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

Thu, 09 Oct 2014 15:42:23 +0200

author
mgerdin
date
Thu, 09 Oct 2014 15:42:23 +0200
changeset 7470
060cdf93040c
parent 6719
8e20ef014b08
child 7476
c2844108a708
permissions
-rw-r--r--

8055479: TLAB stability
Reviewed-by: brutisso, stefank, ahgross

     1 /*
     2  * Copyright (c) 2001, 2014, 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 "precompiled.hpp"
    26 #include "classfile/classLoaderData.hpp"
    27 #include "classfile/symbolTable.hpp"
    28 #include "classfile/systemDictionary.hpp"
    29 #include "code/codeCache.hpp"
    30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
    31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
    32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
    33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
    34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
    35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
    36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
    37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
    38 #include "gc_implementation/parNew/parNewGeneration.hpp"
    39 #include "gc_implementation/shared/collectorCounters.hpp"
    40 #include "gc_implementation/shared/gcTimer.hpp"
    41 #include "gc_implementation/shared/gcTrace.hpp"
    42 #include "gc_implementation/shared/gcTraceTime.hpp"
    43 #include "gc_implementation/shared/isGCActiveMark.hpp"
    44 #include "gc_interface/collectedHeap.inline.hpp"
    45 #include "memory/allocation.hpp"
    46 #include "memory/cardTableRS.hpp"
    47 #include "memory/collectorPolicy.hpp"
    48 #include "memory/gcLocker.inline.hpp"
    49 #include "memory/genCollectedHeap.hpp"
    50 #include "memory/genMarkSweep.hpp"
    51 #include "memory/genOopClosures.inline.hpp"
    52 #include "memory/iterator.hpp"
    53 #include "memory/padded.hpp"
    54 #include "memory/referencePolicy.hpp"
    55 #include "memory/resourceArea.hpp"
    56 #include "memory/tenuredGeneration.hpp"
    57 #include "oops/oop.inline.hpp"
    58 #include "prims/jvmtiExport.hpp"
    59 #include "runtime/globals_extension.hpp"
    60 #include "runtime/handles.inline.hpp"
    61 #include "runtime/java.hpp"
    62 #include "runtime/vmThread.hpp"
    63 #include "services/memoryService.hpp"
    64 #include "services/runtimeService.hpp"
    66 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
    68 // statics
    69 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
    70 bool CMSCollector::_full_gc_requested = false;
    71 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
    73 //////////////////////////////////////////////////////////////////
    74 // In support of CMS/VM thread synchronization
    75 //////////////////////////////////////////////////////////////////
    76 // We split use of the CGC_lock into 2 "levels".
    77 // The low-level locking is of the usual CGC_lock monitor. We introduce
    78 // a higher level "token" (hereafter "CMS token") built on top of the
    79 // low level monitor (hereafter "CGC lock").
    80 // The token-passing protocol gives priority to the VM thread. The
    81 // CMS-lock doesn't provide any fairness guarantees, but clients
    82 // should ensure that it is only held for very short, bounded
    83 // durations.
    84 //
    85 // When either of the CMS thread or the VM thread is involved in
    86 // collection operations during which it does not want the other
    87 // thread to interfere, it obtains the CMS token.
    88 //
    89 // If either thread tries to get the token while the other has
    90 // it, that thread waits. However, if the VM thread and CMS thread
    91 // both want the token, then the VM thread gets priority while the
    92 // CMS thread waits. This ensures, for instance, that the "concurrent"
    93 // phases of the CMS thread's work do not block out the VM thread
    94 // for long periods of time as the CMS thread continues to hog
    95 // the token. (See bug 4616232).
    96 //
    97 // The baton-passing functions are, however, controlled by the
    98 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
    99 // and here the low-level CMS lock, not the high level token,
   100 // ensures mutual exclusion.
   101 //
   102 // Two important conditions that we have to satisfy:
   103 // 1. if a thread does a low-level wait on the CMS lock, then it
   104 //    relinquishes the CMS token if it were holding that token
   105 //    when it acquired the low-level CMS lock.
   106 // 2. any low-level notifications on the low-level lock
   107 //    should only be sent when a thread has relinquished the token.
   108 //
   109 // In the absence of either property, we'd have potential deadlock.
   110 //
   111 // We protect each of the CMS (concurrent and sequential) phases
   112 // with the CMS _token_, not the CMS _lock_.
   113 //
   114 // The only code protected by CMS lock is the token acquisition code
   115 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
   116 // baton-passing code.
   117 //
   118 // Unfortunately, i couldn't come up with a good abstraction to factor and
   119 // hide the naked CGC_lock manipulation in the baton-passing code
   120 // further below. That's something we should try to do. Also, the proof
   121 // of correctness of this 2-level locking scheme is far from obvious,
   122 // and potentially quite slippery. We have an uneasy supsicion, for instance,
   123 // that there may be a theoretical possibility of delay/starvation in the
   124 // low-level lock/wait/notify scheme used for the baton-passing because of
   125 // potential intereference with the priority scheme embodied in the
   126 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
   127 // invocation further below and marked with "XXX 20011219YSR".
   128 // Indeed, as we note elsewhere, this may become yet more slippery
   129 // in the presence of multiple CMS and/or multiple VM threads. XXX
   131 class CMSTokenSync: public StackObj {
   132  private:
   133   bool _is_cms_thread;
   134  public:
   135   CMSTokenSync(bool is_cms_thread):
   136     _is_cms_thread(is_cms_thread) {
   137     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
   138            "Incorrect argument to constructor");
   139     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
   140   }
   142   ~CMSTokenSync() {
   143     assert(_is_cms_thread ?
   144              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
   145              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
   146           "Incorrect state");
   147     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
   148   }
   149 };
   151 // Convenience class that does a CMSTokenSync, and then acquires
   152 // upto three locks.
   153 class CMSTokenSyncWithLocks: public CMSTokenSync {
   154  private:
   155   // Note: locks are acquired in textual declaration order
   156   // and released in the opposite order
   157   MutexLockerEx _locker1, _locker2, _locker3;
   158  public:
   159   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
   160                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
   161     CMSTokenSync(is_cms_thread),
   162     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
   163     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
   164     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
   165   { }
   166 };
   169 // Wrapper class to temporarily disable icms during a foreground cms collection.
   170 class ICMSDisabler: public StackObj {
   171  public:
   172   // The ctor disables icms and wakes up the thread so it notices the change;
   173   // the dtor re-enables icms.  Note that the CMSCollector methods will check
   174   // CMSIncrementalMode.
   175   ICMSDisabler()  { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
   176   ~ICMSDisabler() { CMSCollector::enable_icms(); }
   177 };
   179 //////////////////////////////////////////////////////////////////
   180 //  Concurrent Mark-Sweep Generation /////////////////////////////
   181 //////////////////////////////////////////////////////////////////
   183 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
   185 // This struct contains per-thread things necessary to support parallel
   186 // young-gen collection.
   187 class CMSParGCThreadState: public CHeapObj<mtGC> {
   188  public:
   189   CFLS_LAB lab;
   190   PromotionInfo promo;
   192   // Constructor.
   193   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
   194     promo.setSpace(cfls);
   195   }
   196 };
   198 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
   199      ReservedSpace rs, size_t initial_byte_size, int level,
   200      CardTableRS* ct, bool use_adaptive_freelists,
   201      FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
   202   CardGeneration(rs, initial_byte_size, level, ct),
   203   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
   204   _debug_collection_type(Concurrent_collection_type),
   205   _did_compact(false)
   206 {
   207   HeapWord* bottom = (HeapWord*) _virtual_space.low();
   208   HeapWord* end    = (HeapWord*) _virtual_space.high();
   210   _direct_allocated_words = 0;
   211   NOT_PRODUCT(
   212     _numObjectsPromoted = 0;
   213     _numWordsPromoted = 0;
   214     _numObjectsAllocated = 0;
   215     _numWordsAllocated = 0;
   216   )
   218   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
   219                                            use_adaptive_freelists,
   220                                            dictionaryChoice);
   221   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
   222   if (_cmsSpace == NULL) {
   223     vm_exit_during_initialization(
   224       "CompactibleFreeListSpace allocation failure");
   225   }
   226   _cmsSpace->_gen = this;
   228   _gc_stats = new CMSGCStats();
   230   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
   231   // offsets match. The ability to tell free chunks from objects
   232   // depends on this property.
   233   debug_only(
   234     FreeChunk* junk = NULL;
   235     assert(UseCompressedClassPointers ||
   236            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
   237            "Offset of FreeChunk::_prev within FreeChunk must match"
   238            "  that of OopDesc::_klass within OopDesc");
   239   )
   240   if (CollectedHeap::use_parallel_gc_threads()) {
   241     typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
   242     _par_gc_thread_states =
   243       NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
   244     if (_par_gc_thread_states == NULL) {
   245       vm_exit_during_initialization("Could not allocate par gc structs");
   246     }
   247     for (uint i = 0; i < ParallelGCThreads; i++) {
   248       _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
   249       if (_par_gc_thread_states[i] == NULL) {
   250         vm_exit_during_initialization("Could not allocate par gc structs");
   251       }
   252     }
   253   } else {
   254     _par_gc_thread_states = NULL;
   255   }
   256   _incremental_collection_failed = false;
   257   // The "dilatation_factor" is the expansion that can occur on
   258   // account of the fact that the minimum object size in the CMS
   259   // generation may be larger than that in, say, a contiguous young
   260   //  generation.
   261   // Ideally, in the calculation below, we'd compute the dilatation
   262   // factor as: MinChunkSize/(promoting_gen's min object size)
   263   // Since we do not have such a general query interface for the
   264   // promoting generation, we'll instead just use the mimimum
   265   // object size (which today is a header's worth of space);
   266   // note that all arithmetic is in units of HeapWords.
   267   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
   268   assert(_dilatation_factor >= 1.0, "from previous assert");
   269 }
   272 // The field "_initiating_occupancy" represents the occupancy percentage
   273 // at which we trigger a new collection cycle.  Unless explicitly specified
   274 // via CMSInitiatingOccupancyFraction (argument "io" below), it
   275 // is calculated by:
   276 //
   277 //   Let "f" be MinHeapFreeRatio in
   278 //
   279 //    _intiating_occupancy = 100-f +
   280 //                           f * (CMSTriggerRatio/100)
   281 //   where CMSTriggerRatio is the argument "tr" below.
   282 //
   283 // That is, if we assume the heap is at its desired maximum occupancy at the
   284 // end of a collection, we let CMSTriggerRatio of the (purported) free
   285 // space be allocated before initiating a new collection cycle.
   286 //
   287 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
   288   assert(io <= 100 && tr <= 100, "Check the arguments");
   289   if (io >= 0) {
   290     _initiating_occupancy = (double)io / 100.0;
   291   } else {
   292     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
   293                              (double)(tr * MinHeapFreeRatio) / 100.0)
   294                             / 100.0;
   295   }
   296 }
   298 void ConcurrentMarkSweepGeneration::ref_processor_init() {
   299   assert(collector() != NULL, "no collector");
   300   collector()->ref_processor_init();
   301 }
   303 void CMSCollector::ref_processor_init() {
   304   if (_ref_processor == NULL) {
   305     // Allocate and initialize a reference processor
   306     _ref_processor =
   307       new ReferenceProcessor(_span,                               // span
   308                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
   309                              (int) ParallelGCThreads,             // mt processing degree
   310                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
   311                              (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
   312                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
   313                              &_is_alive_closure);                 // closure for liveness info
   314     // Initialize the _ref_processor field of CMSGen
   315     _cmsGen->set_ref_processor(_ref_processor);
   317   }
   318 }
   320 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
   321   GenCollectedHeap* gch = GenCollectedHeap::heap();
   322   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
   323     "Wrong type of heap");
   324   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
   325     gch->gen_policy()->size_policy();
   326   assert(sp->is_gc_cms_adaptive_size_policy(),
   327     "Wrong type of size policy");
   328   return sp;
   329 }
   331 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
   332   CMSGCAdaptivePolicyCounters* results =
   333     (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
   334   assert(
   335     results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
   336     "Wrong gc policy counter kind");
   337   return results;
   338 }
   341 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
   343   const char* gen_name = "old";
   345   // Generation Counters - generation 1, 1 subspace
   346   _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
   348   _space_counters = new GSpaceCounters(gen_name, 0,
   349                                        _virtual_space.reserved_size(),
   350                                        this, _gen_counters);
   351 }
   353 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
   354   _cms_gen(cms_gen)
   355 {
   356   assert(alpha <= 100, "bad value");
   357   _saved_alpha = alpha;
   359   // Initialize the alphas to the bootstrap value of 100.
   360   _gc0_alpha = _cms_alpha = 100;
   362   _cms_begin_time.update();
   363   _cms_end_time.update();
   365   _gc0_duration = 0.0;
   366   _gc0_period = 0.0;
   367   _gc0_promoted = 0;
   369   _cms_duration = 0.0;
   370   _cms_period = 0.0;
   371   _cms_allocated = 0;
   373   _cms_used_at_gc0_begin = 0;
   374   _cms_used_at_gc0_end = 0;
   375   _allow_duty_cycle_reduction = false;
   376   _valid_bits = 0;
   377   _icms_duty_cycle = CMSIncrementalDutyCycle;
   378 }
   380 double CMSStats::cms_free_adjustment_factor(size_t free) const {
   381   // TBD: CR 6909490
   382   return 1.0;
   383 }
   385 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
   386 }
   388 // If promotion failure handling is on use
   389 // the padded average size of the promotion for each
   390 // young generation collection.
   391 double CMSStats::time_until_cms_gen_full() const {
   392   size_t cms_free = _cms_gen->cmsSpace()->free();
   393   GenCollectedHeap* gch = GenCollectedHeap::heap();
   394   size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
   395                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
   396   if (cms_free > expected_promotion) {
   397     // Start a cms collection if there isn't enough space to promote
   398     // for the next minor collection.  Use the padded average as
   399     // a safety factor.
   400     cms_free -= expected_promotion;
   402     // Adjust by the safety factor.
   403     double cms_free_dbl = (double)cms_free;
   404     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
   405     // Apply a further correction factor which tries to adjust
   406     // for recent occurance of concurrent mode failures.
   407     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
   408     cms_free_dbl = cms_free_dbl * cms_adjustment;
   410     if (PrintGCDetails && Verbose) {
   411       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
   412         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
   413         cms_free, expected_promotion);
   414       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
   415         cms_free_dbl, cms_consumption_rate() + 1.0);
   416     }
   417     // Add 1 in case the consumption rate goes to zero.
   418     return cms_free_dbl / (cms_consumption_rate() + 1.0);
   419   }
   420   return 0.0;
   421 }
   423 // Compare the duration of the cms collection to the
   424 // time remaining before the cms generation is empty.
   425 // Note that the time from the start of the cms collection
   426 // to the start of the cms sweep (less than the total
   427 // duration of the cms collection) can be used.  This
   428 // has been tried and some applications experienced
   429 // promotion failures early in execution.  This was
   430 // possibly because the averages were not accurate
   431 // enough at the beginning.
   432 double CMSStats::time_until_cms_start() const {
   433   // We add "gc0_period" to the "work" calculation
   434   // below because this query is done (mostly) at the
   435   // end of a scavenge, so we need to conservatively
   436   // account for that much possible delay
   437   // in the query so as to avoid concurrent mode failures
   438   // due to starting the collection just a wee bit too
   439   // late.
