src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp

Wed, 27 Aug 2008 11:20:46 -0700

author
ysr
date
Wed, 27 Aug 2008 11:20:46 -0700
changeset 795
5d254928c888
parent 775
ebeb6490b814
parent 791
1ee8caae33af
child 887
00b023ae2d78
permissions
-rw-r--r--

Merge

     1 /*
     2  * Copyright 2001-2008 Sun Microsystems, Inc.  All Rights Reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
    20  * CA 95054 USA or visit www.sun.com if you need additional information or
    21  * have any questions.
    22  *
    23  */
    25 # include "incls/_precompiled.incl"
    26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl"
    28 // statics
    29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
    30 bool          CMSCollector::_full_gc_requested          = false;
    32 //////////////////////////////////////////////////////////////////
    33 // In support of CMS/VM thread synchronization
    34 //////////////////////////////////////////////////////////////////
    35 // We split use of the CGC_lock into 2 "levels".
    36 // The low-level locking is of the usual CGC_lock monitor. We introduce
    37 // a higher level "token" (hereafter "CMS token") built on top of the
    38 // low level monitor (hereafter "CGC lock").
    39 // The token-passing protocol gives priority to the VM thread. The
    40 // CMS-lock doesn't provide any fairness guarantees, but clients
    41 // should ensure that it is only held for very short, bounded
    42 // durations.
    43 //
    44 // When either of the CMS thread or the VM thread is involved in
    45 // collection operations during which it does not want the other
    46 // thread to interfere, it obtains the CMS token.
    47 //
    48 // If either thread tries to get the token while the other has
    49 // it, that thread waits. However, if the VM thread and CMS thread
    50 // both want the token, then the VM thread gets priority while the
    51 // CMS thread waits. This ensures, for instance, that the "concurrent"
    52 // phases of the CMS thread's work do not block out the VM thread
    53 // for long periods of time as the CMS thread continues to hog
    54 // the token. (See bug 4616232).
    55 //
    56 // The baton-passing functions are, however, controlled by the
    57 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
    58 // and here the low-level CMS lock, not the high level token,
    59 // ensures mutual exclusion.
    60 //
    61 // Two important conditions that we have to satisfy:
    62 // 1. if a thread does a low-level wait on the CMS lock, then it
    63 //    relinquishes the CMS token if it were holding that token
    64 //    when it acquired the low-level CMS lock.
    65 // 2. any low-level notifications on the low-level lock
    66 //    should only be sent when a thread has relinquished the token.
    67 //
    68 // In the absence of either property, we'd have potential deadlock.
    69 //
    70 // We protect each of the CMS (concurrent and sequential) phases
    71 // with the CMS _token_, not the CMS _lock_.
    72 //
    73 // The only code protected by CMS lock is the token acquisition code
    74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
    75 // baton-passing code.
    76 //
    77 // Unfortunately, i couldn't come up with a good abstraction to factor and
    78 // hide the naked CGC_lock manipulation in the baton-passing code
    79 // further below. That's something we should try to do. Also, the proof
    80 // of correctness of this 2-level locking scheme is far from obvious,
    81 // and potentially quite slippery. We have an uneasy supsicion, for instance,
    82 // that there may be a theoretical possibility of delay/starvation in the
    83 // low-level lock/wait/notify scheme used for the baton-passing because of
    84 // potential intereference with the priority scheme embodied in the
    85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
    86 // invocation further below and marked with "XXX 20011219YSR".
    87 // Indeed, as we note elsewhere, this may become yet more slippery
    88 // in the presence of multiple CMS and/or multiple VM threads. XXX
    90 class CMSTokenSync: public StackObj {
    91  private:
    92   bool _is_cms_thread;
    93  public:
    94   CMSTokenSync(bool is_cms_thread):
    95     _is_cms_thread(is_cms_thread) {
    96     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
    97            "Incorrect argument to constructor");
    98     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
    99   }
   101   ~CMSTokenSync() {
   102     assert(_is_cms_thread ?
   103              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
   104              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
   105           "Incorrect state");
   106     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
   107   }
   108 };
   110 // Convenience class that does a CMSTokenSync, and then acquires
   111 // upto three locks.
   112 class CMSTokenSyncWithLocks: public CMSTokenSync {
   113  private:
   114   // Note: locks are acquired in textual declaration order
   115   // and released in the opposite order
   116   MutexLockerEx _locker1, _locker2, _locker3;
   117  public:
   118   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
   119                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
   120     CMSTokenSync(is_cms_thread),
   121     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
   122     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
   123     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
   124   { }
   125 };
   128 // Wrapper class to temporarily disable icms during a foreground cms collection.
   129 class ICMSDisabler: public StackObj {
   130  public:
   131   // The ctor disables icms and wakes up the thread so it notices the change;
   132   // the dtor re-enables icms.  Note that the CMSCollector methods will check
   133   // CMSIncrementalMode.
   134   ICMSDisabler()  { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
   135   ~ICMSDisabler() { CMSCollector::enable_icms(); }
   136 };
   138 //////////////////////////////////////////////////////////////////
   139 //  Concurrent Mark-Sweep Generation /////////////////////////////
   140 //////////////////////////////////////////////////////////////////
   142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
   144 // This struct contains per-thread things necessary to support parallel
   145 // young-gen collection.
   146 class CMSParGCThreadState: public CHeapObj {
   147  public:
   148   CFLS_LAB lab;
   149   PromotionInfo promo;
   151   // Constructor.
   152   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
   153     promo.setSpace(cfls);
   154   }
   155 };
   157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
   158      ReservedSpace rs, size_t initial_byte_size, int level,
   159      CardTableRS* ct, bool use_adaptive_freelists,
   160      FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
   161   CardGeneration(rs, initial_byte_size, level, ct),
   162   _dilatation_factor(((double)MinChunkSize)/((double)(oopDesc::header_size()))),
   163   _debug_collection_type(Concurrent_collection_type)
   164 {
   165   HeapWord* bottom = (HeapWord*) _virtual_space.low();
   166   HeapWord* end    = (HeapWord*) _virtual_space.high();
   168   _direct_allocated_words = 0;
   169   NOT_PRODUCT(
   170     _numObjectsPromoted = 0;
   171     _numWordsPromoted = 0;
   172     _numObjectsAllocated = 0;
   173     _numWordsAllocated = 0;
   174   )
   176   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
   177                                            use_adaptive_freelists,
   178                                            dictionaryChoice);
   179   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
   180   if (_cmsSpace == NULL) {
   181     vm_exit_during_initialization(
   182       "CompactibleFreeListSpace allocation failure");
   183   }
   184   _cmsSpace->_gen = this;
   186   _gc_stats = new CMSGCStats();
   188   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
   189   // offsets match. The ability to tell free chunks from objects
   190   // depends on this property.
   191   debug_only(
   192     FreeChunk* junk = NULL;
   193     assert(UseCompressedOops ||
   194            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
   195            "Offset of FreeChunk::_prev within FreeChunk must match"
   196            "  that of OopDesc::_klass within OopDesc");
   197   )
   198   if (ParallelGCThreads > 0) {
   199     typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
   200     _par_gc_thread_states =
   201       NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
   202     if (_par_gc_thread_states == NULL) {
   203       vm_exit_during_initialization("Could not allocate par gc structs");
   204     }
   205     for (uint i = 0; i < ParallelGCThreads; i++) {
   206       _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
   207       if (_par_gc_thread_states[i] == NULL) {
   208         vm_exit_during_initialization("Could not allocate par gc structs");
   209       }
   210     }
   211   } else {
   212     _par_gc_thread_states = NULL;
   213   }
   214   _incremental_collection_failed = false;
   215   // The "dilatation_factor" is the expansion that can occur on
   216   // account of the fact that the minimum object size in the CMS
   217   // generation may be larger than that in, say, a contiguous young
   218   //  generation.
   219   // Ideally, in the calculation below, we'd compute the dilatation
   220   // factor as: MinChunkSize/(promoting_gen's min object size)
   221   // Since we do not have such a general query interface for the
   222   // promoting generation, we'll instead just use the mimimum
   223   // object size (which today is a header's worth of space);
   224   // note that all arithmetic is in units of HeapWords.
   225   assert(MinChunkSize >= oopDesc::header_size(), "just checking");
   226   assert(_dilatation_factor >= 1.0, "from previous assert");
   227 }
   230 // The field "_initiating_occupancy" represents the occupancy percentage
   231 // at which we trigger a new collection cycle.  Unless explicitly specified
   232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
   233 // is calculated by:
   234 //
   235 //   Let "f" be MinHeapFreeRatio in
   236 //
   237 //    _intiating_occupancy = 100-f +
   238 //                           f * (CMSTrigger[Perm]Ratio/100)
   239 //   where CMSTrigger[Perm]Ratio is the argument "tr" below.
   240 //
   241 // That is, if we assume the heap is at its desired maximum occupancy at the
   242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
   243 // space be allocated before initiating a new collection cycle.
   244 //
   245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
   246   assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
   247   if (io >= 0) {
   248     _initiating_occupancy = (double)io / 100.0;
   249   } else {
   250     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
   251                              (double)(tr * MinHeapFreeRatio) / 100.0)
   252                             / 100.0;
   253   }
   254 }
   257 void ConcurrentMarkSweepGeneration::ref_processor_init() {
   258   assert(collector() != NULL, "no collector");
   259   collector()->ref_processor_init();
   260 }
   262 void CMSCollector::ref_processor_init() {
   263   if (_ref_processor == NULL) {
   264     // Allocate and initialize a reference processor
   265     _ref_processor = ReferenceProcessor::create_ref_processor(
   266         _span,                               // span
   267         _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
   268         _cmsGen->refs_discovery_is_mt(),     // mt_discovery
   269         &_is_alive_closure,
   270         ParallelGCThreads,
   271         ParallelRefProcEnabled);
   272     // Initialize the _ref_processor field of CMSGen
   273     _cmsGen->set_ref_processor(_ref_processor);
   275     // Allocate a dummy ref processor for perm gen.
   276     ReferenceProcessor* rp2 = new ReferenceProcessor();
   277     if (rp2 == NULL) {
   278       vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
   279     }
   280     _permGen->set_ref_processor(rp2);
   281   }
   282 }
   284 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
   285   GenCollectedHeap* gch = GenCollectedHeap::heap();
   286   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
   287     "Wrong type of heap");
   288   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
   289     gch->gen_policy()->size_policy();
   290   assert(sp->is_gc_cms_adaptive_size_policy(),
   291     "Wrong type of size policy");
   292   return sp;
   293 }
   295 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
   296   CMSGCAdaptivePolicyCounters* results =
   297     (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
   298   assert(
   299     results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
   300     "Wrong gc policy counter kind");
   301   return results;
   302 }
   305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
   307   const char* gen_name = "old";
   309   // Generation Counters - generation 1, 1 subspace
   310   _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
   312   _space_counters = new GSpaceCounters(gen_name, 0,
   313                                        _virtual_space.reserved_size(),
   314                                        this, _gen_counters);
   315 }
   317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
   318   _cms_gen(cms_gen)
   319 {
   320   assert(alpha <= 100, "bad value");
   321   _saved_alpha = alpha;
   323   // Initialize the alphas to the bootstrap value of 100.
   324   _gc0_alpha = _cms_alpha = 100;
   326   _cms_begin_time.update();
   327   _cms_end_time.update();
   329   _gc0_duration = 0.0;
   330   _gc0_period = 0.0;
   331   _gc0_promoted = 0;
   333   _cms_duration = 0.0;
   334   _cms_period = 0.0;
   335   _cms_allocated = 0;
   337   _cms_used_at_gc0_begin = 0;
   338   _cms_used_at_gc0_end = 0;
   339   _allow_duty_cycle_reduction = false;
   340   _valid_bits = 0;
   341   _icms_duty_cycle = CMSIncrementalDutyCycle;
   342 }
   344 // If promotion failure handling is on use
   345 // the padded average size of the promotion for each
   346 // young generation collection.
   347 double CMSStats::time_until_cms_gen_full() const {
   348   size_t cms_free = _cms_gen->cmsSpace()->free();
   349   GenCollectedHeap* gch = GenCollectedHeap::heap();
   350   size_t expected_promotion = gch->get_gen(0)->capacity();
   351   if (HandlePromotionFailure) {
   352     expected_promotion = MIN2(
   353         (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
   354         expected_promotion);
   355   }
   356   if (cms_free > expected_promotion) {
   357     // Start a cms collection if there isn't enough space to promote
   358     // for the next minor collection.  Use the padded average as
   359     // a safety factor.
   360     cms_free -= expected_promotion;
   362     // Adjust by the safety factor.
   363     double cms_free_dbl = (double)cms_free;
   364     cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
   366     if (PrintGCDetails && Verbose) {
   367       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
   368         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
   369         cms_free, expected_promotion);
   370       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
   371         cms_free_dbl, cms_consumption_rate() + 1.0);
   372     }
   373     // Add 1 in case the consumption rate goes to zero.
   374     return cms_free_dbl / (cms_consumption_rate() + 1.0);
   375   }
   376   return 0.0;
   377 }
   379 // Compare the duration of the cms collection to the
   380 // time remaining before the cms generation is empty.
   381 // Note that the time from the start of the cms collection
   382 // to the start of the cms sweep (less than the total
   383 // duration of the cms collection) can be used.  This
   384 // has been tried and some applications experienced
   385 // promotion failures early in execution.  This was
   386 // possibly because the averages were not accurate
   387 // enough at the beginning.
   388 double CMSStats::time_until_cms_start() const {
   389   // We add "gc0_period" to the "work" calculation
   390   // below because this query is done (mostly) at the
   391   // end of a scavenge, so we need to conservatively
   392   // account for that much possible delay
   393   // in the query so as to avoid concurrent mode failures
   394   // due to starting the collection just a wee bit too
   395   // late.
   396   double work = cms_duration() + gc0_period();
   397   double deadline = time_until_cms_gen_full();
   398   if (work > deadline) {
   399     if (Verbose && PrintGCDetails) {
   400       gclog_or_tty->print(
   401         " CMSCollector: collect because of anticipated promotion "
   402         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
   403         gc0_period(), time_until_cms_gen_full());
   404     }
   405     return 0.0;
   406   }
   407   return work - deadline;
   408 }
   410 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
   411 // amount of change to prevent wild oscillation.
   412 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
   413                                               unsigned int new_duty_cycle) {
   414   assert(old_duty_cycle <= 100, "bad input value");
   415   assert(new_duty_cycle <= 100, "bad input value");
   417   // Note:  use subtraction with caution since it may underflow (values are
   418   // unsigned).  Addition is safe since we're in the range 0-100.
   419   unsigned int damped_duty_cycle = new_duty_cycle;
   420   if (new_duty_cycle < old_duty_cycle) {
   421     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
   422     if (new_duty_cycle + largest_delta < old_duty_cycle) {
   423       damped_duty_cycle = old_duty_cycle - largest_delta;
   424     }
   425   } else if (new_duty_cycle > old_duty_cycle) {
   426     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
   427     if (new_duty_cycle > old_duty_cycle + largest_delta) {
   428       damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
   429     }
   430   }
   431   assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
   433   if (CMSTraceIncrementalPacing) {
   434     gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
   435                            old_duty_cycle, new_duty_cycle, damped_duty_cycle);
   436   }
   437   return damped_duty_cycle;
   438 }
   440 unsigned int CMSStats::icms_update_duty_cycle_impl() {
   441   assert(CMSIncrementalPacing && valid(),
   442          "should be handled in icms_update_duty_cycle()");
   444   double cms_time_so_far = cms_timer().seconds();
   445   double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
   446   double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
   448   // Avoid division by 0.
   449   double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
   450   double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
   452   unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
   453   if (new_duty_cycle > _icms_duty_cycle) {
   454     // Avoid very small duty cycles (1 or 2); 0 is allowed.
   455     if (new_duty_cycle > 2) {
   456       _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
   457                                                 new_duty_cycle);
   458     }
   459   } else if (_allow_duty_cycle_reduction) {
   460     // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
   461     new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
   462     // Respect the minimum duty cycle.
   463     unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
   464     _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
   465   }
   467   if (PrintGCDetails || CMSTraceIncrementalPacing) {
   468     gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
   469   }
   471   _allow_duty_cycle_reduction = false;
   472   return _icms_duty_cycle;
   473 }
   475 #ifndef PRODUCT
   476 void CMSStats::print_on(outputStream *st) const {
   477   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
   478   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
   479                gc0_duration(), gc0_period(), gc0_promoted());
   480   st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
   481             cms_duration(), cms_duration_per_mb(),
   482             cms_period(), cms_allocated());
   483   st->print(",cms_since_beg=%g,cms_since_end=%g",
   484             cms_time_since_begin(), cms_time_since_end());
   485   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
   486             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
   487   if (CMSIncrementalMode) {
   488     st->print(",dc=%d", icms_duty_cycle());
   489   }
   491   if (valid()) {
   492     st->print(",promo_rate=%g,cms_alloc_rate=%g",
   493               promotion_rate(), cms_allocation_rate());
   494     st->print(",cms_consumption_rate=%g,time_until_full=%g",
   495               cms_consumption_rate(), time_until_cms_gen_full());
   496   }
   497   st->print(" ");
   498 }
   499 #endif // #ifndef PRODUCT
   501 CMSCollector::CollectorState CMSCollector::_collectorState =
   502                              CMSCollector::Idling;
   503 bool CMSCollector::_foregroundGCIsActive = false;
   504 bool CMSCollector::_foregroundGCShouldWait = false;
   506 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
   507                            ConcurrentMarkSweepGeneration* permGen,
   508                            CardTableRS*                   ct,
   509                            ConcurrentMarkSweepPolicy*     cp):
   510   _cmsGen(cmsGen),
   511   _permGen(permGen),
   512   _ct(ct),
   513   _ref_processor(NULL),    // will be set later
   514   _conc_workers(NULL),     // may be set later
   515   _abort_preclean(false),
   516   _start_sampling(false),
   517   _between_prologue_and_epilogue(false),
   518   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
   519   _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
   520   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
   521                  -1 /* lock-free */, "No_lock" /* dummy */),
   522   _modUnionClosure(&_modUnionTable),
   523   _modUnionClosurePar(&_modUnionTable),
   524   // Adjust my span to cover old (cms) gen and perm gen
   525   _span(cmsGen->reserved()._union(permGen->reserved())),
   526   // Construct the is_alive_closure with _span & markBitMap
   527   _is_alive_closure(_span, &_markBitMap),
   528   _restart_addr(NULL),
   529   _overflow_list(NULL),
   530   _preserved_oop_stack(NULL),
   531   _preserved_mark_stack(NULL),
   532   _stats(cmsGen),
   533   _eden_chunk_array(NULL),     // may be set in ctor body
   534   _eden_chunk_capacity(0),     // -- ditto --
   535   _eden_chunk_index(0),        // -- ditto --
   536   _survivor_plab_array(NULL),  // -- ditto --
   537   _survivor_chunk_array(NULL), // -- ditto --
   538   _survivor_chunk_capacity(0), // -- ditto --
   539   _survivor_chunk_index(0),    // -- ditto --
   540   _ser_pmc_preclean_ovflw(0),
   541   _ser_pmc_remark_ovflw(0),
   542   _par_pmc_remark_ovflw(0),
   543   _ser_kac_ovflw(0),
   544   _par_kac_ovflw(0),
   545 #ifndef PRODUCT
   546   _num_par_pushes(0),
   547 #endif
   548   _collection_count_start(0),
   549   _verifying(false),
   550   _icms_start_limit(NULL),
   551   _icms_stop_limit(NULL),
   552   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
   553   _completed_initialization(false),
   554   _collector_policy(cp),
   555   _should_unload_classes(false),
   556   _concurrent_cycles_since_last_unload(0),
   557   _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
   558 {
   559   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
   560     ExplicitGCInvokesConcurrent = true;
   561   }
   562   // Now expand the span and allocate the collection support structures
   563   // (MUT, marking bit map etc.) to cover both generations subject to
   564   // collection.
   566   // First check that _permGen is adjacent to _cmsGen and above it.
   567   assert(   _cmsGen->reserved().word_size()  > 0
   568          && _permGen->reserved().word_size() > 0,
   569          "generations should not be of zero size");
   570   assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
   571          "_cmsGen and _permGen should not overlap");
   572   assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
   573          "_cmsGen->end() different from _permGen->start()");
   575   // For use by dirty card to oop closures.
   576   _cmsGen->cmsSpace()->set_collector(this);
   577   _permGen->cmsSpace()->set_collector(this);
   579   // Allocate MUT and marking bit map
   580   {
   581     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
   582     if (!_markBitMap.allocate(_span)) {
   583       warning("Failed to allocate CMS Bit Map");
   584       return;
   585     }
   586     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
   587   }
   588   {
   589     _modUnionTable.allocate(_span);
   590     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
   591   }
   593   if (!_markStack.allocate(CMSMarkStackSize)) {
   594     warning("Failed to allocate CMS Marking Stack");
   595     return;
   596   }
   597   if (!_revisitStack.allocate(CMSRevisitStackSize)) {
   598     warning("Failed to allocate CMS Revisit Stack");
   599     return;
   600   }
   602   // Support for multi-threaded concurrent phases
   603   if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
   604     if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
   605       // just for now
   606       FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
   607     }
   608     if (ParallelCMSThreads > 1) {
   609       _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
   610                                  ParallelCMSThreads, true);
   611       if (_conc_workers == NULL) {
   612         warning("GC/CMS: _conc_workers allocation failure: "
   613               "forcing -CMSConcurrentMTEnabled");
   614         CMSConcurrentMTEnabled = false;
   615       }
   616     } else {
   617       CMSConcurrentMTEnabled = false;
   618     }
   619   }
   620   if (!CMSConcurrentMTEnabled) {
   621     ParallelCMSThreads = 0;
   622   } else {
   623     // Turn off CMSCleanOnEnter optimization temporarily for
   624     // the MT case where it's not fixed yet; see 6178663.
   625     CMSCleanOnEnter = false;
   626   }
   627   assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
   628          "Inconsistency");
   630   // Parallel task queues; these are shared for the
   631   // concurrent and stop-world phases of CMS, but
   632   // are not shared with parallel scavenge (ParNew).
   633   {
   634     uint i;
   635     uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
   637     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
   638          || ParallelRefProcEnabled)
   639         && num_queues > 0) {
   640       _task_queues = new OopTaskQueueSet(num_queues);
   641       if (_task_queues == NULL) {
   642         warning("task_queues allocation failure.");
   643         return;
   644       }
   645       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
   646       if (_hash_seed == NULL) {
   647         warning("_hash_seed array allocation failure");
   648         return;
   649       }
   651       // XXX use a global constant instead of 64!
   652       typedef struct OopTaskQueuePadded {
   653         OopTaskQueue work_queue;
   654         char pad[64 - sizeof(OopTaskQueue)];  // prevent false sharing
   655       } OopTaskQueuePadded;
   657       for (i = 0; i < num_queues; i++) {
   658         OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
   659         if (q_padded == NULL) {
   660           warning("work_queue allocation failure.");
   661           return;
   662         }
   663         _task_queues->register_queue(i, &q_padded->work_queue);
   664       }
   665       for (i = 0; i < num_queues; i++) {
   666         _task_queues->queue(i)->initialize();
   667         _hash_seed[i] = 17;  // copied from ParNew
   668       }
   669     }
   670   }
   672   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
   673   _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
   675   // Clip CMSBootstrapOccupancy between 0 and 100.
   676   _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
   677                          /(double)100;
   679   _full_gcs_since_conc_gc = 0;
   681   // Now tell CMS generations the identity of their collector
   682   ConcurrentMarkSweepGeneration::set_collector(this);
   684   // Create & start a CMS thread for this CMS collector
   685   _cmsThread = ConcurrentMarkSweepThread::start(this);
   686   assert(cmsThread() != NULL, "CMS Thread should have been created");
   687   assert(cmsThread()->collector() == this,
   688          "CMS Thread should refer to this gen");
   689   assert(CGC_lock != NULL, "Where's the CGC_lock?");
   691   // Support for parallelizing young gen rescan
   692   GenCollectedHeap* gch = GenCollectedHeap::heap();
   693   _young_gen = gch->prev_gen(_cmsGen);
   694   if (gch->supports_inline_contig_alloc()) {
   695     _top_addr = gch->top_addr();
   696     _end_addr = gch->end_addr();
   697     assert(_young_gen != NULL, "no _young_gen");
   698     _eden_chunk_index = 0;
   699     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
   700     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
   701     if (_eden_chunk_array == NULL) {
   702       _eden_chunk_capacity = 0;
   703       warning("GC/CMS: _eden_chunk_array allocation failure");
   704     }
   705   }
   706   assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
   708   // Support for parallelizing survivor space rescan
   709   if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
   710     size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
   711     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
   712     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
   713     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
   714     if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
   715         || _cursor == NULL) {
   716       warning("Failed to allocate survivor plab/chunk array");
   717       if (_survivor_plab_array  != NULL) {
   718         FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
   719         _survivor_plab_array = NULL;
   720       }
   721       if (_survivor_chunk_array != NULL) {
   722         FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
   723         _survivor_chunk_array = NULL;
   724       }
   725       if (_cursor != NULL) {
   726         FREE_C_HEAP_ARRAY(size_t, _cursor);
   727         _cursor = NULL;
   728       }
   729     } else {
   730       _survivor_chunk_capacity = 2*max_plab_samples;
   731       for (uint i = 0; i < ParallelGCThreads; i++) {
   732         HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
   733         if (vec == NULL) {
   734           warning("Failed to allocate survivor plab array");
   735           for (int j = i; j > 0; j--) {
   736             FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
   737           }
   738           FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
   739           FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
   740           _survivor_plab_array = NULL;
   741           _survivor_chunk_array = NULL;
   742           _survivor_chunk_capacity = 0;
   743           break;
   744         } else {
   745           ChunkArray* cur =
   746             ::new (&_survivor_plab_array[i]) ChunkArray(vec,
   747                                                         max_plab_samples);
   748           assert(cur->end() == 0, "Should be 0");
   749           assert(cur->array() == vec, "Should be vec");
   750           assert(cur->capacity() == max_plab_samples, "Error");
   751         }
   752       }
   753     }
   754   }
   755   assert(   (   _survivor_plab_array  != NULL
   756              && _survivor_chunk_array != NULL)
   757          || (   _survivor_chunk_capacity == 0
   758              && _survivor_chunk_index == 0),
   759          "Error");
   761   // Choose what strong roots should be scanned depending on verification options
   762   // and perm gen collection mode.
   763   if (!CMSClassUnloadingEnabled) {
   764     // If class unloading is disabled we want to include all classes into the root set.
   765     add_root_scanning_option(SharedHeap::SO_AllClasses);
   766   } else {
   767     add_root_scanning_option(SharedHeap::SO_SystemClasses);
   768   }
   770   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
   771   _gc_counters = new CollectorCounters("CMS", 1);
   772   _completed_initialization = true;
   773   _sweep_timer.start();  // start of time
   774 }
   776 const char* ConcurrentMarkSweepGeneration::name() const {
   777   return "concurrent mark-sweep generation";
   778 }
   779 void ConcurrentMarkSweepGeneration::update_counters() {
   780   if (UsePerfData) {
   781     _space_counters->update_all();
   782     _gen_counters->update_all();
   783   }
   784 }
   786 // this is an optimized version of update_counters(). it takes the
   787 // used value as a parameter rather than computing it.
   788 //
   789 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
   790   if (UsePerfData) {
   791     _space_counters->update_used(used);
   792     _space_counters->update_capacity();
   793     _gen_counters->update_all();
   794   }
   795 }
   797 void ConcurrentMarkSweepGeneration::print() const {
   798   Generation::print();
   799   cmsSpace()->print();
   800 }
   802 #ifndef PRODUCT
   803 void ConcurrentMarkSweepGeneration::print_statistics() {
   804   cmsSpace()->printFLCensus(0);
   805 }
   806 #endif
   808 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
   809   GenCollectedHeap* gch = GenCollectedHeap::heap();
   810   if (PrintGCDetails) {
   811     if (Verbose) {
   812       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
   813         level(), short_name(), s, used(), capacity());
   814     } else {
   815       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
   816         level(), short_name(), s, used() / K, capacity() / K);
   817     }
   818   }
   819   if (Verbose) {
   820     gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
   821               gch->used(), gch->capacity());
   822   } else {
   823     gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
   824               gch->used() / K, gch->capacity() / K);
   825   }
   826 }
   828 size_t
   829 ConcurrentMarkSweepGeneration::contiguous_available() const {
   830   // dld proposes an improvement in precision here. If the committed
   831   // part of the space ends in a free block we should add that to
   832   // uncommitted size in the calculation below. Will make this
   833   // change later, staying with the approximation below for the
   834   // time being. -- ysr.
   835   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
   836 }
   838 size_t
   839 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
   840   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
   841 }
   843 size_t ConcurrentMarkSweepGeneration::max_available() const {
   844   return free() + _virtual_space.uncommitted_size();
   845 }
   847 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
   848     size_t max_promotion_in_bytes,
   849     bool younger_handles_promotion_failure) const {
   851   // This is the most conservative test.  Full promotion is
   852   // guaranteed if this is used. The multiplicative factor is to
   853   // account for the worst case "dilatation".
   854   double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
   855   if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
   856     adjusted_max_promo_bytes = (double)max_uintx;
   857   }
   858   bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
   860   if (younger_handles_promotion_failure && !result) {
   861     // Full promotion is not guaranteed because fragmentation
   862     // of the cms generation can prevent the full promotion.
   863     result = (max_available() >= (size_t)adjusted_max_promo_bytes);
   865     if (!result) {
   866       // With promotion failure handling the test for the ability
   867       // to support the promotion does not have to be guaranteed.
   868       // Use an average of the amount promoted.
   869       result = max_available() >= (size_t)
   870         gc_stats()->avg_promoted()->padded_average();
   871       if (PrintGC && Verbose && result) {
   872         gclog_or_tty->print_cr(
   873           "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
   874           " max_available: " SIZE_FORMAT
   875           " avg_promoted: " SIZE_FORMAT,
   876           max_available(), (size_t)
   877           gc_stats()->avg_promoted()->padded_average());
   878       }
   879     } else {
   880       if (PrintGC && Verbose) {
   881         gclog_or_tty->print_cr(
   882           "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
   883           " max_available: " SIZE_FORMAT
   884           " adj_max_promo_bytes: " SIZE_FORMAT,
   885           max_available(), (size_t)adjusted_max_promo_bytes);
   886       }
   887     }
   888   } else {
   889     if (PrintGC && Verbose) {
   890       gclog_or_tty->print_cr(
   891         "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
   892         " contiguous_available: " SIZE_FORMAT
   893         " adj_max_promo_bytes: " SIZE_FORMAT,
   894         max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
   895     }
   896   }
   897   return result;
   898 }
   900 CompactibleSpace*
   901 ConcurrentMarkSweepGeneration::first_compaction_space() const {
   902   return _cmsSpace;
   903 }
   905 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
   906   // Clear the promotion information.  These pointers can be adjusted
   907   // along with all the other pointers into the heap but
   908   // compaction is expected to be a rare event with
   909   // a heap using cms so don't do it without seeing the need.
   910   if (ParallelGCThreads > 0) {
   911     for (uint i = 0; i < ParallelGCThreads; i++) {
   912       _par_gc_thread_states[i]->promo.reset();
   913     }
   914   }
   915 }
   917 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
   918   blk->do_space(_cmsSpace);
   919 }
   921 void ConcurrentMarkSweepGeneration::compute_new_size() {
   922   assert_locked_or_safepoint(Heap_lock);
   924   // If incremental collection failed, we just want to expand
   925   // to the limit.
   926   if (incremental_collection_failed()) {
   927     clear_incremental_collection_failed();
   928     grow_to_reserved();
   929     return;
   930   }
   932   size_t expand_bytes = 0;
   933   double free_percentage = ((double) free()) / capacity();
   934   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
   935   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
   937   // compute expansion delta needed for reaching desired free percentage
   938   if (free_percentage < desired_free_percentage) {
   939     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   940     assert(desired_capacity >= capacity(), "invalid expansion size");
   941     expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
   942   }
   943   if (expand_bytes > 0) {
   944     if (PrintGCDetails && Verbose) {
   945       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   946       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
   947       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
   948       gclog_or_tty->print_cr("  Desired free fraction %f",
   949         desired_free_percentage);
   950       gclog_or_tty->print_cr("  Maximum free fraction %f",
   951         maximum_free_percentage);
   952       gclog_or_tty->print_cr("  Capactiy "SIZE_FORMAT, capacity()/1000);
   953       gclog_or_tty->print_cr("  Desired capacity "SIZE_FORMAT,
   954         desired_capacity/1000);
   955       int prev_level = level() - 1;
   956       if (prev_level >= 0) {
   957         size_t prev_size = 0;
   958         GenCollectedHeap* gch = GenCollectedHeap::heap();
   959         Generation* prev_gen = gch->_gens[prev_level];
   960         prev_size = prev_gen->capacity();
   961           gclog_or_tty->print_cr("  Younger gen size "SIZE_FORMAT,
   962                                  prev_size/1000);
   963       }
   964       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc "SIZE_FORMAT,
   965         unsafe_max_alloc_nogc()/1000);
   966       gclog_or_tty->print_cr("  contiguous available "SIZE_FORMAT,
   967         contiguous_available()/1000);
   968       gclog_or_tty->print_cr("  Expand by "SIZE_FORMAT" (bytes)",
   969         expand_bytes);
   970     }
   971     // safe if expansion fails
   972     expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
   973     if (PrintGCDetails && Verbose) {
   974       gclog_or_tty->print_cr("  Expanded free fraction %f",
   975         ((double) free()) / capacity());
   976     }
   977   }
   978 }
   980 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
   981   return cmsSpace()->freelistLock();
   982 }
   984 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
   985                                                   bool   tlab) {
   986   CMSSynchronousYieldRequest yr;
   987   MutexLockerEx x(freelistLock(),
   988                   Mutex::_no_safepoint_check_flag);
   989   return have_lock_and_allocate(size, tlab);
   990 }
   992 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
   993                                                   bool   tlab) {
   994   assert_lock_strong(freelistLock());
   995   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
   996   HeapWord* res = cmsSpace()->allocate(adjustedSize);
   997   // Allocate the object live (grey) if the background collector has
   998   // started marking. This is necessary because the marker may
   999   // have passed this address and consequently this object will
  1000   // not otherwise be greyed and would be incorrectly swept up.
  1001   // Note that if this object contains references, the writing
  1002   // of those references will dirty the card containing this object
  1003   // allowing the object to be blackened (and its references scanned)
  1004   // either during a preclean phase or at the final checkpoint.
  1005   if (res != NULL) {
  1006     collector()->direct_allocated(res, adjustedSize);
  1007     _direct_allocated_words += adjustedSize;
  1008     // allocation counters
  1009     NOT_PRODUCT(
  1010       _numObjectsAllocated++;
  1011       _numWordsAllocated += (int)adjustedSize;
  1014   return res;
  1017 // In the case of direct allocation by mutators in a generation that
  1018 // is being concurrently collected, the object must be allocated
  1019 // live (grey) if the background collector has started marking.
  1020 // This is necessary because the marker may
  1021 // have passed this address and consequently this object will
  1022 // not otherwise be greyed and would be incorrectly swept up.
  1023 // Note that if this object contains references, the writing
  1024 // of those references will dirty the card containing this object
  1025 // allowing the object to be blackened (and its references scanned)
  1026 // either during a preclean phase or at the final checkpoint.
  1027 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
  1028   assert(_markBitMap.covers(start, size), "Out of bounds");
  1029   if (_collectorState >= Marking) {
  1030     MutexLockerEx y(_markBitMap.lock(),
  1031                     Mutex::_no_safepoint_check_flag);
  1032     // [see comments preceding SweepClosure::do_blk() below for details]
  1033     // 1. need to mark the object as live so it isn't collected
  1034     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
  1035     // 3. need to mark the end of the object so sweeper can skip over it
  1036     //    if it's uninitialized when the sweeper reaches it.
  1037     _markBitMap.mark(start);          // object is live
  1038     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
  1039     _markBitMap.mark(start + size - 1);
  1040                                       // mark end of object
  1042   // check that oop looks uninitialized
  1043   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
  1046 void CMSCollector::promoted(bool par, HeapWord* start,
  1047                             bool is_obj_array, size_t obj_size) {
  1048   assert(_markBitMap.covers(start), "Out of bounds");
  1049   // See comment in direct_allocated() about when objects should
  1050   // be allocated live.
  1051   if (_collectorState >= Marking) {
  1052     // we already hold the marking bit map lock, taken in
  1053     // the prologue
  1054     if (par) {
  1055       _markBitMap.par_mark(start);
  1056     } else {
  1057       _markBitMap.mark(start);
  1059     // We don't need to mark the object as uninitialized (as
  1060     // in direct_allocated above) because this is being done with the
  1061     // world stopped and the object will be initialized by the
  1062     // time the sweeper gets to look at it.
  1063     assert(SafepointSynchronize::is_at_safepoint(),
  1064            "expect promotion only at safepoints");
  1066     if (_collectorState < Sweeping) {
  1067       // Mark the appropriate cards in the modUnionTable, so that
  1068       // this object gets scanned before the sweep. If this is
  1069       // not done, CMS generation references in the object might
  1070       // not get marked.
  1071       // For the case of arrays, which are otherwise precisely
  1072       // marked, we need to dirty the entire array, not just its head.
  1073       if (is_obj_array) {
  1074         // The [par_]mark_range() method expects mr.end() below to
  1075         // be aligned to the granularity of a bit's representation
  1076         // in the heap. In the case of the MUT below, that's a
  1077         // card size.
  1078         MemRegion mr(start,
  1079                      (HeapWord*)round_to((intptr_t)(start + obj_size),
  1080                         CardTableModRefBS::card_size /* bytes */));
  1081         if (par) {
  1082           _modUnionTable.par_mark_range(mr);
  1083         } else {
  1084           _modUnionTable.mark_range(mr);
  1086       } else {  // not an obj array; we can just mark the head
  1087         if (par) {
  1088           _modUnionTable.par_mark(start);
  1089         } else {
  1090           _modUnionTable.mark(start);
  1097 static inline size_t percent_of_space(Space* space, HeapWord* addr)
  1099   size_t delta = pointer_delta(addr, space->bottom());
  1100   return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
  1103 void CMSCollector::icms_update_allocation_limits()
  1105   Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
  1106   EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
  1108   const unsigned int duty_cycle = stats().icms_update_duty_cycle();
  1109   if (CMSTraceIncrementalPacing) {
  1110     stats().print();
  1113   assert(duty_cycle <= 100, "invalid duty cycle");
  1114   if (duty_cycle != 0) {
  1115     // The duty_cycle is a percentage between 0 and 100; convert to words and
  1116     // then compute the offset from the endpoints of the space.
  1117     size_t free_words = eden->free() / HeapWordSize;
  1118     double free_words_dbl = (double)free_words;
  1119     size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
  1120     size_t offset_words = (free_words - duty_cycle_words) / 2;
  1122     _icms_start_limit = eden->top() + offset_words;
  1123     _icms_stop_limit = eden->end() - offset_words;
  1125     // The limits may be adjusted (shifted to the right) by
  1126     // CMSIncrementalOffset, to allow the application more mutator time after a
  1127     // young gen gc (when all mutators were stopped) and before CMS starts and
  1128     // takes away one or more cpus.
  1129     if (CMSIncrementalOffset != 0) {
  1130       double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
  1131       size_t adjustment = (size_t)adjustment_dbl;
  1132       HeapWord* tmp_stop = _icms_stop_limit + adjustment;
  1133       if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
  1134         _icms_start_limit += adjustment;
  1135         _icms_stop_limit = tmp_stop;
  1139   if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
  1140     _icms_start_limit = _icms_stop_limit = eden->end();
  1143   // Install the new start limit.
  1144   eden->set_soft_end(_icms_start_limit);
  1146   if (CMSTraceIncrementalMode) {
  1147     gclog_or_tty->print(" icms alloc limits:  "
  1148                            PTR_FORMAT "," PTR_FORMAT
  1149                            " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
  1150                            _icms_start_limit, _icms_stop_limit,
  1151                            percent_of_space(eden, _icms_start_limit),
  1152                            percent_of_space(eden, _icms_stop_limit));
  1153     if (Verbose) {
  1154       gclog_or_tty->print("eden:  ");
  1155       eden->print_on(gclog_or_tty);
  1160 // Any changes here should try to maintain the invariant
  1161 // that if this method is called with _icms_start_limit
  1162 // and _icms_stop_limit both NULL, then it should return NULL
  1163 // and not notify the icms thread.
