (obj->is_objArray()) {
6898            size_t sz = obj->size();
6899            HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6900            MemRegion redirty_range = MemRegion(addr, end_card_addr);
6901            assert(!redirty_range.is_empty(), "Arithmetical tautology");
6902            _mod_union_table->mark_range(redirty_range);
6903          } else {
6904            _mod_union_table->mark(addr);
6905          }
6906          _collector->_ser_pmc_preclean_ovflw++;
6907       } else {
6908          // During the remark phase, we need to remember this oop
6909          // in the overflow list.
6910          _collector->push_on_overflow_list(obj);
6911          _collector->_ser_pmc_remark_ovflw++;
6912       }
6913     }
6914   }
6915 }
6916 
6917 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6918                                              MemRegion span,
6919                                              ReferenceDiscoverer* rd,
6920                                              CMSBitMap* bit_map,
6921                                              OopTaskQueue* work_queue):
6922   MetadataAwareOopClosure(rd),
6923   _collector(collector),
6924   _span(span),
6925   _bit_map(bit_map),
6926   _work_queue(work_queue)
6927 {
6928   assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6929 }
6930 
6931 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
6932 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6933 
6934 // Grey object rescan during second checkpoint phase --
6935 // the parallel version.
6936 void ParPushAndMarkClosure::do_oop(oop obj) {
6937   // In the assert below, we ignore the mark word because
6938   // this oop may point to an already visited object that is
6939   // on the overflow stack (in which case the mark word has
6940   // been hijacked for chaining into the overflow stack --
6941   // if this is the last object in the overflow stack then
6942   // its mark word will be NULL). Because this object may
6943   // have been subsequently popped off the global overflow
6944   // stack, and the mark word possibly restored to the prototypical
6945   // value, by the time we get to examined this failing assert in
6946   // the debugger, is_oop_or_null(false) may subsequently start
6947   // to hold.
6948   assert(oopDesc::is_oop_or_null(obj, true),
6949          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6950   HeapWord* addr = (HeapWord*)obj;
6951   // Check if oop points into the CMS generation
6952   // and is not marked
6953   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6954     // a white object ...
6955     // If we manage to "claim" the object, by being the
6956     // first thread to mark it, then we push it on our
6957     // marking stack
6958     if (_bit_map->par_mark(addr)) {     // ... now grey
6959       // push on work queue (grey set)
6960       bool simulate_overflow = false;
6961       NOT_PRODUCT(
6962         if (CMSMarkStackOverflowALot &&
6963             _collector->par_simulate_overflow()) {
6964           // simulate a stack overflow
6965           simulate_overflow = true;
6966         }
6967       )
6968       if (simulate_overflow || !_work_queue->push(obj)) {
6969         _collector->par_push_on_overflow_list(obj);
6970         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
6971       }
6972     } // Else, some other thread got there first
6973   }
6974 }
6975 
6976 void ParPushAndMarkClosure::do_oop(oop* p)       { ParPushAndMarkClosure::do_oop_work(p); }
6977 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6978 
6979 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6980   Mutex* bml = _collector->bitMapLock();
6981   assert_lock_strong(bml);
6982   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6983          "CMS thread should hold CMS token");
6984 
6985   bml->unlock();
6986   ConcurrentMarkSweepThread::desynchronize(true);
6987 
6988   _collector->stopTimer();
6989   _collector->incrementYields();
6990 
6991   // See the comment in coordinator_yield()
6992   for (unsigned i = 0; i < CMSYieldSleepCount &&
6993                        ConcurrentMarkSweepThread::should_yield() &&
6994                        !CMSCollector::foregroundGCIsActive(); ++i) {
6995     os::sleep(Thread::current(), 1, false);
6996   }
6997 
6998   ConcurrentMarkSweepThread::synchronize(true);
6999   bml->lock();
7000 
7001   _collector->startTimer();
7002 }
7003 
7004 bool CMSPrecleanRefsYieldClosure::should_return() {
7005   if (ConcurrentMarkSweepThread::should_yield()) {
7006     do_yield_work();
7007   }
7008   return _collector->foregroundGCIsActive();
7009 }
7010 
7011 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7012   assert(((size_t)mr.start())%CardTable::card_size_in_words == 0,
7013          "mr should be aligned to start at a card boundary");
7014   // We'd like to assert:
7015   // assert(mr.word_size()%CardTable::card_size_in_words == 0,
7016   //        "mr should be a range of cards");
7017   // However, that would be too strong in one case -- the last
7018   // partition ends at _unallocated_block which, in general, can be
7019   // an arbitrary boundary, not necessarily card aligned.
7020   _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words;
7021   _space->object_iterate_mem(mr, &_scan_cl);
7022 }
7023 
7024 SweepClosure::SweepClosure(CMSCollector* collector,
7025                            ConcurrentMarkSweepGeneration* g,
7026                            CMSBitMap* bitMap, bool should_yield) :
7027   _collector(collector),
7028   _g(g),
7029   _sp(g->cmsSpace()),
7030   _limit(_sp->sweep_limit()),
7031   _freelistLock(_sp->freelistLock()),
7032   _bitMap(bitMap),
7033   _yield(should_yield),
7034   _inFreeRange(false),           // No free range at beginning of sweep
7035   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7036   _lastFreeRangeCoalesced(false),
7037   _freeFinger(g->used_region().start())
7038 {
7039   NOT_PRODUCT(
7040     _numObjectsFreed = 0;
7041     _numWordsFreed   = 0;
7042     _numObjectsLive = 0;
7043     _numWordsLive = 0;
7044     _numObjectsAlreadyFree = 0;
7045     _numWordsAlreadyFree = 0;
7046     _last_fc = NULL;
7047 
7048     _sp->initializeIndexedFreeListArrayReturnedBytes();
7049     _sp->dictionary()->initialize_dict_returned_bytes();
7050   )
7051   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7052          "sweep _limit out of bounds");
7053   log_develop_trace(gc, sweep)("====================");
7054   log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7055 }
7056 
7057 void SweepClosure::print_on(outputStream* st) const {
7058   st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7059                p2i(_sp->bottom()), p2i(_sp->end()));
7060   st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7061   st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7062   NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7063   st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7064                _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7065 }
7066 
7067 #ifndef PRODUCT
7068 // Assertion checking only:  no useful work in product mode --
7069 // however, if any of the flags below become product flags,
7070 // you may need to review this code to see if it needs to be
7071 // enabled in product mode.