   440   double work = cms_duration() + gc0_period();
   441   double deadline = time_until_cms_gen_full();
   442   // If a concurrent mode failure occurred recently, we want to be
   443   // more conservative and halve our expected time_until_cms_gen_full()
   444   if (work > deadline) {
   445     if (Verbose && PrintGCDetails) {
   446       gclog_or_tty->print(
   447         " CMSCollector: collect because of anticipated promotion "
   448         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
   449         gc0_period(), time_until_cms_gen_full());
   450     }
   451     return 0.0;
   452   }
   453   return work - deadline;
   454 }
   456 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
   457 // amount of change to prevent wild oscillation.
   458 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
   459                                               unsigned int new_duty_cycle) {
   460   assert(old_duty_cycle <= 100, "bad input value");
   461   assert(new_duty_cycle <= 100, "bad input value");
   463   // Note:  use subtraction with caution since it may underflow (values are
   464   // unsigned).  Addition is safe since we're in the range 0-100.
   465   unsigned int damped_duty_cycle = new_duty_cycle;
   466   if (new_duty_cycle < old_duty_cycle) {
   467     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
   468     if (new_duty_cycle + largest_delta < old_duty_cycle) {
   469       damped_duty_cycle = old_duty_cycle - largest_delta;
   470     }
   471   } else if (new_duty_cycle > old_duty_cycle) {
   472     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
   473     if (new_duty_cycle > old_duty_cycle + largest_delta) {
   474       damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
   475     }
   476   }
   477   assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
   479   if (CMSTraceIncrementalPacing) {
   480     gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
   481                            old_duty_cycle, new_duty_cycle, damped_duty_cycle);
   482   }
   483   return damped_duty_cycle;
   484 }
   486 unsigned int CMSStats::icms_update_duty_cycle_impl() {
   487   assert(CMSIncrementalPacing && valid(),
   488          "should be handled in icms_update_duty_cycle()");
   490   double cms_time_so_far = cms_timer().seconds();
   491   double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
   492   double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
   494   // Avoid division by 0.
   495   double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
   496   double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
   498   unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
   499   if (new_duty_cycle > _icms_duty_cycle) {
   500     // Avoid very small duty cycles (1 or 2); 0 is allowed.
   501     if (new_duty_cycle > 2) {
   502       _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
   503                                                 new_duty_cycle);
   504     }
   505   } else if (_allow_duty_cycle_reduction) {
   506     // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
   507     new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
   508     // Respect the minimum duty cycle.
   509     unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
   510     _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
   511   }
   513   if (PrintGCDetails || CMSTraceIncrementalPacing) {
   514     gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
   515   }
   517   _allow_duty_cycle_reduction = false;
   518   return _icms_duty_cycle;
   519 }
   521 #ifndef PRODUCT
   522 void CMSStats::print_on(outputStream *st) const {
   523   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
   524   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
   525                gc0_duration(), gc0_period(), gc0_promoted());
   526   st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
   527             cms_duration(), cms_duration_per_mb(),
   528             cms_period(), cms_allocated());
   529   st->print(",cms_since_beg=%g,cms_since_end=%g",
   530             cms_time_since_begin(), cms_time_since_end());
   531   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
   532             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
   533   if (CMSIncrementalMode) {
   534     st->print(",dc=%d", icms_duty_cycle());
   535   }
   537   if (valid()) {
   538     st->print(",promo_rate=%g,cms_alloc_rate=%g",
   539               promotion_rate(), cms_allocation_rate());
   540     st->print(",cms_consumption_rate=%g,time_until_full=%g",
   541               cms_consumption_rate(), time_until_cms_gen_full());
   542   }
   543   st->print(" ");
   544 }
   545 #endif // #ifndef PRODUCT
   547 CMSCollector::CollectorState CMSCollector::_collectorState =
   548                              CMSCollector::Idling;
   549 bool CMSCollector::_foregroundGCIsActive = false;
   550 bool CMSCollector::_foregroundGCShouldWait = false;
   552 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
   553                            CardTableRS*                   ct,
   554                            ConcurrentMarkSweepPolicy*     cp):
   555   _cmsGen(cmsGen),
   556   _ct(ct),
   557   _ref_processor(NULL),    // will be set later
   558   _conc_workers(NULL),     // may be set later
   559   _abort_preclean(false),
   560   _start_sampling(false),
   561   _between_prologue_and_epilogue(false),
   562   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
   563   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
   564                  -1 /* lock-free */, "No_lock" /* dummy */),
   565   _modUnionClosure(&_modUnionTable),
   566   _modUnionClosurePar(&_modUnionTable),
   567   // Adjust my span to cover old (cms) gen
   568   _span(cmsGen->reserved()),
   569   // Construct the is_alive_closure with _span & markBitMap
   570   _is_alive_closure(_span, &_markBitMap),
   571   _restart_addr(NULL),
   572   _overflow_list(NULL),
   573   _stats(cmsGen),
   574   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
   575   _eden_chunk_array(NULL),     // may be set in ctor body
   576   _eden_chunk_capacity(0),     // -- ditto --
   577   _eden_chunk_index(0),        // -- ditto --
   578   _survivor_plab_array(NULL),  // -- ditto --
   579   _survivor_chunk_array(NULL), // -- ditto --
   580   _survivor_chunk_capacity(0), // -- ditto --
   581   _survivor_chunk_index(0),    // -- ditto --
   582   _ser_pmc_preclean_ovflw(0),
   583   _ser_kac_preclean_ovflw(0),
   584   _ser_pmc_remark_ovflw(0),
   585   _par_pmc_remark_ovflw(0),
   586   _ser_kac_ovflw(0),
   587   _par_kac_ovflw(0),
   588 #ifndef PRODUCT
   589   _num_par_pushes(0),
   590 #endif
   591   _collection_count_start(0),
   592   _verifying(false),
   593   _icms_start_limit(NULL),
   594   _icms_stop_limit(NULL),
   595   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
   596   _completed_initialization(false),
   597   _collector_policy(cp),
   598   _should_unload_classes(CMSClassUnloadingEnabled),
   599   _concurrent_cycles_since_last_unload(0),
   600   _roots_scanning_options(SharedHeap::SO_None),
   601   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
   602   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
   603   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
   604   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
   605   _cms_start_registered(false)
   606 {
   607   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
   608     ExplicitGCInvokesConcurrent = true;
   609   }
   610   // Now expand the span and allocate the collection support structures
   611   // (MUT, marking bit map etc.) to cover both generations subject to
   612   // collection.
   614   // For use by dirty card to oop closures.
   615   _cmsGen->cmsSpace()->set_collector(this);
   617   // Allocate MUT and marking bit map
   618   {
   619     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
   620     if (!_markBitMap.allocate(_span)) {
   621       warning("Failed to allocate CMS Bit Map");
   622       return;
   623     }
   624     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
   625   }
   626   {
   627     _modUnionTable.allocate(_span);
   628     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
   629   }
   631   if (!_markStack.allocate(MarkStackSize)) {
   632     warning("Failed to allocate CMS Marking Stack");
   633     return;
   634   }
   636   // Support for multi-threaded concurrent phases
   637   if (CMSConcurrentMTEnabled) {
   638     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
   639       // just for now
   640       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
   641     }
   642     if (ConcGCThreads > 1) {
   643       _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
   644                                  ConcGCThreads, true);
   645       if (_conc_workers == NULL) {
   646         warning("GC/CMS: _conc_workers allocation failure: "
   647               "forcing -CMSConcurrentMTEnabled");
   648         CMSConcurrentMTEnabled = false;
   649       } else {
   650         _conc_workers->initialize_workers();
   651       }
   652     } else {
   653       CMSConcurrentMTEnabled = false;
   654     }
   655   }
   656   if (!CMSConcurrentMTEnabled) {
   657     ConcGCThreads = 0;
   658   } else {
   659     // Turn off CMSCleanOnEnter optimization temporarily for
   660     // the MT case where it's not fixed yet; see 6178663.
   661     CMSCleanOnEnter = false;
   662   }
   663   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
   664          "Inconsistency");
   666   // Parallel task queues; these are shared for the
   667   // concurrent and stop-world phases of CMS, but
   668   // are not shared with parallel scavenge (ParNew).
   669   {
   670     uint i;
   671     uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
   673     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
   674          || ParallelRefProcEnabled)
   675         && num_queues > 0) {
   676       _task_queues = new OopTaskQueueSet(num_queues);
   677       if (_task_queues == NULL) {
   678         warning("task_queues allocation failure.");
   679         return;
   680       }
   681       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
   682       if (_hash_seed == NULL) {
   683         warning("_hash_seed array allocation failure");
   684         return;
   685       }
   687       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
   688       for (i = 0; i < num_queues; i++) {
   689         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
   690         if (q == NULL) {
   691           warning("work_queue allocation failure.");
   692           return;
   693         }
   694         _task_queues->register_queue(i, q);
   695       }
   696       for (i = 0; i < num_queues; i++) {
   697         _task_queues->queue(i)->initialize();
   698         _hash_seed[i] = 17;  // copied from ParNew
   699       }
   700     }
   701   }
   703   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
   705   // Clip CMSBootstrapOccupancy between 0 and 100.
   706   _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
   708   _full_gcs_since_conc_gc = 0;
   710   // Now tell CMS generations the identity of their collector
   711   ConcurrentMarkSweepGeneration::set_collector(this);
   713   // Create & start a CMS thread for this CMS collector
   714   _cmsThread = ConcurrentMarkSweepThread::start(this);
   715   assert(cmsThread() != NULL, "CMS Thread should have been created");
   716   assert(cmsThread()->collector() == this,
   717          "CMS Thread should refer to this gen");
   718   assert(CGC_lock != NULL, "Where's the CGC_lock?");
   720   // Support for parallelizing young gen rescan
   721   GenCollectedHeap* gch = GenCollectedHeap::heap();
   722   _young_gen = gch->prev_gen(_cmsGen);
   723   if (gch->supports_inline_contig_alloc()) {
   724     _top_addr = gch->top_addr();
   725     _end_addr = gch->end_addr();
   726     assert(_young_gen != NULL, "no _young_gen");
   727     _eden_chunk_index = 0;
   728     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
   729     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
   730     if (_eden_chunk_array == NULL) {
   731       _eden_chunk_capacity = 0;
   732       warning("GC/CMS: _eden_chunk_array allocation failure");
   733     }
   734   }
   735   assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
   737   // Support for parallelizing survivor space rescan
   738   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
   739     const size_t max_plab_samples =
   740       ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
   742     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
   743     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
   744     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
   745     if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
   746         || _cursor == NULL) {
   747       warning("Failed to allocate survivor plab/chunk array");
   748       if (_survivor_plab_array  != NULL) {
   749         FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
   750         _survivor_plab_array = NULL;
   751       }
   752       if (_survivor_chunk_array != NULL) {
   753         FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
   754         _survivor_chunk_array = NULL;
   755       }
   756       if (_cursor != NULL) {
   757         FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
   758         _cursor = NULL;
   759       }
   760     } else {
   761       _survivor_chunk_capacity = 2*max_plab_samples;
   762       for (uint i = 0; i < ParallelGCThreads; i++) {
   763         HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
   764         if (vec == NULL) {
   765           warning("Failed to allocate survivor plab array");
   766           for (int j = i; j > 0; j--) {
   767             FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
   768           }
   769           FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
   770           FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
   771           _survivor_plab_array = NULL;
   772           _survivor_chunk_array = NULL;
   773           _survivor_chunk_capacity = 0;
   774           break;
   775         } else {
   776           ChunkArray* cur =
   777             ::new (&_survivor_plab_array[i]) ChunkArray(vec,
   778                                                         max_plab_samples);
   779           assert(cur->end() == 0, "Should be 0");
   780           assert(cur->array() == vec, "Should be vec");
   781           assert(cur->capacity() == max_plab_samples, "Error");
   782         }
   783       }
   784     }
   785   }
   786   assert(   (   _survivor_plab_array  != NULL
   787              && _survivor_chunk_array != NULL)
   788          || (   _survivor_chunk_capacity == 0
   789              && _survivor_chunk_index == 0),
   790          "Error");
   792   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
   793   _gc_counters = new CollectorCounters("CMS", 1);
   794   _completed_initialization = true;
   795   _inter_sweep_timer.start();  // start of time
   796 }
   798 size_t CMSCollector::plab_sample_minimum_size() {
   799   // The default value of MinTLABSize is 2k, but there is
   800   // no way to get the default value if the flag has been overridden.
   801   return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
   802 }
   804 const char* ConcurrentMarkSweepGeneration::name() const {
   805   return "concurrent mark-sweep generation";
   806 }
   807 void ConcurrentMarkSweepGeneration::update_counters() {
   808   if (UsePerfData) {
   809     _space_counters->update_all();
   810     _gen_counters->update_all();
   811   }
   812 }
   814 // this is an optimized version of update_counters(). it takes the
   815 // used value as a parameter rather than computing it.
   816 //
   817 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
   818   if (UsePerfData) {
   819     _space_counters->update_used(used);
   820     _space_counters->update_capacity();
   821     _gen_counters->update_all();
   822   }
   823 }
   825 void ConcurrentMarkSweepGeneration::print() const {
   826   Generation::print();
   827   cmsSpace()->print();
   828 }
   830 #ifndef PRODUCT
   831 void ConcurrentMarkSweepGeneration::print_statistics() {
   832   cmsSpace()->printFLCensus(0);
   833 }
   834 #endif
   836 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
   837   GenCollectedHeap* gch = GenCollectedHeap::heap();
   838   if (PrintGCDetails) {
   839     if (Verbose) {
   840       gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
   841         level(), short_name(), s, used(), capacity());
   842     } else {
   843       gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
   844         level(), short_name(), s, used() / K, capacity() / K);
   845     }
   846   }
   847   if (Verbose) {
   848     gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
   849               gch->used(), gch->capacity());
   850   } else {
   851     gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
   852               gch->used() / K, gch->capacity() / K);
   853   }
   854 }
   856 size_t
   857 ConcurrentMarkSweepGeneration::contiguous_available() const {
   858   // dld proposes an improvement in precision here. If the committed
   859   // part of the space ends in a free block we should add that to
   860   // uncommitted size in the calculation below. Will make this
   861   // change later, staying with the approximation below for the
   862   // time being. -- ysr.
   863   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
   864 }
   866 size_t
   867 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
   868   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
   869 }
   871 size_t ConcurrentMarkSweepGeneration::max_available() const {
   872   return free() + _virtual_space.uncommitted_size();
   873 }
   875 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
   876   size_t available = max_available();
   877   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
   878   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
   879   if (Verbose && PrintGCDetails) {
   880     gclog_or_tty->print_cr(
   881       "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
   882       "max_promo("SIZE_FORMAT")",
   883       res? "":" not", available, res? ">=":"<",
   884       av_promo, max_promotion_in_bytes);
   885   }
   886   return res;
   887 }
   889 // At a promotion failure dump information on block layout in heap
   890 // (cms old generation).
   891 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
   892   if (CMSDumpAtPromotionFailure) {
   893     cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
   894   }
   895 }
   897 CompactibleSpace*
   898 ConcurrentMarkSweepGeneration::first_compaction_space() const {
   899   return _cmsSpace;
   900 }
   902 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
   903   // Clear the promotion information.  These pointers can be adjusted
   904   // along with all the other pointers into the heap but
   905   // compaction is expected to be a rare event with
   906   // a heap using cms so don't do it without seeing the need.
   907   if (CollectedHeap::use_parallel_gc_threads()) {
   908     for (uint i = 0; i < ParallelGCThreads; i++) {
   909       _par_gc_thread_states[i]->promo.reset();
   910     }
   911   }
   912 }
   914 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
   915   blk->do_space(_cmsSpace);
   916 }
   918 void ConcurrentMarkSweepGeneration::compute_new_size() {
   919   assert_locked_or_safepoint(Heap_lock);
   921   // If incremental collection failed, we just want to expand
   922   // to the limit.
