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

Sat, 23 Oct 2010 23:03:49 -0700

author
ysr
date
Sat, 23 Oct 2010 23:03:49 -0700
changeset 2243
a7214d79fcf1
parent 2192
c99c53f07c14
child 2314
f95d63e2154a
permissions
-rw-r--r--

6896603: CMS/GCH: collection_attempt_is_safe() ergo should use more recent data
Summary: Deprecated HandlePromotionFailure, removing the ability to turn off that feature, did away with one epoch look-ahead when deciding if a scavenge is likely to fail, relying on current data.
Reviewed-by: jmasa, johnc, poonam

     1 /*
     2  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    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)(CollectedHeap::min_fill_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 (CollectedHeap::use_parallel_gc_threads()) {
   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 >= CollectedHeap::min_fill_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 }
   256 void ConcurrentMarkSweepGeneration::ref_processor_init() {
   257   assert(collector() != NULL, "no collector");
   258   collector()->ref_processor_init();
   259 }
   261 void CMSCollector::ref_processor_init() {
   262   if (_ref_processor == NULL) {
   263     // Allocate and initialize a reference processor
   264     _ref_processor = ReferenceProcessor::create_ref_processor(
   265         _span,                               // span
   266         _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
   267         _cmsGen->refs_discovery_is_mt(),     // mt_discovery
   268         &_is_alive_closure,
   269         ParallelGCThreads,
   270         ParallelRefProcEnabled);
   271     // Initialize the _ref_processor field of CMSGen
   272     _cmsGen->set_ref_processor(_ref_processor);
   274     // Allocate a dummy ref processor for perm gen.
   275     ReferenceProcessor* rp2 = new ReferenceProcessor();
   276     if (rp2 == NULL) {
   277       vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
   278     }
   279     _permGen->set_ref_processor(rp2);
   280   }
   281 }
   283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
   284   GenCollectedHeap* gch = GenCollectedHeap::heap();
   285   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
   286     "Wrong type of heap");
   287   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
   288     gch->gen_policy()->size_policy();
   289   assert(sp->is_gc_cms_adaptive_size_policy(),
   290     "Wrong type of size policy");
   291   return sp;
   292 }
   294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
   295   CMSGCAdaptivePolicyCounters* results =
   296     (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
   297   assert(
   298     results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
   299     "Wrong gc policy counter kind");
   300   return results;
   301 }
   304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
   306   const char* gen_name = "old";
   308   // Generation Counters - generation 1, 1 subspace
   309   _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
   311   _space_counters = new GSpaceCounters(gen_name, 0,
   312                                        _virtual_space.reserved_size(),
   313                                        this, _gen_counters);
   314 }
   316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
   317   _cms_gen(cms_gen)
   318 {
   319   assert(alpha <= 100, "bad value");
   320   _saved_alpha = alpha;
   322   // Initialize the alphas to the bootstrap value of 100.
   323   _gc0_alpha = _cms_alpha = 100;
   325   _cms_begin_time.update();
   326   _cms_end_time.update();
   328   _gc0_duration = 0.0;
   329   _gc0_period = 0.0;
   330   _gc0_promoted = 0;
   332   _cms_duration = 0.0;
   333   _cms_period = 0.0;
   334   _cms_allocated = 0;
   336   _cms_used_at_gc0_begin = 0;
   337   _cms_used_at_gc0_end = 0;
   338   _allow_duty_cycle_reduction = false;
   339   _valid_bits = 0;
   340   _icms_duty_cycle = CMSIncrementalDutyCycle;
   341 }
   343 double CMSStats::cms_free_adjustment_factor(size_t free) const {
   344   // TBD: CR 6909490
   345   return 1.0;
   346 }
   348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
   349 }
   351 // If promotion failure handling is on use
   352 // the padded average size of the promotion for each
   353 // young generation collection.
   354 double CMSStats::time_until_cms_gen_full() const {
   355   size_t cms_free = _cms_gen->cmsSpace()->free();
   356   GenCollectedHeap* gch = GenCollectedHeap::heap();
   357   size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
   358                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
   359   if (cms_free > expected_promotion) {
   360     // Start a cms collection if there isn't enough space to promote
   361     // for the next minor collection.  Use the padded average as
   362     // a safety factor.
   363     cms_free -= expected_promotion;
   365     // Adjust by the safety factor.
   366     double cms_free_dbl = (double)cms_free;
   367     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
   368     // Apply a further correction factor which tries to adjust
   369     // for recent occurance of concurrent mode failures.
   370     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
   371     cms_free_dbl = cms_free_dbl * cms_adjustment;
   373     if (PrintGCDetails && Verbose) {
   374       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
   375         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
   376         cms_free, expected_promotion);
   377       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
   378         cms_free_dbl, cms_consumption_rate() + 1.0);
   379     }
   380     // Add 1 in case the consumption rate goes to zero.
   381     return cms_free_dbl / (cms_consumption_rate() + 1.0);
   382   }
   383   return 0.0;
   384 }
   386 // Compare the duration of the cms collection to the
   387 // time remaining before the cms generation is empty.
   388 // Note that the time from the start of the cms collection
   389 // to the start of the cms sweep (less than the total
   390 // duration of the cms collection) can be used.  This
   391 // has been tried and some applications experienced
   392 // promotion failures early in execution.  This was
   393 // possibly because the averages were not accurate
   394 // enough at the beginning.
   395 double CMSStats::time_until_cms_start() const {
   396   // We add "gc0_period" to the "work" calculation
   397   // below because this query is done (mostly) at the
   398   // end of a scavenge, so we need to conservatively
   399   // account for that much possible delay
   400   // in the query so as to avoid concurrent mode failures
   401   // due to starting the collection just a wee bit too
   402   // late.
   403   double work = cms_duration() + gc0_period();
   404   double deadline = time_until_cms_gen_full();
   405   // If a concurrent mode failure occurred recently, we want to be
   406   // more conservative and halve our expected time_until_cms_gen_full()
   407   if (work > deadline) {
   408     if (Verbose && PrintGCDetails) {
   409       gclog_or_tty->print(
   410         " CMSCollector: collect because of anticipated promotion "
   411         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
   412         gc0_period(), time_until_cms_gen_full());
   413     }
   414     return 0.0;
   415   }
   416   return work - deadline;
   417 }
   419 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
   420 // amount of change to prevent wild oscillation.
   421 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
   422                                               unsigned int new_duty_cycle) {
   423   assert(old_duty_cycle <= 100, "bad input value");
   424   assert(new_duty_cycle <= 100, "bad input value");
   426   // Note:  use subtraction with caution since it may underflow (values are
   427   // unsigned).  Addition is safe since we're in the range 0-100.
   428   unsigned int damped_duty_cycle = new_duty_cycle;
   429   if (new_duty_cycle < old_duty_cycle) {
   430     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
   431     if (new_duty_cycle + largest_delta < old_duty_cycle) {
   432       damped_duty_cycle = old_duty_cycle - largest_delta;
   433     }
   434   } else if (new_duty_cycle > old_duty_cycle) {
   435     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
   436     if (new_duty_cycle > old_duty_cycle + largest_delta) {
   437       damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
   438     }
   439   }
   440   assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
   442   if (CMSTraceIncrementalPacing) {
   443     gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
   444                            old_duty_cycle, new_duty_cycle, damped_duty_cycle);
   445   }
   446   return damped_duty_cycle;
   447 }
   449 unsigned int CMSStats::icms_update_duty_cycle_impl() {
   450   assert(CMSIncrementalPacing && valid(),
   451          "should be handled in icms_update_duty_cycle()");
   453   double cms_time_so_far = cms_timer().seconds();
   454   double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
   455   double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
   457   // Avoid division by 0.
   458   double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
   459   double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
   461   unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
   462   if (new_duty_cycle > _icms_duty_cycle) {
   463     // Avoid very small duty cycles (1 or 2); 0 is allowed.
   464     if (new_duty_cycle > 2) {
   465       _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
   466                                                 new_duty_cycle);
   467     }
   468   } else if (_allow_duty_cycle_reduction) {
   469     // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
   470     new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
   471     // Respect the minimum duty cycle.
   472     unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
   473     _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
   474   }
   476   if (PrintGCDetails || CMSTraceIncrementalPacing) {
   477     gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
   478   }
   480   _allow_duty_cycle_reduction = false;
   481   return _icms_duty_cycle;
   482 }
   484 #ifndef PRODUCT
   485 void CMSStats::print_on(outputStream *st) const {
   486   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
   487   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
   488                gc0_duration(), gc0_period(), gc0_promoted());
   489   st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
   490             cms_duration(), cms_duration_per_mb(),
   491             cms_period(), cms_allocated());
   492   st->print(",cms_since_beg=%g,cms_since_end=%g",
   493             cms_time_since_begin(), cms_time_since_end());
   494   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
   495             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
   496   if (CMSIncrementalMode) {
   497     st->print(",dc=%d", icms_duty_cycle());
   498   }
   500   if (valid()) {
   501     st->print(",promo_rate=%g,cms_alloc_rate=%g",
   502               promotion_rate(), cms_allocation_rate());
   503     st->print(",cms_consumption_rate=%g,time_until_full=%g",
   504               cms_consumption_rate(), time_until_cms_gen_full());
   505   }
   506   st->print(" ");
   507 }
   508 #endif // #ifndef PRODUCT
   510 CMSCollector::CollectorState CMSCollector::_collectorState =
   511                              CMSCollector::Idling;
   512 bool CMSCollector::_foregroundGCIsActive = false;
   513 bool CMSCollector::_foregroundGCShouldWait = false;
   515 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
   516                            ConcurrentMarkSweepGeneration* permGen,
   517                            CardTableRS*                   ct,
   518                            ConcurrentMarkSweepPolicy*     cp):
   519   _cmsGen(cmsGen),
   520   _permGen(permGen),
   521   _ct(ct),
   522   _ref_processor(NULL),    // will be set later
   523   _conc_workers(NULL),     // may be set later
   524   _abort_preclean(false),
   525   _start_sampling(false),
   526   _between_prologue_and_epilogue(false),
   527   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
   528   _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
   529   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
   530                  -1 /* lock-free */, "No_lock" /* dummy */),
   531   _modUnionClosure(&_modUnionTable),
   532   _modUnionClosurePar(&_modUnionTable),
   533   // Adjust my span to cover old (cms) gen and perm gen
   534   _span(cmsGen->reserved()._union(permGen->reserved())),
   535   // Construct the is_alive_closure with _span & markBitMap
   536   _is_alive_closure(_span, &_markBitMap),
   537   _restart_addr(NULL),
   538   _overflow_list(NULL),
   539   _stats(cmsGen),
   540   _eden_chunk_array(NULL),     // may be set in ctor body
   541   _eden_chunk_capacity(0),     // -- ditto --
   542   _eden_chunk_index(0),        // -- ditto --
   543   _survivor_plab_array(NULL),  // -- ditto --
   544   _survivor_chunk_array(NULL), // -- ditto --
   545   _survivor_chunk_capacity(0), // -- ditto --
   546   _survivor_chunk_index(0),    // -- ditto --
   547   _ser_pmc_preclean_ovflw(0),
   548   _ser_kac_preclean_ovflw(0),
   549   _ser_pmc_remark_ovflw(0),
   550   _par_pmc_remark_ovflw(0),
   551   _ser_kac_ovflw(0),
   552   _par_kac_ovflw(0),
   553 #ifndef PRODUCT
   554   _num_par_pushes(0),
   555 #endif
   556   _collection_count_start(0),
   557   _verifying(false),
   558   _icms_start_limit(NULL),
   559   _icms_stop_limit(NULL),
   560   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
   561   _completed_initialization(false),
   562   _collector_policy(cp),
   563   _should_unload_classes(false),
   564   _concurrent_cycles_since_last_unload(0),
   565   _roots_scanning_options(0),
   566   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
   567   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
   568 {
   569   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
   570     ExplicitGCInvokesConcurrent = true;
   571   }
   572   // Now expand the span and allocate the collection support structures
   573   // (MUT, marking bit map etc.) to cover both generations subject to
   574   // collection.
   576   // First check that _permGen is adjacent to _cmsGen and above it.
   577   assert(   _cmsGen->reserved().word_size()  > 0
   578          && _permGen->reserved().word_size() > 0,
   579          "generations should not be of zero size");
   580   assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
   581          "_cmsGen and _permGen should not overlap");
   582   assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
   583          "_cmsGen->end() different from _permGen->start()");
   585   // For use by dirty card to oop closures.
   586   _cmsGen->cmsSpace()->set_collector(this);
   587   _permGen->cmsSpace()->set_collector(this);
   589   // Allocate MUT and marking bit map
   590   {
   591     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
   592     if (!_markBitMap.allocate(_span)) {
   593       warning("Failed to allocate CMS Bit Map");
   594       return;
   595     }
   596     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
   597   }
   598   {
   599     _modUnionTable.allocate(_span);
   600     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
   601   }
   603   if (!_markStack.allocate(MarkStackSize)) {
   604     warning("Failed to allocate CMS Marking Stack");
   605     return;
   606   }
   607   if (!_revisitStack.allocate(CMSRevisitStackSize)) {
   608     warning("Failed to allocate CMS Revisit Stack");
   609     return;
   610   }
   612   // Support for multi-threaded concurrent phases
   613   if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) {
   614     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
   615       // just for now
   616       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
   617     }
   618     if (ConcGCThreads > 1) {
   619       _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
   620                                  ConcGCThreads, true);
   621       if (_conc_workers == NULL) {
   622         warning("GC/CMS: _conc_workers allocation failure: "
   623               "forcing -CMSConcurrentMTEnabled");
   624         CMSConcurrentMTEnabled = false;
   625       } else {
   626         _conc_workers->initialize_workers();
   627       }
   628     } else {
   629       CMSConcurrentMTEnabled = false;
   630     }
   631   }
   632   if (!CMSConcurrentMTEnabled) {
   633     ConcGCThreads = 0;
   634   } else {
   635     // Turn off CMSCleanOnEnter optimization temporarily for
   636     // the MT case where it's not fixed yet; see 6178663.
   637     CMSCleanOnEnter = false;
   638   }
   639   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
   640          "Inconsistency");
   642   // Parallel task queues; these are shared for the
   643   // concurrent and stop-world phases of CMS, but
   644   // are not shared with parallel scavenge (ParNew).
   645   {
   646     uint i;
   647     uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
   649     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
   650          || ParallelRefProcEnabled)
   651         && num_queues > 0) {
   652       _task_queues = new OopTaskQueueSet(num_queues);
   653       if (_task_queues == NULL) {
   654         warning("task_queues allocation failure.");
   655         return;
   656       }
   657       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
   658       if (_hash_seed == NULL) {
   659         warning("_hash_seed array allocation failure");
   660         return;
   661       }
   663       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
   664       for (i = 0; i < num_queues; i++) {
   665         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
   666         if (q == NULL) {
   667           warning("work_queue allocation failure.");
   668           return;
   669         }
   670         _task_queues->register_queue(i, q);
   671       }
   672       for (i = 0; i < num_queues; i++) {
   673         _task_queues->queue(i)->initialize();
   674         _hash_seed[i] = 17;  // copied from ParNew
   675       }
   676     }
   677   }
   679   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
   680   _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
   682   // Clip CMSBootstrapOccupancy between 0 and 100.
   683   _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
   684                          /(double)100;
   686   _full_gcs_since_conc_gc = 0;
   688   // Now tell CMS generations the identity of their collector
   689   ConcurrentMarkSweepGeneration::set_collector(this);
   691   // Create & start a CMS thread for this CMS collector
   692   _cmsThread = ConcurrentMarkSweepThread::start(this);
   693   assert(cmsThread() != NULL, "CMS Thread should have been created");
   694   assert(cmsThread()->collector() == this,
   695          "CMS Thread should refer to this gen");
   696   assert(CGC_lock != NULL, "Where's the CGC_lock?");
   698   // Support for parallelizing young gen rescan
   699   GenCollectedHeap* gch = GenCollectedHeap::heap();
   700   _young_gen = gch->prev_gen(_cmsGen);
   701   if (gch->supports_inline_contig_alloc()) {
   702     _top_addr = gch->top_addr();
   703     _end_addr = gch->end_addr();
   704     assert(_young_gen != NULL, "no _young_gen");
   705     _eden_chunk_index = 0;
   706     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
   707     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
   708     if (_eden_chunk_array == NULL) {
   709       _eden_chunk_capacity = 0;
   710       warning("GC/CMS: _eden_chunk_array allocation failure");
   711     }
   712   }
   713   assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
   715   // Support for parallelizing survivor space rescan
   716   if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
   717     const size_t max_plab_samples =
   718       ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
   720     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
   721     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
   722     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
   723     if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
   724         || _cursor == NULL) {
   725       warning("Failed to allocate survivor plab/chunk array");
   726       if (_survivor_plab_array  != NULL) {
   727         FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
   728         _survivor_plab_array = NULL;
   729       }
   730       if (_survivor_chunk_array != NULL) {
   731         FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
   732         _survivor_chunk_array = NULL;
   733       }
   734       if (_cursor != NULL) {
   735         FREE_C_HEAP_ARRAY(size_t, _cursor);
   736         _cursor = NULL;
   737       }
   738     } else {
   739       _survivor_chunk_capacity = 2*max_plab_samples;
   740       for (uint i = 0; i < ParallelGCThreads; i++) {
   741         HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
   742         if (vec == NULL) {
   743           warning("Failed to allocate survivor plab array");
   744           for (int j = i; j > 0; j--) {
   745             FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
   746           }
   747           FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
   748           FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
   749           _survivor_plab_array = NULL;
   750           _survivor_chunk_array = NULL;
   751           _survivor_chunk_capacity = 0;
   752           break;
   753         } else {
   754           ChunkArray* cur =
   755             ::new (&_survivor_plab_array[i]) ChunkArray(vec,
   756                                                         max_plab_samples);
   757           assert(cur->end() == 0, "Should be 0");
   758           assert(cur->array() == vec, "Should be vec");
   759           assert(cur->capacity() == max_plab_samples, "Error");
   760         }
   761       }
   762     }
   763   }
   764   assert(   (   _survivor_plab_array  != NULL
   765              && _survivor_chunk_array != NULL)
   766          || (   _survivor_chunk_capacity == 0
   767              && _survivor_chunk_index == 0),
   768          "Error");
   770   // Choose what strong roots should be scanned depending on verification options
   771   // and perm gen collection mode.
   772   if (!CMSClassUnloadingEnabled) {
   773     // If class unloading is disabled we want to include all classes into the root set.
   774     add_root_scanning_option(SharedHeap::SO_AllClasses);
   775   } else {
   776     add_root_scanning_option(SharedHeap::SO_SystemClasses);
   777   }
   779   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
   780   _gc_counters = new CollectorCounters("CMS", 1);
   781   _completed_initialization = true;
   782   _inter_sweep_timer.start();  // start of time
   783 #ifdef SPARC
   784   // Issue a stern warning, but allow use for experimentation and debugging.
   785   if (VM_Version::is_sun4v() && UseMemSetInBOT) {
   786     assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
   787     warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
   788             " on sun4v; please understand that you are using at your own risk!");
   789   }
   790 #endif
   791 }
   793 const char* ConcurrentMarkSweepGeneration::name() const {
   794   return "concurrent mark-sweep generation";
   795 }
   796 void ConcurrentMarkSweepGeneration::update_counters() {
   797   if (UsePerfData) {
   798     _space_counters->update_all();
   799     _gen_counters->update_all();
   800   }
   801 }
   803 // this is an optimized version of update_counters(). it takes the
   804 // used value as a parameter rather than computing it.
   805 //
   806 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
   807   if (UsePerfData) {
   808     _space_counters->update_used(used);
   809     _space_counters->update_capacity();
   810     _gen_counters->update_all();
   811   }
   812 }
   814 void ConcurrentMarkSweepGeneration::print() const {
   815   Generation::print();
   816   cmsSpace()->print();
   817 }
   819 #ifndef PRODUCT
   820 void ConcurrentMarkSweepGeneration::print_statistics() {
   821   cmsSpace()->printFLCensus(0);
   822 }
   823 #endif
   825 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
   826   GenCollectedHeap* gch = GenCollectedHeap::heap();
   827   if (PrintGCDetails) {
   828     if (Verbose) {
   829       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
   830         level(), short_name(), s, used(), capacity());
   831     } else {
   832       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
   833         level(), short_name(), s, used() / K, capacity() / K);
   834     }
   835   }
   836   if (Verbose) {
   837     gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
   838               gch->used(), gch->capacity());
   839   } else {
   840     gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
   841               gch->used() / K, gch->capacity() / K);
   842   }
   843 }
   845 size_t
   846 ConcurrentMarkSweepGeneration::contiguous_available() const {
   847   // dld proposes an improvement in precision here. If the committed
   848   // part of the space ends in a free block we should add that to
   849   // uncommitted size in the calculation below. Will make this
   850   // change later, staying with the approximation below for the
   851   // time being. -- ysr.
   852   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
   853 }
   855 size_t
   856 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
   857   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
   858 }
   860 size_t ConcurrentMarkSweepGeneration::max_available() const {
   861   return free() + _virtual_space.uncommitted_size();
   862 }
   864 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
   865   size_t available = max_available();
   866   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
   867   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
   868   if (PrintGC && Verbose) {
   869     gclog_or_tty->print_cr(
   870       "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
   871       "max_promo("SIZE_FORMAT")",
   872       res? "":" not", available, res? ">=":"<",
   873       av_promo, max_promotion_in_bytes);
   874   }
   875   return res;
   876 }
   878 // At a promotion failure dump information on block layout in heap
   879 // (cms old generation).
   880 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
   881   if (CMSDumpAtPromotionFailure) {
   882     cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
   883   }
   884 }
   886 CompactibleSpace*
   887 ConcurrentMarkSweepGeneration::first_compaction_space() const {
   888   return _cmsSpace;
   889 }
   891 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
   892   // Clear the promotion information.  These pointers can be adjusted
   893   // along with all the other pointers into the heap but
   894   // compaction is expected to be a rare event with
   895   // a heap using cms so don't do it without seeing the need.
   896   if (CollectedHeap::use_parallel_gc_threads()) {
   897     for (uint i = 0; i < ParallelGCThreads; i++) {
   898       _par_gc_thread_states[i]->promo.reset();
   899     }
   900   }
   901 }
   903 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
   904   blk->do_space(_cmsSpace);
   905 }
   907 void ConcurrentMarkSweepGeneration::compute_new_size() {
   908   assert_locked_or_safepoint(Heap_lock);
   910   // If incremental collection failed, we just want to expand
   911   // to the limit.
   912   if (incremental_collection_failed()) {
   913     clear_incremental_collection_failed();
   914     grow_to_reserved();
   915     return;
   916   }
   918   size_t expand_bytes = 0;
   919   double free_percentage = ((double) free()) / capacity();
   920   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
   921   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
   923   // compute expansion delta needed for reaching desired free percentage
   924   if (free_percentage < desired_free_percentage) {
   925     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   926     assert(desired_capacity >= capacity(), "invalid expansion size");
   927     expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
   928   }
   929   if (expand_bytes > 0) {
   930     if (PrintGCDetails && Verbose) {
   931       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   932       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
   933       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
   934       gclog_or_tty->print_cr("  Desired free fraction %f",
   935         desired_free_percentage);
   936       gclog_or_tty->print_cr("  Maximum free fraction %f",
   937         maximum_free_percentage);
   938       gclog_or_tty->print_cr("  Capactiy "SIZE_FORMAT, capacity()/1000);
   939       gclog_or_tty->print_cr("  Desired capacity "SIZE_FORMAT,
   940         desired_capacity/1000);
   941       int prev_level = level() - 1;
   942       if (prev_level >= 0) {
   943         size_t prev_size = 0;
   944         GenCollectedHeap* gch = GenCollectedHeap::heap();
   945         Generation* prev_gen = gch->_gens[prev_level];
   946         prev_size = prev_gen->capacity();
   947           gclog_or_tty->print_cr("  Younger gen size "SIZE_FORMAT,
   948                                  prev_size/1000);
   949       }
   950       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc "SIZE_FORMAT,
   951         unsafe_max_alloc_nogc()/1000);
   952       gclog_or_tty->print_cr("  contiguous available "SIZE_FORMAT,
   953         contiguous_available()/1000);
   954       gclog_or_tty->print_cr("  Expand by "SIZE_FORMAT" (bytes)",
   955         expand_bytes);
   956     }
   957     // safe if expansion fails
   958     expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
   959     if (PrintGCDetails && Verbose) {
   960       gclog_or_tty->print_cr("  Expanded free fraction %f",
   961         ((double) free()) / capacity());
   962     }
   963   }
   964 }
   966 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
   967   return cmsSpace()->freelistLock();
   968 }
   970 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
   971                                                   bool   tlab) {
   972   CMSSynchronousYieldRequest yr;
   973   MutexLockerEx x(freelistLock(),
   974                   Mutex::_no_safepoint_check_flag);
   975   return have_lock_and_allocate(size, tlab);
   976 }
   978 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
   979                                                   bool   tlab /* ignored */) {
   980   assert_lock_strong(freelistLock());
   981   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
   982   HeapWord* res = cmsSpace()->allocate(adjustedSize);
   983   // Allocate the object live (grey) if the background collector has
   984   // started marking. This is necessary because the marker may
   985   // have passed this address and consequently this object will
   986   // not otherwise be greyed and would be incorrectly swept up.
   987   // Note that if this object contains references, the writing
   988   // of those references will dirty the card containing this object
   989   // allowing the object to be blackened (and its references scanned)
   990   // either during a preclean phase or at the final checkpoint.
   991   if (res != NULL) {
   992     // We may block here with an uninitialized object with
   993     // its mark-bit or P-bits not yet set. Such objects need
   994     // to be safely navigable by block_start().
   995     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
   996     assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size");
   997     collector()->direct_allocated(res, adjustedSize);
   998     _direct_allocated_words += adjustedSize;
   999     // allocation counters
  1000     NOT_PRODUCT(
  1001       _numObjectsAllocated++;
  1002       _numWordsAllocated += (int)adjustedSize;
  1005   return res;
  1008 // In the case of direct allocation by mutators in a generation that
  1009 // is being concurrently collected, the object must be allocated
  1010 // live (grey) if the background collector has started marking.
  1011 // This is necessary because the marker may
  1012 // have passed this address and consequently this object will
  1013 // not otherwise be greyed and would be incorrectly swept up.
  1014 // Note that if this object contains references, the writing
  1015 // of those references will dirty the card containing this object
  1016 // allowing the object to be blackened (and its references scanned)
  1017 // either during a preclean phase or at the final checkpoint.
  1018 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
  1019   assert(_markBitMap.covers(start, size), "Out of bounds");
  1020   if (_collectorState >= Marking) {
  1021     MutexLockerEx y(_markBitMap.lock(),
  1022                     Mutex::_no_safepoint_check_flag);
  1023     // [see comments preceding SweepClosure::do_blk() below for details]
  1024     // 1. need to mark the object as live so it isn't collected
  1025     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
  1026     // 3. need to mark the end of the object so marking, precleaning or sweeping
  1027     //    can skip over uninitialized or unparsable objects. An allocated
  1028     //    object is considered uninitialized for our purposes as long as
  1029     //    its klass word is NULL. (Unparsable objects are those which are
  1030     //    initialized in the sense just described, but whose sizes can still
  1031     //    not be correctly determined. Note that the class of unparsable objects
  1032     //    can only occur in the perm gen. All old gen objects are parsable
  1033     //    as soon as they are initialized.)
  1034     _markBitMap.mark(start);          // object is live
  1035     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
  1036     _markBitMap.mark(start + size - 1);
  1037                                       // mark end of object
  1039   // check that oop looks uninitialized
  1040   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
  1043 void CMSCollector::promoted(bool par, HeapWord* start,
  1044                             bool is_obj_array, size_t obj_size) {
  1045   assert(_markBitMap.covers(start), "Out of bounds");
  1046   // See comment in direct_allocated() about when objects should
  1047   // be allocated live.
  1048   if (_collectorState >= Marking) {
  1049     // we already hold the marking bit map lock, taken in
  1050     // the prologue
  1051     if (par) {
  1052       _markBitMap.par_mark(start);
  1053     } else {
  1054       _markBitMap.mark(start);
  1056     // We don't need to mark the object as uninitialized (as
  1057     // in direct_allocated above) because this is being done with the
  1058     // world stopped and the object will be initialized by the
  1059     // time the marking, precleaning or sweeping get to look at it.
  1060     // But see the code for copying objects into the CMS generation,
  1061     // where we need to ensure that concurrent readers of the
  1062     // block offset table are able to safely navigate a block that
  1063     // is in flux from being free to being allocated (and in
  1064     // transition while being copied into) and subsequently
  1065     // becoming a bona-fide object when the copy/promotion is complete.
  1066     assert(SafepointSynchronize::is_at_safepoint(),
  1067            "expect promotion only at safepoints");
  1069     if (_collectorState < Sweeping) {
  1070       // Mark the appropriate cards in the modUnionTable, so that
  1071       // this object gets scanned before the sweep. If this is
  1072       // not done, CMS generation references in the object might
  1073       // not get marked.
  1074       // For the case of arrays, which are otherwise precisely
  1075       // marked, we need to dirty the entire array, not just its head.
  1076       if (is_obj_array) {
  1077         // The [par_]mark_range() method expects mr.end() below to
  1078         // be aligned to the granularity of a bit's representation
  1079         // in the heap. In the case of the MUT below, that's a
  1080         // card size.
  1081         MemRegion mr(start,
  1082                      (HeapWord*)round_to((intptr_t)(start + obj_size),
  1083                         CardTableModRefBS::card_size /* bytes */));
  1084         if (par) {
  1085           _modUnionTable.par_mark_range(mr);
  1086         } else {
  1087           _modUnionTable.mark_range(mr);
  1089       } else {  // not an obj array; we can just mark the head
  1090         if (par) {
  1091           _modUnionTable.par_mark(start);
  1092         } else {
  1093           _modUnionTable.mark(start);
  1100 static inline size_t percent_of_space(Space* space, HeapWord* addr)
  1102   size_t delta = pointer_delta(addr, space->bottom());
  1103   return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
  1106 void CMSCollector::icms_update_allocation_limits()
  1108   Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
  1109   EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
  1111   const unsigned int duty_cycle = stats().icms_update_duty_cycle();
  1112   if (CMSTraceIncrementalPacing) {
  1113     stats().print();
  1116   assert(duty_cycle <= 100, "invalid duty cycle");
  1117   if (duty_cycle != 0) {
  1118     // The duty_cycle is a percentage between 0 and 100; convert to words and
  1119     // then compute the offset from the endpoints of the space.
  1120     size_t free_words = eden->free() / HeapWordSize;
  1121     double free_words_dbl = (double)free_words;
  1122     size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
  1123     size_t offset_words = (free_words - duty_cycle_words) / 2;
  1125     _icms_start_limit = eden->top() + offset_words;
  1126     _icms_stop_limit = eden->end() - offset_words;
  1128     // The limits may be adjusted (shifted to the right) by
  1129     // CMSIncrementalOffset, to allow the application more mutator time after a
  1130     // young gen gc (when all mutators were stopped) and before CMS starts and
  1131     // takes away one or more cpus.
  1132     if (CMSIncrementalOffset != 0) {
  1133       double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
  1134       size_t adjustment = (size_t)adjustment_dbl;
  1135       HeapWord* tmp_stop = _icms_stop_limit + adjustment;
  1136       if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
  1137         _icms_start_limit += adjustment;
  1138         _icms_stop_limit = tmp_stop;
  1142   if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
  1143     _icms_start_limit = _icms_stop_limit = eden->end();
  1146   // Install the new start limit.
  1147   eden->set_soft_end(_icms_start_limit);
  1149   if (CMSTraceIncrementalMode) {
  1150     gclog_or_tty->print(" icms alloc limits:  "
  1151                            PTR_FORMAT "," PTR_FORMAT
  1152                            " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
  1153                            _icms_start_limit, _icms_stop_limit,
  1154                            percent_of_space(eden, _icms_start_limit),
  1155                            percent_of_space(eden, _icms_stop_limit));
  1156     if (Verbose) {
  1157       gclog_or_tty->print("eden:  ");
  1158       eden->print_on(gclog_or_tty);
  1163 // Any changes here should try to maintain the invariant
  1164 // that if this method is called with _icms_start_limit
  1165 // and _icms_stop_limit both NULL, then it should return NULL
  1166 // and not notify the icms thread.
  1167 HeapWord*
  1168 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
  1169                                        size_t word_size)
  1171   // A start_limit equal to end() means the duty cycle is 0, so treat that as a
  1172   // nop.
  1173   if (CMSIncrementalMode && _icms_start_limit != space->end()) {
  1174     if (top <= _icms_start_limit) {
  1175       if (CMSTraceIncrementalMode) {
  1176         space->print_on(gclog_or_tty);
  1177         gclog_or_tty->stamp();
  1178         gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
  1179                                ", new limit=" PTR_FORMAT
  1180                                " (" SIZE_FORMAT "%%)",
  1181                                top, _icms_stop_limit,
  1182                                percent_of_space(space, _icms_stop_limit));
  1184       ConcurrentMarkSweepThread::start_icms();
  1185       assert(top < _icms_stop_limit, "Tautology");
  1186       if (word_size < pointer_delta(_icms_stop_limit, top)) {
  1187         return _icms_stop_limit;
  1190       // The allocation will cross both the _start and _stop limits, so do the
  1191       // stop notification also and return end().
  1192       if (CMSTraceIncrementalMode) {
  1193         space->print_on(gclog_or_tty);
  1194         gclog_or_tty->stamp();
  1195         gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
  1196                                ", new limit=" PTR_FORMAT
  1197                                " (" SIZE_FORMAT "%%)",
  1198                                top, space->end(),
  1199                                percent_of_space(space, space->end()));
  1201       ConcurrentMarkSweepThread::stop_icms();
  1202       return space->end();
  1205     if (top <= _icms_stop_limit) {
  1206       if (CMSTraceIncrementalMode) {
  1207         space->print_on(gclog_or_tty);
  1208         gclog_or_tty->stamp();
  1209         gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
  1210                                ", new limit=" PTR_FORMAT
  1211                                " (" SIZE_FORMAT "%%)",
  1212                                top, space->end(),
  1213                                percent_of_space(space, space->end()));
  1215       ConcurrentMarkSweepThread::stop_icms();
  1216       return space->end();
  1219     if (CMSTraceIncrementalMode) {
  1220       space->print_on(gclog_or_tty);
  1221       gclog_or_tty->stamp();
  1222       gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
  1223                              ", new limit=" PTR_FORMAT,
  1224                              top, NULL);
  1228   return NULL;
  1231 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
  1232   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
  1233   // allocate, copy and if necessary update promoinfo --
  1234   // delegate to underlying space.
  1235   assert_lock_strong(freelistLock());
  1237 #ifndef PRODUCT
  1238   if (Universe::heap()->promotion_should_fail()) {
  1239     return NULL;
  1241 #endif  // #ifndef PRODUCT
  1243   oop res = _cmsSpace->promote(obj, obj_size);
  1244   if (res == NULL) {
  1245     // expand and retry
  1246     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
  1247     expand(s*HeapWordSize, MinHeapDeltaBytes,
  1248       CMSExpansionCause::_satisfy_promotion);
  1249     // Since there's currently no next generation, we don't try to promote
  1250     // into a more senior generation.
  1251     assert(next_gen() == NULL, "assumption, based upon which no attempt "
  1252                                "is made to pass on a possibly failing "
  1253                                "promotion to next generation");
  1254     res = _cmsSpace->promote(obj, obj_size);
  1256   if (res != NULL) {
  1257     // See comment in allocate() about when objects should
  1258     // be allocated live.
  1259     assert(obj->is_oop(), "Will dereference klass pointer below");
  1260     collector()->promoted(false,           // Not parallel
  1261                           (HeapWord*)res, obj->is_objArray(), obj_size);
  1262     // promotion counters
  1263     NOT_PRODUCT(
  1264       _numObjectsPromoted++;
  1265       _numWordsPromoted +=
  1266         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
  1269   return res;
  1273 HeapWord*
  1274 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
  1275                                              HeapWord* top,
  1276                                              size_t word_sz)
  1278   return collector()->allocation_limit_reached(space, top, word_sz);
  1281 // IMPORTANT: Notes on object size recognition in CMS.
  1282 // ---------------------------------------------------
  1283 // A block of storage in the CMS generation is always in
  1284 // one of three states. A free block (FREE), an allocated
  1285 // object (OBJECT) whose size() method reports the correct size,
  1286 // and an intermediate state (TRANSIENT) in which its size cannot
  1287 // be accurately determined.
  1288 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
  1289 // -----------------------------------------------------
  1290 // FREE:      klass_word & 1 == 1; mark_word holds block size
  1291 //
  1292 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
  1293 //            obj->size() computes correct size
  1294 //            [Perm Gen objects needs to be "parsable" before they can be navigated]
  1295 //
  1296 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
  1297 //
  1298 // STATE IDENTIFICATION: (64 bit+COOPS)
  1299 // ------------------------------------
  1300 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
  1301 //
  1302 // OBJECT:    klass_word installed; klass_word != 0;
  1303 //            obj->size() computes correct size
  1304 //            [Perm Gen comment above continues to hold]
  1305 //
  1306 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
  1307 //
  1308 //
  1309 // STATE TRANSITION DIAGRAM
  1310 //
  1311 //        mut / parnew                     mut  /  parnew
  1312 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
  1313 //  ^                                                                   |
  1314 //  |------------------------ DEAD <------------------------------------|
  1315 //         sweep                            mut
  1316 //
  1317 // While a block is in TRANSIENT state its size cannot be determined
  1318 // so readers will either need to come back later or stall until
  1319 // the size can be determined. Note that for the case of direct
  1320 // allocation, P-bits, when available, may be used to determine the
  1321 // size of an object that may not yet have been initialized.
  1323 // Things to support parallel young-gen collection.
  1324 oop
  1325 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
  1326                                            oop old, markOop m,
  1327                                            size_t word_sz) {
  1328 #ifndef PRODUCT
  1329   if (Universe::heap()->promotion_should_fail()) {
  1330     return NULL;
  1332 #endif  // #ifndef PRODUCT
  1334   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1335   PromotionInfo* promoInfo = &ps->promo;
  1336   // if we are tracking promotions, then first ensure space for
  1337   // promotion (including spooling space for saving header if necessary).
  1338   // then allocate and copy, then track promoted info if needed.
  1339   // When tracking (see PromotionInfo::track()), the mark word may
  1340   // be displaced and in this case restoration of the mark word
  1341   // occurs in the (oop_since_save_marks_)iterate phase.
  1342   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
  1343     // Out of space for allocating spooling buffers;
  1344     // try expanding and allocating spooling buffers.
  1345     if (!expand_and_ensure_spooling_space(promoInfo)) {
  1346       return NULL;
  1349   assert(promoInfo->has_spooling_space(), "Control point invariant");
  1350   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
  1351   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
  1352   if (obj_ptr == NULL) {
  1353      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
  1354      if (obj_ptr == NULL) {
  1355        return NULL;
  1358   oop obj = oop(obj_ptr);
  1359   OrderAccess::storestore();
  1360   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1361   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
  1362   // IMPORTANT: See note on object initialization for CMS above.
  1363   // Otherwise, copy the object.  Here we must be careful to insert the
  1364   // klass pointer last, since this marks the block as an allocated object.
  1365   // Except with compressed oops it's the mark word.
  1366   HeapWord* old_ptr = (HeapWord*)old;
  1367   // Restore the mark word copied above.
  1368   obj->set_mark(m);
  1369   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1370   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
  1371   OrderAccess::storestore();
  1373   if (UseCompressedOops) {
  1374     // Copy gap missed by (aligned) header size calculation below
  1375     obj->set_klass_gap(old->klass_gap());
  1377   if (word_sz > (size_t)oopDesc::header_size()) {
  1378     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
  1379                                  obj_ptr + oopDesc::header_size(),
  1380                                  word_sz - oopDesc::header_size());
  1383   // Now we can track the promoted object, if necessary.  We take care
  1384   // to delay the transition from uninitialized to full object
  1385   // (i.e., insertion of klass pointer) until after, so that it
  1386   // atomically becomes a promoted object.
  1387   if (promoInfo->tracking()) {
  1388     promoInfo->track((PromotedObject*)obj, old->klass());
  1390   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1391   assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size");
  1392   assert(old->is_oop(), "Will use and dereference old klass ptr below");
  1394   // Finally, install the klass pointer (this should be volatile).
  1395   OrderAccess::storestore();
  1396   obj->set_klass(old->klass());
  1397   // We should now be able to calculate the right size for this object
  1398   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
  1400   collector()->promoted(true,          // parallel
  1401                         obj_ptr, old->is_objArray(), word_sz);
  1403   NOT_PRODUCT(
  1404     Atomic::inc_ptr(&_numObjectsPromoted);
  1405     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
  1408   return obj;
  1411 void
  1412 ConcurrentMarkSweepGeneration::
  1413 par_promote_alloc_undo(int thread_num,
  1414                        HeapWord* obj, size_t word_sz) {
  1415   // CMS does not support promotion undo.