7072 SweepClosure::~SweepClosure() {
7073   assert_lock_strong(_freelistLock);
7074   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7075          "sweep _limit out of bounds");
7076   if (inFreeRange()) {
7077     Log(gc, sweep) log;
7078     log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7079     ResourceMark rm;
7080     LogStream ls(log.error());
7081     print_on(&ls);
7082     ShouldNotReachHere();
7083   }
7084 
7085   if (log_is_enabled(Debug, gc, sweep)) {
7086     log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7087                          _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7088     log_debug(gc, sweep)("Live " SIZE_FORMAT " objects,  " SIZE_FORMAT " bytes  Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7089                          _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7090     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7091     log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7092   }
7093 
7094   if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7095     size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7096     size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7097     size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7098     log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes   Indexed List Returned " SIZE_FORMAT " bytes        Dictionary Returned " SIZE_FORMAT " bytes",
7099                          returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7100   }
7101   log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7102   log_develop_trace(gc, sweep)("================");
7103 }
7104 #endif  // PRODUCT
7105 
7106 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7107     bool freeRangeInFreeLists) {
7108   log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7109                                p2i(freeFinger), freeRangeInFreeLists);
7110   assert(!inFreeRange(), "Trampling existing free range");
7111   set_inFreeRange(true);
7112   set_lastFreeRangeCoalesced(false);
7113 
7114   set_freeFinger(freeFinger);
7115   set_freeRangeInFreeLists(freeRangeInFreeLists);
7116   if (CMSTestInFreeList) {
7117     if (freeRangeInFreeLists) {
7118       FreeChunk* fc = (FreeChunk*) freeFinger;
7119       assert(fc->is_free(), "A chunk on the free list should be free.");
7120       assert(fc->size() > 0, "Free range should have a size");
7121       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7122     }
7123   }
7124 }
7125 
7126 // Note that the sweeper runs concurrently with mutators. Thus,
7127 // it is possible for direct allocation in this generation to happen
7128 // in the middle of the sweep. Note that the sweeper also coalesces
7129 // contiguous free blocks. Thus, unless the sweeper and the allocator
7130 // synchronize appropriately freshly allocated blocks may get swept up.
7131 // This is accomplished by the sweeper locking the free lists while
7132 // it is sweeping. Thus blocks that are determined to be free are
7133 // indeed free. There is however one additional complication:
7134 // blocks that have been allocated since the final checkpoint and
7135 // mark, will not have been marked and so would be treated as
7136 // unreachable and swept up. To prevent this, the allocator marks
7137 // the bit map when allocating during the sweep phase. This leads,
7138 // however, to a further complication -- objects may have been allocated
7139 // but not yet initialized -- in the sense that the header isn't yet
7140 // installed. The sweeper can not then determine the size of the block
7141 // in order to skip over it. To deal with this case, we use a technique
7142 // (due to Printezis) to encode such uninitialized block sizes in the
7143 // bit map. Since the bit map uses a bit per every HeapWord, but the
7144 // CMS generation has a minimum object size of 3 HeapWords, it follows
7145 // that "normal marks" won't be adjacent in the bit map (there will
7146 // always be at least two 0 bits between successive 1 bits). We make use
7147 // of these "unused" bits to represent uninitialized blocks -- the bit
7148 // corresponding to the start of the uninitialized object and the next
7149 // bit are both set. Finally, a 1 bit marks the end of the object that
7150 // started with the two consecutive 1 bits to indicate its potentially
7151 // uninitialized state.
7152 
7153 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7154   FreeChunk* fc = (FreeChunk*)addr;
7155   size_t res;
7156 
7157   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7158   // than "addr == _limit" because although _limit was a block boundary when
7159   // we started the sweep, it may no longer be one because heap expansion
7160   // may have caused us to coalesce the block ending at the address _limit
7161   // with a newly expanded chunk (this happens when _limit was set to the
7162   // previous _end of the space), so we may have stepped past _limit:
7163   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7164   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7165     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7166            "sweep _limit out of bounds");
7167     assert(addr < _sp->end(), "addr out of bounds");
7168     // Flush any free range we might be holding as a single
7169     // coalesced chunk to the appropriate free list.
7170     if (inFreeRange()) {
7171       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7172              "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7173       flush_cur_free_chunk(freeFinger(),
7174                            pointer_delta(addr, freeFinger()));
7175       log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7176                                    p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7177                                    lastFreeRangeCoalesced() ? 1 : 0);
7178     }
7179 
7180     // help the iterator loop finish
7181     return pointer_delta(_sp->end(), addr);
7182   }
7183 
7184   assert(addr < _limit, "sweep invariant");
7185   // check if we should yield
7186   do_yield_check(addr);
7187   if (fc->is_free()) {
7188     // Chunk that is already free
7189     res = fc->size();
7190     do_already_free_chunk(fc);
7191     debug_only(_sp->verifyFreeLists());
7192     // If we flush the chunk at hand in lookahead_and_flush()
7193     // and it's coalesced with a preceding chunk, then the
7194     // process of "mangling" the payload of the coalesced block
7195     // will cause erasure of the size information from the
7196     // (erstwhile) header of all the coalesced blocks but the
7197     // first, so the first disjunct in the assert will not hold
7198     // in that specific case (in which case the second disjunct
7199     // will hold).
7200     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7201            "Otherwise the size info doesn't change at this step");
7202     NOT_PRODUCT(
7203       _numObjectsAlreadyFree++;
7204       _numWordsAlreadyFree += res;
7205     )
7206     NOT_PRODUCT(_last_fc = fc;)
7207   } else if (!_bitMap->isMarked(addr)) {
7208     // Chunk is fresh garbage
7209     res = do_garbage_chunk(fc);
7210     debug_only(_sp->verifyFreeLists());
7211     NOT_PRODUCT(
7212       _numObjectsFreed++;
7213       _numWordsFreed += res;
7214     )
7215   } else {
7216     // Chunk that is alive.