   923   if (incremental_collection_failed()) {
   924     clear_incremental_collection_failed();
   925     grow_to_reserved();
   926     return;
   927   }
   929   // The heap has been compacted but not reset yet.
   930   // Any metric such as free() or used() will be incorrect.
   932   CardGeneration::compute_new_size();
   934   // Reset again after a possible resizing
   935   if (did_compact()) {
   936     cmsSpace()->reset_after_compaction();
   937   }
   938 }
   940 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
   941   assert_locked_or_safepoint(Heap_lock);
   943   // If incremental collection failed, we just want to expand
   944   // to the limit.
   945   if (incremental_collection_failed()) {
   946     clear_incremental_collection_failed();
   947     grow_to_reserved();
   948     return;
   949   }
   951   double free_percentage = ((double) free()) / capacity();
   952   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
   953   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
   955   // compute expansion delta needed for reaching desired free percentage
   956   if (free_percentage < desired_free_percentage) {
   957     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   958     assert(desired_capacity >= capacity(), "invalid expansion size");
   959     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
   960     if (PrintGCDetails && Verbose) {
   961       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   962       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
   963       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
   964       gclog_or_tty->print_cr("  Desired free fraction %f",
   965         desired_free_percentage);
   966       gclog_or_tty->print_cr("  Maximum free fraction %f",
   967         maximum_free_percentage);
   968       gclog_or_tty->print_cr("  Capactiy "SIZE_FORMAT, capacity()/1000);
   969       gclog_or_tty->print_cr("  Desired capacity "SIZE_FORMAT,
   970         desired_capacity/1000);
   971       int prev_level = level() - 1;
   972       if (prev_level >= 0) {
   973         size_t prev_size = 0;
   974         GenCollectedHeap* gch = GenCollectedHeap::heap();
   975         Generation* prev_gen = gch->_gens[prev_level];
   976         prev_size = prev_gen->capacity();
   977           gclog_or_tty->print_cr("  Younger gen size "SIZE_FORMAT,
   978                                  prev_size/1000);
   979       }
   980       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc "SIZE_FORMAT,
   981         unsafe_max_alloc_nogc()/1000);
   982       gclog_or_tty->print_cr("  contiguous available "SIZE_FORMAT,
   983         contiguous_available()/1000);
   984       gclog_or_tty->print_cr("  Expand by "SIZE_FORMAT" (bytes)",
   985         expand_bytes);
   986     }
   987     // safe if expansion fails
   988     expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
   989     if (PrintGCDetails && Verbose) {
   990       gclog_or_tty->print_cr("  Expanded free fraction %f",
   991         ((double) free()) / capacity());
   992     }
   993   } else {
   994     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
   995     assert(desired_capacity <= capacity(), "invalid expansion size");
   996     size_t shrink_bytes = capacity() - desired_capacity;
   997     // Don't shrink unless the delta is greater than the minimum shrink we want
   998     if (shrink_bytes >= MinHeapDeltaBytes) {
   999       shrink_free_list_by(shrink_bytes);
  1004 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
  1005   return cmsSpace()->freelistLock();
  1008 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
  1009                                                   bool   tlab) {
  1010   CMSSynchronousYieldRequest yr;
  1011   MutexLockerEx x(freelistLock(),
  1012                   Mutex::_no_safepoint_check_flag);
  1013   return have_lock_and_allocate(size, tlab);
  1016 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
  1017                                                   bool   tlab /* ignored */) {
  1018   assert_lock_strong(freelistLock());
  1019   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
  1020   HeapWord* res = cmsSpace()->allocate(adjustedSize);
  1021   // Allocate the object live (grey) if the background collector has
  1022   // started marking. This is necessary because the marker may
  1023   // have passed this address and consequently this object will
  1024   // not otherwise be greyed and would be incorrectly swept up.
  1025   // Note that if this object contains references, the writing
  1026   // of those references will dirty the card containing this object
  1027   // allowing the object to be blackened (and its references scanned)
  1028   // either during a preclean phase or at the final checkpoint.
  1029   if (res != NULL) {
  1030     // We may block here with an uninitialized object with
  1031     // its mark-bit or P-bits not yet set. Such objects need
  1032     // to be safely navigable by block_start().
  1033     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
  1034     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
  1035     collector()->direct_allocated(res, adjustedSize);
  1036     _direct_allocated_words += adjustedSize;
  1037     // allocation counters
  1038     NOT_PRODUCT(
  1039       _numObjectsAllocated++;
  1040       _numWordsAllocated += (int)adjustedSize;
  1043   return res;
  1046 // In the case of direct allocation by mutators in a generation that
  1047 // is being concurrently collected, the object must be allocated
  1048 // live (grey) if the background collector has started marking.
  1049 // This is necessary because the marker may
  1050 // have passed this address and consequently this object will
  1051 // not otherwise be greyed and would be incorrectly swept up.
  1052 // Note that if this object contains references, the writing
  1053 // of those references will dirty the card containing this object
  1054 // allowing the object to be blackened (and its references scanned)
  1055 // either during a preclean phase or at the final checkpoint.
  1056 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
  1057   assert(_markBitMap.covers(start, size), "Out of bounds");
  1058   if (_collectorState >= Marking) {
  1059     MutexLockerEx y(_markBitMap.lock(),
  1060                     Mutex::_no_safepoint_check_flag);
  1061     // [see comments preceding SweepClosure::do_blk() below for details]
  1062     //
  1063     // Can the P-bits be deleted now?  JJJ
  1064     //
  1065     // 1. need to mark the object as live so it isn't collected
  1066     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
  1067     // 3. need to mark the end of the object so marking, precleaning or sweeping
  1068     //    can skip over uninitialized or unparsable objects. An allocated
  1069     //    object is considered uninitialized for our purposes as long as
  1070     //    its klass word is NULL.  All old gen objects are parsable
  1071     //    as soon as they are initialized.)
  1072     _markBitMap.mark(start);          // object is live
  1073     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
  1074     _markBitMap.mark(start + size - 1);
  1075                                       // mark end of object
  1077   // check that oop looks uninitialized
  1078   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
  1081 void CMSCollector::promoted(bool par, HeapWord* start,
  1082                             bool is_obj_array, size_t obj_size) {
  1083   assert(_markBitMap.covers(start), "Out of bounds");
  1084   // See comment in direct_allocated() about when objects should
  1085   // be allocated live.
  1086   if (_collectorState >= Marking) {
  1087     // we already hold the marking bit map lock, taken in
  1088     // the prologue
  1089     if (par) {
  1090       _markBitMap.par_mark(start);
  1091     } else {
  1092       _markBitMap.mark(start);
  1094     // We don't need to mark the object as uninitialized (as
  1095     // in direct_allocated above) because this is being done with the
  1096     // world stopped and the object will be initialized by the
  1097     // time the marking, precleaning or sweeping get to look at it.
  1098     // But see the code for copying objects into the CMS generation,
  1099     // where we need to ensure that concurrent readers of the
  1100     // block offset table are able to safely navigate a block that
  1101     // is in flux from being free to being allocated (and in
  1102     // transition while being copied into) and subsequently
  1103     // becoming a bona-fide object when the copy/promotion is complete.
  1104     assert(SafepointSynchronize::is_at_safepoint(),
  1105            "expect promotion only at safepoints");
  1107     if (_collectorState < Sweeping) {
  1108       // Mark the appropriate cards in the modUnionTable, so that
  1109       // this object gets scanned before the sweep. If this is
  1110       // not done, CMS generation references in the object might
  1111       // not get marked.
  1112       // For the case of arrays, which are otherwise precisely
  1113       // marked, we need to dirty the entire array, not just its head.
  1114       if (is_obj_array) {
  1115         // The [par_]mark_range() method expects mr.end() below to
  1116         // be aligned to the granularity of a bit's representation
  1117         // in the heap. In the case of the MUT below, that's a
  1118         // card size.
  1119         MemRegion mr(start,
  1120                      (HeapWord*)round_to((intptr_t)(start + obj_size),
  1121                         CardTableModRefBS::card_size /* bytes */));
  1122         if (par) {
  1123           _modUnionTable.par_mark_range(mr);
  1124         } else {
  1125           _modUnionTable.mark_range(mr);
  1127       } else {  // not an obj array; we can just mark the head
  1128         if (par) {
  1129           _modUnionTable.par_mark(start);
  1130         } else {
  1131           _modUnionTable.mark(start);
  1138 static inline size_t percent_of_space(Space* space, HeapWord* addr)
  1140   size_t delta = pointer_delta(addr, space->bottom());
  1141   return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
  1144 void CMSCollector::icms_update_allocation_limits()
  1146   Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
  1147   EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
  1149   const unsigned int duty_cycle = stats().icms_update_duty_cycle();
  1150   if (CMSTraceIncrementalPacing) {
  1151     stats().print();
  1154   assert(duty_cycle <= 100, "invalid duty cycle");
  1155   if (duty_cycle != 0) {
  1156     // The duty_cycle is a percentage between 0 and 100; convert to words and
  1157     // then compute the offset from the endpoints of the space.
  1158     size_t free_words = eden->free() / HeapWordSize;
  1159     double free_words_dbl = (double)free_words;
  1160     size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
  1161     size_t offset_words = (free_words - duty_cycle_words) / 2;
  1163     _icms_start_limit = eden->top() + offset_words;
  1164     _icms_stop_limit = eden->end() - offset_words;
  1166     // The limits may be adjusted (shifted to the right) by
  1167     // CMSIncrementalOffset, to allow the application more mutator time after a
  1168     // young gen gc (when all mutators were stopped) and before CMS starts and
  1169     // takes away one or more cpus.
  1170     if (CMSIncrementalOffset != 0) {
  1171       double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
  1172       size_t adjustment = (size_t)adjustment_dbl;
  1173       HeapWord* tmp_stop = _icms_stop_limit + adjustment;
  1174       if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
  1175         _icms_start_limit += adjustment;
  1176         _icms_stop_limit = tmp_stop;
  1180   if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
  1181     _icms_start_limit = _icms_stop_limit = eden->end();
  1184   // Install the new start limit.
  1185   eden->set_soft_end(_icms_start_limit);
  1187   if (CMSTraceIncrementalMode) {
  1188     gclog_or_tty->print(" icms alloc limits:  "
  1189                            PTR_FORMAT "," PTR_FORMAT
  1190                            " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
  1191                            p2i(_icms_start_limit), p2i(_icms_stop_limit),
  1192                            percent_of_space(eden, _icms_start_limit),
  1193                            percent_of_space(eden, _icms_stop_limit));
  1194     if (Verbose) {
  1195       gclog_or_tty->print("eden:  ");
  1196       eden->print_on(gclog_or_tty);
  1201 // Any changes here should try to maintain the invariant
  1202 // that if this method is called with _icms_start_limit
  1203 // and _icms_stop_limit both NULL, then it should return NULL
  1204 // and not notify the icms thread.
  1205 HeapWord*
  1206 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
  1207                                        size_t word_size)
  1209   // A start_limit equal to end() means the duty cycle is 0, so treat that as a
  1210   // nop.
  1211   if (CMSIncrementalMode && _icms_start_limit != space->end()) {
  1212     if (top <= _icms_start_limit) {
  1213       if (CMSTraceIncrementalMode) {
  1214         space->print_on(gclog_or_tty);
  1215         gclog_or_tty->stamp();
  1216         gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
  1217                                ", new limit=" PTR_FORMAT
  1218                                " (" SIZE_FORMAT "%%)",
  1219                                p2i(top), p2i(_icms_stop_limit),
  1220                                percent_of_space(space, _icms_stop_limit));
  1222       ConcurrentMarkSweepThread::start_icms();
  1223       assert(top < _icms_stop_limit, "Tautology");
  1224       if (word_size < pointer_delta(_icms_stop_limit, top)) {
  1225         return _icms_stop_limit;
  1228       // The allocation will cross both the _start and _stop limits, so do the
  1229       // stop notification also and return end().
  1230       if (CMSTraceIncrementalMode) {
  1231         space->print_on(gclog_or_tty);
  1232         gclog_or_tty->stamp();
  1233         gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
  1234                                ", new limit=" PTR_FORMAT
  1235                                " (" SIZE_FORMAT "%%)",
  1236                                p2i(top), p2i(space->end()),
  1237                                percent_of_space(space, space->end()));
  1239       ConcurrentMarkSweepThread::stop_icms();
  1240       return space->end();
  1243     if (top <= _icms_stop_limit) {
  1244       if (CMSTraceIncrementalMode) {
  1245         space->print_on(gclog_or_tty);
  1246         gclog_or_tty->stamp();
  1247         gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
  1248                                ", new limit=" PTR_FORMAT
  1249                                " (" SIZE_FORMAT "%%)",
  1250                                top, space->end(),
  1251                                percent_of_space(space, space->end()));
  1253       ConcurrentMarkSweepThread::stop_icms();
  1254       return space->end();
  1257     if (CMSTraceIncrementalMode) {
  1258       space->print_on(gclog_or_tty);
  1259       gclog_or_tty->stamp();
  1260       gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
  1261                              ", new limit=" PTR_FORMAT,
  1262                              top, NULL);
  1266   return NULL;
  1269 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
  1270   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
  1271   // allocate, copy and if necessary update promoinfo --
  1272   // delegate to underlying space.
  1273   assert_lock_strong(freelistLock());
  1275 #ifndef PRODUCT
  1276   if (Universe::heap()->promotion_should_fail()) {
  1277     return NULL;
  1279 #endif  // #ifndef PRODUCT
  1281   oop res = _cmsSpace->promote(obj, obj_size);
  1282   if (res == NULL) {
  1283     // expand and retry
  1284     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
  1285     expand(s*HeapWordSize, MinHeapDeltaBytes,
  1286       CMSExpansionCause::_satisfy_promotion);
  1287     // Since there's currently no next generation, we don't try to promote
  1288     // into a more senior generation.
  1289     assert(next_gen() == NULL, "assumption, based upon which no attempt "
  1290                                "is made to pass on a possibly failing "
  1291                                "promotion to next generation");
  1292     res = _cmsSpace->promote(obj, obj_size);
  1294   if (res != NULL) {
  1295     // See comment in allocate() about when objects should
  1296     // be allocated live.
  1297     assert(obj->is_oop(), "Will dereference klass pointer below");
  1298     collector()->promoted(false,           // Not parallel
  1299                           (HeapWord*)res, obj->is_objArray(), obj_size);
  1300     // promotion counters
  1301     NOT_PRODUCT(
  1302       _numObjectsPromoted++;
  1303       _numWordsPromoted +=
  1304         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
  1307   return res;
  1311 HeapWord*
  1312 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
  1313                                              HeapWord* top,
  1314                                              size_t word_sz)
  1316   return collector()->allocation_limit_reached(space, top, word_sz);
  1319 // IMPORTANT: Notes on object size recognition in CMS.
  1320 // ---------------------------------------------------
  1321 // A block of storage in the CMS generation is always in
  1322 // one of three states. A free block (FREE), an allocated
  1323 // object (OBJECT) whose size() method reports the correct size,
  1324 // and an intermediate state (TRANSIENT) in which its size cannot
  1325 // be accurately determined.