  1416   ShouldNotReachHere();
  1419 void
  1420 ConcurrentMarkSweepGeneration::
  1421 par_promote_alloc_done(int thread_num) {
  1422   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1423   ps->lab.retire(thread_num);
  1426 void
  1427 ConcurrentMarkSweepGeneration::
  1428 par_oop_since_save_marks_iterate_done(int thread_num) {
  1429   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1430   ParScanWithoutBarrierClosure* dummy_cl = NULL;
  1431   ps->promo.promoted_oops_iterate_nv(dummy_cl);
  1434 // XXXPERM
  1435 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
  1436                                                    size_t size,
  1437                                                    bool   tlab)
  1439   // We allow a STW collection only if a full
  1440   // collection was requested.
  1441   return full || should_allocate(size, tlab); // FIX ME !!!
  1442   // This and promotion failure handling are connected at the
  1443   // hip and should be fixed by untying them.
  1446 bool CMSCollector::shouldConcurrentCollect() {
  1447   if (_full_gc_requested) {
  1448     if (Verbose && PrintGCDetails) {
  1449       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
  1450                              " gc request (or gc_locker)");
  1452     return true;
  1455   // For debugging purposes, change the type of collection.
  1456   // If the rotation is not on the concurrent collection
  1457   // type, don't start a concurrent collection.
  1458   NOT_PRODUCT(
  1459     if (RotateCMSCollectionTypes &&
  1460         (_cmsGen->debug_collection_type() !=
  1461           ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
  1462       assert(_cmsGen->debug_collection_type() !=
  1463         ConcurrentMarkSweepGeneration::Unknown_collection_type,
  1464         "Bad cms collection type");
  1465       return false;
  1469   FreelistLocker x(this);
  1470   // ------------------------------------------------------------------
  1471   // Print out lots of information which affects the initiation of
  1472   // a collection.
  1473   if (PrintCMSInitiationStatistics && stats().valid()) {
  1474     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
  1475     gclog_or_tty->stamp();
  1476     gclog_or_tty->print_cr("");
  1477     stats().print_on(gclog_or_tty);
  1478     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
  1479       stats().time_until_cms_gen_full());
  1480     gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
  1481     gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
  1482                            _cmsGen->contiguous_available());
  1483     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
  1484     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
  1485     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
  1486     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
  1487     gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
  1489   // ------------------------------------------------------------------
  1491   // If the estimated time to complete a cms collection (cms_duration())
  1492   // is less than the estimated time remaining until the cms generation
  1493   // is full, start a collection.
  1494   if (!UseCMSInitiatingOccupancyOnly) {
  1495     if (stats().valid()) {
  1496       if (stats().time_until_cms_start() == 0.0) {
  1497         return true;
  1499     } else {
  1500       // We want to conservatively collect somewhat early in order
  1501       // to try and "bootstrap" our CMS/promotion statistics;
  1502       // this branch will not fire after the first successful CMS
  1503       // collection because the stats should then be valid.
  1504       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
  1505         if (Verbose && PrintGCDetails) {
  1506           gclog_or_tty->print_cr(
  1507             " CMSCollector: collect for bootstrapping statistics:"
  1508             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
  1509             _bootstrap_occupancy);
  1511         return true;
  1516   // Otherwise, we start a collection cycle if either the perm gen or
  1517   // old gen want a collection cycle started. Each may use
  1518   // an appropriate criterion for making this decision.
  1519   // XXX We need to make sure that the gen expansion
  1520   // criterion dovetails well with this. XXX NEED TO FIX THIS
  1521   if (_cmsGen->should_concurrent_collect()) {
  1522     if (Verbose && PrintGCDetails) {
  1523       gclog_or_tty->print_cr("CMS old gen initiated");
  1525     return true;
  1528   // We start a collection if we believe an incremental collection may fail;
  1529   // this is not likely to be productive in practice because it's probably too
  1530   // late anyway.
  1531   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1532   assert(gch->collector_policy()->is_two_generation_policy(),
  1533          "You may want to check the correctness of the following");
  1534   if (gch->incremental_collection_will_fail()) {
  1535     if (PrintGCDetails && Verbose) {
  1536       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
  1538     return true;
  1541   if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
  1542     bool res = update_should_unload_classes();
  1543     if (res) {
  1544       if (Verbose && PrintGCDetails) {
  1545         gclog_or_tty->print_cr("CMS perm gen initiated");
  1547       return true;
  1550   return false;
  1553 // Clear _expansion_cause fields of constituent generations
  1554 void CMSCollector::clear_expansion_cause() {
  1555   _cmsGen->clear_expansion_cause();
  1556   _permGen->clear_expansion_cause();
  1559 // We should be conservative in starting a collection cycle.  To
  1560 // start too eagerly runs the risk of collecting too often in the
  1561 // extreme.  To collect too rarely falls back on full collections,
  1562 // which works, even if not optimum in terms of concurrent work.
  1563 // As a work around for too eagerly collecting, use the flag
  1564 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
  1565 // giving the user an easily understandable way of controlling the
  1566 // collections.
  1567 // We want to start a new collection cycle if any of the following
  1568 // conditions hold:
  1569 // . our current occupancy exceeds the configured initiating occupancy
  1570 //   for this generation, or
  1571 // . we recently needed to expand this space and have not, since that
  1572 //   expansion, done a collection of this generation, or
  1573 // . the underlying space believes that it may be a good idea to initiate
  1574 //   a concurrent collection (this may be based on criteria such as the
  1575 //   following: the space uses linear allocation and linear allocation is
  1576 //   going to fail, or there is believed to be excessive fragmentation in
  1577 //   the generation, etc... or ...
  1578 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
  1579 //   the case of the old generation, not the perm generation; see CR 6543076):
  1580 //   we may be approaching a point at which allocation requests may fail because
  1581 //   we will be out of sufficient free space given allocation rate estimates.]
  1582 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
  1584   assert_lock_strong(freelistLock());
  1585   if (occupancy() > initiating_occupancy()) {
  1586     if (PrintGCDetails && Verbose) {
  1587       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
  1588         short_name(), occupancy(), initiating_occupancy());
  1590     return true;
  1592   if (UseCMSInitiatingOccupancyOnly) {
  1593     return false;
  1595   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
  1596     if (PrintGCDetails && Verbose) {
  1597       gclog_or_tty->print(" %s: collect because expanded for allocation ",
  1598         short_name());
  1600     return true;
  1602   if (_cmsSpace->should_concurrent_collect()) {
  1603     if (PrintGCDetails && Verbose) {
  1604       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
  1605         short_name());
  1607     return true;
  1609   return false;
  1612 void ConcurrentMarkSweepGeneration::collect(bool   full,
  1613                                             bool   clear_all_soft_refs,
  1614                                             size_t size,
  1615                                             bool   tlab)
  1617   collector()->collect(full, clear_all_soft_refs, size, tlab);
  1620 void CMSCollector::collect(bool   full,
  1621                            bool   clear_all_soft_refs,
  1622                            size_t size,
  1623                            bool   tlab)
  1625   if (!UseCMSCollectionPassing && _collectorState > Idling) {
  1626     // For debugging purposes skip the collection if the state
  1627     // is not currently idle
  1628     if (TraceCMSState) {
  1629       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
  1630         Thread::current(), full, _collectorState);
  1632     return;
  1635   // The following "if" branch is present for defensive reasons.
  1636   // In the current uses of this interface, it can be replaced with:
  1637   // assert(!GC_locker.is_active(), "Can't be called otherwise");
  1638   // But I am not placing that assert here to allow future
  1639   // generality in invoking this interface.
  1640   if (GC_locker::is_active()) {
  1641     // A consistency test for GC_locker
  1642     assert(GC_locker::needs_gc(), "Should have been set already");
  1643     // Skip this foreground collection, instead
  1644     // expanding the heap if necessary.
  1645     // Need the free list locks for the call to free() in compute_new_size()
  1646     compute_new_size();
  1647     return;
  1649   acquire_control_and_collect(full, clear_all_soft_refs);
  1650   _full_gcs_since_conc_gc++;
  1654 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
  1655   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1656   unsigned int gc_count = gch->total_full_collections();
  1657   if (gc_count == full_gc_count) {
  1658     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
  1659     _full_gc_requested = true;
  1660     CGC_lock->notify();   // nudge CMS thread
  1665 // The foreground and background collectors need to coordinate in order
  1666 // to make sure that they do not mutually interfere with CMS collections.
  1667 // When a background collection is active,
  1668 // the foreground collector may need to take over (preempt) and
  1669 // synchronously complete an ongoing collection. Depending on the
  1670 // frequency of the background collections and the heap usage
  1671 // of the application, this preemption can be seldom or frequent.
  1672 // There are only certain
  1673 // points in the background collection that the "collection-baton"
  1674 // can be passed to the foreground collector.
  1675 //
  1676 // The foreground collector will wait for the baton before
  1677 // starting any part of the collection.  The foreground collector
  1678 // will only wait at one location.
  1679 //
  1680 // The background collector will yield the baton before starting a new
  1681 // phase of the collection (e.g., before initial marking, marking from roots,
  1682 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
  1683 // of the loop which switches the phases. The background collector does some
  1684 // of the phases (initial mark, final re-mark) with the world stopped.
  1685 // Because of locking involved in stopping the world,
  1686 // the foreground collector should not block waiting for the background
  1687 // collector when it is doing a stop-the-world phase.  The background
  1688 // collector will yield the baton at an additional point just before
  1689 // it enters a stop-the-world phase.  Once the world is stopped, the
  1690 // background collector checks the phase of the collection.  If the
  1691 // phase has not changed, it proceeds with the collection.  If the
  1692 // phase has changed, it skips that phase of the collection.  See
  1693 // the comments on the use of the Heap_lock in collect_in_background().
  1694 //
  1695 // Variable used in baton passing.
  1696 //   _foregroundGCIsActive - Set to true by the foreground collector when
  1697 //      it wants the baton.  The foreground clears it when it has finished
  1698 //      the collection.
  1699 //   _foregroundGCShouldWait - Set to true by the background collector
  1700 //        when it is running.  The foreground collector waits while
  1701 //      _foregroundGCShouldWait is true.
  1702 //  CGC_lock - monitor used to protect access to the above variables
  1703 //      and to notify the foreground and background collectors.
  1704 //  _collectorState - current state of the CMS collection.
  1705 //
  1706 // The foreground collector
  1707 //   acquires the CGC_lock
  1708 //   sets _foregroundGCIsActive
  1709 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
  1710 //     various locks acquired in preparation for the collection
  1711 //     are released so as not to block the background collector
  1712 //     that is in the midst of a collection
  1713 //   proceeds with the collection
  1714 //   clears _foregroundGCIsActive
  1715 //   returns
  1716 //
  1717 // The background collector in a loop iterating on the phases of the
  1718 //      collection
  1719 //   acquires the CGC_lock
  1720 //   sets _foregroundGCShouldWait
  1721 //   if _foregroundGCIsActive is set
  1722 //     clears _foregroundGCShouldWait, notifies _CGC_lock
  1723 //     waits on _CGC_lock for _foregroundGCIsActive to become false
  1724 //     and exits the loop.
  1725 //   otherwise
  1726 //     proceed with that phase of the collection
  1727 //     if the phase is a stop-the-world phase,
  1728 //       yield the baton once more just before enqueueing
  1729 //       the stop-world CMS operation (executed by the VM thread).
  1730 //   returns after all phases of the collection are done
  1731 //
  1733 void CMSCollector::acquire_control_and_collect(bool full,
  1734         bool clear_all_soft_refs) {
  1735   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
  1736   assert(!Thread::current()->is_ConcurrentGC_thread(),
  1737          "shouldn't try to acquire control from self!");
  1739   // Start the protocol for acquiring control of the
  1740   // collection from the background collector (aka CMS thread).
  1741   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1742          "VM thread should have CMS token");
  1743   // Remember the possibly interrupted state of an ongoing
  1744   // concurrent collection
  1745   CollectorState first_state = _collectorState;
  1747   // Signal to a possibly ongoing concurrent collection that
  1748   // we want to do a foreground collection.
  1749   _foregroundGCIsActive = true;
  1751   // Disable incremental mode during a foreground collection.
  1752   ICMSDisabler icms_disabler;
  1754   // release locks and wait for a notify from the background collector
  1755   // releasing the locks in only necessary for phases which
  1756   // do yields to improve the granularity of the collection.
  1757   assert_lock_strong(bitMapLock());
  1758   // We need to lock the Free list lock for the space that we are
  1759   // currently collecting.
  1760   assert(haveFreelistLocks(), "Must be holding free list locks");
  1761   bitMapLock()->unlock();
  1762   releaseFreelistLocks();
  1764     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  1765     if (_foregroundGCShouldWait) {
  1766       // We are going to be waiting for action for the CMS thread;
  1767       // it had better not be gone (for instance at shutdown)!
  1768       assert(ConcurrentMarkSweepThread::cmst() != NULL,
  1769              "CMS thread must be running");
  1770       // Wait here until the background collector gives us the go-ahead
  1771       ConcurrentMarkSweepThread::clear_CMS_flag(
  1772         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
  1773       // Get a possibly blocked CMS thread going:
  1774       //   Note that we set _foregroundGCIsActive true above,
  1775       //   without protection of the CGC_lock.
  1776       CGC_lock->notify();
  1777       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
  1778              "Possible deadlock");
  1779       while (_foregroundGCShouldWait) {
  1780         // wait for notification
  1781         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  1782         // Possibility of delay/starvation here, since CMS token does
  1783         // not know to give priority to VM thread? Actually, i think
  1784         // there wouldn't be any delay/starvation, but the proof of
  1785         // that "fact" (?) appears non-trivial. XXX 20011219YSR
  1787       ConcurrentMarkSweepThread::set_CMS_flag(
  1788         ConcurrentMarkSweepThread::CMS_vm_has_token);
  1791   // The CMS_token is already held.  Get back the other locks.
  1792   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1793          "VM thread should have CMS token");
  1794   getFreelistLocks();
  1795   bitMapLock()->lock_without_safepoint_check();
  1796   if (TraceCMSState) {
  1797     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
  1798       INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
  1799     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
  1802   // Check if we need to do a compaction, or if not, whether
  1803   // we need to start the mark-sweep from scratch.
  1804   bool should_compact    = false;
  1805   bool should_start_over = false;
  1806   decide_foreground_collection_type(clear_all_soft_refs,
  1807     &should_compact, &should_start_over);
  1809 NOT_PRODUCT(
  1810   if (RotateCMSCollectionTypes) {
  1811     if (_cmsGen->debug_collection_type() ==
  1812         ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
  1813       should_compact = true;
  1814     } else if (_cmsGen->debug_collection_type() ==
  1815                ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
  1816       should_compact = false;
  1821   if (PrintGCDetails && first_state > Idling) {
  1822     GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
  1823     if (GCCause::is_user_requested_gc(cause) ||
  1824         GCCause::is_serviceability_requested_gc(cause)) {
  1825       gclog_or_tty->print(" (concurrent mode interrupted)");
  1826     } else {
  1827       gclog_or_tty->print(" (concurrent mode failure)");
  1831   if (should_compact) {
  1832     // If the collection is being acquired from the background
  1833     // collector, there may be references on the discovered
  1834     // references lists that have NULL referents (being those
  1835     // that were concurrently cleared by a mutator) or
  1836     // that are no longer active (having been enqueued concurrently
  1837     // by the mutator).
  1838     // Scrub the list of those references because Mark-Sweep-Compact
  1839     // code assumes referents are not NULL and that all discovered
  1840     // Reference objects are active.
  1841     ref_processor()->clean_up_discovered_references();
  1843     do_compaction_work(clear_all_soft_refs);
  1845     // Has the GC time limit been exceeded?
  1846     DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
  1847     size_t max_eden_size = young_gen->max_capacity() -
  1848                            young_gen->to()->capacity() -
  1849                            young_gen->from()->capacity();
  1850     GenCollectedHeap* gch = GenCollectedHeap::heap();
  1851     GCCause::Cause gc_cause = gch->gc_cause();
  1852     size_policy()->check_gc_overhead_limit(_young_gen->used(),
  1853                                            young_gen->eden()->used(),
  1854                                            _cmsGen->max_capacity(),
  1855                                            max_eden_size,
  1856                                            full,
  1857                                            gc_cause,
  1858                                            gch->collector_policy());
  1859   } else {
  1860     do_mark_sweep_work(clear_all_soft_refs, first_state,
  1861       should_start_over);
  1863   // Reset the expansion cause, now that we just completed
  1864   // a collection cycle.
  1865   clear_expansion_cause();
  1866   _foregroundGCIsActive = false;
  1867   return;
  1870 // Resize the perm generation and the tenured generation
  1871 // after obtaining the free list locks for the
  1872 // two generations.
  1873 void CMSCollector::compute_new_size() {
  1874   assert_locked_or_safepoint(Heap_lock);
  1875   FreelistLocker z(this);
  1876   _permGen->compute_new_size();
  1877   _cmsGen->compute_new_size();
  1880 // A work method used by foreground collection to determine
  1881 // what type of collection (compacting or not, continuing or fresh)
  1882 // it should do.
  1883 // NOTE: the intent is to make UseCMSCompactAtFullCollection
  1884 // and CMSCompactWhenClearAllSoftRefs the default in the future
  1885 // and do away with the flags after a suitable period.
  1886 void CMSCollector::decide_foreground_collection_type(
  1887   bool clear_all_soft_refs, bool* should_compact,
  1888   bool* should_start_over) {
  1889   // Normally, we'll compact only if the UseCMSCompactAtFullCollection
  1890   // flag is set, and we have either requested a System.gc() or
  1891   // the number of full gc's since the last concurrent cycle
  1892   // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
  1893   // or if an incremental collection has failed
  1894   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1895   assert(gch->collector_policy()->is_two_generation_policy(),
  1896          "You may want to check the correctness of the following");
  1897   // Inform cms gen if this was due to partial collection failing.
  1898   // The CMS gen may use this fact to determine its expansion policy.
  1899   if (gch->incremental_collection_will_fail()) {
  1900     assert(!_cmsGen->incremental_collection_failed(),
  1901            "Should have been noticed, reacted to and cleared");
  1902     _cmsGen->set_incremental_collection_failed();
  1904   *should_compact =
  1905     UseCMSCompactAtFullCollection &&
  1906     ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
  1907      GCCause::is_user_requested_gc(gch->gc_cause()) ||
  1908      gch->incremental_collection_will_fail());
  1909   *should_start_over = false;
  1910   if (clear_all_soft_refs && !*should_compact) {
  1911     // We are about to do a last ditch collection attempt
  1912     // so it would normally make sense to do a compaction
  1913     // to reclaim as much space as possible.
  1914     if (CMSCompactWhenClearAllSoftRefs) {
  1915       // Default: The rationale is that in this case either
  1916       // we are past the final marking phase, in which case
  1917       // we'd have to start over, or so little has been done
  1918       // that there's little point in saving that work. Compaction
  1919       // appears to be the sensible choice in either case.
  1920       *should_compact = true;
  1921     } else {
  1922       // We have been asked to clear all soft refs, but not to
  1923       // compact. Make sure that we aren't past the final checkpoint
  1924       // phase, for that is where we process soft refs. If we are already
  1925       // past that phase, we'll need to redo the refs discovery phase and
  1926       // if necessary clear soft refs that weren't previously
  1927       // cleared. We do so by remembering the phase in which
  1928       // we came in, and if we are past the refs processing
  1929       // phase, we'll choose to just redo the mark-sweep
  1930       // collection from scratch.
  1931       if (_collectorState > FinalMarking) {
  1932         // We are past the refs processing phase;
  1933         // start over and do a fresh synchronous CMS cycle
  1934         _collectorState = Resetting; // skip to reset to start new cycle
  1935         reset(false /* == !asynch */);
  1936         *should_start_over = true;
  1937       } // else we can continue a possibly ongoing current cycle
  1942 // A work method used by the foreground collector to do
  1943 // a mark-sweep-compact.
  1944 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
  1945   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1946   TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
  1947   if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
  1948     gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
  1949       "collections passed to foreground collector", _full_gcs_since_conc_gc);
  1952   // Sample collection interval time and reset for collection pause.
  1953   if (UseAdaptiveSizePolicy) {
  1954     size_policy()->msc_collection_begin();
  1957   // Temporarily widen the span of the weak reference processing to
  1958   // the entire heap.
  1959   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
  1960   ReferenceProcessorSpanMutator x(ref_processor(), new_span);
  1962   // Temporarily, clear the "is_alive_non_header" field of the
  1963   // reference processor.
  1964   ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
  1966   // Temporarily make reference _processing_ single threaded (non-MT).
  1967   ReferenceProcessorMTProcMutator z(ref_processor(), false);
  1969   // Temporarily make refs discovery atomic
  1970   ReferenceProcessorAtomicMutator w(ref_processor(), true);
  1972   ref_processor()->set_enqueuing_is_done(false);
  1973   ref_processor()->enable_discovery();
  1974   ref_processor()->setup_policy(clear_all_soft_refs);
  1975   // If an asynchronous collection finishes, the _modUnionTable is
  1976   // all clear.  If we are assuming the collection from an asynchronous
  1977   // collection, clear the _modUnionTable.
  1978   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
  1979     "_modUnionTable should be clear if the baton was not passed");
  1980   _modUnionTable.clear_all();
  1982   // We must adjust the allocation statistics being maintained
  1983   // in the free list space. We do so by reading and clearing
  1984   // the sweep timer and updating the block flux rate estimates below.
  1985   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
  1986   if (_inter_sweep_timer.is_active()) {
  1987     _inter_sweep_timer.stop();
  1988     // Note that we do not use this sample to update the _inter_sweep_estimate.
  1989     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
  1990                                             _inter_sweep_estimate.padded_average(),
  1991                                             _intra_sweep_estimate.padded_average());
  1995     TraceCMSMemoryManagerStats();
  1997   GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
  1998     ref_processor(), clear_all_soft_refs);
  1999   #ifdef ASSERT
  2000     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
  2001     size_t free_size = cms_space->free();
  2002     assert(free_size ==
  2003            pointer_delta(cms_space->end(), cms_space->compaction_top())
  2004            * HeapWordSize,
  2005       "All the free space should be compacted into one chunk at top");
  2006     assert(cms_space->dictionary()->totalChunkSize(
  2007                                       debug_only(cms_space->freelistLock())) == 0 ||
  2008            cms_space->totalSizeInIndexedFreeLists() == 0,
  2009       "All the free space should be in a single chunk");
  2010     size_t num = cms_space->totalCount();
  2011     assert((free_size == 0 && num == 0) ||
  2012            (free_size > 0  && (num == 1 || num == 2)),
  2013          "There should be at most 2 free chunks after compaction");
  2014   #endif // ASSERT
  2015   _collectorState = Resetting;
  2016   assert(_restart_addr == NULL,
  2017          "Should have been NULL'd before baton was passed");
  2018   reset(false /* == !asynch */);
  2019   _cmsGen->reset_after_compaction();
  2020   _concurrent_cycles_since_last_unload = 0;
  2022   if (verifying() && !should_unload_classes()) {
  2023     perm_gen_verify_bit_map()->clear_all();
  2026   // Clear any data recorded in the PLAB chunk arrays.
  2027   if (_survivor_plab_array != NULL) {
  2028     reset_survivor_plab_arrays();
  2031   // Adjust the per-size allocation stats for the next epoch.
  2032   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
  2033   // Restart the "inter sweep timer" for the next epoch.
  2034   _inter_sweep_timer.reset();
  2035   _inter_sweep_timer.start();
  2037   // Sample collection pause time and reset for collection interval.
  2038   if (UseAdaptiveSizePolicy) {
  2039     size_policy()->msc_collection_end(gch->gc_cause());
  2042   // For a mark-sweep-compact, compute_new_size() will be called
  2043   // in the heap's do_collection() method.
  2046 // A work method used by the foreground collector to do
  2047 // a mark-sweep, after taking over from a possibly on-going
  2048 // concurrent mark-sweep collection.
  2049 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
  2050   CollectorState first_state, bool should_start_over) {
  2051   if (PrintGC && Verbose) {
  2052     gclog_or_tty->print_cr("Pass concurrent collection to foreground "
  2053       "collector with count %d",
  2054       _full_gcs_since_conc_gc);
  2056   switch (_collectorState) {
  2057     case Idling:
  2058       if (first_state == Idling || should_start_over) {
  2059         // The background GC was not active, or should
  2060         // restarted from scratch;  start the cycle.
  2061         _collectorState = InitialMarking;
  2063       // If first_state was not Idling, then a background GC
  2064       // was in progress and has now finished.  No need to do it
  2065       // again.  Leave the state as Idling.
  2066       break;
  2067     case Precleaning:
  2068       // In the foreground case don't do the precleaning since
  2069       // it is not done concurrently and there is extra work
  2070       // required.
  2071       _collectorState = FinalMarking;
  2073   if (PrintGCDetails &&
  2074       (_collectorState > Idling ||
  2075        !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
  2076     gclog_or_tty->print(" (concurrent mode failure)");
  2078   collect_in_foreground(clear_all_soft_refs);
  2080   // For a mark-sweep, compute_new_size() will be called
  2081   // in the heap's do_collection() method.
  2085 void CMSCollector::getFreelistLocks() const {
  2086   // Get locks for all free lists in all generations that this
  2087   // collector is responsible for
  2088   _cmsGen->freelistLock()->lock_without_safepoint_check();
  2089   _permGen->freelistLock()->lock_without_safepoint_check();
  2092 void CMSCollector::releaseFreelistLocks() const {
  2093   // Release locks for all free lists in all generations that this
  2094   // collector is responsible for
  2095   _cmsGen->freelistLock()->unlock();
  2096   _permGen->freelistLock()->unlock();
  2099 bool CMSCollector::haveFreelistLocks() const {
  2100   // Check locks for all free lists in all generations that this
  2101   // collector is responsible for
  2102   assert_lock_strong(_cmsGen->freelistLock());
  2103   assert_lock_strong(_permGen->freelistLock());
  2104   PRODUCT_ONLY(ShouldNotReachHere());
  2105   return true;
  2108 // A utility class that is used by the CMS collector to
  2109 // temporarily "release" the foreground collector from its
  2110 // usual obligation to wait for the background collector to
  2111 // complete an ongoing phase before proceeding.
  2112 class ReleaseForegroundGC: public StackObj {
  2113  private:
  2114   CMSCollector* _c;
  2115  public:
  2116   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
  2117     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
  2118     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2119     // allow a potentially blocked foreground collector to proceed
  2120     _c->_foregroundGCShouldWait = false;
  2121     if (_c->_foregroundGCIsActive) {
  2122       CGC_lock->notify();
  2124     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2125            "Possible deadlock");
  2128   ~ReleaseForegroundGC() {
  2129     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
  2130     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2131     _c->_foregroundGCShouldWait = true;
  2133 };
  2135 // There are separate collect_in_background and collect_in_foreground because of
  2136 // the different locking requirements of the background collector and the
  2137 // foreground collector.  There was originally an attempt to share
  2138 // one "collect" method between the background collector and the foreground
  2139 // collector but the if-then-else required made it cleaner to have
  2140 // separate methods.
  2141 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
  2142   assert(Thread::current()->is_ConcurrentGC_thread(),
  2143     "A CMS asynchronous collection is only allowed on a CMS thread.");
  2145   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2147     bool safepoint_check = Mutex::_no_safepoint_check_flag;
  2148     MutexLockerEx hl(Heap_lock, safepoint_check);
  2149     FreelistLocker fll(this);
  2150     MutexLockerEx x(CGC_lock, safepoint_check);
  2151     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
  2152       // The foreground collector is active or we're
  2153       // not using asynchronous collections.  Skip this
  2154       // background collection.
  2155       assert(!_foregroundGCShouldWait, "Should be clear");
  2156       return;
  2157     } else {
  2158       assert(_collectorState == Idling, "Should be idling before start.");
  2159       _collectorState = InitialMarking;
  2160       // Reset the expansion cause, now that we are about to begin
  2161       // a new cycle.
  2162       clear_expansion_cause();
  2164     // Decide if we want to enable class unloading as part of the
  2165     // ensuing concurrent GC cycle.
  2166     update_should_unload_classes();
  2167     _full_gc_requested = false;           // acks all outstanding full gc requests
  2168     // Signal that we are about to start a collection
  2169     gch->increment_total_full_collections();  // ... starting a collection cycle
  2170     _collection_count_start = gch->total_full_collections();
  2173   // Used for PrintGC
  2174   size_t prev_used;
  2175   if (PrintGC && Verbose) {
  2176     prev_used = _cmsGen->used(); // XXXPERM
  2179   // The change of the collection state is normally done at this level;
  2180   // the exceptions are phases that are executed while the world is
  2181   // stopped.  For those phases the change of state is done while the
  2182   // world is stopped.  For baton passing purposes this allows the
  2183   // background collector to finish the phase and change state atomically.
  2184   // The foreground collector cannot wait on a phase that is done
  2185   // while the world is stopped because the foreground collector already
  2186   // has the world stopped and would deadlock.
  2187   while (_collectorState != Idling) {
  2188     if (TraceCMSState) {
  2189       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2190         Thread::current(), _collectorState);
  2192     // The foreground collector
  2193     //   holds the Heap_lock throughout its collection.
  2194     //   holds the CMS token (but not the lock)
  2195     //     except while it is waiting for the background collector to yield.
  2196     //
  2197     // The foreground collector should be blocked (not for long)
  2198     //   if the background collector is about to start a phase
  2199     //   executed with world stopped.  If the background
  2200     //   collector has already started such a phase, the
  2201     //   foreground collector is blocked waiting for the
  2202     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
  2203     //   are executed in the VM thread.
  2204     //
  2205     // The locking order is
  2206     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
  2207     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
  2208     //   CMS token  (claimed in
  2209     //                stop_world_and_do() -->
  2210     //                  safepoint_synchronize() -->
  2211     //                    CMSThread::synchronize())
  2214       // Check if the FG collector wants us to yield.
  2215       CMSTokenSync x(true); // is cms thread
  2216       if (waitForForegroundGC()) {
  2217         // We yielded to a foreground GC, nothing more to be
  2218         // done this round.
  2219         assert(_foregroundGCShouldWait == false, "We set it to false in "
  2220                "waitForForegroundGC()");
  2221         if (TraceCMSState) {
  2222           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2223             " exiting collection CMS state %d",
  2224             Thread::current(), _collectorState);
  2226         return;
  2227       } else {
  2228         // The background collector can run but check to see if the
  2229         // foreground collector has done a collection while the
  2230         // background collector was waiting to get the CGC_lock
  2231         // above.  If yes, break so that _foregroundGCShouldWait
  2232         // is cleared before returning.
  2233         if (_collectorState == Idling) {
  2234           break;
  2239     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
  2240       "should be waiting");
  2242     switch (_collectorState) {
  2243       case InitialMarking:
  2245           ReleaseForegroundGC x(this);
  2246           stats().record_cms_begin();
  2248           VM_CMS_Initial_Mark initial_mark_op(this);
  2249           VMThread::execute(&initial_mark_op);
  2251         // The collector state may be any legal state at this point
  2252         // since the background collector may have yielded to the
  2253         // foreground collector.
  2254         break;
  2255       case Marking:
  2256         // initial marking in checkpointRootsInitialWork has been completed
  2257         if (markFromRoots(true)) { // we were successful
  2258           assert(_collectorState == Precleaning, "Collector state should "
  2259             "have changed");
  2260         } else {
  2261           assert(_foregroundGCIsActive, "Internal state inconsistency");
  2263         break;
  2264       case Precleaning:
  2265         if (UseAdaptiveSizePolicy) {
  2266           size_policy()->concurrent_precleaning_begin();
  2268         // marking from roots in markFromRoots has been completed
  2269         preclean();
  2270         if (UseAdaptiveSizePolicy) {
  2271           size_policy()->concurrent_precleaning_end();
  2273         assert(_collectorState == AbortablePreclean ||
  2274                _collectorState == FinalMarking,
  2275                "Collector state should have changed");
  2276         break;
  2277       case AbortablePreclean:
  2278         if (UseAdaptiveSizePolicy) {
  2279         size_policy()->concurrent_phases_resume();
  2281         abortable_preclean();
  2282         if (UseAdaptiveSizePolicy) {
  2283           size_policy()->concurrent_precleaning_end();
  2285         assert(_collectorState == FinalMarking, "Collector state should "
  2286           "have changed");
  2287         break;
  2288       case FinalMarking:
  2290           ReleaseForegroundGC x(this);
  2292           VM_CMS_Final_Remark final_remark_op(this);
  2293           VMThread::execute(&final_remark_op);
  2295         assert(_foregroundGCShouldWait, "block post-condition");
  2296         break;
  2297       case Sweeping:
  2298         if (UseAdaptiveSizePolicy) {
  2299           size_policy()->concurrent_sweeping_begin();
  2301         // final marking in checkpointRootsFinal has been completed
  2302         sweep(true);
  2303         assert(_collectorState == Resizing, "Collector state change "
  2304           "to Resizing must be done under the free_list_lock");
  2305         _full_gcs_since_conc_gc = 0;
  2307         // Stop the timers for adaptive size policy for the concurrent phases
  2308         if (UseAdaptiveSizePolicy) {
  2309           size_policy()->concurrent_sweeping_end();
  2310           size_policy()->concurrent_phases_end(gch->gc_cause(),
  2311                                              gch->prev_gen(_cmsGen)->capacity(),
  2312                                              _cmsGen->free());
  2315       case Resizing: {
  2316         // Sweeping has been completed...
  2317         // At this point the background collection has completed.
  2318         // Don't move the call to compute_new_size() down
  2319         // into code that might be executed if the background
  2320         // collection was preempted.
  2322           ReleaseForegroundGC x(this);   // unblock FG collection
  2323           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
  2324           CMSTokenSync        z(true);   // not strictly needed.
  2325           if (_collectorState == Resizing) {
  2326             compute_new_size();
  2327             _collectorState = Resetting;
  2328           } else {
  2329             assert(_collectorState == Idling, "The state should only change"
  2330                    " because the foreground collector has finished the collection");
  2333         break;
  2335       case Resetting:
  2336         // CMS heap resizing has been completed
  2337         reset(true);
  2338         assert(_collectorState == Idling, "Collector state should "
  2339           "have changed");
  2340         stats().record_cms_end();
  2341         // Don't move the concurrent_phases_end() and compute_new_size()
  2342         // calls to here because a preempted background collection
  2343         // has it's state set to "Resetting".
  2344         break;
  2345       case Idling:
  2346       default:
  2347         ShouldNotReachHere();
  2348         break;
  2350     if (TraceCMSState) {
  2351       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2352         Thread::current(), _collectorState);
  2354     assert(_foregroundGCShouldWait, "block post-condition");
  2357   // Should this be in gc_epilogue?
  2358   collector_policy()->counters()->update_counters();
  2361     // Clear _foregroundGCShouldWait and, in the event that the
  2362     // foreground collector is waiting, notify it, before
  2363     // returning.
  2364     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2365     _foregroundGCShouldWait = false;
  2366     if (_foregroundGCIsActive) {
  2367       CGC_lock->notify();
  2369     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2370            "Possible deadlock");
  2372   if (TraceCMSState) {
  2373     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2374       " exiting collection CMS state %d",
  2375       Thread::current(), _collectorState);
  2377   if (PrintGC && Verbose) {
  2378     _cmsGen->print_heap_change(prev_used);
  2382 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
  2383   assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
  2384          "Foreground collector should be waiting, not executing");
  2385   assert(Thread::current()->is_VM_thread(), "A foreground collection"
  2386     "may only be done by the VM Thread with the world stopped");
  2387   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  2388          "VM thread should have CMS token");
  2390   NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
  2391     true, gclog_or_tty);)
  2392   if (UseAdaptiveSizePolicy) {
  2393     size_policy()->ms_collection_begin();
  2395   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
  2397   HandleMark hm;  // Discard invalid handles created during verification
  2399   if (VerifyBeforeGC &&
  2400       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2401     Universe::verify(true);
  2404   // Snapshot the soft reference policy to be used in this collection cycle.
  2405   ref_processor()->setup_policy(clear_all_soft_refs);
  2407   bool init_mark_was_synchronous = false; // until proven otherwise
  2408   while (_collectorState != Idling) {
  2409     if (TraceCMSState) {
  2410       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2411         Thread::current(), _collectorState);
  2413     switch (_collectorState) {
  2414       case InitialMarking:
  2415         init_mark_was_synchronous = true;  // fact to be exploited in re-mark
  2416         checkpointRootsInitial(false);
  2417         assert(_collectorState == Marking, "Collector state should have changed"
  2418           " within checkpointRootsInitial()");
  2419         break;
  2420       case Marking:
  2421         // initial marking in checkpointRootsInitialWork has been completed
  2422         if (VerifyDuringGC &&
  2423             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2424           gclog_or_tty->print("Verify before initial mark: ");
  2425           Universe::verify(true);
  2428           bool res = markFromRoots(false);
  2429           assert(res && _collectorState == FinalMarking, "Collector state should "
  2430             "have changed");
  2431           break;
  2433       case FinalMarking:
  2434         if (VerifyDuringGC &&
  2435             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2436           gclog_or_tty->print("Verify before re-mark: ");
  2437           Universe::verify(true);
  2439         checkpointRootsFinal(false, clear_all_soft_refs,
  2440                              init_mark_was_synchronous);
  2441         assert(_collectorState == Sweeping, "Collector state should not "
  2442           "have changed within checkpointRootsFinal()");
  2443         break;
  2444       case Sweeping:
  2445         // final marking in checkpointRootsFinal has been completed
  2446         if (VerifyDuringGC &&
  2447             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2448           gclog_or_tty->print("Verify before sweep: ");
  2449           Universe::verify(true);
  2451         sweep(false);
  2452         assert(_collectorState == Resizing, "Incorrect state");
  2453         break;
  2454       case Resizing: {
  2455         // Sweeping has been completed; the actual resize in this case
  2456         // is done separately; nothing to be done in this state.
  2457         _collectorState = Resetting;
  2458         break;
  2460       case Resetting:
  2461         // The heap has been resized.
  2462         if (VerifyDuringGC &&
  2463             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2464           gclog_or_tty->print("Verify before reset: ");
  2465           Universe::verify(true);
  2467         reset(false);
  2468         assert(_collectorState == Idling, "Collector state should "
  2469           "have changed");
  2470         break;
  2471       case Precleaning:
  2472       case AbortablePreclean:
  2473         // Elide the preclean phase
  2474         _collectorState = FinalMarking;
  2475         break;
  2476       default:
  2477         ShouldNotReachHere();
  2479     if (TraceCMSState) {
  2480       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2481         Thread::current(), _collectorState);
  2485   if (UseAdaptiveSizePolicy) {
  2486     GenCollectedHeap* gch = GenCollectedHeap::heap();
  2487     size_policy()->ms_collection_end(gch->gc_cause());
  2490   if (VerifyAfterGC &&
  2491       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2492     Universe::verify(true);
  2494   if (TraceCMSState) {
  2495     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2496       " exiting collection CMS state %d",
  2497       Thread::current(), _collectorState);
  2501 bool CMSCollector::waitForForegroundGC() {
  2502   bool res = false;
  2503   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2504          "CMS thread should have CMS token");
  2505   // Block the foreground collector until the
  2506   // background collectors decides whether to
  2507   // yield.
  2508   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2509   _foregroundGCShouldWait = true;
  2510   if (_foregroundGCIsActive) {
  2511     // The background collector yields to the
  2512     // foreground collector and returns a value
  2513     // indicating that it has yielded.  The foreground
  2514     // collector can proceed.
  2515     res = true;
  2516     _foregroundGCShouldWait = false;
  2517     ConcurrentMarkSweepThread::clear_CMS_flag(
  2518       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2519     ConcurrentMarkSweepThread::set_CMS_flag(
  2520       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2521     // Get a possibly blocked foreground thread going
  2522     CGC_lock->notify();
  2523     if (TraceCMSState) {
  2524       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
  2525         Thread::current(), _collectorState);
  2527     while (_foregroundGCIsActive) {
  2528       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  2530     ConcurrentMarkSweepThread::set_CMS_flag(
  2531       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2532     ConcurrentMarkSweepThread::clear_CMS_flag(
  2533       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2535   if (TraceCMSState) {
  2536     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
  2537       Thread::current(), _collectorState);
  2539   return res;
  2542 // Because of the need to lock the free lists and other structures in
  2543 // the collector, common to all the generations that the collector is
  2544 // collecting, we need the gc_prologues of individual CMS generations
  2545 // delegate to their collector. It may have been simpler had the
  2546 // current infrastructure allowed one to call a prologue on a
  2547 // collector. In the absence of that we have the generation's
  2548 // prologue delegate to the collector, which delegates back
  2549 // some "local" work to a worker method in the individual generations
  2550 // that it's responsible for collecting, while itself doing any
  2551 // work common to all generations it's responsible for. A similar
  2552 // comment applies to the  gc_epilogue()'s.