7217     res = do_live_chunk(fc);
7218     debug_only(_sp->verifyFreeLists());
7219     NOT_PRODUCT(
7220         _numObjectsLive++;
7221         _numWordsLive += res;
7222     )
7223   }
7224   return res;
7225 }
7226 
7227 // For the smart allocation, record following
7228 //  split deaths - a free chunk is removed from its free list because
7229 //      it is being split into two or more chunks.
7230 //  split birth - a free chunk is being added to its free list because
7231 //      a larger free chunk has been split and resulted in this free chunk.
7232 //  coal death - a free chunk is being removed from its free list because
7233 //      it is being coalesced into a large free chunk.
7234 //  coal birth - a free chunk is being added to its free list because
7235 //      it was created when two or more free chunks where coalesced into
7236 //      this free chunk.
7237 //
7238 // These statistics are used to determine the desired number of free
7239 // chunks of a given size.  The desired number is chosen to be relative
7240 // to the end of a CMS sweep.  The desired number at the end of a sweep
7241 // is the
7242 //      count-at-end-of-previous-sweep (an amount that was enough)
7243 //              - count-at-beginning-of-current-sweep  (the excess)
7244 //              + split-births  (gains in this size during interval)
7245 //              - split-deaths  (demands on this size during interval)
7246 // where the interval is from the end of one sweep to the end of the
7247 // next.
7248 //
7249 // When sweeping the sweeper maintains an accumulated chunk which is
7250 // the chunk that is made up of chunks that have been coalesced.  That
7251 // will be termed the left-hand chunk.  A new chunk of garbage that
7252 // is being considered for coalescing will be referred to as the
7253 // right-hand chunk.
7254 //
7255 // When making a decision on whether to coalesce a right-hand chunk with
7256 // the current left-hand chunk, the current count vs. the desired count
7257 // of the left-hand chunk is considered.  Also if the right-hand chunk
7258 // is near the large chunk at the end of the heap (see
7259 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7260 // left-hand chunk is coalesced.
7261 //
7262 // When making a decision about whether to split a chunk, the desired count
7263 // vs. the current count of the candidate to be split is also considered.
7264 // If the candidate is underpopulated (currently fewer chunks than desired)
7265 // a chunk of an overpopulated (currently more chunks than desired) size may
7266 // be chosen.  The "hint" associated with a free list, if non-null, points
7267 // to a free list which may be overpopulated.
7268 //
7269 
7270 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7271   const size_t size = fc->size();
7272   // Chunks that cannot be coalesced are not in the
7273   // free lists.
7274   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7275     assert(_sp->verify_chunk_in_free_list(fc),
7276            "free chunk should be in free lists");
7277   }
7278   // a chunk that is already free, should not have been
7279   // marked in the bit map
7280   HeapWord* const addr = (HeapWord*) fc;
7281   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7282   // Verify that the bit map has no bits marked between
7283   // addr and purported end of this block.
7284   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7285 
7286   // Some chunks cannot be coalesced under any circumstances.
7287   // See the definition of cantCoalesce().
7288   if (!fc->cantCoalesce()) {
7289     // This chunk can potentially be coalesced.
7290     // All the work is done in
7291     do_post_free_or_garbage_chunk(fc, size);
7292     // Note that if the chunk is not coalescable (the else arm
7293     // below), we unconditionally flush, without needing to do
7294     // a "lookahead," as we do below.
7295     if (inFreeRange()) lookahead_and_flush(fc, size);
7296   } else {
7297     // Code path common to both original and adaptive free lists.
7298 
7299     // cant coalesce with previous block; this should be treated
7300     // as the end of a free run if any
7301     if (inFreeRange()) {
7302       // we kicked some butt; time to pick up the garbage
7303       assert(freeFinger() < addr, "freeFinger points too high");
7304       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7305     }
7306     // else, nothing to do, just continue
7307   }
7308 }
7309 
7310 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7311   // This is a chunk of garbage.  It is not in any free list.
7312   // Add it to a free list or let it possibly be coalesced into
7313   // a larger chunk.
7314   HeapWord* const addr = (HeapWord*) fc;
7315   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7316 
7317   // Verify that the bit map has no bits marked between
7318   // addr and purported end of just dead object.
7319   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7320   do_post_free_or_garbage_chunk(fc, size);
7321 
7322   assert(_limit >= addr + size,
7323          "A freshly garbage chunk can't possibly straddle over _limit");
7324   if (inFreeRange()) lookahead_and_flush(fc, size);
7325   return size;
7326 }
7327 
7328 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7329   HeapWord* addr = (HeapWord*) fc;
7330   // The sweeper has just found a live object. Return any accumulated
7331   // left hand chunk to the free lists.
7332   if (inFreeRange()) {
7333     assert(freeFinger() < addr, "freeFinger points too high");
7334     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7335   }
7336 
7337   // This object is live: we'd normally expect this to be
7338   // an oop, and like to assert the following:
7339   // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7340   // However, as we commented above, this may be an object whose
7341   // header hasn't yet been initialized.
7342   size_t size;
7343   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7344   if (_bitMap->isMarked(addr + 1)) {
7345     // Determine the size from the bit map, rather than trying to
7346     // compute it from the object header.
7347     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7348     size = pointer_delta(nextOneAddr + 1, addr);
7349     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7350            "alignment problem");
7351 
7352 #ifdef ASSERT
7353       if (oop(addr)->klass_or_null_acquire() != NULL) {
7354         // Ignore mark word because we are running concurrent with mutators
7355         assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7356         assert(size ==
7357                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7358                "P-mark and computed size do not agree");
7359       }
7360 #endif
7361 
7362   } else {
7363     // This should be an initialized object that's alive.
7364     assert(oop(addr)->klass_or_null_acquire() != NULL,
7365            "Should be an initialized object");
7366     // Ignore mark word because we are running concurrent with mutators
7367     assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7368     // Verify that the bit map has no bits marked between
7369     // addr and purported end of this block.