  1326 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
  1327 // -----------------------------------------------------
  1328 // FREE:      klass_word & 1 == 1; mark_word holds block size
  1329 //
  1330 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
  1331 //            obj->size() computes correct size
  1332 //
  1333 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
  1334 //
  1335 // STATE IDENTIFICATION: (64 bit+COOPS)
  1336 // ------------------------------------
  1337 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
  1338 //
  1339 // OBJECT:    klass_word installed; klass_word != 0;
  1340 //            obj->size() computes correct size
  1341 //
  1342 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
  1343 //
  1344 //
  1345 // STATE TRANSITION DIAGRAM
  1346 //
  1347 //        mut / parnew                     mut  /  parnew
  1348 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
  1349 //  ^                                                                   |
  1350 //  |------------------------ DEAD <------------------------------------|
  1351 //         sweep                            mut
  1352 //
  1353 // While a block is in TRANSIENT state its size cannot be determined
  1354 // so readers will either need to come back later or stall until
  1355 // the size can be determined. Note that for the case of direct
  1356 // allocation, P-bits, when available, may be used to determine the
  1357 // size of an object that may not yet have been initialized.
  1359 // Things to support parallel young-gen collection.
  1360 oop
  1361 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
  1362                                            oop old, markOop m,
  1363                                            size_t word_sz) {
  1364 #ifndef PRODUCT
  1365   if (Universe::heap()->promotion_should_fail()) {
  1366     return NULL;
  1368 #endif  // #ifndef PRODUCT
  1370   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1371   PromotionInfo* promoInfo = &ps->promo;
  1372   // if we are tracking promotions, then first ensure space for
  1373   // promotion (including spooling space for saving header if necessary).
  1374   // then allocate and copy, then track promoted info if needed.
  1375   // When tracking (see PromotionInfo::track()), the mark word may
  1376   // be displaced and in this case restoration of the mark word
  1377   // occurs in the (oop_since_save_marks_)iterate phase.
  1378   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
  1379     // Out of space for allocating spooling buffers;
  1380     // try expanding and allocating spooling buffers.
  1381     if (!expand_and_ensure_spooling_space(promoInfo)) {
  1382       return NULL;
  1385   assert(promoInfo->has_spooling_space(), "Control point invariant");
  1386   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
  1387   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
  1388   if (obj_ptr == NULL) {
  1389      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
  1390      if (obj_ptr == NULL) {
  1391        return NULL;
  1394   oop obj = oop(obj_ptr);
  1395   OrderAccess::storestore();
  1396   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1397   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
  1398   // IMPORTANT: See note on object initialization for CMS above.
  1399   // Otherwise, copy the object.  Here we must be careful to insert the
  1400   // klass pointer last, since this marks the block as an allocated object.
  1401   // Except with compressed oops it's the mark word.
  1402   HeapWord* old_ptr = (HeapWord*)old;
  1403   // Restore the mark word copied above.
  1404   obj->set_mark(m);
  1405   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1406   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
  1407   OrderAccess::storestore();
  1409   if (UseCompressedClassPointers) {
  1410     // Copy gap missed by (aligned) header size calculation below
  1411     obj->set_klass_gap(old->klass_gap());
  1413   if (word_sz > (size_t)oopDesc::header_size()) {
  1414     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
  1415                                  obj_ptr + oopDesc::header_size(),
  1416                                  word_sz - oopDesc::header_size());
  1419   // Now we can track the promoted object, if necessary.  We take care
  1420   // to delay the transition from uninitialized to full object
  1421   // (i.e., insertion of klass pointer) until after, so that it
  1422   // atomically becomes a promoted object.
  1423   if (promoInfo->tracking()) {
  1424     promoInfo->track((PromotedObject*)obj, old->klass());
  1426   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
  1427   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
  1428   assert(old->is_oop(), "Will use and dereference old klass ptr below");
  1430   // Finally, install the klass pointer (this should be volatile).
  1431   OrderAccess::storestore();
  1432   obj->set_klass(old->klass());
  1433   // We should now be able to calculate the right size for this object
  1434   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
  1436   collector()->promoted(true,          // parallel
  1437                         obj_ptr, old->is_objArray(), word_sz);
  1439   NOT_PRODUCT(
  1440     Atomic::inc_ptr(&_numObjectsPromoted);
  1441     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
  1444   return obj;
  1447 void
  1448 ConcurrentMarkSweepGeneration::
  1449 par_promote_alloc_undo(int thread_num,
  1450                        HeapWord* obj, size_t word_sz) {
  1451   // CMS does not support promotion undo.
  1452   ShouldNotReachHere();
  1455 void
  1456 ConcurrentMarkSweepGeneration::
  1457 par_promote_alloc_done(int thread_num) {
  1458   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1459   ps->lab.retire(thread_num);
  1462 void
  1463 ConcurrentMarkSweepGeneration::
  1464 par_oop_since_save_marks_iterate_done(int thread_num) {
  1465   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
  1466   ParScanWithoutBarrierClosure* dummy_cl = NULL;
  1467   ps->promo.promoted_oops_iterate_nv(dummy_cl);
  1470 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
  1471                                                    size_t size,
  1472                                                    bool   tlab)
  1474   // We allow a STW collection only if a full
  1475   // collection was requested.
  1476   return full || should_allocate(size, tlab); // FIX ME !!!
  1477   // This and promotion failure handling are connected at the
  1478   // hip and should be fixed by untying them.
  1481 bool CMSCollector::shouldConcurrentCollect() {
  1482   if (_full_gc_requested) {
  1483     if (Verbose && PrintGCDetails) {
  1484       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
  1485                              " gc request (or gc_locker)");
  1487     return true;
  1490   // For debugging purposes, change the type of collection.
  1491   // If the rotation is not on the concurrent collection
  1492   // type, don't start a concurrent collection.
  1493   NOT_PRODUCT(
  1494     if (RotateCMSCollectionTypes &&
  1495         (_cmsGen->debug_collection_type() !=
  1496           ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
  1497       assert(_cmsGen->debug_collection_type() !=
  1498         ConcurrentMarkSweepGeneration::Unknown_collection_type,
  1499         "Bad cms collection type");
  1500       return false;
  1504   FreelistLocker x(this);
  1505   // ------------------------------------------------------------------
  1506   // Print out lots of information which affects the initiation of
  1507   // a collection.
  1508   if (PrintCMSInitiationStatistics && stats().valid()) {
  1509     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
  1510     gclog_or_tty->stamp();
  1511     gclog_or_tty->cr();
  1512     stats().print_on(gclog_or_tty);
  1513     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
  1514       stats().time_until_cms_gen_full());
  1515     gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
  1516     gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
  1517                            _cmsGen->contiguous_available());
  1518     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
  1519     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
  1520     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
  1521     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
  1522     gclog_or_tty->print_cr("metadata initialized %d",
  1523       MetaspaceGC::should_concurrent_collect());
  1525   // ------------------------------------------------------------------
  1527   // If the estimated time to complete a cms collection (cms_duration())
  1528   // is less than the estimated time remaining until the cms generation
  1529   // is full, start a collection.
  1530   if (!UseCMSInitiatingOccupancyOnly) {
  1531     if (stats().valid()) {
  1532       if (stats().time_until_cms_start() == 0.0) {
  1533         return true;
  1535     } else {
  1536       // We want to conservatively collect somewhat early in order
  1537       // to try and "bootstrap" our CMS/promotion statistics;
  1538       // this branch will not fire after the first successful CMS
  1539       // collection because the stats should then be valid.
  1540       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
  1541         if (Verbose && PrintGCDetails) {
  1542           gclog_or_tty->print_cr(
  1543             " CMSCollector: collect for bootstrapping statistics:"
  1544             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
  1545             _bootstrap_occupancy);
  1547         return true;
  1552   // Otherwise, we start a collection cycle if
  1553   // old gen want a collection cycle started. Each may use
  1554   // an appropriate criterion for making this decision.
  1555   // XXX We need to make sure that the gen expansion
  1556   // criterion dovetails well with this. XXX NEED TO FIX THIS
  1557   if (_cmsGen->should_concurrent_collect()) {
  1558     if (Verbose && PrintGCDetails) {
  1559       gclog_or_tty->print_cr("CMS old gen initiated");
  1561     return true;
  1564   // We start a collection if we believe an incremental collection may fail;
  1565   // this is not likely to be productive in practice because it's probably too
  1566   // late anyway.
  1567   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1568   assert(gch->collector_policy()->is_two_generation_policy(),
  1569          "You may want to check the correctness of the following");
  1570   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
  1571     if (Verbose && PrintGCDetails) {
  1572       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
  1574     return true;
  1577   if (MetaspaceGC::should_concurrent_collect()) {
  1578       if (Verbose && PrintGCDetails) {
  1579       gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
  1581       return true;
  1584   return false;
  1587 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
  1589 // Clear _expansion_cause fields of constituent generations
  1590 void CMSCollector::clear_expansion_cause() {
  1591   _cmsGen->clear_expansion_cause();
  1594 // We should be conservative in starting a collection cycle.  To
  1595 // start too eagerly runs the risk of collecting too often in the
  1596 // extreme.  To collect too rarely falls back on full collections,
  1597 // which works, even if not optimum in terms of concurrent work.
  1598 // As a work around for too eagerly collecting, use the flag
  1599 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
  1600 // giving the user an easily understandable way of controlling the
  1601 // collections.
  1602 // We want to start a new collection cycle if any of the following
  1603 // conditions hold:
  1604 // . our current occupancy exceeds the configured initiating occupancy
  1605 //   for this generation, or
  1606 // . we recently needed to expand this space and have not, since that
  1607 //   expansion, done a collection of this generation, or
  1608 // . the underlying space believes that it may be a good idea to initiate
  1609 //   a concurrent collection (this may be based on criteria such as the
  1610 //   following: the space uses linear allocation and linear allocation is
  1611 //   going to fail, or there is believed to be excessive fragmentation in
  1612 //   the generation, etc... or ...
  1613 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
  1614 //   the case of the old generation; see CR 6543076):
  1615 //   we may be approaching a point at which allocation requests may fail because
  1616 //   we will be out of sufficient free space given allocation rate estimates.]
  1617 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
  1619   assert_lock_strong(freelistLock());
  1620   if (occupancy() > initiating_occupancy()) {
  1621     if (PrintGCDetails && Verbose) {
  1622       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
  1623         short_name(), occupancy(), initiating_occupancy());
  1625     return true;
  1627   if (UseCMSInitiatingOccupancyOnly) {
  1628     return false;
  1630   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
  1631     if (PrintGCDetails && Verbose) {
  1632       gclog_or_tty->print(" %s: collect because expanded for allocation ",
  1633         short_name());
  1635     return true;
  1637   if (_cmsSpace->should_concurrent_collect()) {
  1638     if (PrintGCDetails && Verbose) {
  1639       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
  1640         short_name());
  1642     return true;
  1644   return false;
  1647 void ConcurrentMarkSweepGeneration::collect(bool   full,
  1648                                             bool   clear_all_soft_refs,
  1649                                             size_t size,
  1650                                             bool   tlab)
  1652   collector()->collect(full, clear_all_soft_refs, size, tlab);
  1655 void CMSCollector::collect(bool   full,
  1656                            bool   clear_all_soft_refs,
  1657                            size_t size,
  1658                            bool   tlab)
  1660   if (!UseCMSCollectionPassing && _collectorState > Idling) {
  1661     // For debugging purposes skip the collection if the state
  1662     // is not currently idle
  1663     if (TraceCMSState) {
  1664       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
  1665         Thread::current(), full, _collectorState);
  1667     return;
  1670   // The following "if" branch is present for defensive reasons.
  1671   // In the current uses of this interface, it can be replaced with:
  1672   // assert(!GC_locker.is_active(), "Can't be called otherwise");
  1673   // But I am not placing that assert here to allow future
  1674   // generality in invoking this interface.
  1675   if (GC_locker::is_active()) {
  1676     // A consistency test for GC_locker
  1677     assert(GC_locker::needs_gc(), "Should have been set already");
  1678     // Skip this foreground collection, instead
  1679     // expanding the heap if necessary.
  1680     // Need the free list locks for the call to free() in compute_new_size()
  1681     compute_new_size();
  1682     return;
  1684   acquire_control_and_collect(full, clear_all_soft_refs);
  1685   _full_gcs_since_conc_gc++;
  1688 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
  1689   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1690   unsigned int gc_count = gch->total_full_collections();
  1691   if (gc_count == full_gc_count) {
  1692     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
  1693     _full_gc_requested = true;
  1694     _full_gc_cause = cause;
  1695     CGC_lock->notify();   // nudge CMS thread
  1696   } else {
  1697     assert(gc_count > full_gc_count, "Error: causal loop");
  1701 bool CMSCollector::is_external_interruption() {
  1702   GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
  1703   return GCCause::is_user_requested_gc(cause) ||
  1704          GCCause::is_serviceability_requested_gc(cause);
  1707 void CMSCollector::report_concurrent_mode_interruption() {
  1708   if (is_external_interruption()) {
  1709     if (PrintGCDetails) {
  1710       gclog_or_tty->print(" (concurrent mode interrupted)");
  1712   } else {
  1713     if (PrintGCDetails) {
  1714       gclog_or_tty->print(" (concurrent mode failure)");
  1716     _gc_tracer_cm->report_concurrent_mode_failure();
  1721 // The foreground and background collectors need to coordinate in order
  1722 // to make sure that they do not mutually interfere with CMS collections.
  1723 // When a background collection is active,
  1724 // the foreground collector may need to take over (preempt) and
  1725 // synchronously complete an ongoing collection. Depending on the
  1726 // frequency of the background collections and the heap usage
  1727 // of the application, this preemption can be seldom or frequent.
  1728 // There are only certain
  1729 // points in the background collection that the "collection-baton"
  1730 // can be passed to the foreground collector.
  1731 //
  1732 // The foreground collector will wait for the baton before
  1733 // starting any part of the collection.  The foreground collector
  1734 // will only wait at one location.
  1735 //
  1736 // The background collector will yield the baton before starting a new
  1737 // phase of the collection (e.g., before initial marking, marking from roots,
  1738 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
  1739 // of the loop which switches the phases. The background collector does some
  1740 // of the phases (initial mark, final re-mark) with the world stopped.
  1741 // Because of locking involved in stopping the world,
  1742 // the foreground collector should not block waiting for the background
  1743 // collector when it is doing a stop-the-world phase.  The background
  1744 // collector will yield the baton at an additional point just before
  1745 // it enters a stop-the-world phase.  Once the world is stopped, the
  1746 // background collector checks the phase of the collection.  If the
  1747 // phase has not changed, it proceeds with the collection.  If the
  1748 // phase has changed, it skips that phase of the collection.  See
  1749 // the comments on the use of the Heap_lock in collect_in_background().
  1750 //
  1751 // Variable used in baton passing.
  1752 //   _foregroundGCIsActive - Set to true by the foreground collector when
  1753 //      it wants the baton.  The foreground clears it when it has finished
  1754 //      the collection.
  1755 //   _foregroundGCShouldWait - Set to true by the background collector
  1756 //        when it is running.  The foreground collector waits while
  1757 //      _foregroundGCShouldWait is true.
  1758 //  CGC_lock - monitor used to protect access to the above variables
  1759 //      and to notify the foreground and background collectors.
  1760 //  _collectorState - current state of the CMS collection.