  2553 // The role of the varaible _between_prologue_and_epilogue is to
  2554 // enforce the invocation protocol.
  2555 void CMSCollector::gc_prologue(bool full) {
  2556   // Call gc_prologue_work() for each CMSGen and PermGen that
  2557   // we are responsible for.
  2559   // The following locking discipline assumes that we are only called
  2560   // when the world is stopped.
  2561   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
  2563   // The CMSCollector prologue must call the gc_prologues for the
  2564   // "generations" (including PermGen if any) that it's responsible
  2565   // for.
  2567   assert(   Thread::current()->is_VM_thread()
  2568          || (   CMSScavengeBeforeRemark
  2569              && Thread::current()->is_ConcurrentGC_thread()),
  2570          "Incorrect thread type for prologue execution");
  2572   if (_between_prologue_and_epilogue) {
  2573     // We have already been invoked; this is a gc_prologue delegation
  2574     // from yet another CMS generation that we are responsible for, just
  2575     // ignore it since all relevant work has already been done.
  2576     return;
  2579   // set a bit saying prologue has been called; cleared in epilogue
  2580   _between_prologue_and_epilogue = true;
  2581   // Claim locks for common data structures, then call gc_prologue_work()
  2582   // for each CMSGen and PermGen that we are responsible for.
  2584   getFreelistLocks();   // gets free list locks on constituent spaces
  2585   bitMapLock()->lock_without_safepoint_check();
  2587   // Should call gc_prologue_work() for all cms gens we are responsible for
  2588   bool registerClosure =    _collectorState >= Marking
  2589                          && _collectorState < Sweeping;
  2590   ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
  2591                                                &_modUnionClosurePar
  2592                                                : &_modUnionClosure;
  2593   _cmsGen->gc_prologue_work(full, registerClosure, muc);
  2594   _permGen->gc_prologue_work(full, registerClosure, muc);
  2596   if (!full) {
  2597     stats().record_gc0_begin();
  2601 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
  2602   // Delegate to CMScollector which knows how to coordinate between
  2603   // this and any other CMS generations that it is responsible for
  2604   // collecting.
  2605   collector()->gc_prologue(full);
  2608 // This is a "private" interface for use by this generation's CMSCollector.
  2609 // Not to be called directly by any other entity (for instance,
  2610 // GenCollectedHeap, which calls the "public" gc_prologue method above).
  2611 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
  2612   bool registerClosure, ModUnionClosure* modUnionClosure) {
  2613   assert(!incremental_collection_failed(), "Shouldn't be set yet");
  2614   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
  2615     "Should be NULL");
  2616   if (registerClosure) {
  2617     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
  2619   cmsSpace()->gc_prologue();
  2620   // Clear stat counters
  2621   NOT_PRODUCT(
  2622     assert(_numObjectsPromoted == 0, "check");
  2623     assert(_numWordsPromoted   == 0, "check");
  2624     if (Verbose && PrintGC) {
  2625       gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
  2626                           SIZE_FORMAT" bytes concurrently",
  2627       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
  2629     _numObjectsAllocated = 0;
  2630     _numWordsAllocated   = 0;
  2634 void CMSCollector::gc_epilogue(bool full) {
  2635   // The following locking discipline assumes that we are only called
  2636   // when the world is stopped.
  2637   assert(SafepointSynchronize::is_at_safepoint(),
  2638          "world is stopped assumption");
  2640   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
  2641   // if linear allocation blocks need to be appropriately marked to allow the
  2642   // the blocks to be parsable. We also check here whether we need to nudge the
  2643   // CMS collector thread to start a new cycle (if it's not already active).
  2644   assert(   Thread::current()->is_VM_thread()
  2645          || (   CMSScavengeBeforeRemark
  2646              && Thread::current()->is_ConcurrentGC_thread()),
  2647          "Incorrect thread type for epilogue execution");
  2649   if (!_between_prologue_and_epilogue) {
  2650     // We have already been invoked; this is a gc_epilogue delegation
  2651     // from yet another CMS generation that we are responsible for, just
  2652     // ignore it since all relevant work has already been done.
  2653     return;
  2655   assert(haveFreelistLocks(), "must have freelist locks");
  2656   assert_lock_strong(bitMapLock());
  2658   _cmsGen->gc_epilogue_work(full);
  2659   _permGen->gc_epilogue_work(full);
  2661   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
  2662     // in case sampling was not already enabled, enable it
  2663     _start_sampling = true;
  2665   // reset _eden_chunk_array so sampling starts afresh
  2666   _eden_chunk_index = 0;
  2668   size_t cms_used   = _cmsGen->cmsSpace()->used();
  2669   size_t perm_used  = _permGen->cmsSpace()->used();
  2671   // update performance counters - this uses a special version of
  2672   // update_counters() that allows the utilization to be passed as a
  2673   // parameter, avoiding multiple calls to used().
  2674   //
  2675   _cmsGen->update_counters(cms_used);
  2676   _permGen->update_counters(perm_used);
  2678   if (CMSIncrementalMode) {
  2679     icms_update_allocation_limits();
  2682   bitMapLock()->unlock();
  2683   releaseFreelistLocks();
  2685   _between_prologue_and_epilogue = false;  // ready for next cycle
  2688 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
  2689   collector()->gc_epilogue(full);
  2691   // Also reset promotion tracking in par gc thread states.
  2692   if (CollectedHeap::use_parallel_gc_threads()) {
  2693     for (uint i = 0; i < ParallelGCThreads; i++) {
  2694       _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
  2699 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
  2700   assert(!incremental_collection_failed(), "Should have been cleared");
  2701   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
  2702   cmsSpace()->gc_epilogue();
  2703     // Print stat counters
  2704   NOT_PRODUCT(
  2705     assert(_numObjectsAllocated == 0, "check");
  2706     assert(_numWordsAllocated == 0, "check");
  2707     if (Verbose && PrintGC) {
  2708       gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
  2709                           SIZE_FORMAT" bytes",
  2710                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
  2712     _numObjectsPromoted = 0;
  2713     _numWordsPromoted   = 0;
  2716   if (PrintGC && Verbose) {
  2717     // Call down the chain in contiguous_available needs the freelistLock
  2718     // so print this out before releasing the freeListLock.
  2719     gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
  2720                         contiguous_available());
  2724 #ifndef PRODUCT
  2725 bool CMSCollector::have_cms_token() {
  2726   Thread* thr = Thread::current();
  2727   if (thr->is_VM_thread()) {
  2728     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
  2729   } else if (thr->is_ConcurrentGC_thread()) {
  2730     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
  2731   } else if (thr->is_GC_task_thread()) {
  2732     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
  2733            ParGCRareEvent_lock->owned_by_self();
  2735   return false;
  2737 #endif
  2739 // Check reachability of the given heap address in CMS generation,
  2740 // treating all other generations as roots.
  2741 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
  2742   // We could "guarantee" below, rather than assert, but i'll
  2743   // leave these as "asserts" so that an adventurous debugger
  2744   // could try this in the product build provided some subset of
  2745   // the conditions were met, provided they were intersted in the
  2746   // results and knew that the computation below wouldn't interfere
  2747   // with other concurrent computations mutating the structures
  2748   // being read or written.
  2749   assert(SafepointSynchronize::is_at_safepoint(),
  2750          "Else mutations in object graph will make answer suspect");
  2751   assert(have_cms_token(), "Should hold cms token");
  2752   assert(haveFreelistLocks(), "must hold free list locks");
  2753   assert_lock_strong(bitMapLock());
  2755   // Clear the marking bit map array before starting, but, just
  2756   // for kicks, first report if the given address is already marked
  2757   gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
  2758                 _markBitMap.isMarked(addr) ? "" : " not");
  2760   if (verify_after_remark()) {
  2761     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2762     bool result = verification_mark_bm()->isMarked(addr);
  2763     gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
  2764                            result ? "IS" : "is NOT");
  2765     return result;
  2766   } else {
  2767     gclog_or_tty->print_cr("Could not compute result");
  2768     return false;
  2772 ////////////////////////////////////////////////////////
  2773 // CMS Verification Support
  2774 ////////////////////////////////////////////////////////
  2775 // Following the remark phase, the following invariant
  2776 // should hold -- each object in the CMS heap which is
  2777 // marked in markBitMap() should be marked in the verification_mark_bm().
  2779 class VerifyMarkedClosure: public BitMapClosure {
  2780   CMSBitMap* _marks;
  2781   bool       _failed;
  2783  public:
  2784   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
  2786   bool do_bit(size_t offset) {
  2787     HeapWord* addr = _marks->offsetToHeapWord(offset);
  2788     if (!_marks->isMarked(addr)) {
  2789       oop(addr)->print_on(gclog_or_tty);
  2790       gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
  2791       _failed = true;
  2793     return true;
  2796   bool failed() { return _failed; }
  2797 };
  2799 bool CMSCollector::verify_after_remark() {
  2800   gclog_or_tty->print(" [Verifying CMS Marking... ");
  2801   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2802   static bool init = false;
  2804   assert(SafepointSynchronize::is_at_safepoint(),
  2805          "Else mutations in object graph will make answer suspect");
  2806   assert(have_cms_token(),
  2807          "Else there may be mutual interference in use of "
  2808          " verification data structures");
  2809   assert(_collectorState > Marking && _collectorState <= Sweeping,
  2810          "Else marking info checked here may be obsolete");
  2811   assert(haveFreelistLocks(), "must hold free list locks");
  2812   assert_lock_strong(bitMapLock());
  2815   // Allocate marking bit map if not already allocated
  2816   if (!init) { // first time
  2817     if (!verification_mark_bm()->allocate(_span)) {
  2818       return false;
  2820     init = true;
  2823   assert(verification_mark_stack()->isEmpty(), "Should be empty");
  2825   // Turn off refs discovery -- so we will be tracing through refs.
  2826   // This is as intended, because by this time
  2827   // GC must already have cleared any refs that need to be cleared,
  2828   // and traced those that need to be marked; moreover,
  2829   // the marking done here is not going to intefere in any
  2830   // way with the marking information used by GC.
  2831   NoRefDiscovery no_discovery(ref_processor());
  2833   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  2835   // Clear any marks from a previous round
  2836   verification_mark_bm()->clear_all();
  2837   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
  2838   verify_work_stacks_empty();
  2840   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2841   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  2842   // Update the saved marks which may affect the root scans.
  2843   gch->save_marks();
  2845   if (CMSRemarkVerifyVariant == 1) {
  2846     // In this first variant of verification, we complete
  2847     // all marking, then check if the new marks-verctor is
  2848     // a subset of the CMS marks-vector.
  2849     verify_after_remark_work_1();
  2850   } else if (CMSRemarkVerifyVariant == 2) {
  2851     // In this second variant of verification, we flag an error
  2852     // (i.e. an object reachable in the new marks-vector not reachable
  2853     // in the CMS marks-vector) immediately, also indicating the
  2854     // identify of an object (A) that references the unmarked object (B) --
  2855     // presumably, a mutation to A failed to be picked up by preclean/remark?
  2856     verify_after_remark_work_2();
  2857   } else {
  2858     warning("Unrecognized value %d for CMSRemarkVerifyVariant",
  2859             CMSRemarkVerifyVariant);
  2861   gclog_or_tty->print(" done] ");
  2862   return true;
  2865 void CMSCollector::verify_after_remark_work_1() {
  2866   ResourceMark rm;
  2867   HandleMark  hm;
  2868   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2870   // Mark from roots one level into CMS
  2871   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
  2872   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  2874   gch->gen_process_strong_roots(_cmsGen->level(),
  2875                                 true,   // younger gens are roots
  2876                                 true,   // activate StrongRootsScope
  2877                                 true,   // collecting perm gen
  2878                                 SharedHeap::ScanningOption(roots_scanning_options()),
  2879                                 &notOlder,
  2880                                 true,   // walk code active on stacks
  2881                                 NULL);
  2883   // Now mark from the roots
  2884   assert(_revisitStack.isEmpty(), "Should be empty");
  2885   MarkFromRootsClosure markFromRootsClosure(this, _span,
  2886     verification_mark_bm(), verification_mark_stack(), &_revisitStack,
  2887     false /* don't yield */, true /* verifying */);
  2888   assert(_restart_addr == NULL, "Expected pre-condition");
  2889   verification_mark_bm()->iterate(&markFromRootsClosure);
  2890   while (_restart_addr != NULL) {
  2891     // Deal with stack overflow: by restarting at the indicated
  2892     // address.
  2893     HeapWord* ra = _restart_addr;
  2894     markFromRootsClosure.reset(ra);
  2895     _restart_addr = NULL;
  2896     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  2898   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  2899   verify_work_stacks_empty();
  2900   // Should reset the revisit stack above, since no class tree
  2901   // surgery is forthcoming.
  2902   _revisitStack.reset(); // throwing away all contents
  2904   // Marking completed -- now verify that each bit marked in
  2905   // verification_mark_bm() is also marked in markBitMap(); flag all
  2906   // errors by printing corresponding objects.
  2907   VerifyMarkedClosure vcl(markBitMap());
  2908   verification_mark_bm()->iterate(&vcl);
  2909   if (vcl.failed()) {
  2910     gclog_or_tty->print("Verification failed");
  2911     Universe::heap()->print_on(gclog_or_tty);
  2912     fatal("CMS: failed marking verification after remark");
  2916 void CMSCollector::verify_after_remark_work_2() {
  2917   ResourceMark rm;
  2918   HandleMark  hm;
  2919   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2921   // Mark from roots one level into CMS
  2922   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
  2923                                      markBitMap());
  2924   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  2925   gch->gen_process_strong_roots(_cmsGen->level(),
  2926                                 true,   // younger gens are roots
  2927                                 true,   // activate StrongRootsScope
  2928                                 true,   // collecting perm gen
  2929                                 SharedHeap::ScanningOption(roots_scanning_options()),
  2930                                 &notOlder,
  2931                                 true,   // walk code active on stacks
  2932                                 NULL);
  2934   // Now mark from the roots
  2935   assert(_revisitStack.isEmpty(), "Should be empty");
  2936   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
  2937     verification_mark_bm(), markBitMap(), verification_mark_stack());
  2938   assert(_restart_addr == NULL, "Expected pre-condition");
  2939   verification_mark_bm()->iterate(&markFromRootsClosure);
  2940   while (_restart_addr != NULL) {
  2941     // Deal with stack overflow: by restarting at the indicated
  2942     // address.
  2943     HeapWord* ra = _restart_addr;
  2944     markFromRootsClosure.reset(ra);
  2945     _restart_addr = NULL;
  2946     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  2948   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  2949   verify_work_stacks_empty();
  2950   // Should reset the revisit stack above, since no class tree
  2951   // surgery is forthcoming.
  2952   _revisitStack.reset(); // throwing away all contents
  2954   // Marking completed -- now verify that each bit marked in
  2955   // verification_mark_bm() is also marked in markBitMap(); flag all
  2956   // errors by printing corresponding objects.
  2957   VerifyMarkedClosure vcl(markBitMap());
  2958   verification_mark_bm()->iterate(&vcl);
  2959   assert(!vcl.failed(), "Else verification above should not have succeeded");
  2962 void ConcurrentMarkSweepGeneration::save_marks() {
  2963   // delegate to CMS space
  2964   cmsSpace()->save_marks();
  2965   for (uint i = 0; i < ParallelGCThreads; i++) {
  2966     _par_gc_thread_states[i]->promo.startTrackingPromotions();
  2970 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
  2971   return cmsSpace()->no_allocs_since_save_marks();
  2974 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
  2976 void ConcurrentMarkSweepGeneration::                            \
  2977 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
  2978   cl->set_generation(this);                                     \
  2979   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
  2980   cl->reset_generation();                                       \
  2981   save_marks();                                                 \
  2984 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
  2986 void
  2987 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
  2989   // Not currently implemented; need to do the following. -- ysr.
  2990   // dld -- I think that is used for some sort of allocation profiler.  So it
  2991   // really means the objects allocated by the mutator since the last
  2992   // GC.  We could potentially implement this cheaply by recording only
  2993   // the direct allocations in a side data structure.
  2994   //
  2995   // I think we probably ought not to be required to support these
  2996   // iterations at any arbitrary point; I think there ought to be some
  2997   // call to enable/disable allocation profiling in a generation/space,
  2998   // and the iterator ought to return the objects allocated in the
  2999   // gen/space since the enable call, or the last iterator call (which
  3000   // will probably be at a GC.)  That way, for gens like CM&S that would
  3001   // require some extra data structure to support this, we only pay the
  3002   // cost when it's in use...
  3003   cmsSpace()->object_iterate_since_last_GC(blk);
  3006 void
  3007 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
  3008   cl->set_generation(this);
  3009   younger_refs_in_space_iterate(_cmsSpace, cl);
  3010   cl->reset_generation();
  3013 void
  3014 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
  3015   if (freelistLock()->owned_by_self()) {
  3016     Generation::oop_iterate(mr, cl);
  3017   } else {
  3018     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3019     Generation::oop_iterate(mr, cl);
  3023 void
  3024 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
  3025   if (freelistLock()->owned_by_self()) {
  3026     Generation::oop_iterate(cl);
  3027   } else {
  3028     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3029     Generation::oop_iterate(cl);
  3033 void
  3034 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
  3035   if (freelistLock()->owned_by_self()) {
  3036     Generation::object_iterate(cl);
  3037   } else {
  3038     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3039     Generation::object_iterate(cl);
  3043 void
  3044 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
  3045   if (freelistLock()->owned_by_self()) {
  3046     Generation::safe_object_iterate(cl);
  3047   } else {
  3048     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3049     Generation::safe_object_iterate(cl);
  3053 void
  3054 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
  3057 void
  3058 ConcurrentMarkSweepGeneration::post_compact() {
  3061 void
  3062 ConcurrentMarkSweepGeneration::prepare_for_verify() {
  3063   // Fix the linear allocation blocks to look like free blocks.
  3065   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3066   // are not called when the heap is verified during universe initialization and
  3067   // at vm shutdown.
  3068   if (freelistLock()->owned_by_self()) {
  3069     cmsSpace()->prepare_for_verify();
  3070   } else {
  3071     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3072     cmsSpace()->prepare_for_verify();
  3076 void
  3077 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
  3078   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3079   // are not called when the heap is verified during universe initialization and
  3080   // at vm shutdown.
  3081   if (freelistLock()->owned_by_self()) {
  3082     cmsSpace()->verify(false /* ignored */);
  3083   } else {
  3084     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3085     cmsSpace()->verify(false /* ignored */);
  3089 void CMSCollector::verify(bool allow_dirty /* ignored */) {
  3090   _cmsGen->verify(allow_dirty);
  3091   _permGen->verify(allow_dirty);
  3094 #ifndef PRODUCT
  3095 bool CMSCollector::overflow_list_is_empty() const {
  3096   assert(_num_par_pushes >= 0, "Inconsistency");
  3097   if (_overflow_list == NULL) {
  3098     assert(_num_par_pushes == 0, "Inconsistency");
  3100   return _overflow_list == NULL;
  3103 // The methods verify_work_stacks_empty() and verify_overflow_empty()
  3104 // merely consolidate assertion checks that appear to occur together frequently.
  3105 void CMSCollector::verify_work_stacks_empty() const {
  3106   assert(_markStack.isEmpty(), "Marking stack should be empty");
  3107   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3110 void CMSCollector::verify_overflow_empty() const {
  3111   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3112   assert(no_preserved_marks(), "No preserved marks");
  3114 #endif // PRODUCT
  3116 // Decide if we want to enable class unloading as part of the
  3117 // ensuing concurrent GC cycle. We will collect the perm gen and
  3118 // unload classes if it's the case that:
  3119 // (1) an explicit gc request has been made and the flag
  3120 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
  3121 // (2) (a) class unloading is enabled at the command line, and
  3122 //     (b) (i)   perm gen threshold has been crossed, or
  3123 //         (ii)  old gen is getting really full, or
  3124 //         (iii) the previous N CMS collections did not collect the
  3125 //               perm gen
  3126 // NOTE: Provided there is no change in the state of the heap between
  3127 // calls to this method, it should have idempotent results. Moreover,
  3128 // its results should be monotonically increasing (i.e. going from 0 to 1,
  3129 // but not 1 to 0) between successive calls between which the heap was
  3130 // not collected. For the implementation below, it must thus rely on
  3131 // the property that concurrent_cycles_since_last_unload()
  3132 // will not decrease unless a collection cycle happened and that
  3133 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
  3134 // themselves also monotonic in that sense. See check_monotonicity()
  3135 // below.
  3136 bool CMSCollector::update_should_unload_classes() {
  3137   _should_unload_classes = false;
  3138   // Condition 1 above
  3139   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
  3140     _should_unload_classes = true;
  3141   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
  3142     // Disjuncts 2.b.(i,ii,iii) above
  3143     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
  3144                               CMSClassUnloadingMaxInterval)
  3145                            || _permGen->should_concurrent_collect()
  3146                            || _cmsGen->is_too_full();
  3148   return _should_unload_classes;
  3151 bool ConcurrentMarkSweepGeneration::is_too_full() const {
  3152   bool res = should_concurrent_collect();
  3153   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
  3154   return res;
  3157 void CMSCollector::setup_cms_unloading_and_verification_state() {
  3158   const  bool should_verify =    VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
  3159                              || VerifyBeforeExit;
  3160   const  int  rso           =    SharedHeap::SO_Symbols | SharedHeap::SO_Strings
  3161                              |   SharedHeap::SO_CodeCache;
  3163   if (should_unload_classes()) {   // Should unload classes this cycle
  3164     remove_root_scanning_option(rso);  // Shrink the root set appropriately
  3165     set_verifying(should_verify);    // Set verification state for this cycle
  3166     return;                            // Nothing else needs to be done at this time
  3169   // Not unloading classes this cycle
  3170   assert(!should_unload_classes(), "Inconsitency!");
  3171   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
  3172     // We were not verifying, or we _were_ unloading classes in the last cycle,
  3173     // AND some verification options are enabled this cycle; in this case,
  3174     // we must make sure that the deadness map is allocated if not already so,
  3175     // and cleared (if already allocated previously --
  3176     // CMSBitMap::sizeInBits() is used to determine if it's allocated).
  3177     if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
  3178       if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
  3179         warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
  3180                 "permanent generation verification disabled");
  3181         return;  // Note that we leave verification disabled, so we'll retry this
  3182                  // allocation next cycle. We _could_ remember this failure
  3183                  // and skip further attempts and permanently disable verification
  3184                  // attempts if that is considered more desirable.
  3186       assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
  3187               "_perm_gen_ver_bit_map inconsistency?");
  3188     } else {
  3189       perm_gen_verify_bit_map()->clear_all();
  3191     // Include symbols, strings and code cache elements to prevent their resurrection.
  3192     add_root_scanning_option(rso);
  3193     set_verifying(true);
  3194   } else if (verifying() && !should_verify) {
  3195     // We were verifying, but some verification flags got disabled.
  3196     set_verifying(false);
  3197     // Exclude symbols, strings and code cache elements from root scanning to
  3198     // reduce IM and RM pauses.
  3199     remove_root_scanning_option(rso);
  3204 #ifndef PRODUCT
  3205 HeapWord* CMSCollector::block_start(const void* p) const {
  3206   const HeapWord* addr = (HeapWord*)p;
  3207   if (_span.contains(p)) {
  3208     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
  3209       return _cmsGen->cmsSpace()->block_start(p);
  3210     } else {
  3211       assert(_permGen->cmsSpace()->is_in_reserved(addr),
  3212              "Inconsistent _span?");
  3213       return _permGen->cmsSpace()->block_start(p);
  3216   return NULL;
  3218 #endif
  3220 HeapWord*
  3221 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
  3222                                                    bool   tlab,
  3223                                                    bool   parallel) {
  3224   CMSSynchronousYieldRequest yr;
  3225   assert(!tlab, "Can't deal with TLAB allocation");
  3226   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3227   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
  3228     CMSExpansionCause::_satisfy_allocation);
  3229   if (GCExpandToAllocateDelayMillis > 0) {
  3230     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3232   return have_lock_and_allocate(word_size, tlab);
  3235 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
  3236 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
  3237 // to CardGeneration and share it...
  3238 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
  3239   return CardGeneration::expand(bytes, expand_bytes);
  3242 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
  3243   CMSExpansionCause::Cause cause)
  3246   bool success = expand(bytes, expand_bytes);
  3248   // remember why we expanded; this information is used
  3249   // by shouldConcurrentCollect() when making decisions on whether to start
  3250   // a new CMS cycle.
  3251   if (success) {
  3252     set_expansion_cause(cause);
  3253     if (PrintGCDetails && Verbose) {
  3254       gclog_or_tty->print_cr("Expanded CMS gen for %s",
  3255         CMSExpansionCause::to_string(cause));
  3260 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
  3261   HeapWord* res = NULL;
  3262   MutexLocker x(ParGCRareEvent_lock);
  3263   while (true) {
  3264     // Expansion by some other thread might make alloc OK now:
  3265     res = ps->lab.alloc(word_sz);
  3266     if (res != NULL) return res;
  3267     // If there's not enough expansion space available, give up.
  3268     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
  3269       return NULL;
  3271     // Otherwise, we try expansion.
  3272     expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
  3273       CMSExpansionCause::_allocate_par_lab);
  3274     // Now go around the loop and try alloc again;
  3275     // A competing par_promote might beat us to the expansion space,
  3276     // so we may go around the loop again if promotion fails agaion.
  3277     if (GCExpandToAllocateDelayMillis > 0) {
  3278       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3284 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
  3285   PromotionInfo* promo) {
  3286   MutexLocker x(ParGCRareEvent_lock);
  3287   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
  3288   while (true) {
  3289     // Expansion by some other thread might make alloc OK now:
  3290     if (promo->ensure_spooling_space()) {
  3291       assert(promo->has_spooling_space(),
  3292              "Post-condition of successful ensure_spooling_space()");
  3293       return true;
  3295     // If there's not enough expansion space available, give up.
  3296     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
  3297       return false;
  3299     // Otherwise, we try expansion.
  3300     expand(refill_size_bytes, MinHeapDeltaBytes,
  3301       CMSExpansionCause::_allocate_par_spooling_space);
  3302     // Now go around the loop and try alloc again;
  3303     // A competing allocation might beat us to the expansion space,
  3304     // so we may go around the loop again if allocation fails again.
  3305     if (GCExpandToAllocateDelayMillis > 0) {
  3306       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3313 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
  3314   assert_locked_or_safepoint(Heap_lock);
  3315   size_t size = ReservedSpace::page_align_size_down(bytes);
  3316   if (size > 0) {
  3317     shrink_by(size);
  3321 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
  3322   assert_locked_or_safepoint(Heap_lock);
  3323   bool result = _virtual_space.expand_by(bytes);
  3324   if (result) {
  3325     HeapWord* old_end = _cmsSpace->end();
  3326     size_t new_word_size =
  3327       heap_word_size(_virtual_space.committed_size());
  3328     MemRegion mr(_cmsSpace->bottom(), new_word_size);
  3329     _bts->resize(new_word_size);  // resize the block offset shared array
  3330     Universe::heap()->barrier_set()->resize_covered_region(mr);
  3331     // Hmmmm... why doesn't CFLS::set_end verify locking?
  3332     // This is quite ugly; FIX ME XXX
  3333     _cmsSpace->assert_locked(freelistLock());
  3334     _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  3336     // update the space and generation capacity counters
  3337     if (UsePerfData) {
  3338       _space_counters->update_capacity();
  3339       _gen_counters->update_all();
  3342     if (Verbose && PrintGC) {
  3343       size_t new_mem_size = _virtual_space.committed_size();
  3344       size_t old_mem_size = new_mem_size - bytes;
  3345       gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
  3346                     name(), old_mem_size/K, bytes/K, new_mem_size/K);
  3349   return result;
  3352 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
  3353   assert_locked_or_safepoint(Heap_lock);
  3354   bool success = true;
  3355   const size_t remaining_bytes = _virtual_space.uncommitted_size();
  3356   if (remaining_bytes > 0) {
  3357     success = grow_by(remaining_bytes);
  3358     DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
  3360   return success;
  3363 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
  3364   assert_locked_or_safepoint(Heap_lock);
  3365   assert_lock_strong(freelistLock());
  3366   // XXX Fix when compaction is implemented.
  3367   warning("Shrinking of CMS not yet implemented");
  3368   return;
  3372 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
  3373 // phases.
  3374 class CMSPhaseAccounting: public StackObj {
  3375  public:
  3376   CMSPhaseAccounting(CMSCollector *collector,
  3377                      const char *phase,
  3378                      bool print_cr = true);
  3379   ~CMSPhaseAccounting();
  3381  private:
  3382   CMSCollector *_collector;
  3383   const char *_phase;
  3384   elapsedTimer _wallclock;
  3385   bool _print_cr;
  3387  public:
  3388   // Not MT-safe; so do not pass around these StackObj's
  3389   // where they may be accessed by other threads.
  3390   jlong wallclock_millis() {
  3391     assert(_wallclock.is_active(), "Wall clock should not stop");
  3392     _wallclock.stop();  // to record time
  3393     jlong ret = _wallclock.milliseconds();
  3394     _wallclock.start(); // restart
  3395     return ret;
  3397 };
  3399 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
  3400                                        const char *phase,
  3401                                        bool print_cr) :
  3402   _collector(collector), _phase(phase), _print_cr(print_cr) {
  3404   if (PrintCMSStatistics != 0) {
  3405     _collector->resetYields();
  3407   if (PrintGCDetails && PrintGCTimeStamps) {
  3408     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3409     gclog_or_tty->stamp();
  3410     gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
  3411       _collector->cmsGen()->short_name(), _phase);
  3413   _collector->resetTimer();
  3414   _wallclock.start();
  3415   _collector->startTimer();
  3418 CMSPhaseAccounting::~CMSPhaseAccounting() {
  3419   assert(_wallclock.is_active(), "Wall clock should not have stopped");
  3420   _collector->stopTimer();
  3421   _wallclock.stop();
  3422   if (PrintGCDetails) {
  3423     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3424     if (PrintGCTimeStamps) {
  3425       gclog_or_tty->stamp();
  3426       gclog_or_tty->print(": ");
  3428     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
  3429                  _collector->cmsGen()->short_name(),
  3430                  _phase, _collector->timerValue(), _wallclock.seconds());
  3431     if (_print_cr) {
  3432       gclog_or_tty->print_cr("");
  3434     if (PrintCMSStatistics != 0) {
  3435       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
  3436                     _collector->yields());
  3441 // CMS work
  3443 // Checkpoint the roots into this generation from outside
  3444 // this generation. [Note this initial checkpoint need only
  3445 // be approximate -- we'll do a catch up phase subsequently.]
  3446 void CMSCollector::checkpointRootsInitial(bool asynch) {
  3447   assert(_collectorState == InitialMarking, "Wrong collector state");
  3448   check_correct_thread_executing();
  3449   TraceCMSMemoryManagerStats tms(_collectorState);
  3450   ReferenceProcessor* rp = ref_processor();
  3451   SpecializationStats::clear();
  3452   assert(_restart_addr == NULL, "Control point invariant");
  3453   if (asynch) {
  3454     // acquire locks for subsequent manipulations
  3455     MutexLockerEx x(bitMapLock(),
  3456                     Mutex::_no_safepoint_check_flag);
  3457     checkpointRootsInitialWork(asynch);
  3458     rp->verify_no_references_recorded();
  3459     rp->enable_discovery(); // enable ("weak") refs discovery
  3460     _collectorState = Marking;
  3461   } else {
  3462     // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
  3463     // which recognizes if we are a CMS generation, and doesn't try to turn on
  3464     // discovery; verify that they aren't meddling.
  3465     assert(!rp->discovery_is_atomic(),
  3466            "incorrect setting of discovery predicate");
  3467     assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
  3468            "ref discovery for this generation kind");
  3469     // already have locks
  3470     checkpointRootsInitialWork(asynch);
  3471     rp->enable_discovery(); // now enable ("weak") refs discovery
  3472     _collectorState = Marking;
  3474   SpecializationStats::print();
  3477 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
  3478   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
  3479   assert(_collectorState == InitialMarking, "just checking");
  3481   // If there has not been a GC[n-1] since last GC[n] cycle completed,
  3482   // precede our marking with a collection of all
  3483   // younger generations to keep floating garbage to a minimum.
  3484   // XXX: we won't do this for now -- it's an optimization to be done later.
  3486   // already have locks
  3487   assert_lock_strong(bitMapLock());
  3488   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
  3490   // Setup the verification and class unloading state for this
  3491   // CMS collection cycle.
  3492   setup_cms_unloading_and_verification_state();
  3494   NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
  3495     PrintGCDetails && Verbose, true, gclog_or_tty);)
  3496   if (UseAdaptiveSizePolicy) {
  3497     size_policy()->checkpoint_roots_initial_begin();
  3500   // Reset all the PLAB chunk arrays if necessary.
  3501   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
  3502     reset_survivor_plab_arrays();
  3505   ResourceMark rm;
  3506   HandleMark  hm;
  3508   FalseClosure falseClosure;
  3509   // In the case of a synchronous collection, we will elide the
  3510   // remark step, so it's important to catch all the nmethod oops
  3511   // in this step.
  3512   // The final 'true' flag to gen_process_strong_roots will ensure this.
  3513   // If 'async' is true, we can relax the nmethod tracing.
  3514   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
  3515   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3517   verify_work_stacks_empty();
  3518   verify_overflow_empty();
  3520   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  3521   // Update the saved marks which may affect the root scans.
  3522   gch->save_marks();
  3524   // weak reference processing has not started yet.
  3525   ref_processor()->set_enqueuing_is_done(false);
  3528     // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
  3529     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  3530     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  3531     gch->gen_process_strong_roots(_cmsGen->level(),
  3532                                   true,   // younger gens are roots
  3533                                   true,   // activate StrongRootsScope
  3534                                   true,   // collecting perm gen
  3535                                   SharedHeap::ScanningOption(roots_scanning_options()),
  3536                                   &notOlder,
  3537                                   true,   // walk all of code cache if (so & SO_CodeCache)
  3538                                   NULL);
  3541   // Clear mod-union table; it will be dirtied in the prologue of
  3542   // CMS generation per each younger generation collection.
  3544   assert(_modUnionTable.isAllClear(),
  3545        "Was cleared in most recent final checkpoint phase"
  3546        " or no bits are set in the gc_prologue before the start of the next "
  3547        "subsequent marking phase.");
  3549   // Temporarily disabled, since pre/post-consumption closures don't
  3550   // care about precleaned cards
  3551   #if 0
  3553     MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
  3554                              (HeapWord*)_virtual_space.high());
  3555     _ct->ct_bs()->preclean_dirty_cards(mr);
  3557   #endif
  3559   // Save the end of the used_region of the constituent generations
  3560   // to be used to limit the extent of sweep in each generation.
  3561   save_sweep_limits();
  3562   if (UseAdaptiveSizePolicy) {
  3563     size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
  3565   verify_overflow_empty();
  3568 bool CMSCollector::markFromRoots(bool asynch) {
  3569   // we might be tempted to assert that:
  3570   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
  3571   //        "inconsistent argument?");
  3572   // However that wouldn't be right, because it's possible that
  3573   // a safepoint is indeed in progress as a younger generation
  3574   // stop-the-world GC happens even as we mark in this generation.
  3575   assert(_collectorState == Marking, "inconsistent state?");
  3576   check_correct_thread_executing();
  3577   verify_overflow_empty();
  3579   bool res;
  3580   if (asynch) {
  3582     // Start the timers for adaptive size policy for the concurrent phases
  3583     // Do it here so that the foreground MS can use the concurrent
  3584     // timer since a foreground MS might has the sweep done concurrently
  3585     // or STW.
  3586     if (UseAdaptiveSizePolicy) {
  3587       size_policy()->concurrent_marking_begin();
  3590     // Weak ref discovery note: We may be discovering weak
  3591     // refs in this generation concurrent (but interleaved) with
  3592     // weak ref discovery by a younger generation collector.
  3594     CMSTokenSyncWithLocks ts(true, bitMapLock());
  3595     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  3596     CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
  3597     res = markFromRootsWork(asynch);
  3598     if (res) {
  3599       _collectorState = Precleaning;
  3600     } else { // We failed and a foreground collection wants to take over
  3601       assert(_foregroundGCIsActive, "internal state inconsistency");
  3602       assert(_restart_addr == NULL,  "foreground will restart from scratch");
  3603       if (PrintGCDetails) {
  3604         gclog_or_tty->print_cr("bailing out to foreground collection");
  3607     if (UseAdaptiveSizePolicy) {
  3608       size_policy()->concurrent_marking_end();
  3610   } else {
  3611     assert(SafepointSynchronize::is_at_safepoint(),
  3612            "inconsistent with asynch == false");
  3613     if (UseAdaptiveSizePolicy) {
  3614       size_policy()->ms_collection_marking_begin();
  3616     // already have locks
  3617     res = markFromRootsWork(asynch);
  3618     _collectorState = FinalMarking;
  3619     if (UseAdaptiveSizePolicy) {
  3620       GenCollectedHeap* gch = GenCollectedHeap::heap();
  3621       size_policy()->ms_collection_marking_end(gch->gc_cause());
  3624   verify_overflow_empty();
  3625   return res;
  3628 bool CMSCollector::markFromRootsWork(bool asynch) {
  3629   // iterate over marked bits in bit map, doing a full scan and mark
  3630   // from these roots using the following algorithm:
  3631   // . if oop is to the right of the current scan pointer,
  3632   //   mark corresponding bit (we'll process it later)
  3633   // . else (oop is to left of current scan pointer)
  3634   //   push oop on marking stack
  3635   // . drain the marking stack
  3637   // Note that when we do a marking step we need to hold the
  3638   // bit map lock -- recall that direct allocation (by mutators)
  3639   // and promotion (by younger generation collectors) is also
  3640   // marking the bit map. [the so-called allocate live policy.]
  3641   // Because the implementation of bit map marking is not
  3642   // robust wrt simultaneous marking of bits in the same word,
  3643   // we need to make sure that there is no such interference
  3644   // between concurrent such updates.
  3646   // already have locks
  3647   assert_lock_strong(bitMapLock());
  3649   // Clear the revisit stack, just in case there are any
  3650   // obsolete contents from a short-circuited previous CMS cycle.
  3651   _revisitStack.reset();
  3652   verify_work_stacks_empty();
  3653   verify_overflow_empty();
  3654   assert(_revisitStack.isEmpty(), "tabula rasa");
  3655   DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
  3656   bool result = false;
  3657   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
  3658     result = do_marking_mt(asynch);
  3659   } else {
  3660     result = do_marking_st(asynch);
  3662   return result;
  3665 // Forward decl
  3666 class CMSConcMarkingTask;
  3668 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
  3669   CMSCollector*       _collector;
  3670   CMSConcMarkingTask* _task;
  3671  public:
  3672   virtual void yield();
  3674   // "n_threads" is the number of threads to be terminated.
  3675   // "queue_set" is a set of work queues of other threads.
  3676   // "collector" is the CMS collector associated with this task terminator.
  3677   // "yield" indicates whether we need the gang as a whole to yield.
  3678   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
  3679     ParallelTaskTerminator(n_threads, queue_set),
  3680     _collector(collector) { }
  3682   void set_task(CMSConcMarkingTask* task) {
  3683     _task = task;
  3685 };
  3687 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
  3688   CMSConcMarkingTask* _task;
  3689  public:
  3690   bool should_exit_termination();
  3691   void set_task(CMSConcMarkingTask* task) {
  3692     _task = task;
  3694 };
  3696 // MT Concurrent Marking Task
  3697 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
  3698   CMSCollector* _collector;
  3699   int           _n_workers;                  // requested/desired # workers
  3700   bool          _asynch;
  3701   bool          _result;
  3702   CompactibleFreeListSpace*  _cms_space;
  3703   CompactibleFreeListSpace* _perm_space;
  3704   char          _pad_front[64];   // padding to ...