7370     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7371     assert(size >= 3, "Necessary for Printezis marks to work");
7372     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7373     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7374   }
7375   return size;
7376 }
7377 
7378 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7379                                                  size_t chunkSize) {
7380   // do_post_free_or_garbage_chunk() should only be called in the case
7381   // of the adaptive free list allocator.
7382   const bool fcInFreeLists = fc->is_free();
7383   assert((HeapWord*)fc <= _limit, "sweep invariant");
7384   if (CMSTestInFreeList && fcInFreeLists) {
7385     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7386   }
7387 
7388   log_develop_trace(gc, sweep)("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7389 
7390   HeapWord* const fc_addr = (HeapWord*) fc;
7391 
7392   bool coalesce = false;
7393   const size_t left  = pointer_delta(fc_addr, freeFinger());
7394   const size_t right = chunkSize;
7395   switch (FLSCoalescePolicy) {
7396     // numeric value forms a coalition aggressiveness metric
7397     case 0:  { // never coalesce
7398       coalesce = false;
7399       break;
7400     }
7401     case 1: { // coalesce if left & right chunks on overpopulated lists
7402       coalesce = _sp->coalOverPopulated(left) &&
7403                  _sp->coalOverPopulated(right);
7404       break;
7405     }
7406     case 2: { // coalesce if left chunk on overpopulated list (default)
7407       coalesce = _sp->coalOverPopulated(left);
7408       break;
7409     }
7410     case 3: { // coalesce if left OR right chunk on overpopulated list
7411       coalesce = _sp->coalOverPopulated(left) ||
7412                  _sp->coalOverPopulated(right);
7413       break;
7414     }
7415     case 4: { // always coalesce
7416       coalesce = true;
7417       break;
7418     }
7419     default:
7420      ShouldNotReachHere();
7421   }
7422 
7423   // Should the current free range be coalesced?
7424   // If the chunk is in a free range and either we decided to coalesce above
7425   // or the chunk is near the large block at the end of the heap
7426   // (isNearLargestChunk() returns true), then coalesce this chunk.
7427   const bool doCoalesce = inFreeRange()
7428                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7429   if (doCoalesce) {
7430     // Coalesce the current free range on the left with the new
7431     // chunk on the right.  If either is on a free list,
7432     // it must be removed from the list and stashed in the closure.
7433     if (freeRangeInFreeLists()) {
7434       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7435       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7436              "Size of free range is inconsistent with chunk size.");
7437       if (CMSTestInFreeList) {
7438         assert(_sp->verify_chunk_in_free_list(ffc),
7439                "Chunk is not in free lists");
7440       }
7441       _sp->coalDeath(ffc->size());
7442       _sp->removeFreeChunkFromFreeLists(ffc);
7443       set_freeRangeInFreeLists(false);
7444     }
7445     if (fcInFreeLists) {
7446       _sp->coalDeath(chunkSize);
7447       assert(fc->size() == chunkSize,
7448         "The chunk has the wrong size or is not in the free lists");
7449       _sp->removeFreeChunkFromFreeLists(fc);
7450     }
7451     set_lastFreeRangeCoalesced(true);
7452     print_free_block_coalesced(fc);
7453   } else {  // not in a free range and/or should not coalesce
7454     // Return the current free range and start a new one.
7455     if (inFreeRange()) {
7456       // In a free range but cannot coalesce with the right hand chunk.
7457       // Put the current free range into the free lists.
7458       flush_cur_free_chunk(freeFinger(),
7459                            pointer_delta(fc_addr, freeFinger()));
7460     }
7461     // Set up for new free range.  Pass along whether the right hand
7462     // chunk is in the free lists.
7463     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7464   }
7465 }
7466 
7467 // Lookahead flush:
7468 // If we are tracking a free range, and this is the last chunk that
7469 // we'll look at because its end crosses past _limit, we'll preemptively
7470 // flush it along with any free range we may be holding on to. Note that
7471 // this can be the case only for an already free or freshly garbage
7472 // chunk. If this block is an object, it can never straddle
7473 // over _limit. The "straddling" occurs when _limit is set at
7474 // the previous end of the space when this cycle started, and
7475 // a subsequent heap expansion caused the previously co-terminal
7476 // free block to be coalesced with the newly expanded portion,
7477 // thus rendering _limit a non-block-boundary making it dangerous
7478 // for the sweeper to step over and examine.
7479 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7480   assert(inFreeRange(), "Should only be called if currently in a free range.");
7481   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7482   assert(_sp->used_region().contains(eob - 1),
7483          "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7484          " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7485          " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7486          p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7487   if (eob >= _limit) {
7488     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7489     log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7490                                  "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7491                                  "[" PTR_FORMAT "," PTR_FORMAT ")",
7492                                  p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7493     // Return the storage we are tracking back into the free lists.
7494     log_develop_trace(gc, sweep)("Flushing ... ");
7495     assert(freeFinger() < eob, "Error");
7496     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7497   }
7498 }
7499 
7500 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7501   assert(inFreeRange(), "Should only be called if currently in a free range.");
7502   assert(size > 0,
7503     "A zero sized chunk cannot be added to the free lists.");
7504   if (!freeRangeInFreeLists()) {
7505     if (CMSTestInFreeList) {
7506       FreeChunk* fc = (FreeChunk*) chunk;
7507       fc->set_size(size);
7508       assert(!_sp->verify_chunk_in_free_list(fc),
7509              "chunk should not be in free lists yet");
7510     }
7511     log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7512     // A new free range is going to be starting.  The current
7513     // free range has not been added to the free lists yet or
7514     // was removed so add it back.
7515     // If the current free range was coalesced, then the death
7516     // of the free range was recorded.  Record a birth now.
7517     if (lastFreeRangeCoalesced()) {
7518       _sp->coalBirth(size);
7519     }
7520     _sp->addChunkAndRepairOffsetTable(chunk, size,
7521             lastFreeRangeCoalesced());
7522   } else {
7523     log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7524   }
7525   set_inFreeRange(false);
7526   set_freeRangeInFreeLists(false);
7527 }
7528 
7529 // We take a break if we've been at this for a while,
7530 // so as to avoid monopolizing the locks involved.