  1164 HeapWord*
  1165 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
  1166                                        size_t word_size)
  1168   // A start_limit equal to end() means the duty cycle is 0, so treat that as a
  1169   // nop.
  1170   if (CMSIncrementalMode && _icms_start_limit != space->end()) {
  1171     if (top <= _icms_start_limit) {
  1172       if (CMSTraceIncrementalMode) {
  1173         space->print_on(gclog_or_tty);
  1174         gclog_or_tty->stamp();
  1175         gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
  1176                                ", new limit=" PTR_FORMAT
  1177                                " (" SIZE_FORMAT "%%)",
  1178                                top, _icms_stop_limit,
  1179                                percent_of_space(space, _icms_stop_limit));
  1181       ConcurrentMarkSweepThread::start_icms();
  1182       assert(top < _icms_stop_limit, "Tautology");
  1183       if (word_size < pointer_delta(_icms_stop_limit, top)) {
  1184         return _icms_stop_limit;
  1187       // The allocation will cross both the _start and _stop limits, so do the
  1188       // stop notification also and return end().
  1189       if (CMSTraceIncrementalMode) {
  1190         space->print_on(gclog_or_tty);
  1191         gclog_or_tty->stamp();
  1192         gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
  1193                                ", new limit=" PTR_FORMAT
  1194                                " (" SIZE_FORMAT "%%)",
  1195                                top, space->end(),
  1196                                percent_of_space(space, space->end()));
  1198       ConcurrentMarkSweepThread::stop_icms();
  1199       return space->end();
  1202     if (top <= _icms_stop_limit) {
  1203       if (CMSTraceIncrementalMode) {
  1204         space->print_on(gclog_or_tty);
  1205         gclog_or_tty->stamp();
  1206         gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
  1207                                ", new limit=" PTR_FORMAT
  1208                                " (" SIZE_FORMAT "%%)",
  1209                                top, space->end(),
  1210                                percent_of_space(space, space->end()));
  1212       ConcurrentMarkSweepThread::stop_icms();
  1213       return space->end();
  1216     if (CMSTraceIncrementalMode) {
  1217       space->print_on(gclog_or_tty);
  1218       gclog_or_tty->stamp();
  1219       gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
  1220                              ", new limit=" PTR_FORMAT,
  1221                              top, NULL);
  1225   return NULL;
  1228 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
  1229   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
  1230   // allocate, copy and if necessary update promoinfo --
  1231   // delegate to underlying space.
  1232   assert_lock_strong(freelistLock());
  1234 #ifndef PRODUCT
  1235   if (Universe::heap()->promotion_should_fail()) {
  1236     return NULL;
  1238 #endif  // #ifndef PRODUCT
  1240   oop res = _cmsSpace->promote(obj, obj_size);
  1241   if (res == NULL) {
  1242     // expand and retry
  1243     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
  1244     expand(s*HeapWordSize, MinHeapDeltaBytes,
  1245       CMSExpansionCause::_satisfy_promotion);
  1246     // Since there's currently no next generation, we don't try to promote
  1247     // into a more senior generation.
  1248     assert(next_gen() == NULL, "assumption, based upon which no attempt "
  1249                                "is made to pass on a possibly failing "
  1250                                "promotion to next generation");
  1251     res = _cmsSpace->promote(obj, obj_size);
  1253   if (res != NULL) {
  1254     // See comment in allocate() about when objects should
  1255     // be allocated live.
  1256     assert(obj->is_oop(), "Will dereference klass pointer below");
  1257     collector()->promoted(false,           // Not parallel
  1258                           (HeapWord*)res, obj->is_objArray(), obj_size);
  1259     // promotion counters
  1260     NOT_PRODUCT(
  1261       _numObjectsPromoted++;
  1262       _numWordsPromoted +=
  1263         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
  1266   return res;
  1270 HeapWord*
  1271 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
  1272                                              HeapWord* top,
  1273                                              size_t word_sz)
  1275   return collector()->allocation_limit_reached(space, top, word_sz);
  1278 // Things to support parallel young-gen collection.
  1279 oop
  1280 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
  1281                                            oop old, markOop m,
  1282                                            size_t word_sz) {
  1283 #ifndef PRODUCT
  1284   if (Universe::heap()->promotion_should_fail()) {
  1285     return NULL;
  1287 #endif  // #ifndef PRODUCT
  1289   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1290   PromotionInfo* promoInfo = &ps->promo;
  1291   // if we are tracking promotions, then first ensure space for
  1292   // promotion (including spooling space for saving header if necessary).
  1293   // then allocate and copy, then track promoted info if needed.
  1294   // When tracking (see PromotionInfo::track()), the mark word may
  1295   // be displaced and in this case restoration of the mark word
  1296   // occurs in the (oop_since_save_marks_)iterate phase.
  1297   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
  1298     // Out of space for allocating spooling buffers;
  1299     // try expanding and allocating spooling buffers.
  1300     if (!expand_and_ensure_spooling_space(promoInfo)) {
  1301       return NULL;
  1304   assert(promoInfo->has_spooling_space(), "Control point invariant");
  1305   HeapWord* obj_ptr = ps->lab.alloc(word_sz);
  1306   if (obj_ptr == NULL) {
  1307      obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
  1308      if (obj_ptr == NULL) {
  1309        return NULL;
  1312   oop obj = oop(obj_ptr);
  1313   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1314   // Otherwise, copy the object.  Here we must be careful to insert the
  1315   // klass pointer last, since this marks the block as an allocated object.
  1316   // Except with compressed oops it's the mark word.
  1317   HeapWord* old_ptr = (HeapWord*)old;
  1318   if (word_sz > (size_t)oopDesc::header_size()) {
  1319     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
  1320                                  obj_ptr + oopDesc::header_size(),
  1321                                  word_sz - oopDesc::header_size());
  1324   if (UseCompressedOops) {
  1325     // Copy gap missed by (aligned) header size calculation above
  1326     obj->set_klass_gap(old->klass_gap());
  1329   // Restore the mark word copied above.
  1330   obj->set_mark(m);
  1332   // Now we can track the promoted object, if necessary.  We take care
  1333   // To delay the transition from uninitialized to full object
  1334   // (i.e., insertion of klass pointer) until after, so that it
  1335   // atomically becomes a promoted object.
  1336   if (promoInfo->tracking()) {
  1337     promoInfo->track((PromotedObject*)obj, old->klass());
  1340   // Finally, install the klass pointer (this should be volatile).
  1341   obj->set_klass(old->klass());
  1343   assert(old->is_oop(), "Will dereference klass ptr below");
  1344   collector()->promoted(true,          // parallel
  1345                         obj_ptr, old->is_objArray(), word_sz);
  1347   NOT_PRODUCT(
  1348     Atomic::inc(&_numObjectsPromoted);
  1349     Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
  1350                 &_numWordsPromoted);
  1353   return obj;
  1356 void
  1357 ConcurrentMarkSweepGeneration::
  1358 par_promote_alloc_undo(int thread_num,
  1359                        HeapWord* obj, size_t word_sz) {
  1360   // CMS does not support promotion undo.
  1361   ShouldNotReachHere();
  1364 void
  1365 ConcurrentMarkSweepGeneration::
  1366 par_promote_alloc_done(int thread_num) {
  1367   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1368   ps->lab.retire();
  1369 #if CFLS_LAB_REFILL_STATS
  1370   if (thread_num == 0) {
  1371     _cmsSpace->print_par_alloc_stats();
  1373 #endif
  1376 void
  1377 ConcurrentMarkSweepGeneration::
  1378 par_oop_since_save_marks_iterate_done(int thread_num) {
  1379   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1380   ParScanWithoutBarrierClosure* dummy_cl = NULL;
  1381   ps->promo.promoted_oops_iterate_nv(dummy_cl);
  1384 // XXXPERM
  1385 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
  1386                                                    size_t size,
  1387                                                    bool   tlab)
  1389   // We allow a STW collection only if a full
  1390   // collection was requested.
  1391   return full || should_allocate(size, tlab); // FIX ME !!!
  1392   // This and promotion failure handling are connected at the
  1393   // hip and should be fixed by untying them.
  1396 bool CMSCollector::shouldConcurrentCollect() {
  1397   if (_full_gc_requested) {
  1398     assert(ExplicitGCInvokesConcurrent, "Unexpected state");
  1399     if (Verbose && PrintGCDetails) {
  1400       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
  1401                              " gc request");
  1403     return true;
  1406   // For debugging purposes, change the type of collection.
  1407   // If the rotation is not on the concurrent collection
  1408   // type, don't start a concurrent collection.
  1409   NOT_PRODUCT(
  1410     if (RotateCMSCollectionTypes &&
  1411         (_cmsGen->debug_collection_type() !=
  1412           ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
  1413       assert(_cmsGen->debug_collection_type() !=
  1414         ConcurrentMarkSweepGeneration::Unknown_collection_type,
  1415         "Bad cms collection type");
  1416       return false;
  1420   FreelistLocker x(this);
  1421   // ------------------------------------------------------------------
  1422   // Print out lots of information which affects the initiation of
  1423   // a collection.
  1424   if (PrintCMSInitiationStatistics && stats().valid()) {
  1425     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
  1426     gclog_or_tty->stamp();
  1427     gclog_or_tty->print_cr("");
  1428     stats().print_on(gclog_or_tty);
  1429     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
  1430       stats().time_until_cms_gen_full());
  1431     gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
  1432     gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
  1433                            _cmsGen->contiguous_available());
  1434     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
  1435     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
  1436     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
  1437     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
  1438     gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
  1440   // ------------------------------------------------------------------
  1442   // If the estimated time to complete a cms collection (cms_duration())
  1443   // is less than the estimated time remaining until the cms generation
  1444   // is full, start a collection.
  1445   if (!UseCMSInitiatingOccupancyOnly) {
  1446     if (stats().valid()) {
  1447       if (stats().time_until_cms_start() == 0.0) {
  1448         return true;
  1450     } else {
  1451       // We want to conservatively collect somewhat early in order
  1452       // to try and "bootstrap" our CMS/promotion statistics;
  1453       // this branch will not fire after the first successful CMS
  1454       // collection because the stats should then be valid.
  1455       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
  1456         if (Verbose && PrintGCDetails) {
  1457           gclog_or_tty->print_cr(
  1458             " CMSCollector: collect for bootstrapping statistics:"
  1459             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
  1460             _bootstrap_occupancy);
  1462         return true;
  1467   // Otherwise, we start a collection cycle if either the perm gen or
  1468   // old gen want a collection cycle started. Each may use
  1469   // an appropriate criterion for making this decision.
  1470   // XXX We need to make sure that the gen expansion
  1471   // criterion dovetails well with this. XXX NEED TO FIX THIS
  1472   if (_cmsGen->should_concurrent_collect()) {
  1473     if (Verbose && PrintGCDetails) {
  1474       gclog_or_tty->print_cr("CMS old gen initiated");
  1476     return true;
  1479   // We start a collection if we believe an incremental collection may fail;
  1480   // this is not likely to be productive in practice because it's probably too
  1481   // late anyway.
  1482   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1483   assert(gch->collector_policy()->is_two_generation_policy(),
  1484          "You may want to check the correctness of the following");
  1485   if (gch->incremental_collection_will_fail()) {
  1486     if (PrintGCDetails && Verbose) {
  1487       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
  1489     return true;
  1492   if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
  1493     bool res = update_should_unload_classes();
  1494     if (res) {
  1495       if (Verbose && PrintGCDetails) {
  1496         gclog_or_tty->print_cr("CMS perm gen initiated");
  1498       return true;
  1501   return false;
  1504 // Clear _expansion_cause fields of constituent generations
  1505 void CMSCollector::clear_expansion_cause() {
  1506   _cmsGen->clear_expansion_cause();
  1507   _permGen->clear_expansion_cause();
  1510 // We should be conservative in starting a collection cycle.  To
  1511 // start too eagerly runs the risk of collecting too often in the
  1512 // extreme.  To collect too rarely falls back on full collections,
  1513 // which works, even if not optimum in terms of concurrent work.
  1514 // As a work around for too eagerly collecting, use the flag
  1515 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
  1516 // giving the user an easily understandable way of controlling the
  1517 // collections.
  1518 // We want to start a new collection cycle if any of the following
  1519 // conditions hold:
  1520 // . our current occupancy exceeds the configured initiating occupancy
  1521 //   for this generation, or
  1522 // . we recently needed to expand this space and have not, since that
  1523 //   expansion, done a collection of this generation, or
  1524 // . the underlying space believes that it may be a good idea to initiate
  1525 //   a concurrent collection (this may be based on criteria such as the
  1526 //   following: the space uses linear allocation and linear allocation is
  1527 //   going to fail, or there is believed to be excessive fragmentation in
  1528 //   the generation, etc... or ...
  1529 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
  1530 //   the case of the old generation, not the perm generation; see CR 6543076):
  1531 //   we may be approaching a point at which allocation requests may fail because
  1532 //   we will be out of sufficient free space given allocation rate estimates.]
  1533 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
  1535   assert_lock_strong(freelistLock());
  1536   if (occupancy() > initiating_occupancy()) {
  1537     if (PrintGCDetails && Verbose) {
  1538       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
  1539         short_name(), occupancy(), initiating_occupancy());
  1541     return true;
  1543   if (UseCMSInitiatingOccupancyOnly) {
  1544     return false;
  1546   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
  1547     if (PrintGCDetails && Verbose) {
  1548       gclog_or_tty->print(" %s: collect because expanded for allocation ",
  1549         short_name());
  1551     return true;
  1553   if (_cmsSpace->should_concurrent_collect()) {
  1554     if (PrintGCDetails && Verbose) {
  1555       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
  1556         short_name());
  1558     return true;
  1560   return false;
  1563 void ConcurrentMarkSweepGeneration::collect(bool   full,
  1564                                             bool   clear_all_soft_refs,
  1565                                             size_t size,
  1566                                             bool   tlab)
  1568   collector()->collect(full, clear_all_soft_refs, size, tlab);
  1571 void CMSCollector::collect(bool   full,
  1572                            bool   clear_all_soft_refs,
  1573                            size_t size,
  1574                            bool   tlab)
  1576   if (!UseCMSCollectionPassing && _collectorState > Idling) {
  1577     // For debugging purposes skip the collection if the state
  1578     // is not currently idle
  1579     if (TraceCMSState) {
  1580       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
  1581         Thread::current(), full, _collectorState);
  1583     return;
  1586   // The following "if" branch is present for defensive reasons.
  1587   // In the current uses of this interface, it can be replaced with:
  1588   // assert(!GC_locker.is_active(), "Can't be called otherwise");
  1589   // But I am not placing that assert here to allow future
  1590   // generality in invoking this interface.
  1591   if (GC_locker::is_active()) {
  1592     // A consistency test for GC_locker
  1593     assert(GC_locker::needs_gc(), "Should have been set already");
  1594     // Skip this foreground collection, instead
  1595     // expanding the heap if necessary.
  1596     // Need the free list locks for the call to free() in compute_new_size()
  1597     compute_new_size();
  1598     return;
  1600   acquire_control_and_collect(full, clear_all_soft_refs);
  1601   _full_gcs_since_conc_gc++;
  1605 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
  1606   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1607   unsigned int gc_count = gch->total_full_collections();
  1608   if (gc_count == full_gc_count) {
  1609     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
  1610     _full_gc_requested = true;
  1611     CGC_lock->notify();   // nudge CMS thread
  1616 // The foreground and background collectors need to coordinate in order
  1617 // to make sure that they do not mutually interfere with CMS collections.
  1618 // When a background collection is active,
  1619 // the foreground collector may need to take over (preempt) and
  1620 // synchronously complete an ongoing collection. Depending on the
  1621 // frequency of the background collections and the heap usage
  1622 // of the application, this preemption can be seldom or frequent.
  1623 // There are only certain
  1624 // points in the background collection that the "collection-baton"
  1625 // can be passed to the foreground collector.
  1626 //
  1627 // The foreground collector will wait for the baton before
  1628 // starting any part of the collection.  The foreground collector
  1629 // will only wait at one location.
  1630 //
  1631 // The background collector will yield the baton before starting a new
  1632 // phase of the collection (e.g., before initial marking, marking from roots,
  1633 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
  1634 // of the loop which switches the phases. The background collector does some
  1635 // of the phases (initial mark, final re-mark) with the world stopped.
  1636 // Because of locking involved in stopping the world,
  1637 // the foreground collector should not block waiting for the background
  1638 // collector when it is doing a stop-the-world phase.  The background
  1639 // collector will yield the baton at an additional point just before
  1640 // it enters a stop-the-world phase.  Once the world is stopped, the
  1641 // background collector checks the phase of the collection.  If the
  1642 // phase has not changed, it proceeds with the collection.  If the
  1643 // phase has changed, it skips that phase of the collection.  See
  1644 // the comments on the use of the Heap_lock in collect_in_background().
  1645 //
  1646 // Variable used in baton passing.
  1647 //   _foregroundGCIsActive - Set to true by the foreground collector when
  1648 //      it wants the baton.  The foreground clears it when it has finished
  1649 //      the collection.
  1650 //   _foregroundGCShouldWait - Set to true by the background collector
  1651 //        when it is running.  The foreground collector waits while
  1652 //      _foregroundGCShouldWait is true.
  1653 //  CGC_lock - monitor used to protect access to the above variables
  1654 //      and to notify the foreground and background collectors.
  1655 //  _collectorState - current state of the CMS collection.
  1656 //
  1657 // The foreground collector
  1658 //   acquires the CGC_lock
  1659 //   sets _foregroundGCIsActive
  1660 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
  1661 //     various locks acquired in preparation for the collection
  1662 //     are released so as not to block the background collector
  1663 //     that is in the midst of a collection
  1664 //   proceeds with the collection
  1665 //   clears _foregroundGCIsActive
  1666 //   returns
  1667 //
  1668 // The background collector in a loop iterating on the phases of the
  1669 //      collection
  1670 //   acquires the CGC_lock
  1671 //   sets _foregroundGCShouldWait
  1672 //   if _foregroundGCIsActive is set
  1673 //     clears _foregroundGCShouldWait, notifies _CGC_lock
  1674 //     waits on _CGC_lock for _foregroundGCIsActive to become false
  1675 //     and exits the loop.
  1676 //   otherwise
  1677 //     proceed with that phase of the collection
  1678 //     if the phase is a stop-the-world phase,
  1679 //       yield the baton once more just before enqueueing
  1680 //       the stop-world CMS operation (executed by the VM thread).
  1681 //   returns after all phases of the collection are done
  1682 //
  1684 void CMSCollector::acquire_control_and_collect(bool full,
  1685         bool clear_all_soft_refs) {
  1686   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
  1687   assert(!Thread::current()->is_ConcurrentGC_thread(),
  1688          "shouldn't try to acquire control from self!");
  1690   // Start the protocol for acquiring control of the
  1691   // collection from the background collector (aka CMS thread).
  1692   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1693          "VM thread should have CMS token");
  1694   // Remember the possibly interrupted state of an ongoing
  1695   // concurrent collection
  1696   CollectorState first_state = _collectorState;
  1698   // Signal to a possibly ongoing concurrent collection that
  1699   // we want to do a foreground collection.
  1700   _foregroundGCIsActive = true;
  1702   // Disable incremental mode during a foreground collection.
  1703   ICMSDisabler icms_disabler;
  1705   // release locks and wait for a notify from the background collector
  1706   // releasing the locks in only necessary for phases which
  1707   // do yields to improve the granularity of the collection.
  1708   assert_lock_strong(bitMapLock());
  1709   // We need to lock the Free list lock for the space that we are
  1710   // currently collecting.
  1711   assert(haveFreelistLocks(), "Must be holding free list locks");
  1712   bitMapLock()->unlock();
  1713   releaseFreelistLocks();
  1715     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  1716     if (_foregroundGCShouldWait) {
  1717       // We are going to be waiting for action for the CMS thread;
  1718       // it had better not be gone (for instance at shutdown)!
  1719       assert(ConcurrentMarkSweepThread::cmst() != NULL,
  1720              "CMS thread must be running");
  1721       // Wait here until the background collector gives us the go-ahead
  1722       ConcurrentMarkSweepThread::clear_CMS_flag(
  1723         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
  1724       // Get a possibly blocked CMS thread going:
  1725       //   Note that we set _foregroundGCIsActive true above,
  1726       //   without protection of the CGC_lock.
  1727       CGC_lock->notify();
  1728       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
  1729              "Possible deadlock");
  1730       while (_foregroundGCShouldWait) {
  1731         // wait for notification
  1732         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  1733         // Possibility of delay/starvation here, since CMS token does
  1734         // not know to give priority to VM thread? Actually, i think
  1735         // there wouldn't be any delay/starvation, but the proof of
  1736         // that "fact" (?) appears non-trivial. XXX 20011219YSR
  1738       ConcurrentMarkSweepThread::set_CMS_flag(
  1739         ConcurrentMarkSweepThread::CMS_vm_has_token);
  1742   // The CMS_token is already held.  Get back the other locks.
  1743   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1744          "VM thread should have CMS token");
  1745   getFreelistLocks();
  1746   bitMapLock()->lock_without_safepoint_check();
  1747   if (TraceCMSState) {
  1748     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
  1749       INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
  1750     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
  1753   // Check if we need to do a compaction, or if not, whether
  1754   // we need to start the mark-sweep from scratch.
  1755   bool should_compact    = false;
  1756   bool should_start_over = false;
  1757   decide_foreground_collection_type(clear_all_soft_refs,
  1758     &should_compact, &should_start_over);
  1760 NOT_PRODUCT(
  1761   if (RotateCMSCollectionTypes) {
  1762     if (_cmsGen->debug_collection_type() ==
  1763         ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
  1764       should_compact = true;
  1765     } else if (_cmsGen->debug_collection_type() ==
  1766                ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
  1767       should_compact = false;
  1772   if (PrintGCDetails && first_state > Idling) {
  1773     GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
  1774     if (GCCause::is_user_requested_gc(cause) ||
  1775         GCCause::is_serviceability_requested_gc(cause)) {
  1776       gclog_or_tty->print(" (concurrent mode interrupted)");
  1777     } else {
  1778       gclog_or_tty->print(" (concurrent mode failure)");
  1782   if (should_compact) {
  1783     // If the collection is being acquired from the background
  1784     // collector, there may be references on the discovered
  1785     // references lists that have NULL referents (being those
  1786     // that were concurrently cleared by a mutator) or
  1787     // that are no longer active (having been enqueued concurrently
  1788     // by the mutator).
  1789     // Scrub the list of those references because Mark-Sweep-Compact
  1790     // code assumes referents are not NULL and that all discovered
  1791     // Reference objects are active.
  1792     ref_processor()->clean_up_discovered_references();
  1794     do_compaction_work(clear_all_soft_refs);
  1796     // Has the GC time limit been exceeded?
  1797     check_gc_time_limit();
  1799   } else {
  1800     do_mark_sweep_work(clear_all_soft_refs, first_state,
  1801       should_start_over);
  1803   // Reset the expansion cause, now that we just completed
  1804   // a collection cycle.
  1805   clear_expansion_cause();
  1806   _foregroundGCIsActive = false;
  1807   return;
  1810 void CMSCollector::check_gc_time_limit() {
  1812   // Ignore explicit GC's.  Exiting here does not set the flag and
  1813   // does not reset the count.  Updating of the averages for system
  1814   // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
  1815   GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
  1816   if (GCCause::is_user_requested_gc(gc_cause) ||
  1817       GCCause::is_serviceability_requested_gc(gc_cause)) {
  1818     return;
  1821   // Calculate the fraction of the CMS generation was freed during
  1822   // the last collection.
  1823   // Only consider the STW compacting cost for now.
  1824   //
  1825   // Note that the gc time limit test only works for the collections
  1826   // of the young gen + tenured gen and not for collections of the
  1827   // permanent gen.  That is because the calculation of the space
  1828   // freed by the collection is the free space in the young gen +
  1829   // tenured gen.
  1831   double fraction_free =
  1832     ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
  1833   if ((100.0 * size_policy()->compacting_gc_cost()) >
  1834          ((double) GCTimeLimit) &&
  1835         ((fraction_free * 100) < GCHeapFreeLimit)) {
  1836     size_policy()->inc_gc_time_limit_count();
  1837     if (UseGCOverheadLimit &&
  1838         (size_policy()->gc_time_limit_count() >
  1839          AdaptiveSizePolicyGCTimeLimitThreshold)) {
  1840       size_policy()->set_gc_time_limit_exceeded(true);
  1841       // Avoid consecutive OOM due to the gc time limit by resetting
  1842       // the counter.
  1843       size_policy()->reset_gc_time_limit_count();
  1844       if (PrintGCDetails) {
  1845         gclog_or_tty->print_cr("      GC is exceeding overhead limit "
  1846           "of %d%%", GCTimeLimit);
  1848     } else {
  1849       if (PrintGCDetails) {
  1850         gclog_or_tty->print_cr("      GC would exceed overhead limit "
  1851           "of %d%%", GCTimeLimit);
  1854   } else {
  1855     size_policy()->reset_gc_time_limit_count();
  1859 // Resize the perm generation and the tenured generation
  1860 // after obtaining the free list locks for the
  1861 // two generations.
  1862 void CMSCollector::compute_new_size() {
  1863   assert_locked_or_safepoint(Heap_lock);
  1864   FreelistLocker z(this);
  1865   _permGen->compute_new_size();
  1866   _cmsGen->compute_new_size();
  1869 // A work method used by foreground collection to determine
  1870 // what type of collection (compacting or not, continuing or fresh)
  1871 // it should do.
  1872 // NOTE: the intent is to make UseCMSCompactAtFullCollection
  1873 // and CMSCompactWhenClearAllSoftRefs the default in the future
  1874 // and do away with the flags after a suitable period.
  1875 void CMSCollector::decide_foreground_collection_type(
  1876   bool clear_all_soft_refs, bool* should_compact,
  1877   bool* should_start_over) {
  1878   // Normally, we'll compact only if the UseCMSCompactAtFullCollection
  1879   // flag is set, and we have either requested a System.gc() or
  1880   // the number of full gc's since the last concurrent cycle
  1881   // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
  1882   // or if an incremental collection has failed
  1883   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1884   assert(gch->collector_policy()->is_two_generation_policy(),
  1885          "You may want to check the correctness of the following");
  1886   // Inform cms gen if this was due to partial collection failing.
  1887   // The CMS gen may use this fact to determine its expansion policy.
  1888   if (gch->incremental_collection_will_fail()) {
  1889     assert(!_cmsGen->incremental_collection_failed(),
  1890            "Should have been noticed, reacted to and cleared");
  1891     _cmsGen->set_incremental_collection_failed();
  1893   *should_compact =
  1894     UseCMSCompactAtFullCollection &&
  1895     ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
  1896      GCCause::is_user_requested_gc(gch->gc_cause()) ||
  1897      gch->incremental_collection_will_fail());
  1898   *should_start_over = false;
  1899   if (clear_all_soft_refs && !*should_compact) {
  1900     // We are about to do a last ditch collection attempt
  1901     // so it would normally make sense to do a compaction
  1902     // to reclaim as much space as possible.
  1903     if (CMSCompactWhenClearAllSoftRefs) {
  1904       // Default: The rationale is that in this case either
  1905       // we are past the final marking phase, in which case
  1906       // we'd have to start over, or so little has been done
  1907       // that there's little point in saving that work. Compaction
  1908       // appears to be the sensible choice in either case.
  1909       *should_compact = true;
  1910     } else {
  1911       // We have been asked to clear all soft refs, but not to
  1912       // compact. Make sure that we aren't past the final checkpoint
  1913       // phase, for that is where we process soft refs. If we are already
  1914       // past that phase, we'll need to redo the refs discovery phase and
  1915       // if necessary clear soft refs that weren't previously
  1916       // cleared. We do so by remembering the phase in which
  1917       // we came in, and if we are past the refs processing
  1918       // phase, we'll choose to just redo the mark-sweep
  1919       // collection from scratch.
  1920       if (_collectorState > FinalMarking) {
  1921         // We are past the refs processing phase;
  1922         // start over and do a fresh synchronous CMS cycle
  1923         _collectorState = Resetting; // skip to reset to start new cycle
  1924         reset(false /* == !asynch */);
  1925         *should_start_over = true;
  1926       } // else we can continue a possibly ongoing current cycle
  1931 // A work method used by the foreground collector to do
  1932 // a mark-sweep-compact.
  1933 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
  1934   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1935   TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
  1936   if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
  1937     gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
  1938       "collections passed to foreground collector", _full_gcs_since_conc_gc);
  1941   // Sample collection interval time and reset for collection pause.
  1942   if (UseAdaptiveSizePolicy) {
  1943     size_policy()->msc_collection_begin();
  1946   // Temporarily widen the span of the weak reference processing to
  1947   // the entire heap.
  1948   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
  1949   ReferenceProcessorSpanMutator x(ref_processor(), new_span);
  1951   // Temporarily, clear the "is_alive_non_header" field of the
  1952   // reference processor.
  1953   ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
  1955   // Temporarily make reference _processing_ single threaded (non-MT).
  1956   ReferenceProcessorMTProcMutator z(ref_processor(), false);
  1958   // Temporarily make refs discovery atomic
  1959   ReferenceProcessorAtomicMutator w(ref_processor(), true);
  1961   ref_processor()->set_enqueuing_is_done(false);
  1962   ref_processor()->enable_discovery();
  1963   // If an asynchronous collection finishes, the _modUnionTable is
  1964   // all clear.  If we are assuming the collection from an asynchronous
  1965   // collection, clear the _modUnionTable.
  1966   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
  1967     "_modUnionTable should be clear if the baton was not passed");
  1968   _modUnionTable.clear_all();
  1970   // We must adjust the allocation statistics being maintained
  1971   // in the free list space. We do so by reading and clearing
  1972   // the sweep timer and updating the block flux rate estimates below.
  1973   assert(_sweep_timer.is_active(), "We should never see the timer inactive");
  1974   _sweep_timer.stop();
  1975   // Note that we do not use this sample to update the _sweep_estimate.
  1976   _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
  1977                                           _sweep_estimate.padded_average());
  1979   GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
  1980     ref_processor(), clear_all_soft_refs);
  1981   #ifdef ASSERT
  1982     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
  1983     size_t free_size = cms_space->free();
  1984     assert(free_size ==
  1985            pointer_delta(cms_space->end(), cms_space->compaction_top())
  1986            * HeapWordSize,
  1987       "All the free space should be compacted into one chunk at top");
  1988     assert(cms_space->dictionary()->totalChunkSize(
  1989                                       debug_only(cms_space->freelistLock())) == 0 ||
  1990            cms_space->totalSizeInIndexedFreeLists() == 0,
  1991       "All the free space should be in a single chunk");
  1992     size_t num = cms_space->totalCount();
  1993     assert((free_size == 0 && num == 0) ||
  1994            (free_size > 0  && (num == 1 || num == 2)),
  1995          "There should be at most 2 free chunks after compaction");
  1996   #endif // ASSERT
  1997   _collectorState = Resetting;
  1998   assert(_restart_addr == NULL,
  1999          "Should have been NULL'd before baton was passed");
  2000   reset(false /* == !asynch */);
  2001   _cmsGen->reset_after_compaction();
  2002   _concurrent_cycles_since_last_unload = 0;
  2004   if (verifying() && !should_unload_classes()) {
  2005     perm_gen_verify_bit_map()->clear_all();
  2008   // Clear any data recorded in the PLAB chunk arrays.
  2009   if (_survivor_plab_array != NULL) {
  2010     reset_survivor_plab_arrays();
  2013   // Adjust the per-size allocation stats for the next epoch.
  2014   _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
  2015   // Restart the "sweep timer" for next epoch.
  2016   _sweep_timer.reset();
  2017   _sweep_timer.start();
  2019   // Sample collection pause time and reset for collection interval.
  2020   if (UseAdaptiveSizePolicy) {
  2021     size_policy()->msc_collection_end(gch->gc_cause());
  2024   // For a mark-sweep-compact, compute_new_size() will be called
  2025   // in the heap's do_collection() method.
  2028 // A work method used by the foreground collector to do
  2029 // a mark-sweep, after taking over from a possibly on-going
  2030 // concurrent mark-sweep collection.
  2031 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
  2032   CollectorState first_state, bool should_start_over) {
  2033   if (PrintGC && Verbose) {
  2034     gclog_or_tty->print_cr("Pass concurrent collection to foreground "
  2035       "collector with count %d",
  2036       _full_gcs_since_conc_gc);
  2038   switch (_collectorState) {
  2039     case Idling:
  2040       if (first_state == Idling || should_start_over) {
  2041         // The background GC was not active, or should
  2042         // restarted from scratch;  start the cycle.
  2043         _collectorState = InitialMarking;
  2045       // If first_state was not Idling, then a background GC
  2046       // was in progress and has now finished.  No need to do it
  2047       // again.  Leave the state as Idling.
  2048       break;
  2049     case Precleaning:
  2050       // In the foreground case don't do the precleaning since
  2051       // it is not done concurrently and there is extra work
  2052       // required.
  2053       _collectorState = FinalMarking;
  2055   if (PrintGCDetails &&
  2056       (_collectorState > Idling ||
  2057        !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
  2058     gclog_or_tty->print(" (concurrent mode failure)");
  2060   collect_in_foreground(clear_all_soft_refs);
  2062   // For a mark-sweep, compute_new_size() will be called
  2063   // in the heap's do_collection() method.
  2067 void CMSCollector::getFreelistLocks() const {
  2068   // Get locks for all free lists in all generations that this
  2069   // collector is responsible for
  2070   _cmsGen->freelistLock()->lock_without_safepoint_check();
  2071   _permGen->freelistLock()->lock_without_safepoint_check();
  2074 void CMSCollector::releaseFreelistLocks() const {
  2075   // Release locks for all free lists in all generations that this
  2076   // collector is responsible for
  2077   _cmsGen->freelistLock()->unlock();
  2078   _permGen->freelistLock()->unlock();
  2081 bool CMSCollector::haveFreelistLocks() const {
  2082   // Check locks for all free lists in all generations that this
  2083   // collector is responsible for
  2084   assert_lock_strong(_cmsGen->freelistLock());
  2085   assert_lock_strong(_permGen->freelistLock());
  2086   PRODUCT_ONLY(ShouldNotReachHere());
  2087   return true;
  2090 // A utility class that is used by the CMS collector to
  2091 // temporarily "release" the foreground collector from its
  2092 // usual obligation to wait for the background collector to
  2093 // complete an ongoing phase before proceeding.
  2094 class ReleaseForegroundGC: public StackObj {
  2095  private:
  2096   CMSCollector* _c;
  2097  public:
  2098   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
  2099     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
  2100     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2101     // allow a potentially blocked foreground collector to proceed
  2102     _c->_foregroundGCShouldWait = false;
  2103     if (_c->_foregroundGCIsActive) {
  2104       CGC_lock->notify();
  2106     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2107            "Possible deadlock");
  2110   ~ReleaseForegroundGC() {
  2111     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
  2112     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2113     _c->_foregroundGCShouldWait = true;
  2115 };
  2117 // There are separate collect_in_background and collect_in_foreground because of
  2118 // the different locking requirements of the background collector and the
  2119 // foreground collector.  There was originally an attempt to share
  2120 // one "collect" method between the background collector and the foreground
  2121 // collector but the if-then-else required made it cleaner to have
  2122 // separate methods.
  2123 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
  2124   assert(Thread::current()->is_ConcurrentGC_thread(),
  2125     "A CMS asynchronous collection is only allowed on a CMS thread.");
  2127   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2129     bool safepoint_check = Mutex::_no_safepoint_check_flag;
  2130     MutexLockerEx hl(Heap_lock, safepoint_check);
  2131     FreelistLocker fll(this);
  2132     MutexLockerEx x(CGC_lock, safepoint_check);
  2133     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
  2134       // The foreground collector is active or we're
  2135       // not using asynchronous collections.  Skip this
  2136       // background collection.
  2137       assert(!_foregroundGCShouldWait, "Should be clear");
  2138       return;
  2139     } else {
  2140       assert(_collectorState == Idling, "Should be idling before start.");
  2141       _collectorState = InitialMarking;
  2142       // Reset the expansion cause, now that we are about to begin
  2143       // a new cycle.
  2144       clear_expansion_cause();
  2146     // Decide if we want to enable class unloading as part of the
  2147     // ensuing concurrent GC cycle.
  2148     update_should_unload_classes();
  2149     _full_gc_requested = false;           // acks all outstanding full gc requests
  2150     // Signal that we are about to start a collection
  2151     gch->increment_total_full_collections();  // ... starting a collection cycle
  2152     _collection_count_start = gch->total_full_collections();
  2155   // Used for PrintGC
  2156   size_t prev_used;
  2157   if (PrintGC && Verbose) {
  2158     prev_used = _cmsGen->used(); // XXXPERM
  2161   // The change of the collection state is normally done at this level;
  2162   // the exceptions are phases that are executed while the world is
  2163   // stopped.  For those phases the change of state is done while the
  2164   // world is stopped.  For baton passing purposes this allows the
  2165   // background collector to finish the phase and change state atomically.
  2166   // The foreground collector cannot wait on a phase that is done
  2167   // while the world is stopped because the foreground collector already
  2168   // has the world stopped and would deadlock.
  2169   while (_collectorState != Idling) {
  2170     if (TraceCMSState) {
  2171       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2172         Thread::current(), _collectorState);
  2174     // The foreground collector
  2175     //   holds the Heap_lock throughout its collection.
  2176     //   holds the CMS token (but not the lock)
  2177     //     except while it is waiting for the background collector to yield.
  2178     //
  2179     // The foreground collector should be blocked (not for long)
  2180     //   if the background collector is about to start a phase
  2181     //   executed with world stopped.  If the background
  2182     //   collector has already started such a phase, the
  2183     //   foreground collector is blocked waiting for the
  2184     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
  2185     //   are executed in the VM thread.
  2186     //
  2187     // The locking order is
  2188     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
  2189     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
  2190     //   CMS token  (claimed in
  2191     //                stop_world_and_do() -->
  2192     //                  safepoint_synchronize() -->
  2193     //                    CMSThread::synchronize())
  2196       // Check if the FG collector wants us to yield.
  2197       CMSTokenSync x(true); // is cms thread
  2198       if (waitForForegroundGC()) {
  2199         // We yielded to a foreground GC, nothing more to be
  2200         // done this round.
  2201         assert(_foregroundGCShouldWait == false, "We set it to false in "
  2202                "waitForForegroundGC()");
  2203         if (TraceCMSState) {
  2204           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2205             " exiting collection CMS state %d",
  2206             Thread::current(), _collectorState);
  2208         return;
  2209       } else {
  2210         // The background collector can run but check to see if the
  2211         // foreground collector has done a collection while the
  2212         // background collector was waiting to get the CGC_lock
  2213         // above.  If yes, break so that _foregroundGCShouldWait
  2214         // is cleared before returning.
  2215         if (_collectorState == Idling) {
  2216           break;
  2221     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
  2222       "should be waiting");
  2224     switch (_collectorState) {
  2225       case InitialMarking:
  2227           ReleaseForegroundGC x(this);
  2228           stats().record_cms_begin();
  2230           VM_CMS_Initial_Mark initial_mark_op(this);
  2231           VMThread::execute(&initial_mark_op);
  2233         // The collector state may be any legal state at this point
  2234         // since the background collector may have yielded to the
  2235         // foreground collector.