  1761 //
  1762 // The foreground collector
  1763 //   acquires the CGC_lock
  1764 //   sets _foregroundGCIsActive
  1765 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
  1766 //     various locks acquired in preparation for the collection
  1767 //     are released so as not to block the background collector
  1768 //     that is in the midst of a collection
  1769 //   proceeds with the collection
  1770 //   clears _foregroundGCIsActive
  1771 //   returns
  1772 //
  1773 // The background collector in a loop iterating on the phases of the
  1774 //      collection
  1775 //   acquires the CGC_lock
  1776 //   sets _foregroundGCShouldWait
  1777 //   if _foregroundGCIsActive is set
  1778 //     clears _foregroundGCShouldWait, notifies _CGC_lock
  1779 //     waits on _CGC_lock for _foregroundGCIsActive to become false
  1780 //     and exits the loop.
  1781 //   otherwise
  1782 //     proceed with that phase of the collection
  1783 //     if the phase is a stop-the-world phase,
  1784 //       yield the baton once more just before enqueueing
  1785 //       the stop-world CMS operation (executed by the VM thread).
  1786 //   returns after all phases of the collection are done
  1787 //
  1789 void CMSCollector::acquire_control_and_collect(bool full,
  1790         bool clear_all_soft_refs) {
  1791   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
  1792   assert(!Thread::current()->is_ConcurrentGC_thread(),
  1793          "shouldn't try to acquire control from self!");
  1795   // Start the protocol for acquiring control of the
  1796   // collection from the background collector (aka CMS thread).
  1797   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1798          "VM thread should have CMS token");
  1799   // Remember the possibly interrupted state of an ongoing
  1800   // concurrent collection
  1801   CollectorState first_state = _collectorState;
  1803   // Signal to a possibly ongoing concurrent collection that
  1804   // we want to do a foreground collection.
  1805   _foregroundGCIsActive = true;
  1807   // Disable incremental mode during a foreground collection.
  1808   ICMSDisabler icms_disabler;
  1810   // release locks and wait for a notify from the background collector
  1811   // releasing the locks in only necessary for phases which
  1812   // do yields to improve the granularity of the collection.
  1813   assert_lock_strong(bitMapLock());
  1814   // We need to lock the Free list lock for the space that we are
  1815   // currently collecting.
  1816   assert(haveFreelistLocks(), "Must be holding free list locks");
  1817   bitMapLock()->unlock();
  1818   releaseFreelistLocks();
  1820     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  1821     if (_foregroundGCShouldWait) {
  1822       // We are going to be waiting for action for the CMS thread;
  1823       // it had better not be gone (for instance at shutdown)!
  1824       assert(ConcurrentMarkSweepThread::cmst() != NULL,
  1825              "CMS thread must be running");
  1826       // Wait here until the background collector gives us the go-ahead
  1827       ConcurrentMarkSweepThread::clear_CMS_flag(
  1828         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
  1829       // Get a possibly blocked CMS thread going:
  1830       //   Note that we set _foregroundGCIsActive true above,
  1831       //   without protection of the CGC_lock.
  1832       CGC_lock->notify();
  1833       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
  1834              "Possible deadlock");
  1835       while (_foregroundGCShouldWait) {
  1836         // wait for notification
  1837         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  1838         // Possibility of delay/starvation here, since CMS token does
  1839         // not know to give priority to VM thread? Actually, i think
  1840         // there wouldn't be any delay/starvation, but the proof of
  1841         // that "fact" (?) appears non-trivial. XXX 20011219YSR
  1843       ConcurrentMarkSweepThread::set_CMS_flag(
  1844         ConcurrentMarkSweepThread::CMS_vm_has_token);
  1847   // The CMS_token is already held.  Get back the other locks.
  1848   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  1849          "VM thread should have CMS token");
  1850   getFreelistLocks();
  1851   bitMapLock()->lock_without_safepoint_check();
  1852   if (TraceCMSState) {
  1853     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
  1854       INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
  1855     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
  1858   // Check if we need to do a compaction, or if not, whether
  1859   // we need to start the mark-sweep from scratch.
  1860   bool should_compact    = false;
  1861   bool should_start_over = false;
  1862   decide_foreground_collection_type(clear_all_soft_refs,
  1863     &should_compact, &should_start_over);
  1865 NOT_PRODUCT(
  1866   if (RotateCMSCollectionTypes) {
  1867     if (_cmsGen->debug_collection_type() ==
  1868         ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
  1869       should_compact = true;
  1870     } else if (_cmsGen->debug_collection_type() ==
  1871                ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
  1872       should_compact = false;
  1877   if (first_state > Idling) {
  1878     report_concurrent_mode_interruption();
  1881   set_did_compact(should_compact);
  1882   if (should_compact) {
  1883     // If the collection is being acquired from the background
  1884     // collector, there may be references on the discovered
  1885     // references lists that have NULL referents (being those
  1886     // that were concurrently cleared by a mutator) or
  1887     // that are no longer active (having been enqueued concurrently
  1888     // by the mutator).
  1889     // Scrub the list of those references because Mark-Sweep-Compact
  1890     // code assumes referents are not NULL and that all discovered
  1891     // Reference objects are active.
  1892     ref_processor()->clean_up_discovered_references();
  1894     if (first_state > Idling) {
  1895       save_heap_summary();
  1898     do_compaction_work(clear_all_soft_refs);
  1900     // Has the GC time limit been exceeded?
  1901     DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
  1902     size_t max_eden_size = young_gen->max_capacity() -
  1903                            young_gen->to()->capacity() -
  1904                            young_gen->from()->capacity();
  1905     GenCollectedHeap* gch = GenCollectedHeap::heap();
  1906     GCCause::Cause gc_cause = gch->gc_cause();
  1907     size_policy()->check_gc_overhead_limit(_young_gen->used(),
  1908                                            young_gen->eden()->used(),
  1909                                            _cmsGen->max_capacity(),
  1910                                            max_eden_size,
  1911                                            full,
  1912                                            gc_cause,
  1913                                            gch->collector_policy());
  1914   } else {
  1915     do_mark_sweep_work(clear_all_soft_refs, first_state,
  1916       should_start_over);
  1918   // Reset the expansion cause, now that we just completed
  1919   // a collection cycle.
  1920   clear_expansion_cause();
  1921   _foregroundGCIsActive = false;
  1922   return;
  1925 // Resize the tenured generation
  1926 // after obtaining the free list locks for the
  1927 // two generations.
  1928 void CMSCollector::compute_new_size() {
  1929   assert_locked_or_safepoint(Heap_lock);
  1930   FreelistLocker z(this);
  1931   MetaspaceGC::compute_new_size();
  1932   _cmsGen->compute_new_size_free_list();
  1935 // A work method used by foreground collection to determine
  1936 // what type of collection (compacting or not, continuing or fresh)
  1937 // it should do.
  1938 // NOTE: the intent is to make UseCMSCompactAtFullCollection
  1939 // and CMSCompactWhenClearAllSoftRefs the default in the future
  1940 // and do away with the flags after a suitable period.
  1941 void CMSCollector::decide_foreground_collection_type(
  1942   bool clear_all_soft_refs, bool* should_compact,
  1943   bool* should_start_over) {
  1944   // Normally, we'll compact only if the UseCMSCompactAtFullCollection
  1945   // flag is set, and we have either requested a System.gc() or
  1946   // the number of full gc's since the last concurrent cycle
  1947   // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
  1948   // or if an incremental collection has failed
  1949   GenCollectedHeap* gch = GenCollectedHeap::heap();
  1950   assert(gch->collector_policy()->is_two_generation_policy(),
  1951          "You may want to check the correctness of the following");
  1952   // Inform cms gen if this was due to partial collection failing.
  1953   // The CMS gen may use this fact to determine its expansion policy.
  1954   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
  1955     assert(!_cmsGen->incremental_collection_failed(),
  1956            "Should have been noticed, reacted to and cleared");
  1957     _cmsGen->set_incremental_collection_failed();
  1959   *should_compact =
  1960     UseCMSCompactAtFullCollection &&
  1961     ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
  1962      GCCause::is_user_requested_gc(gch->gc_cause()) ||
  1963      gch->incremental_collection_will_fail(true /* consult_young */));
  1964   *should_start_over = false;
  1965   if (clear_all_soft_refs && !*should_compact) {
  1966     // We are about to do a last ditch collection attempt
  1967     // so it would normally make sense to do a compaction
  1968     // to reclaim as much space as possible.
  1969     if (CMSCompactWhenClearAllSoftRefs) {
  1970       // Default: The rationale is that in this case either
  1971       // we are past the final marking phase, in which case
  1972       // we'd have to start over, or so little has been done
  1973       // that there's little point in saving that work. Compaction
  1974       // appears to be the sensible choice in either case.
  1975       *should_compact = true;
  1976     } else {
  1977       // We have been asked to clear all soft refs, but not to
  1978       // compact. Make sure that we aren't past the final checkpoint
  1979       // phase, for that is where we process soft refs. If we are already
  1980       // past that phase, we'll need to redo the refs discovery phase and
  1981       // if necessary clear soft refs that weren't previously
  1982       // cleared. We do so by remembering the phase in which
  1983       // we came in, and if we are past the refs processing
  1984       // phase, we'll choose to just redo the mark-sweep
  1985       // collection from scratch.
  1986       if (_collectorState > FinalMarking) {
  1987         // We are past the refs processing phase;
  1988         // start over and do a fresh synchronous CMS cycle
  1989         _collectorState = Resetting; // skip to reset to start new cycle
  1990         reset(false /* == !asynch */);
  1991         *should_start_over = true;
  1992       } // else we can continue a possibly ongoing current cycle
  1997 // A work method used by the foreground collector to do
  1998 // a mark-sweep-compact.
  1999 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
  2000   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2002   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
  2003   gc_timer->register_gc_start();
  2005   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
  2006   gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
  2008   GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
  2009   if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
  2010     gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
  2011       "collections passed to foreground collector", _full_gcs_since_conc_gc);
  2014   // Sample collection interval time and reset for collection pause.
  2015   if (UseAdaptiveSizePolicy) {
  2016     size_policy()->msc_collection_begin();
  2019   // Temporarily widen the span of the weak reference processing to
  2020   // the entire heap.
  2021   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
  2022   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
  2023   // Temporarily, clear the "is_alive_non_header" field of the
  2024   // reference processor.
  2025   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
  2026   // Temporarily make reference _processing_ single threaded (non-MT).
  2027   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
  2028   // Temporarily make refs discovery atomic
  2029   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
  2030   // Temporarily make reference _discovery_ single threaded (non-MT)
  2031   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
  2033   ref_processor()->set_enqueuing_is_done(false);
  2034   ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
  2035   ref_processor()->setup_policy(clear_all_soft_refs);
  2036   // If an asynchronous collection finishes, the _modUnionTable is
  2037   // all clear.  If we are assuming the collection from an asynchronous
  2038   // collection, clear the _modUnionTable.
  2039   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
  2040     "_modUnionTable should be clear if the baton was not passed");
  2041   _modUnionTable.clear_all();
  2042   assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
  2043     "mod union for klasses should be clear if the baton was passed");
  2044   _ct->klass_rem_set()->clear_mod_union();
  2046   // We must adjust the allocation statistics being maintained
  2047   // in the free list space. We do so by reading and clearing
  2048   // the sweep timer and updating the block flux rate estimates below.
  2049   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
  2050   if (_inter_sweep_timer.is_active()) {
  2051     _inter_sweep_timer.stop();
  2052     // Note that we do not use this sample to update the _inter_sweep_estimate.
  2053     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
  2054                                             _inter_sweep_estimate.padded_average(),
  2055                                             _intra_sweep_estimate.padded_average());
  2058   GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
  2059     ref_processor(), clear_all_soft_refs);
  2060   #ifdef ASSERT
  2061     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
  2062     size_t free_size = cms_space->free();
  2063     assert(free_size ==
  2064            pointer_delta(cms_space->end(), cms_space->compaction_top())
  2065            * HeapWordSize,
  2066       "All the free space should be compacted into one chunk at top");
  2067     assert(cms_space->dictionary()->total_chunk_size(
  2068                                       debug_only(cms_space->freelistLock())) == 0 ||
  2069            cms_space->totalSizeInIndexedFreeLists() == 0,
  2070       "All the free space should be in a single chunk");
  2071     size_t num = cms_space->totalCount();
  2072     assert((free_size == 0 && num == 0) ||
  2073            (free_size > 0  && (num == 1 || num == 2)),
  2074          "There should be at most 2 free chunks after compaction");
  2075   #endif // ASSERT
  2076   _collectorState = Resetting;
  2077   assert(_restart_addr == NULL,
  2078          "Should have been NULL'd before baton was passed");
  2079   reset(false /* == !asynch */);
  2080   _cmsGen->reset_after_compaction();
  2081   _concurrent_cycles_since_last_unload = 0;
  2083   // Clear any data recorded in the PLAB chunk arrays.
  2084   if (_survivor_plab_array != NULL) {
  2085     reset_survivor_plab_arrays();
  2088   // Adjust the per-size allocation stats for the next epoch.
  2089   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
  2090   // Restart the "inter sweep timer" for the next epoch.
  2091   _inter_sweep_timer.reset();
  2092   _inter_sweep_timer.start();
  2094   // Sample collection pause time and reset for collection interval.
  2095   if (UseAdaptiveSizePolicy) {
  2096     size_policy()->msc_collection_end(gch->gc_cause());
  2099   gc_timer->register_gc_end();
  2101   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
  2103   // For a mark-sweep-compact, compute_new_size() will be called
  2104   // in the heap's do_collection() method.
  2107 // A work method used by the foreground collector to do
  2108 // a mark-sweep, after taking over from a possibly on-going
  2109 // concurrent mark-sweep collection.
  2110 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
  2111   CollectorState first_state, bool should_start_over) {
  2112   if (PrintGC && Verbose) {
  2113     gclog_or_tty->print_cr("Pass concurrent collection to foreground "
  2114       "collector with count %d",
  2115       _full_gcs_since_conc_gc);
  2117   switch (_collectorState) {
  2118     case Idling:
  2119       if (first_state == Idling || should_start_over) {
  2120         // The background GC was not active, or should
  2121         // restarted from scratch;  start the cycle.
  2122         _collectorState = InitialMarking;
  2124       // If first_state was not Idling, then a background GC
  2125       // was in progress and has now finished.  No need to do it
  2126       // again.  Leave the state as Idling.
  2127       break;
  2128     case Precleaning:
  2129       // In the foreground case don't do the precleaning since
  2130       // it is not done concurrently and there is extra work
  2131       // required.
  2132       _collectorState = FinalMarking;
  2134   collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
  2136   // For a mark-sweep, compute_new_size() will be called
  2137   // in the heap's do_collection() method.
  2141 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
  2142   DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
  2143   EdenSpace* eden_space = dng->eden();
  2144   ContiguousSpace* from_space = dng->from();
  2145   ContiguousSpace* to_space   = dng->to();
  2146   // Eden
  2147   if (_eden_chunk_array != NULL) {
  2148     gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
  2149                            eden_space->bottom(), eden_space->top(),
  2150                            eden_space->end(), eden_space->capacity());
  2151     gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
  2152                            "_eden_chunk_capacity=" SIZE_FORMAT,
  2153                            _eden_chunk_index, _eden_chunk_capacity);
  2154     for (size_t i = 0; i < _eden_chunk_index; i++) {
  2155       gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
  2156                              i, _eden_chunk_array[i]);
  2159   // Survivor
  2160   if (_survivor_chunk_array != NULL) {
  2161     gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
  2162                            from_space->bottom(), from_space->top(),
  2163                            from_space->end(), from_space->capacity());
  2164     gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
  2165                            "_survivor_chunk_capacity=" SIZE_FORMAT,
  2166                            _survivor_chunk_index, _survivor_chunk_capacity);
  2167     for (size_t i = 0; i < _survivor_chunk_index; i++) {
  2168       gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
  2169                              i, _survivor_chunk_array[i]);
  2174 void CMSCollector::getFreelistLocks() const {
  2175   // Get locks for all free lists in all generations that this
  2176   // collector is responsible for
  2177   _cmsGen->freelistLock()->lock_without_safepoint_check();
  2180 void CMSCollector::releaseFreelistLocks() const {
  2181   // Release locks for all free lists in all generations that this
  2182   // collector is responsible for
  2183   _cmsGen->freelistLock()->unlock();
  2186 bool CMSCollector::haveFreelistLocks() const {
  2187   // Check locks for all free lists in all generations that this
  2188   // collector is responsible for
  2189   assert_lock_strong(_cmsGen->freelistLock());
  2190   PRODUCT_ONLY(ShouldNotReachHere());
  2191   return true;
  2194 // A utility class that is used by the CMS collector to
  2195 // temporarily "release" the foreground collector from its
  2196 // usual obligation to wait for the background collector to
  2197 // complete an ongoing phase before proceeding.