7531 void SweepClosure::do_yield_work(HeapWord* addr) {
7532   // Return current free chunk being used for coalescing (if any)
7533   // to the appropriate freelist.  After yielding, the next
7534   // free block encountered will start a coalescing range of
7535   // free blocks.  If the next free block is adjacent to the
7536   // chunk just flushed, they will need to wait for the next
7537   // sweep to be coalesced.
7538   if (inFreeRange()) {
7539     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7540   }
7541 
7542   // First give up the locks, then yield, then re-lock.
7543   // We should probably use a constructor/destructor idiom to
7544   // do this unlock/lock or modify the MutexUnlocker class to
7545   // serve our purpose. XXX
7546   assert_lock_strong(_bitMap->lock());
7547   assert_lock_strong(_freelistLock);
7548   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7549          "CMS thread should hold CMS token");
7550   _bitMap->lock()->unlock();
7551   _freelistLock->unlock();
7552   ConcurrentMarkSweepThread::desynchronize(true);
7553   _collector->stopTimer();
7554   _collector->incrementYields();
7555 
7556   // See the comment in coordinator_yield()
7557   for (unsigned i = 0; i < CMSYieldSleepCount &&
7558                        ConcurrentMarkSweepThread::should_yield() &&
7559                        !CMSCollector::foregroundGCIsActive(); ++i) {
7560     os::sleep(Thread::current(), 1, false);
7561   }
7562 
7563   ConcurrentMarkSweepThread::synchronize(true);
7564   _freelistLock->lock();
7565   _bitMap->lock()->lock_without_safepoint_check();
7566   _collector->startTimer();
7567 }
7568 
7569 #ifndef PRODUCT
7570 // This is actually very useful in a product build if it can
7571 // be called from the debugger.  Compile it into the product
7572 // as needed.
7573 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7574   return debug_cms_space->verify_chunk_in_free_list(fc);
7575 }
7576 #endif
7577 
7578 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7579   log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7580                                p2i(fc), fc->size());
7581 }
7582 
7583 // CMSIsAliveClosure
7584 bool CMSIsAliveClosure::do_object_b(oop obj) {
7585   HeapWord* addr = (HeapWord*)obj;
7586   return addr != NULL &&
7587          (!_span.contains(addr) || _bit_map->isMarked(addr));
7588 }
7589 
7590 
7591 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7592                       MemRegion span,
7593                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7594                       bool cpc):
7595   _collector(collector),
7596   _span(span),
7597   _bit_map(bit_map),
7598   _mark_stack(mark_stack),
7599   _concurrent_precleaning(cpc) {
7600   assert(!_span.is_empty(), "Empty span could spell trouble");
7601 }
7602 
7603 
7604 // CMSKeepAliveClosure: the serial version
7605 void CMSKeepAliveClosure::do_oop(oop obj) {
7606   HeapWord* addr = (HeapWord*)obj;
7607   if (_span.contains(addr) &&
7608       !_bit_map->isMarked(addr)) {
7609     _bit_map->mark(addr);
7610     bool simulate_overflow = false;
7611     NOT_PRODUCT(
7612       if (CMSMarkStackOverflowALot &&
7613           _collector->simulate_overflow()) {
7614         // simulate a stack overflow
7615         simulate_overflow = true;
7616       }
7617     )
7618     if (simulate_overflow || !_mark_stack->push(obj)) {
7619       if (_concurrent_precleaning) {
7620         // We dirty the overflown object and let the remark
7621         // phase deal with it.
7622         assert(_collector->overflow_list_is_empty(), "Error");
7623         // In the case of object arrays, we need to dirty all of
7624         // the cards that the object spans. No locking or atomics
7625         // are needed since no one else can be mutating the mod union
7626         // table.
7627         if (obj->is_objArray()) {
7628           size_t sz = obj->size();
7629           HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
7630           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7631           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7632           _collector->_modUnionTable.mark_range(redirty_range);
7633         } else {
7634           _collector->_modUnionTable.mark(addr);
7635         }
7636         _collector->_ser_kac_preclean_ovflw++;
7637       } else {
7638         _collector->push_on_overflow_list(obj);
7639         _collector->_ser_kac_ovflw++;
7640       }
7641     }
7642   }
7643 }
7644 
7645 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
7646 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7647 
7648 // CMSParKeepAliveClosure: a parallel version of the above.
7649 // The work queues are private to each closure (thread),
7650 // but (may be) available for stealing by other threads.
7651 void CMSParKeepAliveClosure::do_oop(oop obj) {
7652   HeapWord* addr = (HeapWord*)obj;
7653   if (_span.contains(addr) &&
7654       !_bit_map->isMarked(addr)) {
7655     // In general, during recursive tracing, several threads
7656     // may be concurrently getting here; the first one to
7657     // "tag" it, claims it.
7658     if (_bit_map->par_mark(addr)) {
7659       bool res = _work_queue->push(obj);
7660       assert(res, "Low water mark should be much less than capacity");
7661       // Do a recursive trim in the hope that this will keep
7662       // stack usage lower, but leave some oops for potential stealers
7663       trim_queue(_low_water_mark);
7664     } // Else, another thread got there first
7665   }
7666 }
7667 
7668 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
7669 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7670 
7671 void CMSParKeepAliveClosure::trim_queue(uint max) {
7672   while (_work_queue->size() > max) {
7673     oop new_oop;
7674     if (_work_queue->pop_local(new_oop)) {
7675       assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7676       assert(_bit_map->isMarked((HeapWord*)new_oop),
7677              "no white objects on this stack!");
7678       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7679       // iterate over the oops in this oop, marking and pushing
7680       // the ones in CMS heap (i.e. in _span).