  3705   HeapWord*     _global_finger;   // ... avoid sharing cache line
  3706   char          _pad_back[64];
  3707   HeapWord*     _restart_addr;
  3709   //  Exposed here for yielding support
  3710   Mutex* const _bit_map_lock;
  3712   // The per thread work queues, available here for stealing
  3713   OopTaskQueueSet*  _task_queues;
  3715   // Termination (and yielding) support
  3716   CMSConcMarkingTerminator _term;
  3717   CMSConcMarkingTerminatorTerminator _term_term;
  3719  public:
  3720   CMSConcMarkingTask(CMSCollector* collector,
  3721                  CompactibleFreeListSpace* cms_space,
  3722                  CompactibleFreeListSpace* perm_space,
  3723                  bool asynch,
  3724                  YieldingFlexibleWorkGang* workers,
  3725                  OopTaskQueueSet* task_queues):
  3726     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
  3727     _collector(collector),
  3728     _cms_space(cms_space),
  3729     _perm_space(perm_space),
  3730     _asynch(asynch), _n_workers(0), _result(true),
  3731     _task_queues(task_queues),
  3732     _term(_n_workers, task_queues, _collector),
  3733     _bit_map_lock(collector->bitMapLock())
  3735     _requested_size = _n_workers;
  3736     _term.set_task(this);
  3737     _term_term.set_task(this);
  3738     assert(_cms_space->bottom() < _perm_space->bottom(),
  3739            "Finger incorrectly initialized below");
  3740     _restart_addr = _global_finger = _cms_space->bottom();
  3744   OopTaskQueueSet* task_queues()  { return _task_queues; }
  3746   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  3748   HeapWord** global_finger_addr() { return &_global_finger; }
  3750   CMSConcMarkingTerminator* terminator() { return &_term; }
  3752   virtual void set_for_termination(int active_workers) {
  3753     terminator()->reset_for_reuse(active_workers);
  3756   void work(int i);
  3757   bool should_yield() {
  3758     return    ConcurrentMarkSweepThread::should_yield()
  3759            && !_collector->foregroundGCIsActive()
  3760            && _asynch;
  3763   virtual void coordinator_yield();  // stuff done by coordinator
  3764   bool result() { return _result; }
  3766   void reset(HeapWord* ra) {
  3767     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
  3768     assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
  3769     assert(ra             <  _perm_space->end(), "ra too large");
  3770     _restart_addr = _global_finger = ra;
  3771     _term.reset_for_reuse();
  3774   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  3775                                            OopTaskQueue* work_q);
  3777  private:
  3778   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
  3779   void do_work_steal(int i);
  3780   void bump_global_finger(HeapWord* f);
  3781 };
  3783 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
  3784   assert(_task != NULL, "Error");
  3785   return _task->yielding();
  3786   // Note that we do not need the disjunct || _task->should_yield() above
  3787   // because we want terminating threads to yield only if the task
  3788   // is already in the midst of yielding, which happens only after at least one
  3789   // thread has yielded.
  3792 void CMSConcMarkingTerminator::yield() {
  3793   if (_task->should_yield()) {
  3794     _task->yield();
  3795   } else {
  3796     ParallelTaskTerminator::yield();
  3800 ////////////////////////////////////////////////////////////////
  3801 // Concurrent Marking Algorithm Sketch
  3802 ////////////////////////////////////////////////////////////////
  3803 // Until all tasks exhausted (both spaces):
  3804 // -- claim next available chunk
  3805 // -- bump global finger via CAS
  3806 // -- find first object that starts in this chunk
  3807 //    and start scanning bitmap from that position
  3808 // -- scan marked objects for oops
  3809 // -- CAS-mark target, and if successful:
  3810 //    . if target oop is above global finger (volatile read)
  3811 //      nothing to do
  3812 //    . if target oop is in chunk and above local finger
  3813 //        then nothing to do
  3814 //    . else push on work-queue
  3815 // -- Deal with possible overflow issues:
  3816 //    . local work-queue overflow causes stuff to be pushed on
  3817 //      global (common) overflow queue
  3818 //    . always first empty local work queue
  3819 //    . then get a batch of oops from global work queue if any
  3820 //    . then do work stealing
  3821 // -- When all tasks claimed (both spaces)
  3822 //    and local work queue empty,
  3823 //    then in a loop do:
  3824 //    . check global overflow stack; steal a batch of oops and trace
  3825 //    . try to steal from other threads oif GOS is empty
  3826 //    . if neither is available, offer termination
  3827 // -- Terminate and return result
  3828 //
  3829 void CMSConcMarkingTask::work(int i) {
  3830   elapsedTimer _timer;
  3831   ResourceMark rm;
  3832   HandleMark hm;
  3834   DEBUG_ONLY(_collector->verify_overflow_empty();)
  3836   // Before we begin work, our work queue should be empty
  3837   assert(work_queue(i)->size() == 0, "Expected to be empty");
  3838   // Scan the bitmap covering _cms_space, tracing through grey objects.
  3839   _timer.start();
  3840   do_scan_and_mark(i, _cms_space);
  3841   _timer.stop();
  3842   if (PrintCMSStatistics != 0) {
  3843     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
  3844       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3847   // ... do the same for the _perm_space
  3848   _timer.reset();
  3849   _timer.start();
  3850   do_scan_and_mark(i, _perm_space);
  3851   _timer.stop();
  3852   if (PrintCMSStatistics != 0) {
  3853     gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
  3854       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3857   // ... do work stealing
  3858   _timer.reset();
  3859   _timer.start();
  3860   do_work_steal(i);
  3861   _timer.stop();
  3862   if (PrintCMSStatistics != 0) {
  3863     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
  3864       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
  3866   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
  3867   assert(work_queue(i)->size() == 0, "Should have been emptied");
  3868   // Note that under the current task protocol, the
  3869   // following assertion is true even of the spaces
  3870   // expanded since the completion of the concurrent
  3871   // marking. XXX This will likely change under a strict
  3872   // ABORT semantics.
  3873   assert(_global_finger >  _cms_space->end() &&
  3874          _global_finger >= _perm_space->end(),
  3875          "All tasks have been completed");
  3876   DEBUG_ONLY(_collector->verify_overflow_empty();)
  3879 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
  3880   HeapWord* read = _global_finger;
  3881   HeapWord* cur  = read;
  3882   while (f > read) {
  3883     cur = read;
  3884     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
  3885     if (cur == read) {
  3886       // our cas succeeded
  3887       assert(_global_finger >= f, "protocol consistency");
  3888       break;
  3893 // This is really inefficient, and should be redone by
  3894 // using (not yet available) block-read and -write interfaces to the
  3895 // stack and the work_queue. XXX FIX ME !!!
  3896 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  3897                                                       OopTaskQueue* work_q) {
  3898   // Fast lock-free check
  3899   if (ovflw_stk->length() == 0) {
  3900     return false;
  3902   assert(work_q->size() == 0, "Shouldn't steal");
  3903   MutexLockerEx ml(ovflw_stk->par_lock(),
  3904                    Mutex::_no_safepoint_check_flag);
  3905   // Grab up to 1/4 the size of the work queue
  3906   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  3907                     (size_t)ParGCDesiredObjsFromOverflowList);
  3908   num = MIN2(num, ovflw_stk->length());
  3909   for (int i = (int) num; i > 0; i--) {
  3910     oop cur = ovflw_stk->pop();
  3911     assert(cur != NULL, "Counted wrong?");
  3912     work_q->push(cur);
  3914   return num > 0;
  3917 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
  3918   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  3919   int n_tasks = pst->n_tasks();
  3920   // We allow that there may be no tasks to do here because
  3921   // we are restarting after a stack overflow.
  3922   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
  3923   int nth_task = 0;
  3925   HeapWord* aligned_start = sp->bottom();
  3926   if (sp->used_region().contains(_restart_addr)) {
  3927     // Align down to a card boundary for the start of 0th task
  3928     // for this space.
  3929     aligned_start =
  3930       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
  3931                                  CardTableModRefBS::card_size);
  3934   size_t chunk_size = sp->marking_task_size();
  3935   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  3936     // Having claimed the nth task in this space,
  3937     // compute the chunk that it corresponds to:
  3938     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
  3939                                aligned_start + (nth_task+1)*chunk_size);
  3940     // Try and bump the global finger via a CAS;
  3941     // note that we need to do the global finger bump
  3942     // _before_ taking the intersection below, because
  3943     // the task corresponding to that region will be
  3944     // deemed done even if the used_region() expands
  3945     // because of allocation -- as it almost certainly will
  3946     // during start-up while the threads yield in the
  3947     // closure below.
  3948     HeapWord* finger = span.end();
  3949     bump_global_finger(finger);   // atomically
  3950     // There are null tasks here corresponding to chunks
  3951     // beyond the "top" address of the space.
  3952     span = span.intersection(sp->used_region());
  3953     if (!span.is_empty()) {  // Non-null task
  3954       HeapWord* prev_obj;
  3955       assert(!span.contains(_restart_addr) || nth_task == 0,
  3956              "Inconsistency");
  3957       if (nth_task == 0) {
  3958         // For the 0th task, we'll not need to compute a block_start.
  3959         if (span.contains(_restart_addr)) {
  3960           // In the case of a restart because of stack overflow,
  3961           // we might additionally skip a chunk prefix.
  3962           prev_obj = _restart_addr;
  3963         } else {
  3964           prev_obj = span.start();
  3966       } else {
  3967         // We want to skip the first object because
  3968         // the protocol is to scan any object in its entirety
  3969         // that _starts_ in this span; a fortiori, any
  3970         // object starting in an earlier span is scanned
  3971         // as part of an earlier claimed task.
  3972         // Below we use the "careful" version of block_start
  3973         // so we do not try to navigate uninitialized objects.
  3974         prev_obj = sp->block_start_careful(span.start());
  3975         // Below we use a variant of block_size that uses the
  3976         // Printezis bits to avoid waiting for allocated
  3977         // objects to become initialized/parsable.
  3978         while (prev_obj < span.start()) {
  3979           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
  3980           if (sz > 0) {
  3981             prev_obj += sz;
  3982           } else {
  3983             // In this case we may end up doing a bit of redundant
  3984             // scanning, but that appears unavoidable, short of
  3985             // locking the free list locks; see bug 6324141.
  3986             break;
  3990       if (prev_obj < span.end()) {
  3991         MemRegion my_span = MemRegion(prev_obj, span.end());
  3992         // Do the marking work within a non-empty span --
  3993         // the last argument to the constructor indicates whether the
  3994         // iteration should be incremental with periodic yields.
  3995         Par_MarkFromRootsClosure cl(this, _collector, my_span,
  3996                                     &_collector->_markBitMap,
  3997                                     work_queue(i),
  3998                                     &_collector->_markStack,
  3999                                     &_collector->_revisitStack,
  4000                                     _asynch);
  4001         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
  4002       } // else nothing to do for this task
  4003     }   // else nothing to do for this task
  4005   // We'd be tempted to assert here that since there are no
  4006   // more tasks left to claim in this space, the global_finger
  4007   // must exceed space->top() and a fortiori space->end(). However,
  4008   // that would not quite be correct because the bumping of
  4009   // global_finger occurs strictly after the claiming of a task,
  4010   // so by the time we reach here the global finger may not yet
  4011   // have been bumped up by the thread that claimed the last
  4012   // task.
  4013   pst->all_tasks_completed();
  4016 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
  4017  private:
  4018   CMSConcMarkingTask* _task;
  4019   MemRegion     _span;
  4020   CMSBitMap*    _bit_map;
  4021   CMSMarkStack* _overflow_stack;
  4022   OopTaskQueue* _work_queue;
  4023  protected:
  4024   DO_OOP_WORK_DEFN
  4025  public:
  4026   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
  4027                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
  4028                          CMSMarkStack* revisit_stack):
  4029     Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
  4030     _task(task),
  4031     _span(collector->_span),
  4032     _work_queue(work_queue),
  4033     _bit_map(bit_map),
  4034     _overflow_stack(overflow_stack)
  4035   { }
  4036   virtual void do_oop(oop* p);
  4037   virtual void do_oop(narrowOop* p);
  4038   void trim_queue(size_t max);
  4039   void handle_stack_overflow(HeapWord* lost);
  4040   void do_yield_check() {
  4041     if (_task->should_yield()) {
  4042       _task->yield();
  4045 };
  4047 // Grey object scanning during work stealing phase --
  4048 // the salient assumption here is that any references
  4049 // that are in these stolen objects being scanned must
  4050 // already have been initialized (else they would not have
  4051 // been published), so we do not need to check for
  4052 // uninitialized objects before pushing here.
  4053 void Par_ConcMarkingClosure::do_oop(oop obj) {
  4054   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  4055   HeapWord* addr = (HeapWord*)obj;
  4056   // Check if oop points into the CMS generation
  4057   // and is not marked
  4058   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  4059     // a white object ...
  4060     // If we manage to "claim" the object, by being the
  4061     // first thread to mark it, then we push it on our
  4062     // marking stack
  4063     if (_bit_map->par_mark(addr)) {     // ... now grey
  4064       // push on work queue (grey set)
  4065       bool simulate_overflow = false;
  4066       NOT_PRODUCT(
  4067         if (CMSMarkStackOverflowALot &&
  4068             _collector->simulate_overflow()) {
  4069           // simulate a stack overflow
  4070           simulate_overflow = true;
  4073       if (simulate_overflow ||
  4074           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  4075         // stack overflow
  4076         if (PrintCMSStatistics != 0) {
  4077           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  4078                                  SIZE_FORMAT, _overflow_stack->capacity());
  4080         // We cannot assert that the overflow stack is full because
  4081         // it may have been emptied since.
  4082         assert(simulate_overflow ||
  4083                _work_queue->size() == _work_queue->max_elems(),
  4084               "Else push should have succeeded");
  4085         handle_stack_overflow(addr);
  4087     } // Else, some other thread got there first
  4088     do_yield_check();
  4092 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
  4093 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
  4095 void Par_ConcMarkingClosure::trim_queue(size_t max) {
  4096   while (_work_queue->size() > max) {
  4097     oop new_oop;
  4098     if (_work_queue->pop_local(new_oop)) {
  4099       assert(new_oop->is_oop(), "Should be an oop");
  4100       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
  4101       assert(_span.contains((HeapWord*)new_oop), "Not in span");
  4102       assert(new_oop->is_parsable(), "Should be parsable");
  4103       new_oop->oop_iterate(this);  // do_oop() above
  4104       do_yield_check();
  4109 // Upon stack overflow, we discard (part of) the stack,
  4110 // remembering the least address amongst those discarded
  4111 // in CMSCollector's _restart_address.
  4112 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
  4113   // We need to do this under a mutex to prevent other
  4114   // workers from interfering with the work done below.
  4115   MutexLockerEx ml(_overflow_stack->par_lock(),
  4116                    Mutex::_no_safepoint_check_flag);
  4117   // Remember the least grey address discarded
  4118   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  4119   _collector->lower_restart_addr(ra);
  4120   _overflow_stack->reset();  // discard stack contents
  4121   _overflow_stack->expand(); // expand the stack if possible
  4125 void CMSConcMarkingTask::do_work_steal(int i) {
  4126   OopTaskQueue* work_q = work_queue(i);
  4127   oop obj_to_scan;
  4128   CMSBitMap* bm = &(_collector->_markBitMap);
  4129   CMSMarkStack* ovflw = &(_collector->_markStack);
  4130   CMSMarkStack* revisit = &(_collector->_revisitStack);
  4131   int* seed = _collector->hash_seed(i);
  4132   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
  4133   while (true) {
  4134     cl.trim_queue(0);
  4135     assert(work_q->size() == 0, "Should have been emptied above");
  4136     if (get_work_from_overflow_stack(ovflw, work_q)) {
  4137       // Can't assert below because the work obtained from the
  4138       // overflow stack may already have been stolen from us.
  4139       // assert(work_q->size() > 0, "Work from overflow stack");
  4140       continue;
  4141     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  4142       assert(obj_to_scan->is_oop(), "Should be an oop");
  4143       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
  4144       obj_to_scan->oop_iterate(&cl);
  4145     } else if (terminator()->offer_termination(&_term_term)) {
  4146       assert(work_q->size() == 0, "Impossible!");
  4147       break;
  4148     } else if (yielding() || should_yield()) {
  4149       yield();
  4154 // This is run by the CMS (coordinator) thread.
  4155 void CMSConcMarkingTask::coordinator_yield() {
  4156   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4157          "CMS thread should hold CMS token");
  4158   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  4159   // First give up the locks, then yield, then re-lock
  4160   // We should probably use a constructor/destructor idiom to
  4161   // do this unlock/lock or modify the MutexUnlocker class to
  4162   // serve our purpose. XXX
  4163   assert_lock_strong(_bit_map_lock);
  4164   _bit_map_lock->unlock();
  4165   ConcurrentMarkSweepThread::desynchronize(true);
  4166   ConcurrentMarkSweepThread::acknowledge_yield_request();
  4167   _collector->stopTimer();
  4168   if (PrintCMSStatistics != 0) {
  4169     _collector->incrementYields();
  4171   _collector->icms_wait();
  4173   // It is possible for whichever thread initiated the yield request
  4174   // not to get a chance to wake up and take the bitmap lock between
  4175   // this thread releasing it and reacquiring it. So, while the
  4176   // should_yield() flag is on, let's sleep for a bit to give the
  4177   // other thread a chance to wake up. The limit imposed on the number
  4178   // of iterations is defensive, to avoid any unforseen circumstances
  4179   // putting us into an infinite loop. Since it's always been this
  4180   // (coordinator_yield()) method that was observed to cause the
  4181   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
  4182   // which is by default non-zero. For the other seven methods that
  4183   // also perform the yield operation, as are using a different
  4184   // parameter (CMSYieldSleepCount) which is by default zero. This way we
  4185   // can enable the sleeping for those methods too, if necessary.
  4186   // See 6442774.
  4187   //
  4188   // We really need to reconsider the synchronization between the GC
  4189   // thread and the yield-requesting threads in the future and we
  4190   // should really use wait/notify, which is the recommended
  4191   // way of doing this type of interaction. Additionally, we should
  4192   // consolidate the eight methods that do the yield operation and they
  4193   // are almost identical into one for better maintenability and
  4194   // readability. See 6445193.
  4195   //
  4196   // Tony 2006.06.29
  4197   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
  4198                    ConcurrentMarkSweepThread::should_yield() &&
  4199                    !CMSCollector::foregroundGCIsActive(); ++i) {
  4200     os::sleep(Thread::current(), 1, false);
  4201     ConcurrentMarkSweepThread::acknowledge_yield_request();
  4204   ConcurrentMarkSweepThread::synchronize(true);
  4205   _bit_map_lock->lock_without_safepoint_check();
  4206   _collector->startTimer();
  4209 bool CMSCollector::do_marking_mt(bool asynch) {
  4210   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
  4211   // In the future this would be determined ergonomically, based
  4212   // on #cpu's, # active mutator threads (and load), and mutation rate.
  4213   int num_workers = ConcGCThreads;
  4215   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  4216   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
  4218   CMSConcMarkingTask tsk(this,
  4219                          cms_space,
  4220                          perm_space,
  4221                          asynch,
  4222                          conc_workers(),
  4223                          task_queues());
  4225   // Since the actual number of workers we get may be different
  4226   // from the number we requested above, do we need to do anything different
  4227   // below? In particular, may be we need to subclass the SequantialSubTasksDone
  4228   // class?? XXX
  4229   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
  4230   perm_space->initialize_sequential_subtasks_for_marking(num_workers);
  4232   // Refs discovery is already non-atomic.
  4233   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
  4234   // Mutate the Refs discovery so it is MT during the
  4235   // multi-threaded marking phase.
  4236   ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
  4237   DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
  4238   conc_workers()->start_task(&tsk);
  4239   while (tsk.yielded()) {
  4240     tsk.coordinator_yield();
  4241     conc_workers()->continue_task(&tsk);
  4243   // If the task was aborted, _restart_addr will be non-NULL
  4244   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
  4245   while (_restart_addr != NULL) {
  4246     // XXX For now we do not make use of ABORTED state and have not
  4247     // yet implemented the right abort semantics (even in the original
  4248     // single-threaded CMS case). That needs some more investigation
  4249     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
  4250     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
  4251     // If _restart_addr is non-NULL, a marking stack overflow
  4252     // occurred; we need to do a fresh marking iteration from the
  4253     // indicated restart address.
  4254     if (_foregroundGCIsActive && asynch) {
  4255       // We may be running into repeated stack overflows, having
  4256       // reached the limit of the stack size, while making very
  4257       // slow forward progress. It may be best to bail out and
  4258       // let the foreground collector do its job.
  4259       // Clear _restart_addr, so that foreground GC
  4260       // works from scratch. This avoids the headache of
  4261       // a "rescan" which would otherwise be needed because
  4262       // of the dirty mod union table & card table.
  4263       _restart_addr = NULL;
  4264       return false;
  4266     // Adjust the task to restart from _restart_addr
  4267     tsk.reset(_restart_addr);
  4268     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
  4269                   _restart_addr);
  4270     perm_space->initialize_sequential_subtasks_for_marking(num_workers,
  4271                   _restart_addr);
  4272     _restart_addr = NULL;
  4273     // Get the workers going again
  4274     conc_workers()->start_task(&tsk);
  4275     while (tsk.yielded()) {
  4276       tsk.coordinator_yield();
  4277       conc_workers()->continue_task(&tsk);
  4280   assert(tsk.completed(), "Inconsistency");
  4281   assert(tsk.result() == true, "Inconsistency");
  4282   return true;
  4285 bool CMSCollector::do_marking_st(bool asynch) {
  4286   ResourceMark rm;
  4287   HandleMark   hm;
  4289   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
  4290     &_markStack, &_revisitStack, CMSYield && asynch);
  4291   // the last argument to iterate indicates whether the iteration
  4292   // should be incremental with periodic yields.
  4293   _markBitMap.iterate(&markFromRootsClosure);
  4294   // If _restart_addr is non-NULL, a marking stack overflow
  4295   // occurred; we need to do a fresh iteration from the
  4296   // indicated restart address.
  4297   while (_restart_addr != NULL) {
  4298     if (_foregroundGCIsActive && asynch) {
  4299       // We may be running into repeated stack overflows, having
  4300       // reached the limit of the stack size, while making very
  4301       // slow forward progress. It may be best to bail out and
  4302       // let the foreground collector do its job.
  4303       // Clear _restart_addr, so that foreground GC
  4304       // works from scratch. This avoids the headache of
  4305       // a "rescan" which would otherwise be needed because
  4306       // of the dirty mod union table & card table.
  4307       _restart_addr = NULL;
  4308       return false;  // indicating failure to complete marking
  4310     // Deal with stack overflow:
  4311     // we restart marking from _restart_addr
  4312     HeapWord* ra = _restart_addr;
  4313     markFromRootsClosure.reset(ra);
  4314     _restart_addr = NULL;
  4315     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
  4317   return true;
  4320 void CMSCollector::preclean() {
  4321   check_correct_thread_executing();
  4322   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
  4323   verify_work_stacks_empty();
  4324   verify_overflow_empty();
  4325   _abort_preclean = false;
  4326   if (CMSPrecleaningEnabled) {
  4327     // Precleaning is currently not MT but the reference processor
  4328     // may be set for MT.  Disable it temporarily here.
  4329     ReferenceProcessor* rp = ref_processor();
  4330     ReferenceProcessorMTProcMutator z(rp, false);
  4331     _eden_chunk_index = 0;
  4332     size_t used = get_eden_used();
  4333     size_t capacity = get_eden_capacity();
  4334     // Don't start sampling unless we will get sufficiently
  4335     // many samples.
  4336     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
  4337                 * CMSScheduleRemarkEdenPenetration)) {
  4338       _start_sampling = true;
  4339     } else {
  4340       _start_sampling = false;
  4342     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4343     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
  4344     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
  4346   CMSTokenSync x(true); // is cms thread
  4347   if (CMSPrecleaningEnabled) {
  4348     sample_eden();
  4349     _collectorState = AbortablePreclean;
  4350   } else {
  4351     _collectorState = FinalMarking;
  4353   verify_work_stacks_empty();
  4354   verify_overflow_empty();
  4357 // Try and schedule the remark such that young gen
  4358 // occupancy is CMSScheduleRemarkEdenPenetration %.
  4359 void CMSCollector::abortable_preclean() {
  4360   check_correct_thread_executing();
  4361   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
  4362   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
  4364   // If Eden's current occupancy is below this threshold,
  4365   // immediately schedule the remark; else preclean
  4366   // past the next scavenge in an effort to
  4367   // schedule the pause as described avove. By choosing
  4368   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
  4369   // we will never do an actual abortable preclean cycle.
  4370   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
  4371     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4372     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
  4373     // We need more smarts in the abortable preclean
  4374     // loop below to deal with cases where allocation
  4375     // in young gen is very very slow, and our precleaning
  4376     // is running a losing race against a horde of
  4377     // mutators intent on flooding us with CMS updates
  4378     // (dirty cards).
  4379     // One, admittedly dumb, strategy is to give up
  4380     // after a certain number of abortable precleaning loops
  4381     // or after a certain maximum time. We want to make
  4382     // this smarter in the next iteration.
  4383     // XXX FIX ME!!! YSR
  4384     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
  4385     while (!(should_abort_preclean() ||
  4386              ConcurrentMarkSweepThread::should_terminate())) {
  4387       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
  4388       cumworkdone += workdone;
  4389       loops++;
  4390       // Voluntarily terminate abortable preclean phase if we have
  4391       // been at it for too long.
  4392       if ((CMSMaxAbortablePrecleanLoops != 0) &&
  4393           loops >= CMSMaxAbortablePrecleanLoops) {
  4394         if (PrintGCDetails) {
  4395           gclog_or_tty->print(" CMS: abort preclean due to loops ");
  4397         break;
  4399       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
  4400         if (PrintGCDetails) {
  4401           gclog_or_tty->print(" CMS: abort preclean due to time ");
  4403         break;
  4405       // If we are doing little work each iteration, we should
  4406       // take a short break.
  4407       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
  4408         // Sleep for some time, waiting for work to accumulate
  4409         stopTimer();
  4410         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
  4411         startTimer();
  4412         waited++;
  4415     if (PrintCMSStatistics > 0) {
  4416       gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
  4417                           loops, waited, cumworkdone);
  4420   CMSTokenSync x(true); // is cms thread
  4421   if (_collectorState != Idling) {
  4422     assert(_collectorState == AbortablePreclean,
  4423            "Spontaneous state transition?");
  4424     _collectorState = FinalMarking;
  4425   } // Else, a foreground collection completed this CMS cycle.
  4426   return;
  4429 // Respond to an Eden sampling opportunity
  4430 void CMSCollector::sample_eden() {
  4431   // Make sure a young gc cannot sneak in between our
  4432   // reading and recording of a sample.
  4433   assert(Thread::current()->is_ConcurrentGC_thread(),
  4434          "Only the cms thread may collect Eden samples");
  4435   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4436          "Should collect samples while holding CMS token");
  4437   if (!_start_sampling) {
  4438     return;
  4440   if (_eden_chunk_array) {
  4441     if (_eden_chunk_index < _eden_chunk_capacity) {
  4442       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
  4443       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
  4444              "Unexpected state of Eden");
  4445       // We'd like to check that what we just sampled is an oop-start address;
  4446       // however, we cannot do that here since the object may not yet have been
  4447       // initialized. So we'll instead do the check when we _use_ this sample
  4448       // later.
  4449       if (_eden_chunk_index == 0 ||
  4450           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
  4451                          _eden_chunk_array[_eden_chunk_index-1])
  4452            >= CMSSamplingGrain)) {
  4453         _eden_chunk_index++;  // commit sample
  4457   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
  4458     size_t used = get_eden_used();
  4459     size_t capacity = get_eden_capacity();
  4460     assert(used <= capacity, "Unexpected state of Eden");
  4461     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
  4462       _abort_preclean = true;
  4468 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
  4469   assert(_collectorState == Precleaning ||
  4470          _collectorState == AbortablePreclean, "incorrect state");
  4471   ResourceMark rm;
  4472   HandleMark   hm;
  4473   // Do one pass of scrubbing the discovered reference lists
  4474   // to remove any reference objects with strongly-reachable
  4475   // referents.
  4476   if (clean_refs) {
  4477     ReferenceProcessor* rp = ref_processor();
  4478     CMSPrecleanRefsYieldClosure yield_cl(this);
  4479     assert(rp->span().equals(_span), "Spans should be equal");
  4480     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
  4481                                    &_markStack, &_revisitStack,
  4482                                    true /* preclean */);
  4483     CMSDrainMarkingStackClosure complete_trace(this,
  4484                                    _span, &_markBitMap, &_markStack,
  4485                                    &keep_alive, true /* preclean */);
  4487     // We don't want this step to interfere with a young
  4488     // collection because we don't want to take CPU
  4489     // or memory bandwidth away from the young GC threads
  4490     // (which may be as many as there are CPUs).
  4491     // Note that we don't need to protect ourselves from
  4492     // interference with mutators because they can't
  4493     // manipulate the discovered reference lists nor affect
  4494     // the computed reachability of the referents, the
  4495     // only properties manipulated by the precleaning
  4496     // of these reference lists.
  4497     stopTimer();
  4498     CMSTokenSyncWithLocks x(true /* is cms thread */,
  4499                             bitMapLock());
  4500     startTimer();
  4501     sample_eden();
  4503     // The following will yield to allow foreground
  4504     // collection to proceed promptly. XXX YSR:
  4505     // The code in this method may need further
  4506     // tweaking for better performance and some restructuring
  4507     // for cleaner interfaces.
  4508     rp->preclean_discovered_references(
  4509           rp->is_alive_non_header(), &keep_alive, &complete_trace,
  4510           &yield_cl, should_unload_classes());
  4513   if (clean_survivor) {  // preclean the active survivor space(s)
  4514     assert(_young_gen->kind() == Generation::DefNew ||
  4515            _young_gen->kind() == Generation::ParNew ||
  4516            _young_gen->kind() == Generation::ASParNew,
  4517          "incorrect type for cast");
  4518     DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  4519     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
  4520                              &_markBitMap, &_modUnionTable,
  4521                              &_markStack, &_revisitStack,
  4522                              true /* precleaning phase */);
  4523     stopTimer();
  4524     CMSTokenSyncWithLocks ts(true /* is cms thread */,
  4525                              bitMapLock());
  4526     startTimer();
  4527     unsigned int before_count =
  4528       GenCollectedHeap::heap()->total_collections();
  4529     SurvivorSpacePrecleanClosure
  4530       sss_cl(this, _span, &_markBitMap, &_markStack,
  4531              &pam_cl, before_count, CMSYield);
  4532     DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
  4533     dng->from()->object_iterate_careful(&sss_cl);
  4534     dng->to()->object_iterate_careful(&sss_cl);
  4536   MarkRefsIntoAndScanClosure
  4537     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
  4538              &_markStack, &_revisitStack, this, CMSYield,
  4539              true /* precleaning phase */);
  4540   // CAUTION: The following closure has persistent state that may need to
  4541   // be reset upon a decrease in the sequence of addresses it
  4542   // processes.
  4543   ScanMarkedObjectsAgainCarefullyClosure
  4544     smoac_cl(this, _span,
  4545       &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
  4547   // Preclean dirty cards in ModUnionTable and CardTable using
  4548   // appropriate convergence criterion;
  4549   // repeat CMSPrecleanIter times unless we find that
  4550   // we are losing.
  4551   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
  4552   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
  4553          "Bad convergence multiplier");
  4554   assert(CMSPrecleanThreshold >= 100,
  4555          "Unreasonably low CMSPrecleanThreshold");
  4557   size_t numIter, cumNumCards, lastNumCards, curNumCards;
  4558   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
  4559        numIter < CMSPrecleanIter;
  4560        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
  4561     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
  4562     if (CMSPermGenPrecleaningEnabled) {
  4563       curNumCards  += preclean_mod_union_table(_permGen, &smoac_cl);
  4565     if (Verbose && PrintGCDetails) {
  4566       gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
  4568     // Either there are very few dirty cards, so re-mark
  4569     // pause will be small anyway, or our pre-cleaning isn't
  4570     // that much faster than the rate at which cards are being
  4571     // dirtied, so we might as well stop and re-mark since
  4572     // precleaning won't improve our re-mark time by much.
  4573     if (curNumCards <= CMSPrecleanThreshold ||
  4574         (numIter > 0 &&
  4575          (curNumCards * CMSPrecleanDenominator >
  4576          lastNumCards * CMSPrecleanNumerator))) {
  4577       numIter++;
  4578       cumNumCards += curNumCards;
  4579       break;
  4582   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
  4583   if (CMSPermGenPrecleaningEnabled) {
  4584     curNumCards += preclean_card_table(_permGen, &smoac_cl);
  4586   cumNumCards += curNumCards;
  4587   if (PrintGCDetails && PrintCMSStatistics != 0) {
  4588     gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
  4589                   curNumCards, cumNumCards, numIter);
  4591   return cumNumCards;   // as a measure of useful work done
  4594 // PRECLEANING NOTES:
  4595 // Precleaning involves:
  4596 // . reading the bits of the modUnionTable and clearing the set bits.
  4597 // . For the cards corresponding to the set bits, we scan the
  4598 //   objects on those cards. This means we need the free_list_lock
  4599 //   so that we can safely iterate over the CMS space when scanning
  4600 //   for oops.
  4601 // . When we scan the objects, we'll be both reading and setting
  4602 //   marks in the marking bit map, so we'll need the marking bit map.
  4603 // . For protecting _collector_state transitions, we take the CGC_lock.
  4604 //   Note that any races in the reading of of card table entries by the
  4605 //   CMS thread on the one hand and the clearing of those entries by the
  4606 //   VM thread or the setting of those entries by the mutator threads on the
  4607 //   other are quite benign. However, for efficiency it makes sense to keep
  4608 //   the VM thread from racing with the CMS thread while the latter is
  4609 //   dirty card info to the modUnionTable. We therefore also use the
  4610 //   CGC_lock to protect the reading of the card table and the mod union
  4611 //   table by the CM thread.
  4612 // . We run concurrently with mutator updates, so scanning
  4613 //   needs to be done carefully  -- we should not try to scan
  4614 //   potentially uninitialized objects.
  4615 //
  4616 // Locking strategy: While holding the CGC_lock, we scan over and
  4617 // reset a maximal dirty range of the mod union / card tables, then lock
  4618 // the free_list_lock and bitmap lock to do a full marking, then
  4619 // release these locks; and repeat the cycle. This allows for a
  4620 // certain amount of fairness in the sharing of these locks between
  4621 // the CMS collector on the one hand, and the VM thread and the
  4622 // mutators on the other.
  4624 // NOTE: preclean_mod_union_table() and preclean_card_table()
  4625 // further below are largely identical; if you need to modify
  4626 // one of these methods, please check the other method too.
  4628 size_t CMSCollector::preclean_mod_union_table(
  4629   ConcurrentMarkSweepGeneration* gen,
  4630   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4631   verify_work_stacks_empty();
  4632   verify_overflow_empty();
  4634   // Turn off checking for this method but turn it back on
  4635   // selectively.  There are yield points in this method
  4636   // but it is difficult to turn the checking off just around
  4637   // the yield points.  It is simpler to selectively turn
  4638   // it on.
  4639   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  4641   // strategy: starting with the first card, accumulate contiguous
  4642   // ranges of dirty cards; clear these cards, then scan the region
  4643   // covered by these cards.
  4645   // Since all of the MUT is committed ahead, we can just use
  4646   // that, in case the generations expand while we are precleaning.
  4647   // It might also be fine to just use the committed part of the
  4648   // generation, but we might potentially miss cards when the
  4649   // generation is rapidly expanding while we are in the midst
  4650   // of precleaning.
  4651   HeapWord* startAddr = gen->reserved().start();
  4652   HeapWord* endAddr   = gen->reserved().end();
  4654   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4656   size_t numDirtyCards, cumNumDirtyCards;
  4657   HeapWord *nextAddr, *lastAddr;
  4658   for (cumNumDirtyCards = numDirtyCards = 0,
  4659        nextAddr = lastAddr = startAddr;
  4660        nextAddr < endAddr;
  4661        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4663     ResourceMark rm;
  4664     HandleMark   hm;
  4666     MemRegion dirtyRegion;
  4668       stopTimer();
  4669       // Potential yield point
  4670       CMSTokenSync ts(true);
  4671       startTimer();
  4672       sample_eden();
  4673       // Get dirty region starting at nextOffset (inclusive),
  4674       // simultaneously clearing it.
  4675       dirtyRegion =
  4676         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
  4677       assert(dirtyRegion.start() >= nextAddr,
  4678              "returned region inconsistent?");
  4680     // Remember where the next search should begin.
  4681     // The returned region (if non-empty) is a right open interval,
  4682     // so lastOffset is obtained from the right end of that
  4683     // interval.
  4684     lastAddr = dirtyRegion.end();
  4685     // Should do something more transparent and less hacky XXX
  4686     numDirtyCards =
  4687       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
  4689     // We'll scan the cards in the dirty region (with periodic
  4690     // yields for foreground GC as needed).
  4691     if (!dirtyRegion.is_empty()) {
  4692       assert(numDirtyCards > 0, "consistency check");
  4693       HeapWord* stop_point = NULL;
  4694       stopTimer();
  4695       // Potential yield point
  4696       CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
  4697                                bitMapLock());
  4698       startTimer();
  4700         verify_work_stacks_empty();
  4701         verify_overflow_empty();
  4702         sample_eden();
  4703         DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
  4704         stop_point =
  4705           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4707       if (stop_point != NULL) {
  4708         // The careful iteration stopped early either because it found an
  4709         // uninitialized object, or because we were in the midst of an
  4710         // "abortable preclean", which should now be aborted. Redirty
  4711         // the bits corresponding to the partially-scanned or unscanned
  4712         // cards. We'll either restart at the next block boundary or
  4713         // abort the preclean.
  4714         assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
  4715                (_collectorState == AbortablePreclean && should_abort_preclean()),
  4716                "Unparsable objects should only be in perm gen.");
  4717         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
  4718         if (should_abort_preclean()) {
  4719           break; // out of preclean loop
  4720         } else {
  4721           // Compute the next address at which preclean should pick up;
  4722           // might need bitMapLock in order to read P-bits.
  4723           lastAddr = next_card_start_after_block(stop_point);
  4726     } else {
  4727       assert(lastAddr == endAddr, "consistency check");
  4728       assert(numDirtyCards == 0, "consistency check");
  4729       break;
  4732   verify_work_stacks_empty();
  4733   verify_overflow_empty();
  4734   return cumNumDirtyCards;
  4737 // NOTE: preclean_mod_union_table() above and preclean_card_table()
  4738 // below are largely identical; if you need to modify
  4739 // one of these methods, please check the other method too.
  4741 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
  4742   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4743   // strategy: it's similar to precleamModUnionTable above, in that
  4744   // we accumulate contiguous ranges of dirty cards, mark these cards
  4745   // precleaned, then scan the region covered by these cards.
  4746   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
  4747   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
  4749   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4751   size_t numDirtyCards, cumNumDirtyCards;
  4752   HeapWord *lastAddr, *nextAddr;
  4754   for (cumNumDirtyCards = numDirtyCards = 0,
  4755        nextAddr = lastAddr = startAddr;
  4756        nextAddr < endAddr;
  4757        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4759     ResourceMark rm;
  4760     HandleMark   hm;
  4762     MemRegion dirtyRegion;
  4764       // See comments in "Precleaning notes" above on why we
  4765       // do this locking. XXX Could the locking overheads be
  4766       // too high when dirty cards are sparse? [I don't think so.]
  4767       stopTimer();
  4768       CMSTokenSync x(true); // is cms thread
  4769       startTimer();
  4770       sample_eden();
  4771       // Get and clear dirty region from card table
  4772       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
  4773                                     MemRegion(nextAddr, endAddr),
  4774                                     true,
  4775                                     CardTableModRefBS::precleaned_card_val());
  4777       assert(dirtyRegion.start() >= nextAddr,
  4778              "returned region inconsistent?");
  4780     lastAddr = dirtyRegion.end();
  4781     numDirtyCards =
  4782       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
  4784     if (!dirtyRegion.is_empty()) {
  4785       stopTimer();
  4786       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
  4787       startTimer();
  4788       sample_eden();
  4789       verify_work_stacks_empty();
  4790       verify_overflow_empty();
  4791       DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
  4792       HeapWord* stop_point =
  4793         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4794       if (stop_point != NULL) {
  4795         // The careful iteration stopped early because it found an
  4796         // uninitialized object.  Redirty the bits corresponding to the
  4797         // partially-scanned or unscanned cards, and start again at the
  4798         // next block boundary.
  4799         assert(CMSPermGenPrecleaningEnabled ||
  4800                (_collectorState == AbortablePreclean && should_abort_preclean()),
  4801                "Unparsable objects should only be in perm gen.");
  4802         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
  4803         if (should_abort_preclean()) {
  4804           break; // out of preclean loop
  4805         } else {
  4806           // Compute the next address at which preclean should pick up.