  2198 class ReleaseForegroundGC: public StackObj {
  2199  private:
  2200   CMSCollector* _c;
  2201  public:
  2202   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
  2203     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
  2204     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2205     // allow a potentially blocked foreground collector to proceed
  2206     _c->_foregroundGCShouldWait = false;
  2207     if (_c->_foregroundGCIsActive) {
  2208       CGC_lock->notify();
  2210     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2211            "Possible deadlock");
  2214   ~ReleaseForegroundGC() {
  2215     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
  2216     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2217     _c->_foregroundGCShouldWait = true;
  2219 };
  2221 // There are separate collect_in_background and collect_in_foreground because of
  2222 // the different locking requirements of the background collector and the
  2223 // foreground collector.  There was originally an attempt to share
  2224 // one "collect" method between the background collector and the foreground
  2225 // collector but the if-then-else required made it cleaner to have
  2226 // separate methods.
  2227 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
  2228   assert(Thread::current()->is_ConcurrentGC_thread(),
  2229     "A CMS asynchronous collection is only allowed on a CMS thread.");
  2231   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2233     bool safepoint_check = Mutex::_no_safepoint_check_flag;
  2234     MutexLockerEx hl(Heap_lock, safepoint_check);
  2235     FreelistLocker fll(this);
  2236     MutexLockerEx x(CGC_lock, safepoint_check);
  2237     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
  2238       // The foreground collector is active or we're
  2239       // not using asynchronous collections.  Skip this
  2240       // background collection.
  2241       assert(!_foregroundGCShouldWait, "Should be clear");
  2242       return;
  2243     } else {
  2244       assert(_collectorState == Idling, "Should be idling before start.");
  2245       _collectorState = InitialMarking;
  2246       register_gc_start(cause);
  2247       // Reset the expansion cause, now that we are about to begin
  2248       // a new cycle.
  2249       clear_expansion_cause();
  2251       // Clear the MetaspaceGC flag since a concurrent collection
  2252       // is starting but also clear it after the collection.
  2253       MetaspaceGC::set_should_concurrent_collect(false);
  2255     // Decide if we want to enable class unloading as part of the
  2256     // ensuing concurrent GC cycle.
  2257     update_should_unload_classes();
  2258     _full_gc_requested = false;           // acks all outstanding full gc requests
  2259     _full_gc_cause = GCCause::_no_gc;
  2260     // Signal that we are about to start a collection
  2261     gch->increment_total_full_collections();  // ... starting a collection cycle
  2262     _collection_count_start = gch->total_full_collections();
  2265   // Used for PrintGC
  2266   size_t prev_used;
  2267   if (PrintGC && Verbose) {
  2268     prev_used = _cmsGen->used(); // XXXPERM
  2271   // The change of the collection state is normally done at this level;
  2272   // the exceptions are phases that are executed while the world is
  2273   // stopped.  For those phases the change of state is done while the
  2274   // world is stopped.  For baton passing purposes this allows the
  2275   // background collector to finish the phase and change state atomically.
  2276   // The foreground collector cannot wait on a phase that is done
  2277   // while the world is stopped because the foreground collector already
  2278   // has the world stopped and would deadlock.
  2279   while (_collectorState != Idling) {
  2280     if (TraceCMSState) {
  2281       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2282         Thread::current(), _collectorState);
  2284     // The foreground collector
  2285     //   holds the Heap_lock throughout its collection.
  2286     //   holds the CMS token (but not the lock)
  2287     //     except while it is waiting for the background collector to yield.
  2288     //
  2289     // The foreground collector should be blocked (not for long)
  2290     //   if the background collector is about to start a phase
  2291     //   executed with world stopped.  If the background
  2292     //   collector has already started such a phase, the
  2293     //   foreground collector is blocked waiting for the
  2294     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
  2295     //   are executed in the VM thread.
  2296     //
  2297     // The locking order is
  2298     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
  2299     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
  2300     //   CMS token  (claimed in
  2301     //                stop_world_and_do() -->
  2302     //                  safepoint_synchronize() -->
  2303     //                    CMSThread::synchronize())
  2306       // Check if the FG collector wants us to yield.
  2307       CMSTokenSync x(true); // is cms thread
  2308       if (waitForForegroundGC()) {
  2309         // We yielded to a foreground GC, nothing more to be
  2310         // done this round.
  2311         assert(_foregroundGCShouldWait == false, "We set it to false in "
  2312                "waitForForegroundGC()");
  2313         if (TraceCMSState) {
  2314           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2315             " exiting collection CMS state %d",
  2316             Thread::current(), _collectorState);
  2318         return;
  2319       } else {
  2320         // The background collector can run but check to see if the
  2321         // foreground collector has done a collection while the
  2322         // background collector was waiting to get the CGC_lock
  2323         // above.  If yes, break so that _foregroundGCShouldWait
  2324         // is cleared before returning.
  2325         if (_collectorState == Idling) {
  2326           break;
  2331     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
  2332       "should be waiting");
  2334     switch (_collectorState) {
  2335       case InitialMarking:
  2337           ReleaseForegroundGC x(this);
  2338           stats().record_cms_begin();
  2339           VM_CMS_Initial_Mark initial_mark_op(this);
  2340           VMThread::execute(&initial_mark_op);
  2342         // The collector state may be any legal state at this point
  2343         // since the background collector may have yielded to the
  2344         // foreground collector.
  2345         break;
  2346       case Marking:
  2347         // initial marking in checkpointRootsInitialWork has been completed
  2348         if (markFromRoots(true)) { // we were successful
  2349           assert(_collectorState == Precleaning, "Collector state should "
  2350             "have changed");
  2351         } else {
  2352           assert(_foregroundGCIsActive, "Internal state inconsistency");
  2354         break;
  2355       case Precleaning:
  2356         if (UseAdaptiveSizePolicy) {
  2357           size_policy()->concurrent_precleaning_begin();
  2359         // marking from roots in markFromRoots has been completed
  2360         preclean();
  2361         if (UseAdaptiveSizePolicy) {
  2362           size_policy()->concurrent_precleaning_end();
  2364         assert(_collectorState == AbortablePreclean ||
  2365                _collectorState == FinalMarking,
  2366                "Collector state should have changed");
  2367         break;
  2368       case AbortablePreclean:
  2369         if (UseAdaptiveSizePolicy) {
  2370         size_policy()->concurrent_phases_resume();
  2372         abortable_preclean();
  2373         if (UseAdaptiveSizePolicy) {
  2374           size_policy()->concurrent_precleaning_end();
  2376         assert(_collectorState == FinalMarking, "Collector state should "
  2377           "have changed");
  2378         break;
  2379       case FinalMarking:
  2381           ReleaseForegroundGC x(this);
  2383           VM_CMS_Final_Remark final_remark_op(this);
  2384           VMThread::execute(&final_remark_op);
  2386         assert(_foregroundGCShouldWait, "block post-condition");
  2387         break;
  2388       case Sweeping:
  2389         if (UseAdaptiveSizePolicy) {
  2390           size_policy()->concurrent_sweeping_begin();
  2392         // final marking in checkpointRootsFinal has been completed
  2393         sweep(true);
  2394         assert(_collectorState == Resizing, "Collector state change "
  2395           "to Resizing must be done under the free_list_lock");
  2396         _full_gcs_since_conc_gc = 0;
  2398         // Stop the timers for adaptive size policy for the concurrent phases
  2399         if (UseAdaptiveSizePolicy) {
  2400           size_policy()->concurrent_sweeping_end();
  2401           size_policy()->concurrent_phases_end(gch->gc_cause(),
  2402                                              gch->prev_gen(_cmsGen)->capacity(),
  2403                                              _cmsGen->free());
  2406       case Resizing: {
  2407         // Sweeping has been completed...
  2408         // At this point the background collection has completed.
  2409         // Don't move the call to compute_new_size() down
  2410         // into code that might be executed if the background
  2411         // collection was preempted.
  2413           ReleaseForegroundGC x(this);   // unblock FG collection
  2414           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
  2415           CMSTokenSync        z(true);   // not strictly needed.
  2416           if (_collectorState == Resizing) {
  2417             compute_new_size();
  2418             save_heap_summary();
  2419             _collectorState = Resetting;
  2420           } else {
  2421             assert(_collectorState == Idling, "The state should only change"
  2422                    " because the foreground collector has finished the collection");
  2425         break;
  2427       case Resetting:
  2428         // CMS heap resizing has been completed
  2429         reset(true);
  2430         assert(_collectorState == Idling, "Collector state should "
  2431           "have changed");
  2433         MetaspaceGC::set_should_concurrent_collect(false);
  2435         stats().record_cms_end();
  2436         // Don't move the concurrent_phases_end() and compute_new_size()
  2437         // calls to here because a preempted background collection
  2438         // has it's state set to "Resetting".
  2439         break;
  2440       case Idling:
  2441       default:
  2442         ShouldNotReachHere();
  2443         break;
  2445     if (TraceCMSState) {
  2446       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2447         Thread::current(), _collectorState);
  2449     assert(_foregroundGCShouldWait, "block post-condition");
  2452   // Should this be in gc_epilogue?
  2453   collector_policy()->counters()->update_counters();
  2456     // Clear _foregroundGCShouldWait and, in the event that the
  2457     // foreground collector is waiting, notify it, before
  2458     // returning.
  2459     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2460     _foregroundGCShouldWait = false;
  2461     if (_foregroundGCIsActive) {
  2462       CGC_lock->notify();
  2464     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2465            "Possible deadlock");
  2467   if (TraceCMSState) {
  2468     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2469       " exiting collection CMS state %d",
  2470       Thread::current(), _collectorState);
  2472   if (PrintGC && Verbose) {
  2473     _cmsGen->print_heap_change(prev_used);
  2477 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
  2478   if (!_cms_start_registered) {
  2479     register_gc_start(cause);
  2483 void CMSCollector::register_gc_start(GCCause::Cause cause) {
  2484   _cms_start_registered = true;
  2485   _gc_timer_cm->register_gc_start();
  2486   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
  2489 void CMSCollector::register_gc_end() {
  2490   if (_cms_start_registered) {
  2491     report_heap_summary(GCWhen::AfterGC);
  2493     _gc_timer_cm->register_gc_end();
  2494     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
  2495     _cms_start_registered = false;
  2499 void CMSCollector::save_heap_summary() {
  2500   GenCollectedHeap* gch = GenCollectedHeap::heap();
  2501   _last_heap_summary = gch->create_heap_summary();
  2502   _last_metaspace_summary = gch->create_metaspace_summary();
  2505 void CMSCollector::report_heap_summary(GCWhen::Type when) {
  2506   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
  2507   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
  2510 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
  2511   assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
  2512          "Foreground collector should be waiting, not executing");
  2513   assert(Thread::current()->is_VM_thread(), "A foreground collection"
  2514     "may only be done by the VM Thread with the world stopped");
  2515   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
  2516          "VM thread should have CMS token");
  2518   NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
  2519     true, NULL);)
  2520   if (UseAdaptiveSizePolicy) {
  2521     size_policy()->ms_collection_begin();
  2523   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
  2525   HandleMark hm;  // Discard invalid handles created during verification
  2527   if (VerifyBeforeGC &&
  2528       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2529     Universe::verify();
  2532   // Snapshot the soft reference policy to be used in this collection cycle.
  2533   ref_processor()->setup_policy(clear_all_soft_refs);
  2535   // Decide if class unloading should be done
  2536   update_should_unload_classes();
  2538   bool init_mark_was_synchronous = false; // until proven otherwise
  2539   while (_collectorState != Idling) {
  2540     if (TraceCMSState) {
  2541       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
  2542         Thread::current(), _collectorState);
  2544     switch (_collectorState) {
  2545       case InitialMarking:
  2546         register_foreground_gc_start(cause);
  2547         init_mark_was_synchronous = true;  // fact to be exploited in re-mark
  2548         checkpointRootsInitial(false);
  2549         assert(_collectorState == Marking, "Collector state should have changed"
  2550           " within checkpointRootsInitial()");
  2551         break;
  2552       case Marking:
  2553         // initial marking in checkpointRootsInitialWork has been completed
  2554         if (VerifyDuringGC &&
  2555             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2556           Universe::verify("Verify before initial mark: ");
  2559           bool res = markFromRoots(false);
  2560           assert(res && _collectorState == FinalMarking, "Collector state should "
  2561             "have changed");
  2562           break;
  2564       case FinalMarking:
  2565         if (VerifyDuringGC &&
  2566             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2567           Universe::verify("Verify before re-mark: ");
  2569         checkpointRootsFinal(false, clear_all_soft_refs,
  2570                              init_mark_was_synchronous);
  2571         assert(_collectorState == Sweeping, "Collector state should not "
  2572           "have changed within checkpointRootsFinal()");
  2573         break;
  2574       case Sweeping:
  2575         // final marking in checkpointRootsFinal has been completed
  2576         if (VerifyDuringGC &&
  2577             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2578           Universe::verify("Verify before sweep: ");
  2580         sweep(false);
  2581         assert(_collectorState == Resizing, "Incorrect state");
  2582         break;
  2583       case Resizing: {
  2584         // Sweeping has been completed; the actual resize in this case
  2585         // is done separately; nothing to be done in this state.
  2586         _collectorState = Resetting;
  2587         break;
  2589       case Resetting:
  2590         // The heap has been resized.
  2591         if (VerifyDuringGC &&
  2592             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2593           Universe::verify("Verify before reset: ");
  2595         save_heap_summary();
  2596         reset(false);
  2597         assert(_collectorState == Idling, "Collector state should "
  2598           "have changed");
  2599         break;
  2600       case Precleaning:
  2601       case AbortablePreclean:
  2602         // Elide the preclean phase
  2603         _collectorState = FinalMarking;
  2604         break;
  2605       default:
  2606         ShouldNotReachHere();
  2608     if (TraceCMSState) {
  2609       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
  2610         Thread::current(), _collectorState);
  2614   if (UseAdaptiveSizePolicy) {
  2615     GenCollectedHeap* gch = GenCollectedHeap::heap();
  2616     size_policy()->ms_collection_end(gch->gc_cause());
  2619   if (VerifyAfterGC &&
  2620       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  2621     Universe::verify();
  2623   if (TraceCMSState) {
  2624     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
  2625       " exiting collection CMS state %d",
  2626       Thread::current(), _collectorState);
  2630 bool CMSCollector::waitForForegroundGC() {
  2631   bool res = false;
  2632   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  2633          "CMS thread should have CMS token");
  2634   // Block the foreground collector until the
  2635   // background collectors decides whether to
  2636   // yield.