7681       new_oop->oop_iterate(&_mark_and_push);
7682     }
7683   }
7684 }
7685 
7686 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7687                                 CMSCollector* collector,
7688                                 MemRegion span, CMSBitMap* bit_map,
7689                                 OopTaskQueue* work_queue):
7690   _collector(collector),
7691   _span(span),
7692   _bit_map(bit_map),
7693   _work_queue(work_queue) { }
7694 
7695 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7696   HeapWord* addr = (HeapWord*)obj;
7697   if (_span.contains(addr) &&
7698       !_bit_map->isMarked(addr)) {
7699     if (_bit_map->par_mark(addr)) {
7700       bool simulate_overflow = false;
7701       NOT_PRODUCT(
7702         if (CMSMarkStackOverflowALot &&
7703             _collector->par_simulate_overflow()) {
7704           // simulate a stack overflow
7705           simulate_overflow = true;
7706         }
7707       )
7708       if (simulate_overflow || !_work_queue->push(obj)) {
7709         _collector->par_push_on_overflow_list(obj);
7710         _collector->_par_kac_ovflw++;
7711       }
7712     } // Else another thread got there already
7713   }
7714 }
7715 
7716 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7717 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7718 
7719 //////////////////////////////////////////////////////////////////
7720 //  CMSExpansionCause                /////////////////////////////
7721 //////////////////////////////////////////////////////////////////
7722 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7723   switch (cause) {
7724     case _no_expansion:
7725       return "No expansion";
7726     case _satisfy_free_ratio:
7727       return "Free ratio";
7728     case _satisfy_promotion:
7729       return "Satisfy promotion";
7730     case _satisfy_allocation:
7731       return "allocation";
7732     case _allocate_par_lab:
7733       return "Par LAB";
7734     case _allocate_par_spooling_space:
7735       return "Par Spooling Space";
7736     case _adaptive_size_policy:
7737       return "Ergonomics";
7738     default:
7739       return "unknown";
7740   }
7741 }
7742 
7743 void CMSDrainMarkingStackClosure::do_void() {
7744   // the max number to take from overflow list at a time
7745   const size_t num = _mark_stack->capacity()/4;
7746   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7747          "Overflow list should be NULL during concurrent phases");
7748   while (!_mark_stack->isEmpty() ||
7749          // if stack is empty, check the overflow list
7750          _collector->take_from_overflow_list(num, _mark_stack)) {
7751     oop obj = _mark_stack->pop();
7752     HeapWord* addr = (HeapWord*)obj;
7753     assert(_span.contains(addr), "Should be within span");
7754     assert(_bit_map->isMarked(addr), "Should be marked");
7755     assert(oopDesc::is_oop(obj), "Should be an oop");
7756     obj->oop_iterate(_keep_alive);
7757   }
7758 }
7759 
7760 void CMSParDrainMarkingStackClosure::do_void() {
7761   // drain queue
7762   trim_queue(0);
7763 }
7764 
7765 // Trim our work_queue so its length is below max at return
7766 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7767   while (_work_queue->size() > max) {
7768     oop new_oop;
7769     if (_work_queue->pop_local(new_oop)) {
7770       assert(oopDesc::is_oop(new_oop), "Expected an oop");
7771       assert(_bit_map->isMarked((HeapWord*)new_oop),
7772              "no white objects on this stack!");
7773       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7774       // iterate over the oops in this oop, marking and pushing
7775       // the ones in CMS heap (i.e. in _span).
7776       new_oop->oop_iterate(&_mark_and_push);
7777     }
7778   }
7779 }
7780 
7781 ////////////////////////////////////////////////////////////////////
7782 // Support for Marking Stack Overflow list handling and related code
7783 ////////////////////////////////////////////////////////////////////
7784 // Much of the following code is similar in shape and spirit to the
7785 // code used in ParNewGC. We should try and share that code
7786 // as much as possible in the future.
7787 
7788 #ifndef PRODUCT
7789 // Debugging support for CMSStackOverflowALot
7790 
7791 // It's OK to call this multi-threaded;  the worst thing
7792 // that can happen is that we'll get a bunch of closely
7793 // spaced simulated overflows, but that's OK, in fact
7794 // probably good as it would exercise the overflow code
7795 // under contention.
7796 bool CMSCollector::simulate_overflow() {
7797   if (_overflow_counter-- <= 0) { // just being defensive
7798     _overflow_counter = CMSMarkStackOverflowInterval;
7799     return true;
7800   } else {
7801     return false;
7802   }
7803 }
7804 
7805 bool CMSCollector::par_simulate_overflow() {
7806   return simulate_overflow();
7807 }
7808 #endif
7809 
7810 // Single-threaded
7811 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7812   assert(stack->isEmpty(), "Expected precondition");
7813   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7814   size_t i = num;
7815   oop  cur = _overflow_list;
7816   const markOop proto = markOopDesc::prototype();
7817   NOT_PRODUCT(ssize_t n = 0;)
7818   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7819     next = oop(cur->mark_raw());
7820     cur->set_mark_raw(proto);   // until proven otherwise
7821     assert(oopDesc::is_oop(cur), "Should be an oop");
7822     bool res = stack->push(cur);
7823     assert(res, "Bit off more than can chew?");
7824     NOT_PRODUCT(n++;)
7825   }
7826   _overflow_list = cur;
7827 #ifndef PRODUCT
7828   assert(_num_par_pushes >= n, "Too many pops?");
7829   _num_par_pushes -=n;
7830 #endif
7831   return !stack->isEmpty();
7832 }
7833 
7834 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
7835 // (MT-safe) Get a prefix of at most "num" from the list.
7836 // The overflow list is chained through the mark word of
7837 // each object in the list. We fetch the entire list,
7838 // break off a prefix of the right size and return the
7839 // remainder. If other threads try to take objects from
7840 // the overflow list at that time, they will wait for
7841 // some time to see if data becomes available. If (and
7842 // only if) another thread places one or more object(s)
7843 // on the global list before we have returned the suffix
7844 // to the global list, we will walk down our local list
7845 // to find its end and append the global list to
7846 // our suffix before returning it. This suffix walk can
7847 // prove to be expensive (quadratic in the amount of traffic)
7848 // when there are many objects in the overflow list and
7849 // there is much producer-consumer contention on the list.
7850 // *NOTE*: The overflow list manipulation code here and
7851 // in ParNewGeneration:: are very similar in shape,
7852 // except that in the ParNew case we use the old (from/eden)
7853 // copy of the object to thread the list via its klass word.