  2236         break;
  2237       case Marking:
  2238         // initial marking in checkpointRootsInitialWork has been completed
  2239         if (markFromRoots(true)) { // we were successful
  2240           assert(_collectorState == Precleaning, "Collector state should "
  2241             "have changed");
  2242         } else {
  2243           assert(_foregroundGCIsActive, "Internal state inconsistency");
  2245         break;
  2246       case Precleaning:
  2247         if (UseAdaptiveSizePolicy) {
  2248           size_policy()->concurrent_precleaning_begin();
  2250         // marking from roots in markFromRoots has been completed
  2251         preclean();
  2252         if (UseAdaptiveSizePolicy) {
  2253           size_policy()->concurrent_precleaning_end();
  2255         assert(_collectorState == AbortablePreclean ||
  2256                _collectorState == FinalMarking,
  2257                "Collector state should have changed");
  2258         break;
  2259       case AbortablePreclean:
  2260         if (UseAdaptiveSizePolicy) {
  2261         size_policy()->concurrent_phases_resume();
  2263         abortable_preclean();
  2264         if (UseAdaptiveSizePolicy) {
  2265           size_policy()->concurrent_precleaning_end();
  2267         assert(_collectorState == FinalMarking, "Collector state should "
  2268           "have changed");
  2269         break;
  2270       case FinalMarking:
  2272           ReleaseForegroundGC x(this);
  2274           VM_CMS_Final_Remark final_remark_op(this);
  2275           VMThread::execute(&final_remark_op);
  2277         assert(_foregroundGCShouldWait, "block post-condition");
  2278         break;
  2279       case Sweeping:
  2280         if (UseAdaptiveSizePolicy) {
  2281           size_policy()->concurrent_sweeping_begin();
  2283         // final marking in checkpointRootsFinal has been completed
  2284         sweep(true);
  2285         assert(_collectorState == Resizing, "Collector state change "
  2286           "to Resizing must be done under the free_list_lock");
  2287         _full_gcs_since_conc_gc = 0;
  2289         // Stop the timers for adaptive size policy for the concurrent phases
  2290         if (UseAdaptiveSizePolicy) {
  2291           size_policy()->concurrent_sweeping_end();
  2292           size_policy()->concurrent_phases_end(gch->gc_cause(),
  2293                                              gch->prev_gen(_cmsGen)->capacity(),
  2294                                              _cmsGen->free());
  2297       case Resizing: {
  2298         // Sweeping has been completed...
  2299         // At this point the background collection has completed.
  2300         // Don't move the call to compute_new_size() down
  2301         // into code that might be executed if the background
  2302         // collection was preempted.
  2304           ReleaseForegroundGC x(this);   // unblock FG collection
  2305           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
  2306           CMSTokenSync        z(true);   // not strictly needed.
  2307           if (_collectorState == Resizing) {
  2308             compute_new_size();
  2309             _collectorState = Resetting;
  2310           } else {
  2311             assert(_collectorState == Idling, "The state should only change"
  2312                    " because the foreground collector has finished the collection");
  2315         break;
  2317       case Resetting:
  2318         // CMS heap resizing has been completed
  2319         reset(true);
  2320         assert(_collectorState == Idling, "Collector state should "
  2321           "have changed");
  2322         stats().record_cms_end();
  2323         // Don't move the concurrent_phases_end() and compute_new_size()
  2324         // calls to here because a preempted background collection
  2325         // has it's state set to "Resetting".
  2326         break;
  2327       case Idling:
  2328       default:
  2329         ShouldNotReachHere();
  2330         break;
  2332     if (TraceCMSState) {
  2333       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2334         Thread::current(), _collectorState);
  2336     assert(_foregroundGCShouldWait, "block post-condition");
  2339   // Should this be in gc_epilogue?
  2340   collector_policy()->counters()->update_counters();
  2343     // Clear _foregroundGCShouldWait and, in the event that the
  2344     // foreground collector is waiting, notify it, before
  2345     // returning.
  2346     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2347     _foregroundGCShouldWait = false;
  2348     if (_foregroundGCIsActive) {
  2349       CGC_lock->notify();
  2351     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2352            "Possible deadlock");
  2354   if (TraceCMSState) {
  2355     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2356       " exiting collection CMS state %d",
  2357       Thread::current(), _collectorState);
  2359   if (PrintGC && Verbose) {
  2360     _cmsGen->print_heap_change(prev_used);
  2364 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
  2365   assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
  2366          "Foreground collector should be waiting, not executing");
  2367   assert(Thread::current()->is_VM_thread(), "A foreground collection"
  2368     "may only be done by the VM Thread with the world stopped");
  2369   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  2370          "VM thread should have CMS token");
  2372   NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
  2373     true, gclog_or_tty);)
  2374   if (UseAdaptiveSizePolicy) {
  2375     size_policy()->ms_collection_begin();
  2377   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
  2379   HandleMark hm;  // Discard invalid handles created during verification
  2381   if (VerifyBeforeGC &&
  2382       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2383     Universe::verify(true);
  2386   bool init_mark_was_synchronous = false; // until proven otherwise
  2387   while (_collectorState != Idling) {
  2388     if (TraceCMSState) {
  2389       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2390         Thread::current(), _collectorState);
  2392     switch (_collectorState) {
  2393       case InitialMarking:
  2394         init_mark_was_synchronous = true;  // fact to be exploited in re-mark
  2395         checkpointRootsInitial(false);
  2396         assert(_collectorState == Marking, "Collector state should have changed"
  2397           " within checkpointRootsInitial()");
  2398         break;
  2399       case Marking:
  2400         // initial marking in checkpointRootsInitialWork has been completed
  2401         if (VerifyDuringGC &&
  2402             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2403           gclog_or_tty->print("Verify before initial mark: ");
  2404           Universe::verify(true);
  2407           bool res = markFromRoots(false);
  2408           assert(res && _collectorState == FinalMarking, "Collector state should "
  2409             "have changed");
  2410           break;
  2412       case FinalMarking:
  2413         if (VerifyDuringGC &&
  2414             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2415           gclog_or_tty->print("Verify before re-mark: ");
  2416           Universe::verify(true);
  2418         checkpointRootsFinal(false, clear_all_soft_refs,
  2419                              init_mark_was_synchronous);
  2420         assert(_collectorState == Sweeping, "Collector state should not "
  2421           "have changed within checkpointRootsFinal()");
  2422         break;
  2423       case Sweeping:
  2424         // final marking in checkpointRootsFinal has been completed
  2425         if (VerifyDuringGC &&
  2426             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2427           gclog_or_tty->print("Verify before sweep: ");
  2428           Universe::verify(true);
  2430         sweep(false);
  2431         assert(_collectorState == Resizing, "Incorrect state");
  2432         break;
  2433       case Resizing: {
  2434         // Sweeping has been completed; the actual resize in this case
  2435         // is done separately; nothing to be done in this state.
  2436         _collectorState = Resetting;
  2437         break;
  2439       case Resetting:
  2440         // The heap has been resized.
  2441         if (VerifyDuringGC &&
  2442             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2443           gclog_or_tty->print("Verify before reset: ");
  2444           Universe::verify(true);
  2446         reset(false);
  2447         assert(_collectorState == Idling, "Collector state should "
  2448           "have changed");
  2449         break;
  2450       case Precleaning:
  2451       case AbortablePreclean:
  2452         // Elide the preclean phase
  2453         _collectorState = FinalMarking;
  2454         break;
  2455       default:
  2456         ShouldNotReachHere();
  2458     if (TraceCMSState) {
  2459       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2460         Thread::current(), _collectorState);
  2464   if (UseAdaptiveSizePolicy) {
  2465     GenCollectedHeap* gch = GenCollectedHeap::heap();
  2466     size_policy()->ms_collection_end(gch->gc_cause());
  2469   if (VerifyAfterGC &&
  2470       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2471     Universe::verify(true);
  2473   if (TraceCMSState) {
  2474     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2475       " exiting collection CMS state %d",
  2476       Thread::current(), _collectorState);
  2480 bool CMSCollector::waitForForegroundGC() {
  2481   bool res = false;
  2482   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2483          "CMS thread should have CMS token");
  2484   // Block the foreground collector until the
  2485   // background collectors decides whether to
  2486   // yield.
  2487   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2488   _foregroundGCShouldWait = true;
  2489   if (_foregroundGCIsActive) {
  2490     // The background collector yields to the
  2491     // foreground collector and returns a value
  2492     // indicating that it has yielded.  The foreground
  2493     // collector can proceed.
  2494     res = true;
  2495     _foregroundGCShouldWait = false;
  2496     ConcurrentMarkSweepThread::clear_CMS_flag(
  2497       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2498     ConcurrentMarkSweepThread::set_CMS_flag(
  2499       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2500     // Get a possibly blocked foreground thread going
  2501     CGC_lock->notify();
  2502     if (TraceCMSState) {
  2503       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
  2504         Thread::current(), _collectorState);
  2506     while (_foregroundGCIsActive) {
  2507       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  2509     ConcurrentMarkSweepThread::set_CMS_flag(
  2510       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2511     ConcurrentMarkSweepThread::clear_CMS_flag(
  2512       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2514   if (TraceCMSState) {
  2515     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
  2516       Thread::current(), _collectorState);
  2518   return res;
  2521 // Because of the need to lock the free lists and other structures in
  2522 // the collector, common to all the generations that the collector is
  2523 // collecting, we need the gc_prologues of individual CMS generations
  2524 // delegate to their collector. It may have been simpler had the
  2525 // current infrastructure allowed one to call a prologue on a
  2526 // collector. In the absence of that we have the generation's
  2527 // prologue delegate to the collector, which delegates back
  2528 // some "local" work to a worker method in the individual generations
  2529 // that it's responsible for collecting, while itself doing any
  2530 // work common to all generations it's responsible for. A similar
  2531 // comment applies to the  gc_epilogue()'s.
  2532 // The role of the varaible _between_prologue_and_epilogue is to
  2533 // enforce the invocation protocol.
  2534 void CMSCollector::gc_prologue(bool full) {
  2535   // Call gc_prologue_work() for each CMSGen and PermGen that
  2536   // we are responsible for.
  2538   // The following locking discipline assumes that we are only called
  2539   // when the world is stopped.
  2540   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
  2542   // The CMSCollector prologue must call the gc_prologues for the
  2543   // "generations" (including PermGen if any) that it's responsible
  2544   // for.
  2546   assert(   Thread::current()->is_VM_thread()
  2547          || (   CMSScavengeBeforeRemark
  2548              && Thread::current()->is_ConcurrentGC_thread()),
  2549          "Incorrect thread type for prologue execution");
  2551   if (_between_prologue_and_epilogue) {
  2552     // We have already been invoked; this is a gc_prologue delegation
  2553     // from yet another CMS generation that we are responsible for, just
  2554     // ignore it since all relevant work has already been done.
  2555     return;
  2558   // set a bit saying prologue has been called; cleared in epilogue
  2559   _between_prologue_and_epilogue = true;
  2560   // Claim locks for common data structures, then call gc_prologue_work()
  2561   // for each CMSGen and PermGen that we are responsible for.
  2563   getFreelistLocks();   // gets free list locks on constituent spaces
  2564   bitMapLock()->lock_without_safepoint_check();
  2566   // Should call gc_prologue_work() for all cms gens we are responsible for
  2567   bool registerClosure =    _collectorState >= Marking
  2568                          && _collectorState < Sweeping;
  2569   ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
  2570                                                : &_modUnionClosure;
  2571   _cmsGen->gc_prologue_work(full, registerClosure, muc);
  2572   _permGen->gc_prologue_work(full, registerClosure, muc);
  2574   if (!full) {
  2575     stats().record_gc0_begin();
  2579 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
  2580   // Delegate to CMScollector which knows how to coordinate between
  2581   // this and any other CMS generations that it is responsible for
  2582   // collecting.
  2583   collector()->gc_prologue(full);
  2586 // This is a "private" interface for use by this generation's CMSCollector.
  2587 // Not to be called directly by any other entity (for instance,
  2588 // GenCollectedHeap, which calls the "public" gc_prologue method above).
  2589 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
  2590   bool registerClosure, ModUnionClosure* modUnionClosure) {
  2591   assert(!incremental_collection_failed(), "Shouldn't be set yet");
  2592   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
  2593     "Should be NULL");
  2594   if (registerClosure) {
  2595     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
  2597   cmsSpace()->gc_prologue();
  2598   // Clear stat counters
  2599   NOT_PRODUCT(
  2600     assert(_numObjectsPromoted == 0, "check");
  2601     assert(_numWordsPromoted   == 0, "check");
  2602     if (Verbose && PrintGC) {
  2603       gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
  2604                           SIZE_FORMAT" bytes concurrently",
  2605       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
  2607     _numObjectsAllocated = 0;
  2608     _numWordsAllocated   = 0;
  2612 void CMSCollector::gc_epilogue(bool full) {
  2613   // The following locking discipline assumes that we are only called
  2614   // when the world is stopped.
  2615   assert(SafepointSynchronize::is_at_safepoint(),
  2616          "world is stopped assumption");
  2618   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
  2619   // if linear allocation blocks need to be appropriately marked to allow the
  2620   // the blocks to be parsable. We also check here whether we need to nudge the
  2621   // CMS collector thread to start a new cycle (if it's not already active).
  2622   assert(   Thread::current()->is_VM_thread()
  2623          || (   CMSScavengeBeforeRemark
  2624              && Thread::current()->is_ConcurrentGC_thread()),
  2625          "Incorrect thread type for epilogue execution");
  2627   if (!_between_prologue_and_epilogue) {
  2628     // We have already been invoked; this is a gc_epilogue delegation
  2629     // from yet another CMS generation that we are responsible for, just
  2630     // ignore it since all relevant work has already been done.
  2631     return;
  2633   assert(haveFreelistLocks(), "must have freelist locks");
  2634   assert_lock_strong(bitMapLock());
  2636   _cmsGen->gc_epilogue_work(full);
  2637   _permGen->gc_epilogue_work(full);
  2639   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
  2640     // in case sampling was not already enabled, enable it
  2641     _start_sampling = true;
  2643   // reset _eden_chunk_array so sampling starts afresh
  2644   _eden_chunk_index = 0;
  2646   size_t cms_used   = _cmsGen->cmsSpace()->used();
  2647   size_t perm_used  = _permGen->cmsSpace()->used();
  2649   // update performance counters - this uses a special version of
  2650   // update_counters() that allows the utilization to be passed as a
  2651   // parameter, avoiding multiple calls to used().
  2652   //
  2653   _cmsGen->update_counters(cms_used);
  2654   _permGen->update_counters(perm_used);
  2656   if (CMSIncrementalMode) {
  2657     icms_update_allocation_limits();
  2660   bitMapLock()->unlock();
  2661   releaseFreelistLocks();
  2663   _between_prologue_and_epilogue = false;  // ready for next cycle
  2666 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
  2667   collector()->gc_epilogue(full);
  2669   // Also reset promotion tracking in par gc thread states.
  2670   if (ParallelGCThreads > 0) {
  2671     for (uint i = 0; i < ParallelGCThreads; i++) {
  2672       _par_gc_thread_states[i]->promo.stopTrackingPromotions();
  2677 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
  2678   assert(!incremental_collection_failed(), "Should have been cleared");
  2679   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
  2680   cmsSpace()->gc_epilogue();
  2681     // Print stat counters
  2682   NOT_PRODUCT(
  2683     assert(_numObjectsAllocated == 0, "check");
  2684     assert(_numWordsAllocated == 0, "check");
  2685     if (Verbose && PrintGC) {
  2686       gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
  2687                           SIZE_FORMAT" bytes",
  2688                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
  2690     _numObjectsPromoted = 0;
  2691     _numWordsPromoted   = 0;
  2694   if (PrintGC && Verbose) {
  2695     // Call down the chain in contiguous_available needs the freelistLock
  2696     // so print this out before releasing the freeListLock.
  2697     gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
  2698                         contiguous_available());
  2702 #ifndef PRODUCT
  2703 bool CMSCollector::have_cms_token() {
  2704   Thread* thr = Thread::current();
  2705   if (thr->is_VM_thread()) {
  2706     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
  2707   } else if (thr->is_ConcurrentGC_thread()) {
  2708     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
  2709   } else if (thr->is_GC_task_thread()) {
  2710     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
  2711            ParGCRareEvent_lock->owned_by_self();
  2713   return false;
  2715 #endif
  2717 // Check reachability of the given heap address in CMS generation,
  2718 // treating all other generations as roots.
  2719 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
  2720   // We could "guarantee" below, rather than assert, but i'll
  2721   // leave these as "asserts" so that an adventurous debugger
  2722   // could try this in the product build provided some subset of
  2723   // the conditions were met, provided they were intersted in the
  2724   // results and knew that the computation below wouldn't interfere
  2725   // with other concurrent computations mutating the structures
  2726   // being read or written.
  2727   assert(SafepointSynchronize::is_at_safepoint(),
  2728          "Else mutations in object graph will make answer suspect");
  2729   assert(have_cms_token(), "Should hold cms token");
  2730   assert(haveFreelistLocks(), "must hold free list locks");
  2731   assert_lock_strong(bitMapLock());
  2733   // Clear the marking bit map array before starting, but, just
  2734   // for kicks, first report if the given address is already marked
  2735   gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
  2736                 _markBitMap.isMarked(addr) ? "" : " not");
  2738   if (verify_after_remark()) {
  2739     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2740     bool result = verification_mark_bm()->isMarked(addr);
  2741     gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
  2742                            result ? "IS" : "is NOT");
  2743     return result;
  2744   } else {
  2745     gclog_or_tty->print_cr("Could not compute result");
  2746     return false;
  2750 ////////////////////////////////////////////////////////
  2751 // CMS Verification Support
  2752 ////////////////////////////////////////////////////////
  2753 // Following the remark phase, the following invariant
  2754 // should hold -- each object in the CMS heap which is
  2755 // marked in markBitMap() should be marked in the verification_mark_bm().
  2757 class VerifyMarkedClosure: public BitMapClosure {
  2758   CMSBitMap* _marks;
  2759   bool       _failed;
  2761  public:
  2762   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
  2764   bool do_bit(size_t offset) {
  2765     HeapWord* addr = _marks->offsetToHeapWord(offset);
  2766     if (!_marks->isMarked(addr)) {
  2767       oop(addr)->print();
  2768       gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
  2769       _failed = true;
  2771     return true;
  2774   bool failed() { return _failed; }
  2775 };
  2777 bool CMSCollector::verify_after_remark() {
  2778   gclog_or_tty->print(" [Verifying CMS Marking... ");
  2779   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2780   static bool init = false;
  2782   assert(SafepointSynchronize::is_at_safepoint(),
  2783          "Else mutations in object graph will make answer suspect");
  2784   assert(have_cms_token(),
  2785          "Else there may be mutual interference in use of "
  2786          " verification data structures");
  2787   assert(_collectorState > Marking && _collectorState <= Sweeping,
  2788          "Else marking info checked here may be obsolete");
  2789   assert(haveFreelistLocks(), "must hold free list locks");
  2790   assert_lock_strong(bitMapLock());
  2793   // Allocate marking bit map if not already allocated
  2794   if (!init) { // first time
  2795     if (!verification_mark_bm()->allocate(_span)) {
  2796       return false;
  2798     init = true;
  2801   assert(verification_mark_stack()->isEmpty(), "Should be empty");
  2803   // Turn off refs discovery -- so we will be tracing through refs.
  2804   // This is as intended, because by this time
  2805   // GC must already have cleared any refs that need to be cleared,
  2806   // and traced those that need to be marked; moreover,
  2807   // the marking done here is not going to intefere in any
  2808   // way with the marking information used by GC.
  2809   NoRefDiscovery no_discovery(ref_processor());
  2811   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  2813   // Clear any marks from a previous round
  2814   verification_mark_bm()->clear_all();
  2815   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
  2816   assert(overflow_list_is_empty(), "overflow list should be empty");
  2818   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2819   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  2820   // Update the saved marks which may affect the root scans.
  2821   gch->save_marks();
  2823   if (CMSRemarkVerifyVariant == 1) {
  2824     // In this first variant of verification, we complete
  2825     // all marking, then check if the new marks-verctor is
  2826     // a subset of the CMS marks-vector.
  2827     verify_after_remark_work_1();
  2828   } else if (CMSRemarkVerifyVariant == 2) {
  2829     // In this second variant of verification, we flag an error
  2830     // (i.e. an object reachable in the new marks-vector not reachable
  2831     // in the CMS marks-vector) immediately, also indicating the
  2832     // identify of an object (A) that references the unmarked object (B) --
  2833     // presumably, a mutation to A failed to be picked up by preclean/remark?
  2834     verify_after_remark_work_2();
  2835   } else {
  2836     warning("Unrecognized value %d for CMSRemarkVerifyVariant",
  2837             CMSRemarkVerifyVariant);
  2839   gclog_or_tty->print(" done] ");
  2840   return true;
  2843 void CMSCollector::verify_after_remark_work_1() {
  2844   ResourceMark rm;
  2845   HandleMark  hm;
  2846   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2848   // Mark from roots one level into CMS
  2849   MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
  2850   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  2852   gch->gen_process_strong_roots(_cmsGen->level(),
  2853                                 true,   // younger gens are roots
  2854                                 true,   // collecting perm gen
  2855                                 SharedHeap::ScanningOption(roots_scanning_options()),
  2856                                 NULL, &notOlder);
  2858   // Now mark from the roots
  2859   assert(_revisitStack.isEmpty(), "Should be empty");
  2860   MarkFromRootsClosure markFromRootsClosure(this, _span,
  2861     verification_mark_bm(), verification_mark_stack(), &_revisitStack,
  2862     false /* don't yield */, true /* verifying */);
  2863   assert(_restart_addr == NULL, "Expected pre-condition");
  2864   verification_mark_bm()->iterate(&markFromRootsClosure);
  2865   while (_restart_addr != NULL) {
  2866     // Deal with stack overflow: by restarting at the indicated
  2867     // address.
  2868     HeapWord* ra = _restart_addr;
  2869     markFromRootsClosure.reset(ra);
  2870     _restart_addr = NULL;
  2871     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  2873   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  2874   verify_work_stacks_empty();
  2875   // Should reset the revisit stack above, since no class tree
  2876   // surgery is forthcoming.
  2877   _revisitStack.reset(); // throwing away all contents
  2879   // Marking completed -- now verify that each bit marked in
  2880   // verification_mark_bm() is also marked in markBitMap(); flag all
  2881   // errors by printing corresponding objects.
  2882   VerifyMarkedClosure vcl(markBitMap());
  2883   verification_mark_bm()->iterate(&vcl);
  2884   if (vcl.failed()) {
  2885     gclog_or_tty->print("Verification failed");
  2886     Universe::heap()->print();
  2887     fatal(" ... aborting");
  2891 void CMSCollector::verify_after_remark_work_2() {
  2892   ResourceMark rm;
  2893   HandleMark  hm;
  2894   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2896   // Mark from roots one level into CMS
  2897   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
  2898                                      markBitMap(), true /* nmethods */);
  2899   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  2900   gch->gen_process_strong_roots(_cmsGen->level(),
  2901                                 true,   // younger gens are roots
  2902                                 true,   // collecting perm gen
  2903                                 SharedHeap::ScanningOption(roots_scanning_options()),
  2904                                 NULL, &notOlder);
  2906   // Now mark from the roots
  2907   assert(_revisitStack.isEmpty(), "Should be empty");
  2908   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
  2909     verification_mark_bm(), markBitMap(), verification_mark_stack());
  2910   assert(_restart_addr == NULL, "Expected pre-condition");
  2911   verification_mark_bm()->iterate(&markFromRootsClosure);
  2912   while (_restart_addr != NULL) {
  2913     // Deal with stack overflow: by restarting at the indicated
  2914     // address.
  2915     HeapWord* ra = _restart_addr;
  2916     markFromRootsClosure.reset(ra);
  2917     _restart_addr = NULL;
  2918     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  2920   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  2921   verify_work_stacks_empty();
  2922   // Should reset the revisit stack above, since no class tree
  2923   // surgery is forthcoming.
  2924   _revisitStack.reset(); // throwing away all contents
  2926   // Marking completed -- now verify that each bit marked in
  2927   // verification_mark_bm() is also marked in markBitMap(); flag all
  2928   // errors by printing corresponding objects.
  2929   VerifyMarkedClosure vcl(markBitMap());
  2930   verification_mark_bm()->iterate(&vcl);
  2931   assert(!vcl.failed(), "Else verification above should not have succeeded");
  2934 void ConcurrentMarkSweepGeneration::save_marks() {
  2935   // delegate to CMS space
  2936   cmsSpace()->save_marks();
  2937   for (uint i = 0; i < ParallelGCThreads; i++) {
  2938     _par_gc_thread_states[i]->promo.startTrackingPromotions();
  2942 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
  2943   return cmsSpace()->no_allocs_since_save_marks();
  2946 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
  2948 void ConcurrentMarkSweepGeneration::                            \
  2949 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
  2950   cl->set_generation(this);                                     \
  2951   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
  2952   cl->reset_generation();                                       \
  2953   save_marks();                                                 \
  2956 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
  2958 void
  2959 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
  2961   // Not currently implemented; need to do the following. -- ysr.
  2962   // dld -- I think that is used for some sort of allocation profiler.  So it
  2963   // really means the objects allocated by the mutator since the last
  2964   // GC.  We could potentially implement this cheaply by recording only
  2965   // the direct allocations in a side data structure.
  2966   //
  2967   // I think we probably ought not to be required to support these
  2968   // iterations at any arbitrary point; I think there ought to be some
  2969   // call to enable/disable allocation profiling in a generation/space,
  2970   // and the iterator ought to return the objects allocated in the
  2971   // gen/space since the enable call, or the last iterator call (which
  2972   // will probably be at a GC.)  That way, for gens like CM&S that would
  2973   // require some extra data structure to support this, we only pay the
  2974   // cost when it's in use...
  2975   cmsSpace()->object_iterate_since_last_GC(blk);
  2978 void
  2979 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
  2980   cl->set_generation(this);
  2981   younger_refs_in_space_iterate(_cmsSpace, cl);
  2982   cl->reset_generation();
  2985 void
  2986 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
  2987   if (freelistLock()->owned_by_self()) {
  2988     Generation::oop_iterate(mr, cl);
  2989   } else {
  2990     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  2991     Generation::oop_iterate(mr, cl);
  2995 void
  2996 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
  2997   if (freelistLock()->owned_by_self()) {
  2998     Generation::oop_iterate(cl);
  2999   } else {
  3000     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3001     Generation::oop_iterate(cl);
  3005 void
  3006 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
  3007   if (freelistLock()->owned_by_self()) {
  3008     Generation::object_iterate(cl);
  3009   } else {
  3010     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3011     Generation::object_iterate(cl);
  3015 void
  3016 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
  3019 void
  3020 ConcurrentMarkSweepGeneration::post_compact() {
  3023 void
  3024 ConcurrentMarkSweepGeneration::prepare_for_verify() {
  3025   // Fix the linear allocation blocks to look like free blocks.
  3027   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3028   // are not called when the heap is verified during universe initialization and
  3029   // at vm shutdown.
  3030   if (freelistLock()->owned_by_self()) {
  3031     cmsSpace()->prepare_for_verify();
  3032   } else {
  3033     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3034     cmsSpace()->prepare_for_verify();
  3038 void
  3039 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
  3040   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3041   // are not called when the heap is verified during universe initialization and
  3042   // at vm shutdown.
  3043   if (freelistLock()->owned_by_self()) {
  3044     cmsSpace()->verify(false /* ignored */);
  3045   } else {
  3046     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3047     cmsSpace()->verify(false /* ignored */);
  3051 void CMSCollector::verify(bool allow_dirty /* ignored */) {
  3052   _cmsGen->verify(allow_dirty);
  3053   _permGen->verify(allow_dirty);
  3056 #ifndef PRODUCT
  3057 bool CMSCollector::overflow_list_is_empty() const {
  3058   assert(_num_par_pushes >= 0, "Inconsistency");
  3059   if (_overflow_list == NULL) {
  3060     assert(_num_par_pushes == 0, "Inconsistency");
  3062   return _overflow_list == NULL;
  3065 // The methods verify_work_stacks_empty() and verify_overflow_empty()
  3066 // merely consolidate assertion checks that appear to occur together frequently.
  3067 void CMSCollector::verify_work_stacks_empty() const {
  3068   assert(_markStack.isEmpty(), "Marking stack should be empty");
  3069   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3072 void CMSCollector::verify_overflow_empty() const {
  3073   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3074   assert(no_preserved_marks(), "No preserved marks");
  3076 #endif // PRODUCT
  3078 // Decide if we want to enable class unloading as part of the
  3079 // ensuing concurrent GC cycle. We will collect the perm gen and
  3080 // unload classes if it's the case that:
  3081 // (1) an explicit gc request has been made and the flag
  3082 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
  3083 // (2) (a) class unloading is enabled at the command line, and
  3084 //     (b) (i)   perm gen threshold has been crossed, or
  3085 //         (ii)  old gen is getting really full, or
  3086 //         (iii) the previous N CMS collections did not collect the
  3087 //               perm gen
  3088 // NOTE: Provided there is no change in the state of the heap between
  3089 // calls to this method, it should have idempotent results. Moreover,
  3090 // its results should be monotonically increasing (i.e. going from 0 to 1,
  3091 // but not 1 to 0) between successive calls between which the heap was
  3092 // not collected. For the implementation below, it must thus rely on
  3093 // the property that concurrent_cycles_since_last_unload()
  3094 // will not decrease unless a collection cycle happened and that
  3095 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
  3096 // themselves also monotonic in that sense. See check_monotonicity()
  3097 // below.
  3098 bool CMSCollector::update_should_unload_classes() {
  3099   _should_unload_classes = false;
  3100   // Condition 1 above
  3101   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
  3102     _should_unload_classes = true;
  3103   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
  3104     // Disjuncts 2.b.(i,ii,iii) above
  3105     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
  3106                               CMSClassUnloadingMaxInterval)
  3107                            || _permGen->should_concurrent_collect()
  3108                            || _cmsGen->is_too_full();
  3110   return _should_unload_classes;
  3113 bool ConcurrentMarkSweepGeneration::is_too_full() const {
  3114   bool res = should_concurrent_collect();
  3115   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
  3116   return res;
  3119 void CMSCollector::setup_cms_unloading_and_verification_state() {
  3120   const  bool should_verify =    VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
  3121                              || VerifyBeforeExit;
  3122   const  int  rso           =    SharedHeap::SO_Symbols | SharedHeap::SO_Strings
  3123                              |   SharedHeap::SO_CodeCache;
  3125   if (should_unload_classes()) {   // Should unload classes this cycle
  3126     remove_root_scanning_option(rso);  // Shrink the root set appropriately
  3127     set_verifying(should_verify);    // Set verification state for this cycle
  3128     return;                            // Nothing else needs to be done at this time
  3131   // Not unloading classes this cycle
  3132   assert(!should_unload_classes(), "Inconsitency!");
  3133   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
  3134     // We were not verifying, or we _were_ unloading classes in the last cycle,
  3135     // AND some verification options are enabled this cycle; in this case,
  3136     // we must make sure that the deadness map is allocated if not already so,
  3137     // and cleared (if already allocated previously --
  3138     // CMSBitMap::sizeInBits() is used to determine if it's allocated).
  3139     if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
  3140       if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
  3141         warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
  3142                 "permanent generation verification disabled");
  3143         return;  // Note that we leave verification disabled, so we'll retry this
  3144                  // allocation next cycle. We _could_ remember this failure
  3145                  // and skip further attempts and permanently disable verification
  3146                  // attempts if that is considered more desirable.
  3148       assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
  3149               "_perm_gen_ver_bit_map inconsistency?");
  3150     } else {
  3151       perm_gen_verify_bit_map()->clear_all();
  3153     // Include symbols, strings and code cache elements to prevent their resurrection.
  3154     add_root_scanning_option(rso);
  3155     set_verifying(true);
  3156   } else if (verifying() && !should_verify) {
  3157     // We were verifying, but some verification flags got disabled.
  3158     set_verifying(false);
  3159     // Exclude symbols, strings and code cache elements from root scanning to
  3160     // reduce IM and RM pauses.
  3161     remove_root_scanning_option(rso);
  3166 #ifndef PRODUCT
  3167 HeapWord* CMSCollector::block_start(const void* p) const {
  3168   const HeapWord* addr = (HeapWord*)p;
  3169   if (_span.contains(p)) {
  3170     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
  3171       return _cmsGen->cmsSpace()->block_start(p);
  3172     } else {
  3173       assert(_permGen->cmsSpace()->is_in_reserved(addr),
  3174              "Inconsistent _span?");
  3175       return _permGen->cmsSpace()->block_start(p);
  3178   return NULL;
  3180 #endif
  3182 HeapWord*
  3183 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
  3184                                                    bool   tlab,
  3185                                                    bool   parallel) {
  3186   assert(!tlab, "Can't deal with TLAB allocation");
  3187   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3188   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
  3189     CMSExpansionCause::_satisfy_allocation);
  3190   if (GCExpandToAllocateDelayMillis > 0) {
  3191     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3193   return have_lock_and_allocate(word_size, tlab);
  3196 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
  3197 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
  3198 // to CardGeneration and share it...
  3199 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
  3200   return CardGeneration::expand(bytes, expand_bytes);
  3203 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
  3204   CMSExpansionCause::Cause cause)
  3207   bool success = expand(bytes, expand_bytes);
  3209   // remember why we expanded; this information is used
  3210   // by shouldConcurrentCollect() when making decisions on whether to start
  3211   // a new CMS cycle.
  3212   if (success) {
  3213     set_expansion_cause(cause);
  3214     if (PrintGCDetails && Verbose) {
  3215       gclog_or_tty->print_cr("Expanded CMS gen for %s",
  3216         CMSExpansionCause::to_string(cause));
  3221 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
  3222   HeapWord* res = NULL;
  3223   MutexLocker x(ParGCRareEvent_lock);
  3224   while (true) {
  3225     // Expansion by some other thread might make alloc OK now:
  3226     res = ps->lab.alloc(word_sz);
  3227     if (res != NULL) return res;
  3228     // If there's not enough expansion space available, give up.
  3229     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
  3230       return NULL;
  3232     // Otherwise, we try expansion.
  3233     expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
  3234       CMSExpansionCause::_allocate_par_lab);
  3235     // Now go around the loop and try alloc again;
  3236     // A competing par_promote might beat us to the expansion space,
  3237     // so we may go around the loop again if promotion fails agaion.
  3238     if (GCExpandToAllocateDelayMillis > 0) {
  3239       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3245 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
  3246   PromotionInfo* promo) {
  3247   MutexLocker x(ParGCRareEvent_lock);
  3248   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
  3249   while (true) {
  3250     // Expansion by some other thread might make alloc OK now:
  3251     if (promo->ensure_spooling_space()) {
  3252       assert(promo->has_spooling_space(),
  3253              "Post-condition of successful ensure_spooling_space()");
  3254       return true;
  3256     // If there's not enough expansion space available, give up.
  3257     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
  3258       return false;
  3260     // Otherwise, we try expansion.
  3261     expand(refill_size_bytes, MinHeapDeltaBytes,
  3262       CMSExpansionCause::_allocate_par_spooling_space);
  3263     // Now go around the loop and try alloc again;
  3264     // A competing allocation might beat us to the expansion space,
  3265     // so we may go around the loop again if allocation fails again.
  3266     if (GCExpandToAllocateDelayMillis > 0) {
  3267       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3274 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
  3275   assert_locked_or_safepoint(Heap_lock);
  3276   size_t size = ReservedSpace::page_align_size_down(bytes);
  3277   if (size > 0) {
  3278     shrink_by(size);
  3282 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
  3283   assert_locked_or_safepoint(Heap_lock);
  3284   bool result = _virtual_space.expand_by(bytes);
  3285   if (result) {
  3286     HeapWord* old_end = _cmsSpace->end();
  3287     size_t new_word_size =
  3288       heap_word_size(_virtual_space.committed_size());
  3289     MemRegion mr(_cmsSpace->bottom(), new_word_size);
  3290     _bts->resize(new_word_size);  // resize the block offset shared array
  3291     Universe::heap()->barrier_set()->resize_covered_region(mr);
  3292     // Hmmmm... why doesn't CFLS::set_end verify locking?
  3293     // This is quite ugly; FIX ME XXX
  3294     _cmsSpace->assert_locked();
  3295     _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  3297     // update the space and generation capacity counters
  3298     if (UsePerfData) {
  3299       _space_counters->update_capacity();
  3300       _gen_counters->update_all();
  3303     if (Verbose && PrintGC) {
  3304       size_t new_mem_size = _virtual_space.committed_size();
  3305       size_t old_mem_size = new_mem_size - bytes;
  3306       gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
  3307                     name(), old_mem_size/K, bytes/K, new_mem_size/K);
  3310   return result;
  3313 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
  3314   assert_locked_or_safepoint(Heap_lock);
  3315   bool success = true;
  3316   const size_t remaining_bytes = _virtual_space.uncommitted_size();
  3317   if (remaining_bytes > 0) {
  3318     success = grow_by(remaining_bytes);
  3319     DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
  3321   return success;
  3324 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
  3325   assert_locked_or_safepoint(Heap_lock);
  3326   assert_lock_strong(freelistLock());
  3327   // XXX Fix when compaction is implemented.
  3328   warning("Shrinking of CMS not yet implemented");
  3329   return;
  3333 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
  3334 // phases.
  3335 class CMSPhaseAccounting: public StackObj {
  3336  public:
  3337   CMSPhaseAccounting(CMSCollector *collector,
  3338                      const char *phase,
  3339                      bool print_cr = true);
  3340   ~CMSPhaseAccounting();
  3342  private:
  3343   CMSCollector *_collector;
  3344   const char *_phase;
  3345   elapsedTimer _wallclock;
  3346   bool _print_cr;
  3348  public:
  3349   // Not MT-safe; so do not pass around these StackObj's
  3350   // where they may be accessed by other threads.
  3351   jlong wallclock_millis() {
  3352     assert(_wallclock.is_active(), "Wall clock should not stop");
  3353     _wallclock.stop();  // to record time
  3354     jlong ret = _wallclock.milliseconds();
  3355     _wallclock.start(); // restart
  3356     return ret;
  3358 };
  3360 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
  3361                                        const char *phase,
  3362                                        bool print_cr) :
  3363   _collector(collector), _phase(phase), _print_cr(print_cr) {
  3365   if (PrintCMSStatistics != 0) {
  3366     _collector->resetYields();
  3368   if (PrintGCDetails && PrintGCTimeStamps) {
  3369     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3370     gclog_or_tty->stamp();
  3371     gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
  3372       _collector->cmsGen()->short_name(), _phase);
  3374   _collector->resetTimer();
  3375   _wallclock.start();
  3376   _collector->startTimer();
  3379 CMSPhaseAccounting::~CMSPhaseAccounting() {
  3380   assert(_wallclock.is_active(), "Wall clock should not have stopped");
  3381   _collector->stopTimer();
  3382   _wallclock.stop();
  3383   if (PrintGCDetails) {
  3384     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3385     if (PrintGCTimeStamps) {
  3386       gclog_or_tty->stamp();
  3387       gclog_or_tty->print(": ");
  3389     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
  3390                  _collector->cmsGen()->short_name(),
  3391                  _phase, _collector->timerValue(), _wallclock.seconds());
  3392     if (_print_cr) {
  3393       gclog_or_tty->print_cr("");
  3395     if (PrintCMSStatistics != 0) {
  3396       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
  3397                     _collector->yields());
  3402 // CMS work
  3404 // Checkpoint the roots into this generation from outside
  3405 // this generation. [Note this initial checkpoint need only
  3406 // be approximate -- we'll do a catch up phase subsequently.]
  3407 void CMSCollector::checkpointRootsInitial(bool asynch) {
  3408   assert(_collectorState == InitialMarking, "Wrong collector state");
  3409   check_correct_thread_executing();
  3410   ReferenceProcessor* rp = ref_processor();
  3411   SpecializationStats::clear();
  3412   assert(_restart_addr == NULL, "Control point invariant");
  3413   if (asynch) {
  3414     // acquire locks for subsequent manipulations
  3415     MutexLockerEx x(bitMapLock(),
  3416                     Mutex::_no_safepoint_check_flag);
  3417     checkpointRootsInitialWork(asynch);
  3418     rp->verify_no_references_recorded();
  3419     rp->enable_discovery(); // enable ("weak") refs discovery
  3420     _collectorState = Marking;
  3421   } else {
  3422     // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
  3423     // which recognizes if we are a CMS generation, and doesn't try to turn on
  3424     // discovery; verify that they aren't meddling.
  3425     assert(!rp->discovery_is_atomic(),
  3426            "incorrect setting of discovery predicate");
  3427     assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
  3428            "ref discovery for this generation kind");
  3429     // already have locks
  3430     checkpointRootsInitialWork(asynch);
  3431     rp->enable_discovery(); // now enable ("weak") refs discovery
  3432     _collectorState = Marking;
  3434   SpecializationStats::print();
  3437 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
  3438   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
  3439   assert(_collectorState == InitialMarking, "just checking");
  3441   // If there has not been a GC[n-1] since last GC[n] cycle completed,
  3442   // precede our marking with a collection of all
  3443   // younger generations to keep floating garbage to a minimum.