  2637   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  2638   _foregroundGCShouldWait = true;
  2639   if (_foregroundGCIsActive) {
  2640     // The background collector yields to the
  2641     // foreground collector and returns a value
  2642     // indicating that it has yielded.  The foreground
  2643     // collector can proceed.
  2644     res = true;
  2645     _foregroundGCShouldWait = false;
  2646     ConcurrentMarkSweepThread::clear_CMS_flag(
  2647       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2648     ConcurrentMarkSweepThread::set_CMS_flag(
  2649       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2650     // Get a possibly blocked foreground thread going
  2651     CGC_lock->notify();
  2652     if (TraceCMSState) {
  2653       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
  2654         Thread::current(), _collectorState);
  2656     while (_foregroundGCIsActive) {
  2657       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  2659     ConcurrentMarkSweepThread::set_CMS_flag(
  2660       ConcurrentMarkSweepThread::CMS_cms_has_token);
  2661     ConcurrentMarkSweepThread::clear_CMS_flag(
  2662       ConcurrentMarkSweepThread::CMS_cms_wants_token);
  2664   if (TraceCMSState) {
  2665     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
  2666       Thread::current(), _collectorState);
  2668   return res;
  2671 // Because of the need to lock the free lists and other structures in
  2672 // the collector, common to all the generations that the collector is
  2673 // collecting, we need the gc_prologues of individual CMS generations
  2674 // delegate to their collector. It may have been simpler had the
  2675 // current infrastructure allowed one to call a prologue on a
  2676 // collector. In the absence of that we have the generation's
  2677 // prologue delegate to the collector, which delegates back
  2678 // some "local" work to a worker method in the individual generations
  2679 // that it's responsible for collecting, while itself doing any
  2680 // work common to all generations it's responsible for. A similar
  2681 // comment applies to the  gc_epilogue()'s.
  2682 // The role of the varaible _between_prologue_and_epilogue is to
  2683 // enforce the invocation protocol.
  2684 void CMSCollector::gc_prologue(bool full) {
  2685   // Call gc_prologue_work() for the CMSGen
  2686   // we are responsible for.
  2688   // The following locking discipline assumes that we are only called
  2689   // when the world is stopped.
  2690   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
  2692   // The CMSCollector prologue must call the gc_prologues for the
  2693   // "generations" that it's responsible
  2694   // for.
  2696   assert(   Thread::current()->is_VM_thread()
  2697          || (   CMSScavengeBeforeRemark
  2698              && Thread::current()->is_ConcurrentGC_thread()),
  2699          "Incorrect thread type for prologue execution");
  2701   if (_between_prologue_and_epilogue) {
  2702     // We have already been invoked; this is a gc_prologue delegation
  2703     // from yet another CMS generation that we are responsible for, just
  2704     // ignore it since all relevant work has already been done.
  2705     return;
  2708   // set a bit saying prologue has been called; cleared in epilogue
  2709   _between_prologue_and_epilogue = true;
  2710   // Claim locks for common data structures, then call gc_prologue_work()
  2711   // for each CMSGen.
  2713   getFreelistLocks();   // gets free list locks on constituent spaces
  2714   bitMapLock()->lock_without_safepoint_check();
  2716   // Should call gc_prologue_work() for all cms gens we are responsible for
  2717   bool duringMarking =    _collectorState >= Marking
  2718                          && _collectorState < Sweeping;
  2720   // The young collections clear the modified oops state, which tells if
  2721   // there are any modified oops in the class. The remark phase also needs
  2722   // that information. Tell the young collection to save the union of all
  2723   // modified klasses.
  2724   if (duringMarking) {
  2725     _ct->klass_rem_set()->set_accumulate_modified_oops(true);
  2728   bool registerClosure = duringMarking;
  2730   ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
  2731                                                &_modUnionClosurePar
  2732                                                : &_modUnionClosure;
  2733   _cmsGen->gc_prologue_work(full, registerClosure, muc);
  2735   if (!full) {
  2736     stats().record_gc0_begin();
  2740 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
  2742   _capacity_at_prologue = capacity();
  2743   _used_at_prologue = used();
  2745   // Delegate to CMScollector which knows how to coordinate between
  2746   // this and any other CMS generations that it is responsible for
  2747   // collecting.
  2748   collector()->gc_prologue(full);
  2751 // This is a "private" interface for use by this generation's CMSCollector.
  2752 // Not to be called directly by any other entity (for instance,
  2753 // GenCollectedHeap, which calls the "public" gc_prologue method above).
  2754 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
  2755   bool registerClosure, ModUnionClosure* modUnionClosure) {
  2756   assert(!incremental_collection_failed(), "Shouldn't be set yet");
  2757   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
  2758     "Should be NULL");
  2759   if (registerClosure) {
  2760     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
  2762   cmsSpace()->gc_prologue();
  2763   // Clear stat counters
  2764   NOT_PRODUCT(
  2765     assert(_numObjectsPromoted == 0, "check");
  2766     assert(_numWordsPromoted   == 0, "check");
  2767     if (Verbose && PrintGC) {
  2768       gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
  2769                           SIZE_FORMAT" bytes concurrently",
  2770       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
  2772     _numObjectsAllocated = 0;
  2773     _numWordsAllocated   = 0;
  2777 void CMSCollector::gc_epilogue(bool full) {
  2778   // The following locking discipline assumes that we are only called
  2779   // when the world is stopped.
  2780   assert(SafepointSynchronize::is_at_safepoint(),
  2781          "world is stopped assumption");
  2783   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
  2784   // if linear allocation blocks need to be appropriately marked to allow the
  2785   // the blocks to be parsable. We also check here whether we need to nudge the
  2786   // CMS collector thread to start a new cycle (if it's not already active).
  2787   assert(   Thread::current()->is_VM_thread()
  2788          || (   CMSScavengeBeforeRemark
  2789              && Thread::current()->is_ConcurrentGC_thread()),
  2790          "Incorrect thread type for epilogue execution");
  2792   if (!_between_prologue_and_epilogue) {
  2793     // We have already been invoked; this is a gc_epilogue delegation
  2794     // from yet another CMS generation that we are responsible for, just
  2795     // ignore it since all relevant work has already been done.
  2796     return;
  2798   assert(haveFreelistLocks(), "must have freelist locks");
  2799   assert_lock_strong(bitMapLock());
  2801   _ct->klass_rem_set()->set_accumulate_modified_oops(false);
  2803   _cmsGen->gc_epilogue_work(full);
  2805   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
  2806     // in case sampling was not already enabled, enable it
  2807     _start_sampling = true;
  2809   // reset _eden_chunk_array so sampling starts afresh
  2810   _eden_chunk_index = 0;
  2812   size_t cms_used   = _cmsGen->cmsSpace()->used();
  2814   // update performance counters - this uses a special version of
  2815   // update_counters() that allows the utilization to be passed as a
  2816   // parameter, avoiding multiple calls to used().
  2817   //
  2818   _cmsGen->update_counters(cms_used);
  2820   if (CMSIncrementalMode) {
  2821     icms_update_allocation_limits();
  2824   bitMapLock()->unlock();
  2825   releaseFreelistLocks();
  2827   if (!CleanChunkPoolAsync) {
  2828     Chunk::clean_chunk_pool();
  2831   set_did_compact(false);
  2832   _between_prologue_and_epilogue = false;  // ready for next cycle
  2835 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
  2836   collector()->gc_epilogue(full);
  2838   // Also reset promotion tracking in par gc thread states.
  2839   if (CollectedHeap::use_parallel_gc_threads()) {
  2840     for (uint i = 0; i < ParallelGCThreads; i++) {
  2841       _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
  2846 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
  2847   assert(!incremental_collection_failed(), "Should have been cleared");
  2848   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
  2849   cmsSpace()->gc_epilogue();
  2850     // Print stat counters
  2851   NOT_PRODUCT(
  2852     assert(_numObjectsAllocated == 0, "check");
  2853     assert(_numWordsAllocated == 0, "check");
  2854     if (Verbose && PrintGC) {
  2855       gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
  2856                           SIZE_FORMAT" bytes",
  2857                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
  2859     _numObjectsPromoted = 0;
  2860     _numWordsPromoted   = 0;
  2863   if (PrintGC && Verbose) {
  2864     // Call down the chain in contiguous_available needs the freelistLock
  2865     // so print this out before releasing the freeListLock.
  2866     gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
  2867                         contiguous_available());
  2871 #ifndef PRODUCT
  2872 bool CMSCollector::have_cms_token() {
  2873   Thread* thr = Thread::current();
  2874   if (thr->is_VM_thread()) {
  2875     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
  2876   } else if (thr->is_ConcurrentGC_thread()) {
  2877     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
  2878   } else if (thr->is_GC_task_thread()) {
  2879     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
  2880            ParGCRareEvent_lock->owned_by_self();
  2882   return false;
  2884 #endif
  2886 // Check reachability of the given heap address in CMS generation,
  2887 // treating all other generations as roots.
  2888 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
  2889   // We could "guarantee" below, rather than assert, but i'll
  2890   // leave these as "asserts" so that an adventurous debugger
  2891   // could try this in the product build provided some subset of
  2892   // the conditions were met, provided they were intersted in the
  2893   // results and knew that the computation below wouldn't interfere
  2894   // with other concurrent computations mutating the structures
  2895   // being read or written.
  2896   assert(SafepointSynchronize::is_at_safepoint(),
  2897          "Else mutations in object graph will make answer suspect");
  2898   assert(have_cms_token(), "Should hold cms token");
  2899   assert(haveFreelistLocks(), "must hold free list locks");
  2900   assert_lock_strong(bitMapLock());
  2902   // Clear the marking bit map array before starting, but, just
  2903   // for kicks, first report if the given address is already marked
  2904   gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
  2905                 _markBitMap.isMarked(addr) ? "" : " not");
  2907   if (verify_after_remark()) {
  2908     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2909     bool result = verification_mark_bm()->isMarked(addr);
  2910     gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
  2911                            result ? "IS" : "is NOT");
  2912     return result;
  2913   } else {
  2914     gclog_or_tty->print_cr("Could not compute result");
  2915     return false;
  2920 void
  2921 CMSCollector::print_on_error(outputStream* st) {
  2922   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
  2923   if (collector != NULL) {
  2924     CMSBitMap* bitmap = &collector->_markBitMap;
  2925     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
  2926     bitmap->print_on_error(st, " Bits: ");
  2928     st->cr();
  2930     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
  2931     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
  2932     mut_bitmap->print_on_error(st, " Bits: ");
  2936 ////////////////////////////////////////////////////////
  2937 // CMS Verification Support
  2938 ////////////////////////////////////////////////////////
  2939 // Following the remark phase, the following invariant
  2940 // should hold -- each object in the CMS heap which is
  2941 // marked in markBitMap() should be marked in the verification_mark_bm().
  2943 class VerifyMarkedClosure: public BitMapClosure {
  2944   CMSBitMap* _marks;
  2945   bool       _failed;
  2947  public:
  2948   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
  2950   bool do_bit(size_t offset) {
  2951     HeapWord* addr = _marks->offsetToHeapWord(offset);
  2952     if (!_marks->isMarked(addr)) {
  2953       oop(addr)->print_on(gclog_or_tty);
  2954       gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
  2955       _failed = true;
  2957     return true;
  2960   bool failed() { return _failed; }
  2961 };
  2963 bool CMSCollector::verify_after_remark(bool silent) {
  2964   if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
  2965   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
  2966   static bool init = false;
  2968   assert(SafepointSynchronize::is_at_safepoint(),
  2969          "Else mutations in object graph will make answer suspect");
  2970   assert(have_cms_token(),
  2971          "Else there may be mutual interference in use of "
  2972          " verification data structures");
  2973   assert(_collectorState > Marking && _collectorState <= Sweeping,
  2974          "Else marking info checked here may be obsolete");
  2975   assert(haveFreelistLocks(), "must hold free list locks");
  2976   assert_lock_strong(bitMapLock());
  2979   // Allocate marking bit map if not already allocated
  2980   if (!init) { // first time
  2981     if (!verification_mark_bm()->allocate(_span)) {
  2982       return false;
  2984     init = true;
  2987   assert(verification_mark_stack()->isEmpty(), "Should be empty");
  2989   // Turn off refs discovery -- so we will be tracing through refs.
  2990   // This is as intended, because by this time
  2991   // GC must already have cleared any refs that need to be cleared,
  2992   // and traced those that need to be marked; moreover,
  2993   // the marking done here is not going to intefere in any
  2994   // way with the marking information used by GC.
  2995   NoRefDiscovery no_discovery(ref_processor());
  2997   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  2999   // Clear any marks from a previous round
  3000   verification_mark_bm()->clear_all();
  3001   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
  3002   verify_work_stacks_empty();
  3004   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3005   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  3006   // Update the saved marks which may affect the root scans.
  3007   gch->save_marks();
  3009   if (CMSRemarkVerifyVariant == 1) {
  3010     // In this first variant of verification, we complete
  3011     // all marking, then check if the new marks-verctor is
  3012     // a subset of the CMS marks-vector.
  3013     verify_after_remark_work_1();
  3014   } else if (CMSRemarkVerifyVariant == 2) {
  3015     // In this second variant of verification, we flag an error
  3016     // (i.e. an object reachable in the new marks-vector not reachable
  3017     // in the CMS marks-vector) immediately, also indicating the
  3018     // identify of an object (A) that references the unmarked object (B) --
  3019     // presumably, a mutation to A failed to be picked up by preclean/remark?
  3020     verify_after_remark_work_2();
  3021   } else {
  3022     warning("Unrecognized value %d for CMSRemarkVerifyVariant",
  3023             CMSRemarkVerifyVariant);
  3025   if (!silent) gclog_or_tty->print(" done] ");
  3026   return true;
  3029 void CMSCollector::verify_after_remark_work_1() {
  3030   ResourceMark rm;
  3031   HandleMark  hm;
  3032   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3034   // Get a clear set of claim bits for the strong roots processing to work with.
  3035   ClassLoaderDataGraph::clear_claimed_marks();
  3037   // Mark from roots one level into CMS
  3038   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
  3039   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  3041   gch->gen_process_strong_roots(_cmsGen->level(),
  3042                                 true,   // younger gens are roots
  3043                                 true,   // activate StrongRootsScope
  3044                                 false,  // not scavenging
  3045                                 SharedHeap::ScanningOption(roots_scanning_options()),
  3046                                 &notOlder,
  3047                                 true,   // walk code active on stacks
  3048                                 NULL,
  3049                                 NULL); // SSS: Provide correct closure
  3051   // Now mark from the roots
  3052   MarkFromRootsClosure markFromRootsClosure(this, _span,
  3053     verification_mark_bm(), verification_mark_stack(),
  3054     false /* don't yield */, true /* verifying */);
  3055   assert(_restart_addr == NULL, "Expected pre-condition");
  3056   verification_mark_bm()->iterate(&markFromRootsClosure);
  3057   while (_restart_addr != NULL) {
  3058     // Deal with stack overflow: by restarting at the indicated
  3059     // address.
  3060     HeapWord* ra = _restart_addr;
  3061     markFromRootsClosure.reset(ra);
  3062     _restart_addr = NULL;
  3063     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  3065   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  3066   verify_work_stacks_empty();
  3068   // Marking completed -- now verify that each bit marked in
  3069   // verification_mark_bm() is also marked in markBitMap(); flag all
  3070   // errors by printing corresponding objects.