7854 // Because of the common code, if you make any changes in
7855 // the code below, please check the ParNew version to see if
7856 // similar changes might be needed.
7857 // CR 6797058 has been filed to consolidate the common code.
7858 bool CMSCollector::par_take_from_overflow_list(size_t num,
7859                                                OopTaskQueue* work_q,
7860                                                int no_of_gc_threads) {
7861   assert(work_q->size() == 0, "First empty local work queue");
7862   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7863   if (_overflow_list == NULL) {
7864     return false;
7865   }
7866   // Grab the entire list; we'll put back a suffix
7867   oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7868   Thread* tid = Thread::current();
7869   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7870   // set to ParallelGCThreads.
7871   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7872   size_t sleep_time_millis = MAX2((size_t)1, num/100);
7873   // If the list is busy, we spin for a short while,
7874   // sleeping between attempts to get the list.
7875   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7876     os::sleep(tid, sleep_time_millis, false);
7877     if (_overflow_list == NULL) {
7878       // Nothing left to take
7879       return false;
7880     } else if (_overflow_list != BUSY) {
7881       // Try and grab the prefix
7882       prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7883     }
7884   }
7885   // If the list was found to be empty, or we spun long
7886   // enough, we give up and return empty-handed. If we leave
7887   // the list in the BUSY state below, it must be the case that
7888   // some other thread holds the overflow list and will set it
7889   // to a non-BUSY state in the future.
7890   if (prefix == NULL || prefix == BUSY) {
7891      // Nothing to take or waited long enough
7892      if (prefix == NULL) {
7893        // Write back the NULL in case we overwrote it with BUSY above
7894        // and it is still the same value.
7895        Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7896      }
7897      return false;
7898   }
7899   assert(prefix != NULL && prefix != BUSY, "Error");
7900   size_t i = num;
7901   oop cur = prefix;
7902   // Walk down the first "num" objects, unless we reach the end.
7903   for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--);
7904   if (cur->mark_raw() == NULL) {
7905     // We have "num" or fewer elements in the list, so there
7906     // is nothing to return to the global list.
7907     // Write back the NULL in lieu of the BUSY we wrote
7908     // above, if it is still the same value.
7909     if (_overflow_list == BUSY) {
7910       Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7911     }
7912   } else {
7913     // Chop off the suffix and return it to the global list.
7914     assert(cur->mark_raw() != BUSY, "Error");
7915     oop suffix_head = cur->mark_raw(); // suffix will be put back on global list
7916     cur->set_mark_raw(NULL);           // break off suffix
7917     // It's possible that the list is still in the empty(busy) state
7918     // we left it in a short while ago; in that case we may be
7919     // able to place back the suffix without incurring the cost
7920     // of a walk down the list.
7921     oop observed_overflow_list = _overflow_list;
7922     oop cur_overflow_list = observed_overflow_list;
7923     bool attached = false;
7924     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7925       observed_overflow_list =
7926         Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7927       if (cur_overflow_list == observed_overflow_list) {
7928         attached = true;
7929         break;
7930       } else cur_overflow_list = observed_overflow_list;
7931     }
7932     if (!attached) {
7933       // Too bad, someone else sneaked in (at least) an element; we'll need
7934       // to do a splice. Find tail of suffix so we can prepend suffix to global
7935       // list.
7936       for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw()));
7937       oop suffix_tail = cur;
7938       assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL,
7939              "Tautology");
7940       observed_overflow_list = _overflow_list;
7941       do {
7942         cur_overflow_list = observed_overflow_list;
7943         if (cur_overflow_list != BUSY) {
7944           // Do the splice ...
7945           suffix_tail->set_mark_raw(markOop(cur_overflow_list));
7946         } else { // cur_overflow_list == BUSY
7947           suffix_tail->set_mark_raw(NULL);
7948         }
7949         // ... and try to place spliced list back on overflow_list ...
7950         observed_overflow_list =
7951           Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7952       } while (cur_overflow_list != observed_overflow_list);
7953       // ... until we have succeeded in doing so.
7954     }
7955   }
7956 
7957   // Push the prefix elements on work_q
7958   assert(prefix != NULL, "control point invariant");
7959   const markOop proto = markOopDesc::prototype();
7960   oop next;
7961   NOT_PRODUCT(ssize_t n = 0;)
7962   for (cur = prefix; cur != NULL; cur = next) {
7963     next = oop(cur->mark_raw());
7964     cur->set_mark_raw(proto);   // until proven otherwise
7965     assert(oopDesc::is_oop(cur), "Should be an oop");
7966     bool res = work_q->push(cur);
7967     assert(res, "Bit off more than we can chew?");
7968     NOT_PRODUCT(n++;)
7969   }
7970 #ifndef PRODUCT
7971   assert(_num_par_pushes >= n, "Too many pops?");
7972   Atomic::sub(n, &_num_par_pushes);
7973 #endif
7974   return true;
7975 }
7976 
7977 // Single-threaded
7978 void CMSCollector::push_on_overflow_list(oop p) {
7979   NOT_PRODUCT(_num_par_pushes++;)
7980   assert(oopDesc::is_oop(p), "Not an oop");
7981   preserve_mark_if_necessary(p);
7982   p->set_mark_raw((markOop)_overflow_list);
7983   _overflow_list = p;
7984 }
7985 
7986 // Multi-threaded; use CAS to prepend to overflow list
7987 void CMSCollector::par_push_on_overflow_list(oop p) {
7988   NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7989   assert(oopDesc::is_oop(p), "Not an oop");
7990   par_preserve_mark_if_necessary(p);
7991   oop observed_overflow_list = _overflow_list;
7992   oop cur_overflow_list;
7993   do {
7994     cur_overflow_list = observed_overflow_list;
7995     if (cur_overflow_list != BUSY) {
7996       p->set_mark_raw(markOop(cur_overflow_list));
7997     } else {
7998       p->set_mark_raw(NULL);
7999     }
8000     observed_overflow_list =
8001       Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
8002   } while (cur_overflow_list != observed_overflow_list);
8003 }
8004 #undef BUSY
8005 
8006 // Single threaded
8007 // General Note on GrowableArray: pushes may silently fail
8008 // because we are (temporarily) out of C-heap for expanding
8009 // the stack. The problem is quite ubiquitous and affects
8010 // a lot of code in the JVM. The prudent thing for GrowableArray
8011 // to do (for now) is to exit with an error. However, that may
8012 // be too draconian in some cases because the caller may be
8013 // able to recover without much harm. For such cases, we
8014 // should probably introduce a "soft_push" method which returns
8015 // an indication of success or failure with the assumption that
8016 // the caller may be able to recover from a failure; code in
8017 // the VM can then be changed, incrementally, to deal with such
8018 // failures where possible, thus, incrementally hardening the VM
8019 // in such low resource situations.