  4807           lastAddr = next_card_start_after_block(stop_point);
  4810     } else {
  4811       break;
  4814   verify_work_stacks_empty();
  4815   verify_overflow_empty();
  4816   return cumNumDirtyCards;
  4819 void CMSCollector::checkpointRootsFinal(bool asynch,
  4820   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  4821   assert(_collectorState == FinalMarking, "incorrect state transition?");
  4822   check_correct_thread_executing();
  4823   // world is stopped at this checkpoint
  4824   assert(SafepointSynchronize::is_at_safepoint(),
  4825          "world should be stopped");
  4826   TraceCMSMemoryManagerStats tms(_collectorState);
  4827   verify_work_stacks_empty();
  4828   verify_overflow_empty();
  4830   SpecializationStats::clear();
  4831   if (PrintGCDetails) {
  4832     gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
  4833                         _young_gen->used() / K,
  4834                         _young_gen->capacity() / K);
  4836   if (asynch) {
  4837     if (CMSScavengeBeforeRemark) {
  4838       GenCollectedHeap* gch = GenCollectedHeap::heap();
  4839       // Temporarily set flag to false, GCH->do_collection will
  4840       // expect it to be false and set to true
  4841       FlagSetting fl(gch->_is_gc_active, false);
  4842       NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
  4843         PrintGCDetails && Verbose, true, gclog_or_tty);)
  4844       int level = _cmsGen->level() - 1;
  4845       if (level >= 0) {
  4846         gch->do_collection(true,        // full (i.e. force, see below)
  4847                            false,       // !clear_all_soft_refs
  4848                            0,           // size
  4849                            false,       // is_tlab
  4850                            level        // max_level
  4851                           );
  4854     FreelistLocker x(this);
  4855     MutexLockerEx y(bitMapLock(),
  4856                     Mutex::_no_safepoint_check_flag);
  4857     assert(!init_mark_was_synchronous, "but that's impossible!");
  4858     checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
  4859   } else {
  4860     // already have all the locks
  4861     checkpointRootsFinalWork(asynch, clear_all_soft_refs,
  4862                              init_mark_was_synchronous);
  4864   verify_work_stacks_empty();
  4865   verify_overflow_empty();
  4866   SpecializationStats::print();
  4869 void CMSCollector::checkpointRootsFinalWork(bool asynch,
  4870   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  4872   NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
  4874   assert(haveFreelistLocks(), "must have free list locks");
  4875   assert_lock_strong(bitMapLock());
  4877   if (UseAdaptiveSizePolicy) {
  4878     size_policy()->checkpoint_roots_final_begin();
  4881   ResourceMark rm;
  4882   HandleMark   hm;
  4884   GenCollectedHeap* gch = GenCollectedHeap::heap();
  4886   if (should_unload_classes()) {
  4887     CodeCache::gc_prologue();
  4889   assert(haveFreelistLocks(), "must have free list locks");
  4890   assert_lock_strong(bitMapLock());
  4892   DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
  4893   if (!init_mark_was_synchronous) {
  4894     // We might assume that we need not fill TLAB's when
  4895     // CMSScavengeBeforeRemark is set, because we may have just done
  4896     // a scavenge which would have filled all TLAB's -- and besides
  4897     // Eden would be empty. This however may not always be the case --
  4898     // for instance although we asked for a scavenge, it may not have
  4899     // happened because of a JNI critical section. We probably need
  4900     // a policy for deciding whether we can in that case wait until
  4901     // the critical section releases and then do the remark following
  4902     // the scavenge, and skip it here. In the absence of that policy,
  4903     // or of an indication of whether the scavenge did indeed occur,
  4904     // we cannot rely on TLAB's having been filled and must do
  4905     // so here just in case a scavenge did not happen.
  4906     gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
  4907     // Update the saved marks which may affect the root scans.
  4908     gch->save_marks();
  4911       COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  4913       // Note on the role of the mod union table:
  4914       // Since the marker in "markFromRoots" marks concurrently with
  4915       // mutators, it is possible for some reachable objects not to have been
  4916       // scanned. For instance, an only reference to an object A was
  4917       // placed in object B after the marker scanned B. Unless B is rescanned,
  4918       // A would be collected. Such updates to references in marked objects
  4919       // are detected via the mod union table which is the set of all cards
  4920       // dirtied since the first checkpoint in this GC cycle and prior to
  4921       // the most recent young generation GC, minus those cleaned up by the
  4922       // concurrent precleaning.
  4923       if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
  4924         TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
  4925         do_remark_parallel();
  4926       } else {
  4927         TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
  4928                     gclog_or_tty);
  4929         do_remark_non_parallel();
  4932   } else {
  4933     assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
  4934     // The initial mark was stop-world, so there's no rescanning to
  4935     // do; go straight on to the next step below.
  4937   verify_work_stacks_empty();
  4938   verify_overflow_empty();
  4941     NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
  4942     refProcessingWork(asynch, clear_all_soft_refs);
  4944   verify_work_stacks_empty();
  4945   verify_overflow_empty();
  4947   if (should_unload_classes()) {
  4948     CodeCache::gc_epilogue();
  4951   // If we encountered any (marking stack / work queue) overflow
  4952   // events during the current CMS cycle, take appropriate
  4953   // remedial measures, where possible, so as to try and avoid
  4954   // recurrence of that condition.
  4955   assert(_markStack.isEmpty(), "No grey objects");
  4956   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
  4957                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
  4958   if (ser_ovflw > 0) {
  4959     if (PrintCMSStatistics != 0) {
  4960       gclog_or_tty->print_cr("Marking stack overflow (benign) "
  4961         "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
  4962         ", kac_preclean="SIZE_FORMAT")",
  4963         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
  4964         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
  4966     _markStack.expand();
  4967     _ser_pmc_remark_ovflw = 0;
  4968     _ser_pmc_preclean_ovflw = 0;
  4969     _ser_kac_preclean_ovflw = 0;
  4970     _ser_kac_ovflw = 0;
  4972   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
  4973     if (PrintCMSStatistics != 0) {
  4974       gclog_or_tty->print_cr("Work queue overflow (benign) "
  4975         "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
  4976         _par_pmc_remark_ovflw, _par_kac_ovflw);
  4978     _par_pmc_remark_ovflw = 0;
  4979     _par_kac_ovflw = 0;
  4981   if (PrintCMSStatistics != 0) {
  4982      if (_markStack._hit_limit > 0) {
  4983        gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
  4984                               _markStack._hit_limit);
  4986      if (_markStack._failed_double > 0) {
  4987        gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
  4988                               " current capacity "SIZE_FORMAT,
  4989                               _markStack._failed_double,
  4990                               _markStack.capacity());
  4993   _markStack._hit_limit = 0;
  4994   _markStack._failed_double = 0;
  4996   // Check that all the klasses have been checked
  4997   assert(_revisitStack.isEmpty(), "Not all klasses revisited");
  4999   if ((VerifyAfterGC || VerifyDuringGC) &&
  5000       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  5001     verify_after_remark();
  5004   // Change under the freelistLocks.
  5005   _collectorState = Sweeping;
  5006   // Call isAllClear() under bitMapLock
  5007   assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
  5008     " final marking");
  5009   if (UseAdaptiveSizePolicy) {
  5010     size_policy()->checkpoint_roots_final_end(gch->gc_cause());
  5014 // Parallel remark task
  5015 class CMSParRemarkTask: public AbstractGangTask {
  5016   CMSCollector* _collector;
  5017   int           _n_workers;
  5018   CompactibleFreeListSpace* _cms_space;
  5019   CompactibleFreeListSpace* _perm_space;
  5021   // The per-thread work queues, available here for stealing.
  5022   OopTaskQueueSet*       _task_queues;
  5023   ParallelTaskTerminator _term;
  5025  public:
  5026   CMSParRemarkTask(CMSCollector* collector,
  5027                    CompactibleFreeListSpace* cms_space,
  5028                    CompactibleFreeListSpace* perm_space,
  5029                    int n_workers, FlexibleWorkGang* workers,
  5030                    OopTaskQueueSet* task_queues):
  5031     AbstractGangTask("Rescan roots and grey objects in parallel"),
  5032     _collector(collector),
  5033     _cms_space(cms_space), _perm_space(perm_space),
  5034     _n_workers(n_workers),
  5035     _task_queues(task_queues),
  5036     _term(n_workers, task_queues) { }
  5038   OopTaskQueueSet* task_queues() { return _task_queues; }
  5040   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  5042   ParallelTaskTerminator* terminator() { return &_term; }
  5043   int n_workers() { return _n_workers; }
  5045   void work(int i);
  5047  private:
  5048   // Work method in support of parallel rescan ... of young gen spaces
  5049   void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
  5050                              ContiguousSpace* space,
  5051                              HeapWord** chunk_array, size_t chunk_top);
  5053   // ... of  dirty cards in old space
  5054   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
  5055                                   Par_MarkRefsIntoAndScanClosure* cl);
  5057   // ... work stealing for the above
  5058   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
  5059 };
  5061 // work_queue(i) is passed to the closure
  5062 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
  5063 // also is passed to do_dirty_card_rescan_tasks() and to
  5064 // do_work_steal() to select the i-th task_queue.
  5066 void CMSParRemarkTask::work(int i) {
  5067   elapsedTimer _timer;
  5068   ResourceMark rm;
  5069   HandleMark   hm;
  5071   // ---------- rescan from roots --------------
  5072   _timer.start();
  5073   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5074   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
  5075     _collector->_span, _collector->ref_processor(),
  5076     &(_collector->_markBitMap),
  5077     work_queue(i), &(_collector->_revisitStack));
  5079   // Rescan young gen roots first since these are likely
  5080   // coarsely partitioned and may, on that account, constitute
  5081   // the critical path; thus, it's best to start off that
  5082   // work first.
  5083   // ---------- young gen roots --------------
  5085     DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
  5086     EdenSpace* eden_space = dng->eden();
  5087     ContiguousSpace* from_space = dng->from();
  5088     ContiguousSpace* to_space   = dng->to();
  5090     HeapWord** eca = _collector->_eden_chunk_array;
  5091     size_t     ect = _collector->_eden_chunk_index;
  5092     HeapWord** sca = _collector->_survivor_chunk_array;
  5093     size_t     sct = _collector->_survivor_chunk_index;
  5095     assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
  5096     assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
  5098     do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
  5099     do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
  5100     do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
  5102     _timer.stop();
  5103     if (PrintCMSStatistics != 0) {
  5104       gclog_or_tty->print_cr(
  5105         "Finished young gen rescan work in %dth thread: %3.3f sec",
  5106         i, _timer.seconds());
  5110   // ---------- remaining roots --------------
  5111   _timer.reset();
  5112   _timer.start();
  5113   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
  5114                                 false,     // yg was scanned above
  5115                                 false,     // this is parallel code
  5116                                 true,      // collecting perm gen
  5117                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
  5118                                 &par_mrias_cl,
  5119                                 true,   // walk all of code cache if (so & SO_CodeCache)
  5120                                 NULL);
  5121   assert(_collector->should_unload_classes()
  5122          || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
  5123          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
  5124   _timer.stop();
  5125   if (PrintCMSStatistics != 0) {
  5126     gclog_or_tty->print_cr(
  5127       "Finished remaining root rescan work in %dth thread: %3.3f sec",
  5128       i, _timer.seconds());
  5131   // ---------- rescan dirty cards ------------
  5132   _timer.reset();
  5133   _timer.start();
  5135   // Do the rescan tasks for each of the two spaces
  5136   // (cms_space and perm_space) in turn.
  5137   // "i" is passed to select the "i-th" task_queue
  5138   do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
  5139   do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
  5140   _timer.stop();
  5141   if (PrintCMSStatistics != 0) {
  5142     gclog_or_tty->print_cr(
  5143       "Finished dirty card rescan work in %dth thread: %3.3f sec",
  5144       i, _timer.seconds());
  5147   // ---------- steal work from other threads ...
  5148   // ---------- ... and drain overflow list.
  5149   _timer.reset();
  5150   _timer.start();
  5151   do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
  5152   _timer.stop();
  5153   if (PrintCMSStatistics != 0) {
  5154     gclog_or_tty->print_cr(
  5155       "Finished work stealing in %dth thread: %3.3f sec",
  5156       i, _timer.seconds());
  5160 // Note that parameter "i" is not used.
  5161 void
  5162 CMSParRemarkTask::do_young_space_rescan(int i,
  5163   Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
  5164   HeapWord** chunk_array, size_t chunk_top) {
  5165   // Until all tasks completed:
  5166   // . claim an unclaimed task
  5167   // . compute region boundaries corresponding to task claimed
  5168   //   using chunk_array
  5169   // . par_oop_iterate(cl) over that region
  5171   ResourceMark rm;
  5172   HandleMark   hm;
  5174   SequentialSubTasksDone* pst = space->par_seq_tasks();
  5175   assert(pst->valid(), "Uninitialized use?");
  5177   int nth_task = 0;
  5178   int n_tasks  = pst->n_tasks();
  5180   HeapWord *start, *end;
  5181   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5182     // We claimed task # nth_task; compute its boundaries.
  5183     if (chunk_top == 0) {  // no samples were taken
  5184       assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
  5185       start = space->bottom();
  5186       end   = space->top();
  5187     } else if (nth_task == 0) {
  5188       start = space->bottom();
  5189       end   = chunk_array[nth_task];
  5190     } else if (nth_task < (jint)chunk_top) {
  5191       assert(nth_task >= 1, "Control point invariant");
  5192       start = chunk_array[nth_task - 1];
  5193       end   = chunk_array[nth_task];
  5194     } else {
  5195       assert(nth_task == (jint)chunk_top, "Control point invariant");
  5196       start = chunk_array[chunk_top - 1];
  5197       end   = space->top();
  5199     MemRegion mr(start, end);
  5200     // Verify that mr is in space
  5201     assert(mr.is_empty() || space->used_region().contains(mr),
  5202            "Should be in space");
  5203     // Verify that "start" is an object boundary
  5204     assert(mr.is_empty() || oop(mr.start())->is_oop(),
  5205            "Should be an oop");
  5206     space->par_oop_iterate(mr, cl);
  5208   pst->all_tasks_completed();
  5211 void
  5212 CMSParRemarkTask::do_dirty_card_rescan_tasks(
  5213   CompactibleFreeListSpace* sp, int i,
  5214   Par_MarkRefsIntoAndScanClosure* cl) {
  5215   // Until all tasks completed:
  5216   // . claim an unclaimed task
  5217   // . compute region boundaries corresponding to task claimed
  5218   // . transfer dirty bits ct->mut for that region
  5219   // . apply rescanclosure to dirty mut bits for that region
  5221   ResourceMark rm;
  5222   HandleMark   hm;
  5224   OopTaskQueue* work_q = work_queue(i);
  5225   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
  5226   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
  5227   // CAUTION: This closure has state that persists across calls to
  5228   // the work method dirty_range_iterate_clear() in that it has
  5229   // imbedded in it a (subtype of) UpwardsObjectClosure. The
  5230   // use of that state in the imbedded UpwardsObjectClosure instance
  5231   // assumes that the cards are always iterated (even if in parallel
  5232   // by several threads) in monotonically increasing order per each
  5233   // thread. This is true of the implementation below which picks
  5234   // card ranges (chunks) in monotonically increasing order globally
  5235   // and, a-fortiori, in monotonically increasing order per thread
  5236   // (the latter order being a subsequence of the former).
  5237   // If the work code below is ever reorganized into a more chaotic
  5238   // work-partitioning form than the current "sequential tasks"
  5239   // paradigm, the use of that persistent state will have to be
  5240   // revisited and modified appropriately. See also related
  5241   // bug 4756801 work on which should examine this code to make
  5242   // sure that the changes there do not run counter to the
  5243   // assumptions made here and necessary for correctness and
  5244   // efficiency. Note also that this code might yield inefficient
  5245   // behaviour in the case of very large objects that span one or
  5246   // more work chunks. Such objects would potentially be scanned
  5247   // several times redundantly. Work on 4756801 should try and
  5248   // address that performance anomaly if at all possible. XXX
  5249   MemRegion  full_span  = _collector->_span;
  5250   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
  5251   CMSMarkStack* rs = &(_collector->_revisitStack);   // shared
  5252   MarkFromDirtyCardsClosure
  5253     greyRescanClosure(_collector, full_span, // entire span of interest
  5254                       sp, bm, work_q, rs, cl);
  5256   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  5257   assert(pst->valid(), "Uninitialized use?");
  5258   int nth_task = 0;
  5259   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
  5260   MemRegion span = sp->used_region();
  5261   HeapWord* start_addr = span.start();
  5262   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
  5263                                            alignment);
  5264   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
  5265   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
  5266          start_addr, "Check alignment");
  5267   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
  5268          chunk_size, "Check alignment");
  5270   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5271     // Having claimed the nth_task, compute corresponding mem-region,
  5272     // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
  5273     // The alignment restriction ensures that we do not need any
  5274     // synchronization with other gang-workers while setting or
  5275     // clearing bits in thus chunk of the MUT.
  5276     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
  5277                                     start_addr + (nth_task+1)*chunk_size);
  5278     // The last chunk's end might be way beyond end of the
  5279     // used region. In that case pull back appropriately.
  5280     if (this_span.end() > end_addr) {
  5281       this_span.set_end(end_addr);
  5282       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
  5284     // Iterate over the dirty cards covering this chunk, marking them
  5285     // precleaned, and setting the corresponding bits in the mod union
  5286     // table. Since we have been careful to partition at Card and MUT-word
  5287     // boundaries no synchronization is needed between parallel threads.
  5288     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
  5289                                                  &modUnionClosure);
  5291     // Having transferred these marks into the modUnionTable,
  5292     // rescan the marked objects on the dirty cards in the modUnionTable.
  5293     // Even if this is at a synchronous collection, the initial marking
  5294     // may have been done during an asynchronous collection so there
  5295     // may be dirty bits in the mod-union table.
  5296     _collector->_modUnionTable.dirty_range_iterate_clear(
  5297                   this_span, &greyRescanClosure);
  5298     _collector->_modUnionTable.verifyNoOneBitsInRange(
  5299                                  this_span.start(),
  5300                                  this_span.end());
  5302   pst->all_tasks_completed();  // declare that i am done
  5305 // . see if we can share work_queues with ParNew? XXX
  5306 void
  5307 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
  5308                                 int* seed) {
  5309   OopTaskQueue* work_q = work_queue(i);
  5310   NOT_PRODUCT(int num_steals = 0;)
  5311   oop obj_to_scan;
  5312   CMSBitMap* bm = &(_collector->_markBitMap);
  5314   while (true) {
  5315     // Completely finish any left over work from (an) earlier round(s)
  5316     cl->trim_queue(0);
  5317     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  5318                                          (size_t)ParGCDesiredObjsFromOverflowList);
  5319     // Now check if there's any work in the overflow list
  5320     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
  5321     // only affects the number of attempts made to get work from the
  5322     // overflow list and does not affect the number of workers.  Just
  5323     // pass ParallelGCThreads so this behavior is unchanged.
  5324     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  5325                                                 work_q,
  5326                                                 ParallelGCThreads)) {
  5327       // found something in global overflow list;
  5328       // not yet ready to go stealing work from others.
  5329       // We'd like to assert(work_q->size() != 0, ...)
  5330       // because we just took work from the overflow list,
  5331       // but of course we can't since all of that could have
  5332       // been already stolen from us.
  5333       // "He giveth and He taketh away."
  5334       continue;
  5336     // Verify that we have no work before we resort to stealing
  5337     assert(work_q->size() == 0, "Have work, shouldn't steal");
  5338     // Try to steal from other queues that have work
  5339     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  5340       NOT_PRODUCT(num_steals++;)
  5341       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  5342       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  5343       // Do scanning work
  5344       obj_to_scan->oop_iterate(cl);
  5345       // Loop around, finish this work, and try to steal some more
  5346     } else if (terminator()->offer_termination()) {
  5347         break;  // nirvana from the infinite cycle
  5350   NOT_PRODUCT(
  5351     if (PrintCMSStatistics != 0) {
  5352       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  5355   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
  5356          "Else our work is not yet done");
  5359 // Return a thread-local PLAB recording array, as appropriate.
  5360 void* CMSCollector::get_data_recorder(int thr_num) {
  5361   if (_survivor_plab_array != NULL &&
  5362       (CMSPLABRecordAlways ||
  5363        (_collectorState > Marking && _collectorState < FinalMarking))) {
  5364     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
  5365     ChunkArray* ca = &_survivor_plab_array[thr_num];
  5366     ca->reset();   // clear it so that fresh data is recorded
  5367     return (void*) ca;
  5368   } else {
  5369     return NULL;
  5373 // Reset all the thread-local PLAB recording arrays
  5374 void CMSCollector::reset_survivor_plab_arrays() {
  5375   for (uint i = 0; i < ParallelGCThreads; i++) {
  5376     _survivor_plab_array[i].reset();
  5380 // Merge the per-thread plab arrays into the global survivor chunk
  5381 // array which will provide the partitioning of the survivor space
  5382 // for CMS rescan.
  5383 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
  5384                                               int no_of_gc_threads) {
  5385   assert(_survivor_plab_array  != NULL, "Error");
  5386   assert(_survivor_chunk_array != NULL, "Error");
  5387   assert(_collectorState == FinalMarking, "Error");
  5388   for (int j = 0; j < no_of_gc_threads; j++) {
  5389     _cursor[j] = 0;
  5391   HeapWord* top = surv->top();
  5392   size_t i;
  5393   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
  5394     HeapWord* min_val = top;          // Higher than any PLAB address
  5395     uint      min_tid = 0;            // position of min_val this round
  5396     for (int j = 0; j < no_of_gc_threads; j++) {
  5397       ChunkArray* cur_sca = &_survivor_plab_array[j];
  5398       if (_cursor[j] == cur_sca->end()) {
  5399         continue;
  5401       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
  5402       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
  5403       assert(surv->used_region().contains(cur_val), "Out of bounds value");
  5404       if (cur_val < min_val) {
  5405         min_tid = j;
  5406         min_val = cur_val;
  5407       } else {
  5408         assert(cur_val < top, "All recorded addresses should be less");
  5411     // At this point min_val and min_tid are respectively
  5412     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
  5413     // and the thread (j) that witnesses that address.
  5414     // We record this address in the _survivor_chunk_array[i]
  5415     // and increment _cursor[min_tid] prior to the next round i.
  5416     if (min_val == top) {
  5417       break;
  5419     _survivor_chunk_array[i] = min_val;
  5420     _cursor[min_tid]++;
  5422   // We are all done; record the size of the _survivor_chunk_array
  5423   _survivor_chunk_index = i; // exclusive: [0, i)
  5424   if (PrintCMSStatistics > 0) {
  5425     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
  5427   // Verify that we used up all the recorded entries
  5428   #ifdef ASSERT
  5429     size_t total = 0;
  5430     for (int j = 0; j < no_of_gc_threads; j++) {
  5431       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
  5432       total += _cursor[j];
  5434     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
  5435     // Check that the merged array is in sorted order
  5436     if (total > 0) {
  5437       for (size_t i = 0; i < total - 1; i++) {
  5438         if (PrintCMSStatistics > 0) {
  5439           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
  5440                               i, _survivor_chunk_array[i]);
  5442         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
  5443                "Not sorted");
  5446   #endif // ASSERT
  5449 // Set up the space's par_seq_tasks structure for work claiming
  5450 // for parallel rescan of young gen.
  5451 // See ParRescanTask where this is currently used.
  5452 void
  5453 CMSCollector::
  5454 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
  5455   assert(n_threads > 0, "Unexpected n_threads argument");
  5456   DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  5458   // Eden space
  5460     SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
  5461     assert(!pst->valid(), "Clobbering existing data?");
  5462     // Each valid entry in [0, _eden_chunk_index) represents a task.
  5463     size_t n_tasks = _eden_chunk_index + 1;
  5464     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
  5465     // Sets the condition for completion of the subtask (how many threads
  5466     // need to finish in order to be done).
  5467     pst->set_n_threads(n_threads);
  5468     pst->set_n_tasks((int)n_tasks);
  5471   // Merge the survivor plab arrays into _survivor_chunk_array
  5472   if (_survivor_plab_array != NULL) {
  5473     merge_survivor_plab_arrays(dng->from(), n_threads);
  5474   } else {
  5475     assert(_survivor_chunk_index == 0, "Error");
  5478   // To space
  5480     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
  5481     assert(!pst->valid(), "Clobbering existing data?");
  5482     // Sets the condition for completion of the subtask (how many threads
  5483     // need to finish in order to be done).
  5484     pst->set_n_threads(n_threads);
  5485     pst->set_n_tasks(1);
  5486     assert(pst->valid(), "Error");
  5489   // From space
  5491     SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
  5492     assert(!pst->valid(), "Clobbering existing data?");
  5493     size_t n_tasks = _survivor_chunk_index + 1;
  5494     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
  5495     // Sets the condition for completion of the subtask (how many threads
  5496     // need to finish in order to be done).
  5497     pst->set_n_threads(n_threads);
  5498     pst->set_n_tasks((int)n_tasks);
  5499     assert(pst->valid(), "Error");
  5503 // Parallel version of remark
  5504 void CMSCollector::do_remark_parallel() {
  5505   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5506   FlexibleWorkGang* workers = gch->workers();
  5507   assert(workers != NULL, "Need parallel worker threads.");
  5508   int n_workers = workers->total_workers();
  5509   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  5510   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
  5512   CMSParRemarkTask tsk(this,
  5513     cms_space, perm_space,
  5514     n_workers, workers, task_queues());
  5516   // Set up for parallel process_strong_roots work.
  5517   gch->set_par_threads(n_workers);
  5518   // We won't be iterating over the cards in the card table updating
  5519   // the younger_gen cards, so we shouldn't call the following else
  5520   // the verification code as well as subsequent younger_refs_iterate
  5521   // code would get confused. XXX
  5522   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
  5524   // The young gen rescan work will not be done as part of
  5525   // process_strong_roots (which currently doesn't knw how to
  5526   // parallelize such a scan), but rather will be broken up into
  5527   // a set of parallel tasks (via the sampling that the [abortable]
  5528   // preclean phase did of EdenSpace, plus the [two] tasks of
  5529   // scanning the [two] survivor spaces. Further fine-grain
  5530   // parallelization of the scanning of the survivor spaces
  5531   // themselves, and of precleaning of the younger gen itself
  5532   // is deferred to the future.
  5533   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
  5535   // The dirty card rescan work is broken up into a "sequence"
  5536   // of parallel tasks (per constituent space) that are dynamically
  5537   // claimed by the parallel threads.
  5538   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
  5539   perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
  5541   // It turns out that even when we're using 1 thread, doing the work in a
  5542   // separate thread causes wide variance in run times.  We can't help this
  5543   // in the multi-threaded case, but we special-case n=1 here to get
  5544   // repeatable measurements of the 1-thread overhead of the parallel code.
  5545   if (n_workers > 1) {
  5546     // Make refs discovery MT-safe
  5547     ReferenceProcessorMTMutator mt(ref_processor(), true);
  5548     GenCollectedHeap::StrongRootsScope srs(gch);
  5549     workers->run_task(&tsk);
  5550   } else {
  5551     GenCollectedHeap::StrongRootsScope srs(gch);
  5552     tsk.work(0);
  5554   gch->set_par_threads(0);  // 0 ==> non-parallel.
  5555   // restore, single-threaded for now, any preserved marks
  5556   // as a result of work_q overflow
  5557   restore_preserved_marks_if_any();
  5560 // Non-parallel version of remark
  5561 void CMSCollector::do_remark_non_parallel() {
  5562   ResourceMark rm;
  5563   HandleMark   hm;
  5564   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5565   MarkRefsIntoAndScanClosure
  5566     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
  5567              &_markStack, &_revisitStack, this,
  5568              false /* should_yield */, false /* not precleaning */);
  5569   MarkFromDirtyCardsClosure
  5570     markFromDirtyCardsClosure(this, _span,
  5571                               NULL,  // space is set further below
  5572                               &_markBitMap, &_markStack, &_revisitStack,
  5573                               &mrias_cl);
  5575     TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
  5576     // Iterate over the dirty cards, setting the corresponding bits in the
  5577     // mod union table.
  5579       ModUnionClosure modUnionClosure(&_modUnionTable);
  5580       _ct->ct_bs()->dirty_card_iterate(
  5581                       _cmsGen->used_region(),
  5582                       &modUnionClosure);
  5583       _ct->ct_bs()->dirty_card_iterate(
  5584                       _permGen->used_region(),
  5585                       &modUnionClosure);
  5587     // Having transferred these marks into the modUnionTable, we just need
  5588     // to rescan the marked objects on the dirty cards in the modUnionTable.
  5589     // The initial marking may have been done during an asynchronous
  5590     // collection so there may be dirty bits in the mod-union table.
  5591     const int alignment =
  5592       CardTableModRefBS::card_size * BitsPerWord;
  5594       // ... First handle dirty cards in CMS gen
  5595       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
  5596       MemRegion ur = _cmsGen->used_region();
  5597       HeapWord* lb = ur.start();
  5598       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
  5599       MemRegion cms_span(lb, ub);
  5600       _modUnionTable.dirty_range_iterate_clear(cms_span,
  5601                                                &markFromDirtyCardsClosure);
  5602       verify_work_stacks_empty();
  5603       if (PrintCMSStatistics != 0) {
  5604         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
  5605           markFromDirtyCardsClosure.num_dirty_cards());
  5609       // .. and then repeat for dirty cards in perm gen
  5610       markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
  5611       MemRegion ur = _permGen->used_region();
  5612       HeapWord* lb = ur.start();
  5613       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
  5614       MemRegion perm_span(lb, ub);
  5615       _modUnionTable.dirty_range_iterate_clear(perm_span,
  5616                                                &markFromDirtyCardsClosure);
  5617       verify_work_stacks_empty();
  5618       if (PrintCMSStatistics != 0) {
  5619         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
  5620           markFromDirtyCardsClosure.num_dirty_cards());
  5624   if (VerifyDuringGC &&
  5625       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  5626     HandleMark hm;  // Discard invalid handles created during verification
  5627     Universe::verify(true);
  5630     TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
  5632     verify_work_stacks_empty();
  5634     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  5635     GenCollectedHeap::StrongRootsScope srs(gch);
  5636     gch->gen_process_strong_roots(_cmsGen->level(),
  5637                                   true,  // younger gens as roots
  5638                                   false, // use the local StrongRootsScope
  5639                                   true,  // collecting perm gen
  5640                                   SharedHeap::ScanningOption(roots_scanning_options()),
  5641                                   &mrias_cl,
  5642                                   true,   // walk code active on stacks
  5643                                   NULL);
  5644     assert(should_unload_classes()
  5645            || (roots_scanning_options() & SharedHeap::SO_CodeCache),
  5646            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
  5648   verify_work_stacks_empty();
  5649   // Restore evacuated mark words, if any, used for overflow list links
  5650   if (!CMSOverflowEarlyRestoration) {
  5651     restore_preserved_marks_if_any();
  5653   verify_overflow_empty();
  5656 ////////////////////////////////////////////////////////
  5657 // Parallel Reference Processing Task Proxy Class
  5658 ////////////////////////////////////////////////////////
  5659 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
  5660   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
  5661   CMSCollector*          _collector;
  5662   CMSBitMap*             _mark_bit_map;
  5663   const MemRegion        _span;
  5664   ProcessTask&           _task;
  5666 public:
  5667   CMSRefProcTaskProxy(ProcessTask&     task,
  5668                       CMSCollector*    collector,
  5669                       const MemRegion& span,
  5670                       CMSBitMap*       mark_bit_map,
  5671                       AbstractWorkGang* workers,
  5672                       OopTaskQueueSet* task_queues):
  5673     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
  5674       task_queues),
  5675     _task(task),
  5676     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
  5678       assert(_collector->_span.equals(_span) && !_span.is_empty(),
  5679              "Inconsistency in _span");
  5682   OopTaskQueueSet* task_queues() { return queues(); }
  5684   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  5686   void do_work_steal(int i,
  5687                      CMSParDrainMarkingStackClosure* drain,
  5688                      CMSParKeepAliveClosure* keep_alive,
  5689                      int* seed);
  5691   virtual void work(int i);
  5692 };
  5694 void CMSRefProcTaskProxy::work(int i) {
  5695   assert(_collector->_span.equals(_span), "Inconsistency in _span");
  5696   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
  5697                                         _mark_bit_map,
  5698                                         &_collector->_revisitStack,
  5699                                         work_queue(i));
  5700   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
  5701                                                  _mark_bit_map,
  5702                                                  &_collector->_revisitStack,
  5703                                                  work_queue(i));
  5704   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
  5705   _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
  5706   if (_task.marks_oops_alive()) {
  5707     do_work_steal(i, &par_drain_stack, &par_keep_alive,
  5708                   _collector->hash_seed(i));
  5710   assert(work_queue(i)->size() == 0, "work_queue should be empty");
  5711   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
  5714 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
  5715   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
  5716   EnqueueTask& _task;
  5718 public:
  5719   CMSRefEnqueueTaskProxy(EnqueueTask& task)
  5720     : AbstractGangTask("Enqueue reference objects in parallel"),
  5721       _task(task)
  5722   { }
  5724   virtual void work(int i)
  5726     _task.work(i);
  5728 };
  5730 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
  5731   MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
  5732   OopTaskQueue* work_queue):
  5733    Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
  5734    _span(span),
  5735    _bit_map(bit_map),
  5736    _work_queue(work_queue),
  5737    _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
  5738    _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  5739                         (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
  5740 { }
  5742 // . see if we can share work_queues with ParNew? XXX
  5743 void CMSRefProcTaskProxy::do_work_steal(int i,
  5744   CMSParDrainMarkingStackClosure* drain,
  5745   CMSParKeepAliveClosure* keep_alive,
  5746   int* seed) {
  5747   OopTaskQueue* work_q = work_queue(i);
  5748   NOT_PRODUCT(int num_steals = 0;)
  5749   oop obj_to_scan;
  5751   while (true) {
  5752     // Completely finish any left over work from (an) earlier round(s)
  5753     drain->trim_queue(0);
  5754     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  5755                                          (size_t)ParGCDesiredObjsFromOverflowList);
  5756     // Now check if there's any work in the overflow list
  5757     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
  5758     // only affects the number of attempts made to get work from the
  5759     // overflow list and does not affect the number of workers.  Just
  5760     // pass ParallelGCThreads so this behavior is unchanged.
  5761     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  5762                                                 work_q,
  5763                                                 ParallelGCThreads)) {
  5764       // Found something in global overflow list;
  5765       // not yet ready to go stealing work from others.
  5766       // We'd like to assert(work_q->size() != 0, ...)
  5767       // because we just took work from the overflow list,
  5768       // but of course we can't, since all of that might have
  5769       // been already stolen from us.
  5770       continue;
  5772     // Verify that we have no work before we resort to stealing
  5773     assert(work_q->size() == 0, "Have work, shouldn't steal");
  5774     // Try to steal from other queues that have work
  5775     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  5776       NOT_PRODUCT(num_steals++;)
  5777       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  5778       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  5779       // Do scanning work
  5780       obj_to_scan->oop_iterate(keep_alive);
  5781       // Loop around, finish this work, and try to steal some more
  5782     } else if (terminator()->offer_termination()) {
  5783       break;  // nirvana from the infinite cycle
  5786   NOT_PRODUCT(
  5787     if (PrintCMSStatistics != 0) {
  5788       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  5793 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
  5795   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5796   FlexibleWorkGang* workers = gch->workers();
  5797   assert(workers != NULL, "Need parallel worker threads.");
  5798   CMSRefProcTaskProxy rp_task(task, &_collector,
  5799                               _collector.ref_processor()->span(),
  5800                               _collector.markBitMap(),
  5801                               workers, _collector.task_queues());
  5802   workers->run_task(&rp_task);
  5805 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
  5808   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5809   FlexibleWorkGang* workers = gch->workers();
  5810   assert(workers != NULL, "Need parallel worker threads.");
  5811   CMSRefEnqueueTaskProxy enq_task(task);
  5812   workers->run_task(&enq_task);
  5815 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
  5817   ResourceMark rm;
  5818   HandleMark   hm;
  5820   ReferenceProcessor* rp = ref_processor();
  5821   assert(rp->span().equals(_span), "Spans should be equal");
  5822   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
  5823   // Process weak references.
  5824   rp->setup_policy(clear_all_soft_refs);
  5825   verify_work_stacks_empty();
  5827   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
  5828                                           &_markStack, &_revisitStack,
  5829                                           false /* !preclean */);
  5830   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
  5831                                 _span, &_markBitMap, &_markStack,
  5832                                 &cmsKeepAliveClosure, false /* !preclean */);
  5834     TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
  5835     if (rp->processing_is_mt()) {
  5836       // Set the degree of MT here.  If the discovery is done MT, there
  5837       // may have been a different number of threads doing the discovery
  5838       // and a different number of discovered lists may have Ref objects.
  5839       // That is OK as long as the Reference lists are balanced (see
  5840       // balance_all_queues() and balance_queues()).
  5843       rp->set_mt_degree(ParallelGCThreads);
  5844       CMSRefProcTaskExecutor task_executor(*this);
  5845       rp->process_discovered_references(&_is_alive_closure,
  5846                                         &cmsKeepAliveClosure,
  5847                                         &cmsDrainMarkingStackClosure,
  5848                                         &task_executor);
  5849     } else {
  5850       rp->process_discovered_references(&_is_alive_closure,
  5851                                         &cmsKeepAliveClosure,
  5852                                         &cmsDrainMarkingStackClosure,
  5853                                         NULL);
  5855     verify_work_stacks_empty();
  5858   if (should_unload_classes()) {
  5860       TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
  5862       // Follow SystemDictionary roots and unload classes
  5863       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
  5865       // Follow CodeCache roots and unload any methods marked for unloading
  5866       CodeCache::do_unloading(&_is_alive_closure,
  5867                               &cmsKeepAliveClosure,
  5868                               purged_class);
  5870       cmsDrainMarkingStackClosure.do_void();
  5871       verify_work_stacks_empty();
  5873       // Update subklass/sibling/implementor links in KlassKlass descendants
  5874       assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
  5875       oop k;
  5876       while ((k = _revisitStack.pop()) != NULL) {
  5877         ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
  5878                        &_is_alive_closure,
  5879                        &cmsKeepAliveClosure);
  5881       assert(!ClassUnloading ||
  5882              (_markStack.isEmpty() && overflow_list_is_empty()),
  5883              "Should not have found new reachable objects");
  5884       assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
  5885       cmsDrainMarkingStackClosure.do_void();
  5886       verify_work_stacks_empty();
  5890       TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
  5891       // Now clean up stale oops in SymbolTable and StringTable
  5892       SymbolTable::unlink(&_is_alive_closure);
  5893       StringTable::unlink(&_is_alive_closure);
  5897   verify_work_stacks_empty();
  5898   // Restore any preserved marks as a result of mark stack or
  5899   // work queue overflow
  5900   restore_preserved_marks_if_any();  // done single-threaded for now
  5902   rp->set_enqueuing_is_done(true);
  5903   if (rp->processing_is_mt()) {
  5904     rp->balance_all_queues();
  5905     CMSRefProcTaskExecutor task_executor(*this);
  5906     rp->enqueue_discovered_references(&task_executor);
  5907   } else {
  5908     rp->enqueue_discovered_references(NULL);
  5910   rp->verify_no_references_recorded();
  5911   assert(!rp->discovery_enabled(), "should have been disabled");
  5913   // JVMTI object tagging is based on JNI weak refs. If any of these
  5914   // refs were cleared then JVMTI needs to update its maps and
  5915   // maybe post ObjectFrees to agents.
  5916   JvmtiExport::cms_ref_processing_epilogue();
  5919 #ifndef PRODUCT
  5920 void CMSCollector::check_correct_thread_executing() {
  5921   Thread* t = Thread::current();
  5922   // Only the VM thread or the CMS thread should be here.
  5923   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
  5924          "Unexpected thread type");
  5925   // If this is the vm thread, the foreground process
  5926   // should not be waiting.  Note that _foregroundGCIsActive is
  5927   // true while the foreground collector is waiting.
  5928   if (_foregroundGCShouldWait) {
  5929     // We cannot be the VM thread
  5930     assert(t->is_ConcurrentGC_thread(),
  5931            "Should be CMS thread");
  5932   } else {
  5933     // We can be the CMS thread only if we are in a stop-world
  5934     // phase of CMS collection.
  5935     if (t->is_ConcurrentGC_thread()) {
  5936       assert(_collectorState == InitialMarking ||
  5937              _collectorState == FinalMarking,
  5938              "Should be a stop-world phase");
  5939       // The CMS thread should be holding the CMS_token.
  5940       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  5941              "Potential interference with concurrently "
  5942              "executing VM thread");
  5946 #endif
  5948 void CMSCollector::sweep(bool asynch) {
  5949   assert(_collectorState == Sweeping, "just checking");
  5950   check_correct_thread_executing();
  5951   verify_work_stacks_empty();
  5952   verify_overflow_empty();
  5953   increment_sweep_count();
  5954   TraceCMSMemoryManagerStats tms(_collectorState);
  5956   _inter_sweep_timer.stop();
  5957   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
  5958   size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
  5960   // PermGen verification support: If perm gen sweeping is disabled in
  5961   // this cycle, we preserve the perm gen object "deadness" information
  5962   // in the perm_gen_verify_bit_map. In order to do that we traverse
  5963   // all blocks in perm gen and mark all dead objects.