  3444   // XXX: we won't do this for now -- it's an optimization to be done later.
  3446   // already have locks
  3447   assert_lock_strong(bitMapLock());
  3448   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
  3450   // Setup the verification and class unloading state for this
  3451   // CMS collection cycle.
  3452   setup_cms_unloading_and_verification_state();
  3454   NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
  3455     PrintGCDetails && Verbose, true, gclog_or_tty);)
  3456   if (UseAdaptiveSizePolicy) {
  3457     size_policy()->checkpoint_roots_initial_begin();
  3460   // Reset all the PLAB chunk arrays if necessary.
  3461   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
  3462     reset_survivor_plab_arrays();
  3465   ResourceMark rm;
  3466   HandleMark  hm;
  3468   FalseClosure falseClosure;
  3469   // In the case of a synchronous collection, we will elide the
  3470   // remark step, so it's important to catch all the nmethod oops
  3471   // in this step; hence the last argument to the constrcutor below.
  3472   MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
  3473   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3475   verify_work_stacks_empty();
  3476   verify_overflow_empty();
  3478   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  3479   // Update the saved marks which may affect the root scans.
  3480   gch->save_marks();
  3482   // weak reference processing has not started yet.
  3483   ref_processor()->set_enqueuing_is_done(false);
  3486     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  3487     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  3488     gch->gen_process_strong_roots(_cmsGen->level(),
  3489                                   true,   // younger gens are roots
  3490                                   true,   // collecting perm gen
  3491                                   SharedHeap::ScanningOption(roots_scanning_options()),
  3492                                   NULL, &notOlder);
  3495   // Clear mod-union table; it will be dirtied in the prologue of
  3496   // CMS generation per each younger generation collection.
  3498   assert(_modUnionTable.isAllClear(),
  3499        "Was cleared in most recent final checkpoint phase"
  3500        " or no bits are set in the gc_prologue before the start of the next "
  3501        "subsequent marking phase.");
  3503   // Temporarily disabled, since pre/post-consumption closures don't
  3504   // care about precleaned cards
  3505   #if 0
  3507     MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
  3508                              (HeapWord*)_virtual_space.high());
  3509     _ct->ct_bs()->preclean_dirty_cards(mr);
  3511   #endif
  3513   // Save the end of the used_region of the constituent generations
  3514   // to be used to limit the extent of sweep in each generation.
  3515   save_sweep_limits();
  3516   if (UseAdaptiveSizePolicy) {
  3517     size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
  3519   verify_overflow_empty();
  3522 bool CMSCollector::markFromRoots(bool asynch) {
  3523   // we might be tempted to assert that:
  3524   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
  3525   //        "inconsistent argument?");
  3526   // However that wouldn't be right, because it's possible that
  3527   // a safepoint is indeed in progress as a younger generation
  3528   // stop-the-world GC happens even as we mark in this generation.
  3529   assert(_collectorState == Marking, "inconsistent state?");
  3530   check_correct_thread_executing();
  3531   verify_overflow_empty();
  3533   bool res;
  3534   if (asynch) {
  3536     // Start the timers for adaptive size policy for the concurrent phases
  3537     // Do it here so that the foreground MS can use the concurrent
  3538     // timer since a foreground MS might has the sweep done concurrently
  3539     // or STW.
  3540     if (UseAdaptiveSizePolicy) {
  3541       size_policy()->concurrent_marking_begin();
  3544     // Weak ref discovery note: We may be discovering weak
  3545     // refs in this generation concurrent (but interleaved) with
  3546     // weak ref discovery by a younger generation collector.
  3548     CMSTokenSyncWithLocks ts(true, bitMapLock());
  3549     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  3550     CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
  3551     res = markFromRootsWork(asynch);
  3552     if (res) {
  3553       _collectorState = Precleaning;
  3554     } else { // We failed and a foreground collection wants to take over
  3555       assert(_foregroundGCIsActive, "internal state inconsistency");
  3556       assert(_restart_addr == NULL,  "foreground will restart from scratch");
  3557       if (PrintGCDetails) {
  3558         gclog_or_tty->print_cr("bailing out to foreground collection");
  3561     if (UseAdaptiveSizePolicy) {
  3562       size_policy()->concurrent_marking_end();
  3564   } else {
  3565     assert(SafepointSynchronize::is_at_safepoint(),
  3566            "inconsistent with asynch == false");
  3567     if (UseAdaptiveSizePolicy) {
  3568       size_policy()->ms_collection_marking_begin();
  3570     // already have locks
  3571     res = markFromRootsWork(asynch);
  3572     _collectorState = FinalMarking;
  3573     if (UseAdaptiveSizePolicy) {
  3574       GenCollectedHeap* gch = GenCollectedHeap::heap();
  3575       size_policy()->ms_collection_marking_end(gch->gc_cause());
  3578   verify_overflow_empty();
  3579   return res;
  3582 bool CMSCollector::markFromRootsWork(bool asynch) {
  3583   // iterate over marked bits in bit map, doing a full scan and mark
  3584   // from these roots using the following algorithm:
  3585   // . if oop is to the right of the current scan pointer,
  3586   //   mark corresponding bit (we'll process it later)
  3587   // . else (oop is to left of current scan pointer)
  3588   //   push oop on marking stack
  3589   // . drain the marking stack
  3591   // Note that when we do a marking step we need to hold the
  3592   // bit map lock -- recall that direct allocation (by mutators)
  3593   // and promotion (by younger generation collectors) is also
  3594   // marking the bit map. [the so-called allocate live policy.]
  3595   // Because the implementation of bit map marking is not
  3596   // robust wrt simultaneous marking of bits in the same word,
  3597   // we need to make sure that there is no such interference
  3598   // between concurrent such updates.
  3600   // already have locks
  3601   assert_lock_strong(bitMapLock());
  3603   // Clear the revisit stack, just in case there are any
  3604   // obsolete contents from a short-circuited previous CMS cycle.
  3605   _revisitStack.reset();
  3606   verify_work_stacks_empty();
  3607   verify_overflow_empty();
  3608   assert(_revisitStack.isEmpty(), "tabula rasa");
  3610   bool result = false;
  3611   if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
  3612     result = do_marking_mt(asynch);
  3613   } else {
  3614     result = do_marking_st(asynch);
  3616   return result;
  3619 // Forward decl
  3620 class CMSConcMarkingTask;
  3622 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
  3623   CMSCollector*       _collector;
  3624   CMSConcMarkingTask* _task;
  3625   bool _yield;
  3626  protected:
  3627   virtual void yield();
  3628  public:
  3629   // "n_threads" is the number of threads to be terminated.
  3630   // "queue_set" is a set of work queues of other threads.
  3631   // "collector" is the CMS collector associated with this task terminator.
  3632   // "yield" indicates whether we need the gang as a whole to yield.
  3633   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
  3634                            CMSCollector* collector, bool yield) :
  3635     ParallelTaskTerminator(n_threads, queue_set),
  3636     _collector(collector),
  3637     _yield(yield) { }
  3639   void set_task(CMSConcMarkingTask* task) {
  3640     _task = task;
  3642 };
  3644 // MT Concurrent Marking Task
  3645 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
  3646   CMSCollector* _collector;
  3647   YieldingFlexibleWorkGang* _workers;        // the whole gang
  3648   int           _n_workers;                  // requested/desired # workers
  3649   bool          _asynch;
  3650   bool          _result;
  3651   CompactibleFreeListSpace*  _cms_space;
  3652   CompactibleFreeListSpace* _perm_space;
  3653   HeapWord*     _global_finger;
  3654   HeapWord*     _restart_addr;
  3656   //  Exposed here for yielding support
  3657   Mutex* const _bit_map_lock;
  3659   // The per thread work queues, available here for stealing
  3660   OopTaskQueueSet*  _task_queues;
  3661   CMSConcMarkingTerminator _term;
  3663  public:
  3664   CMSConcMarkingTask(CMSCollector* collector,
  3665                  CompactibleFreeListSpace* cms_space,
  3666                  CompactibleFreeListSpace* perm_space,
  3667                  bool asynch, int n_workers,
  3668                  YieldingFlexibleWorkGang* workers,
  3669                  OopTaskQueueSet* task_queues):
  3670     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
  3671     _collector(collector),
  3672     _cms_space(cms_space),
  3673     _perm_space(perm_space),
  3674     _asynch(asynch), _n_workers(n_workers), _result(true),
  3675     _workers(workers), _task_queues(task_queues),
  3676     _term(n_workers, task_queues, _collector, asynch),
  3677     _bit_map_lock(collector->bitMapLock())
  3679     assert(n_workers <= workers->total_workers(),
  3680            "Else termination won't work correctly today"); // XXX FIX ME!
  3681     _requested_size = n_workers;
  3682     _term.set_task(this);
  3683     assert(_cms_space->bottom() < _perm_space->bottom(),
  3684            "Finger incorrectly initialized below");
  3685     _restart_addr = _global_finger = _cms_space->bottom();
  3689   OopTaskQueueSet* task_queues()  { return _task_queues; }
  3691   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  3693   HeapWord** global_finger_addr() { return &_global_finger; }
  3695   CMSConcMarkingTerminator* terminator() { return &_term; }
  3697   void work(int i);
  3699   virtual void coordinator_yield();  // stuff done by coordinator
  3700   bool result() { return _result; }
  3702   void reset(HeapWord* ra) {
  3703     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
  3704     assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
  3705     assert(ra             <  _perm_space->end(), "ra too large");
  3706     _restart_addr = _global_finger = ra;
  3707     _term.reset_for_reuse();
  3710   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  3711                                            OopTaskQueue* work_q);
  3713  private:
  3714   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
  3715   void do_work_steal(int i);
  3716   void bump_global_finger(HeapWord* f);
  3717 };
  3719 void CMSConcMarkingTerminator::yield() {
  3720   if (ConcurrentMarkSweepThread::should_yield() &&
  3721       !_collector->foregroundGCIsActive() &&
  3722       _yield) {
  3723     _task->yield();
  3724   } else {
  3725     ParallelTaskTerminator::yield();
  3729 ////////////////////////////////////////////////////////////////
  3730 // Concurrent Marking Algorithm Sketch
  3731 ////////////////////////////////////////////////////////////////
  3732 // Until all tasks exhausted (both spaces):
  3733 // -- claim next available chunk
  3734 // -- bump global finger via CAS
  3735 // -- find first object that starts in this chunk
  3736 //    and start scanning bitmap from that position
  3737 // -- scan marked objects for oops
  3738 // -- CAS-mark target, and if successful:
  3739 //    . if target oop is above global finger (volatile read)
  3740 //      nothing to do
  3741 //    . if target oop is in chunk and above local finger
  3742 //        then nothing to do
  3743 //    . else push on work-queue
  3744 // -- Deal with possible overflow issues:
  3745 //    . local work-queue overflow causes stuff to be pushed on
  3746 //      global (common) overflow queue
  3747 //    . always first empty local work queue
  3748 //    . then get a batch of oops from global work queue if any
  3749 //    . then do work stealing
  3750 // -- When all tasks claimed (both spaces)
  3751 //    and local work queue empty,
  3752 //    then in a loop do:
  3753 //    . check global overflow stack; steal a batch of oops and trace
  3754 //    . try to steal from other threads oif GOS is empty
  3755 //    . if neither is available, offer termination
  3756 // -- Terminate and return result
  3757 //
  3758 void CMSConcMarkingTask::work(int i) {
  3759   elapsedTimer _timer;
  3760   ResourceMark rm;
  3761   HandleMark hm;
  3763   DEBUG_ONLY(_collector->verify_overflow_empty();)
  3765   // Before we begin work, our work queue should be empty
  3766   assert(work_queue(i)->size() == 0, "Expected to be empty");
  3767   // Scan the bitmap covering _cms_space, tracing through grey objects.
  3768   _timer.start();
  3769   do_scan_and_mark(i, _cms_space);
  3770   _timer.stop();
  3771   if (PrintCMSStatistics != 0) {
  3772     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
  3773       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3776   // ... do the same for the _perm_space
  3777   _timer.reset();
  3778   _timer.start();
  3779   do_scan_and_mark(i, _perm_space);
  3780   _timer.stop();
  3781   if (PrintCMSStatistics != 0) {
  3782     gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
  3783       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3786   // ... do work stealing
  3787   _timer.reset();
  3788   _timer.start();
  3789   do_work_steal(i);
  3790   _timer.stop();
  3791   if (PrintCMSStatistics != 0) {
  3792     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
  3793       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3795   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
  3796   assert(work_queue(i)->size() == 0, "Should have been emptied");
  3797   // Note that under the current task protocol, the
  3798   // following assertion is true even of the spaces
  3799   // expanded since the completion of the concurrent
  3800   // marking. XXX This will likely change under a strict
  3801   // ABORT semantics.
  3802   assert(_global_finger >  _cms_space->end() &&
  3803          _global_finger >= _perm_space->end(),
  3804          "All tasks have been completed");
  3805   DEBUG_ONLY(_collector->verify_overflow_empty();)
  3808 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
  3809   HeapWord* read = _global_finger;
  3810   HeapWord* cur  = read;
  3811   while (f > read) {
  3812     cur = read;
  3813     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
  3814     if (cur == read) {
  3815       // our cas succeeded
  3816       assert(_global_finger >= f, "protocol consistency");
  3817       break;
  3822 // This is really inefficient, and should be redone by
  3823 // using (not yet available) block-read and -write interfaces to the
  3824 // stack and the work_queue. XXX FIX ME !!!
  3825 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  3826                                                       OopTaskQueue* work_q) {
  3827   // Fast lock-free check
  3828   if (ovflw_stk->length() == 0) {
  3829     return false;
  3831   assert(work_q->size() == 0, "Shouldn't steal");
  3832   MutexLockerEx ml(ovflw_stk->par_lock(),
  3833                    Mutex::_no_safepoint_check_flag);
  3834   // Grab up to 1/4 the size of the work queue
  3835   size_t num = MIN2((size_t)work_q->max_elems()/4,
  3836                     (size_t)ParGCDesiredObjsFromOverflowList);
  3837   num = MIN2(num, ovflw_stk->length());
  3838   for (int i = (int) num; i > 0; i--) {
  3839     oop cur = ovflw_stk->pop();
  3840     assert(cur != NULL, "Counted wrong?");
  3841     work_q->push(cur);
  3843   return num > 0;
  3846 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
  3847   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  3848   int n_tasks = pst->n_tasks();
  3849   // We allow that there may be no tasks to do here because
  3850   // we are restarting after a stack overflow.
  3851   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
  3852   int nth_task = 0;
  3854   HeapWord* aligned_start = sp->bottom();
  3855   if (sp->used_region().contains(_restart_addr)) {
  3856     // Align down to a card boundary for the start of 0th task
  3857     // for this space.
  3858     aligned_start =
  3859       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
  3860                                  CardTableModRefBS::card_size);
  3863   size_t chunk_size = sp->marking_task_size();
  3864   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  3865     // Having claimed the nth task in this space,
  3866     // compute the chunk that it corresponds to:
  3867     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
  3868                                aligned_start + (nth_task+1)*chunk_size);
  3869     // Try and bump the global finger via a CAS;
  3870     // note that we need to do the global finger bump
  3871     // _before_ taking the intersection below, because
  3872     // the task corresponding to that region will be
  3873     // deemed done even if the used_region() expands
  3874     // because of allocation -- as it almost certainly will
  3875     // during start-up while the threads yield in the
  3876     // closure below.
  3877     HeapWord* finger = span.end();
  3878     bump_global_finger(finger);   // atomically
  3879     // There are null tasks here corresponding to chunks
  3880     // beyond the "top" address of the space.
  3881     span = span.intersection(sp->used_region());
  3882     if (!span.is_empty()) {  // Non-null task
  3883       HeapWord* prev_obj;
  3884       assert(!span.contains(_restart_addr) || nth_task == 0,
  3885              "Inconsistency");
  3886       if (nth_task == 0) {
  3887         // For the 0th task, we'll not need to compute a block_start.
  3888         if (span.contains(_restart_addr)) {
  3889           // In the case of a restart because of stack overflow,
  3890           // we might additionally skip a chunk prefix.
  3891           prev_obj = _restart_addr;
  3892         } else {
  3893           prev_obj = span.start();
  3895       } else {
  3896         // We want to skip the first object because
  3897         // the protocol is to scan any object in its entirety
  3898         // that _starts_ in this span; a fortiori, any
  3899         // object starting in an earlier span is scanned
  3900         // as part of an earlier claimed task.
  3901         // Below we use the "careful" version of block_start
  3902         // so we do not try to navigate uninitialized objects.
  3903         prev_obj = sp->block_start_careful(span.start());
  3904         // Below we use a variant of block_size that uses the
  3905         // Printezis bits to avoid waiting for allocated
  3906         // objects to become initialized/parsable.
  3907         while (prev_obj < span.start()) {
  3908           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
  3909           if (sz > 0) {
  3910             prev_obj += sz;
  3911           } else {
  3912             // In this case we may end up doing a bit of redundant
  3913             // scanning, but that appears unavoidable, short of
  3914             // locking the free list locks; see bug 6324141.
  3915             break;
  3919       if (prev_obj < span.end()) {
  3920         MemRegion my_span = MemRegion(prev_obj, span.end());
  3921         // Do the marking work within a non-empty span --
  3922         // the last argument to the constructor indicates whether the
  3923         // iteration should be incremental with periodic yields.
  3924         Par_MarkFromRootsClosure cl(this, _collector, my_span,
  3925                                     &_collector->_markBitMap,
  3926                                     work_queue(i),
  3927                                     &_collector->_markStack,
  3928                                     &_collector->_revisitStack,
  3929                                     _asynch);
  3930         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
  3931       } // else nothing to do for this task
  3932     }   // else nothing to do for this task
  3934   // We'd be tempted to assert here that since there are no
  3935   // more tasks left to claim in this space, the global_finger
  3936   // must exceed space->top() and a fortiori space->end(). However,
  3937   // that would not quite be correct because the bumping of
  3938   // global_finger occurs strictly after the claiming of a task,
  3939   // so by the time we reach here the global finger may not yet
  3940   // have been bumped up by the thread that claimed the last
  3941   // task.
  3942   pst->all_tasks_completed();
  3945 class Par_ConcMarkingClosure: public OopClosure {
  3946  private:
  3947   CMSCollector* _collector;
  3948   MemRegion     _span;
  3949   CMSBitMap*    _bit_map;
  3950   CMSMarkStack* _overflow_stack;
  3951   CMSMarkStack* _revisit_stack;     // XXXXXX Check proper use
  3952   OopTaskQueue* _work_queue;
  3953  protected:
  3954   DO_OOP_WORK_DEFN
  3955  public:
  3956   Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
  3957                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
  3958     _collector(collector),
  3959     _span(_collector->_span),
  3960     _work_queue(work_queue),
  3961     _bit_map(bit_map),
  3962     _overflow_stack(overflow_stack) { }   // need to initialize revisit stack etc.
  3963   virtual void do_oop(oop* p);
  3964   virtual void do_oop(narrowOop* p);
  3965   void trim_queue(size_t max);
  3966   void handle_stack_overflow(HeapWord* lost);
  3967 };
  3969 // Grey object scanning during work stealing phase --
  3970 // the salient assumption here is that any references
  3971 // that are in these stolen objects being scanned must
  3972 // already have been initialized (else they would not have
  3973 // been published), so we do not need to check for
  3974 // uninitialized objects before pushing here.
  3975 void Par_ConcMarkingClosure::do_oop(oop obj) {
  3976   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  3977   HeapWord* addr = (HeapWord*)obj;
  3978   // Check if oop points into the CMS generation
  3979   // and is not marked
  3980   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  3981     // a white object ...
  3982     // If we manage to "claim" the object, by being the
  3983     // first thread to mark it, then we push it on our
  3984     // marking stack
  3985     if (_bit_map->par_mark(addr)) {     // ... now grey
  3986       // push on work queue (grey set)
  3987       bool simulate_overflow = false;
  3988       NOT_PRODUCT(
  3989         if (CMSMarkStackOverflowALot &&
  3990             _collector->simulate_overflow()) {
  3991           // simulate a stack overflow
  3992           simulate_overflow = true;
  3995       if (simulate_overflow ||
  3996           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  3997         // stack overflow
  3998         if (PrintCMSStatistics != 0) {
  3999           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  4000                                  SIZE_FORMAT, _overflow_stack->capacity());
  4002         // We cannot assert that the overflow stack is full because
  4003         // it may have been emptied since.
  4004         assert(simulate_overflow ||
  4005                _work_queue->size() == _work_queue->max_elems(),
  4006               "Else push should have succeeded");
  4007         handle_stack_overflow(addr);
  4009     } // Else, some other thread got there first
  4013 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
  4014 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
  4016 void Par_ConcMarkingClosure::trim_queue(size_t max) {
  4017   while (_work_queue->size() > max) {
  4018     oop new_oop;
  4019     if (_work_queue->pop_local(new_oop)) {
  4020       assert(new_oop->is_oop(), "Should be an oop");
  4021       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
  4022       assert(_span.contains((HeapWord*)new_oop), "Not in span");
  4023       assert(new_oop->is_parsable(), "Should be parsable");
  4024       new_oop->oop_iterate(this);  // do_oop() above
  4029 // Upon stack overflow, we discard (part of) the stack,
  4030 // remembering the least address amongst those discarded
  4031 // in CMSCollector's _restart_address.
  4032 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
  4033   // We need to do this under a mutex to prevent other
  4034   // workers from interfering with the work done below.
  4035   MutexLockerEx ml(_overflow_stack->par_lock(),
  4036                    Mutex::_no_safepoint_check_flag);
  4037   // Remember the least grey address discarded
  4038   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  4039   _collector->lower_restart_addr(ra);
  4040   _overflow_stack->reset();  // discard stack contents
  4041   _overflow_stack->expand(); // expand the stack if possible
  4045 void CMSConcMarkingTask::do_work_steal(int i) {
  4046   OopTaskQueue* work_q = work_queue(i);
  4047   oop obj_to_scan;
  4048   CMSBitMap* bm = &(_collector->_markBitMap);
  4049   CMSMarkStack* ovflw = &(_collector->_markStack);
  4050   int* seed = _collector->hash_seed(i);
  4051   Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
  4052   while (true) {
  4053     cl.trim_queue(0);
  4054     assert(work_q->size() == 0, "Should have been emptied above");
  4055     if (get_work_from_overflow_stack(ovflw, work_q)) {
  4056       // Can't assert below because the work obtained from the
  4057       // overflow stack may already have been stolen from us.
  4058       // assert(work_q->size() > 0, "Work from overflow stack");
  4059       continue;
  4060     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  4061       assert(obj_to_scan->is_oop(), "Should be an oop");
  4062       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
  4063       obj_to_scan->oop_iterate(&cl);
  4064     } else if (terminator()->offer_termination()) {
  4065       assert(work_q->size() == 0, "Impossible!");
  4066       break;
  4071 // This is run by the CMS (coordinator) thread.
  4072 void CMSConcMarkingTask::coordinator_yield() {
  4073   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4074          "CMS thread should hold CMS token");
  4076   // First give up the locks, then yield, then re-lock
  4077   // We should probably use a constructor/destructor idiom to
  4078   // do this unlock/lock or modify the MutexUnlocker class to
  4079   // serve our purpose. XXX
  4080   assert_lock_strong(_bit_map_lock);
  4081   _bit_map_lock->unlock();
  4082   ConcurrentMarkSweepThread::desynchronize(true);
  4083   ConcurrentMarkSweepThread::acknowledge_yield_request();
  4084   _collector->stopTimer();
  4085   if (PrintCMSStatistics != 0) {
  4086     _collector->incrementYields();
  4088   _collector->icms_wait();
  4090   // It is possible for whichever thread initiated the yield request
  4091   // not to get a chance to wake up and take the bitmap lock between
  4092   // this thread releasing it and reacquiring it. So, while the
  4093   // should_yield() flag is on, let's sleep for a bit to give the
  4094   // other thread a chance to wake up. The limit imposed on the number
  4095   // of iterations is defensive, to avoid any unforseen circumstances
  4096   // putting us into an infinite loop. Since it's always been this
  4097   // (coordinator_yield()) method that was observed to cause the
  4098   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
  4099   // which is by default non-zero. For the other seven methods that
  4100   // also perform the yield operation, as are using a different
  4101   // parameter (CMSYieldSleepCount) which is by default zero. This way we
  4102   // can enable the sleeping for those methods too, if necessary.
  4103   // See 6442774.
  4104   //
  4105   // We really need to reconsider the synchronization between the GC
  4106   // thread and the yield-requesting threads in the future and we
  4107   // should really use wait/notify, which is the recommended
  4108   // way of doing this type of interaction. Additionally, we should
  4109   // consolidate the eight methods that do the yield operation and they
  4110   // are almost identical into one for better maintenability and
  4111   // readability. See 6445193.
  4112   //
  4113   // Tony 2006.06.29
  4114   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
  4115                    ConcurrentMarkSweepThread::should_yield() &&
  4116                    !CMSCollector::foregroundGCIsActive(); ++i) {
  4117     os::sleep(Thread::current(), 1, false);
  4118     ConcurrentMarkSweepThread::acknowledge_yield_request();
  4121   ConcurrentMarkSweepThread::synchronize(true);
  4122   _bit_map_lock->lock_without_safepoint_check();
  4123   _collector->startTimer();
  4126 bool CMSCollector::do_marking_mt(bool asynch) {
  4127   assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
  4128   // In the future this would be determined ergonomically, based
  4129   // on #cpu's, # active mutator threads (and load), and mutation rate.
  4130   int num_workers = ParallelCMSThreads;
  4132   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  4133   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
  4135   CMSConcMarkingTask tsk(this, cms_space, perm_space,
  4136                          asynch, num_workers /* number requested XXX */,
  4137                          conc_workers(), task_queues());
  4139   // Since the actual number of workers we get may be different
  4140   // from the number we requested above, do we need to do anything different
  4141   // below? In particular, may be we need to subclass the SequantialSubTasksDone
  4142   // class?? XXX
  4143   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
  4144   perm_space->initialize_sequential_subtasks_for_marking(num_workers);
  4146   // Refs discovery is already non-atomic.
  4147   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
  4148   // Mutate the Refs discovery so it is MT during the
  4149   // multi-threaded marking phase.
  4150   ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
  4152   conc_workers()->start_task(&tsk);
  4153   while (tsk.yielded()) {
  4154     tsk.coordinator_yield();
  4155     conc_workers()->continue_task(&tsk);
  4157   // If the task was aborted, _restart_addr will be non-NULL
  4158   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
  4159   while (_restart_addr != NULL) {
  4160     // XXX For now we do not make use of ABORTED state and have not
  4161     // yet implemented the right abort semantics (even in the original
  4162     // single-threaded CMS case). That needs some more investigation
  4163     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
  4164     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
  4165     // If _restart_addr is non-NULL, a marking stack overflow
  4166     // occured; we need to do a fresh marking iteration from the
  4167     // indicated restart address.
  4168     if (_foregroundGCIsActive && asynch) {
  4169       // We may be running into repeated stack overflows, having
  4170       // reached the limit of the stack size, while making very
  4171       // slow forward progress. It may be best to bail out and
  4172       // let the foreground collector do its job.
  4173       // Clear _restart_addr, so that foreground GC
  4174       // works from scratch. This avoids the headache of
  4175       // a "rescan" which would otherwise be needed because
  4176       // of the dirty mod union table & card table.
  4177       _restart_addr = NULL;
  4178       return false;
  4180     // Adjust the task to restart from _restart_addr
  4181     tsk.reset(_restart_addr);
  4182     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
  4183                   _restart_addr);
  4184     perm_space->initialize_sequential_subtasks_for_marking(num_workers,
  4185                   _restart_addr);
  4186     _restart_addr = NULL;
  4187     // Get the workers going again
  4188     conc_workers()->start_task(&tsk);
  4189     while (tsk.yielded()) {
  4190       tsk.coordinator_yield();
  4191       conc_workers()->continue_task(&tsk);
  4194   assert(tsk.completed(), "Inconsistency");
  4195   assert(tsk.result() == true, "Inconsistency");
  4196   return true;
  4199 bool CMSCollector::do_marking_st(bool asynch) {
  4200   ResourceMark rm;
  4201   HandleMark   hm;
  4203   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
  4204     &_markStack, &_revisitStack, CMSYield && asynch);
  4205   // the last argument to iterate indicates whether the iteration
  4206   // should be incremental with periodic yields.
  4207   _markBitMap.iterate(&markFromRootsClosure);
  4208   // If _restart_addr is non-NULL, a marking stack overflow
  4209   // occured; we need to do a fresh iteration from the
  4210   // indicated restart address.
  4211   while (_restart_addr != NULL) {
  4212     if (_foregroundGCIsActive && asynch) {
  4213       // We may be running into repeated stack overflows, having
  4214       // reached the limit of the stack size, while making very
  4215       // slow forward progress. It may be best to bail out and
  4216       // let the foreground collector do its job.
  4217       // Clear _restart_addr, so that foreground GC
  4218       // works from scratch. This avoids the headache of
  4219       // a "rescan" which would otherwise be needed because
  4220       // of the dirty mod union table & card table.
  4221       _restart_addr = NULL;
  4222       return false;  // indicating failure to complete marking
  4224     // Deal with stack overflow:
  4225     // we restart marking from _restart_addr
  4226     HeapWord* ra = _restart_addr;
  4227     markFromRootsClosure.reset(ra);
  4228     _restart_addr = NULL;
  4229     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
  4231   return true;
  4234 void CMSCollector::preclean() {
  4235   check_correct_thread_executing();
  4236   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
  4237   verify_work_stacks_empty();
  4238   verify_overflow_empty();
  4239   _abort_preclean = false;
  4240   if (CMSPrecleaningEnabled) {
  4241     _eden_chunk_index = 0;
  4242     size_t used = get_eden_used();
  4243     size_t capacity = get_eden_capacity();
  4244     // Don't start sampling unless we will get sufficiently
  4245     // many samples.
  4246     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
  4247                 * CMSScheduleRemarkEdenPenetration)) {
  4248       _start_sampling = true;
  4249     } else {
  4250       _start_sampling = false;
  4252     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4253     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
  4254     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
  4256   CMSTokenSync x(true); // is cms thread
  4257   if (CMSPrecleaningEnabled) {
  4258     sample_eden();
  4259     _collectorState = AbortablePreclean;
  4260   } else {
  4261     _collectorState = FinalMarking;
  4263   verify_work_stacks_empty();
  4264   verify_overflow_empty();
  4267 // Try and schedule the remark such that young gen
  4268 // occupancy is CMSScheduleRemarkEdenPenetration %.
  4269 void CMSCollector::abortable_preclean() {
  4270   check_correct_thread_executing();
  4271   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
  4272   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
  4274   // If Eden's current occupancy is below this threshold,
  4275   // immediately schedule the remark; else preclean
  4276   // past the next scavenge in an effort to
  4277   // schedule the pause as described avove. By choosing
  4278   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
  4279   // we will never do an actual abortable preclean cycle.
  4280   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
  4281     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4282     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
  4283     // We need more smarts in the abortable preclean
  4284     // loop below to deal with cases where allocation
  4285     // in young gen is very very slow, and our precleaning
  4286     // is running a losing race against a horde of
  4287     // mutators intent on flooding us with CMS updates
  4288     // (dirty cards).
  4289     // One, admittedly dumb, strategy is to give up
  4290     // after a certain number of abortable precleaning loops
  4291     // or after a certain maximum time. We want to make
  4292     // this smarter in the next iteration.
  4293     // XXX FIX ME!!! YSR
  4294     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
  4295     while (!(should_abort_preclean() ||
  4296              ConcurrentMarkSweepThread::should_terminate())) {
  4297       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
  4298       cumworkdone += workdone;
  4299       loops++;
  4300       // Voluntarily terminate abortable preclean phase if we have
  4301       // been at it for too long.
  4302       if ((CMSMaxAbortablePrecleanLoops != 0) &&
  4303           loops >= CMSMaxAbortablePrecleanLoops) {
  4304         if (PrintGCDetails) {
  4305           gclog_or_tty->print(" CMS: abort preclean due to loops ");
  4307         break;
  4309       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
  4310         if (PrintGCDetails) {
  4311           gclog_or_tty->print(" CMS: abort preclean due to time ");
  4313         break;
  4315       // If we are doing little work each iteration, we should
  4316       // take a short break.
  4317       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
  4318         // Sleep for some time, waiting for work to accumulate
  4319         stopTimer();
  4320         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
  4321         startTimer();
  4322         waited++;
  4325     if (PrintCMSStatistics > 0) {
  4326       gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
  4327                           loops, waited, cumworkdone);
  4330   CMSTokenSync x(true); // is cms thread
  4331   if (_collectorState != Idling) {
  4332     assert(_collectorState == AbortablePreclean,
  4333            "Spontaneous state transition?");
  4334     _collectorState = FinalMarking;
  4335   } // Else, a foreground collection completed this CMS cycle.
  4336   return;
  4339 // Respond to an Eden sampling opportunity
  4340 void CMSCollector::sample_eden() {
  4341   // Make sure a young gc cannot sneak in between our
  4342   // reading and recording of a sample.
  4343   assert(Thread::current()->is_ConcurrentGC_thread(),
  4344          "Only the cms thread may collect Eden samples");
  4345   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4346          "Should collect samples while holding CMS token");
  4347   if (!_start_sampling) {
  4348     return;
  4350   if (_eden_chunk_array) {
  4351     if (_eden_chunk_index < _eden_chunk_capacity) {
  4352       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
  4353       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
  4354              "Unexpected state of Eden");
  4355       // We'd like to check that what we just sampled is an oop-start address;
  4356       // however, we cannot do that here since the object may not yet have been
  4357       // initialized. So we'll instead do the check when we _use_ this sample
  4358       // later.
  4359       if (_eden_chunk_index == 0 ||
  4360           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
  4361                          _eden_chunk_array[_eden_chunk_index-1])
  4362            >= CMSSamplingGrain)) {
  4363         _eden_chunk_index++;  // commit sample
  4367   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
  4368     size_t used = get_eden_used();
  4369     size_t capacity = get_eden_capacity();
  4370     assert(used <= capacity, "Unexpected state of Eden");
  4371     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
  4372       _abort_preclean = true;
  4378 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
  4379   assert(_collectorState == Precleaning ||
  4380          _collectorState == AbortablePreclean, "incorrect state");
  4381   ResourceMark rm;
  4382   HandleMark   hm;
  4383   // Do one pass of scrubbing the discovered reference lists
  4384   // to remove any reference objects with strongly-reachable
  4385   // referents.
  4386   if (clean_refs) {
  4387     ReferenceProcessor* rp = ref_processor();
  4388     CMSPrecleanRefsYieldClosure yield_cl(this);
  4389     assert(rp->span().equals(_span), "Spans should be equal");
  4390     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
  4391                                    &_markStack);
  4392     CMSDrainMarkingStackClosure complete_trace(this,
  4393                                   _span, &_markBitMap, &_markStack,
  4394                                   &keep_alive);
  4396     // We don't want this step to interfere with a young
  4397     // collection because we don't want to take CPU
  4398     // or memory bandwidth away from the young GC threads
  4399     // (which may be as many as there are CPUs).
  4400     // Note that we don't need to protect ourselves from
  4401     // interference with mutators because they can't
  4402     // manipulate the discovered reference lists nor affect
  4403     // the computed reachability of the referents, the
  4404     // only properties manipulated by the precleaning
  4405     // of these reference lists.
  4406     stopTimer();
  4407     CMSTokenSyncWithLocks x(true /* is cms thread */,
  4408                             bitMapLock());
  4409     startTimer();
  4410     sample_eden();
  4411     // The following will yield to allow foreground
  4412     // collection to proceed promptly. XXX YSR:
  4413     // The code in this method may need further
  4414     // tweaking for better performance and some restructuring
  4415     // for cleaner interfaces.
  4416     rp->preclean_discovered_references(
  4417           rp->is_alive_non_header(), &keep_alive, &complete_trace,
  4418           &yield_cl);
  4421   if (clean_survivor) {  // preclean the active survivor space(s)
  4422     assert(_young_gen->kind() == Generation::DefNew ||
  4423            _young_gen->kind() == Generation::ParNew ||
  4424            _young_gen->kind() == Generation::ASParNew,
  4425          "incorrect type for cast");
  4426     DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  4427     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
  4428                              &_markBitMap, &_modUnionTable,
  4429                              &_markStack, &_revisitStack,
  4430                              true /* precleaning phase */);
  4431     stopTimer();
  4432     CMSTokenSyncWithLocks ts(true /* is cms thread */,
  4433                              bitMapLock());
  4434     startTimer();
  4435     unsigned int before_count =
  4436       GenCollectedHeap::heap()->total_collections();
  4437     SurvivorSpacePrecleanClosure
  4438       sss_cl(this, _span, &_markBitMap, &_markStack,
  4439              &pam_cl, before_count, CMSYield);
  4440     dng->from()->object_iterate_careful(&sss_cl);
  4441     dng->to()->object_iterate_careful(&sss_cl);
  4443   MarkRefsIntoAndScanClosure
  4444     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
  4445              &_markStack, &_revisitStack, this, CMSYield,
  4446              true /* precleaning phase */);
  4447   // CAUTION: The following closure has persistent state that may need to
  4448   // be reset upon a decrease in the sequence of addresses it
  4449   // processes.
  4450   ScanMarkedObjectsAgainCarefullyClosure
  4451     smoac_cl(this, _span,
  4452       &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
  4454   // Preclean dirty cards in ModUnionTable and CardTable using
  4455   // appropriate convergence criterion;
  4456   // repeat CMSPrecleanIter times unless we find that
  4457   // we are losing.
  4458   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
  4459   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
  4460          "Bad convergence multiplier");
  4461   assert(CMSPrecleanThreshold >= 100,
  4462          "Unreasonably low CMSPrecleanThreshold");
  4464   size_t numIter, cumNumCards, lastNumCards, curNumCards;
  4465   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
  4466        numIter < CMSPrecleanIter;
  4467        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
  4468     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
  4469     if (CMSPermGenPrecleaningEnabled) {
  4470       curNumCards  += preclean_mod_union_table(_permGen, &smoac_cl);
  4472     if (Verbose && PrintGCDetails) {
  4473       gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
  4475     // Either there are very few dirty cards, so re-mark
  4476     // pause will be small anyway, or our pre-cleaning isn't
  4477     // that much faster than the rate at which cards are being
  4478     // dirtied, so we might as well stop and re-mark since
  4479     // precleaning won't improve our re-mark time by much.
  4480     if (curNumCards <= CMSPrecleanThreshold ||
  4481         (numIter > 0 &&
  4482          (curNumCards * CMSPrecleanDenominator >
  4483          lastNumCards * CMSPrecleanNumerator))) {
  4484       numIter++;
  4485       cumNumCards += curNumCards;
  4486       break;
  4489   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
  4490   if (CMSPermGenPrecleaningEnabled) {
  4491     curNumCards += preclean_card_table(_permGen, &smoac_cl);
  4493   cumNumCards += curNumCards;
  4494   if (PrintGCDetails && PrintCMSStatistics != 0) {
  4495     gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
  4496                   curNumCards, cumNumCards, numIter);
  4498   return cumNumCards;   // as a measure of useful work done
  4501 // PRECLEANING NOTES:
  4502 // Precleaning involves:
  4503 // . reading the bits of the modUnionTable and clearing the set bits.
  4504 // . For the cards corresponding to the set bits, we scan the
  4505 //   objects on those cards. This means we need the free_list_lock
  4506 //   so that we can safely iterate over the CMS space when scanning
  4507 //   for oops.
  4508 // . When we scan the objects, we'll be both reading and setting
  4509 //   marks in the marking bit map, so we'll need the marking bit map.
  4510 // . For protecting _collector_state transitions, we take the CGC_lock.
  4511 //   Note that any races in the reading of of card table entries by the
  4512 //   CMS thread on the one hand and the clearing of those entries by the
  4513 //   VM thread or the setting of those entries by the mutator threads on the
  4514 //   other are quite benign. However, for efficiency it makes sense to keep
  4515 //   the VM thread from racing with the CMS thread while the latter is
  4516 //   dirty card info to the modUnionTable. We therefore also use the
  4517 //   CGC_lock to protect the reading of the card table and the mod union
  4518 //   table by the CM thread.
  4519 // . We run concurrently with mutator updates, so scanning
  4520 //   needs to be done carefully  -- we should not try to scan
  4521 //   potentially uninitialized objects.