8020 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8021   _preserved_oop_stack.push(p);
8022   _preserved_mark_stack.push(m);
8023   assert(m == p->mark_raw(), "Mark word changed");
8024   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8025          "bijection");
8026 }
8027 
8028 // Single threaded
8029 void CMSCollector::preserve_mark_if_necessary(oop p) {
8030   markOop m = p->mark_raw();
8031   if (m->must_be_preserved(p)) {
8032     preserve_mark_work(p, m);
8033   }
8034 }
8035 
8036 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8037   markOop m = p->mark_raw();
8038   if (m->must_be_preserved(p)) {
8039     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8040     // Even though we read the mark word without holding
8041     // the lock, we are assured that it will not change
8042     // because we "own" this oop, so no other thread can
8043     // be trying to push it on the overflow list; see
8044     // the assertion in preserve_mark_work() that checks
8045     // that m == p->mark_raw().
8046     preserve_mark_work(p, m);
8047   }
8048 }
8049 
8050 // We should be able to do this multi-threaded,
8051 // a chunk of stack being a task (this is
8052 // correct because each oop only ever appears
8053 // once in the overflow list. However, it's
8054 // not very easy to completely overlap this with
8055 // other operations, so will generally not be done
8056 // until all work's been completed. Because we
8057 // expect the preserved oop stack (set) to be small,
8058 // it's probably fine to do this single-threaded.
8059 // We can explore cleverer concurrent/overlapped/parallel
8060 // processing of preserved marks if we feel the
8061 // need for this in the future. Stack overflow should
8062 // be so rare in practice and, when it happens, its
8063 // effect on performance so great that this will
8064 // likely just be in the noise anyway.
8065 void CMSCollector::restore_preserved_marks_if_any() {
8066   assert(SafepointSynchronize::is_at_safepoint(),
8067          "world should be stopped");
8068   assert(Thread::current()->is_ConcurrentGC_thread() ||
8069          Thread::current()->is_VM_thread(),
8070          "should be single-threaded");
8071   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8072          "bijection");
8073 
8074   while (!_preserved_oop_stack.is_empty()) {
8075     oop p = _preserved_oop_stack.pop();
8076     assert(oopDesc::is_oop(p), "Should be an oop");
8077     assert(_span.contains(p), "oop should be in _span");
8078     assert(p->mark_raw() == markOopDesc::prototype(),
8079            "Set when taken from overflow list");
8080     markOop m = _preserved_mark_stack.pop();
8081     p->set_mark_raw(m);
8082   }
8083   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8084          "stacks were cleared above");
8085 }
8086 
8087 #ifndef PRODUCT
8088 bool CMSCollector::no_preserved_marks() const {
8089   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8090 }
8091 #endif
8092 
8093 // Transfer some number of overflown objects to usual marking
8094 // stack. Return true if some objects were transferred.
8095 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8096   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8097                     (size_t)ParGCDesiredObjsFromOverflowList);
8098 
8099   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8100   assert(_collector->overflow_list_is_empty() || res,
8101          "If list is not empty, we should have taken something");
8102   assert(!res || !_mark_stack->isEmpty(),
8103          "If we took something, it should now be on our stack");
8104   return res;
8105 }
8106 
8107 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8108   size_t res = _sp->block_size_no_stall(addr, _collector);
8109   if (_sp->block_is_obj(addr)) {
8110     if (_live_bit_map->isMarked(addr)) {
8111       // It can't have been dead in a previous cycle
8112       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8113     } else {
8114       _dead_bit_map->mark(addr);      // mark the dead object
8115     }
8116   }
8117   // Could be 0, if the block size could not be computed without stalling.
8118   return res;
8119 }
8120 
8121 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8122   GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8123   switch (phase) {
8124     case CMSCollector::InitialMarking:
8125       initialize(manager /* GC manager */ ,
8126                  cause   /* cause of the GC */,
8127                  true    /* recordGCBeginTime */,
8128                  true    /* recordPreGCUsage */,
8129                  false   /* recordPeakUsage */,
8130                  false   /* recordPostGCusage */,
8131                  true    /* recordAccumulatedGCTime */,
8132                  false   /* recordGCEndTime */,
8133                  false   /* countCollection */  );
8134       break;
8135 
8136     case CMSCollector::FinalMarking:
8137       initialize(manager /* GC manager */ ,
8138                  cause   /* cause of the GC */,
8139                  false   /* recordGCBeginTime */,
8140                  false   /* recordPreGCUsage */,
8141                  false   /* recordPeakUsage */,
8142                  false   /* recordPostGCusage */,
8143                  true    /* recordAccumulatedGCTime */,
8144                  false   /* recordGCEndTime */,
8145                  false   /* countCollection */  );
8146       break;
8147 
8148     case CMSCollector::Sweeping:
8149       initialize(manager /* GC manager */ ,
8150                  cause   /* cause of the GC */,
8151                  false   /* recordGCBeginTime */,
8152                  false   /* recordPreGCUsage */,
8153                  true    /* recordPeakUsage */,
8154                  true    /* recordPostGCusage */,
8155                  false   /* recordAccumulatedGCTime */,
8156                  true    /* recordGCEndTime */,
8157                  true    /* countCollection */  );
8158       break;
8159 
8160     default:
8161       ShouldNotReachHere();
8162   }
8163 }