  5964   if (verifying() && !should_unload_classes()) {
  5965     assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
  5966            "Should have already been allocated");
  5967     MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
  5968                                markBitMap(), perm_gen_verify_bit_map());
  5969     if (asynch) {
  5970       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
  5971                                bitMapLock());
  5972       _permGen->cmsSpace()->blk_iterate(&mdo);
  5973     } else {
  5974       // In the case of synchronous sweep, we already have
  5975       // the requisite locks/tokens.
  5976       _permGen->cmsSpace()->blk_iterate(&mdo);
  5980   assert(!_intra_sweep_timer.is_active(), "Should not be active");
  5981   _intra_sweep_timer.reset();
  5982   _intra_sweep_timer.start();
  5983   if (asynch) {
  5984     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  5985     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
  5986     // First sweep the old gen then the perm gen
  5988       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
  5989                                bitMapLock());
  5990       sweepWork(_cmsGen, asynch);
  5993     // Now repeat for perm gen
  5994     if (should_unload_classes()) {
  5995       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
  5996                              bitMapLock());
  5997       sweepWork(_permGen, asynch);
  6000     // Update Universe::_heap_*_at_gc figures.
  6001     // We need all the free list locks to make the abstract state
  6002     // transition from Sweeping to Resetting. See detailed note
  6003     // further below.
  6005       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
  6006                                _permGen->freelistLock());
  6007       // Update heap occupancy information which is used as
  6008       // input to soft ref clearing policy at the next gc.
  6009       Universe::update_heap_info_at_gc();
  6010       _collectorState = Resizing;
  6012   } else {
  6013     // already have needed locks
  6014     sweepWork(_cmsGen,  asynch);
  6016     if (should_unload_classes()) {
  6017       sweepWork(_permGen, asynch);
  6019     // Update heap occupancy information which is used as
  6020     // input to soft ref clearing policy at the next gc.
  6021     Universe::update_heap_info_at_gc();
  6022     _collectorState = Resizing;
  6024   verify_work_stacks_empty();
  6025   verify_overflow_empty();
  6027   _intra_sweep_timer.stop();
  6028   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
  6030   _inter_sweep_timer.reset();
  6031   _inter_sweep_timer.start();
  6033   update_time_of_last_gc(os::javaTimeMillis());
  6035   // NOTE on abstract state transitions:
  6036   // Mutators allocate-live and/or mark the mod-union table dirty
  6037   // based on the state of the collection.  The former is done in
  6038   // the interval [Marking, Sweeping] and the latter in the interval
  6039   // [Marking, Sweeping).  Thus the transitions into the Marking state
  6040   // and out of the Sweeping state must be synchronously visible
  6041   // globally to the mutators.
  6042   // The transition into the Marking state happens with the world
  6043   // stopped so the mutators will globally see it.  Sweeping is
  6044   // done asynchronously by the background collector so the transition
  6045   // from the Sweeping state to the Resizing state must be done
  6046   // under the freelistLock (as is the check for whether to
  6047   // allocate-live and whether to dirty the mod-union table).
  6048   assert(_collectorState == Resizing, "Change of collector state to"
  6049     " Resizing must be done under the freelistLocks (plural)");
  6051   // Now that sweeping has been completed, we clear
  6052   // the incremental_collection_failed flag,
  6053   // thus inviting a younger gen collection to promote into
  6054   // this generation. If such a promotion may still fail,
  6055   // the flag will be set again when a young collection is
  6056   // attempted.
  6057   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6058   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
  6059   gch->update_full_collections_completed(_collection_count_start);
  6062 // FIX ME!!! Looks like this belongs in CFLSpace, with
  6063 // CMSGen merely delegating to it.
  6064 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
  6065   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
  6066   HeapWord*  minAddr        = _cmsSpace->bottom();
  6067   HeapWord*  largestAddr    =
  6068     (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
  6069   if (largestAddr == NULL) {
  6070     // The dictionary appears to be empty.  In this case
  6071     // try to coalesce at the end of the heap.
  6072     largestAddr = _cmsSpace->end();
  6074   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
  6075   size_t nearLargestOffset =
  6076     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
  6077   if (PrintFLSStatistics != 0) {
  6078     gclog_or_tty->print_cr(
  6079       "CMS: Large Block: " PTR_FORMAT ";"
  6080       " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
  6081       largestAddr,
  6082       _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
  6084   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
  6087 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
  6088   return addr >= _cmsSpace->nearLargestChunk();
  6091 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
  6092   return _cmsSpace->find_chunk_at_end();
  6095 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
  6096                                                     bool full) {
  6097   // The next lower level has been collected.  Gather any statistics
  6098   // that are of interest at this point.
  6099   if (!full && (current_level + 1) == level()) {
  6100     // Gather statistics on the young generation collection.
  6101     collector()->stats().record_gc0_end(used());
  6105 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
  6106   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6107   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
  6108     "Wrong type of heap");
  6109   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
  6110     gch->gen_policy()->size_policy();
  6111   assert(sp->is_gc_cms_adaptive_size_policy(),
  6112     "Wrong type of size policy");
  6113   return sp;
  6116 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
  6117   if (PrintGCDetails && Verbose) {
  6118     gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
  6120   _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
  6121   _debug_collection_type =
  6122     (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
  6123   if (PrintGCDetails && Verbose) {
  6124     gclog_or_tty->print_cr("to %d ", _debug_collection_type);
  6128 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
  6129   bool asynch) {
  6130   // We iterate over the space(s) underlying this generation,
  6131   // checking the mark bit map to see if the bits corresponding
  6132   // to specific blocks are marked or not. Blocks that are
  6133   // marked are live and are not swept up. All remaining blocks
  6134   // are swept up, with coalescing on-the-fly as we sweep up
  6135   // contiguous free and/or garbage blocks:
  6136   // We need to ensure that the sweeper synchronizes with allocators
  6137   // and stop-the-world collectors. In particular, the following
  6138   // locks are used:
  6139   // . CMS token: if this is held, a stop the world collection cannot occur
  6140   // . freelistLock: if this is held no allocation can occur from this
  6141   //                 generation by another thread
  6142   // . bitMapLock: if this is held, no other thread can access or update
  6143   //
  6145   // Note that we need to hold the freelistLock if we use
  6146   // block iterate below; else the iterator might go awry if
  6147   // a mutator (or promotion) causes block contents to change
  6148   // (for instance if the allocator divvies up a block).
  6149   // If we hold the free list lock, for all practical purposes
  6150   // young generation GC's can't occur (they'll usually need to
  6151   // promote), so we might as well prevent all young generation
  6152   // GC's while we do a sweeping step. For the same reason, we might
  6153   // as well take the bit map lock for the entire duration
  6155   // check that we hold the requisite locks
  6156   assert(have_cms_token(), "Should hold cms token");
  6157   assert(   (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
  6158          || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
  6159         "Should possess CMS token to sweep");
  6160   assert_lock_strong(gen->freelistLock());
  6161   assert_lock_strong(bitMapLock());
  6163   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
  6164   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
  6165   gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
  6166                                       _inter_sweep_estimate.padded_average(),
  6167                                       _intra_sweep_estimate.padded_average());
  6168   gen->setNearLargestChunk();
  6171     SweepClosure sweepClosure(this, gen, &_markBitMap,
  6172                             CMSYield && asynch);
  6173     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
  6174     // We need to free-up/coalesce garbage/blocks from a
  6175     // co-terminal free run. This is done in the SweepClosure
  6176     // destructor; so, do not remove this scope, else the
  6177     // end-of-sweep-census below will be off by a little bit.
  6179   gen->cmsSpace()->sweep_completed();
  6180   gen->cmsSpace()->endSweepFLCensus(sweep_count());
  6181   if (should_unload_classes()) {                // unloaded classes this cycle,
  6182     _concurrent_cycles_since_last_unload = 0;   // ... reset count
  6183   } else {                                      // did not unload classes,
  6184     _concurrent_cycles_since_last_unload++;     // ... increment count
  6188 // Reset CMS data structures (for now just the marking bit map)
  6189 // preparatory for the next cycle.
  6190 void CMSCollector::reset(bool asynch) {
  6191   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6192   CMSAdaptiveSizePolicy* sp = size_policy();
  6193   AdaptiveSizePolicyOutput(sp, gch->total_collections());
  6194   if (asynch) {
  6195     CMSTokenSyncWithLocks ts(true, bitMapLock());
  6197     // If the state is not "Resetting", the foreground  thread
  6198     // has done a collection and the resetting.
  6199     if (_collectorState != Resetting) {
  6200       assert(_collectorState == Idling, "The state should only change"
  6201         " because the foreground collector has finished the collection");
  6202       return;
  6205     // Clear the mark bitmap (no grey objects to start with)
  6206     // for the next cycle.
  6207     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6208     CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
  6210     HeapWord* curAddr = _markBitMap.startWord();
  6211     while (curAddr < _markBitMap.endWord()) {
  6212       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
  6213       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
  6214       _markBitMap.clear_large_range(chunk);
  6215       if (ConcurrentMarkSweepThread::should_yield() &&
  6216           !foregroundGCIsActive() &&
  6217           CMSYield) {
  6218         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6219                "CMS thread should hold CMS token");
  6220         assert_lock_strong(bitMapLock());
  6221         bitMapLock()->unlock();
  6222         ConcurrentMarkSweepThread::desynchronize(true);
  6223         ConcurrentMarkSweepThread::acknowledge_yield_request();
  6224         stopTimer();
  6225         if (PrintCMSStatistics != 0) {
  6226           incrementYields();
  6228         icms_wait();
  6230         // See the comment in coordinator_yield()
  6231         for (unsigned i = 0; i < CMSYieldSleepCount &&
  6232                          ConcurrentMarkSweepThread::should_yield() &&
  6233                          !CMSCollector::foregroundGCIsActive(); ++i) {
  6234           os::sleep(Thread::current(), 1, false);
  6235           ConcurrentMarkSweepThread::acknowledge_yield_request();
  6238         ConcurrentMarkSweepThread::synchronize(true);
  6239         bitMapLock()->lock_without_safepoint_check();
  6240         startTimer();
  6242       curAddr = chunk.end();
  6244     // A successful mostly concurrent collection has been done.
  6245     // Because only the full (i.e., concurrent mode failure) collections
  6246     // are being measured for gc overhead limits, clean the "near" flag
  6247     // and count.
  6248     sp->reset_gc_overhead_limit_count();
  6249     _collectorState = Idling;
  6250   } else {
  6251     // already have the lock
  6252     assert(_collectorState == Resetting, "just checking");
  6253     assert_lock_strong(bitMapLock());
  6254     _markBitMap.clear_all();
  6255     _collectorState = Idling;
  6258   // Stop incremental mode after a cycle completes, so that any future cycles
  6259   // are triggered by allocation.
  6260   stop_icms();
  6262   NOT_PRODUCT(
  6263     if (RotateCMSCollectionTypes) {
  6264       _cmsGen->rotate_debug_collection_type();
  6269 void CMSCollector::do_CMS_operation(CMS_op_type op) {
  6270   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
  6271   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6272   TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
  6273   TraceCollectorStats tcs(counters());
  6275   switch (op) {
  6276     case CMS_op_checkpointRootsInitial: {
  6277       checkpointRootsInitial(true);       // asynch
  6278       if (PrintGC) {
  6279         _cmsGen->printOccupancy("initial-mark");
  6281       break;
  6283     case CMS_op_checkpointRootsFinal: {
  6284       checkpointRootsFinal(true,    // asynch
  6285                            false,   // !clear_all_soft_refs
  6286                            false);  // !init_mark_was_synchronous
  6287       if (PrintGC) {
  6288         _cmsGen->printOccupancy("remark");
  6290       break;
  6292     default:
  6293       fatal("No such CMS_op");
  6297 #ifndef PRODUCT
  6298 size_t const CMSCollector::skip_header_HeapWords() {
  6299   return FreeChunk::header_size();
  6302 // Try and collect here conditions that should hold when
  6303 // CMS thread is exiting. The idea is that the foreground GC
  6304 // thread should not be blocked if it wants to terminate
  6305 // the CMS thread and yet continue to run the VM for a while
  6306 // after that.
  6307 void CMSCollector::verify_ok_to_terminate() const {
  6308   assert(Thread::current()->is_ConcurrentGC_thread(),
  6309          "should be called by CMS thread");
  6310   assert(!_foregroundGCShouldWait, "should be false");
  6311   // We could check here that all the various low-level locks
  6312   // are not held by the CMS thread, but that is overkill; see
  6313   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
  6314   // is checked.
  6316 #endif
  6318 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
  6319    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
  6320           "missing Printezis mark?");
  6321   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6322   size_t size = pointer_delta(nextOneAddr + 1, addr);
  6323   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6324          "alignment problem");
  6325   assert(size >= 3, "Necessary for Printezis marks to work");
  6326   return size;
  6329 // A variant of the above (block_size_using_printezis_bits()) except
  6330 // that we return 0 if the P-bits are not yet set.
  6331 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
  6332   if (_markBitMap.isMarked(addr)) {
  6333     assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
  6334     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6335     size_t size = pointer_delta(nextOneAddr + 1, addr);
  6336     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6337            "alignment problem");
  6338     assert(size >= 3, "Necessary for Printezis marks to work");
  6339     return size;
  6340   } else {
  6341     assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
  6342     return 0;
  6346 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
  6347   size_t sz = 0;
  6348   oop p = (oop)addr;
  6349   if (p->klass_or_null() != NULL && p->is_parsable()) {
  6350     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
  6351   } else {
  6352     sz = block_size_using_printezis_bits(addr);
  6354   assert(sz > 0, "size must be nonzero");
  6355   HeapWord* next_block = addr + sz;
  6356   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
  6357                                              CardTableModRefBS::card_size);
  6358   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
  6359          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
  6360          "must be different cards");
  6361   return next_card;
  6365 // CMS Bit Map Wrapper /////////////////////////////////////////
  6367 // Construct a CMS bit map infrastructure, but don't create the
  6368 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
  6369 // further below.
  6370 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
  6371   _bm(),
  6372   _shifter(shifter),
  6373   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
  6375   _bmStartWord = 0;
  6376   _bmWordSize  = 0;
  6379 bool CMSBitMap::allocate(MemRegion mr) {
  6380   _bmStartWord = mr.start();
  6381   _bmWordSize  = mr.word_size();
  6382   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
  6383                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
  6384   if (!brs.is_reserved()) {
  6385     warning("CMS bit map allocation failure");
  6386     return false;
  6388   // For now we'll just commit all of the bit map up fromt.
  6389   // Later on we'll try to be more parsimonious with swap.
  6390   if (!_virtual_space.initialize(brs, brs.size())) {
  6391     warning("CMS bit map backing store failure");
  6392     return false;
  6394   assert(_virtual_space.committed_size() == brs.size(),
  6395          "didn't reserve backing store for all of CMS bit map?");
  6396   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
  6397   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
  6398          _bmWordSize, "inconsistency in bit map sizing");
  6399   _bm.set_size(_bmWordSize >> _shifter);
  6401   // bm.clear(); // can we rely on getting zero'd memory? verify below
  6402   assert(isAllClear(),
  6403          "Expected zero'd memory from ReservedSpace constructor");
  6404   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
  6405          "consistency check");
  6406   return true;
  6409 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
  6410   HeapWord *next_addr, *end_addr, *last_addr;
  6411   assert_locked();
  6412   assert(covers(mr), "out-of-range error");
  6413   // XXX assert that start and end are appropriately aligned
  6414   for (next_addr = mr.start(), end_addr = mr.end();
  6415        next_addr < end_addr; next_addr = last_addr) {
  6416     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
  6417     last_addr = dirty_region.end();
  6418     if (!dirty_region.is_empty()) {
  6419       cl->do_MemRegion(dirty_region);
  6420     } else {
  6421       assert(last_addr == end_addr, "program logic");
  6422       return;
  6427 #ifndef PRODUCT
  6428 void CMSBitMap::assert_locked() const {
  6429   CMSLockVerifier::assert_locked(lock());
  6432 bool CMSBitMap::covers(MemRegion mr) const {
  6433   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  6434   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
  6435          "size inconsistency");
  6436   return (mr.start() >= _bmStartWord) &&
  6437          (mr.end()   <= endWord());
  6440 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
  6441     return (start >= _bmStartWord && (start + size) <= endWord());
  6444 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
  6445   // verify that there are no 1 bits in the interval [left, right)
  6446   FalseBitMapClosure falseBitMapClosure;
  6447   iterate(&falseBitMapClosure, left, right);
  6450 void CMSBitMap::region_invariant(MemRegion mr)
  6452   assert_locked();
  6453   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
  6454   assert(!mr.is_empty(), "unexpected empty region");
  6455   assert(covers(mr), "mr should be covered by bit map");
  6456   // convert address range into offset range
  6457   size_t start_ofs = heapWordToOffset(mr.start());
  6458   // Make sure that end() is appropriately aligned
  6459   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
  6460                         (1 << (_shifter+LogHeapWordSize))),
  6461          "Misaligned mr.end()");
  6462   size_t end_ofs   = heapWordToOffset(mr.end());
  6463   assert(end_ofs > start_ofs, "Should mark at least one bit");
  6466 #endif
  6468 bool CMSMarkStack::allocate(size_t size) {
  6469   // allocate a stack of the requisite depth
  6470   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6471                    size * sizeof(oop)));
  6472   if (!rs.is_reserved()) {
  6473     warning("CMSMarkStack allocation failure");
  6474     return false;
  6476   if (!_virtual_space.initialize(rs, rs.size())) {
  6477     warning("CMSMarkStack backing store failure");
  6478     return false;
  6480   assert(_virtual_space.committed_size() == rs.size(),
  6481          "didn't reserve backing store for all of CMS stack?");
  6482   _base = (oop*)(_virtual_space.low());
  6483   _index = 0;
  6484   _capacity = size;
  6485   NOT_PRODUCT(_max_depth = 0);
  6486   return true;
  6489 // XXX FIX ME !!! In the MT case we come in here holding a
  6490 // leaf lock. For printing we need to take a further lock
  6491 // which has lower rank. We need to recallibrate the two
  6492 // lock-ranks involved in order to be able to rpint the
  6493 // messages below. (Or defer the printing to the caller.
  6494 // For now we take the expedient path of just disabling the
  6495 // messages for the problematic case.)
  6496 void CMSMarkStack::expand() {
  6497   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
  6498   if (_capacity == MarkStackSizeMax) {
  6499     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6500       // We print a warning message only once per CMS cycle.
  6501       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
  6503     return;
  6505   // Double capacity if possible
  6506   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
  6507   // Do not give up existing stack until we have managed to
  6508   // get the double capacity that we desired.
  6509   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6510                    new_capacity * sizeof(oop)));
  6511   if (rs.is_reserved()) {
  6512     // Release the backing store associated with old stack
  6513     _virtual_space.release();
  6514     // Reinitialize virtual space for new stack
  6515     if (!_virtual_space.initialize(rs, rs.size())) {
  6516       fatal("Not enough swap for expanded marking stack");
  6518     _base = (oop*)(_virtual_space.low());
  6519     _index = 0;
  6520     _capacity = new_capacity;
  6521   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6522     // Failed to double capacity, continue;
  6523     // we print a detail message only once per CMS cycle.
  6524     gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
  6525             SIZE_FORMAT"K",
  6526             _capacity / K, new_capacity / K);
  6531 // Closures
  6532 // XXX: there seems to be a lot of code  duplication here;
  6533 // should refactor and consolidate common code.
  6535 // This closure is used to mark refs into the CMS generation in
  6536 // the CMS bit map. Called at the first checkpoint. This closure
  6537 // assumes that we do not need to re-mark dirty cards; if the CMS
  6538 // generation on which this is used is not an oldest (modulo perm gen)
  6539 // generation then this will lose younger_gen cards!
  6541 MarkRefsIntoClosure::MarkRefsIntoClosure(
  6542   MemRegion span, CMSBitMap* bitMap):
  6543     _span(span),
  6544     _bitMap(bitMap)
  6546     assert(_ref_processor == NULL, "deliberately left NULL");
  6547     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
  6550 void MarkRefsIntoClosure::do_oop(oop obj) {
  6551   // if p points into _span, then mark corresponding bit in _markBitMap
  6552   assert(obj->is_oop(), "expected an oop");
  6553   HeapWord* addr = (HeapWord*)obj;
  6554   if (_span.contains(addr)) {
  6555     // this should be made more efficient
  6556     _bitMap->mark(addr);
  6560 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
  6561 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
  6563 // A variant of the above, used for CMS marking verification.
  6564 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
  6565   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
  6566     _span(span),
  6567     _verification_bm(verification_bm),
  6568     _cms_bm(cms_bm)
  6570     assert(_ref_processor == NULL, "deliberately left NULL");
  6571     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
  6574 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
  6575   // if p points into _span, then mark corresponding bit in _markBitMap
  6576   assert(obj->is_oop(), "expected an oop");
  6577   HeapWord* addr = (HeapWord*)obj;
  6578   if (_span.contains(addr)) {
  6579     _verification_bm->mark(addr);
  6580     if (!_cms_bm->isMarked(addr)) {
  6581       oop(addr)->print();
  6582       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
  6583       fatal("... aborting");
  6588 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6589 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6591 //////////////////////////////////////////////////
  6592 // MarkRefsIntoAndScanClosure
  6593 //////////////////////////////////////////////////
  6595 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
  6596                                                        ReferenceProcessor* rp,
  6597                                                        CMSBitMap* bit_map,
  6598                                                        CMSBitMap* mod_union_table,
  6599                                                        CMSMarkStack*  mark_stack,
  6600                                                        CMSMarkStack*  revisit_stack,
  6601                                                        CMSCollector* collector,
  6602                                                        bool should_yield,
  6603                                                        bool concurrent_precleaning):
  6604   _collector(collector),
  6605   _span(span),
  6606   _bit_map(bit_map),
  6607   _mark_stack(mark_stack),
  6608   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
  6609                       mark_stack, revisit_stack, concurrent_precleaning),
  6610   _yield(should_yield),
  6611   _concurrent_precleaning(concurrent_precleaning),
  6612   _freelistLock(NULL)
  6614   _ref_processor = rp;
  6615   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  6618 // This closure is used to mark refs into the CMS generation at the
  6619 // second (final) checkpoint, and to scan and transitively follow
  6620 // the unmarked oops. It is also used during the concurrent precleaning
  6621 // phase while scanning objects on dirty cards in the CMS generation.
  6622 // The marks are made in the marking bit map and the marking stack is
  6623 // used for keeping the (newly) grey objects during the scan.
  6624 // The parallel version (Par_...) appears further below.
  6625 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  6626   if (obj != NULL) {
  6627     assert(obj->is_oop(), "expected an oop");
  6628     HeapWord* addr = (HeapWord*)obj;
  6629     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  6630     assert(_collector->overflow_list_is_empty(),
  6631            "overflow list should be empty");
  6632     if (_span.contains(addr) &&
  6633         !_bit_map->isMarked(addr)) {
  6634       // mark bit map (object is now grey)
  6635       _bit_map->mark(addr);
  6636       // push on marking stack (stack should be empty), and drain the
  6637       // stack by applying this closure to the oops in the oops popped
  6638       // from the stack (i.e. blacken the grey objects)
  6639       bool res = _mark_stack->push(obj);
  6640       assert(res, "Should have space to push on empty stack");
  6641       do {
  6642         oop new_oop = _mark_stack->pop();
  6643         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  6644         assert(new_oop->is_parsable(), "Found unparsable oop");
  6645         assert(_bit_map->isMarked((HeapWord*)new_oop),
  6646                "only grey objects on this stack");
  6647         // iterate over the oops in this oop, marking and pushing
  6648         // the ones in CMS heap (i.e. in _span).
  6649         new_oop->oop_iterate(&_pushAndMarkClosure);
  6650         // check if it's time to yield
  6651         do_yield_check();
  6652       } while (!_mark_stack->isEmpty() ||
  6653                (!_concurrent_precleaning && take_from_overflow_list()));
  6654         // if marking stack is empty, and we are not doing this
  6655         // during precleaning, then check the overflow list
  6657     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  6658     assert(_collector->overflow_list_is_empty(),
  6659            "overflow list was drained above");
  6660     // We could restore evacuated mark words, if any, used for
  6661     // overflow list links here because the overflow list is
  6662     // provably empty here. That would reduce the maximum
  6663     // size requirements for preserved_{oop,mark}_stack.
  6664     // But we'll just postpone it until we are all done
  6665     // so we can just stream through.
  6666     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
  6667       _collector->restore_preserved_marks_if_any();
  6668       assert(_collector->no_preserved_marks(), "No preserved marks");
  6670     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
  6671            "All preserved marks should have been restored above");
  6675 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6676 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6678 void MarkRefsIntoAndScanClosure::do_yield_work() {
  6679   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6680          "CMS thread should hold CMS token");
  6681   assert_lock_strong(_freelistLock);
  6682   assert_lock_strong(_bit_map->lock());
  6683   // relinquish the free_list_lock and bitMaplock()
  6684   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  6685   _bit_map->lock()->unlock();
  6686   _freelistLock->unlock();
  6687   ConcurrentMarkSweepThread::desynchronize(true);
  6688   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6689   _collector->stopTimer();
  6690   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6691   if (PrintCMSStatistics != 0) {
  6692     _collector->incrementYields();
  6694   _collector->icms_wait();
  6696   // See the comment in coordinator_yield()
  6697   for (unsigned i = 0;
  6698        i < CMSYieldSleepCount &&
  6699        ConcurrentMarkSweepThread::should_yield() &&
  6700        !CMSCollector::foregroundGCIsActive();
  6701        ++i) {
  6702     os::sleep(Thread::current(), 1, false);
  6703     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6706   ConcurrentMarkSweepThread::synchronize(true);
  6707   _freelistLock->lock_without_safepoint_check();
  6708   _bit_map->lock()->lock_without_safepoint_check();
  6709   _collector->startTimer();
  6712 ///////////////////////////////////////////////////////////
  6713 // Par_MarkRefsIntoAndScanClosure: a parallel version of
  6714 //                                 MarkRefsIntoAndScanClosure
  6715 ///////////////////////////////////////////////////////////
  6716 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
  6717   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
  6718   CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack*  revisit_stack):
  6719   _span(span),
  6720   _bit_map(bit_map),
  6721   _work_queue(work_queue),
  6722   _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  6723                        (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
  6724   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
  6725                           revisit_stack)
  6727   _ref_processor = rp;
  6728   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  6731 // This closure is used to mark refs into the CMS generation at the
  6732 // second (final) checkpoint, and to scan and transitively follow
  6733 // the unmarked oops. The marks are made in the marking bit map and
  6734 // the work_queue is used for keeping the (newly) grey objects during
  6735 // the scan phase whence they are also available for stealing by parallel
  6736 // threads. Since the marking bit map is shared, updates are
  6737 // synchronized (via CAS).
  6738 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  6739   if (obj != NULL) {
  6740     // Ignore mark word because this could be an already marked oop
  6741     // that may be chained at the end of the overflow list.
  6742     assert(obj->is_oop(true), "expected an oop");
  6743     HeapWord* addr = (HeapWord*)obj;
  6744     if (_span.contains(addr) &&
  6745         !_bit_map->isMarked(addr)) {
  6746       // mark bit map (object will become grey):
  6747       // It is possible for several threads to be
  6748       // trying to "claim" this object concurrently;
  6749       // the unique thread that succeeds in marking the
  6750       // object first will do the subsequent push on
  6751       // to the work queue (or overflow list).
  6752       if (_bit_map->par_mark(addr)) {
  6753         // push on work_queue (which may not be empty), and trim the
  6754         // queue to an appropriate length by applying this closure to
  6755         // the oops in the oops popped from the stack (i.e. blacken the
  6756         // grey objects)
  6757         bool res = _work_queue->push(obj);
  6758         assert(res, "Low water mark should be less than capacity?");
  6759         trim_queue(_low_water_mark);
  6760       } // Else, another thread claimed the object
  6765 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6766 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  6768 // This closure is used to rescan the marked objects on the dirty cards
  6769 // in the mod union table and the card table proper.
  6770 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
  6771   oop p, MemRegion mr) {
  6773   size_t size = 0;
  6774   HeapWord* addr = (HeapWord*)p;
  6775   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6776   assert(_span.contains(addr), "we are scanning the CMS generation");
  6777   // check if it's time to yield
  6778   if (do_yield_check()) {
  6779     // We yielded for some foreground stop-world work,
  6780     // and we have been asked to abort this ongoing preclean cycle.
  6781     return 0;
  6783   if (_bitMap->isMarked(addr)) {
  6784     // it's marked; is it potentially uninitialized?
  6785     if (p->klass_or_null() != NULL) {
  6786       // If is_conc_safe is false, the object may be undergoing
  6787       // change by the VM outside a safepoint.  Don't try to
  6788       // scan it, but rather leave it for the remark phase.
  6789       if (CMSPermGenPrecleaningEnabled &&
  6790           (!p->is_conc_safe() || !p->is_parsable())) {
  6791         // Signal precleaning to redirty the card since
  6792         // the klass pointer is already installed.
  6793         assert(size == 0, "Initial value");
  6794       } else {
  6795         assert(p->is_parsable(), "must be parsable.");
  6796         // an initialized object; ignore mark word in verification below
  6797         // since we are running concurrent with mutators
  6798         assert(p->is_oop(true), "should be an oop");
  6799         if (p->is_objArray()) {
  6800           // objArrays are precisely marked; restrict scanning
  6801           // to dirty cards only.
  6802           size = CompactibleFreeListSpace::adjustObjectSize(
  6803                    p->oop_iterate(_scanningClosure, mr));
  6804         } else {
  6805           // A non-array may have been imprecisely marked; we need
  6806           // to scan object in its entirety.
  6807           size = CompactibleFreeListSpace::adjustObjectSize(
  6808                    p->oop_iterate(_scanningClosure));
  6810         #ifdef DEBUG
  6811           size_t direct_size =
  6812             CompactibleFreeListSpace::adjustObjectSize(p->size());
  6813           assert(size == direct_size, "Inconsistency in size");
  6814           assert(size >= 3, "Necessary for Printezis marks to work");
  6815           if (!_bitMap->isMarked(addr+1)) {
  6816             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
  6817           } else {
  6818             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
  6819             assert(_bitMap->isMarked(addr+size-1),
  6820                    "inconsistent Printezis mark");
  6822         #endif // DEBUG
  6824     } else {
  6825       // an unitialized object
  6826       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
  6827       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  6828       size = pointer_delta(nextOneAddr + 1, addr);
  6829       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6830              "alignment problem");
  6831       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
  6832       // will dirty the card when the klass pointer is installed in the
  6833       // object (signalling the completion of initialization).
  6835   } else {
  6836     // Either a not yet marked object or an uninitialized object
  6837     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  6838       // An uninitialized object, skip to the next card, since
  6839       // we may not be able to read its P-bits yet.
  6840       assert(size == 0, "Initial value");
  6841     } else {
  6842       // An object not (yet) reached by marking: we merely need to
  6843       // compute its size so as to go look at the next block.
  6844       assert(p->is_oop(true), "should be an oop");
  6845       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
  6848   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6849   return size;
  6852 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
  6853   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6854          "CMS thread should hold CMS token");
  6855   assert_lock_strong(_freelistLock);
  6856   assert_lock_strong(_bitMap->lock());
  6857   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  6858   // relinquish the free_list_lock and bitMaplock()
  6859   _bitMap->lock()->unlock();
  6860   _freelistLock->unlock();
  6861   ConcurrentMarkSweepThread::desynchronize(true);
  6862   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6863   _collector->stopTimer();
  6864   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6865   if (PrintCMSStatistics != 0) {
  6866     _collector->incrementYields();
  6868   _collector->icms_wait();
  6870   // See the comment in coordinator_yield()
  6871   for (unsigned i = 0; i < CMSYieldSleepCount &&
  6872                    ConcurrentMarkSweepThread::should_yield() &&
  6873                    !CMSCollector::foregroundGCIsActive(); ++i) {
  6874     os::sleep(Thread::current(), 1, false);
  6875     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6878   ConcurrentMarkSweepThread::synchronize(true);
  6879   _freelistLock->lock_without_safepoint_check();
  6880   _bitMap->lock()->lock_without_safepoint_check();
  6881   _collector->startTimer();
  6885 //////////////////////////////////////////////////////////////////
  6886 // SurvivorSpacePrecleanClosure
  6887 //////////////////////////////////////////////////////////////////
  6888 // This (single-threaded) closure is used to preclean the oops in
  6889 // the survivor spaces.
  6890 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
  6892   HeapWord* addr = (HeapWord*)p;
  6893   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  6894   assert(!_span.contains(addr), "we are scanning the survivor spaces");
  6895   assert(p->klass_or_null() != NULL, "object should be initializd");
  6896   assert(p->is_parsable(), "must be parsable.");
  6897   // an initialized object; ignore mark word in verification below
  6898   // since we are running concurrent with mutators
  6899   assert(p->is_oop(true), "should be an oop");
  6900   // Note that we do not yield while we iterate over
  6901   // the interior oops of p, pushing the relevant ones
  6902   // on our marking stack.
  6903   size_t size = p->oop_iterate(_scanning_closure);
  6904   do_yield_check();
  6905   // Observe that below, we do not abandon the preclean
  6906   // phase as soon as we should; rather we empty the
  6907   // marking stack before returning. This is to satisfy
  6908   // some existing assertions. In general, it may be a
  6909   // good idea to abort immediately and complete the marking
  6910   // from the grey objects at a later time.
  6911   while (!_mark_stack->isEmpty()) {
  6912     oop new_oop = _mark_stack->pop();
  6913     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  6914     assert(new_oop->is_parsable(), "Found unparsable oop");
  6915     assert(_bit_map->isMarked((HeapWord*)new_oop),
  6916            "only grey objects on this stack");
  6917     // iterate over the oops in this oop, marking and pushing
  6918     // the ones in CMS heap (i.e. in _span).
  6919     new_oop->oop_iterate(_scanning_closure);
  6920     // check if it's time to yield
  6921     do_yield_check();
  6923   unsigned int after_count =
  6924     GenCollectedHeap::heap()->total_collections();
  6925   bool abort = (_before_count != after_count) ||
  6926                _collector->should_abort_preclean();
  6927   return abort ? 0 : size;
  6930 void SurvivorSpacePrecleanClosure::do_yield_work() {
  6931   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6932          "CMS thread should hold CMS token");
  6933   assert_lock_strong(_bit_map->lock());
  6934   DEBUG_ONLY(RememberKlassesChecker smx(false);)
  6935   // Relinquish the bit map lock
  6936   _bit_map->lock()->unlock();
  6937   ConcurrentMarkSweepThread::desynchronize(true);
  6938   ConcurrentMarkSweepThread::acknowledge_yield_request();
  6939   _collector->stopTimer();
  6940   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  6941   if (PrintCMSStatistics != 0) {
  6942     _collector->incrementYields();
  6944   _collector->icms_wait();
  6946   // See the comment in coordinator_yield()
  6947   for (unsigned i = 0; i < CMSYieldSleepCount &&
  6948                        ConcurrentMarkSweepThread::should_yield() &&
  6949                        !CMSCollector::foregroundGCIsActive(); ++i) {
  6950     os::sleep(Thread::current(), 1, false);
  6951     ConcurrentMarkSweepThread::acknowledge_yield_request();
  6954   ConcurrentMarkSweepThread::synchronize(true);
  6955   _bit_map->lock()->lock_without_safepoint_check();
  6956   _collector->startTimer();
  6959 // This closure is used to rescan the marked objects on the dirty cards
  6960 // in the mod union table and the card table proper. In the parallel
  6961 // case, although the bitMap is shared, we do a single read so the
  6962 // isMarked() query is "safe".
  6963 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
  6964   // Ignore mark word because we are running concurrent with mutators
  6965   assert(p->is_oop_or_null(true), "expected an oop or null");
  6966   HeapWord* addr = (HeapWord*)p;
  6967   assert(_span.contains(addr), "we are scanning the CMS generation");
  6968   bool is_obj_array = false;
  6969   #ifdef DEBUG
  6970     if (!_parallel) {
  6971       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  6972       assert(_collector->overflow_list_is_empty(),
  6973              "overflow list should be empty");
  6976   #endif // DEBUG
  6977   if (_bit_map->isMarked(addr)) {
  6978     // Obj arrays are precisely marked, non-arrays are not;
  6979     // so we scan objArrays precisely and non-arrays in their
  6980     // entirety.
  6981     if (p->is_objArray()) {
  6982       is_obj_array = true;
  6983       if (_parallel) {
  6984         p->oop_iterate(_par_scan_closure, mr);
  6985       } else {
  6986         p->oop_iterate(_scan_closure, mr);
  6988     } else {
  6989       if (_parallel) {
  6990         p->oop_iterate(_par_scan_closure);
  6991       } else {
  6992         p->oop_iterate(_scan_closure);
  6996   #ifdef DEBUG
  6997     if (!_parallel) {
  6998       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  6999       assert(_collector->overflow_list_is_empty(),
  7000              "overflow list should be empty");
  7003   #endif // DEBUG
  7004   return is_obj_array;
  7007 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
  7008                         MemRegion span,
  7009                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
  7010                         CMSMarkStack*  revisitStack,
  7011                         bool should_yield, bool verifying):
  7012   _collector(collector),
  7013   _span(span),
  7014   _bitMap(bitMap),
  7015   _mut(&collector->_modUnionTable),
  7016   _markStack(markStack),
  7017   _revisitStack(revisitStack),
  7018   _yield(should_yield),
  7019   _skipBits(0)
  7021   assert(_markStack->isEmpty(), "stack should be empty");
  7022   _finger = _bitMap->startWord();
  7023   _threshold = _finger;
  7024   assert(_collector->_restart_addr == NULL, "Sanity check");
  7025   assert(_span.contains(_finger), "Out of bounds _finger?");
  7026   DEBUG_ONLY(_verifying = verifying;)
  7029 void MarkFromRootsClosure::reset(HeapWord* addr) {
  7030   assert(_markStack->isEmpty(), "would cause duplicates on stack");
  7031   assert(_span.contains(addr), "Out of bounds _finger?");
  7032   _finger = addr;
  7033   _threshold = (HeapWord*)round_to(
  7034                  (intptr_t)_finger, CardTableModRefBS::card_size);
  7037 // Should revisit to see if this should be restructured for
  7038 // greater efficiency.
  7039 bool MarkFromRootsClosure::do_bit(size_t offset) {
  7040   if (_skipBits > 0) {
  7041     _skipBits--;
  7042     return true;
  7044   // convert offset into a HeapWord*
  7045   HeapWord* addr = _bitMap->startWord() + offset;
  7046   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
  7047          "address out of range");
  7048   assert(_bitMap->isMarked(addr), "tautology");
  7049   if (_bitMap->isMarked(addr+1)) {
  7050     // this is an allocated but not yet initialized object
  7051     assert(_skipBits == 0, "tautology");
  7052     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
  7053     oop p = oop(addr);
  7054     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  7055       DEBUG_ONLY(if (!_verifying) {)
  7056         // We re-dirty the cards on which this object lies and increase
  7057         // the _threshold so that we'll come back to scan this object
  7058         // during the preclean or remark phase. (CMSCleanOnEnter)
  7059         if (CMSCleanOnEnter) {
  7060           size_t sz = _collector->block_size_using_printezis_bits(addr);
  7061           HeapWord* end_card_addr   = (HeapWord*)round_to(
  7062                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  7063           MemRegion redirty_range = MemRegion(addr, end_card_addr);
  7064           assert(!redirty_range.is_empty(), "Arithmetical tautology");
  7065           // Bump _threshold to end_card_addr; note that
  7066           // _threshold cannot possibly exceed end_card_addr, anyhow.
  7067           // This prevents future clearing of the card as the scan proceeds
  7068           // to the right.
  7069           assert(_threshold <= end_card_addr,
  7070                  "Because we are just scanning into this object");
  7071           if (_threshold < end_card_addr) {
  7072             _threshold = end_card_addr;
  7074           if (p->klass_or_null() != NULL) {
  7075             // Redirty the range of cards...
  7076             _mut->mark_range(redirty_range);
  7077           } // ...else the setting of klass will dirty the card anyway.
  7079       DEBUG_ONLY(})
  7080       return true;
  7083   scanOopsInOop(addr);
  7084   return true;
  7087 // We take a break if we've been at this for a while,
  7088 // so as to avoid monopolizing the locks involved.