  4522 //
  4523 // Locking strategy: While holding the CGC_lock, we scan over and
  4524 // reset a maximal dirty range of the mod union / card tables, then lock
  4525 // the free_list_lock and bitmap lock to do a full marking, then
  4526 // release these locks; and repeat the cycle. This allows for a
  4527 // certain amount of fairness in the sharing of these locks between
  4528 // the CMS collector on the one hand, and the VM thread and the
  4529 // mutators on the other.
  4531 // NOTE: preclean_mod_union_table() and preclean_card_table()
  4532 // further below are largely identical; if you need to modify
  4533 // one of these methods, please check the other method too.
  4535 size_t CMSCollector::preclean_mod_union_table(
  4536   ConcurrentMarkSweepGeneration* gen,
  4537   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4538   verify_work_stacks_empty();
  4539   verify_overflow_empty();
  4541   // strategy: starting with the first card, accumulate contiguous
  4542   // ranges of dirty cards; clear these cards, then scan the region
  4543   // covered by these cards.
  4545   // Since all of the MUT is committed ahead, we can just use
  4546   // that, in case the generations expand while we are precleaning.
  4547   // It might also be fine to just use the committed part of the
  4548   // generation, but we might potentially miss cards when the
  4549   // generation is rapidly expanding while we are in the midst
  4550   // of precleaning.
  4551   HeapWord* startAddr = gen->reserved().start();
  4552   HeapWord* endAddr   = gen->reserved().end();
  4554   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4556   size_t numDirtyCards, cumNumDirtyCards;
  4557   HeapWord *nextAddr, *lastAddr;
  4558   for (cumNumDirtyCards = numDirtyCards = 0,
  4559        nextAddr = lastAddr = startAddr;
  4560        nextAddr < endAddr;
  4561        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4563     ResourceMark rm;
  4564     HandleMark   hm;
  4566     MemRegion dirtyRegion;
  4568       stopTimer();
  4569       CMSTokenSync ts(true);
  4570       startTimer();
  4571       sample_eden();
  4572       // Get dirty region starting at nextOffset (inclusive),
  4573       // simultaneously clearing it.
  4574       dirtyRegion =
  4575         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
  4576       assert(dirtyRegion.start() >= nextAddr,
  4577              "returned region inconsistent?");
  4579     // Remember where the next search should begin.
  4580     // The returned region (if non-empty) is a right open interval,
  4581     // so lastOffset is obtained from the right end of that
  4582     // interval.
  4583     lastAddr = dirtyRegion.end();
  4584     // Should do something more transparent and less hacky XXX
  4585     numDirtyCards =
  4586       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
  4588     // We'll scan the cards in the dirty region (with periodic
  4589     // yields for foreground GC as needed).
  4590     if (!dirtyRegion.is_empty()) {
  4591       assert(numDirtyCards > 0, "consistency check");
  4592       HeapWord* stop_point = NULL;
  4594         stopTimer();
  4595         CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
  4596                                  bitMapLock());
  4597         startTimer();
  4598         verify_work_stacks_empty();
  4599         verify_overflow_empty();
  4600         sample_eden();
  4601         stop_point =
  4602           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4604       if (stop_point != NULL) {
  4605         // The careful iteration stopped early either because it found an
  4606         // uninitialized object, or because we were in the midst of an
  4607         // "abortable preclean", which should now be aborted. Redirty
  4608         // the bits corresponding to the partially-scanned or unscanned
  4609         // cards. We'll either restart at the next block boundary or
  4610         // abort the preclean.
  4611         assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
  4612                (_collectorState == AbortablePreclean && should_abort_preclean()),
  4613                "Unparsable objects should only be in perm gen.");
  4615         stopTimer();
  4616         CMSTokenSyncWithLocks ts(true, bitMapLock());
  4617         startTimer();
  4618         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
  4619         if (should_abort_preclean()) {
  4620           break; // out of preclean loop
  4621         } else {
  4622           // Compute the next address at which preclean should pick up;
  4623           // might need bitMapLock in order to read P-bits.
  4624           lastAddr = next_card_start_after_block(stop_point);
  4627     } else {
  4628       assert(lastAddr == endAddr, "consistency check");
  4629       assert(numDirtyCards == 0, "consistency check");
  4630       break;
  4633   verify_work_stacks_empty();
  4634   verify_overflow_empty();
  4635   return cumNumDirtyCards;
  4638 // NOTE: preclean_mod_union_table() above and preclean_card_table()
  4639 // below are largely identical; if you need to modify
  4640 // one of these methods, please check the other method too.
  4642 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
  4643   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4644   // strategy: it's similar to precleamModUnionTable above, in that
  4645   // we accumulate contiguous ranges of dirty cards, mark these cards
  4646   // precleaned, then scan the region covered by these cards.
  4647   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
  4648   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
  4650   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4652   size_t numDirtyCards, cumNumDirtyCards;
  4653   HeapWord *lastAddr, *nextAddr;
  4655   for (cumNumDirtyCards = numDirtyCards = 0,
  4656        nextAddr = lastAddr = startAddr;
  4657        nextAddr < endAddr;
  4658        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4660     ResourceMark rm;
  4661     HandleMark   hm;
  4663     MemRegion dirtyRegion;
  4665       // See comments in "Precleaning notes" above on why we
  4666       // do this locking. XXX Could the locking overheads be
  4667       // too high when dirty cards are sparse? [I don't think so.]
  4668       stopTimer();
  4669       CMSTokenSync x(true); // is cms thread
  4670       startTimer();
  4671       sample_eden();
  4672       // Get and clear dirty region from card table
  4673       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
  4674                                     MemRegion(nextAddr, endAddr),
  4675                                     true,
  4676                                     CardTableModRefBS::precleaned_card_val());
  4678       assert(dirtyRegion.start() >= nextAddr,
  4679              "returned region inconsistent?");
  4681     lastAddr = dirtyRegion.end();
  4682     numDirtyCards =
  4683       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
  4685     if (!dirtyRegion.is_empty()) {
  4686       stopTimer();
  4687       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
  4688       startTimer();
  4689       sample_eden();
  4690       verify_work_stacks_empty();
  4691       verify_overflow_empty();
  4692       HeapWord* stop_point =
  4693         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4694       if (stop_point != NULL) {
  4695         // The careful iteration stopped early because it found an
  4696         // uninitialized object.  Redirty the bits corresponding to the
  4697         // partially-scanned or unscanned cards, and start again at the
  4698         // next block boundary.
  4699         assert(CMSPermGenPrecleaningEnabled ||
  4700                (_collectorState == AbortablePreclean && should_abort_preclean()),
  4701                "Unparsable objects should only be in perm gen.");
  4702         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
  4703         if (should_abort_preclean()) {
  4704           break; // out of preclean loop
  4705         } else {
  4706           // Compute the next address at which preclean should pick up.
  4707           lastAddr = next_card_start_after_block(stop_point);
  4710     } else {
  4711       break;
  4714   verify_work_stacks_empty();
  4715   verify_overflow_empty();
  4716   return cumNumDirtyCards;
  4719 void CMSCollector::checkpointRootsFinal(bool asynch,
  4720   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  4721   assert(_collectorState == FinalMarking, "incorrect state transition?");
  4722   check_correct_thread_executing();
  4723   // world is stopped at this checkpoint
  4724   assert(SafepointSynchronize::is_at_safepoint(),
  4725          "world should be stopped");
  4726   verify_work_stacks_empty();
  4727   verify_overflow_empty();
  4729   SpecializationStats::clear();
  4730   if (PrintGCDetails) {
  4731     gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
  4732                         _young_gen->used() / K,
  4733                         _young_gen->capacity() / K);
  4735   if (asynch) {
  4736     if (CMSScavengeBeforeRemark) {
  4737       GenCollectedHeap* gch = GenCollectedHeap::heap();
  4738       // Temporarily set flag to false, GCH->do_collection will
  4739       // expect it to be false and set to true
  4740       FlagSetting fl(gch->_is_gc_active, false);
  4741       NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
  4742         PrintGCDetails && Verbose, true, gclog_or_tty);)
  4743       int level = _cmsGen->level() - 1;
  4744       if (level >= 0) {
  4745         gch->do_collection(true,        // full (i.e. force, see below)
  4746                            false,       // !clear_all_soft_refs
  4747                            0,           // size
  4748                            false,       // is_tlab
  4749                            level        // max_level
  4750                           );
  4753     FreelistLocker x(this);
  4754     MutexLockerEx y(bitMapLock(),
  4755                     Mutex::_no_safepoint_check_flag);
  4756     assert(!init_mark_was_synchronous, "but that's impossible!");
  4757     checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
  4758   } else {
  4759     // already have all the locks
  4760     checkpointRootsFinalWork(asynch, clear_all_soft_refs,
  4761                              init_mark_was_synchronous);
  4763   verify_work_stacks_empty();
  4764   verify_overflow_empty();
  4765   SpecializationStats::print();
  4768 void CMSCollector::checkpointRootsFinalWork(bool asynch,
  4769   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  4771   NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
  4773   assert(haveFreelistLocks(), "must have free list locks");
  4774   assert_lock_strong(bitMapLock());
  4776   if (UseAdaptiveSizePolicy) {
  4777     size_policy()->checkpoint_roots_final_begin();
  4780   ResourceMark rm;
  4781   HandleMark   hm;
  4783   GenCollectedHeap* gch = GenCollectedHeap::heap();
  4785   if (should_unload_classes()) {
  4786     CodeCache::gc_prologue();
  4788   assert(haveFreelistLocks(), "must have free list locks");
  4789   assert_lock_strong(bitMapLock());
  4791   if (!init_mark_was_synchronous) {
  4792     // We might assume that we need not fill TLAB's when
  4793     // CMSScavengeBeforeRemark is set, because we may have just done
  4794     // a scavenge which would have filled all TLAB's -- and besides
  4795     // Eden would be empty. This however may not always be the case --
  4796     // for instance although we asked for a scavenge, it may not have
  4797     // happened because of a JNI critical section. We probably need
  4798     // a policy for deciding whether we can in that case wait until
  4799     // the critical section releases and then do the remark following
  4800     // the scavenge, and skip it here. In the absence of that policy,
  4801     // or of an indication of whether the scavenge did indeed occur,
  4802     // we cannot rely on TLAB's having been filled and must do
  4803     // so here just in case a scavenge did not happen.
  4804     gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
  4805     // Update the saved marks which may affect the root scans.
  4806     gch->save_marks();
  4809       COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  4811       // Note on the role of the mod union table:
  4812       // Since the marker in "markFromRoots" marks concurrently with
  4813       // mutators, it is possible for some reachable objects not to have been
  4814       // scanned. For instance, an only reference to an object A was
  4815       // placed in object B after the marker scanned B. Unless B is rescanned,
  4816       // A would be collected. Such updates to references in marked objects
  4817       // are detected via the mod union table which is the set of all cards
  4818       // dirtied since the first checkpoint in this GC cycle and prior to
  4819       // the most recent young generation GC, minus those cleaned up by the
  4820       // concurrent precleaning.
  4821       if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
  4822         TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
  4823         do_remark_parallel();
  4824       } else {
  4825         TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
  4826                     gclog_or_tty);
  4827         do_remark_non_parallel();
  4830   } else {
  4831     assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
  4832     // The initial mark was stop-world, so there's no rescanning to
  4833     // do; go straight on to the next step below.
  4835   verify_work_stacks_empty();
  4836   verify_overflow_empty();
  4839     NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
  4840     refProcessingWork(asynch, clear_all_soft_refs);
  4842   verify_work_stacks_empty();
  4843   verify_overflow_empty();
  4845   if (should_unload_classes()) {
  4846     CodeCache::gc_epilogue();
  4849   // If we encountered any (marking stack / work queue) overflow
  4850   // events during the current CMS cycle, take appropriate
  4851   // remedial measures, where possible, so as to try and avoid
  4852   // recurrence of that condition.
  4853   assert(_markStack.isEmpty(), "No grey objects");
  4854   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
  4855                      _ser_kac_ovflw;
  4856   if (ser_ovflw > 0) {
  4857     if (PrintCMSStatistics != 0) {
  4858       gclog_or_tty->print_cr("Marking stack overflow (benign) "
  4859         "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
  4860         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
  4861         _ser_kac_ovflw);
  4863     _markStack.expand();
  4864     _ser_pmc_remark_ovflw = 0;
  4865     _ser_pmc_preclean_ovflw = 0;
  4866     _ser_kac_ovflw = 0;
  4868   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
  4869     if (PrintCMSStatistics != 0) {
  4870       gclog_or_tty->print_cr("Work queue overflow (benign) "
  4871         "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
  4872         _par_pmc_remark_ovflw, _par_kac_ovflw);
  4874     _par_pmc_remark_ovflw = 0;
  4875     _par_kac_ovflw = 0;
  4877   if (PrintCMSStatistics != 0) {
  4878      if (_markStack._hit_limit > 0) {
  4879        gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
  4880                               _markStack._hit_limit);
  4882      if (_markStack._failed_double > 0) {
  4883        gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
  4884                               " current capacity "SIZE_FORMAT,
  4885                               _markStack._failed_double,
  4886                               _markStack.capacity());
  4889   _markStack._hit_limit = 0;
  4890   _markStack._failed_double = 0;
  4892   if ((VerifyAfterGC || VerifyDuringGC) &&
  4893       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  4894     verify_after_remark();
  4897   // Change under the freelistLocks.
  4898   _collectorState = Sweeping;
  4899   // Call isAllClear() under bitMapLock
  4900   assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
  4901     " final marking");
  4902   if (UseAdaptiveSizePolicy) {
  4903     size_policy()->checkpoint_roots_final_end(gch->gc_cause());
  4907 // Parallel remark task
  4908 class CMSParRemarkTask: public AbstractGangTask {
  4909   CMSCollector* _collector;
  4910   WorkGang*     _workers;
  4911   int           _n_workers;
  4912   CompactibleFreeListSpace* _cms_space;
  4913   CompactibleFreeListSpace* _perm_space;
  4915   // The per-thread work queues, available here for stealing.
  4916   OopTaskQueueSet*       _task_queues;
  4917   ParallelTaskTerminator _term;
  4919  public:
  4920   CMSParRemarkTask(CMSCollector* collector,
  4921                    CompactibleFreeListSpace* cms_space,
  4922                    CompactibleFreeListSpace* perm_space,
  4923                    int n_workers, WorkGang* workers,
  4924                    OopTaskQueueSet* task_queues):
  4925     AbstractGangTask("Rescan roots and grey objects in parallel"),
  4926     _collector(collector),
  4927     _cms_space(cms_space), _perm_space(perm_space),
  4928     _n_workers(n_workers),
  4929     _workers(workers),
  4930     _task_queues(task_queues),
  4931     _term(workers->total_workers(), task_queues) { }
  4933   OopTaskQueueSet* task_queues() { return _task_queues; }
  4935   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  4937   ParallelTaskTerminator* terminator() { return &_term; }
  4939   void work(int i);
  4941  private:
  4942   // Work method in support of parallel rescan ... of young gen spaces
  4943   void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
  4944                              ContiguousSpace* space,
  4945                              HeapWord** chunk_array, size_t chunk_top);
  4947   // ... of  dirty cards in old space
  4948   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
  4949                                   Par_MarkRefsIntoAndScanClosure* cl);
  4951   // ... work stealing for the above
  4952   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
  4953 };
  4955 void CMSParRemarkTask::work(int i) {
  4956   elapsedTimer _timer;
  4957   ResourceMark rm;
  4958   HandleMark   hm;
  4960   // ---------- rescan from roots --------------
  4961   _timer.start();
  4962   GenCollectedHeap* gch = GenCollectedHeap::heap();
  4963   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
  4964     _collector->_span, _collector->ref_processor(),
  4965     &(_collector->_markBitMap),
  4966     work_queue(i), &(_collector->_revisitStack));
  4968   // Rescan young gen roots first since these are likely
  4969   // coarsely partitioned and may, on that account, constitute
  4970   // the critical path; thus, it's best to start off that
  4971   // work first.
  4972   // ---------- young gen roots --------------
  4974     DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
  4975     EdenSpace* eden_space = dng->eden();
  4976     ContiguousSpace* from_space = dng->from();
  4977     ContiguousSpace* to_space   = dng->to();
  4979     HeapWord** eca = _collector->_eden_chunk_array;
  4980     size_t     ect = _collector->_eden_chunk_index;
  4981     HeapWord** sca = _collector->_survivor_chunk_array;
  4982     size_t     sct = _collector->_survivor_chunk_index;
  4984     assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
  4985     assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
  4987     do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
  4988     do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
  4989     do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
  4991     _timer.stop();
  4992     if (PrintCMSStatistics != 0) {
  4993       gclog_or_tty->print_cr(
  4994         "Finished young gen rescan work in %dth thread: %3.3f sec",
  4995         i, _timer.seconds());
  4999   // ---------- remaining roots --------------
  5000   _timer.reset();
  5001   _timer.start();
  5002   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
  5003                                 false,     // yg was scanned above
  5004                                 true,      // collecting perm gen
  5005                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
  5006                                 NULL, &par_mrias_cl);
  5007   _timer.stop();
  5008   if (PrintCMSStatistics != 0) {
  5009     gclog_or_tty->print_cr(
  5010       "Finished remaining root rescan work in %dth thread: %3.3f sec",
  5011       i, _timer.seconds());
  5014   // ---------- rescan dirty cards ------------
  5015   _timer.reset();
  5016   _timer.start();
  5018   // Do the rescan tasks for each of the two spaces
  5019   // (cms_space and perm_space) in turn.
  5020   do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
  5021   do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
  5022   _timer.stop();
  5023   if (PrintCMSStatistics != 0) {
  5024     gclog_or_tty->print_cr(
  5025       "Finished dirty card rescan work in %dth thread: %3.3f sec",
  5026       i, _timer.seconds());
  5029   // ---------- steal work from other threads ...
  5030   // ---------- ... and drain overflow list.
  5031   _timer.reset();
  5032   _timer.start();
  5033   do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
  5034   _timer.stop();
  5035   if (PrintCMSStatistics != 0) {
  5036     gclog_or_tty->print_cr(
  5037       "Finished work stealing in %dth thread: %3.3f sec",
  5038       i, _timer.seconds());
  5042 void
  5043 CMSParRemarkTask::do_young_space_rescan(int i,
  5044   Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
  5045   HeapWord** chunk_array, size_t chunk_top) {
  5046   // Until all tasks completed:
  5047   // . claim an unclaimed task
  5048   // . compute region boundaries corresponding to task claimed
  5049   //   using chunk_array
  5050   // . par_oop_iterate(cl) over that region
  5052   ResourceMark rm;
  5053   HandleMark   hm;
  5055   SequentialSubTasksDone* pst = space->par_seq_tasks();
  5056   assert(pst->valid(), "Uninitialized use?");
  5058   int nth_task = 0;
  5059   int n_tasks  = pst->n_tasks();
  5061   HeapWord *start, *end;
  5062   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5063     // We claimed task # nth_task; compute its boundaries.
  5064     if (chunk_top == 0) {  // no samples were taken
  5065       assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
  5066       start = space->bottom();
  5067       end   = space->top();
  5068     } else if (nth_task == 0) {
  5069       start = space->bottom();
  5070       end   = chunk_array[nth_task];
  5071     } else if (nth_task < (jint)chunk_top) {
  5072       assert(nth_task >= 1, "Control point invariant");
  5073       start = chunk_array[nth_task - 1];
  5074       end   = chunk_array[nth_task];
  5075     } else {
  5076       assert(nth_task == (jint)chunk_top, "Control point invariant");
  5077       start = chunk_array[chunk_top - 1];
  5078       end   = space->top();
  5080     MemRegion mr(start, end);
  5081     // Verify that mr is in space
  5082     assert(mr.is_empty() || space->used_region().contains(mr),
  5083            "Should be in space");
  5084     // Verify that "start" is an object boundary
  5085     assert(mr.is_empty() || oop(mr.start())->is_oop(),
  5086            "Should be an oop");
  5087     space->par_oop_iterate(mr, cl);
  5089   pst->all_tasks_completed();
  5092 void
  5093 CMSParRemarkTask::do_dirty_card_rescan_tasks(
  5094   CompactibleFreeListSpace* sp, int i,
  5095   Par_MarkRefsIntoAndScanClosure* cl) {
  5096   // Until all tasks completed:
  5097   // . claim an unclaimed task
  5098   // . compute region boundaries corresponding to task claimed
  5099   // . transfer dirty bits ct->mut for that region
  5100   // . apply rescanclosure to dirty mut bits for that region
  5102   ResourceMark rm;
  5103   HandleMark   hm;
  5105   OopTaskQueue* work_q = work_queue(i);
  5106   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
  5107   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
  5108   // CAUTION: This closure has state that persists across calls to
  5109   // the work method dirty_range_iterate_clear() in that it has
  5110   // imbedded in it a (subtype of) UpwardsObjectClosure. The
  5111   // use of that state in the imbedded UpwardsObjectClosure instance
  5112   // assumes that the cards are always iterated (even if in parallel
  5113   // by several threads) in monotonically increasing order per each
  5114   // thread. This is true of the implementation below which picks
  5115   // card ranges (chunks) in monotonically increasing order globally
  5116   // and, a-fortiori, in monotonically increasing order per thread
  5117   // (the latter order being a subsequence of the former).
  5118   // If the work code below is ever reorganized into a more chaotic
  5119   // work-partitioning form than the current "sequential tasks"
  5120   // paradigm, the use of that persistent state will have to be
  5121   // revisited and modified appropriately. See also related
  5122   // bug 4756801 work on which should examine this code to make
  5123   // sure that the changes there do not run counter to the
  5124   // assumptions made here and necessary for correctness and
  5125   // efficiency. Note also that this code might yield inefficient
  5126   // behaviour in the case of very large objects that span one or
  5127   // more work chunks. Such objects would potentially be scanned
  5128   // several times redundantly. Work on 4756801 should try and
  5129   // address that performance anomaly if at all possible. XXX
  5130   MemRegion  full_span  = _collector->_span;
  5131   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
  5132   CMSMarkStack* rs = &(_collector->_revisitStack);   // shared
  5133   MarkFromDirtyCardsClosure
  5134     greyRescanClosure(_collector, full_span, // entire span of interest
  5135                       sp, bm, work_q, rs, cl);
  5137   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  5138   assert(pst->valid(), "Uninitialized use?");
  5139   int nth_task = 0;
  5140   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
  5141   MemRegion span = sp->used_region();
  5142   HeapWord* start_addr = span.start();
  5143   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
  5144                                            alignment);
  5145   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
  5146   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
  5147          start_addr, "Check alignment");
  5148   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
  5149          chunk_size, "Check alignment");
  5151   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5152     // Having claimed the nth_task, compute corresponding mem-region,
  5153     // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
  5154     // The alignment restriction ensures that we do not need any
  5155     // synchronization with other gang-workers while setting or
  5156     // clearing bits in thus chunk of the MUT.
  5157     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
  5158                                     start_addr + (nth_task+1)*chunk_size);
  5159     // The last chunk's end might be way beyond end of the
  5160     // used region. In that case pull back appropriately.
  5161     if (this_span.end() > end_addr) {
  5162       this_span.set_end(end_addr);
  5163       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
  5165     // Iterate over the dirty cards covering this chunk, marking them
  5166     // precleaned, and setting the corresponding bits in the mod union
  5167     // table. Since we have been careful to partition at Card and MUT-word
  5168     // boundaries no synchronization is needed between parallel threads.
  5169     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
  5170                                                  &modUnionClosure);
  5172     // Having transferred these marks into the modUnionTable,
  5173     // rescan the marked objects on the dirty cards in the modUnionTable.
  5174     // Even if this is at a synchronous collection, the initial marking
  5175     // may have been done during an asynchronous collection so there
  5176     // may be dirty bits in the mod-union table.
  5177     _collector->_modUnionTable.dirty_range_iterate_clear(
  5178                   this_span, &greyRescanClosure);
  5179     _collector->_modUnionTable.verifyNoOneBitsInRange(
  5180                                  this_span.start(),
  5181                                  this_span.end());
  5183   pst->all_tasks_completed();  // declare that i am done
  5186 // . see if we can share work_queues with ParNew? XXX
  5187 void
  5188 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
  5189                                 int* seed) {
  5190   OopTaskQueue* work_q = work_queue(i);
  5191   NOT_PRODUCT(int num_steals = 0;)
  5192   oop obj_to_scan;
  5193   CMSBitMap* bm = &(_collector->_markBitMap);
  5194   size_t num_from_overflow_list =
  5195            MIN2((size_t)work_q->max_elems()/4,
  5196                 (size_t)ParGCDesiredObjsFromOverflowList);
  5198   while (true) {
  5199     // Completely finish any left over work from (an) earlier round(s)
  5200     cl->trim_queue(0);
  5201     // Now check if there's any work in the overflow list
  5202     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  5203                                                 work_q)) {
  5204       // found something in global overflow list;
  5205       // not yet ready to go stealing work from others.
  5206       // We'd like to assert(work_q->size() != 0, ...)
  5207       // because we just took work from the overflow list,
  5208       // but of course we can't since all of that could have
  5209       // been already stolen from us.
  5210       // "He giveth and He taketh away."
  5211       continue;
  5213     // Verify that we have no work before we resort to stealing
  5214     assert(work_q->size() == 0, "Have work, shouldn't steal");
  5215     // Try to steal from other queues that have work
  5216     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  5217       NOT_PRODUCT(num_steals++;)
  5218       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  5219       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  5220       // Do scanning work
  5221       obj_to_scan->oop_iterate(cl);
  5222       // Loop around, finish this work, and try to steal some more
  5223     } else if (terminator()->offer_termination()) {
  5224         break;  // nirvana from the infinite cycle
  5227   NOT_PRODUCT(
  5228     if (PrintCMSStatistics != 0) {
  5229       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  5232   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
  5233          "Else our work is not yet done");
  5236 // Return a thread-local PLAB recording array, as appropriate.
  5237 void* CMSCollector::get_data_recorder(int thr_num) {
  5238   if (_survivor_plab_array != NULL &&
  5239       (CMSPLABRecordAlways ||
  5240        (_collectorState > Marking && _collectorState < FinalMarking))) {
  5241     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
  5242     ChunkArray* ca = &_survivor_plab_array[thr_num];
  5243     ca->reset();   // clear it so that fresh data is recorded
  5244     return (void*) ca;
  5245   } else {
  5246     return NULL;
  5250 // Reset all the thread-local PLAB recording arrays
  5251 void CMSCollector::reset_survivor_plab_arrays() {
  5252   for (uint i = 0; i < ParallelGCThreads; i++) {
  5253     _survivor_plab_array[i].reset();
  5257 // Merge the per-thread plab arrays into the global survivor chunk
  5258 // array which will provide the partitioning of the survivor space
  5259 // for CMS rescan.
  5260 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
  5261   assert(_survivor_plab_array  != NULL, "Error");
  5262   assert(_survivor_chunk_array != NULL, "Error");
  5263   assert(_collectorState == FinalMarking, "Error");
  5264   for (uint j = 0; j < ParallelGCThreads; j++) {
  5265     _cursor[j] = 0;
  5267   HeapWord* top = surv->top();
  5268   size_t i;
  5269   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
  5270     HeapWord* min_val = top;          // Higher than any PLAB address
  5271     uint      min_tid = 0;            // position of min_val this round
  5272     for (uint j = 0; j < ParallelGCThreads; j++) {
  5273       ChunkArray* cur_sca = &_survivor_plab_array[j];
  5274       if (_cursor[j] == cur_sca->end()) {
  5275         continue;
  5277       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
  5278       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
  5279       assert(surv->used_region().contains(cur_val), "Out of bounds value");
  5280       if (cur_val < min_val) {
  5281         min_tid = j;
  5282         min_val = cur_val;
  5283       } else {
  5284         assert(cur_val < top, "All recorded addresses should be less");
  5287     // At this point min_val and min_tid are respectively
  5288     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
  5289     // and the thread (j) that witnesses that address.
  5290     // We record this address in the _survivor_chunk_array[i]
  5291     // and increment _cursor[min_tid] prior to the next round i.
  5292     if (min_val == top) {
  5293       break;
  5295     _survivor_chunk_array[i] = min_val;
  5296     _cursor[min_tid]++;
  5298   // We are all done; record the size of the _survivor_chunk_array
  5299   _survivor_chunk_index = i; // exclusive: [0, i)
  5300   if (PrintCMSStatistics > 0) {
  5301     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
  5303   // Verify that we used up all the recorded entries
  5304   #ifdef ASSERT
  5305     size_t total = 0;
  5306     for (uint j = 0; j < ParallelGCThreads; j++) {
  5307       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
  5308       total += _cursor[j];
  5310     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
  5311     // Check that the merged array is in sorted order
  5312     if (total > 0) {
  5313       for (size_t i = 0; i < total - 1; i++) {
  5314         if (PrintCMSStatistics > 0) {
  5315           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
  5316                               i, _survivor_chunk_array[i]);
  5318         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
  5319                "Not sorted");
  5322   #endif // ASSERT
  5325 // Set up the space's par_seq_tasks structure for work claiming
  5326 // for parallel rescan of young gen.
  5327 // See ParRescanTask where this is currently used.
  5328 void
  5329 CMSCollector::
  5330 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
  5331   assert(n_threads > 0, "Unexpected n_threads argument");
  5332   DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  5334   // Eden space
  5336     SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
  5337     assert(!pst->valid(), "Clobbering existing data?");
  5338     // Each valid entry in [0, _eden_chunk_index) represents a task.
  5339     size_t n_tasks = _eden_chunk_index + 1;
  5340     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
  5341     pst->set_par_threads(n_threads);
  5342     pst->set_n_tasks((int)n_tasks);
  5345   // Merge the survivor plab arrays into _survivor_chunk_array
  5346   if (_survivor_plab_array != NULL) {
  5347     merge_survivor_plab_arrays(dng->from());
  5348   } else {
  5349     assert(_survivor_chunk_index == 0, "Error");
  5352   // To space
  5354     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
  5355     assert(!pst->valid(), "Clobbering existing data?");
  5356     pst->set_par_threads(n_threads);
  5357     pst->set_n_tasks(1);
  5358     assert(pst->valid(), "Error");
  5361   // From space
  5363     SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
  5364     assert(!pst->valid(), "Clobbering existing data?");
  5365     size_t n_tasks = _survivor_chunk_index + 1;
  5366     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
  5367     pst->set_par_threads(n_threads);
  5368     pst->set_n_tasks((int)n_tasks);
  5369     assert(pst->valid(), "Error");
  5373 // Parallel version of remark
  5374 void CMSCollector::do_remark_parallel() {
  5375   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5376   WorkGang* workers = gch->workers();
  5377   assert(workers != NULL, "Need parallel worker threads.");
  5378   int n_workers = workers->total_workers();
  5379   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  5380   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
  5382   CMSParRemarkTask tsk(this,
  5383     cms_space, perm_space,
  5384     n_workers, workers, task_queues());
  5386   // Set up for parallel process_strong_roots work.
  5387   gch->set_par_threads(n_workers);
  5388   gch->change_strong_roots_parity();
  5389   // We won't be iterating over the cards in the card table updating
  5390   // the younger_gen cards, so we shouldn't call the following else
  5391   // the verification code as well as subsequent younger_refs_iterate
  5392   // code would get confused. XXX
  5393   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
  5395   // The young gen rescan work will not be done as part of
  5396   // process_strong_roots (which currently doesn't knw how to
  5397   // parallelize such a scan), but rather will be broken up into
  5398   // a set of parallel tasks (via the sampling that the [abortable]
  5399   // preclean phase did of EdenSpace, plus the [two] tasks of
  5400   // scanning the [two] survivor spaces. Further fine-grain
  5401   // parallelization of the scanning of the survivor spaces
  5402   // themselves, and of precleaning of the younger gen itself
  5403   // is deferred to the future.
  5404   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
  5406   // The dirty card rescan work is broken up into a "sequence"
  5407   // of parallel tasks (per constituent space) that are dynamically
  5408   // claimed by the parallel threads.
  5409   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
  5410   perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
  5412   // It turns out that even when we're using 1 thread, doing the work in a
  5413   // separate thread causes wide variance in run times.  We can't help this
  5414   // in the multi-threaded case, but we special-case n=1 here to get
  5415   // repeatable measurements of the 1-thread overhead of the parallel code.
  5416   if (n_workers > 1) {
  5417     // Make refs discovery MT-safe
  5418     ReferenceProcessorMTMutator mt(ref_processor(), true);
  5419     workers->run_task(&tsk);
  5420   } else {
  5421     tsk.work(0);
  5423   gch->set_par_threads(0);  // 0 ==> non-parallel.
  5424   // restore, single-threaded for now, any preserved marks
  5425   // as a result of work_q overflow
  5426   restore_preserved_marks_if_any();
  5429 // Non-parallel version of remark
  5430 void CMSCollector::do_remark_non_parallel() {
  5431   ResourceMark rm;
  5432   HandleMark   hm;
  5433   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5434   MarkRefsIntoAndScanClosure
  5435     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
  5436              &_markStack, &_revisitStack, this,
  5437              false /* should_yield */, false /* not precleaning */);
  5438   MarkFromDirtyCardsClosure
  5439     markFromDirtyCardsClosure(this, _span,
  5440                               NULL,  // space is set further below
  5441                               &_markBitMap, &_markStack, &_revisitStack,
  5442                               &mrias_cl);
  5444     TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
  5445     // Iterate over the dirty cards, setting the corresponding bits in the
  5446     // mod union table.
  5448       ModUnionClosure modUnionClosure(&_modUnionTable);
  5449       _ct->ct_bs()->dirty_card_iterate(
  5450                       _cmsGen->used_region(),
  5451                       &modUnionClosure);
  5452       _ct->ct_bs()->dirty_card_iterate(
  5453                       _permGen->used_region(),
  5454                       &modUnionClosure);
  5456     // Having transferred these marks into the modUnionTable, we just need
  5457     // to rescan the marked objects on the dirty cards in the modUnionTable.
  5458     // The initial marking may have been done during an asynchronous
  5459     // collection so there may be dirty bits in the mod-union table.
  5460     const int alignment =
  5461       CardTableModRefBS::card_size * BitsPerWord;
  5463       // ... First handle dirty cards in CMS gen
  5464       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
  5465       MemRegion ur = _cmsGen->used_region();
  5466       HeapWord* lb = ur.start();
  5467       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
  5468       MemRegion cms_span(lb, ub);
  5469       _modUnionTable.dirty_range_iterate_clear(cms_span,
  5470                                                &markFromDirtyCardsClosure);
  5471       verify_work_stacks_empty();
  5472       if (PrintCMSStatistics != 0) {
  5473         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
  5474           markFromDirtyCardsClosure.num_dirty_cards());
  5478       // .. and then repeat for dirty cards in perm gen
  5479       markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
  5480       MemRegion ur = _permGen->used_region();
  5481       HeapWord* lb = ur.start();
  5482       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
  5483       MemRegion perm_span(lb, ub);
  5484       _modUnionTable.dirty_range_iterate_clear(perm_span,
  5485                                                &markFromDirtyCardsClosure);
  5486       verify_work_stacks_empty();
  5487       if (PrintCMSStatistics != 0) {
  5488         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
  5489           markFromDirtyCardsClosure.num_dirty_cards());
  5493   if (VerifyDuringGC &&
  5494       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  5495     HandleMark hm;  // Discard invalid handles created during verification
  5496     Universe::verify(true);
  5499     TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
  5501     verify_work_stacks_empty();
  5503     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  5504     gch->gen_process_strong_roots(_cmsGen->level(),
  5505                                   true,  // younger gens as roots
  5506                                   true,  // collecting perm gen
  5507                                   SharedHeap::ScanningOption(roots_scanning_options()),
  5508                                   NULL, &mrias_cl);
  5510   verify_work_stacks_empty();
  5511   // Restore evacuated mark words, if any, used for overflow list links
  5512   if (!CMSOverflowEarlyRestoration) {
  5513     restore_preserved_marks_if_any();
  5515   verify_overflow_empty();
  5518 ////////////////////////////////////////////////////////
  5519 // Parallel Reference Processing Task Proxy Class
  5520 ////////////////////////////////////////////////////////
  5521 class CMSRefProcTaskProxy: public AbstractGangTask {
  5522   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
  5523   CMSCollector*          _collector;
  5524   CMSBitMap*             _mark_bit_map;
  5525   const MemRegion        _span;
  5526   OopTaskQueueSet*       _task_queues;
  5527   ParallelTaskTerminator _term;
  5528   ProcessTask&           _task;
  5530 public:
  5531   CMSRefProcTaskProxy(ProcessTask&     task,
  5532                       CMSCollector*    collector,
  5533                       const MemRegion& span,
  5534                       CMSBitMap*       mark_bit_map,
  5535                       int              total_workers,
  5536                       OopTaskQueueSet* task_queues):
  5537     AbstractGangTask("Process referents by policy in parallel"),
  5538     _task(task),
  5539     _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
  5540     _task_queues(task_queues),
  5541     _term(total_workers, task_queues)
  5543       assert(_collector->_span.equals(_span) && !_span.is_empty(),
  5544              "Inconsistency in _span");
  5547   OopTaskQueueSet* task_queues() { return _task_queues; }
  5549   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  5551   ParallelTaskTerminator* terminator() { return &_term; }
  5553   void do_work_steal(int i,
  5554                      CMSParDrainMarkingStackClosure* drain,
  5555                      CMSParKeepAliveClosure* keep_alive,
  5556                      int* seed);
  5558   virtual void work(int i);
  5559 };
  5561 void CMSRefProcTaskProxy::work(int i) {
  5562   assert(_collector->_span.equals(_span), "Inconsistency in _span");
  5563   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
  5564                                         _mark_bit_map, work_queue(i));
  5565   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
  5566                                                  _mark_bit_map, work_queue(i));
  5567   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
  5568   _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
  5569   if (_task.marks_oops_alive()) {
  5570     do_work_steal(i, &par_drain_stack, &par_keep_alive,
  5571                   _collector->hash_seed(i));
  5573   assert(work_queue(i)->size() == 0, "work_queue should be empty");
  5574   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
  5577 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
  5578   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
  5579   EnqueueTask& _task;
  5581 public:
  5582   CMSRefEnqueueTaskProxy(EnqueueTask& task)
  5583     : AbstractGangTask("Enqueue reference objects in parallel"),
  5584       _task(task)
  5585   { }
  5587   virtual void work(int i)
  5589     _task.work(i);
  5591 };
  5593 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
  5594   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
  5595    _collector(collector),
  5596    _span(span),
  5597    _bit_map(bit_map),
  5598    _work_queue(work_queue),
  5599    _mark_and_push(collector, span, bit_map, work_queue),
  5600    _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  5601                         (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
  5602 { }
  5604 // . see if we can share work_queues with ParNew? XXX
  5605 void CMSRefProcTaskProxy::do_work_steal(int i,
  5606   CMSParDrainMarkingStackClosure* drain,
  5607   CMSParKeepAliveClosure* keep_alive,
  5608   int* seed) {
  5609   OopTaskQueue* work_q = work_queue(i);
  5610   NOT_PRODUCT(int num_steals = 0;)
  5611   oop obj_to_scan;
  5612   size_t num_from_overflow_list =
  5613            MIN2((size_t)work_q->max_elems()/4,
  5614                 (size_t)ParGCDesiredObjsFromOverflowList);
  5616   while (true) {
  5617     // Completely finish any left over work from (an) earlier round(s)
  5618     drain->trim_queue(0);
  5619     // Now check if there's any work in the overflow list
  5620     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  5621                                                 work_q)) {
  5622       // Found something in global overflow list;
  5623       // not yet ready to go stealing work from others.
  5624       // We'd like to assert(work_q->size() != 0, ...)
  5625       // because we just took work from the overflow list,
  5626       // but of course we can't, since all of that might have
  5627       // been already stolen from us.
  5628       continue;
  5630     // Verify that we have no work before we resort to stealing
  5631     assert(work_q->size() == 0, "Have work, shouldn't steal");
  5632     // Try to steal from other queues that have work
  5633     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  5634       NOT_PRODUCT(num_steals++;)
  5635       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  5636       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  5637       // Do scanning work
  5638       obj_to_scan->oop_iterate(keep_alive);
  5639       // Loop around, finish this work, and try to steal some more
  5640     } else if (terminator()->offer_termination()) {
  5641       break;  // nirvana from the infinite cycle
  5644   NOT_PRODUCT(
  5645     if (PrintCMSStatistics != 0) {
  5646       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  5651 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
  5653   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5654   WorkGang* workers = gch->workers();
  5655   assert(workers != NULL, "Need parallel worker threads.");
  5656   int n_workers = workers->total_workers();
  5657   CMSRefProcTaskProxy rp_task(task, &_collector,
  5658                               _collector.ref_processor()->span(),
  5659                               _collector.markBitMap(),
  5660                               n_workers, _collector.task_queues());
  5661   workers->run_task(&rp_task);
  5664 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
  5667   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5668   WorkGang* workers = gch->workers();
  5669   assert(workers != NULL, "Need parallel worker threads.");
  5670   CMSRefEnqueueTaskProxy enq_task(task);
  5671   workers->run_task(&enq_task);
  5674 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
  5676   ResourceMark rm;
  5677   HandleMark   hm;
  5678   ReferencePolicy* soft_ref_policy;
  5680   assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
  5681   // Process weak references.
  5682   if (clear_all_soft_refs) {
  5683     soft_ref_policy = new AlwaysClearPolicy();
  5684   } else {
  5685 #ifdef COMPILER2
  5686     soft_ref_policy = new LRUMaxHeapPolicy();
  5687 #else
  5688     soft_ref_policy = new LRUCurrentHeapPolicy();
  5689 #endif // COMPILER2
  5691   verify_work_stacks_empty();
  5693   ReferenceProcessor* rp = ref_processor();
  5694   assert(rp->span().equals(_span), "Spans should be equal");
  5695   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
  5696                                           &_markStack);
  5697   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
  5698                                 _span, &_markBitMap, &_markStack,
  5699                                 &cmsKeepAliveClosure);
  5701     TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
  5702     if (rp->processing_is_mt()) {
  5703       CMSRefProcTaskExecutor task_executor(*this);
  5704       rp->process_discovered_references(soft_ref_policy,
  5705                                         &_is_alive_closure,
  5706                                         &cmsKeepAliveClosure,
  5707                                         &cmsDrainMarkingStackClosure,
  5708                                         &task_executor);
  5709     } else {
  5710       rp->process_discovered_references(soft_ref_policy,
  5711                                         &_is_alive_closure,
  5712                                         &cmsKeepAliveClosure,
  5713                                         &cmsDrainMarkingStackClosure,
  5714                                         NULL);
  5716     verify_work_stacks_empty();
  5719   if (should_unload_classes()) {
  5721       TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
  5723       // Follow SystemDictionary roots and unload classes
  5724       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
  5726       // Follow CodeCache roots and unload any methods marked for unloading
  5727       CodeCache::do_unloading(&_is_alive_closure,
  5728                               &cmsKeepAliveClosure,
  5729                               purged_class);
  5731       cmsDrainMarkingStackClosure.do_void();
  5732       verify_work_stacks_empty();
  5734       // Update subklass/sibling/implementor links in KlassKlass descendants
  5735       assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
  5736       oop k;
  5737       while ((k = _revisitStack.pop()) != NULL) {
  5738         ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
  5739                        &_is_alive_closure,
  5740                        &cmsKeepAliveClosure);
  5742       assert(!ClassUnloading ||
  5743              (_markStack.isEmpty() && overflow_list_is_empty()),
  5744              "Should not have found new reachable objects");
  5745       assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
  5746       cmsDrainMarkingStackClosure.do_void();
  5747       verify_work_stacks_empty();
  5751       TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
  5752       // Now clean up stale oops in SymbolTable and StringTable
  5753       SymbolTable::unlink(&_is_alive_closure);
  5754       StringTable::unlink(&_is_alive_closure);
  5758   verify_work_stacks_empty();
  5759   // Restore any preserved marks as a result of mark stack or
  5760   // work queue overflow
  5761   restore_preserved_marks_if_any();  // done single-threaded for now
  5763   rp->set_enqueuing_is_done(true);
  5764   if (rp->processing_is_mt()) {
  5765     CMSRefProcTaskExecutor task_executor(*this);
  5766     rp->enqueue_discovered_references(&task_executor);
  5767   } else {
  5768     rp->enqueue_discovered_references(NULL);
  5770   rp->verify_no_references_recorded();
  5771   assert(!rp->discovery_enabled(), "should have been disabled");
  5773   // JVMTI object tagging is based on JNI weak refs. If any of these
  5774   // refs were cleared then JVMTI needs to update its maps and
  5775   // maybe post ObjectFrees to agents.
  5776   JvmtiExport::cms_ref_processing_epilogue();
  5779 #ifndef PRODUCT
  5780 void CMSCollector::check_correct_thread_executing() {
  5781   Thread* t = Thread::current();
  5782   // Only the VM thread or the CMS thread should be here.
  5783   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
  5784          "Unexpected thread type");
  5785   // If this is the vm thread, the foreground process
  5786   // should not be waiting.  Note that _foregroundGCIsActive is
  5787   // true while the foreground collector is waiting.
  5788   if (_foregroundGCShouldWait) {
  5789     // We cannot be the VM thread
  5790     assert(t->is_ConcurrentGC_thread(),
  5791            "Should be CMS thread");
  5792   } else {
  5793     // We can be the CMS thread only if we are in a stop-world
  5794     // phase of CMS collection.
  5795     if (t->is_ConcurrentGC_thread()) {
  5796       assert(_collectorState == InitialMarking ||
  5797              _collectorState == FinalMarking,
  5798              "Should be a stop-world phase");
  5799       // The CMS thread should be holding the CMS_token.
  5800       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  5801              "Potential interference with concurrently "
  5802              "executing VM thread");
  5806 #endif
  5808 void CMSCollector::sweep(bool asynch) {
  5809   assert(_collectorState == Sweeping, "just checking");
  5810   check_correct_thread_executing();
  5811   verify_work_stacks_empty();
  5812   verify_overflow_empty();
  5813   incrementSweepCount();
  5814   _sweep_timer.stop();
  5815   _sweep_estimate.sample(_sweep_timer.seconds());
  5816   size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
  5818   // PermGen verification support: If perm gen sweeping is disabled in
  5819   // this cycle, we preserve the perm gen object "deadness" information
  5820   // in the perm_gen_verify_bit_map. In order to do that we traverse
  5821   // all blocks in perm gen and mark all dead objects.
  5822   if (verifying() && !should_unload_classes()) {
  5823     assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
  5824            "Should have already been allocated");
  5825     MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
  5826                                markBitMap(), perm_gen_verify_bit_map());
  5827     if (asynch) {
  5828       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
  5829                                bitMapLock());
  5830       _permGen->cmsSpace()->blk_iterate(&mdo);
  5831     } else {
  5832       // In the case of synchronous sweep, we already have
  5833       // the requisite locks/tokens.
  5834       _permGen->cmsSpace()->blk_iterate(&mdo);
  5838   if (asynch) {
  5839     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  5840     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
  5841     // First sweep the old gen then the perm gen
  5843       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
  5844                                bitMapLock());
  5845       sweepWork(_cmsGen, asynch);
  5848     // Now repeat for perm gen
  5849     if (should_unload_classes()) {
  5850       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
  5851                              bitMapLock());
  5852       sweepWork(_permGen, asynch);
  5855     // Update Universe::_heap_*_at_gc figures.
  5856     // We need all the free list locks to make the abstract state
  5857     // transition from Sweeping to Resetting. See detailed note
  5858     // further below.
  5860       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
  5861                                _permGen->freelistLock());
  5862       // Update heap occupancy information which is used as
  5863       // input to soft ref clearing policy at the next gc.
  5864       Universe::update_heap_info_at_gc();
  5865       _collectorState = Resizing;
  5867   } else {
  5868     // already have needed locks
  5869     sweepWork(_cmsGen,  asynch);
  5871     if (should_unload_classes()) {
  5872       sweepWork(_permGen, asynch);
  5874     // Update heap occupancy information which is used as
  5875     // input to soft ref clearing policy at the next gc.
  5876     Universe::update_heap_info_at_gc();
  5877     _collectorState = Resizing;
  5879   verify_work_stacks_empty();
  5880   verify_overflow_empty();
  5882   _sweep_timer.reset();
  5883   _sweep_timer.start();
  5885   update_time_of_last_gc(os::javaTimeMillis());
  5887   // NOTE on abstract state transitions:
  5888   // Mutators allocate-live and/or mark the mod-union table dirty
  5889   // based on the state of the collection.  The former is done in
  5890   // the interval [Marking, Sweeping] and the latter in the interval
  5891   // [Marking, Sweeping).  Thus the transitions into the Marking state
  5892   // and out of the Sweeping state must be synchronously visible
  5893   // globally to the mutators.
  5894   // The transition into the Marking state happens with the world
  5895   // stopped so the mutators will globally see it.  Sweeping is
  5896   // done asynchronously by the background collector so the transition
  5897   // from the Sweeping state to the Resizing state must be done
  5898   // under the freelistLock (as is the check for whether to
  5899   // allocate-live and whether to dirty the mod-union table).
  5900   assert(_collectorState == Resizing, "Change of collector state to"
  5901     " Resizing must be done under the freelistLocks (plural)");
  5903   // Now that sweeping has been completed, if the GCH's
  5904   // incremental_collection_will_fail flag is set, clear it,
  5905   // thus inviting a younger gen collection to promote into
  5906   // this generation. If such a promotion may still fail,
  5907   // the flag will be set again when a young collection is
  5908   // attempted.
  5909   // I think the incremental_collection_will_fail flag's use
  5910   // is specific to a 2 generation collection policy, so i'll
  5911   // assert that that's the configuration we are operating within.
  5912   // The use of the flag can and should be generalized appropriately
  5913   // in the future to deal with a general n-generation system.
  5915   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5916   assert(gch->collector_policy()->is_two_generation_policy(),
  5917          "Resetting of incremental_collection_will_fail flag"
  5918          " may be incorrect otherwise");
  5919   gch->clear_incremental_collection_will_fail();
  5920   gch->update_full_collections_completed(_collection_count_start);
  5923 // FIX ME!!! Looks like this belongs in CFLSpace, with
  5924 // CMSGen merely delegating to it.
  5925 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
  5926   double nearLargestPercent = 0.999;
  5927   HeapWord*  minAddr        = _cmsSpace->bottom();
  5928   HeapWord*  largestAddr    =
  5929     (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
  5930   if (largestAddr == 0) {
  5931     // The dictionary appears to be empty.  In this case
  5932     // try to coalesce at the end of the heap.
  5933     largestAddr = _cmsSpace->end();
  5935   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
  5936   size_t nearLargestOffset =
  5937     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
  5938   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
  5941 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
  5942   return addr >= _cmsSpace->nearLargestChunk();
  5945 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
  5946   return _cmsSpace->find_chunk_at_end();
  5949 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
  5950                                                     bool full) {
  5951   // The next lower level has been collected.  Gather any statistics
  5952   // that are of interest at this point.
  5953   if (!full && (current_level + 1) == level()) {
  5954     // Gather statistics on the young generation collection.
  5955     collector()->stats().record_gc0_end(used());
  5959 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
  5960   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5961   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
  5962     "Wrong type of heap");
  5963   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
  5964     gch->gen_policy()->size_policy();
  5965   assert(sp->is_gc_cms_adaptive_size_policy(),
  5966     "Wrong type of size policy");
  5967   return sp;
  5970 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
  5971   if (PrintGCDetails && Verbose) {
  5972     gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
  5974   _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
  5975   _debug_collection_type =
  5976     (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
  5977   if (PrintGCDetails && Verbose) {
  5978     gclog_or_tty->print_cr("to %d ", _debug_collection_type);
  5982 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
  5983   bool asynch) {
  5984   // We iterate over the space(s) underlying this generation,
  5985   // checking the mark bit map to see if the bits corresponding
  5986   // to specific blocks are marked or not. Blocks that are
  5987   // marked are live and are not swept up. All remaining blocks
  5988   // are swept up, with coalescing on-the-fly as we sweep up
  5989   // contiguous free and/or garbage blocks:
  5990   // We need to ensure that the sweeper synchronizes with allocators
  5991   // and stop-the-world collectors. In particular, the following
  5992   // locks are used:
  5993   // . CMS token: if this is held, a stop the world collection cannot occur
  5994   // . freelistLock: if this is held no allocation can occur from this
  5995   //                 generation by another thread
  5996   // . bitMapLock: if this is held, no other thread can access or update
  5997   //
  5999   // Note that we need to hold the freelistLock if we use
  6000   // block iterate below; else the iterator might go awry if
  6001   // a mutator (or promotion) causes block contents to change
  6002   // (for instance if the allocator divvies up a block).
  6003   // If we hold the free list lock, for all practical purposes
  6004   // young generation GC's can't occur (they'll usually need to
  6005   // promote), so we might as well prevent all young generation
  6006   // GC's while we do a sweeping step. For the same reason, we might
  6007   // as well take the bit map lock for the entire duration
  6009   // check that we hold the requisite locks
  6010   assert(have_cms_token(), "Should hold cms token");
  6011   assert(   (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
  6012          || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
  6013         "Should possess CMS token to sweep");
  6014   assert_lock_strong(gen->freelistLock());
  6015   assert_lock_strong(bitMapLock());
  6017   assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
  6018   gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
  6019                                       _sweep_estimate.padded_average());
  6020   gen->setNearLargestChunk();
  6023     SweepClosure sweepClosure(this, gen, &_markBitMap,
  6024                             CMSYield && asynch);
  6025     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
  6026     // We need to free-up/coalesce garbage/blocks from a
  6027     // co-terminal free run. This is done in the SweepClosure
  6028     // destructor; so, do not remove this scope, else the
  6029     // end-of-sweep-census below will be off by a little bit.
  6031   gen->cmsSpace()->sweep_completed();
  6032   gen->cmsSpace()->endSweepFLCensus(sweepCount());
  6033   if (should_unload_classes()) {                // unloaded classes this cycle,
  6034     _concurrent_cycles_since_last_unload = 0;   // ... reset count
  6035   } else {                                      // did not unload classes,
  6036     _concurrent_cycles_since_last_unload++;     // ... increment count
  6040 // Reset CMS data structures (for now just the marking bit map)
  6041 // preparatory for the next cycle.
  6042 void CMSCollector::reset(bool asynch) {
  6043   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6044   CMSAdaptiveSizePolicy* sp = size_policy();
  6045   AdaptiveSizePolicyOutput(sp, gch->total_collections());
  6046   if (asynch) {
  6047     CMSTokenSyncWithLocks ts(true, bitMapLock());
  6049     // If the state is not "Resetting", the foreground  thread
  6050     // has done a collection and the resetting.
  6051     if (_collectorState != Resetting) {
  6052       assert(_collectorState == Idling, "The state should only change"
  6053         " because the foreground collector has finished the collection");
  6054       return;
  6057     // Clear the mark bitmap (no grey objects to start with)
  6058     // for the next cycle.
  6059     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6060     CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
  6062     HeapWord* curAddr = _markBitMap.startWord();
  6063     while (curAddr < _markBitMap.endWord()) {
  6064       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
  6065       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
  6066       _markBitMap.clear_large_range(chunk);
  6067       if (ConcurrentMarkSweepThread::should_yield() &&
  6068           !foregroundGCIsActive() &&
  6069           CMSYield) {
  6070         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6071                "CMS thread should hold CMS token");
  6072         assert_lock_strong(bitMapLock());
  6073         bitMapLock()->unlock();
  6074         ConcurrentMarkSweepThread::desynchronize(true);
  6075         ConcurrentMarkSweepThread::acknowledge_yield_request();
  6076         stopTimer();
  6077         if (PrintCMSStatistics != 0) {
  6078           incrementYields();
  6080         icms_wait();
  6082         // See the comment in coordinator_yield()
  6083         for (unsigned i = 0; i < CMSYieldSleepCount &&
  6084                          ConcurrentMarkSweepThread::should_yield() &&
  6085                          !CMSCollector::foregroundGCIsActive(); ++i) {
  6086           os::sleep(Thread::current(), 1, false);
  6087           ConcurrentMarkSweepThread::acknowledge_yield_request();
  6090         ConcurrentMarkSweepThread::synchronize(true);
  6091         bitMapLock()->lock_without_safepoint_check();
  6092         startTimer();
  6094       curAddr = chunk.end();
  6096     _collectorState = Idling;
  6097   } else {
  6098     // already have the lock
  6099     assert(_collectorState == Resetting, "just checking");
  6100     assert_lock_strong(bitMapLock());
  6101     _markBitMap.clear_all();
  6102     _collectorState = Idling;
  6105   // Stop incremental mode after a cycle completes, so that any future cycles
  6106   // are triggered by allocation.
  6107   stop_icms();
  6109   NOT_PRODUCT(
  6110     if (RotateCMSCollectionTypes) {
  6111       _cmsGen->rotate_debug_collection_type();
  6116 void CMSCollector::do_CMS_operation(CMS_op_type op) {
  6117   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
  6118   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6119   TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
  6120   TraceCollectorStats tcs(counters());
  6122   switch (op) {
  6123     case CMS_op_checkpointRootsInitial: {
  6124       checkpointRootsInitial(true);       // asynch
  6125       if (PrintGC) {
  6126         _cmsGen->printOccupancy("initial-mark");
  6128       break;
  6130     case CMS_op_checkpointRootsFinal: {
  6131       checkpointRootsFinal(true,    // asynch
  6132                            false,   // !clear_all_soft_refs
  6133                            false);  // !init_mark_was_synchronous
  6134       if (PrintGC) {
  6135         _cmsGen->printOccupancy("remark");
  6137       break;
  6139     default:
  6140       fatal("No such CMS_op");
  6144 #ifndef PRODUCT
  6145 size_t const CMSCollector::skip_header_HeapWords() {
  6146   return FreeChunk::header_size();
  6149 // Try and collect here conditions that should hold when
  6150 // CMS thread is exiting. The idea is that the foreground GC
  6151 // thread should not be blocked if it wants to terminate
  6152 // the CMS thread and yet continue to run the VM for a while
  6153 // after that.
  6154 void CMSCollector::verify_ok_to_terminate() const {
  6155   assert(Thread::current()->is_ConcurrentGC_thread(),
  6156          "should be called by CMS thread");
  6157   assert(!_foregroundGCShouldWait, "should be false");
  6158   // We could check here that all the various low-level locks
  6159   // are not held by the CMS thread, but that is overkill; see
  6160   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
  6161   // is checked.
  6163 #endif
  6165 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
  6166   assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
  6167          "missing Printezis mark?");
  6168   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6169   size_t size = pointer_delta(nextOneAddr + 1, addr);
  6170   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6171          "alignment problem");
  6172   assert(size >= 3, "Necessary for Printezis marks to work");
  6173   return size;
  6176 // A variant of the above (block_size_using_printezis_bits()) except
  6177 // that we return 0 if the P-bits are not yet set.
  6178 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
  6179   if (_markBitMap.isMarked(addr)) {
  6180     assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
  6181     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6182     size_t size = pointer_delta(nextOneAddr + 1, addr);
  6183     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6184            "alignment problem");
  6185     assert(size >= 3, "Necessary for Printezis marks to work");
  6186     return size;
  6187   } else {
  6188     assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
  6189     return 0;
  6193 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
  6194   size_t sz = 0;
  6195   oop p = (oop)addr;
  6196   if (p->klass_or_null() != NULL && p->is_parsable()) {
  6197     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
  6198   } else {
  6199     sz = block_size_using_printezis_bits(addr);
  6201   assert(sz > 0, "size must be nonzero");
  6202   HeapWord* next_block = addr + sz;
  6203   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
  6204                                              CardTableModRefBS::card_size);
  6205   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
  6206          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
  6207          "must be different cards");
  6208   return next_card;
  6212 // CMS Bit Map Wrapper /////////////////////////////////////////
  6214 // Construct a CMS bit map infrastructure, but don't create the
  6215 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
  6216 // further below.
  6217 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
  6218   _bm(),
  6219   _shifter(shifter),
  6220   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
  6222   _bmStartWord = 0;
  6223   _bmWordSize  = 0;
  6226 bool CMSBitMap::allocate(MemRegion mr) {
  6227   _bmStartWord = mr.start();
  6228   _bmWordSize  = mr.word_size();
  6229   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
  6230                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
  6231   if (!brs.is_reserved()) {
  6232     warning("CMS bit map allocation failure");
  6233     return false;
  6235   // For now we'll just commit all of the bit map up fromt.
  6236   // Later on we'll try to be more parsimonious with swap.
  6237   if (!_virtual_space.initialize(brs, brs.size())) {
  6238     warning("CMS bit map backing store failure");
  6239     return false;
  6241   assert(_virtual_space.committed_size() == brs.size(),
  6242          "didn't reserve backing store for all of CMS bit map?");
  6243   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
  6244   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
  6245          _bmWordSize, "inconsistency in bit map sizing");
  6246   _bm.set_size(_bmWordSize >> _shifter);
  6248   // bm.clear(); // can we rely on getting zero'd memory? verify below
  6249   assert(isAllClear(),
  6250          "Expected zero'd memory from ReservedSpace constructor");
  6251   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
  6252          "consistency check");
  6253   return true;
  6256 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
  6257   HeapWord *next_addr, *end_addr, *last_addr;
  6258   assert_locked();
  6259   assert(covers(mr), "out-of-range error");
  6260   // XXX assert that start and end are appropriately aligned
  6261   for (next_addr = mr.start(), end_addr = mr.end();
  6262        next_addr < end_addr; next_addr = last_addr) {
  6263     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
  6264     last_addr = dirty_region.end();
  6265     if (!dirty_region.is_empty()) {
  6266       cl->do_MemRegion(dirty_region);
  6267     } else {
  6268       assert(last_addr == end_addr, "program logic");
  6269       return;
  6274 #ifndef PRODUCT
  6275 void CMSBitMap::assert_locked() const {
  6276   CMSLockVerifier::assert_locked(lock());
  6279 bool CMSBitMap::covers(MemRegion mr) const {
  6280   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  6281   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
  6282          "size inconsistency");
  6283   return (mr.start() >= _bmStartWord) &&
  6284          (mr.end()   <= endWord());
  6287 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
  6288     return (start >= _bmStartWord && (start + size) <= endWord());
  6291 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
  6292   // verify that there are no 1 bits in the interval [left, right)
  6293   FalseBitMapClosure falseBitMapClosure;
  6294   iterate(&falseBitMapClosure, left, right);
  6297 void CMSBitMap::region_invariant(MemRegion mr)
  6299   assert_locked();
  6300   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
  6301   assert(!mr.is_empty(), "unexpected empty region");
  6302   assert(covers(mr), "mr should be covered by bit map");
  6303   // convert address range into offset range
  6304   size_t start_ofs = heapWordToOffset(mr.start());
  6305   // Make sure that end() is appropriately aligned
  6306   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
  6307                         (1 << (_shifter+LogHeapWordSize))),
  6308          "Misaligned mr.end()");
  6309   size_t end_ofs   = heapWordToOffset(mr.end());
  6310   assert(end_ofs > start_ofs, "Should mark at least one bit");
  6313 #endif
  6315 bool CMSMarkStack::allocate(size_t size) {
  6316   // allocate a stack of the requisite depth
  6317   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6318                    size * sizeof(oop)));
  6319   if (!rs.is_reserved()) {
  6320     warning("CMSMarkStack allocation failure");
  6321     return false;
  6323   if (!_virtual_space.initialize(rs, rs.size())) {
  6324     warning("CMSMarkStack backing store failure");
  6325     return false;
  6327   assert(_virtual_space.committed_size() == rs.size(),
  6328          "didn't reserve backing store for all of CMS stack?");
  6329   _base = (oop*)(_virtual_space.low());
  6330   _index = 0;
  6331   _capacity = size;
  6332   NOT_PRODUCT(_max_depth = 0);
  6333   return true;
  6336 // XXX FIX ME !!! In the MT case we come in here holding a
  6337 // leaf lock. For printing we need to take a further lock
  6338 // which has lower rank. We need to recallibrate the two
  6339 // lock-ranks involved in order to be able to rpint the
  6340 // messages below. (Or defer the printing to the caller.
  6341 // For now we take the expedient path of just disabling the
  6342 // messages for the problematic case.)
  6343 void CMSMarkStack::expand() {
  6344   assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
  6345   if (_capacity == CMSMarkStackSizeMax) {
  6346     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6347       // We print a warning message only once per CMS cycle.
  6348       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
  6350     return;
  6352   // Double capacity if possible
  6353   size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
  6354   // Do not give up existing stack until we have managed to
  6355   // get the double capacity that we desired.
  6356   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6357                    new_capacity * sizeof(oop)));
  6358   if (rs.is_reserved()) {
  6359     // Release the backing store associated with old stack
  6360     _virtual_space.release();
  6361     // Reinitialize virtual space for new stack
  6362     if (!_virtual_space.initialize(rs, rs.size())) {
  6363       fatal("Not enough swap for expanded marking stack");
  6365     _base = (oop*)(_virtual_space.low());
  6366     _index = 0;
  6367     _capacity = new_capacity;
  6368   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6369     // Failed to double capacity, continue;
  6370     // we print a detail message only once per CMS cycle.
  6371     gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
  6372             SIZE_FORMAT"K",
  6373             _capacity / K, new_capacity / K);
  6378 // Closures
  6379 // XXX: there seems to be a lot of code  duplication here;
  6380 // should refactor and consolidate common code.
  6382 // This closure is used to mark refs into the CMS generation in
  6383 // the CMS bit map. Called at the first checkpoint. This closure
  6384 // assumes that we do not need to re-mark dirty cards; if the CMS
  6385 // generation on which this is used is not an oldest (modulo perm gen)
  6386 // generation then this will lose younger_gen cards!
  6388 MarkRefsIntoClosure::MarkRefsIntoClosure(
  6389   MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
  6390     _span(span),
  6391     _bitMap(bitMap),
  6392     _should_do_nmethods(should_do_nmethods)
  6394     assert(_ref_processor == NULL, "deliberately left NULL");
  6395     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
  6398 void MarkRefsIntoClosure::do_oop(oop obj) {
  6399   // if p points into _span, then mark corresponding bit in _markBitMap
  6400   assert(obj->is_oop(), "expected an oop");
  6401   HeapWord* addr = (HeapWord*)obj;
  6402   if (_span.contains(addr)) {
  6403     // this should be made more efficient
  6404     _bitMap->mark(addr);
  6408 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
  6409 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
  6411 // A variant of the above, used for CMS marking verification.
  6412 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
  6413   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  6414   bool should_do_nmethods):
  6415     _span(span),
  6416     _verification_bm(verification_bm),
  6417     _cms_bm(cms_bm),
  6418     _should_do_nmethods(should_do_nmethods) {
  6419     assert(_ref_processor == NULL, "deliberately left NULL");
  6420     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
  6423 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
  6424   // if p points into _span, then mark corresponding bit in _markBitMap
  6425   assert(obj->is_oop(), "expected an oop");
  6426   HeapWord* addr = (HeapWord*)obj;
  6427   if (_span.contains(addr)) {
  6428     _verification_bm->mark(addr);
  6429     if (!_cms_bm->isMarked(addr)) {
  6430       oop(addr)->print();
  6431       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
  6432       fatal("... aborting");
  6437 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6438 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6440 //////////////////////////////////////////////////
  6441 // MarkRefsIntoAndScanClosure
  6442 //////////////////////////////////////////////////
  6444 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
  6445                                                        ReferenceProcessor* rp,
  6446                                                        CMSBitMap* bit_map,
  6447                                                        CMSBitMap* mod_union_table,
  6448                                                        CMSMarkStack*  mark_stack,
  6449                                                        CMSMarkStack*  revisit_stack,
  6450                                                        CMSCollector* collector,
  6451                                                        bool should_yield,
  6452                                                        bool concurrent_precleaning):
  6453   _collector(collector),
  6454   _span(span),
  6455   _bit_map(bit_map),
  6456   _mark_stack(mark_stack),
  6457   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
  6458                       mark_stack, revisit_stack, concurrent_precleaning),
  6459   _yield(should_yield),
  6460   _concurrent_precleaning(concurrent_precleaning),
  6461   _freelistLock(NULL)
  6463   _ref_processor = rp;
  6464   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  6467 // This closure is used to mark refs into the CMS generation at the
  6468 // second (final) checkpoint, and to scan and transitively follow
  6469 // the unmarked oops. It is also used during the concurrent precleaning
  6470 // phase while scanning objects on dirty cards in the CMS generation.
  6471 // The marks are made in the marking bit map and the marking stack is
  6472 // used for keeping the (newly) grey objects during the scan.
  6473 // The parallel version (Par_...) appears further below.
  6474 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  6475   if (obj != NULL) {
  6476     assert(obj->is_oop(), "expected an oop");
  6477     HeapWord* addr = (HeapWord*)obj;
  6478     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  6479     assert(_collector->overflow_list_is_empty(),
  6480            "overflow list should be empty");
  6481     if (_span.contains(addr) &&
  6482         !_bit_map->isMarked(addr)) {
  6483       // mark bit map (object is now grey)
  6484       _bit_map->mark(addr);
  6485       // push on marking stack (stack should be empty), and drain the
  6486       // stack by applying this closure to the oops in the oops popped
  6487       // from the stack (i.e. blacken the grey objects)
  6488       bool res = _mark_stack->push(obj);
  6489       assert(res, "Should have space to push on empty stack");
  6490       do {
  6491         oop new_oop = _mark_stack->pop();
  6492         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  6493         assert(new_oop->is_parsable(), "Found unparsable oop");
  6494         assert(_bit_map->isMarked((HeapWord*)new_oop),
  6495                "only grey objects on this stack");
  6496         // iterate over the oops in this oop, marking and pushing
  6497         // the ones in CMS heap (i.e. in _span).
  6498         new_oop->oop_iterate(&_pushAndMarkClosure);
  6499         // check if it's time to yield
  6500         do_yield_check();
  6501       } while (!_mark_stack->isEmpty() ||
  6502                (!_concurrent_precleaning && take_from_overflow_list()));
  6503         // if marking stack is empty, and we are not doing this
  6504         // during precleaning, then check the overflow list
  6506     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  6507     assert(_collector->overflow_list_is_empty(),
  6508            "overflow list was drained above");
  6509     // We could restore evacuated mark words, if any, used for
  6510     // overflow list links here because the overflow list is
  6511     // provably empty here. That would reduce the maximum
  6512     // size requirements for preserved_{oop,mark}_stack.
  6513     // But we'll just postpone it until we are all done
  6514     // so we can just stream through.
  6515     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
  6516       _collector->restore_preserved_marks_if_any();
  6517       assert(_collector->no_preserved_marks(), "No preserved marks");
  6519     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
  6520            "All preserved marks should have been restored above");
  6524 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6525 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6527 void MarkRefsIntoAndScanClosure::do_yield_work() {
  6528   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6529          "CMS thread should hold CMS token");
  6530   assert_lock_strong(_freelistLock);
  6531   assert_lock_strong(_bit_map->lock());
  6532   // relinquish the free_list_lock and bitMaplock()
  6533   _bit_map->lock()->unlock();
  6534   _freelistLock->unlock();
  6535   ConcurrentMarkSweepThread::desynchronize(true);
  6536   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6537   _collector->stopTimer();
  6538   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6539   if (PrintCMSStatistics != 0) {
  6540     _collector->incrementYields();
  6542   _collector->icms_wait();
  6544   // See the comment in coordinator_yield()
  6545   for (unsigned i = 0;
  6546        i < CMSYieldSleepCount &&
  6547        ConcurrentMarkSweepThread::should_yield() &&
  6548        !CMSCollector::foregroundGCIsActive();
  6549        ++i) {
  6550     os::sleep(Thread::current(), 1, false);
  6551     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6554   ConcurrentMarkSweepThread::synchronize(true);
  6555   _freelistLock->lock_without_safepoint_check();
  6556   _bit_map->lock()->lock_without_safepoint_check();
  6557   _collector->startTimer();
  6560 ///////////////////////////////////////////////////////////
  6561 // Par_MarkRefsIntoAndScanClosure: a parallel version of
  6562 //                                 MarkRefsIntoAndScanClosure
  6563 ///////////////////////////////////////////////////////////
  6564 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
  6565   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
  6566   CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack*  revisit_stack):
  6567   _span(span),
  6568   _bit_map(bit_map),
  6569   _work_queue(work_queue),
  6570   _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  6571                        (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
  6572   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
  6573                           revisit_stack)
  6575   _ref_processor = rp;
  6576   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  6579 // This closure is used to mark refs into the CMS generation at the
  6580 // second (final) checkpoint, and to scan and transitively follow
  6581 // the unmarked oops. The marks are made in the marking bit map and
  6582 // the work_queue is used for keeping the (newly) grey objects during
  6583 // the scan phase whence they are also available for stealing by parallel
  6584 // threads. Since the marking bit map is shared, updates are
  6585 // synchronized (via CAS).
  6586 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  6587   if (obj != NULL) {
  6588     // Ignore mark word because this could be an already marked oop
  6589     // that may be chained at the end of the overflow list.
  6590     assert(obj->is_oop(true), "expected an oop");
  6591     HeapWord* addr = (HeapWord*)obj;
  6592     if (_span.contains(addr) &&
  6593         !_bit_map->isMarked(addr)) {
  6594       // mark bit map (object will become grey):
  6595       // It is possible for several threads to be
  6596       // trying to "claim" this object concurrently;
  6597       // the unique thread that succeeds in marking the
  6598       // object first will do the subsequent push on
  6599       // to the work queue (or overflow list).
  6600       if (_bit_map->par_mark(addr)) {
  6601         // push on work_queue (which may not be empty), and trim the
  6602         // queue to an appropriate length by applying this closure to
  6603         // the oops in the oops popped from the stack (i.e. blacken the
  6604         // grey objects)
  6605         bool res = _work_queue->push(obj);
  6606         assert(res, "Low water mark should be less than capacity?");
  6607         trim_queue(_low_water_mark);
  6608       } // Else, another thread claimed the object
  6613 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6614 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6616 // This closure is used to rescan the marked objects on the dirty cards
  6617 // in the mod union table and the card table proper.
  6618 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
  6619   oop p, MemRegion mr) {
  6621   size_t size = 0;
  6622   HeapWord* addr = (HeapWord*)p;
  6623   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6624   assert(_span.contains(addr), "we are scanning the CMS generation");
  6625   // check if it's time to yield
  6626   if (do_yield_check()) {
  6627     // We yielded for some foreground stop-world work,
  6628     // and we have been asked to abort this ongoing preclean cycle.
  6629     return 0;
  6631   if (_bitMap->isMarked(addr)) {
  6632     // it's marked; is it potentially uninitialized?
  6633     if (p->klass_or_null() != NULL) {
  6634       if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
  6635         // Signal precleaning to redirty the card since
  6636         // the klass pointer is already installed.
  6637         assert(size == 0, "Initial value");
  6638       } else {
  6639         assert(p->is_parsable(), "must be parsable.");
  6640         // an initialized object; ignore mark word in verification below
  6641         // since we are running concurrent with mutators
  6642         assert(p->is_oop(true), "should be an oop");
  6643         if (p->is_objArray()) {
  6644           // objArrays are precisely marked; restrict scanning
  6645           // to dirty cards only.
  6646           size = CompactibleFreeListSpace::adjustObjectSize(
  6647                    p->oop_iterate(_scanningClosure, mr));
  6648         } else {
  6649           // A non-array may have been imprecisely marked; we need
  6650           // to scan object in its entirety.
  6651           size = CompactibleFreeListSpace::adjustObjectSize(
  6652                    p->oop_iterate(_scanningClosure));
  6654         #ifdef DEBUG
  6655           size_t direct_size =
  6656             CompactibleFreeListSpace::adjustObjectSize(p->size());
  6657           assert(size == direct_size, "Inconsistency in size");
  6658           assert(size >= 3, "Necessary for Printezis marks to work");
  6659           if (!_bitMap->isMarked(addr+1)) {
  6660             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
  6661           } else {
  6662             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
  6663             assert(_bitMap->isMarked(addr+size-1),
  6664                    "inconsistent Printezis mark");
  6666         #endif // DEBUG
  6668     } else {
  6669       // an unitialized object
  6670       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
  6671       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  6672       size = pointer_delta(nextOneAddr + 1, addr);
  6673       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6674              "alignment problem");
  6675       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
  6676       // will dirty the card when the klass pointer is installed in the
  6677       // object (signalling the completion of initialization).
  6679   } else {
  6680     // Either a not yet marked object or an uninitialized object
  6681     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  6682       // An uninitialized object, skip to the next card, since
  6683       // we may not be able to read its P-bits yet.
  6684       assert(size == 0, "Initial value");
  6685     } else {
  6686       // An object not (yet) reached by marking: we merely need to
  6687       // compute its size so as to go look at the next block.
  6688       assert(p->is_oop(true), "should be an oop");
  6689       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
  6692   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6693   return size;
  6696 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
  6697   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6698          "CMS thread should hold CMS token");
  6699   assert_lock_strong(_freelistLock);
  6700   assert_lock_strong(_bitMap->lock());
  6701   // relinquish the free_list_lock and bitMaplock()
  6702   _bitMap->lock()->unlock();
  6703   _freelistLock->unlock();
  6704   ConcurrentMarkSweepThread::desynchronize(true);
  6705   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6706   _collector->stopTimer();
  6707   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6708   if (PrintCMSStatistics != 0) {
  6709     _collector->incrementYields();
  6711   _collector->icms_wait();
  6713   // See the comment in coordinator_yield()
  6714   for (unsigned i = 0; i < CMSYieldSleepCount &&
  6715                    ConcurrentMarkSweepThread::should_yield() &&
  6716                    !CMSCollector::foregroundGCIsActive(); ++i) {
  6717     os::sleep(Thread::current(), 1, false);
  6718     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6721   ConcurrentMarkSweepThread::synchronize(true);
  6722   _freelistLock->lock_without_safepoint_check();
  6723   _bitMap->lock()->lock_without_safepoint_check();
  6724   _collector->startTimer();
  6728 //////////////////////////////////////////////////////////////////
  6729 // SurvivorSpacePrecleanClosure
  6730 //////////////////////////////////////////////////////////////////
  6731 // This (single-threaded) closure is used to preclean the oops in
  6732 // the survivor spaces.
  6733 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
  6735   HeapWord* addr = (HeapWord*)p;
  6736   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6737   assert(!_span.contains(addr), "we are scanning the survivor spaces");
  6738   assert(p->klass_or_null() != NULL, "object should be initializd");
  6739   assert(p->is_parsable(), "must be parsable.");
  6740   // an initialized object; ignore mark word in verification below
  6741   // since we are running concurrent with mutators
  6742   assert(p->is_oop(true), "should be an oop");
  6743   // Note that we do not yield while we iterate over
  6744   // the interior oops of p, pushing the relevant ones
  6745   // on our marking stack.
  6746   size_t size = p->oop_iterate(_scanning_closure);
  6747   do_yield_check();
  6748   // Observe that below, we do not abandon the preclean
  6749   // phase as soon as we should; rather we empty the
  6750   // marking stack before returning. This is to satisfy
  6751   // some existing assertions. In general, it may be a
  6752   // good idea to abort immediately and complete the marking
  6753   // from the grey objects at a later time.
  6754   while (!_mark_stack->isEmpty()) {
  6755     oop new_oop = _mark_stack->pop();
  6756     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  6757     assert(new_oop->is_parsable(), "Found unparsable oop");
  6758     assert(_bit_map->isMarked((HeapWord*)new_oop),
  6759            "only grey objects on this stack");
  6760     // iterate over the oops in this oop, marking and pushing
  6761     // the ones in CMS heap (i.e. in _span).
  6762     new_oop->oop_iterate(_scanning_closure);
  6763     // check if it's time to yield
  6764     do_yield_check();
  6766   unsigned int after_count =
  6767     GenCollectedHeap::heap()->total_collections();
  6768   bool abort = (_before_count != after_count) ||
  6769                _collector->should_abort_preclean();
  6770   return abort ? 0 : size;
  6773 void SurvivorSpacePrecleanClosure::do_yield_work() {
  6774   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6775          "CMS thread should hold CMS token");
  6776   assert_lock_strong(_bit_map->lock());
  6777   // Relinquish the bit map lock
  6778   _bit_map->lock()->unlock();
  6779   ConcurrentMarkSweepThread::desynchronize(true);
  6780   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6781   _collector->stopTimer();
  6782   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6783   if (PrintCMSStatistics != 0) {
  6784     _collector->incrementYields();
  6786   _collector->icms_wait();
  6788   // See the comment in coordinator_yield()
  6789   for (unsigned i = 0; i < CMSYieldSleepCount &&
  6790                        ConcurrentMarkSweepThread::should_yield() &&
  6791                        !CMSCollector::foregroundGCIsActive(); ++i) {
  6792     os::sleep(Thread::current(), 1, false);
  6793     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6796   ConcurrentMarkSweepThread::synchronize(true);
  6797   _bit_map->lock()->lock_without_safepoint_check();
  6798   _collector->startTimer();
  6801 // This closure is used to rescan the marked objects on the dirty cards
  6802 // in the mod union table and the card table proper. In the parallel
  6803 // case, although the bitMap is shared, we do a single read so the
  6804 // isMarked() query is "safe".