  7089 void MarkFromRootsClosure::do_yield_work() {
  7090   // First give up the locks, then yield, then re-lock
  7091   // We should probably use a constructor/destructor idiom to
  7092   // do this unlock/lock or modify the MutexUnlocker class to
  7093   // serve our purpose. XXX
  7094   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7095          "CMS thread should hold CMS token");
  7096   assert_lock_strong(_bitMap->lock());
  7097   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  7098   _bitMap->lock()->unlock();
  7099   ConcurrentMarkSweepThread::desynchronize(true);
  7100   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7101   _collector->stopTimer();
  7102   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7103   if (PrintCMSStatistics != 0) {
  7104     _collector->incrementYields();
  7106   _collector->icms_wait();
  7108   // See the comment in coordinator_yield()
  7109   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7110                        ConcurrentMarkSweepThread::should_yield() &&
  7111                        !CMSCollector::foregroundGCIsActive(); ++i) {
  7112     os::sleep(Thread::current(), 1, false);
  7113     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7116   ConcurrentMarkSweepThread::synchronize(true);
  7117   _bitMap->lock()->lock_without_safepoint_check();
  7118   _collector->startTimer();
  7121 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
  7122   assert(_bitMap->isMarked(ptr), "expected bit to be set");
  7123   assert(_markStack->isEmpty(),
  7124          "should drain stack to limit stack usage");
  7125   // convert ptr to an oop preparatory to scanning
  7126   oop obj = oop(ptr);
  7127   // Ignore mark word in verification below, since we
  7128   // may be running concurrent with mutators.
  7129   assert(obj->is_oop(true), "should be an oop");
  7130   assert(_finger <= ptr, "_finger runneth ahead");
  7131   // advance the finger to right end of this object
  7132   _finger = ptr + obj->size();
  7133   assert(_finger > ptr, "we just incremented it above");
  7134   // On large heaps, it may take us some time to get through
  7135   // the marking phase (especially if running iCMS). During
  7136   // this time it's possible that a lot of mutations have
  7137   // accumulated in the card table and the mod union table --
  7138   // these mutation records are redundant until we have
  7139   // actually traced into the corresponding card.
  7140   // Here, we check whether advancing the finger would make
  7141   // us cross into a new card, and if so clear corresponding
  7142   // cards in the MUT (preclean them in the card-table in the
  7143   // future).
  7145   DEBUG_ONLY(if (!_verifying) {)
  7146     // The clean-on-enter optimization is disabled by default,
  7147     // until we fix 6178663.
  7148     if (CMSCleanOnEnter && (_finger > _threshold)) {
  7149       // [_threshold, _finger) represents the interval
  7150       // of cards to be cleared  in MUT (or precleaned in card table).
  7151       // The set of cards to be cleared is all those that overlap
  7152       // with the interval [_threshold, _finger); note that
  7153       // _threshold is always kept card-aligned but _finger isn't
  7154       // always card-aligned.
  7155       HeapWord* old_threshold = _threshold;
  7156       assert(old_threshold == (HeapWord*)round_to(
  7157               (intptr_t)old_threshold, CardTableModRefBS::card_size),
  7158              "_threshold should always be card-aligned");
  7159       _threshold = (HeapWord*)round_to(
  7160                      (intptr_t)_finger, CardTableModRefBS::card_size);
  7161       MemRegion mr(old_threshold, _threshold);
  7162       assert(!mr.is_empty(), "Control point invariant");
  7163       assert(_span.contains(mr), "Should clear within span");
  7164       // XXX When _finger crosses from old gen into perm gen
  7165       // we may be doing unnecessary cleaning; do better in the
  7166       // future by detecting that condition and clearing fewer
  7167       // MUT/CT entries.
  7168       _mut->clear_range(mr);
  7170   DEBUG_ONLY(})
  7171   // Note: the finger doesn't advance while we drain
  7172   // the stack below.
  7173   PushOrMarkClosure pushOrMarkClosure(_collector,
  7174                                       _span, _bitMap, _markStack,
  7175                                       _revisitStack,
  7176                                       _finger, this);
  7177   bool res = _markStack->push(obj);
  7178   assert(res, "Empty non-zero size stack should have space for single push");
  7179   while (!_markStack->isEmpty()) {
  7180     oop new_oop = _markStack->pop();
  7181     // Skip verifying header mark word below because we are
  7182     // running concurrent with mutators.
  7183     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7184     // now scan this oop's oops
  7185     new_oop->oop_iterate(&pushOrMarkClosure);
  7186     do_yield_check();
  7188   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
  7191 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
  7192                        CMSCollector* collector, MemRegion span,
  7193                        CMSBitMap* bit_map,
  7194                        OopTaskQueue* work_queue,
  7195                        CMSMarkStack*  overflow_stack,
  7196                        CMSMarkStack*  revisit_stack,
  7197                        bool should_yield):
  7198   _collector(collector),
  7199   _whole_span(collector->_span),
  7200   _span(span),
  7201   _bit_map(bit_map),
  7202   _mut(&collector->_modUnionTable),
  7203   _work_queue(work_queue),
  7204   _overflow_stack(overflow_stack),
  7205   _revisit_stack(revisit_stack),
  7206   _yield(should_yield),
  7207   _skip_bits(0),
  7208   _task(task)
  7210   assert(_work_queue->size() == 0, "work_queue should be empty");
  7211   _finger = span.start();
  7212   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
  7213   assert(_span.contains(_finger), "Out of bounds _finger?");
  7216 // Should revisit to see if this should be restructured for
  7217 // greater efficiency.
  7218 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
  7219   if (_skip_bits > 0) {
  7220     _skip_bits--;
  7221     return true;
  7223   // convert offset into a HeapWord*
  7224   HeapWord* addr = _bit_map->startWord() + offset;
  7225   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
  7226          "address out of range");
  7227   assert(_bit_map->isMarked(addr), "tautology");
  7228   if (_bit_map->isMarked(addr+1)) {
  7229     // this is an allocated object that might not yet be initialized
  7230     assert(_skip_bits == 0, "tautology");
  7231     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
  7232     oop p = oop(addr);
  7233     if (p->klass_or_null() == NULL || !p->is_parsable()) {
  7234       // in the case of Clean-on-Enter optimization, redirty card
  7235       // and avoid clearing card by increasing  the threshold.
  7236       return true;
  7239   scan_oops_in_oop(addr);
  7240   return true;
  7243 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
  7244   assert(_bit_map->isMarked(ptr), "expected bit to be set");
  7245   // Should we assert that our work queue is empty or
  7246   // below some drain limit?
  7247   assert(_work_queue->size() == 0,
  7248          "should drain stack to limit stack usage");
  7249   // convert ptr to an oop preparatory to scanning
  7250   oop obj = oop(ptr);
  7251   // Ignore mark word in verification below, since we
  7252   // may be running concurrent with mutators.
  7253   assert(obj->is_oop(true), "should be an oop");
  7254   assert(_finger <= ptr, "_finger runneth ahead");
  7255   // advance the finger to right end of this object
  7256   _finger = ptr + obj->size();
  7257   assert(_finger > ptr, "we just incremented it above");
  7258   // On large heaps, it may take us some time to get through
  7259   // the marking phase (especially if running iCMS). During
  7260   // this time it's possible that a lot of mutations have
  7261   // accumulated in the card table and the mod union table --
  7262   // these mutation records are redundant until we have
  7263   // actually traced into the corresponding card.
  7264   // Here, we check whether advancing the finger would make
  7265   // us cross into a new card, and if so clear corresponding
  7266   // cards in the MUT (preclean them in the card-table in the
  7267   // future).
  7269   // The clean-on-enter optimization is disabled by default,
  7270   // until we fix 6178663.
  7271   if (CMSCleanOnEnter && (_finger > _threshold)) {
  7272     // [_threshold, _finger) represents the interval
  7273     // of cards to be cleared  in MUT (or precleaned in card table).
  7274     // The set of cards to be cleared is all those that overlap
  7275     // with the interval [_threshold, _finger); note that
  7276     // _threshold is always kept card-aligned but _finger isn't
  7277     // always card-aligned.
  7278     HeapWord* old_threshold = _threshold;
  7279     assert(old_threshold == (HeapWord*)round_to(
  7280             (intptr_t)old_threshold, CardTableModRefBS::card_size),
  7281            "_threshold should always be card-aligned");
  7282     _threshold = (HeapWord*)round_to(
  7283                    (intptr_t)_finger, CardTableModRefBS::card_size);
  7284     MemRegion mr(old_threshold, _threshold);
  7285     assert(!mr.is_empty(), "Control point invariant");
  7286     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
  7287     // XXX When _finger crosses from old gen into perm gen
  7288     // we may be doing unnecessary cleaning; do better in the
  7289     // future by detecting that condition and clearing fewer
  7290     // MUT/CT entries.
  7291     _mut->clear_range(mr);
  7294   // Note: the local finger doesn't advance while we drain
  7295   // the stack below, but the global finger sure can and will.
  7296   HeapWord** gfa = _task->global_finger_addr();
  7297   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
  7298                                       _span, _bit_map,
  7299                                       _work_queue,
  7300                                       _overflow_stack,
  7301                                       _revisit_stack,
  7302                                       _finger,
  7303                                       gfa, this);
  7304   bool res = _work_queue->push(obj);   // overflow could occur here
  7305   assert(res, "Will hold once we use workqueues");
  7306   while (true) {
  7307     oop new_oop;
  7308     if (!_work_queue->pop_local(new_oop)) {
  7309       // We emptied our work_queue; check if there's stuff that can
  7310       // be gotten from the overflow stack.
  7311       if (CMSConcMarkingTask::get_work_from_overflow_stack(
  7312             _overflow_stack, _work_queue)) {
  7313         do_yield_check();
  7314         continue;
  7315       } else {  // done
  7316         break;
  7319     // Skip verifying header mark word below because we are
  7320     // running concurrent with mutators.
  7321     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7322     // now scan this oop's oops
  7323     new_oop->oop_iterate(&pushOrMarkClosure);
  7324     do_yield_check();
  7326   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
  7329 // Yield in response to a request from VM Thread or
  7330 // from mutators.
  7331 void Par_MarkFromRootsClosure::do_yield_work() {
  7332   assert(_task != NULL, "sanity");
  7333   _task->yield();
  7336 // A variant of the above used for verifying CMS marking work.
  7337 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
  7338                         MemRegion span,
  7339                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7340                         CMSMarkStack*  mark_stack):
  7341   _collector(collector),
  7342   _span(span),
  7343   _verification_bm(verification_bm),
  7344   _cms_bm(cms_bm),
  7345   _mark_stack(mark_stack),
  7346   _pam_verify_closure(collector, span, verification_bm, cms_bm,
  7347                       mark_stack)
  7349   assert(_mark_stack->isEmpty(), "stack should be empty");
  7350   _finger = _verification_bm->startWord();
  7351   assert(_collector->_restart_addr == NULL, "Sanity check");
  7352   assert(_span.contains(_finger), "Out of bounds _finger?");
  7355 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
  7356   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
  7357   assert(_span.contains(addr), "Out of bounds _finger?");
  7358   _finger = addr;
  7361 // Should revisit to see if this should be restructured for
  7362 // greater efficiency.
  7363 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
  7364   // convert offset into a HeapWord*
  7365   HeapWord* addr = _verification_bm->startWord() + offset;
  7366   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
  7367          "address out of range");
  7368   assert(_verification_bm->isMarked(addr), "tautology");
  7369   assert(_cms_bm->isMarked(addr), "tautology");
  7371   assert(_mark_stack->isEmpty(),
  7372          "should drain stack to limit stack usage");
  7373   // convert addr to an oop preparatory to scanning
  7374   oop obj = oop(addr);
  7375   assert(obj->is_oop(), "should be an oop");
  7376   assert(_finger <= addr, "_finger runneth ahead");
  7377   // advance the finger to right end of this object
  7378   _finger = addr + obj->size();
  7379   assert(_finger > addr, "we just incremented it above");
  7380   // Note: the finger doesn't advance while we drain
  7381   // the stack below.
  7382   bool res = _mark_stack->push(obj);
  7383   assert(res, "Empty non-zero size stack should have space for single push");
  7384   while (!_mark_stack->isEmpty()) {
  7385     oop new_oop = _mark_stack->pop();
  7386     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
  7387     // now scan this oop's oops
  7388     new_oop->oop_iterate(&_pam_verify_closure);
  7390   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
  7391   return true;
  7394 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
  7395   CMSCollector* collector, MemRegion span,
  7396   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7397   CMSMarkStack*  mark_stack):
  7398   OopClosure(collector->ref_processor()),
  7399   _collector(collector),
  7400   _span(span),
  7401   _verification_bm(verification_bm),
  7402   _cms_bm(cms_bm),
  7403   _mark_stack(mark_stack)
  7404 { }
  7406 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
  7407 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
  7409 // Upon stack overflow, we discard (part of) the stack,
  7410 // remembering the least address amongst those discarded
  7411 // in CMSCollector's _restart_address.
  7412 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
  7413   // Remember the least grey address discarded
  7414   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
  7415   _collector->lower_restart_addr(ra);
  7416   _mark_stack->reset();  // discard stack contents
  7417   _mark_stack->expand(); // expand the stack if possible
  7420 void PushAndMarkVerifyClosure::do_oop(oop obj) {
  7421   assert(obj->is_oop_or_null(), "expected an oop or NULL");
  7422   HeapWord* addr = (HeapWord*)obj;
  7423   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
  7424     // Oop lies in _span and isn't yet grey or black
  7425     _verification_bm->mark(addr);            // now grey
  7426     if (!_cms_bm->isMarked(addr)) {
  7427       oop(addr)->print();
  7428       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
  7429                              addr);
  7430       fatal("... aborting");
  7433     if (!_mark_stack->push(obj)) { // stack overflow
  7434       if (PrintCMSStatistics != 0) {
  7435         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7436                                SIZE_FORMAT, _mark_stack->capacity());
  7438       assert(_mark_stack->isFull(), "Else push should have succeeded");
  7439       handle_stack_overflow(addr);
  7441     // anything including and to the right of _finger
  7442     // will be scanned as we iterate over the remainder of the
  7443     // bit map
  7447 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
  7448                      MemRegion span,
  7449                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
  7450                      CMSMarkStack*  revisitStack,
  7451                      HeapWord* finger, MarkFromRootsClosure* parent) :
  7452   KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
  7453   _span(span),
  7454   _bitMap(bitMap),
  7455   _markStack(markStack),
  7456   _finger(finger),
  7457   _parent(parent)
  7458 { }
  7460 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
  7461                      MemRegion span,
  7462                      CMSBitMap* bit_map,
  7463                      OopTaskQueue* work_queue,
  7464                      CMSMarkStack*  overflow_stack,
  7465                      CMSMarkStack*  revisit_stack,
  7466                      HeapWord* finger,
  7467                      HeapWord** global_finger_addr,
  7468                      Par_MarkFromRootsClosure* parent) :
  7469   Par_KlassRememberingOopClosure(collector,
  7470                             collector->ref_processor(),
  7471                             revisit_stack),
  7472   _whole_span(collector->_span),
  7473   _span(span),
  7474   _bit_map(bit_map),
  7475   _work_queue(work_queue),
  7476   _overflow_stack(overflow_stack),
  7477   _finger(finger),
  7478   _global_finger_addr(global_finger_addr),
  7479   _parent(parent)
  7480 { }
  7482 // Assumes thread-safe access by callers, who are
  7483 // responsible for mutual exclusion.
  7484 void CMSCollector::lower_restart_addr(HeapWord* low) {
  7485   assert(_span.contains(low), "Out of bounds addr");
  7486   if (_restart_addr == NULL) {
  7487     _restart_addr = low;
  7488   } else {
  7489     _restart_addr = MIN2(_restart_addr, low);
  7493 // Upon stack overflow, we discard (part of) the stack,
  7494 // remembering the least address amongst those discarded
  7495 // in CMSCollector's _restart_address.
  7496 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7497   // Remember the least grey address discarded
  7498   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
  7499   _collector->lower_restart_addr(ra);
  7500   _markStack->reset();  // discard stack contents
  7501   _markStack->expand(); // expand the stack if possible
  7504 // Upon stack overflow, we discard (part of) the stack,
  7505 // remembering the least address amongst those discarded
  7506 // in CMSCollector's _restart_address.
  7507 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7508   // We need to do this under a mutex to prevent other
  7509   // workers from interfering with the work done below.
  7510   MutexLockerEx ml(_overflow_stack->par_lock(),
  7511                    Mutex::_no_safepoint_check_flag);
  7512   // Remember the least grey address discarded
  7513   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  7514   _collector->lower_restart_addr(ra);
  7515   _overflow_stack->reset();  // discard stack contents
  7516   _overflow_stack->expand(); // expand the stack if possible
  7519 void PushOrMarkClosure::do_oop(oop obj) {
  7520   // Ignore mark word because we are running concurrent with mutators.
  7521   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7522   HeapWord* addr = (HeapWord*)obj;
  7523   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
  7524     // Oop lies in _span and isn't yet grey or black
  7525     _bitMap->mark(addr);            // now grey
  7526     if (addr < _finger) {
  7527       // the bit map iteration has already either passed, or
  7528       // sampled, this bit in the bit map; we'll need to
  7529       // use the marking stack to scan this oop's oops.
  7530       bool simulate_overflow = false;
  7531       NOT_PRODUCT(
  7532         if (CMSMarkStackOverflowALot &&
  7533             _collector->simulate_overflow()) {
  7534           // simulate a stack overflow
  7535           simulate_overflow = true;
  7538       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
  7539         if (PrintCMSStatistics != 0) {
  7540           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7541                                  SIZE_FORMAT, _markStack->capacity());
  7543         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
  7544         handle_stack_overflow(addr);
  7547     // anything including and to the right of _finger
  7548     // will be scanned as we iterate over the remainder of the
  7549     // bit map
  7550     do_yield_check();
  7554 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
  7555 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
  7557 void Par_PushOrMarkClosure::do_oop(oop obj) {
  7558   // Ignore mark word because we are running concurrent with mutators.
  7559   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7560   HeapWord* addr = (HeapWord*)obj;
  7561   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7562     // Oop lies in _span and isn't yet grey or black
  7563     // We read the global_finger (volatile read) strictly after marking oop
  7564     bool res = _bit_map->par_mark(addr);    // now grey
  7565     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
  7566     // Should we push this marked oop on our stack?
  7567     // -- if someone else marked it, nothing to do
  7568     // -- if target oop is above global finger nothing to do
  7569     // -- if target oop is in chunk and above local finger
  7570     //      then nothing to do
  7571     // -- else push on work queue
  7572     if (   !res       // someone else marked it, they will deal with it
  7573         || (addr >= *gfa)  // will be scanned in a later task
  7574         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
  7575       return;
  7577     // the bit map iteration has already either passed, or
  7578     // sampled, this bit in the bit map; we'll need to
  7579     // use the marking stack to scan this oop's oops.
  7580     bool simulate_overflow = false;
  7581     NOT_PRODUCT(
  7582       if (CMSMarkStackOverflowALot &&
  7583           _collector->simulate_overflow()) {
  7584         // simulate a stack overflow
  7585         simulate_overflow = true;
  7588     if (simulate_overflow ||
  7589         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  7590       // stack overflow
  7591       if (PrintCMSStatistics != 0) {
  7592         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7593                                SIZE_FORMAT, _overflow_stack->capacity());
  7595       // We cannot assert that the overflow stack is full because
  7596       // it may have been emptied since.
  7597       assert(simulate_overflow ||
  7598              _work_queue->size() == _work_queue->max_elems(),
  7599             "Else push should have succeeded");
  7600       handle_stack_overflow(addr);
  7602     do_yield_check();
  7606 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
  7607 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
  7609 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
  7610                                              ReferenceProcessor* rp,
  7611                                              CMSMarkStack* revisit_stack) :
  7612   OopClosure(rp),
  7613   _collector(collector),
  7614   _revisit_stack(revisit_stack),
  7615   _should_remember_klasses(collector->should_unload_classes()) {}
  7617 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
  7618                                        MemRegion span,
  7619                                        ReferenceProcessor* rp,
  7620                                        CMSBitMap* bit_map,
  7621                                        CMSBitMap* mod_union_table,
  7622                                        CMSMarkStack*  mark_stack,
  7623                                        CMSMarkStack*  revisit_stack,
  7624                                        bool           concurrent_precleaning):
  7625   KlassRememberingOopClosure(collector, rp, revisit_stack),
  7626   _span(span),
  7627   _bit_map(bit_map),
  7628   _mod_union_table(mod_union_table),
  7629   _mark_stack(mark_stack),
  7630   _concurrent_precleaning(concurrent_precleaning)
  7632   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7635 // Grey object rescan during pre-cleaning and second checkpoint phases --
  7636 // the non-parallel version (the parallel version appears further below.)
  7637 void PushAndMarkClosure::do_oop(oop obj) {
  7638   // Ignore mark word verification. If during concurrent precleaning,
  7639   // the object monitor may be locked. If during the checkpoint
  7640   // phases, the object may already have been reached by a  different
  7641   // path and may be at the end of the global overflow list (so
  7642   // the mark word may be NULL).
  7643   assert(obj->is_oop_or_null(true /* ignore mark word */),
  7644          "expected an oop or NULL");
  7645   HeapWord* addr = (HeapWord*)obj;
  7646   // Check if oop points into the CMS generation
  7647   // and is not marked
  7648   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7649     // a white object ...
  7650     _bit_map->mark(addr);         // ... now grey
  7651     // push on the marking stack (grey set)
  7652     bool simulate_overflow = false;
  7653     NOT_PRODUCT(
  7654       if (CMSMarkStackOverflowALot &&
  7655           _collector->simulate_overflow()) {
  7656         // simulate a stack overflow
  7657         simulate_overflow = true;
  7660     if (simulate_overflow || !_mark_stack->push(obj)) {
  7661       if (_concurrent_precleaning) {
  7662          // During precleaning we can just dirty the appropriate card(s)
  7663          // in the mod union table, thus ensuring that the object remains
  7664          // in the grey set  and continue. In the case of object arrays
  7665          // we need to dirty all of the cards that the object spans,
  7666          // since the rescan of object arrays will be limited to the
  7667          // dirty cards.
  7668          // Note that no one can be intefering with us in this action
  7669          // of dirtying the mod union table, so no locking or atomics
  7670          // are required.
  7671          if (obj->is_objArray()) {
  7672            size_t sz = obj->size();
  7673            HeapWord* end_card_addr = (HeapWord*)round_to(
  7674                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  7675            MemRegion redirty_range = MemRegion(addr, end_card_addr);
  7676            assert(!redirty_range.is_empty(), "Arithmetical tautology");
  7677            _mod_union_table->mark_range(redirty_range);
  7678          } else {
  7679            _mod_union_table->mark(addr);
  7681          _collector->_ser_pmc_preclean_ovflw++;
  7682       } else {
  7683          // During the remark phase, we need to remember this oop
  7684          // in the overflow list.
  7685          _collector->push_on_overflow_list(obj);
  7686          _collector->_ser_pmc_remark_ovflw++;
  7692 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
  7693                                                MemRegion span,
  7694                                                ReferenceProcessor* rp,
  7695                                                CMSBitMap* bit_map,
  7696                                                OopTaskQueue* work_queue,
  7697                                                CMSMarkStack* revisit_stack):
  7698   Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
  7699   _span(span),
  7700   _bit_map(bit_map),
  7701   _work_queue(work_queue)
  7703   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7706 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
  7707 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
  7709 // Grey object rescan during second checkpoint phase --
  7710 // the parallel version.
  7711 void Par_PushAndMarkClosure::do_oop(oop obj) {
  7712   // In the assert below, we ignore the mark word because
  7713   // this oop may point to an already visited object that is
  7714   // on the overflow stack (in which case the mark word has
  7715   // been hijacked for chaining into the overflow stack --
  7716   // if this is the last object in the overflow stack then
  7717   // its mark word will be NULL). Because this object may
  7718   // have been subsequently popped off the global overflow
  7719   // stack, and the mark word possibly restored to the prototypical
  7720   // value, by the time we get to examined this failing assert in
  7721   // the debugger, is_oop_or_null(false) may subsequently start
  7722   // to hold.
  7723   assert(obj->is_oop_or_null(true),
  7724          "expected an oop or NULL");
  7725   HeapWord* addr = (HeapWord*)obj;
  7726   // Check if oop points into the CMS generation
  7727   // and is not marked
  7728   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7729     // a white object ...
  7730     // If we manage to "claim" the object, by being the
  7731     // first thread to mark it, then we push it on our
  7732     // marking stack
  7733     if (_bit_map->par_mark(addr)) {     // ... now grey
  7734       // push on work queue (grey set)
  7735       bool simulate_overflow = false;
  7736       NOT_PRODUCT(
  7737         if (CMSMarkStackOverflowALot &&
  7738             _collector->par_simulate_overflow()) {
  7739           // simulate a stack overflow
  7740           simulate_overflow = true;
  7743       if (simulate_overflow || !_work_queue->push(obj)) {
  7744         _collector->par_push_on_overflow_list(obj);
  7745         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
  7747     } // Else, some other thread got there first
  7751 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
  7752 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
  7754 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
  7755   // TBD
  7758 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
  7759   // TBD
  7762 void CMSPrecleanRefsYieldClosure::do_yield_work() {
  7763   DEBUG_ONLY(RememberKlassesChecker mux(false);)
  7764   Mutex* bml = _collector->bitMapLock();
  7765   assert_lock_strong(bml);
  7766   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7767          "CMS thread should hold CMS token");
  7769   bml->unlock();
  7770   ConcurrentMarkSweepThread::desynchronize(true);
  7772   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7774   _collector->stopTimer();
  7775   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7776   if (PrintCMSStatistics != 0) {
  7777     _collector->incrementYields();
  7779   _collector->icms_wait();
  7781   // See the comment in coordinator_yield()
  7782   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7783                        ConcurrentMarkSweepThread::should_yield() &&
  7784                        !CMSCollector::foregroundGCIsActive(); ++i) {
  7785     os::sleep(Thread::current(), 1, false);
  7786     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7789   ConcurrentMarkSweepThread::synchronize(true);
  7790   bml->lock();
  7792   _collector->startTimer();
  7795 bool CMSPrecleanRefsYieldClosure::should_return() {
  7796   if (ConcurrentMarkSweepThread::should_yield()) {
  7797     do_yield_work();
  7799   return _collector->foregroundGCIsActive();
  7802 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
  7803   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
  7804          "mr should be aligned to start at a card boundary");
  7805   // We'd like to assert:
  7806   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
  7807   //        "mr should be a range of cards");
  7808   // However, that would be too strong in one case -- the last
  7809   // partition ends at _unallocated_block which, in general, can be
  7810   // an arbitrary boundary, not necessarily card aligned.
  7811   if (PrintCMSStatistics != 0) {
  7812     _num_dirty_cards +=
  7813          mr.word_size()/CardTableModRefBS::card_size_in_words;
  7815   _space->object_iterate_mem(mr, &_scan_cl);
  7818 SweepClosure::SweepClosure(CMSCollector* collector,
  7819                            ConcurrentMarkSweepGeneration* g,
  7820                            CMSBitMap* bitMap, bool should_yield) :
  7821   _collector(collector),
  7822   _g(g),
  7823   _sp(g->cmsSpace()),
  7824   _limit(_sp->sweep_limit()),
  7825   _freelistLock(_sp->freelistLock()),
  7826   _bitMap(bitMap),
  7827   _yield(should_yield),
  7828   _inFreeRange(false),           // No free range at beginning of sweep
  7829   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
  7830   _lastFreeRangeCoalesced(false),
  7831   _freeFinger(g->used_region().start())
  7833   NOT_PRODUCT(
  7834     _numObjectsFreed = 0;
  7835     _numWordsFreed   = 0;
  7836     _numObjectsLive = 0;
  7837     _numWordsLive = 0;
  7838     _numObjectsAlreadyFree = 0;
  7839     _numWordsAlreadyFree = 0;
  7840     _last_fc = NULL;
  7842     _sp->initializeIndexedFreeListArrayReturnedBytes();
  7843     _sp->dictionary()->initializeDictReturnedBytes();
  7845   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  7846          "sweep _limit out of bounds");
  7847   if (CMSTraceSweeper) {
  7848     gclog_or_tty->print("\n====================\nStarting new sweep\n");
  7852 // We need this destructor to reclaim any space at the end
  7853 // of the space, which do_blk below may not have added back to
  7854 // the free lists. [basically dealing with the "fringe effect"]
  7855 SweepClosure::~SweepClosure() {
  7856   assert_lock_strong(_freelistLock);
  7857   // this should be treated as the end of a free run if any
  7858   // The current free range should be returned to the free lists
  7859   // as one coalesced chunk.
  7860   if (inFreeRange()) {
  7861     flushCurFreeChunk(freeFinger(),
  7862       pointer_delta(_limit, freeFinger()));
  7863     assert(freeFinger() < _limit, "the finger pointeth off base");
  7864     if (CMSTraceSweeper) {
  7865       gclog_or_tty->print("destructor:");
  7866       gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
  7867                  "[coalesced:"SIZE_FORMAT"]\n",
  7868                  freeFinger(), pointer_delta(_limit, freeFinger()),
  7869                  lastFreeRangeCoalesced());
  7872   NOT_PRODUCT(
  7873     if (Verbose && PrintGC) {
  7874       gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
  7875                           SIZE_FORMAT " bytes",
  7876                  _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
  7877       gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects,  "
  7878                              SIZE_FORMAT" bytes  "
  7879         "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
  7880         _numObjectsLive, _numWordsLive*sizeof(HeapWord),
  7881         _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
  7882       size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
  7883         sizeof(HeapWord);
  7884       gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
  7886       if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
  7887         size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
  7888         size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
  7889         size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
  7890         gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
  7891         gclog_or_tty->print("   Indexed List Returned "SIZE_FORMAT" bytes",
  7892           indexListReturnedBytes);
  7893         gclog_or_tty->print_cr("        Dictionary Returned "SIZE_FORMAT" bytes",
  7894           dictReturnedBytes);
  7898   // Now, in debug mode, just null out the sweep_limit
  7899   NOT_PRODUCT(_sp->clear_sweep_limit();)
  7900   if (CMSTraceSweeper) {
  7901     gclog_or_tty->print("end of sweep\n================\n");
  7905 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
  7906     bool freeRangeInFreeLists) {
  7907   if (CMSTraceSweeper) {
  7908     gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
  7909                freeFinger, _sp->block_size(freeFinger),
  7910                freeRangeInFreeLists);
  7912   assert(!inFreeRange(), "Trampling existing free range");
  7913   set_inFreeRange(true);
  7914   set_lastFreeRangeCoalesced(false);
  7916   set_freeFinger(freeFinger);
  7917   set_freeRangeInFreeLists(freeRangeInFreeLists);
  7918   if (CMSTestInFreeList) {
  7919     if (freeRangeInFreeLists) {
  7920       FreeChunk* fc = (FreeChunk*) freeFinger;
  7921       assert(fc->isFree(), "A chunk on the free list should be free.");
  7922       assert(fc->size() > 0, "Free range should have a size");
  7923       assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
  7928 // Note that the sweeper runs concurrently with mutators. Thus,
  7929 // it is possible for direct allocation in this generation to happen
  7930 // in the middle of the sweep. Note that the sweeper also coalesces
  7931 // contiguous free blocks. Thus, unless the sweeper and the allocator
  7932 // synchronize appropriately freshly allocated blocks may get swept up.
  7933 // This is accomplished by the sweeper locking the free lists while
  7934 // it is sweeping. Thus blocks that are determined to be free are
  7935 // indeed free. There is however one additional complication:
  7936 // blocks that have been allocated since the final checkpoint and
  7937 // mark, will not have been marked and so would be treated as
  7938 // unreachable and swept up. To prevent this, the allocator marks
  7939 // the bit map when allocating during the sweep phase. This leads,
  7940 // however, to a further complication -- objects may have been allocated
  7941 // but not yet initialized -- in the sense that the header isn't yet
  7942 // installed. The sweeper can not then determine the size of the block
  7943 // in order to skip over it. To deal with this case, we use a technique
  7944 // (due to Printezis) to encode such uninitialized block sizes in the
  7945 // bit map. Since the bit map uses a bit per every HeapWord, but the
  7946 // CMS generation has a minimum object size of 3 HeapWords, it follows
  7947 // that "normal marks" won't be adjacent in the bit map (there will
  7948 // always be at least two 0 bits between successive 1 bits). We make use
  7949 // of these "unused" bits to represent uninitialized blocks -- the bit
  7950 // corresponding to the start of the uninitialized object and the next
  7951 // bit are both set. Finally, a 1 bit marks the end of the object that
  7952 // started with the two consecutive 1 bits to indicate its potentially
  7953 // uninitialized state.
  7955 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
  7956   FreeChunk* fc = (FreeChunk*)addr;
  7957   size_t res;
  7959   // Check if we are done sweeping. Below we check "addr >= _limit" rather
  7960   // than "addr == _limit" because although _limit was a block boundary when
  7961   // we started the sweep, it may no longer be one because heap expansion
  7962   // may have caused us to coalesce the block ending at the address _limit
  7963   // with a newly expanded chunk (this happens when _limit was set to the
  7964   // previous _end of the space), so we may have stepped past _limit; see CR 6977970.
  7965   if (addr >= _limit) { // we have swept up to or past the limit, do nothing more
  7966     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  7967            "sweep _limit out of bounds");
  7968     assert(addr < _sp->end(), "addr out of bounds");
  7969     // help the closure application finish
  7970     return pointer_delta(_sp->end(), addr);
  7972   assert(addr < _limit, "sweep invariant");
  7974   // check if we should yield
  7975   do_yield_check(addr);
  7976   if (fc->isFree()) {
  7977     // Chunk that is already free
  7978     res = fc->size();
  7979     doAlreadyFreeChunk(fc);
  7980     debug_only(_sp->verifyFreeLists());
  7981     assert(res == fc->size(), "Don't expect the size to change");
  7982     NOT_PRODUCT(
  7983       _numObjectsAlreadyFree++;
  7984       _numWordsAlreadyFree += res;
  7986     NOT_PRODUCT(_last_fc = fc;)
  7987   } else if (!_bitMap->isMarked(addr)) {
  7988     // Chunk is fresh garbage
  7989     res = doGarbageChunk(fc);
  7990     debug_only(_sp->verifyFreeLists());
  7991     NOT_PRODUCT(
  7992       _numObjectsFreed++;
  7993       _numWordsFreed += res;
  7995   } else {
  7996     // Chunk that is alive.
  7997     res = doLiveChunk(fc);
  7998     debug_only(_sp->verifyFreeLists());
  7999     NOT_PRODUCT(
  8000         _numObjectsLive++;
  8001         _numWordsLive += res;
  8004   return res;
  8007 // For the smart allocation, record following
  8008 //  split deaths - a free chunk is removed from its free list because
  8009 //      it is being split into two or more chunks.
  8010 //  split birth - a free chunk is being added to its free list because
  8011 //      a larger free chunk has been split and resulted in this free chunk.
  8012 //  coal death - a free chunk is being removed from its free list because
  8013 //      it is being coalesced into a large free chunk.
  8014 //  coal birth - a free chunk is being added to its free list because
  8015 //      it was created when two or more free chunks where coalesced into
  8016 //      this free chunk.
  8017 //
  8018 // These statistics are used to determine the desired number of free
  8019 // chunks of a given size.  The desired number is chosen to be relative
  8020 // to the end of a CMS sweep.  The desired number at the end of a sweep
  8021 // is the
  8022 //      count-at-end-of-previous-sweep (an amount that was enough)
  8023 //              - count-at-beginning-of-current-sweep  (the excess)
  8024 //              + split-births  (gains in this size during interval)
  8025 //              - split-deaths  (demands on this size during interval)
  8026 // where the interval is from the end of one sweep to the end of the
  8027 // next.
  8028 //
  8029 // When sweeping the sweeper maintains an accumulated chunk which is
  8030 // the chunk that is made up of chunks that have been coalesced.  That
  8031 // will be termed the left-hand chunk.  A new chunk of garbage that
  8032 // is being considered for coalescing will be referred to as the
  8033 // right-hand chunk.
  8034 //
  8035 // When making a decision on whether to coalesce a right-hand chunk with
  8036 // the current left-hand chunk, the current count vs. the desired count
  8037 // of the left-hand chunk is considered.  Also if the right-hand chunk
  8038 // is near the large chunk at the end of the heap (see
  8039 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
  8040 // left-hand chunk is coalesced.
  8041 //
  8042 // When making a decision about whether to split a chunk, the desired count
  8043 // vs. the current count of the candidate to be split is also considered.
  8044 // If the candidate is underpopulated (currently fewer chunks than desired)
  8045 // a chunk of an overpopulated (currently more chunks than desired) size may
  8046 // be chosen.  The "hint" associated with a free list, if non-null, points
  8047 // to a free list which may be overpopulated.
  8048 //
  8050 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
  8051   size_t size = fc->size();
  8052   // Chunks that cannot be coalesced are not in the
  8053   // free lists.
  8054   if (CMSTestInFreeList && !fc->cantCoalesce()) {
  8055     assert(_sp->verifyChunkInFreeLists(fc),
  8056       "free chunk should be in free lists");
  8058   // a chunk that is already free, should not have been
  8059   // marked in the bit map
  8060   HeapWord* addr = (HeapWord*) fc;
  8061   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
  8062   // Verify that the bit map has no bits marked between
  8063   // addr and purported end of this block.
  8064   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8066   // Some chunks cannot be coalesced in under any circumstances.
  8067   // See the definition of cantCoalesce().
  8068   if (!fc->cantCoalesce()) {
  8069     // This chunk can potentially be coalesced.
  8070     if (_sp->adaptive_freelists()) {
  8071       // All the work is done in
  8072       doPostIsFreeOrGarbageChunk(fc, size);
  8073     } else {  // Not adaptive free lists
  8074       // this is a free chunk that can potentially be coalesced by the sweeper;
  8075       if (!inFreeRange()) {
  8076         // if the next chunk is a free block that can't be coalesced
  8077         // it doesn't make sense to remove this chunk from the free lists
  8078         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
  8079         assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
  8080         if ((HeapWord*)nextChunk < _limit  &&    // there's a next chunk...
  8081             nextChunk->isFree()    &&            // which is free...
  8082             nextChunk->cantCoalesce()) {         // ... but cant be coalesced
  8083           // nothing to do
  8084         } else {
  8085           // Potentially the start of a new free range:
  8086           // Don't eagerly remove it from the free lists.
  8087           // No need to remove it if it will just be put
  8088           // back again.  (Also from a pragmatic point of view
  8089           // if it is a free block in a region that is beyond
  8090           // any allocated blocks, an assertion will fail)
  8091           // Remember the start of a free run.
  8092           initialize_free_range(addr, true);
  8093           // end - can coalesce with next chunk
  8095       } else {
  8096         // the midst of a free range, we are coalescing
  8097         debug_only(record_free_block_coalesced(fc);)
  8098         if (CMSTraceSweeper) {
  8099           gclog_or_tty->print("  -- pick up free block 0x%x (%d)\n", fc, size);
  8101         // remove it from the free lists
  8102         _sp->removeFreeChunkFromFreeLists(fc);
  8103         set_lastFreeRangeCoalesced(true);
  8104         // If the chunk is being coalesced and the current free range is
  8105         // in the free lists, remove the current free range so that it
  8106         // will be returned to the free lists in its entirety - all
  8107         // the coalesced pieces included.
  8108         if (freeRangeInFreeLists()) {
  8109           FreeChunk* ffc = (FreeChunk*) freeFinger();
  8110           assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8111             "Size of free range is inconsistent with chunk size.");
  8112           if (CMSTestInFreeList) {
  8113             assert(_sp->verifyChunkInFreeLists(ffc),
  8114               "free range is not in free lists");
  8116           _sp->removeFreeChunkFromFreeLists(ffc);
  8117           set_freeRangeInFreeLists(false);
  8121   } else {
  8122     // Code path common to both original and adaptive free lists.
  8124     // cant coalesce with previous block; this should be treated
  8125     // as the end of a free run if any
  8126     if (inFreeRange()) {
  8127       // we kicked some butt; time to pick up the garbage
  8128       assert(freeFinger() < addr, "the finger pointeth off base");
  8129       flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8131     // else, nothing to do, just continue
  8135 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
  8136   // This is a chunk of garbage.  It is not in any free list.
  8137   // Add it to a free list or let it possibly be coalesced into
  8138   // a larger chunk.
  8139   HeapWord* addr = (HeapWord*) fc;
  8140   size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  8142   if (_sp->adaptive_freelists()) {
  8143     // Verify that the bit map has no bits marked between
  8144     // addr and purported end of just dead object.
  8145     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8147     doPostIsFreeOrGarbageChunk(fc, size);
  8148   } else {
  8149     if (!inFreeRange()) {
  8150       // start of a new free range
  8151       assert(size > 0, "A free range should have a size");
  8152       initialize_free_range(addr, false);
  8154     } else {
  8155       // this will be swept up when we hit the end of the
  8156       // free range
  8157       if (CMSTraceSweeper) {
  8158         gclog_or_tty->print("  -- pick up garbage 0x%x (%d) \n", fc, size);
  8160       // If the chunk is being coalesced and the current free range is
  8161       // in the free lists, remove the current free range so that it
  8162       // will be returned to the free lists in its entirety - all
  8163       // the coalesced pieces included.