  6805 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
  6806   // Ignore mark word because we are running concurrent with mutators
  6807   assert(p->is_oop_or_null(true), "expected an oop or null");
  6808   HeapWord* addr = (HeapWord*)p;
  6809   assert(_span.contains(addr), "we are scanning the CMS generation");
  6810   bool is_obj_array = false;
  6811   #ifdef DEBUG
  6812     if (!_parallel) {
  6813       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  6814       assert(_collector->overflow_list_is_empty(),
  6815              "overflow list should be empty");
  6818   #endif // DEBUG
  6819   if (_bit_map->isMarked(addr)) {
  6820     // Obj arrays are precisely marked, non-arrays are not;
  6821     // so we scan objArrays precisely and non-arrays in their
  6822     // entirety.
  6823     if (p->is_objArray()) {
  6824       is_obj_array = true;
  6825       if (_parallel) {
  6826         p->oop_iterate(_par_scan_closure, mr);
  6827       } else {
  6828         p->oop_iterate(_scan_closure, mr);
  6830     } else {
  6831       if (_parallel) {
  6832         p->oop_iterate(_par_scan_closure);
  6833       } else {
  6834         p->oop_iterate(_scan_closure);
  6838   #ifdef DEBUG
  6839     if (!_parallel) {
  6840       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  6841       assert(_collector->overflow_list_is_empty(),
  6842              "overflow list should be empty");
  6845   #endif // DEBUG
  6846   return is_obj_array;
  6849 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
  6850                         MemRegion span,
  6851                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
  6852                         CMSMarkStack*  revisitStack,
  6853                         bool should_yield, bool verifying):
  6854   _collector(collector),
  6855   _span(span),
  6856   _bitMap(bitMap),
  6857   _mut(&collector->_modUnionTable),
  6858   _markStack(markStack),
  6859   _revisitStack(revisitStack),
  6860   _yield(should_yield),
  6861   _skipBits(0)
  6863   assert(_markStack->isEmpty(), "stack should be empty");
  6864   _finger = _bitMap->startWord();
  6865   _threshold = _finger;
  6866   assert(_collector->_restart_addr == NULL, "Sanity check");
  6867   assert(_span.contains(_finger), "Out of bounds _finger?");
  6868   DEBUG_ONLY(_verifying = verifying;)
  6871 void MarkFromRootsClosure::reset(HeapWord* addr) {
  6872   assert(_markStack->isEmpty(), "would cause duplicates on stack");
  6873   assert(_span.contains(addr), "Out of bounds _finger?");
  6874   _finger = addr;
  6875   _threshold = (HeapWord*)round_to(
  6876                  (intptr_t)_finger, CardTableModRefBS::card_size);
  6879 // Should revisit to see if this should be restructured for
  6880 // greater efficiency.
  6881 bool MarkFromRootsClosure::do_bit(size_t offset) {
  6882   if (_skipBits > 0) {
  6883     _skipBits--;
  6884     return true;
  6886   // convert offset into a HeapWord*
  6887   HeapWord* addr = _bitMap->startWord() + offset;
  6888   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
  6889          "address out of range");
  6890   assert(_bitMap->isMarked(addr), "tautology");
  6891   if (_bitMap->isMarked(addr+1)) {
  6892     // this is an allocated but not yet initialized object
  6893     assert(_skipBits == 0, "tautology");
  6894     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
  6895     oop p = oop(addr);
  6896     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  6897       DEBUG_ONLY(if (!_verifying) {)
  6898         // We re-dirty the cards on which this object lies and increase
  6899         // the _threshold so that we'll come back to scan this object
  6900         // during the preclean or remark phase. (CMSCleanOnEnter)
  6901         if (CMSCleanOnEnter) {
  6902           size_t sz = _collector->block_size_using_printezis_bits(addr);
  6903           HeapWord* end_card_addr   = (HeapWord*)round_to(
  6904                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  6905           MemRegion redirty_range = MemRegion(addr, end_card_addr);
  6906           assert(!redirty_range.is_empty(), "Arithmetical tautology");
  6907           // Bump _threshold to end_card_addr; note that
  6908           // _threshold cannot possibly exceed end_card_addr, anyhow.
  6909           // This prevents future clearing of the card as the scan proceeds
  6910           // to the right.
  6911           assert(_threshold <= end_card_addr,
  6912                  "Because we are just scanning into this object");
  6913           if (_threshold < end_card_addr) {
  6914             _threshold = end_card_addr;
  6916           if (p->klass_or_null() != NULL) {
  6917             // Redirty the range of cards...
  6918             _mut->mark_range(redirty_range);
  6919           } // ...else the setting of klass will dirty the card anyway.
  6921       DEBUG_ONLY(})
  6922       return true;
  6925   scanOopsInOop(addr);
  6926   return true;
  6929 // We take a break if we've been at this for a while,
  6930 // so as to avoid monopolizing the locks involved.
  6931 void MarkFromRootsClosure::do_yield_work() {
  6932   // First give up the locks, then yield, then re-lock
  6933   // We should probably use a constructor/destructor idiom to
  6934   // do this unlock/lock or modify the MutexUnlocker class to
  6935   // serve our purpose. XXX
  6936   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6937          "CMS thread should hold CMS token");
  6938   assert_lock_strong(_bitMap->lock());
  6939   _bitMap->lock()->unlock();
  6940   ConcurrentMarkSweepThread::desynchronize(true);
  6941   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6942   _collector->stopTimer();
  6943   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6944   if (PrintCMSStatistics != 0) {
  6945     _collector->incrementYields();
  6947   _collector->icms_wait();
  6949   // See the comment in coordinator_yield()
  6950   for (unsigned i = 0; i < CMSYieldSleepCount &&
  6951                        ConcurrentMarkSweepThread::should_yield() &&
  6952                        !CMSCollector::foregroundGCIsActive(); ++i) {
  6953     os::sleep(Thread::current(), 1, false);
  6954     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6957   ConcurrentMarkSweepThread::synchronize(true);
  6958   _bitMap->lock()->lock_without_safepoint_check();
  6959   _collector->startTimer();
  6962 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
  6963   assert(_bitMap->isMarked(ptr), "expected bit to be set");
  6964   assert(_markStack->isEmpty(),
  6965          "should drain stack to limit stack usage");
  6966   // convert ptr to an oop preparatory to scanning
  6967   oop obj = oop(ptr);
  6968   // Ignore mark word in verification below, since we
  6969   // may be running concurrent with mutators.
  6970   assert(obj->is_oop(true), "should be an oop");
  6971   assert(_finger <= ptr, "_finger runneth ahead");
  6972   // advance the finger to right end of this object
  6973   _finger = ptr + obj->size();
  6974   assert(_finger > ptr, "we just incremented it above");
  6975   // On large heaps, it may take us some time to get through
  6976   // the marking phase (especially if running iCMS). During
  6977   // this time it's possible that a lot of mutations have
  6978   // accumulated in the card table and the mod union table --
  6979   // these mutation records are redundant until we have
  6980   // actually traced into the corresponding card.
  6981   // Here, we check whether advancing the finger would make
  6982   // us cross into a new card, and if so clear corresponding
  6983   // cards in the MUT (preclean them in the card-table in the
  6984   // future).
  6986   DEBUG_ONLY(if (!_verifying) {)
  6987     // The clean-on-enter optimization is disabled by default,
  6988     // until we fix 6178663.
  6989     if (CMSCleanOnEnter && (_finger > _threshold)) {
  6990       // [_threshold, _finger) represents the interval
  6991       // of cards to be cleared  in MUT (or precleaned in card table).
  6992       // The set of cards to be cleared is all those that overlap
  6993       // with the interval [_threshold, _finger); note that
  6994       // _threshold is always kept card-aligned but _finger isn't
  6995       // always card-aligned.
  6996       HeapWord* old_threshold = _threshold;
  6997       assert(old_threshold == (HeapWord*)round_to(
  6998               (intptr_t)old_threshold, CardTableModRefBS::card_size),
  6999              "_threshold should always be card-aligned");
  7000       _threshold = (HeapWord*)round_to(
  7001                      (intptr_t)_finger, CardTableModRefBS::card_size);
  7002       MemRegion mr(old_threshold, _threshold);
  7003       assert(!mr.is_empty(), "Control point invariant");
  7004       assert(_span.contains(mr), "Should clear within span");
  7005       // XXX When _finger crosses from old gen into perm gen
  7006       // we may be doing unnecessary cleaning; do better in the
  7007       // future by detecting that condition and clearing fewer
  7008       // MUT/CT entries.
  7009       _mut->clear_range(mr);
  7011   DEBUG_ONLY(})
  7013   // Note: the finger doesn't advance while we drain
  7014   // the stack below.
  7015   PushOrMarkClosure pushOrMarkClosure(_collector,
  7016                                       _span, _bitMap, _markStack,
  7017                                       _revisitStack,
  7018                                       _finger, this);
  7019   bool res = _markStack->push(obj);
  7020   assert(res, "Empty non-zero size stack should have space for single push");
  7021   while (!_markStack->isEmpty()) {
  7022     oop new_oop = _markStack->pop();
  7023     // Skip verifying header mark word below because we are
  7024     // running concurrent with mutators.
  7025     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7026     // now scan this oop's oops
  7027     new_oop->oop_iterate(&pushOrMarkClosure);
  7028     do_yield_check();
  7030   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
  7033 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
  7034                        CMSCollector* collector, MemRegion span,
  7035                        CMSBitMap* bit_map,
  7036                        OopTaskQueue* work_queue,
  7037                        CMSMarkStack*  overflow_stack,
  7038                        CMSMarkStack*  revisit_stack,
  7039                        bool should_yield):
  7040   _collector(collector),
  7041   _whole_span(collector->_span),
  7042   _span(span),
  7043   _bit_map(bit_map),
  7044   _mut(&collector->_modUnionTable),
  7045   _work_queue(work_queue),
  7046   _overflow_stack(overflow_stack),
  7047   _revisit_stack(revisit_stack),
  7048   _yield(should_yield),
  7049   _skip_bits(0),
  7050   _task(task)
  7052   assert(_work_queue->size() == 0, "work_queue should be empty");
  7053   _finger = span.start();
  7054   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
  7055   assert(_span.contains(_finger), "Out of bounds _finger?");
  7058 // Should revisit to see if this should be restructured for
  7059 // greater efficiency.
  7060 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
  7061   if (_skip_bits > 0) {
  7062     _skip_bits--;
  7063     return true;
  7065   // convert offset into a HeapWord*
  7066   HeapWord* addr = _bit_map->startWord() + offset;
  7067   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
  7068          "address out of range");
  7069   assert(_bit_map->isMarked(addr), "tautology");
  7070   if (_bit_map->isMarked(addr+1)) {
  7071     // this is an allocated object that might not yet be initialized
  7072     assert(_skip_bits == 0, "tautology");
  7073     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
  7074     oop p = oop(addr);
  7075     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  7076       // in the case of Clean-on-Enter optimization, redirty card
  7077       // and avoid clearing card by increasing  the threshold.
  7078       return true;
  7081   scan_oops_in_oop(addr);
  7082   return true;
  7085 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
  7086   assert(_bit_map->isMarked(ptr), "expected bit to be set");
  7087   // Should we assert that our work queue is empty or
  7088   // below some drain limit?
  7089   assert(_work_queue->size() == 0,
  7090          "should drain stack to limit stack usage");
  7091   // convert ptr to an oop preparatory to scanning
  7092   oop obj = oop(ptr);
  7093   // Ignore mark word in verification below, since we
  7094   // may be running concurrent with mutators.
  7095   assert(obj->is_oop(true), "should be an oop");
  7096   assert(_finger <= ptr, "_finger runneth ahead");
  7097   // advance the finger to right end of this object
  7098   _finger = ptr + obj->size();
  7099   assert(_finger > ptr, "we just incremented it above");
  7100   // On large heaps, it may take us some time to get through
  7101   // the marking phase (especially if running iCMS). During
  7102   // this time it's possible that a lot of mutations have
  7103   // accumulated in the card table and the mod union table --
  7104   // these mutation records are redundant until we have
  7105   // actually traced into the corresponding card.
  7106   // Here, we check whether advancing the finger would make
  7107   // us cross into a new card, and if so clear corresponding
  7108   // cards in the MUT (preclean them in the card-table in the
  7109   // future).
  7111   // The clean-on-enter optimization is disabled by default,
  7112   // until we fix 6178663.
  7113   if (CMSCleanOnEnter && (_finger > _threshold)) {
  7114     // [_threshold, _finger) represents the interval
  7115     // of cards to be cleared  in MUT (or precleaned in card table).
  7116     // The set of cards to be cleared is all those that overlap
  7117     // with the interval [_threshold, _finger); note that
  7118     // _threshold is always kept card-aligned but _finger isn't
  7119     // always card-aligned.
  7120     HeapWord* old_threshold = _threshold;
  7121     assert(old_threshold == (HeapWord*)round_to(
  7122             (intptr_t)old_threshold, CardTableModRefBS::card_size),
  7123            "_threshold should always be card-aligned");
  7124     _threshold = (HeapWord*)round_to(
  7125                    (intptr_t)_finger, CardTableModRefBS::card_size);
  7126     MemRegion mr(old_threshold, _threshold);
  7127     assert(!mr.is_empty(), "Control point invariant");
  7128     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
  7129     // XXX When _finger crosses from old gen into perm gen
  7130     // we may be doing unnecessary cleaning; do better in the
  7131     // future by detecting that condition and clearing fewer
  7132     // MUT/CT entries.
  7133     _mut->clear_range(mr);
  7136   // Note: the local finger doesn't advance while we drain
  7137   // the stack below, but the global finger sure can and will.
  7138   HeapWord** gfa = _task->global_finger_addr();
  7139   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
  7140                                       _span, _bit_map,
  7141                                       _work_queue,
  7142                                       _overflow_stack,
  7143                                       _revisit_stack,
  7144                                       _finger,
  7145                                       gfa, this);
  7146   bool res = _work_queue->push(obj);   // overflow could occur here
  7147   assert(res, "Will hold once we use workqueues");
  7148   while (true) {
  7149     oop new_oop;
  7150     if (!_work_queue->pop_local(new_oop)) {
  7151       // We emptied our work_queue; check if there's stuff that can
  7152       // be gotten from the overflow stack.
  7153       if (CMSConcMarkingTask::get_work_from_overflow_stack(
  7154             _overflow_stack, _work_queue)) {
  7155         do_yield_check();
  7156         continue;
  7157       } else {  // done
  7158         break;
  7161     // Skip verifying header mark word below because we are
  7162     // running concurrent with mutators.
  7163     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7164     // now scan this oop's oops
  7165     new_oop->oop_iterate(&pushOrMarkClosure);
  7166     do_yield_check();
  7168   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
  7171 // Yield in response to a request from VM Thread or
  7172 // from mutators.
  7173 void Par_MarkFromRootsClosure::do_yield_work() {
  7174   assert(_task != NULL, "sanity");
  7175   _task->yield();
  7178 // A variant of the above used for verifying CMS marking work.
  7179 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
  7180                         MemRegion span,
  7181                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7182                         CMSMarkStack*  mark_stack):
  7183   _collector(collector),
  7184   _span(span),
  7185   _verification_bm(verification_bm),
  7186   _cms_bm(cms_bm),
  7187   _mark_stack(mark_stack),
  7188   _pam_verify_closure(collector, span, verification_bm, cms_bm,
  7189                       mark_stack)
  7191   assert(_mark_stack->isEmpty(), "stack should be empty");
  7192   _finger = _verification_bm->startWord();
  7193   assert(_collector->_restart_addr == NULL, "Sanity check");
  7194   assert(_span.contains(_finger), "Out of bounds _finger?");
  7197 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
  7198   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
  7199   assert(_span.contains(addr), "Out of bounds _finger?");
  7200   _finger = addr;
  7203 // Should revisit to see if this should be restructured for
  7204 // greater efficiency.
  7205 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
  7206   // convert offset into a HeapWord*
  7207   HeapWord* addr = _verification_bm->startWord() + offset;
  7208   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
  7209          "address out of range");
  7210   assert(_verification_bm->isMarked(addr), "tautology");
  7211   assert(_cms_bm->isMarked(addr), "tautology");
  7213   assert(_mark_stack->isEmpty(),
  7214          "should drain stack to limit stack usage");
  7215   // convert addr to an oop preparatory to scanning
  7216   oop obj = oop(addr);
  7217   assert(obj->is_oop(), "should be an oop");
  7218   assert(_finger <= addr, "_finger runneth ahead");
  7219   // advance the finger to right end of this object
  7220   _finger = addr + obj->size();
  7221   assert(_finger > addr, "we just incremented it above");
  7222   // Note: the finger doesn't advance while we drain
  7223   // the stack below.
  7224   bool res = _mark_stack->push(obj);
  7225   assert(res, "Empty non-zero size stack should have space for single push");
  7226   while (!_mark_stack->isEmpty()) {
  7227     oop new_oop = _mark_stack->pop();
  7228     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
  7229     // now scan this oop's oops
  7230     new_oop->oop_iterate(&_pam_verify_closure);
  7232   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
  7233   return true;
  7236 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
  7237   CMSCollector* collector, MemRegion span,
  7238   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7239   CMSMarkStack*  mark_stack):
  7240   OopClosure(collector->ref_processor()),
  7241   _collector(collector),
  7242   _span(span),
  7243   _verification_bm(verification_bm),
  7244   _cms_bm(cms_bm),
  7245   _mark_stack(mark_stack)
  7246 { }
  7248 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
  7249 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
  7251 // Upon stack overflow, we discard (part of) the stack,
  7252 // remembering the least address amongst those discarded
  7253 // in CMSCollector's _restart_address.
  7254 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
  7255   // Remember the least grey address discarded
  7256   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
  7257   _collector->lower_restart_addr(ra);
  7258   _mark_stack->reset();  // discard stack contents
  7259   _mark_stack->expand(); // expand the stack if possible
  7262 void PushAndMarkVerifyClosure::do_oop(oop obj) {
  7263   assert(obj->is_oop_or_null(), "expected an oop or NULL");
  7264   HeapWord* addr = (HeapWord*)obj;
  7265   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
  7266     // Oop lies in _span and isn't yet grey or black
  7267     _verification_bm->mark(addr);            // now grey
  7268     if (!_cms_bm->isMarked(addr)) {
  7269       oop(addr)->print();
  7270       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
  7271                              addr);
  7272       fatal("... aborting");
  7275     if (!_mark_stack->push(obj)) { // stack overflow
  7276       if (PrintCMSStatistics != 0) {
  7277         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7278                                SIZE_FORMAT, _mark_stack->capacity());
  7280       assert(_mark_stack->isFull(), "Else push should have succeeded");
  7281       handle_stack_overflow(addr);
  7283     // anything including and to the right of _finger
  7284     // will be scanned as we iterate over the remainder of the
  7285     // bit map
  7289 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
  7290                      MemRegion span,
  7291                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
  7292                      CMSMarkStack*  revisitStack,
  7293                      HeapWord* finger, MarkFromRootsClosure* parent) :
  7294   OopClosure(collector->ref_processor()),
  7295   _collector(collector),
  7296   _span(span),
  7297   _bitMap(bitMap),
  7298   _markStack(markStack),
  7299   _revisitStack(revisitStack),
  7300   _finger(finger),
  7301   _parent(parent),
  7302   _should_remember_klasses(collector->should_unload_classes())
  7303 { }
  7305 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
  7306                      MemRegion span,
  7307                      CMSBitMap* bit_map,
  7308                      OopTaskQueue* work_queue,
  7309                      CMSMarkStack*  overflow_stack,
  7310                      CMSMarkStack*  revisit_stack,
  7311                      HeapWord* finger,
  7312                      HeapWord** global_finger_addr,
  7313                      Par_MarkFromRootsClosure* parent) :
  7314   OopClosure(collector->ref_processor()),
  7315   _collector(collector),
  7316   _whole_span(collector->_span),
  7317   _span(span),
  7318   _bit_map(bit_map),
  7319   _work_queue(work_queue),
  7320   _overflow_stack(overflow_stack),
  7321   _revisit_stack(revisit_stack),
  7322   _finger(finger),
  7323   _global_finger_addr(global_finger_addr),
  7324   _parent(parent),
  7325   _should_remember_klasses(collector->should_unload_classes())
  7326 { }
  7328 // Assumes thread-safe access by callers, who are
  7329 // responsible for mutual exclusion.
  7330 void CMSCollector::lower_restart_addr(HeapWord* low) {
  7331   assert(_span.contains(low), "Out of bounds addr");
  7332   if (_restart_addr == NULL) {
  7333     _restart_addr = low;
  7334   } else {
  7335     _restart_addr = MIN2(_restart_addr, low);
  7339 // Upon stack overflow, we discard (part of) the stack,
  7340 // remembering the least address amongst those discarded
  7341 // in CMSCollector's _restart_address.
  7342 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7343   // Remember the least grey address discarded
  7344   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
  7345   _collector->lower_restart_addr(ra);
  7346   _markStack->reset();  // discard stack contents
  7347   _markStack->expand(); // expand the stack if possible
  7350 // Upon stack overflow, we discard (part of) the stack,
  7351 // remembering the least address amongst those discarded
  7352 // in CMSCollector's _restart_address.
  7353 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7354   // We need to do this under a mutex to prevent other
  7355   // workers from interfering with the work done below.
  7356   MutexLockerEx ml(_overflow_stack->par_lock(),
  7357                    Mutex::_no_safepoint_check_flag);
  7358   // Remember the least grey address discarded
  7359   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  7360   _collector->lower_restart_addr(ra);
  7361   _overflow_stack->reset();  // discard stack contents
  7362   _overflow_stack->expand(); // expand the stack if possible
  7365 void PushOrMarkClosure::do_oop(oop obj) {
  7366   // Ignore mark word because we are running concurrent with mutators.
  7367   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7368   HeapWord* addr = (HeapWord*)obj;
  7369   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
  7370     // Oop lies in _span and isn't yet grey or black
  7371     _bitMap->mark(addr);            // now grey
  7372     if (addr < _finger) {
  7373       // the bit map iteration has already either passed, or
  7374       // sampled, this bit in the bit map; we'll need to
  7375       // use the marking stack to scan this oop's oops.
  7376       bool simulate_overflow = false;
  7377       NOT_PRODUCT(
  7378         if (CMSMarkStackOverflowALot &&
  7379             _collector->simulate_overflow()) {
  7380           // simulate a stack overflow
  7381           simulate_overflow = true;
  7384       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
  7385         if (PrintCMSStatistics != 0) {
  7386           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7387                                  SIZE_FORMAT, _markStack->capacity());
  7389         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
  7390         handle_stack_overflow(addr);
  7393     // anything including and to the right of _finger
  7394     // will be scanned as we iterate over the remainder of the
  7395     // bit map
  7396     do_yield_check();
  7400 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
  7401 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
  7403 void Par_PushOrMarkClosure::do_oop(oop obj) {
  7404   // Ignore mark word because we are running concurrent with mutators.
  7405   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7406   HeapWord* addr = (HeapWord*)obj;
  7407   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7408     // Oop lies in _span and isn't yet grey or black
  7409     // We read the global_finger (volatile read) strictly after marking oop
  7410     bool res = _bit_map->par_mark(addr);    // now grey
  7411     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
  7412     // Should we push this marked oop on our stack?
  7413     // -- if someone else marked it, nothing to do
  7414     // -- if target oop is above global finger nothing to do
  7415     // -- if target oop is in chunk and above local finger
  7416     //      then nothing to do
  7417     // -- else push on work queue
  7418     if (   !res       // someone else marked it, they will deal with it
  7419         || (addr >= *gfa)  // will be scanned in a later task
  7420         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
  7421       return;
  7423     // the bit map iteration has already either passed, or
  7424     // sampled, this bit in the bit map; we'll need to
  7425     // use the marking stack to scan this oop's oops.
  7426     bool simulate_overflow = false;
  7427     NOT_PRODUCT(
  7428       if (CMSMarkStackOverflowALot &&
  7429           _collector->simulate_overflow()) {
  7430         // simulate a stack overflow
  7431         simulate_overflow = true;
  7434     if (simulate_overflow ||
  7435         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  7436       // stack overflow
  7437       if (PrintCMSStatistics != 0) {
  7438         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7439                                SIZE_FORMAT, _overflow_stack->capacity());
  7441       // We cannot assert that the overflow stack is full because
  7442       // it may have been emptied since.
  7443       assert(simulate_overflow ||
  7444              _work_queue->size() == _work_queue->max_elems(),
  7445             "Else push should have succeeded");
  7446       handle_stack_overflow(addr);
  7448     do_yield_check();
  7452 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
  7453 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
  7455 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
  7456                                        MemRegion span,
  7457                                        ReferenceProcessor* rp,
  7458                                        CMSBitMap* bit_map,
  7459                                        CMSBitMap* mod_union_table,
  7460                                        CMSMarkStack*  mark_stack,
  7461                                        CMSMarkStack*  revisit_stack,
  7462                                        bool           concurrent_precleaning):
  7463   OopClosure(rp),
  7464   _collector(collector),
  7465   _span(span),
  7466   _bit_map(bit_map),
  7467   _mod_union_table(mod_union_table),
  7468   _mark_stack(mark_stack),
  7469   _revisit_stack(revisit_stack),
  7470   _concurrent_precleaning(concurrent_precleaning),
  7471   _should_remember_klasses(collector->should_unload_classes())
  7473   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7476 // Grey object rescan during pre-cleaning and second checkpoint phases --
  7477 // the non-parallel version (the parallel version appears further below.)
  7478 void PushAndMarkClosure::do_oop(oop obj) {
  7479   // Ignore mark word verification. If during concurrent precleaning,
  7480   // the object monitor may be locked. If during the checkpoint
  7481   // phases, the object may already have been reached by a  different
  7482   // path and may be at the end of the global overflow list (so
  7483   // the mark word may be NULL).
  7484   assert(obj->is_oop_or_null(true /* ignore mark word */),
  7485          "expected an oop or NULL");
  7486   HeapWord* addr = (HeapWord*)obj;
  7487   // Check if oop points into the CMS generation
  7488   // and is not marked
  7489   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7490     // a white object ...
  7491     _bit_map->mark(addr);         // ... now grey
  7492     // push on the marking stack (grey set)
  7493     bool simulate_overflow = false;
  7494     NOT_PRODUCT(
  7495       if (CMSMarkStackOverflowALot &&
  7496           _collector->simulate_overflow()) {
  7497         // simulate a stack overflow
  7498         simulate_overflow = true;
  7501     if (simulate_overflow || !_mark_stack->push(obj)) {
  7502       if (_concurrent_precleaning) {
  7503          // During precleaning we can just dirty the appropriate card(s)
  7504          // in the mod union table, thus ensuring that the object remains
  7505          // in the grey set  and continue. In the case of object arrays
  7506          // we need to dirty all of the cards that the object spans,
  7507          // since the rescan of object arrays will be limited to the
  7508          // dirty cards.
  7509          // Note that no one can be intefering with us in this action
  7510          // of dirtying the mod union table, so no locking or atomics
  7511          // are required.
  7512          if (obj->is_objArray()) {
  7513            size_t sz = obj->size();
  7514            HeapWord* end_card_addr = (HeapWord*)round_to(
  7515                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  7516            MemRegion redirty_range = MemRegion(addr, end_card_addr);
  7517            assert(!redirty_range.is_empty(), "Arithmetical tautology");
  7518            _mod_union_table->mark_range(redirty_range);
  7519          } else {
  7520            _mod_union_table->mark(addr);
  7522          _collector->_ser_pmc_preclean_ovflw++;
  7523       } else {
  7524          // During the remark phase, we need to remember this oop
  7525          // in the overflow list.
  7526          _collector->push_on_overflow_list(obj);
  7527          _collector->_ser_pmc_remark_ovflw++;
  7533 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
  7534                                                MemRegion span,
  7535                                                ReferenceProcessor* rp,
  7536                                                CMSBitMap* bit_map,
  7537                                                OopTaskQueue* work_queue,
  7538                                                CMSMarkStack* revisit_stack):
  7539   OopClosure(rp),
  7540   _collector(collector),
  7541   _span(span),
  7542   _bit_map(bit_map),
  7543   _work_queue(work_queue),
  7544   _revisit_stack(revisit_stack),
  7545   _should_remember_klasses(collector->should_unload_classes())
  7547   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7550 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
  7551 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
  7553 // Grey object rescan during second checkpoint phase --
  7554 // the parallel version.
  7555 void Par_PushAndMarkClosure::do_oop(oop obj) {
  7556   // In the assert below, we ignore the mark word because
  7557   // this oop may point to an already visited object that is
  7558   // on the overflow stack (in which case the mark word has
  7559   // been hijacked for chaining into the overflow stack --
  7560   // if this is the last object in the overflow stack then
  7561   // its mark word will be NULL). Because this object may
  7562   // have been subsequently popped off the global overflow
  7563   // stack, and the mark word possibly restored to the prototypical
  7564   // value, by the time we get to examined this failing assert in
  7565   // the debugger, is_oop_or_null(false) may subsequently start
  7566   // to hold.
  7567   assert(obj->is_oop_or_null(true),
  7568          "expected an oop or NULL");
  7569   HeapWord* addr = (HeapWord*)obj;
  7570   // Check if oop points into the CMS generation
  7571   // and is not marked
  7572   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7573     // a white object ...
  7574     // If we manage to "claim" the object, by being the
  7575     // first thread to mark it, then we push it on our
  7576     // marking stack
  7577     if (_bit_map->par_mark(addr)) {     // ... now grey
  7578       // push on work queue (grey set)
  7579       bool simulate_overflow = false;
  7580       NOT_PRODUCT(
  7581         if (CMSMarkStackOverflowALot &&
  7582             _collector->par_simulate_overflow()) {
  7583           // simulate a stack overflow
  7584           simulate_overflow = true;
  7587       if (simulate_overflow || !_work_queue->push(obj)) {
  7588         _collector->par_push_on_overflow_list(obj);
  7589         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
  7591     } // Else, some other thread got there first
  7595 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
  7596 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
  7598 void PushAndMarkClosure::remember_klass(Klass* k) {
  7599   if (!_revisit_stack->push(oop(k))) {
  7600     fatal("Revisit stack overflowed in PushAndMarkClosure");
  7604 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
  7605   if (!_revisit_stack->par_push(oop(k))) {
  7606     fatal("Revist stack overflowed in Par_PushAndMarkClosure");
  7610 void CMSPrecleanRefsYieldClosure::do_yield_work() {
  7611   Mutex* bml = _collector->bitMapLock();
  7612   assert_lock_strong(bml);
  7613   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7614          "CMS thread should hold CMS token");
  7616   bml->unlock();
  7617   ConcurrentMarkSweepThread::desynchronize(true);
  7619   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7621   _collector->stopTimer();
  7622   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7623   if (PrintCMSStatistics != 0) {
  7624     _collector->incrementYields();
  7626   _collector->icms_wait();
  7628   // See the comment in coordinator_yield()
  7629   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7630                        ConcurrentMarkSweepThread::should_yield() &&
  7631                        !CMSCollector::foregroundGCIsActive(); ++i) {
  7632     os::sleep(Thread::current(), 1, false);
  7633     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7636   ConcurrentMarkSweepThread::synchronize(true);
  7637   bml->lock();
  7639   _collector->startTimer();
  7642 bool CMSPrecleanRefsYieldClosure::should_return() {
  7643   if (ConcurrentMarkSweepThread::should_yield()) {
  7644     do_yield_work();
  7646   return _collector->foregroundGCIsActive();
  7649 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
  7650   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
  7651          "mr should be aligned to start at a card boundary");
  7652   // We'd like to assert:
  7653   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
  7654   //        "mr should be a range of cards");
  7655   // However, that would be too strong in one case -- the last
  7656   // partition ends at _unallocated_block which, in general, can be
  7657   // an arbitrary boundary, not necessarily card aligned.
  7658   if (PrintCMSStatistics != 0) {
  7659     _num_dirty_cards +=
  7660          mr.word_size()/CardTableModRefBS::card_size_in_words;
  7662   _space->object_iterate_mem(mr, &_scan_cl);
  7665 SweepClosure::SweepClosure(CMSCollector* collector,
  7666                            ConcurrentMarkSweepGeneration* g,
  7667                            CMSBitMap* bitMap, bool should_yield) :
  7668   _collector(collector),
  7669   _g(g),
  7670   _sp(g->cmsSpace()),
  7671   _limit(_sp->sweep_limit()),
  7672   _freelistLock(_sp->freelistLock()),
  7673   _bitMap(bitMap),
  7674   _yield(should_yield),
  7675   _inFreeRange(false),           // No free range at beginning of sweep
  7676   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
  7677   _lastFreeRangeCoalesced(false),
  7678   _freeFinger(g->used_region().start())
  7680   NOT_PRODUCT(
  7681     _numObjectsFreed = 0;
  7682     _numWordsFreed   = 0;
  7683     _numObjectsLive = 0;
  7684     _numWordsLive = 0;
  7685     _numObjectsAlreadyFree = 0;
  7686     _numWordsAlreadyFree = 0;
  7687     _last_fc = NULL;
  7689     _sp->initializeIndexedFreeListArrayReturnedBytes();
  7690     _sp->dictionary()->initializeDictReturnedBytes();
  7692   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  7693          "sweep _limit out of bounds");
  7694   if (CMSTraceSweeper) {
  7695     gclog_or_tty->print("\n====================\nStarting new sweep\n");
  7699 // We need this destructor to reclaim any space at the end
  7700 // of the space, which do_blk below may not have added back to
  7701 // the free lists. [basically dealing with the "fringe effect"]
  7702 SweepClosure::~SweepClosure() {
  7703   assert_lock_strong(_freelistLock);
  7704   // this should be treated as the end of a free run if any
  7705   // The current free range should be returned to the free lists
  7706   // as one coalesced chunk.
  7707   if (inFreeRange()) {
  7708     flushCurFreeChunk(freeFinger(),
  7709       pointer_delta(_limit, freeFinger()));
  7710     assert(freeFinger() < _limit, "the finger pointeth off base");
  7711     if (CMSTraceSweeper) {
  7712       gclog_or_tty->print("destructor:");
  7713       gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
  7714                  "[coalesced:"SIZE_FORMAT"]\n",
  7715                  freeFinger(), pointer_delta(_limit, freeFinger()),
  7716                  lastFreeRangeCoalesced());
  7719   NOT_PRODUCT(
  7720     if (Verbose && PrintGC) {
  7721       gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
  7722                           SIZE_FORMAT " bytes",
  7723                  _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
  7724       gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects,  "
  7725                              SIZE_FORMAT" bytes  "
  7726         "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
  7727         _numObjectsLive, _numWordsLive*sizeof(HeapWord),
  7728         _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
  7729       size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
  7730         sizeof(HeapWord);
  7731       gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
  7733       if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
  7734         size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
  7735         size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
  7736         size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
  7737         gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
  7738         gclog_or_tty->print("   Indexed List Returned "SIZE_FORMAT" bytes",
  7739           indexListReturnedBytes);
  7740         gclog_or_tty->print_cr("        Dictionary Returned "SIZE_FORMAT" bytes",
  7741           dictReturnedBytes);
  7745   // Now, in debug mode, just null out the sweep_limit
  7746   NOT_PRODUCT(_sp->clear_sweep_limit();)
  7747   if (CMSTraceSweeper) {
  7748     gclog_or_tty->print("end of sweep\n================\n");
  7752 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
  7753     bool freeRangeInFreeLists) {
  7754   if (CMSTraceSweeper) {
  7755     gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
  7756                freeFinger, _sp->block_size(freeFinger),
  7757                freeRangeInFreeLists);
  7759   assert(!inFreeRange(), "Trampling existing free range");
  7760   set_inFreeRange(true);
  7761   set_lastFreeRangeCoalesced(false);
  7763   set_freeFinger(freeFinger);
  7764   set_freeRangeInFreeLists(freeRangeInFreeLists);
  7765   if (CMSTestInFreeList) {
  7766     if (freeRangeInFreeLists) {
  7767       FreeChunk* fc = (FreeChunk*) freeFinger;
  7768       assert(fc->isFree(), "A chunk on the free list should be free.");
  7769       assert(fc->size() > 0, "Free range should have a size");
  7770       assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
  7775 // Note that the sweeper runs concurrently with mutators. Thus,
  7776 // it is possible for direct allocation in this generation to happen
  7777 // in the middle of the sweep. Note that the sweeper also coalesces
  7778 // contiguous free blocks. Thus, unless the sweeper and the allocator
  7779 // synchronize appropriately freshly allocated blocks may get swept up.
  7780 // This is accomplished by the sweeper locking the free lists while
  7781 // it is sweeping. Thus blocks that are determined to be free are
  7782 // indeed free. There is however one additional complication:
  7783 // blocks that have been allocated since the final checkpoint and
  7784 // mark, will not have been marked and so would be treated as
  7785 // unreachable and swept up. To prevent this, the allocator marks
  7786 // the bit map when allocating during the sweep phase. This leads,
  7787 // however, to a further complication -- objects may have been allocated
  7788 // but not yet initialized -- in the sense that the header isn't yet
  7789 // installed. The sweeper can not then determine the size of the block
  7790 // in order to skip over it. To deal with this case, we use a technique
  7791 // (due to Printezis) to encode such uninitialized block sizes in the
  7792 // bit map. Since the bit map uses a bit per every HeapWord, but the
  7793 // CMS generation has a minimum object size of 3 HeapWords, it follows
  7794 // that "normal marks" won't be adjacent in the bit map (there will
  7795 // always be at least two 0 bits between successive 1 bits). We make use
  7796 // of these "unused" bits to represent uninitialized blocks -- the bit
  7797 // corresponding to the start of the uninitialized object and the next
  7798 // bit are both set. Finally, a 1 bit marks the end of the object that
  7799 // started with the two consecutive 1 bits to indicate its potentially
  7800 // uninitialized state.
  7802 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
  7803   FreeChunk* fc = (FreeChunk*)addr;
  7804   size_t res;
  7806   // check if we are done sweepinrg
  7807   if (addr == _limit) { // we have swept up to the limit, do nothing more
  7808     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  7809            "sweep _limit out of bounds");
  7810     // help the closure application finish
  7811     return pointer_delta(_sp->end(), _limit);
  7813   assert(addr <= _limit, "sweep invariant");
  7815   // check if we should yield
  7816   do_yield_check(addr);
  7817   if (fc->isFree()) {
  7818     // Chunk that is already free
  7819     res = fc->size();
  7820     doAlreadyFreeChunk(fc);
  7821     debug_only(_sp->verifyFreeLists());
  7822     assert(res == fc->size(), "Don't expect the size to change");
  7823     NOT_PRODUCT(
  7824       _numObjectsAlreadyFree++;
  7825       _numWordsAlreadyFree += res;
  7827     NOT_PRODUCT(_last_fc = fc;)
  7828   } else if (!_bitMap->isMarked(addr)) {
  7829     // Chunk is fresh garbage
  7830     res = doGarbageChunk(fc);
  7831     debug_only(_sp->verifyFreeLists());
  7832     NOT_PRODUCT(
  7833       _numObjectsFreed++;
  7834       _numWordsFreed += res;
  7836   } else {
  7837     // Chunk that is alive.
  7838     res = doLiveChunk(fc);
  7839     debug_only(_sp->verifyFreeLists());
  7840     NOT_PRODUCT(
  7841         _numObjectsLive++;
  7842         _numWordsLive += res;
  7845   return res;
  7848 // For the smart allocation, record following
  7849 //  split deaths - a free chunk is removed from its free list because
  7850 //      it is being split into two or more chunks.
  7851 //  split birth - a free chunk is being added to its free list because
  7852 //      a larger free chunk has been split and resulted in this free chunk.
  7853 //  coal death - a free chunk is being removed from its free list because
  7854 //      it is being coalesced into a large free chunk.
  7855 //  coal birth - a free chunk is being added to its free list because
  7856 //      it was created when two or more free chunks where coalesced into
  7857 //      this free chunk.
  7858 //
  7859 // These statistics are used to determine the desired number of free
  7860 // chunks of a given size.  The desired number is chosen to be relative
  7861 // to the end of a CMS sweep.  The desired number at the end of a sweep
  7862 // is the
  7863 //      count-at-end-of-previous-sweep (an amount that was enough)
  7864 //              - count-at-beginning-of-current-sweep  (the excess)
  7865 //              + split-births  (gains in this size during interval)
  7866 //              - split-deaths  (demands on this size during interval)
  7867 // where the interval is from the end of one sweep to the end of the
  7868 // next.
  7869 //
  7870 // When sweeping the sweeper maintains an accumulated chunk which is
  7871 // the chunk that is made up of chunks that have been coalesced.  That
  7872 // will be termed the left-hand chunk.  A new chunk of garbage that
  7873 // is being considered for coalescing will be referred to as the
  7874 // right-hand chunk.
  7875 //
  7876 // When making a decision on whether to coalesce a right-hand chunk with
  7877 // the current left-hand chunk, the current count vs. the desired count
  7878 // of the left-hand chunk is considered.  Also if the right-hand chunk
  7879 // is near the large chunk at the end of the heap (see
  7880 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
  7881 // left-hand chunk is coalesced.
  7882 //
  7883 // When making a decision about whether to split a chunk, the desired count
  7884 // vs. the current count of the candidate to be split is also considered.
  7885 // If the candidate is underpopulated (currently fewer chunks than desired)
  7886 // a chunk of an overpopulated (currently more chunks than desired) size may
  7887 // be chosen.  The "hint" associated with a free list, if non-null, points
  7888 // to a free list which may be overpopulated.
  7889 //
  7891 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
  7892   size_t size = fc->size();
  7893   // Chunks that cannot be coalesced are not in the
  7894   // free lists.
  7895   if (CMSTestInFreeList && !fc->cantCoalesce()) {
  7896     assert(_sp->verifyChunkInFreeLists(fc),
  7897       "free chunk should be in free lists");
  7899   // a chunk that is already free, should not have been
  7900   // marked in the bit map
  7901   HeapWord* addr = (HeapWord*) fc;
  7902   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
  7903   // Verify that the bit map has no bits marked between
  7904   // addr and purported end of this block.
  7905   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  7907   // Some chunks cannot be coalesced in under any circumstances.
  7908   // See the definition of cantCoalesce().
  7909   if (!fc->cantCoalesce()) {
  7910     // This chunk can potentially be coalesced.
  7911     if (_sp->adaptive_freelists()) {
  7912       // All the work is done in
  7913       doPostIsFreeOrGarbageChunk(fc, size);
  7914     } else {  // Not adaptive free lists
  7915       // this is a free chunk that can potentially be coalesced by the sweeper;
  7916       if (!inFreeRange()) {
  7917         // if the next chunk is a free block that can't be coalesced
  7918         // it doesn't make sense to remove this chunk from the free lists
  7919         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
  7920         assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
  7921         if ((HeapWord*)nextChunk < _limit  &&    // there's a next chunk...
  7922             nextChunk->isFree()    &&            // which is free...
  7923             nextChunk->cantCoalesce()) {         // ... but cant be coalesced
  7924           // nothing to do
  7925         } else {
  7926           // Potentially the start of a new free range:
  7927           // Don't eagerly remove it from the free lists.
  7928           // No need to remove it if it will just be put
  7929           // back again.  (Also from a pragmatic point of view
  7930           // if it is a free block in a region that is beyond
  7931           // any allocated blocks, an assertion will fail)
  7932           // Remember the start of a free run.
  7933           initialize_free_range(addr, true);
  7934           // end - can coalesce with next chunk
  7936       } else {
  7937         // the midst of a free range, we are coalescing
  7938         debug_only(record_free_block_coalesced(fc);)
  7939         if (CMSTraceSweeper) {
  7940           gclog_or_tty->print("  -- pick up free block 0x%x (%d)\n", fc, size);
  7942         // remove it from the free lists
  7943         _sp->removeFreeChunkFromFreeLists(fc);
  7944         set_lastFreeRangeCoalesced(true);
  7945         // If the chunk is being coalesced and the current free range is
  7946         // in the free lists, remove the current free range so that it
  7947         // will be returned to the free lists in its entirety - all
  7948         // the coalesced pieces included.
  7949         if (freeRangeInFreeLists()) {
  7950           FreeChunk* ffc = (FreeChunk*) freeFinger();
  7951           assert(ffc->size() == pointer_delta(addr, freeFinger()),
  7952             "Size of free range is inconsistent with chunk size.");
  7953           if (CMSTestInFreeList) {
  7954             assert(_sp->verifyChunkInFreeLists(ffc),
  7955               "free range is not in free lists");
  7957           _sp->removeFreeChunkFromFreeLists(ffc);
  7958           set_freeRangeInFreeLists(false);
  7962   } else {
  7963     // Code path common to both original and adaptive free lists.
  7965     // cant coalesce with previous block; this should be treated
  7966     // as the end of a free run if any
  7967     if (inFreeRange()) {
  7968       // we kicked some butt; time to pick up the garbage
  7969       assert(freeFinger() < addr, "the finger pointeth off base");
  7970       flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
  7972     // else, nothing to do, just continue
  7976 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
  7977   // This is a chunk of garbage.  It is not in any free list.
  7978   // Add it to a free list or let it possibly be coalesced into
  7979   // a larger chunk.
  7980   HeapWord* addr = (HeapWord*) fc;
  7981   size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  7983   if (_sp->adaptive_freelists()) {
  7984     // Verify that the bit map has no bits marked between
  7985     // addr and purported end of just dead object.
  7986     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  7988     doPostIsFreeOrGarbageChunk(fc, size);
  7989   } else {
  7990     if (!inFreeRange()) {
  7991       // start of a new free range
  7992       assert(size > 0, "A free range should have a size");
  7993       initialize_free_range(addr, false);
  7995     } else {
  7996       // this will be swept up when we hit the end of the
  7997       // free range
  7998       if (CMSTraceSweeper) {
  7999         gclog_or_tty->print("  -- pick up garbage 0x%x (%d) \n", fc, size);
  8001       // If the chunk is being coalesced and the current free range is
  8002       // in the free lists, remove the current free range so that it
  8003       // will be returned to the free lists in its entirety - all
  8004       // the coalesced pieces included.
  8005       if (freeRangeInFreeLists()) {
  8006         FreeChunk* ffc = (FreeChunk*)freeFinger();
  8007         assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8008           "Size of free range is inconsistent with chunk size.");
  8009         if (CMSTestInFreeList) {
  8010           assert(_sp->verifyChunkInFreeLists(ffc),
  8011             "free range is not in free lists");
  8013         _sp->removeFreeChunkFromFreeLists(ffc);
  8014         set_freeRangeInFreeLists(false);
  8016       set_lastFreeRangeCoalesced(true);
  8018     // this will be swept up when we hit the end of the free range
  8020     // Verify that the bit map has no bits marked between
  8021     // addr and purported end of just dead object.
  8022     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8024   return size;
  8027 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
  8028   HeapWord* addr = (HeapWord*) fc;
  8029   // The sweeper has just found a live object. Return any accumulated
  8030   // left hand chunk to the free lists.
  8031   if (inFreeRange()) {
  8032     if (_sp->adaptive_freelists()) {
  8033       flushCurFreeChunk(freeFinger(),
  8034                         pointer_delta(addr, freeFinger()));
  8035     } else { // not adaptive freelists
  8036       set_inFreeRange(false);
  8037       // Add the free range back to the free list if it is not already
  8038       // there.
  8039       if (!freeRangeInFreeLists()) {
  8040         assert(freeFinger() < addr, "the finger pointeth off base");
  8041         if (CMSTraceSweeper) {
  8042           gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
  8043             "[coalesced:%d]\n",
  8044             freeFinger(), pointer_delta(addr, freeFinger()),
  8045             lastFreeRangeCoalesced());
  8047         _sp->addChunkAndRepairOffsetTable(freeFinger(),
  8048           pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
  8053   // Common code path for original and adaptive free lists.
  8055   // this object is live: we'd normally expect this to be
  8056   // an oop, and like to assert the following:
  8057   // assert(oop(addr)->is_oop(), "live block should be an oop");
  8058   // However, as we commented above, this may be an object whose
  8059   // header hasn't yet been initialized.
  8060   size_t size;
  8061   assert(_bitMap->isMarked(addr), "Tautology for this control point");
  8062   if (_bitMap->isMarked(addr + 1)) {
  8063     // Determine the size from the bit map, rather than trying to
  8064     // compute it from the object header.
  8065     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  8066     size = pointer_delta(nextOneAddr + 1, addr);
  8067     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  8068            "alignment problem");
  8070     #ifdef DEBUG
  8071       if (oop(addr)->klass_or_null() != NULL &&
  8072           (   !_collector->should_unload_classes()
  8073            || oop(addr)->is_parsable())) {
  8074         // Ignore mark word because we are running concurrent with mutators
  8075         assert(oop(addr)->is_oop(true), "live block should be an oop");
  8076         assert(size ==
  8077                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
  8078                "P-mark and computed size do not agree");
  8080     #endif
  8082   } else {
  8083     // This should be an initialized object that's alive.
  8084     assert(oop(addr)->klass_or_null() != NULL &&
  8085            (!_collector->should_unload_classes()
  8086             || oop(addr)->is_parsable()),
  8087            "Should be an initialized object");
  8088     // Ignore mark word because we are running concurrent with mutators
  8089     assert(oop(addr)->is_oop(true), "live block should be an oop");
  8090     // Verify that the bit map has no bits marked between
  8091     // addr and purported end of this block.
  8092     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  8093     assert(size >= 3, "Necessary for Printezis marks to work");
  8094     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
  8095     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
  8097   return size;
  8100 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
  8101                                             size_t chunkSize) {
  8102   // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
  8103   // scheme.
  8104   bool fcInFreeLists = fc->isFree();
  8105   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
  8106   assert((HeapWord*)fc <= _limit, "sweep invariant");
  8107   if (CMSTestInFreeList && fcInFreeLists) {
  8108     assert(_sp->verifyChunkInFreeLists(fc),
  8109       "free chunk is not in free lists");
  8113   if (CMSTraceSweeper) {
  8114     gclog_or_tty->print_cr("  -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
  8117   HeapWord* addr = (HeapWord*) fc;
  8119   bool coalesce;
  8120   size_t left  = pointer_delta(addr, freeFinger());
  8121   size_t right = chunkSize;
  8122   switch (FLSCoalescePolicy) {
  8123     // numeric value forms a coalition aggressiveness metric
  8124     case 0:  { // never coalesce
  8125       coalesce = false;
  8126       break;
  8128     case 1: { // coalesce if left & right chunks on overpopulated lists
  8129       coalesce = _sp->coalOverPopulated(left) &&
  8130                  _sp->coalOverPopulated(right);
  8131       break;
  8133     case 2: { // coalesce if left chunk on overpopulated list (default)
  8134       coalesce = _sp->coalOverPopulated(left);
  8135       break;
  8137     case 3: { // coalesce if left OR right chunk on overpopulated list
  8138       coalesce = _sp->coalOverPopulated(left) ||
  8139                  _sp->coalOverPopulated(right);
  8140       break;
  8142     case 4: { // always coalesce
  8143       coalesce = true;
  8144       break;
  8146     default:
  8147      ShouldNotReachHere();
  8150   // Should the current free range be coalesced?
  8151   // If the chunk is in a free range and either we decided to coalesce above
  8152   // or the chunk is near the large block at the end of the heap
  8153   // (isNearLargestChunk() returns true), then coalesce this chunk.
  8154   bool doCoalesce = inFreeRange() &&
  8155     (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
  8156   if (doCoalesce) {
  8157     // Coalesce the current free range on the left with the new
  8158     // chunk on the right.  If either is on a free list,
  8159     // it must be removed from the list and stashed in the closure.
  8160     if (freeRangeInFreeLists()) {
  8161       FreeChunk* ffc = (FreeChunk*)freeFinger();
  8162       assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8163         "Size of free range is inconsistent with chunk size.");
  8164       if (CMSTestInFreeList) {
  8165         assert(_sp->verifyChunkInFreeLists(ffc),
  8166           "Chunk is not in free lists");
  8168       _sp->coalDeath(ffc->size());
  8169       _sp->removeFreeChunkFromFreeLists(ffc);
  8170       set_freeRangeInFreeLists(false);
  8172     if (fcInFreeLists) {
  8173       _sp->coalDeath(chunkSize);
  8174       assert(fc->size() == chunkSize,
  8175         "The chunk has the wrong size or is not in the free lists");
  8176       _sp->removeFreeChunkFromFreeLists(fc);
  8178     set_lastFreeRangeCoalesced(true);
  8179   } else {  // not in a free range and/or should not coalesce
  8180     // Return the current free range and start a new one.
  8181     if (inFreeRange()) {
  8182       // In a free range but cannot coalesce with the right hand chunk.
  8183       // Put the current free range into the free lists.
  8184       flushCurFreeChunk(freeFinger(),
  8185         pointer_delta(addr, freeFinger()));
  8187     // Set up for new free range.  Pass along whether the right hand
  8188     // chunk is in the free lists.
  8189     initialize_free_range((HeapWord*)fc, fcInFreeLists);
  8192 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
  8193   assert(inFreeRange(), "Should only be called if currently in a free range.");
  8194   assert(size > 0,
  8195     "A zero sized chunk cannot be added to the free lists.");
  8196   if (!freeRangeInFreeLists()) {
  8197     if(CMSTestInFreeList) {
  8198       FreeChunk* fc = (FreeChunk*) chunk;
  8199       fc->setSize(size);
  8200       assert(!_sp->verifyChunkInFreeLists(fc),
  8201         "chunk should not be in free lists yet");
  8203     if (CMSTraceSweeper) {
  8204       gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
  8205                     chunk, size);
  8207     // A new free range is going to be starting.  The current
  8208     // free range has not been added to the free lists yet or
  8209     // was removed so add it back.
  8210     // If the current free range was coalesced, then the death
  8211     // of the free range was recorded.  Record a birth now.
  8212     if (lastFreeRangeCoalesced()) {
  8213       _sp->coalBirth(size);
  8215     _sp->addChunkAndRepairOffsetTable(chunk, size,
  8216             lastFreeRangeCoalesced());
  8218   set_inFreeRange(false);
  8219   set_freeRangeInFreeLists(false);
  8222 // We take a break if we've been at this for a while,
  8223 // so as to avoid monopolizing the locks involved.
  8224 void SweepClosure::do_yield_work(HeapWord* addr) {
  8225   // Return current free chunk being used for coalescing (if any)
  8226   // to the appropriate freelist.  After yielding, the next
  8227   // free block encountered will start a coalescing range of
  8228   // free blocks.  If the next free block is adjacent to the
  8229   // chunk just flushed, they will need to wait for the next
  8230   // sweep to be coalesced.
  8231   if (inFreeRange()) {
  8232     flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8235   // First give up the locks, then yield, then re-lock.
  8236   // We should probably use a constructor/destructor idiom to
  8237   // do this unlock/lock or modify the MutexUnlocker class to
  8238   // serve our purpose. XXX
  8239   assert_lock_strong(_bitMap->lock());
  8240   assert_lock_strong(_freelistLock);
  8241   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  8242          "CMS thread should hold CMS token");
  8243   _bitMap->lock()->unlock();
  8244   _freelistLock->unlock();
  8245   ConcurrentMarkSweepThread::desynchronize(true);
  8246   ConcurrentMarkSweepThread::acknowledge_yield_request();
  8247   _collector->stopTimer();
  8248   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  8249   if (PrintCMSStatistics != 0) {
  8250     _collector->incrementYields();
  8252   _collector->icms_wait();
  8254   // See the comment in coordinator_yield()
  8255   for (unsigned i = 0; i < CMSYieldSleepCount &&
  8256                        ConcurrentMarkSweepThread::should_yield() &&
  8257                        !CMSCollector::foregroundGCIsActive(); ++i) {
  8258     os::sleep(Thread::current(), 1, false);
  8259     ConcurrentMarkSweepThread::acknowledge_yield_request();
  8262   ConcurrentMarkSweepThread::synchronize(true);
  8263   _freelistLock->lock();
  8264   _bitMap->lock()->lock_without_safepoint_check();
  8265   _collector->startTimer();
  8268 #ifndef PRODUCT
  8269 // This is actually very useful in a product build if it can
  8270 // be called from the debugger.  Compile it into the product
  8271 // as needed.
  8272 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
  8273   return debug_cms_space->verifyChunkInFreeLists(fc);
  8276 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
  8277   if (CMSTraceSweeper) {
  8278     gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
  8281 #endif
  8283 // CMSIsAliveClosure
  8284 bool CMSIsAliveClosure::do_object_b(oop obj) {
  8285   HeapWord* addr = (HeapWord*)obj;
  8286   return addr != NULL &&
  8287          (!_span.contains(addr) || _bit_map->isMarked(addr));
  8290 // CMSKeepAliveClosure: the serial version
  8291 void CMSKeepAliveClosure::do_oop(oop obj) {
  8292   HeapWord* addr = (HeapWord*)obj;
  8293   if (_span.contains(addr) &&
  8294       !_bit_map->isMarked(addr)) {
  8295     _bit_map->mark(addr);
  8296     bool simulate_overflow = false;
  8297     NOT_PRODUCT(
  8298       if (CMSMarkStackOverflowALot &&
  8299           _collector->simulate_overflow()) {
  8300         // simulate a stack overflow
  8301         simulate_overflow = true;
  8304     if (simulate_overflow || !_mark_stack->push(obj)) {
  8305       _collector->push_on_overflow_list(obj);
  8306       _collector->_ser_kac_ovflw++;
  8311 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
  8312 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
  8314 // CMSParKeepAliveClosure: a parallel version of the above.
  8315 // The work queues are private to each closure (thread),
  8316 // but (may be) available for stealing by other threads.
  8317 void CMSParKeepAliveClosure::do_oop(oop obj) {
  8318   HeapWord* addr = (HeapWord*)obj;
  8319   if (_span.contains(addr) &&
  8320       !_bit_map->isMarked(addr)) {
  8321     // In general, during recursive tracing, several threads
  8322     // may be concurrently getting here; the first one to
  8323     // "tag" it, claims it.
  8324     if (_bit_map->par_mark(addr)) {
  8325       bool res = _work_queue->push(obj);
  8326       assert(res, "Low water mark should be much less than capacity");
  8327       // Do a recursive trim in the hope that this will keep
  8328       // stack usage lower, but leave some oops for potential stealers
  8329       trim_queue(_low_water_mark);
  8330     } // Else, another thread got there first
  8334 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
  8335 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
  8337 void CMSParKeepAliveClosure::trim_queue(uint max) {
  8338   while (_work_queue->size() > max) {
  8339     oop new_oop;
  8340     if (_work_queue->pop_local(new_oop)) {
  8341       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  8342       assert(_bit_map->isMarked((HeapWord*)new_oop),
  8343              "no white objects on this stack!");
  8344       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  8345       // iterate over the oops in this oop, marking and pushing
  8346       // the ones in CMS heap (i.e. in _span).
  8347       new_oop->oop_iterate(&_mark_and_push);
  8352 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
  8353   HeapWord* addr = (HeapWord*)obj;
  8354   if (_span.contains(addr) &&
  8355       !_bit_map->isMarked(addr)) {
  8356     if (_bit_map->par_mark(addr)) {
  8357       bool simulate_overflow = false;
  8358       NOT_PRODUCT(
  8359         if (CMSMarkStackOverflowALot &&
  8360             _collector->par_simulate_overflow()) {
  8361           // simulate a stack overflow
  8362           simulate_overflow = true;
  8365       if (simulate_overflow || !_work_queue->push(obj)) {
  8366         _collector->par_push_on_overflow_list(obj);
  8367         _collector->_par_kac_ovflw++;
  8369     } // Else another thread got there already
  8373 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8374 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8376 //////////////////////////////////////////////////////////////////
  8377 //  CMSExpansionCause                /////////////////////////////
  8378 //////////////////////////////////////////////////////////////////
  8379 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
  8380   switch (cause) {
  8381     case _no_expansion:
  8382       return "No expansion";
  8383     case _satisfy_free_ratio:
  8384       return "Free ratio";
  8385     case _satisfy_promotion:
  8386       return "Satisfy promotion";
  8387     case _satisfy_allocation:
  8388       return "allocation";
  8389     case _allocate_par_lab:
  8390       return "Par LAB";
  8391     case _allocate_par_spooling_space:
  8392       return "Par Spooling Space";
  8393     case _adaptive_size_policy:
  8394       return "Ergonomics";
  8395     default:
  8396       return "unknown";
  8400 void CMSDrainMarkingStackClosure::do_void() {
  8401   // the max number to take from overflow list at a time
  8402   const size_t num = _mark_stack->capacity()/4;
  8403   while (!_mark_stack->isEmpty() ||
  8404          // if stack is empty, check the overflow list
  8405          _collector->take_from_overflow_list(num, _mark_stack)) {
  8406     oop obj = _mark_stack->pop();
  8407     HeapWord* addr = (HeapWord*)obj;
  8408     assert(_span.contains(addr), "Should be within span");
  8409     assert(_bit_map->isMarked(addr), "Should be marked");
  8410     assert(obj->is_oop(), "Should be an oop");
  8411     obj->oop_iterate(_keep_alive);
  8415 void CMSParDrainMarkingStackClosure::do_void() {
  8416   // drain queue
  8417   trim_queue(0);
  8420 // Trim our work_queue so its length is below max at return
  8421 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
  8422   while (_work_queue->size() > max) {
  8423     oop new_oop;
  8424     if (_work_queue->pop_local(new_oop)) {
  8425       assert(new_oop->is_oop(), "Expected an oop");
  8426       assert(_bit_map->isMarked((HeapWord*)new_oop),
  8427              "no white objects on this stack!");
  8428       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  8429       // iterate over the oops in this oop, marking and pushing
  8430       // the ones in CMS heap (i.e. in _span).
  8431       new_oop->oop_iterate(&_mark_and_push);
  8436 ////////////////////////////////////////////////////////////////////
  8437 // Support for Marking Stack Overflow list handling and related code
  8438 ////////////////////////////////////////////////////////////////////
  8439 // Much of the following code is similar in shape and spirit to the
  8440 // code used in ParNewGC. We should try and share that code
  8441 // as much as possible in the future.
  8443 #ifndef PRODUCT
  8444 // Debugging support for CMSStackOverflowALot
  8446 // It's OK to call this multi-threaded;  the worst thing
  8447 // that can happen is that we'll get a bunch of closely
  8448 // spaced simulated oveflows, but that's OK, in fact
  8449 // probably good as it would exercise the overflow code
  8450 // under contention.
  8451 bool CMSCollector::simulate_overflow() {
  8452   if (_overflow_counter-- <= 0) { // just being defensive
  8453     _overflow_counter = CMSMarkStackOverflowInterval;
  8454     return true;
  8455   } else {
  8456     return false;
  8460 bool CMSCollector::par_simulate_overflow() {
  8461   return simulate_overflow();
  8463 #endif
  8465 // Single-threaded
  8466 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
  8467   assert(stack->isEmpty(), "Expected precondition");
  8468   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
  8469   size_t i = num;
  8470   oop  cur = _overflow_list;
  8471   const markOop proto = markOopDesc::prototype();
  8472   NOT_PRODUCT(size_t n = 0;)
  8473   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
  8474     next = oop(cur->mark());
  8475     cur->set_mark(proto);   // until proven otherwise
  8476     assert(cur->is_oop(), "Should be an oop");
  8477     bool res = stack->push(cur);
  8478     assert(res, "Bit off more than can chew?");
  8479     NOT_PRODUCT(n++;)
  8481   _overflow_list = cur;
  8482 #ifndef PRODUCT
  8483   assert(_num_par_pushes >= n, "Too many pops?");
  8484   _num_par_pushes -=n;
  8485 #endif
  8486   return !stack->isEmpty();
  8489 // Multi-threaded; use CAS to break off a prefix
  8490 bool CMSCollector::par_take_from_overflow_list(size_t num,
  8491                                                OopTaskQueue* work_q) {
  8492   assert(work_q->size() == 0, "That's the current policy");
  8493   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
  8494   if (_overflow_list == NULL) {
  8495     return false;
  8497   // Grab the entire list; we'll put back a suffix
  8498   oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
  8499   if (prefix == NULL) {  // someone grabbed it before we did ...
  8500     // ... we could spin for a short while, but for now we don't
  8501     return false;
  8503   size_t i = num;
  8504   oop cur = prefix;
  8505   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
  8506   if (cur->mark() != NULL) {
  8507     oop suffix_head = cur->mark(); // suffix will be put back on global list
  8508     cur->set_mark(NULL);           // break off suffix
  8509     // Find tail of suffix so we can prepend suffix to global list
  8510     for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
  8511     oop suffix_tail = cur;
  8512     assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
  8513            "Tautology");
  8514     oop observed_overflow_list = _overflow_list;
  8515     do {
  8516       cur = observed_overflow_list;
  8517       suffix_tail->set_mark(markOop(cur));
  8518       observed_overflow_list =
  8519         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
  8520     } while (cur != observed_overflow_list);
  8523   // Push the prefix elements on work_q
  8524   assert(prefix != NULL, "control point invariant");
  8525   const markOop proto = markOopDesc::prototype();
  8526   oop next;
  8527   NOT_PRODUCT(size_t n = 0;)
  8528   for (cur = prefix; cur != NULL; cur = next) {
  8529     next = oop(cur->mark());
  8530     cur->set_mark(proto);   // until proven otherwise
  8531     assert(cur->is_oop(), "Should be an oop");
  8532     bool res = work_q->push(cur);
  8533     assert(res, "Bit off more than we can chew?");
  8534     NOT_PRODUCT(n++;)
  8536 #ifndef PRODUCT
  8537   assert(_num_par_pushes >= n, "Too many pops?");
  8538   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
  8539 #endif
  8540   return true;
  8543 // Single-threaded
  8544 void CMSCollector::push_on_overflow_list(oop p) {
  8545   NOT_PRODUCT(_num_par_pushes++;)
  8546   assert(p->is_oop(), "Not an oop");
  8547   preserve_mark_if_necessary(p);
  8548   p->set_mark((markOop)_overflow_list);
  8549   _overflow_list = p;
  8552 // Multi-threaded; use CAS to prepend to overflow list
  8553 void CMSCollector::par_push_on_overflow_list(oop p) {
  8554   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
  8555   assert(p->is_oop(), "Not an oop");
  8556   par_preserve_mark_if_necessary(p);
  8557   oop observed_overflow_list = _overflow_list;
  8558   oop cur_overflow_list;
  8559   do {
  8560     cur_overflow_list = observed_overflow_list;
  8561     p->set_mark(markOop(cur_overflow_list));
  8562     observed_overflow_list =
  8563       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
  8564   } while (cur_overflow_list != observed_overflow_list);
  8567 // Single threaded
  8568 // General Note on GrowableArray: pushes may silently fail
  8569 // because we are (temporarily) out of C-heap for expanding
  8570 // the stack. The problem is quite ubiquitous and affects
  8571 // a lot of code in the JVM. The prudent thing for GrowableArray
  8572 // to do (for now) is to exit with an error. However, that may
  8573 // be too draconian in some cases because the caller may be
  8574 // able to recover without much harm. For suych cases, we
  8575 // should probably introduce a "soft_push" method which returns
  8576 // an indication of success or failure with the assumption that
  8577 // the caller may be able to recover from a failure; code in
  8578 // the VM can then be changed, incrementally, to deal with such
  8579 // failures where possible, thus, incrementally hardening the VM
  8580 // in such low resource situations.
  8581 void CMSCollector::preserve_mark_work(oop p, markOop m) {
  8582   int PreserveMarkStackSize = 128;
  8584   if (_preserved_oop_stack == NULL) {
  8585     assert(_preserved_mark_stack == NULL,
  8586            "bijection with preserved_oop_stack");
  8587     // Allocate the stacks
  8588     _preserved_oop_stack  = new (ResourceObj::C_HEAP)
  8589       GrowableArray<oop>(PreserveMarkStackSize, true);
  8590     _preserved_mark_stack = new (ResourceObj::C_HEAP)
  8591       GrowableArray<markOop>(PreserveMarkStackSize, true);
  8592     if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
  8593       vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
  8594                             "Preserved Mark/Oop Stack for CMS (C-heap)");
  8597   _preserved_oop_stack->push(p);
  8598   _preserved_mark_stack->push(m);
  8599   assert(m == p->mark(), "Mark word changed");
  8600   assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
  8601          "bijection");
  8604 // Single threaded
  8605 void CMSCollector::preserve_mark_if_necessary(oop p) {
  8606   markOop m = p->mark();
  8607   if (m->must_be_preserved(p)) {
  8608     preserve_mark_work(p, m);
  8612 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
  8613   markOop m = p->mark();
  8614   if (m->must_be_preserved(p)) {
  8615     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
  8616     // Even though we read the mark word without holding
  8617     // the lock, we are assured that it will not change
  8618     // because we "own" this oop, so no other thread can
  8619     // be trying to push it on the overflow list; see
  8620     // the assertion in preserve_mark_work() that checks
  8621     // that m == p->mark().
  8622     preserve_mark_work(p, m);
  8626 // We should be able to do this multi-threaded,
  8627 // a chunk of stack being a task (this is
  8628 // correct because each oop only ever appears
  8629 // once in the overflow list. However, it's
  8630 // not very easy to completely overlap this with
  8631 // other operations, so will generally not be done
  8632 // until all work's been completed. Because we
  8633 // expect the preserved oop stack (set) to be small,
  8634 // it's probably fine to do this single-threaded.
  8635 // We can explore cleverer concurrent/overlapped/parallel
  8636 // processing of preserved marks if we feel the
  8637 // need for this in the future. Stack overflow should
  8638 // be so rare in practice and, when it happens, its
  8639 // effect on performance so great that this will
  8640 // likely just be in the noise anyway.
  8641 void CMSCollector::restore_preserved_marks_if_any() {
  8642   if (_preserved_oop_stack == NULL) {
  8643     assert(_preserved_mark_stack == NULL,
  8644            "bijection with preserved_oop_stack");
  8645     return;
  8648   assert(SafepointSynchronize::is_at_safepoint(),
  8649          "world should be stopped");
  8650   assert(Thread::current()->is_ConcurrentGC_thread() ||
  8651          Thread::current()->is_VM_thread(),
  8652          "should be single-threaded");
  8654   int length = _preserved_oop_stack->length();
  8655   assert(_preserved_mark_stack->length() == length, "bijection");
  8656   for (int i = 0; i < length; i++) {
  8657     oop p = _preserved_oop_stack->at(i);
  8658     assert(p->is_oop(), "Should be an oop");
  8659     assert(_span.contains(p), "oop should be in _span");
  8660     assert(p->mark() == markOopDesc::prototype(),
  8661            "Set when taken from overflow list");
  8662     markOop m = _preserved_mark_stack->at(i);
  8663     p->set_mark(m);
  8665   _preserved_mark_stack->clear();
  8666   _preserved_oop_stack->clear();
  8667   assert(_preserved_mark_stack->is_empty() &&
  8668          _preserved_oop_stack->is_empty(),
  8669          "stacks were cleared above");
  8672 #ifndef PRODUCT
  8673 bool CMSCollector::no_preserved_marks() const {
  8674   return (   (   _preserved_mark_stack == NULL
  8675               && _preserved_oop_stack == NULL)
  8676           || (   _preserved_mark_stack->is_empty()
  8677               && _preserved_oop_stack->is_empty()));
  8679 #endif
  8681 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
  8683   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
  8684   CMSAdaptiveSizePolicy* size_policy =
  8685     (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
  8686   assert(size_policy->is_gc_cms_adaptive_size_policy(),
  8687     "Wrong type for size policy");
  8688   return size_policy;
  8691 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
  8692                                            size_t desired_promo_size) {
  8693   if (cur_promo_size < desired_promo_size) {
  8694     size_t expand_bytes = desired_promo_size - cur_promo_size;
  8695     if (PrintAdaptiveSizePolicy && Verbose) {
  8696       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  8697         "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
  8698         expand_bytes);
  8700     expand(expand_bytes,
  8701            MinHeapDeltaBytes,
  8702            CMSExpansionCause::_adaptive_size_policy);
  8703   } else if (desired_promo_size < cur_promo_size) {
  8704     size_t shrink_bytes = cur_promo_size - desired_promo_size;
  8705     if (PrintAdaptiveSizePolicy && Verbose) {
  8706       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  8707         "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
  8708         shrink_bytes);
  8710     shrink(shrink_bytes);
  8714 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
  8715   GenCollectedHeap* gch = GenCollectedHeap::heap();
  8716   CMSGCAdaptivePolicyCounters* counters =
  8717     (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
  8718   assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
  8719     "Wrong kind of counters");
  8720   return counters;
  8724 void ASConcurrentMarkSweepGeneration::update_counters() {
  8725   if (UsePerfData) {
  8726     _space_counters->update_all();
  8727     _gen_counters->update_all();
  8728     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  8729     GenCollectedHeap* gch = GenCollectedHeap::heap();
  8730     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  8731     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  8732       "Wrong gc statistics type");
  8733     counters->update_counters(gc_stats_l);
  8737 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
  8738   if (UsePerfData) {
  8739     _space_counters->update_used(used);
  8740     _space_counters->update_capacity();
  8741     _gen_counters->update_all();
  8743     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  8744     GenCollectedHeap* gch = GenCollectedHeap::heap();
  8745     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  8746     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  8747       "Wrong gc statistics type");
  8748     counters->update_counters(gc_stats_l);
  8752 // The desired expansion delta is computed so that:
  8753 // . desired free percentage or greater is used
  8754 void ASConcurrentMarkSweepGeneration::compute_new_size() {
  8755   assert_locked_or_safepoint(Heap_lock);
  8757   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
  8759   // If incremental collection failed, we just want to expand
  8760   // to the limit.
  8761   if (incremental_collection_failed()) {
  8762     clear_incremental_collection_failed();
  8763     grow_to_reserved();
  8764     return;
  8767   assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
  8769   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
  8770     "Wrong type of heap");
  8771   int prev_level = level() - 1;
  8772   assert(prev_level >= 0, "The cms generation is the lowest generation");
  8773   Generation* prev_gen = gch->get_gen(prev_level);
  8774   assert(prev_gen->kind() == Generation::ASParNew,
  8775     "Wrong type of young generation");
  8776   ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
  8777   size_t cur_eden = younger_gen->eden()->capacity();
  8778   CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
  8779   size_t cur_promo = free();
  8780   size_policy->compute_tenured_generation_free_space(cur_promo,
  8781                                                        max_available(),
  8782                                                        cur_eden);
  8783   resize(cur_promo, size_policy->promo_size());
  8785   // Record the new size of the space in the cms generation
  8786   // that is available for promotions.  This is temporary.
  8787   // It should be the desired promo size.
  8788   size_policy->avg_cms_promo()->sample(free());
  8789   size_policy->avg_old_live()->sample(used());
  8791   if (UsePerfData) {
  8792     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  8793     counters->update_cms_capacity_counter(capacity());
  8797 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
  8798   assert_locked_or_safepoint(Heap_lock);
  8799   assert_lock_strong(freelistLock());
  8800   HeapWord* old_end = _cmsSpace->end();
  8801   HeapWord* unallocated_start = _cmsSpace->unallocated_block();
  8802   assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
  8803   FreeChunk* chunk_at_end = find_chunk_at_end();
  8804   if (chunk_at_end == NULL) {
  8805     // No room to shrink
  8806     if (PrintGCDetails && Verbose) {
  8807       gclog_or_tty->print_cr("No room to shrink: old_end  "
  8808         PTR_FORMAT "  unallocated_start  " PTR_FORMAT
  8809         " chunk_at_end  " PTR_FORMAT,
  8810         old_end, unallocated_start, chunk_at_end);
  8812     return;
  8813   } else {
  8815     // Find the chunk at the end of the space and determine
  8816     // how much it can be shrunk.
  8817     size_t shrinkable_size_in_bytes = chunk_at_end->size();
  8818     size_t aligned_shrinkable_size_in_bytes =
  8819       align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
  8820     assert(unallocated_start <= chunk_at_end->end(),
  8821       "Inconsistent chunk at end of space");
  8822     size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
  8823     size_t word_size_before = heap_word_size(_virtual_space.committed_size());
  8825     // Shrink the underlying space
  8826     _virtual_space.shrink_by(bytes);
  8827     if (PrintGCDetails && Verbose) {
  8828       gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
  8829         " desired_bytes " SIZE_FORMAT
  8830         " shrinkable_size_in_bytes " SIZE_FORMAT
  8831         " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
  8832         "  bytes  " SIZE_FORMAT,
  8833         desired_bytes, shrinkable_size_in_bytes,
  8834         aligned_shrinkable_size_in_bytes, bytes);
  8835       gclog_or_tty->print_cr("          old_end  " SIZE_FORMAT
  8836         "  unallocated_start  " SIZE_FORMAT,
  8837         old_end, unallocated_start);
  8840     // If the space did shrink (shrinking is not guaranteed),
  8841     // shrink the chunk at the end by the appropriate amount.
  8842     if (((HeapWord*)_virtual_space.high()) < old_end) {
  8843       size_t new_word_size =
  8844         heap_word_size(_virtual_space.committed_size());
  8846       // Have to remove the chunk from the dictionary because it is changing
  8847       // size and might be someplace elsewhere in the dictionary.
  8849       // Get the chunk at end, shrink it, and put it
  8850       // back.
  8851       _cmsSpace->removeChunkFromDictionary(chunk_at_end);
  8852       size_t word_size_change = word_size_before - new_word_size;
  8853       size_t chunk_at_end_old_size = chunk_at_end->size();
  8854       assert(chunk_at_end_old_size >= word_size_change,
  8855         "Shrink is too large");
  8856       chunk_at_end->setSize(chunk_at_end_old_size -
  8857                           word_size_change);
  8858       _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
  8859         word_size_change);
  8861       _cmsSpace->returnChunkToDictionary(chunk_at_end);
  8863       MemRegion mr(_cmsSpace->bottom(), new_word_size);
  8864       _bts->resize(new_word_size);  // resize the block offset shared array
  8865       Universe::heap()->barrier_set()->resize_covered_region(mr);
  8866       _cmsSpace->assert_locked();
  8867       _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  8869       NOT_PRODUCT(_cmsSpace->dictionary()->verify());
  8871       // update the space and generation capacity counters
  8872       if (UsePerfData) {
  8873         _space_counters->update_capacity();
  8874         _gen_counters->update_all();
  8877       if (Verbose && PrintGCDetails) {
  8878         size_t new_mem_size = _virtual_space.committed_size();
  8879         size_t old_mem_size = new_mem_size + bytes;
  8880         gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
  8881                       name(), old_mem_size/K, bytes/K, new_mem_size/K);
  8885     assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
  8886       "Inconsistency at end of space");
  8887     assert(chunk_at_end->end() == _cmsSpace->end(),
  8888       "Shrinking is inconsistent");
  8889     return;
  8893 // Transfer some number of overflown objects to usual marking
  8894 // stack. Return true if some objects were transferred.
  8895 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
  8896   size_t num = MIN2((size_t)_mark_stack->capacity()/4,
  8897                     (size_t)ParGCDesiredObjsFromOverflowList);
  8899   bool res = _collector->take_from_overflow_list(num, _mark_stack);
  8900   assert(_collector->overflow_list_is_empty() || res,
  8901          "If list is not empty, we should have taken something");
  8902   assert(!res || !_mark_stack->isEmpty(),
  8903          "If we took something, it should now be on our stack");
  8904   return res;
  8907 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
  8908   size_t res = _sp->block_size_no_stall(addr, _collector);
  8909   assert(res != 0, "Should always be able to compute a size");
  8910   if (_sp->block_is_obj(addr)) {
  8911     if (_live_bit_map->isMarked(addr)) {
  8912       // It can't have been dead in a previous cycle
  8913       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
  8914     } else {
  8915       _dead_bit_map->mark(addr);      // mark the dead object
  8918   return res;

mercurial