  8164       if (freeRangeInFreeLists()) {
  8165         FreeChunk* ffc = (FreeChunk*)freeFinger();
  8166         assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8167           "Size of free range is inconsistent with chunk size.");
  8168         if (CMSTestInFreeList) {
  8169           assert(_sp->verifyChunkInFreeLists(ffc),
  8170             "free range is not in free lists");
  8172         _sp->removeFreeChunkFromFreeLists(ffc);
  8173         set_freeRangeInFreeLists(false);
  8175       set_lastFreeRangeCoalesced(true);
  8177     // this will be swept up when we hit the end of the free range
  8179     // Verify that the bit map has no bits marked between
  8180     // addr and purported end of just dead object.
  8181     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8183   return size;
  8186 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
  8187   HeapWord* addr = (HeapWord*) fc;
  8188   // The sweeper has just found a live object. Return any accumulated
  8189   // left hand chunk to the free lists.
  8190   if (inFreeRange()) {
  8191     if (_sp->adaptive_freelists()) {
  8192       flushCurFreeChunk(freeFinger(),
  8193                         pointer_delta(addr, freeFinger()));
  8194     } else { // not adaptive freelists
  8195       set_inFreeRange(false);
  8196       // Add the free range back to the free list if it is not already
  8197       // there.
  8198       if (!freeRangeInFreeLists()) {
  8199         assert(freeFinger() < addr, "the finger pointeth off base");
  8200         if (CMSTraceSweeper) {
  8201           gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
  8202             "[coalesced:%d]\n",
  8203             freeFinger(), pointer_delta(addr, freeFinger()),
  8204             lastFreeRangeCoalesced());
  8206         _sp->addChunkAndRepairOffsetTable(freeFinger(),
  8207           pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
  8212   // Common code path for original and adaptive free lists.
  8214   // this object is live: we'd normally expect this to be
  8215   // an oop, and like to assert the following:
  8216   // assert(oop(addr)->is_oop(), "live block should be an oop");
  8217   // However, as we commented above, this may be an object whose
  8218   // header hasn't yet been initialized.
  8219   size_t size;
  8220   assert(_bitMap->isMarked(addr), "Tautology for this control point");
  8221   if (_bitMap->isMarked(addr + 1)) {
  8222     // Determine the size from the bit map, rather than trying to
  8223     // compute it from the object header.
  8224     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  8225     size = pointer_delta(nextOneAddr + 1, addr);
  8226     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  8227            "alignment problem");
  8229     #ifdef DEBUG
  8230       if (oop(addr)->klass_or_null() != NULL &&
  8231           (   !_collector->should_unload_classes()
  8232            || (oop(addr)->is_parsable()) &&
  8233                oop(addr)->is_conc_safe())) {
  8234         // Ignore mark word because we are running concurrent with mutators
  8235         assert(oop(addr)->is_oop(true), "live block should be an oop");
  8236         // is_conc_safe is checked before performing this assertion
  8237         // because an object that is not is_conc_safe may yet have
  8238         // the return from size() correct.
  8239         assert(size ==
  8240                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
  8241                "P-mark and computed size do not agree");
  8243     #endif
  8245   } else {
  8246     // This should be an initialized object that's alive.
  8247     assert(oop(addr)->klass_or_null() != NULL &&
  8248            (!_collector->should_unload_classes()
  8249             || oop(addr)->is_parsable()),
  8250            "Should be an initialized object");
  8251     // Note that there are objects used during class redefinition
  8252     // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite()
  8253     // which are discarded with their is_conc_safe state still
  8254     // false.  These object may be floating garbage so may be
  8255     // seen here.  If they are floating garbage their size
  8256     // should be attainable from their klass.  Do not that
  8257     // is_conc_safe() is true for oop(addr).
  8258     // Ignore mark word because we are running concurrent with mutators
  8259     assert(oop(addr)->is_oop(true), "live block should be an oop");
  8260     // Verify that the bit map has no bits marked between
  8261     // addr and purported end of this block.
  8262     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  8263     assert(size >= 3, "Necessary for Printezis marks to work");
  8264     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
  8265     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
  8267   return size;
  8270 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
  8271                                             size_t chunkSize) {
  8272   // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
  8273   // scheme.
  8274   bool fcInFreeLists = fc->isFree();
  8275   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
  8276   assert((HeapWord*)fc <= _limit, "sweep invariant");
  8277   if (CMSTestInFreeList && fcInFreeLists) {
  8278     assert(_sp->verifyChunkInFreeLists(fc),
  8279       "free chunk is not in free lists");
  8283   if (CMSTraceSweeper) {
  8284     gclog_or_tty->print_cr("  -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
  8287   HeapWord* addr = (HeapWord*) fc;
  8289   bool coalesce;
  8290   size_t left  = pointer_delta(addr, freeFinger());
  8291   size_t right = chunkSize;
  8292   switch (FLSCoalescePolicy) {
  8293     // numeric value forms a coalition aggressiveness metric
  8294     case 0:  { // never coalesce
  8295       coalesce = false;
  8296       break;
  8298     case 1: { // coalesce if left & right chunks on overpopulated lists
  8299       coalesce = _sp->coalOverPopulated(left) &&
  8300                  _sp->coalOverPopulated(right);
  8301       break;
  8303     case 2: { // coalesce if left chunk on overpopulated list (default)
  8304       coalesce = _sp->coalOverPopulated(left);
  8305       break;
  8307     case 3: { // coalesce if left OR right chunk on overpopulated list
  8308       coalesce = _sp->coalOverPopulated(left) ||
  8309                  _sp->coalOverPopulated(right);
  8310       break;
  8312     case 4: { // always coalesce
  8313       coalesce = true;
  8314       break;
  8316     default:
  8317      ShouldNotReachHere();
  8320   // Should the current free range be coalesced?
  8321   // If the chunk is in a free range and either we decided to coalesce above
  8322   // or the chunk is near the large block at the end of the heap
  8323   // (isNearLargestChunk() returns true), then coalesce this chunk.
  8324   bool doCoalesce = inFreeRange() &&
  8325     (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
  8326   if (doCoalesce) {
  8327     // Coalesce the current free range on the left with the new
  8328     // chunk on the right.  If either is on a free list,
  8329     // it must be removed from the list and stashed in the closure.
  8330     if (freeRangeInFreeLists()) {
  8331       FreeChunk* ffc = (FreeChunk*)freeFinger();
  8332       assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8333         "Size of free range is inconsistent with chunk size.");
  8334       if (CMSTestInFreeList) {
  8335         assert(_sp->verifyChunkInFreeLists(ffc),
  8336           "Chunk is not in free lists");
  8338       _sp->coalDeath(ffc->size());
  8339       _sp->removeFreeChunkFromFreeLists(ffc);
  8340       set_freeRangeInFreeLists(false);
  8342     if (fcInFreeLists) {
  8343       _sp->coalDeath(chunkSize);
  8344       assert(fc->size() == chunkSize,
  8345         "The chunk has the wrong size or is not in the free lists");
  8346       _sp->removeFreeChunkFromFreeLists(fc);
  8348     set_lastFreeRangeCoalesced(true);
  8349   } else {  // not in a free range and/or should not coalesce
  8350     // Return the current free range and start a new one.
  8351     if (inFreeRange()) {
  8352       // In a free range but cannot coalesce with the right hand chunk.
  8353       // Put the current free range into the free lists.
  8354       flushCurFreeChunk(freeFinger(),
  8355         pointer_delta(addr, freeFinger()));
  8357     // Set up for new free range.  Pass along whether the right hand
  8358     // chunk is in the free lists.
  8359     initialize_free_range((HeapWord*)fc, fcInFreeLists);
  8362 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
  8363   assert(inFreeRange(), "Should only be called if currently in a free range.");
  8364   assert(size > 0,
  8365     "A zero sized chunk cannot be added to the free lists.");
  8366   if (!freeRangeInFreeLists()) {
  8367     if(CMSTestInFreeList) {
  8368       FreeChunk* fc = (FreeChunk*) chunk;
  8369       fc->setSize(size);
  8370       assert(!_sp->verifyChunkInFreeLists(fc),
  8371         "chunk should not be in free lists yet");
  8373     if (CMSTraceSweeper) {
  8374       gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
  8375                     chunk, size);
  8377     // A new free range is going to be starting.  The current
  8378     // free range has not been added to the free lists yet or
  8379     // was removed so add it back.
  8380     // If the current free range was coalesced, then the death
  8381     // of the free range was recorded.  Record a birth now.
  8382     if (lastFreeRangeCoalesced()) {
  8383       _sp->coalBirth(size);
  8385     _sp->addChunkAndRepairOffsetTable(chunk, size,
  8386             lastFreeRangeCoalesced());
  8388   set_inFreeRange(false);
  8389   set_freeRangeInFreeLists(false);
  8392 // We take a break if we've been at this for a while,
  8393 // so as to avoid monopolizing the locks involved.
  8394 void SweepClosure::do_yield_work(HeapWord* addr) {
  8395   // Return current free chunk being used for coalescing (if any)
  8396   // to the appropriate freelist.  After yielding, the next
  8397   // free block encountered will start a coalescing range of
  8398   // free blocks.  If the next free block is adjacent to the
  8399   // chunk just flushed, they will need to wait for the next
  8400   // sweep to be coalesced.
  8401   if (inFreeRange()) {
  8402     flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8405   // First give up the locks, then yield, then re-lock.
  8406   // We should probably use a constructor/destructor idiom to
  8407   // do this unlock/lock or modify the MutexUnlocker class to
  8408   // serve our purpose. XXX
  8409   assert_lock_strong(_bitMap->lock());
  8410   assert_lock_strong(_freelistLock);
  8411   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  8412          "CMS thread should hold CMS token");
  8413   _bitMap->lock()->unlock();
  8414   _freelistLock->unlock();
  8415   ConcurrentMarkSweepThread::desynchronize(true);
  8416   ConcurrentMarkSweepThread::acknowledge_yield_request();
  8417   _collector->stopTimer();
  8418   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  8419   if (PrintCMSStatistics != 0) {
  8420     _collector->incrementYields();
  8422   _collector->icms_wait();
  8424   // See the comment in coordinator_yield()
  8425   for (unsigned i = 0; i < CMSYieldSleepCount &&
  8426                        ConcurrentMarkSweepThread::should_yield() &&
  8427                        !CMSCollector::foregroundGCIsActive(); ++i) {
  8428     os::sleep(Thread::current(), 1, false);
  8429     ConcurrentMarkSweepThread::acknowledge_yield_request();
  8432   ConcurrentMarkSweepThread::synchronize(true);
  8433   _freelistLock->lock();
  8434   _bitMap->lock()->lock_without_safepoint_check();
  8435   _collector->startTimer();
  8438 #ifndef PRODUCT
  8439 // This is actually very useful in a product build if it can
  8440 // be called from the debugger.  Compile it into the product
  8441 // as needed.
  8442 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
  8443   return debug_cms_space->verifyChunkInFreeLists(fc);
  8446 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
  8447   if (CMSTraceSweeper) {
  8448     gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
  8451 #endif
  8453 // CMSIsAliveClosure
  8454 bool CMSIsAliveClosure::do_object_b(oop obj) {
  8455   HeapWord* addr = (HeapWord*)obj;
  8456   return addr != NULL &&
  8457          (!_span.contains(addr) || _bit_map->isMarked(addr));
  8460 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
  8461                       MemRegion span,
  8462                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
  8463                       CMSMarkStack* revisit_stack, bool cpc):
  8464   KlassRememberingOopClosure(collector, NULL, revisit_stack),
  8465   _span(span),
  8466   _bit_map(bit_map),
  8467   _mark_stack(mark_stack),
  8468   _concurrent_precleaning(cpc) {
  8469   assert(!_span.is_empty(), "Empty span could spell trouble");
  8473 // CMSKeepAliveClosure: the serial version
  8474 void CMSKeepAliveClosure::do_oop(oop obj) {
  8475   HeapWord* addr = (HeapWord*)obj;
  8476   if (_span.contains(addr) &&
  8477       !_bit_map->isMarked(addr)) {
  8478     _bit_map->mark(addr);
  8479     bool simulate_overflow = false;
  8480     NOT_PRODUCT(
  8481       if (CMSMarkStackOverflowALot &&
  8482           _collector->simulate_overflow()) {
  8483         // simulate a stack overflow
  8484         simulate_overflow = true;
  8487     if (simulate_overflow || !_mark_stack->push(obj)) {
  8488       if (_concurrent_precleaning) {
  8489         // We dirty the overflown object and let the remark
  8490         // phase deal with it.
  8491         assert(_collector->overflow_list_is_empty(), "Error");
  8492         // In the case of object arrays, we need to dirty all of
  8493         // the cards that the object spans. No locking or atomics
  8494         // are needed since no one else can be mutating the mod union
  8495         // table.
  8496         if (obj->is_objArray()) {
  8497           size_t sz = obj->size();
  8498           HeapWord* end_card_addr =
  8499             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
  8500           MemRegion redirty_range = MemRegion(addr, end_card_addr);
  8501           assert(!redirty_range.is_empty(), "Arithmetical tautology");
  8502           _collector->_modUnionTable.mark_range(redirty_range);
  8503         } else {
  8504           _collector->_modUnionTable.mark(addr);
  8506         _collector->_ser_kac_preclean_ovflw++;
  8507       } else {
  8508         _collector->push_on_overflow_list(obj);
  8509         _collector->_ser_kac_ovflw++;
  8515 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
  8516 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
  8518 // CMSParKeepAliveClosure: a parallel version of the above.
  8519 // The work queues are private to each closure (thread),
  8520 // but (may be) available for stealing by other threads.
  8521 void CMSParKeepAliveClosure::do_oop(oop obj) {
  8522   HeapWord* addr = (HeapWord*)obj;
  8523   if (_span.contains(addr) &&
  8524       !_bit_map->isMarked(addr)) {
  8525     // In general, during recursive tracing, several threads
  8526     // may be concurrently getting here; the first one to
  8527     // "tag" it, claims it.
  8528     if (_bit_map->par_mark(addr)) {
  8529       bool res = _work_queue->push(obj);
  8530       assert(res, "Low water mark should be much less than capacity");
  8531       // Do a recursive trim in the hope that this will keep
  8532       // stack usage lower, but leave some oops for potential stealers
  8533       trim_queue(_low_water_mark);
  8534     } // Else, another thread got there first
  8538 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
  8539 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
  8541 void CMSParKeepAliveClosure::trim_queue(uint max) {
  8542   while (_work_queue->size() > max) {
  8543     oop new_oop;
  8544     if (_work_queue->pop_local(new_oop)) {
  8545       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  8546       assert(_bit_map->isMarked((HeapWord*)new_oop),
  8547              "no white objects on this stack!");
  8548       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  8549       // iterate over the oops in this oop, marking and pushing
  8550       // the ones in CMS heap (i.e. in _span).
  8551       new_oop->oop_iterate(&_mark_and_push);
  8556 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
  8557                                 CMSCollector* collector,
  8558                                 MemRegion span, CMSBitMap* bit_map,
  8559                                 CMSMarkStack* revisit_stack,
  8560                                 OopTaskQueue* work_queue):
  8561   Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
  8562   _span(span),
  8563   _bit_map(bit_map),
  8564   _work_queue(work_queue) { }
  8566 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
  8567   HeapWord* addr = (HeapWord*)obj;
  8568   if (_span.contains(addr) &&
  8569       !_bit_map->isMarked(addr)) {
  8570     if (_bit_map->par_mark(addr)) {
  8571       bool simulate_overflow = false;
  8572       NOT_PRODUCT(
  8573         if (CMSMarkStackOverflowALot &&
  8574             _collector->par_simulate_overflow()) {
  8575           // simulate a stack overflow
  8576           simulate_overflow = true;
  8579       if (simulate_overflow || !_work_queue->push(obj)) {
  8580         _collector->par_push_on_overflow_list(obj);
  8581         _collector->_par_kac_ovflw++;
  8583     } // Else another thread got there already
  8587 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8588 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8590 //////////////////////////////////////////////////////////////////
  8591 //  CMSExpansionCause                /////////////////////////////
  8592 //////////////////////////////////////////////////////////////////
  8593 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
  8594   switch (cause) {
  8595     case _no_expansion:
  8596       return "No expansion";
  8597     case _satisfy_free_ratio:
  8598       return "Free ratio";
  8599     case _satisfy_promotion:
  8600       return "Satisfy promotion";
  8601     case _satisfy_allocation:
  8602       return "allocation";
  8603     case _allocate_par_lab:
  8604       return "Par LAB";
  8605     case _allocate_par_spooling_space:
  8606       return "Par Spooling Space";
  8607     case _adaptive_size_policy:
  8608       return "Ergonomics";
  8609     default:
  8610       return "unknown";
  8614 void CMSDrainMarkingStackClosure::do_void() {
  8615   // the max number to take from overflow list at a time
  8616   const size_t num = _mark_stack->capacity()/4;
  8617   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
  8618          "Overflow list should be NULL during concurrent phases");
  8619   while (!_mark_stack->isEmpty() ||
  8620          // if stack is empty, check the overflow list
  8621          _collector->take_from_overflow_list(num, _mark_stack)) {
  8622     oop obj = _mark_stack->pop();
  8623     HeapWord* addr = (HeapWord*)obj;
  8624     assert(_span.contains(addr), "Should be within span");
  8625     assert(_bit_map->isMarked(addr), "Should be marked");
  8626     assert(obj->is_oop(), "Should be an oop");
  8627     obj->oop_iterate(_keep_alive);
  8631 void CMSParDrainMarkingStackClosure::do_void() {
  8632   // drain queue
  8633   trim_queue(0);
  8636 // Trim our work_queue so its length is below max at return
  8637 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
  8638   while (_work_queue->size() > max) {
  8639     oop new_oop;
  8640     if (_work_queue->pop_local(new_oop)) {
  8641       assert(new_oop->is_oop(), "Expected an oop");
  8642       assert(_bit_map->isMarked((HeapWord*)new_oop),
  8643              "no white objects on this stack!");
  8644       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  8645       // iterate over the oops in this oop, marking and pushing
  8646       // the ones in CMS heap (i.e. in _span).
  8647       new_oop->oop_iterate(&_mark_and_push);
  8652 ////////////////////////////////////////////////////////////////////
  8653 // Support for Marking Stack Overflow list handling and related code
  8654 ////////////////////////////////////////////////////////////////////
  8655 // Much of the following code is similar in shape and spirit to the
  8656 // code used in ParNewGC. We should try and share that code
  8657 // as much as possible in the future.
  8659 #ifndef PRODUCT
  8660 // Debugging support for CMSStackOverflowALot
  8662 // It's OK to call this multi-threaded;  the worst thing
  8663 // that can happen is that we'll get a bunch of closely
  8664 // spaced simulated oveflows, but that's OK, in fact
  8665 // probably good as it would exercise the overflow code
  8666 // under contention.
  8667 bool CMSCollector::simulate_overflow() {
  8668   if (_overflow_counter-- <= 0) { // just being defensive
  8669     _overflow_counter = CMSMarkStackOverflowInterval;
  8670     return true;
  8671   } else {
  8672     return false;
  8676 bool CMSCollector::par_simulate_overflow() {
  8677   return simulate_overflow();
  8679 #endif
  8681 // Single-threaded
  8682 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
  8683   assert(stack->isEmpty(), "Expected precondition");
  8684   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
  8685   size_t i = num;
  8686   oop  cur = _overflow_list;
  8687   const markOop proto = markOopDesc::prototype();
  8688   NOT_PRODUCT(ssize_t n = 0;)
  8689   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
  8690     next = oop(cur->mark());
  8691     cur->set_mark(proto);   // until proven otherwise
  8692     assert(cur->is_oop(), "Should be an oop");
  8693     bool res = stack->push(cur);
  8694     assert(res, "Bit off more than can chew?");
  8695     NOT_PRODUCT(n++;)
  8697   _overflow_list = cur;
  8698 #ifndef PRODUCT
  8699   assert(_num_par_pushes >= n, "Too many pops?");
  8700   _num_par_pushes -=n;
  8701 #endif
  8702   return !stack->isEmpty();
  8705 #define BUSY  (oop(0x1aff1aff))
  8706 // (MT-safe) Get a prefix of at most "num" from the list.
  8707 // The overflow list is chained through the mark word of
  8708 // each object in the list. We fetch the entire list,
  8709 // break off a prefix of the right size and return the
  8710 // remainder. If other threads try to take objects from
  8711 // the overflow list at that time, they will wait for
  8712 // some time to see if data becomes available. If (and
  8713 // only if) another thread places one or more object(s)
  8714 // on the global list before we have returned the suffix
  8715 // to the global list, we will walk down our local list
  8716 // to find its end and append the global list to
  8717 // our suffix before returning it. This suffix walk can
  8718 // prove to be expensive (quadratic in the amount of traffic)
  8719 // when there are many objects in the overflow list and
  8720 // there is much producer-consumer contention on the list.
  8721 // *NOTE*: The overflow list manipulation code here and
  8722 // in ParNewGeneration:: are very similar in shape,
  8723 // except that in the ParNew case we use the old (from/eden)
  8724 // copy of the object to thread the list via its klass word.
  8725 // Because of the common code, if you make any changes in
  8726 // the code below, please check the ParNew version to see if
  8727 // similar changes might be needed.
  8728 // CR 6797058 has been filed to consolidate the common code.
  8729 bool CMSCollector::par_take_from_overflow_list(size_t num,
  8730                                                OopTaskQueue* work_q,
  8731                                                int no_of_gc_threads) {
  8732   assert(work_q->size() == 0, "First empty local work queue");
  8733   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
  8734   if (_overflow_list == NULL) {
  8735     return false;
  8737   // Grab the entire list; we'll put back a suffix
  8738   oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
  8739   Thread* tid = Thread::current();
  8740   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
  8741   // set to ParallelGCThreads.
  8742   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
  8743   size_t sleep_time_millis = MAX2((size_t)1, num/100);
  8744   // If the list is busy, we spin for a short while,
  8745   // sleeping between attempts to get the list.
  8746   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
  8747     os::sleep(tid, sleep_time_millis, false);
  8748     if (_overflow_list == NULL) {
  8749       // Nothing left to take
  8750       return false;
  8751     } else if (_overflow_list != BUSY) {
  8752       // Try and grab the prefix
  8753       prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
  8756   // If the list was found to be empty, or we spun long
  8757   // enough, we give up and return empty-handed. If we leave
  8758   // the list in the BUSY state below, it must be the case that
  8759   // some other thread holds the overflow list and will set it
  8760   // to a non-BUSY state in the future.
  8761   if (prefix == NULL || prefix == BUSY) {
  8762      // Nothing to take or waited long enough
  8763      if (prefix == NULL) {
  8764        // Write back the NULL in case we overwrote it with BUSY above
  8765        // and it is still the same value.
  8766        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
  8768      return false;
  8770   assert(prefix != NULL && prefix != BUSY, "Error");
  8771   size_t i = num;
  8772   oop cur = prefix;
  8773   // Walk down the first "num" objects, unless we reach the end.
  8774   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
  8775   if (cur->mark() == NULL) {
  8776     // We have "num" or fewer elements in the list, so there
  8777     // is nothing to return to the global list.
  8778     // Write back the NULL in lieu of the BUSY we wrote
  8779     // above, if it is still the same value.
  8780     if (_overflow_list == BUSY) {
  8781       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
  8783   } else {
  8784     // Chop off the suffix and rerturn it to the global list.
  8785     assert(cur->mark() != BUSY, "Error");
  8786     oop suffix_head = cur->mark(); // suffix will be put back on global list
  8787     cur->set_mark(NULL);           // break off suffix
  8788     // It's possible that the list is still in the empty(busy) state
  8789     // we left it in a short while ago; in that case we may be
  8790     // able to place back the suffix without incurring the cost
  8791     // of a walk down the list.
  8792     oop observed_overflow_list = _overflow_list;
  8793     oop cur_overflow_list = observed_overflow_list;
  8794     bool attached = false;
  8795     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
  8796       observed_overflow_list =
  8797         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
  8798       if (cur_overflow_list == observed_overflow_list) {
  8799         attached = true;
  8800         break;
  8801       } else cur_overflow_list = observed_overflow_list;
  8803     if (!attached) {
  8804       // Too bad, someone else sneaked in (at least) an element; we'll need
  8805       // to do a splice. Find tail of suffix so we can prepend suffix to global
  8806       // list.
  8807       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
  8808       oop suffix_tail = cur;
  8809       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
  8810              "Tautology");
  8811       observed_overflow_list = _overflow_list;
  8812       do {
  8813         cur_overflow_list = observed_overflow_list;
  8814         if (cur_overflow_list != BUSY) {
  8815           // Do the splice ...
  8816           suffix_tail->set_mark(markOop(cur_overflow_list));
  8817         } else { // cur_overflow_list == BUSY
  8818           suffix_tail->set_mark(NULL);
  8820         // ... and try to place spliced list back on overflow_list ...
  8821         observed_overflow_list =
  8822           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
  8823       } while (cur_overflow_list != observed_overflow_list);
  8824       // ... until we have succeeded in doing so.
  8828   // Push the prefix elements on work_q
  8829   assert(prefix != NULL, "control point invariant");
  8830   const markOop proto = markOopDesc::prototype();
  8831   oop next;
  8832   NOT_PRODUCT(ssize_t n = 0;)
  8833   for (cur = prefix; cur != NULL; cur = next) {
  8834     next = oop(cur->mark());
  8835     cur->set_mark(proto);   // until proven otherwise
  8836     assert(cur->is_oop(), "Should be an oop");
  8837     bool res = work_q->push(cur);
  8838     assert(res, "Bit off more than we can chew?");
  8839     NOT_PRODUCT(n++;)
  8841 #ifndef PRODUCT
  8842   assert(_num_par_pushes >= n, "Too many pops?");
  8843   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
  8844 #endif
  8845   return true;
  8848 // Single-threaded
  8849 void CMSCollector::push_on_overflow_list(oop p) {
  8850   NOT_PRODUCT(_num_par_pushes++;)
  8851   assert(p->is_oop(), "Not an oop");
  8852   preserve_mark_if_necessary(p);
  8853   p->set_mark((markOop)_overflow_list);
  8854   _overflow_list = p;
  8857 // Multi-threaded; use CAS to prepend to overflow list
  8858 void CMSCollector::par_push_on_overflow_list(oop p) {
  8859   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
  8860   assert(p->is_oop(), "Not an oop");
  8861   par_preserve_mark_if_necessary(p);
  8862   oop observed_overflow_list = _overflow_list;
  8863   oop cur_overflow_list;
  8864   do {
  8865     cur_overflow_list = observed_overflow_list;
  8866     if (cur_overflow_list != BUSY) {
  8867       p->set_mark(markOop(cur_overflow_list));
  8868     } else {
  8869       p->set_mark(NULL);
  8871     observed_overflow_list =
  8872       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
  8873   } while (cur_overflow_list != observed_overflow_list);
  8875 #undef BUSY
  8877 // Single threaded
  8878 // General Note on GrowableArray: pushes may silently fail
  8879 // because we are (temporarily) out of C-heap for expanding
  8880 // the stack. The problem is quite ubiquitous and affects
  8881 // a lot of code in the JVM. The prudent thing for GrowableArray
  8882 // to do (for now) is to exit with an error. However, that may
  8883 // be too draconian in some cases because the caller may be
  8884 // able to recover without much harm. For such cases, we
  8885 // should probably introduce a "soft_push" method which returns
  8886 // an indication of success or failure with the assumption that
  8887 // the caller may be able to recover from a failure; code in
  8888 // the VM can then be changed, incrementally, to deal with such
  8889 // failures where possible, thus, incrementally hardening the VM
  8890 // in such low resource situations.
  8891 void CMSCollector::preserve_mark_work(oop p, markOop m) {
  8892   _preserved_oop_stack.push(p);
  8893   _preserved_mark_stack.push(m);
  8894   assert(m == p->mark(), "Mark word changed");
  8895   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
  8896          "bijection");
  8899 // Single threaded
  8900 void CMSCollector::preserve_mark_if_necessary(oop p) {
  8901   markOop m = p->mark();
  8902   if (m->must_be_preserved(p)) {
  8903     preserve_mark_work(p, m);
  8907 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
  8908   markOop m = p->mark();
  8909   if (m->must_be_preserved(p)) {
  8910     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
  8911     // Even though we read the mark word without holding
  8912     // the lock, we are assured that it will not change
  8913     // because we "own" this oop, so no other thread can
  8914     // be trying to push it on the overflow list; see
  8915     // the assertion in preserve_mark_work() that checks
  8916     // that m == p->mark().
  8917     preserve_mark_work(p, m);
  8921 // We should be able to do this multi-threaded,
  8922 // a chunk of stack being a task (this is
  8923 // correct because each oop only ever appears
  8924 // once in the overflow list. However, it's
  8925 // not very easy to completely overlap this with
  8926 // other operations, so will generally not be done
  8927 // until all work's been completed. Because we
  8928 // expect the preserved oop stack (set) to be small,
  8929 // it's probably fine to do this single-threaded.
  8930 // We can explore cleverer concurrent/overlapped/parallel
  8931 // processing of preserved marks if we feel the
  8932 // need for this in the future. Stack overflow should
  8933 // be so rare in practice and, when it happens, its
  8934 // effect on performance so great that this will
  8935 // likely just be in the noise anyway.
  8936 void CMSCollector::restore_preserved_marks_if_any() {
  8937   assert(SafepointSynchronize::is_at_safepoint(),
  8938          "world should be stopped");
  8939   assert(Thread::current()->is_ConcurrentGC_thread() ||
  8940          Thread::current()->is_VM_thread(),
  8941          "should be single-threaded");
  8942   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
  8943          "bijection");
  8945   while (!_preserved_oop_stack.is_empty()) {
  8946     oop p = _preserved_oop_stack.pop();
  8947     assert(p->is_oop(), "Should be an oop");
  8948     assert(_span.contains(p), "oop should be in _span");
  8949     assert(p->mark() == markOopDesc::prototype(),
  8950            "Set when taken from overflow list");
  8951     markOop m = _preserved_mark_stack.pop();
  8952     p->set_mark(m);
  8954   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
  8955          "stacks were cleared above");
  8958 #ifndef PRODUCT
  8959 bool CMSCollector::no_preserved_marks() const {
  8960   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
  8962 #endif
  8964 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
  8966   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
  8967   CMSAdaptiveSizePolicy* size_policy =
  8968     (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
  8969   assert(size_policy->is_gc_cms_adaptive_size_policy(),
  8970     "Wrong type for size policy");
  8971   return size_policy;
  8974 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
  8975                                            size_t desired_promo_size) {
  8976   if (cur_promo_size < desired_promo_size) {
  8977     size_t expand_bytes = desired_promo_size - cur_promo_size;
  8978     if (PrintAdaptiveSizePolicy && Verbose) {
  8979       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  8980         "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
  8981         expand_bytes);
  8983     expand(expand_bytes,
  8984            MinHeapDeltaBytes,
  8985            CMSExpansionCause::_adaptive_size_policy);
  8986   } else if (desired_promo_size < cur_promo_size) {
  8987     size_t shrink_bytes = cur_promo_size - desired_promo_size;
  8988     if (PrintAdaptiveSizePolicy && Verbose) {
  8989       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  8990         "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
  8991         shrink_bytes);
  8993     shrink(shrink_bytes);
  8997 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
  8998   GenCollectedHeap* gch = GenCollectedHeap::heap();
  8999   CMSGCAdaptivePolicyCounters* counters =
  9000     (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
  9001   assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
  9002     "Wrong kind of counters");
  9003   return counters;
  9007 void ASConcurrentMarkSweepGeneration::update_counters() {
  9008   if (UsePerfData) {
  9009     _space_counters->update_all();
  9010     _gen_counters->update_all();
  9011     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  9012     GenCollectedHeap* gch = GenCollectedHeap::heap();
  9013     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  9014     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  9015       "Wrong gc statistics type");
  9016     counters->update_counters(gc_stats_l);
  9020 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
  9021   if (UsePerfData) {
  9022     _space_counters->update_used(used);
  9023     _space_counters->update_capacity();
  9024     _gen_counters->update_all();
  9026     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  9027     GenCollectedHeap* gch = GenCollectedHeap::heap();
  9028     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  9029     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  9030       "Wrong gc statistics type");
  9031     counters->update_counters(gc_stats_l);
  9035 // The desired expansion delta is computed so that:
  9036 // . desired free percentage or greater is used
  9037 void ASConcurrentMarkSweepGeneration::compute_new_size() {
  9038   assert_locked_or_safepoint(Heap_lock);
  9040   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
  9042   // If incremental collection failed, we just want to expand
  9043   // to the limit.
  9044   if (incremental_collection_failed()) {
  9045     clear_incremental_collection_failed();
  9046     grow_to_reserved();
  9047     return;
  9050   assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
  9052   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
  9053     "Wrong type of heap");
  9054   int prev_level = level() - 1;
  9055   assert(prev_level >= 0, "The cms generation is the lowest generation");
  9056   Generation* prev_gen = gch->get_gen(prev_level);
  9057   assert(prev_gen->kind() == Generation::ASParNew,
  9058     "Wrong type of young generation");
  9059   ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
  9060   size_t cur_eden = younger_gen->eden()->capacity();
  9061   CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
  9062   size_t cur_promo = free();
  9063   size_policy->compute_tenured_generation_free_space(cur_promo,
  9064                                                        max_available(),
  9065                                                        cur_eden);
  9066   resize(cur_promo, size_policy->promo_size());
  9068   // Record the new size of the space in the cms generation
  9069   // that is available for promotions.  This is temporary.
  9070   // It should be the desired promo size.
  9071   size_policy->avg_cms_promo()->sample(free());
  9072   size_policy->avg_old_live()->sample(used());
  9074   if (UsePerfData) {
  9075     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  9076     counters->update_cms_capacity_counter(capacity());
  9080 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
  9081   assert_locked_or_safepoint(Heap_lock);
  9082   assert_lock_strong(freelistLock());
  9083   HeapWord* old_end = _cmsSpace->end();
  9084   HeapWord* unallocated_start = _cmsSpace->unallocated_block();
  9085   assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
  9086   FreeChunk* chunk_at_end = find_chunk_at_end();
  9087   if (chunk_at_end == NULL) {
  9088     // No room to shrink
  9089     if (PrintGCDetails && Verbose) {
  9090       gclog_or_tty->print_cr("No room to shrink: old_end  "
  9091         PTR_FORMAT "  unallocated_start  " PTR_FORMAT
  9092         " chunk_at_end  " PTR_FORMAT,
  9093         old_end, unallocated_start, chunk_at_end);
  9095     return;
  9096   } else {
  9098     // Find the chunk at the end of the space and determine
  9099     // how much it can be shrunk.
  9100     size_t shrinkable_size_in_bytes = chunk_at_end->size();
  9101     size_t aligned_shrinkable_size_in_bytes =
  9102       align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
  9103     assert(unallocated_start <= chunk_at_end->end(),
  9104       "Inconsistent chunk at end of space");
  9105     size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
  9106     size_t word_size_before = heap_word_size(_virtual_space.committed_size());
  9108     // Shrink the underlying space
  9109     _virtual_space.shrink_by(bytes);
  9110     if (PrintGCDetails && Verbose) {
  9111       gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
  9112         " desired_bytes " SIZE_FORMAT
  9113         " shrinkable_size_in_bytes " SIZE_FORMAT
  9114         " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
  9115         "  bytes  " SIZE_FORMAT,
  9116         desired_bytes, shrinkable_size_in_bytes,
  9117         aligned_shrinkable_size_in_bytes, bytes);
  9118       gclog_or_tty->print_cr("          old_end  " SIZE_FORMAT
  9119         "  unallocated_start  " SIZE_FORMAT,
  9120         old_end, unallocated_start);
  9123     // If the space did shrink (shrinking is not guaranteed),
  9124     // shrink the chunk at the end by the appropriate amount.
  9125     if (((HeapWord*)_virtual_space.high()) < old_end) {
  9126       size_t new_word_size =
  9127         heap_word_size(_virtual_space.committed_size());
  9129       // Have to remove the chunk from the dictionary because it is changing
  9130       // size and might be someplace elsewhere in the dictionary.
  9132       // Get the chunk at end, shrink it, and put it
  9133       // back.
  9134       _cmsSpace->removeChunkFromDictionary(chunk_at_end);
  9135       size_t word_size_change = word_size_before - new_word_size;
  9136       size_t chunk_at_end_old_size = chunk_at_end->size();
  9137       assert(chunk_at_end_old_size >= word_size_change,
  9138         "Shrink is too large");
  9139       chunk_at_end->setSize(chunk_at_end_old_size -
  9140                           word_size_change);
  9141       _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
  9142         word_size_change);
  9144       _cmsSpace->returnChunkToDictionary(chunk_at_end);
  9146       MemRegion mr(_cmsSpace->bottom(), new_word_size);
  9147       _bts->resize(new_word_size);  // resize the block offset shared array
  9148       Universe::heap()->barrier_set()->resize_covered_region(mr);
  9149       _cmsSpace->assert_locked();
  9150       _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  9152       NOT_PRODUCT(_cmsSpace->dictionary()->verify());
  9154       // update the space and generation capacity counters
  9155       if (UsePerfData) {
  9156         _space_counters->update_capacity();
  9157         _gen_counters->update_all();
  9160       if (Verbose && PrintGCDetails) {
  9161         size_t new_mem_size = _virtual_space.committed_size();
  9162         size_t old_mem_size = new_mem_size + bytes;
  9163         gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
  9164                       name(), old_mem_size/K, bytes/K, new_mem_size/K);
  9168     assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
  9169       "Inconsistency at end of space");
  9170     assert(chunk_at_end->end() == _cmsSpace->end(),
  9171       "Shrinking is inconsistent");
  9172     return;
  9176 // Transfer some number of overflown objects to usual marking
  9177 // stack. Return true if some objects were transferred.
  9178 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
  9179   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
  9180                     (size_t)ParGCDesiredObjsFromOverflowList);
  9182   bool res = _collector->take_from_overflow_list(num, _mark_stack);
  9183   assert(_collector->overflow_list_is_empty() || res,
  9184          "If list is not empty, we should have taken something");
  9185   assert(!res || !_mark_stack->isEmpty(),
  9186          "If we took something, it should now be on our stack");
  9187   return res;
  9190 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
  9191   size_t res = _sp->block_size_no_stall(addr, _collector);
  9192   assert(res != 0, "Should always be able to compute a size");
  9193   if (_sp->block_is_obj(addr)) {
  9194     if (_live_bit_map->isMarked(addr)) {
  9195       // It can't have been dead in a previous cycle
  9196       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
  9197     } else {
  9198       _dead_bit_map->mark(addr);      // mark the dead object
  9201   return res;
  9204 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase): TraceMemoryManagerStats() {
  9206   switch (phase) {
  9207     case CMSCollector::InitialMarking:
  9208       initialize(true  /* fullGC */ ,
  9209                  true  /* recordGCBeginTime */,
  9210                  true  /* recordPreGCUsage */,
  9211                  false /* recordPeakUsage */,
  9212                  false /* recordPostGCusage */,
  9213                  true  /* recordAccumulatedGCTime */,
  9214                  false /* recordGCEndTime */,
  9215                  false /* countCollection */  );
  9216       break;
  9218     case CMSCollector::FinalMarking:
  9219       initialize(true  /* fullGC */ ,
  9220                  false /* recordGCBeginTime */,
  9221                  false /* recordPreGCUsage */,
  9222                  false /* recordPeakUsage */,
  9223                  false /* recordPostGCusage */,
  9224                  true  /* recordAccumulatedGCTime */,
  9225                  false /* recordGCEndTime */,
  9226                  false /* countCollection */  );
  9227       break;
  9229     case CMSCollector::Sweeping:
  9230       initialize(true  /* fullGC */ ,
  9231                  false /* recordGCBeginTime */,
  9232                  false /* recordPreGCUsage */,
  9233                  true  /* recordPeakUsage */,
  9234                  true  /* recordPostGCusage */,
  9235                  false /* recordAccumulatedGCTime */,
  9236                  true  /* recordGCEndTime */,
  9237                  true  /* countCollection */  );
  9238       break;
  9240     default:
  9241       ShouldNotReachHere();
  9245 // when bailing out of cms in concurrent mode failure
  9246 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(): TraceMemoryManagerStats() {
  9247   initialize(true /* fullGC */ ,
  9248              true /* recordGCBeginTime */,
  9249              true /* recordPreGCUsage */,
  9250              true /* recordPeakUsage */,
  9251              true /* recordPostGCusage */,
  9252              true /* recordAccumulatedGCTime */,
  9253              true /* recordGCEndTime */,
  9254              true /* countCollection */ );

mercurial