  3071   VerifyMarkedClosure vcl(markBitMap());
  3072   verification_mark_bm()->iterate(&vcl);
  3073   if (vcl.failed()) {
  3074     gclog_or_tty->print("Verification failed");
  3075     Universe::heap()->print_on(gclog_or_tty);
  3076     fatal("CMS: failed marking verification after remark");
  3080 class VerifyKlassOopsKlassClosure : public KlassClosure {
  3081   class VerifyKlassOopsClosure : public OopClosure {
  3082     CMSBitMap* _bitmap;
  3083    public:
  3084     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
  3085     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
  3086     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
  3087   } _oop_closure;
  3088  public:
  3089   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
  3090   void do_klass(Klass* k) {
  3091     k->oops_do(&_oop_closure);
  3093 };
  3095 void CMSCollector::verify_after_remark_work_2() {
  3096   ResourceMark rm;
  3097   HandleMark  hm;
  3098   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3100   // Get a clear set of claim bits for the strong roots processing to work with.
  3101   ClassLoaderDataGraph::clear_claimed_marks();
  3103   // Mark from roots one level into CMS
  3104   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
  3105                                      markBitMap());
  3106   CMKlassClosure klass_closure(&notOlder);
  3108   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  3109   gch->gen_process_strong_roots(_cmsGen->level(),
  3110                                 true,   // younger gens are roots
  3111                                 true,   // activate StrongRootsScope
  3112                                 false,  // not scavenging
  3113                                 SharedHeap::ScanningOption(roots_scanning_options()),
  3114                                 &notOlder,
  3115                                 true,   // walk code active on stacks
  3116                                 NULL,
  3117                                 &klass_closure);
  3119   // Now mark from the roots
  3120   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
  3121     verification_mark_bm(), markBitMap(), verification_mark_stack());
  3122   assert(_restart_addr == NULL, "Expected pre-condition");
  3123   verification_mark_bm()->iterate(&markFromRootsClosure);
  3124   while (_restart_addr != NULL) {
  3125     // Deal with stack overflow: by restarting at the indicated
  3126     // address.
  3127     HeapWord* ra = _restart_addr;
  3128     markFromRootsClosure.reset(ra);
  3129     _restart_addr = NULL;
  3130     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
  3132   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
  3133   verify_work_stacks_empty();
  3135   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
  3136   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
  3138   // Marking completed -- now verify that each bit marked in
  3139   // verification_mark_bm() is also marked in markBitMap(); flag all
  3140   // errors by printing corresponding objects.
  3141   VerifyMarkedClosure vcl(markBitMap());
  3142   verification_mark_bm()->iterate(&vcl);
  3143   assert(!vcl.failed(), "Else verification above should not have succeeded");
  3146 void ConcurrentMarkSweepGeneration::save_marks() {
  3147   // delegate to CMS space
  3148   cmsSpace()->save_marks();
  3149   for (uint i = 0; i < ParallelGCThreads; i++) {
  3150     _par_gc_thread_states[i]->promo.startTrackingPromotions();
  3154 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
  3155   return cmsSpace()->no_allocs_since_save_marks();
  3158 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
  3160 void ConcurrentMarkSweepGeneration::                            \
  3161 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
  3162   cl->set_generation(this);                                     \
  3163   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
  3164   cl->reset_generation();                                       \
  3165   save_marks();                                                 \
  3168 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
  3170 void
  3171 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
  3172   cl->set_generation(this);
  3173   younger_refs_in_space_iterate(_cmsSpace, cl);
  3174   cl->reset_generation();
  3177 void
  3178 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
  3179   if (freelistLock()->owned_by_self()) {
  3180     Generation::oop_iterate(mr, cl);
  3181   } else {
  3182     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3183     Generation::oop_iterate(mr, cl);
  3187 void
  3188 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
  3189   if (freelistLock()->owned_by_self()) {
  3190     Generation::oop_iterate(cl);
  3191   } else {
  3192     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3193     Generation::oop_iterate(cl);
  3197 void
  3198 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
  3199   if (freelistLock()->owned_by_self()) {
  3200     Generation::object_iterate(cl);
  3201   } else {
  3202     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3203     Generation::object_iterate(cl);
  3207 void
  3208 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
  3209   if (freelistLock()->owned_by_self()) {
  3210     Generation::safe_object_iterate(cl);
  3211   } else {
  3212     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3213     Generation::safe_object_iterate(cl);
  3217 void
  3218 ConcurrentMarkSweepGeneration::post_compact() {
  3221 void
  3222 ConcurrentMarkSweepGeneration::prepare_for_verify() {
  3223   // Fix the linear allocation blocks to look like free blocks.
  3225   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3226   // are not called when the heap is verified during universe initialization and
  3227   // at vm shutdown.
  3228   if (freelistLock()->owned_by_self()) {
  3229     cmsSpace()->prepare_for_verify();
  3230   } else {
  3231     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3232     cmsSpace()->prepare_for_verify();
  3236 void
  3237 ConcurrentMarkSweepGeneration::verify() {
  3238   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
  3239   // are not called when the heap is verified during universe initialization and
  3240   // at vm shutdown.
  3241   if (freelistLock()->owned_by_self()) {
  3242     cmsSpace()->verify();
  3243   } else {
  3244     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
  3245     cmsSpace()->verify();
  3249 void CMSCollector::verify() {
  3250   _cmsGen->verify();
  3253 #ifndef PRODUCT
  3254 bool CMSCollector::overflow_list_is_empty() const {
  3255   assert(_num_par_pushes >= 0, "Inconsistency");
  3256   if (_overflow_list == NULL) {
  3257     assert(_num_par_pushes == 0, "Inconsistency");
  3259   return _overflow_list == NULL;
  3262 // The methods verify_work_stacks_empty() and verify_overflow_empty()
  3263 // merely consolidate assertion checks that appear to occur together frequently.
  3264 void CMSCollector::verify_work_stacks_empty() const {
  3265   assert(_markStack.isEmpty(), "Marking stack should be empty");
  3266   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3269 void CMSCollector::verify_overflow_empty() const {
  3270   assert(overflow_list_is_empty(), "Overflow list should be empty");
  3271   assert(no_preserved_marks(), "No preserved marks");
  3273 #endif // PRODUCT
  3275 // Decide if we want to enable class unloading as part of the
  3276 // ensuing concurrent GC cycle. We will collect and
  3277 // unload classes if it's the case that:
  3278 // (1) an explicit gc request has been made and the flag
  3279 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
  3280 // (2) (a) class unloading is enabled at the command line, and
  3281 //     (b) old gen is getting really full
  3282 // NOTE: Provided there is no change in the state of the heap between
  3283 // calls to this method, it should have idempotent results. Moreover,
  3284 // its results should be monotonically increasing (i.e. going from 0 to 1,
  3285 // but not 1 to 0) between successive calls between which the heap was
  3286 // not collected. For the implementation below, it must thus rely on
  3287 // the property that concurrent_cycles_since_last_unload()
  3288 // will not decrease unless a collection cycle happened and that
  3289 // _cmsGen->is_too_full() are
  3290 // themselves also monotonic in that sense. See check_monotonicity()
  3291 // below.
  3292 void CMSCollector::update_should_unload_classes() {
  3293   _should_unload_classes = false;
  3294   // Condition 1 above
  3295   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
  3296     _should_unload_classes = true;
  3297   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
  3298     // Disjuncts 2.b.(i,ii,iii) above
  3299     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
  3300                               CMSClassUnloadingMaxInterval)
  3301                            || _cmsGen->is_too_full();
  3305 bool ConcurrentMarkSweepGeneration::is_too_full() const {
  3306   bool res = should_concurrent_collect();
  3307   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
  3308   return res;
  3311 void CMSCollector::setup_cms_unloading_and_verification_state() {
  3312   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
  3313                              || VerifyBeforeExit;
  3314   const  int  rso           =   SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
  3316   // We set the proper root for this CMS cycle here.
  3317   if (should_unload_classes()) {   // Should unload classes this cycle
  3318     remove_root_scanning_option(SharedHeap::SO_AllClasses);
  3319     add_root_scanning_option(SharedHeap::SO_SystemClasses);
  3320     remove_root_scanning_option(rso);  // Shrink the root set appropriately
  3321     set_verifying(should_verify);    // Set verification state for this cycle
  3322     return;                            // Nothing else needs to be done at this time
  3325   // Not unloading classes this cycle
  3326   assert(!should_unload_classes(), "Inconsitency!");
  3327   remove_root_scanning_option(SharedHeap::SO_SystemClasses);
  3328   add_root_scanning_option(SharedHeap::SO_AllClasses);
  3330   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
  3331     // Include symbols, strings and code cache elements to prevent their resurrection.
  3332     add_root_scanning_option(rso);
  3333     set_verifying(true);
  3334   } else if (verifying() && !should_verify) {
  3335     // We were verifying, but some verification flags got disabled.
  3336     set_verifying(false);
  3337     // Exclude symbols, strings and code cache elements from root scanning to
  3338     // reduce IM and RM pauses.
  3339     remove_root_scanning_option(rso);
  3344 #ifndef PRODUCT
  3345 HeapWord* CMSCollector::block_start(const void* p) const {
  3346   const HeapWord* addr = (HeapWord*)p;
  3347   if (_span.contains(p)) {
  3348     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
  3349       return _cmsGen->cmsSpace()->block_start(p);
  3352   return NULL;
  3354 #endif
  3356 HeapWord*
  3357 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
  3358                                                    bool   tlab,
  3359                                                    bool   parallel) {
  3360   CMSSynchronousYieldRequest yr;
  3361   assert(!tlab, "Can't deal with TLAB allocation");
  3362   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
  3363   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
  3364     CMSExpansionCause::_satisfy_allocation);
  3365   if (GCExpandToAllocateDelayMillis > 0) {
  3366     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3368   return have_lock_and_allocate(word_size, tlab);
  3371 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
  3372 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
  3373 // to CardGeneration and share it...
  3374 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
  3375   return CardGeneration::expand(bytes, expand_bytes);
  3378 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
  3379   CMSExpansionCause::Cause cause)
  3382   bool success = expand(bytes, expand_bytes);
  3384   // remember why we expanded; this information is used
  3385   // by shouldConcurrentCollect() when making decisions on whether to start
  3386   // a new CMS cycle.
  3387   if (success) {
  3388     set_expansion_cause(cause);
  3389     if (PrintGCDetails && Verbose) {
  3390       gclog_or_tty->print_cr("Expanded CMS gen for %s",
  3391         CMSExpansionCause::to_string(cause));
  3396 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
  3397   HeapWord* res = NULL;
  3398   MutexLocker x(ParGCRareEvent_lock);
  3399   while (true) {
  3400     // Expansion by some other thread might make alloc OK now:
  3401     res = ps->lab.alloc(word_sz);
  3402     if (res != NULL) return res;
  3403     // If there's not enough expansion space available, give up.
  3404     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
  3405       return NULL;
  3407     // Otherwise, we try expansion.
  3408     expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
  3409       CMSExpansionCause::_allocate_par_lab);
  3410     // Now go around the loop and try alloc again;
  3411     // A competing par_promote might beat us to the expansion space,
  3412     // so we may go around the loop again if promotion fails agaion.
  3413     if (GCExpandToAllocateDelayMillis > 0) {
  3414       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3420 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
  3421   PromotionInfo* promo) {
  3422   MutexLocker x(ParGCRareEvent_lock);
  3423   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
  3424   while (true) {
  3425     // Expansion by some other thread might make alloc OK now:
  3426     if (promo->ensure_spooling_space()) {
  3427       assert(promo->has_spooling_space(),
  3428              "Post-condition of successful ensure_spooling_space()");
  3429       return true;
  3431     // If there's not enough expansion space available, give up.
  3432     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
  3433       return false;
  3435     // Otherwise, we try expansion.
  3436     expand(refill_size_bytes, MinHeapDeltaBytes,
  3437       CMSExpansionCause::_allocate_par_spooling_space);
  3438     // Now go around the loop and try alloc again;
  3439     // A competing allocation might beat us to the expansion space,
  3440     // so we may go around the loop again if allocation fails again.
  3441     if (GCExpandToAllocateDelayMillis > 0) {
  3442       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
  3448 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
  3449   assert_locked_or_safepoint(ExpandHeap_lock);
  3450   // Shrink committed space
  3451   _virtual_space.shrink_by(bytes);
  3452   // Shrink space; this also shrinks the space's BOT
  3453   _cmsSpace->set_end((HeapWord*) _virtual_space.high());
  3454   size_t new_word_size = heap_word_size(_cmsSpace->capacity());
  3455   // Shrink the shared block offset array
  3456   _bts->resize(new_word_size);
  3457   MemRegion mr(_cmsSpace->bottom(), new_word_size);
  3458   // Shrink the card table
  3459   Universe::heap()->barrier_set()->resize_covered_region(mr);
  3461   if (Verbose && PrintGC) {
  3462     size_t new_mem_size = _virtual_space.committed_size();
  3463     size_t old_mem_size = new_mem_size + bytes;
  3464     gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
  3465                   name(), old_mem_size/K, new_mem_size/K);
  3469 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
  3470   assert_locked_or_safepoint(Heap_lock);
  3471   size_t size = ReservedSpace::page_align_size_down(bytes);
  3472   // Only shrink if a compaction was done so that all the free space
  3473   // in the generation is in a contiguous block at the end.
  3474   if (size > 0 && did_compact()) {
  3475     shrink_by(size);
  3479 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
  3480   assert_locked_or_safepoint(Heap_lock);
  3481   bool result = _virtual_space.expand_by(bytes);
  3482   if (result) {
  3483     size_t new_word_size =
  3484       heap_word_size(_virtual_space.committed_size());
  3485     MemRegion mr(_cmsSpace->bottom(), new_word_size);
  3486     _bts->resize(new_word_size);  // resize the block offset shared array
  3487     Universe::heap()->barrier_set()->resize_covered_region(mr);
  3488     // Hmmmm... why doesn't CFLS::set_end verify locking?
  3489     // This is quite ugly; FIX ME XXX
  3490     _cmsSpace->assert_locked(freelistLock());
  3491     _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  3493     // update the space and generation capacity counters
  3494     if (UsePerfData) {
  3495       _space_counters->update_capacity();
  3496       _gen_counters->update_all();
  3499     if (Verbose && PrintGC) {
  3500       size_t new_mem_size = _virtual_space.committed_size();
  3501       size_t old_mem_size = new_mem_size - bytes;
  3502       gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
  3503                     name(), old_mem_size/K, bytes/K, new_mem_size/K);
  3506   return result;
  3509 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
  3510   assert_locked_or_safepoint(Heap_lock);
  3511   bool success = true;
  3512   const size_t remaining_bytes = _virtual_space.uncommitted_size();
  3513   if (remaining_bytes > 0) {
  3514     success = grow_by(remaining_bytes);
  3515     DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
  3517   return success;
  3520 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
  3521   assert_locked_or_safepoint(Heap_lock);
  3522   assert_lock_strong(freelistLock());
  3523   if (PrintGCDetails && Verbose) {
  3524     warning("Shrinking of CMS not yet implemented");
  3526   return;
  3530 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
  3531 // phases.
  3532 class CMSPhaseAccounting: public StackObj {
  3533  public:
  3534   CMSPhaseAccounting(CMSCollector *collector,
  3535                      const char *phase,
  3536                      bool print_cr = true);
  3537   ~CMSPhaseAccounting();
  3539  private:
  3540   CMSCollector *_collector;
  3541   const char *_phase;
  3542   elapsedTimer _wallclock;
  3543   bool _print_cr;
  3545  public:
  3546   // Not MT-safe; so do not pass around these StackObj's
  3547   // where they may be accessed by other threads.
  3548   jlong wallclock_millis() {
  3549     assert(_wallclock.is_active(), "Wall clock should not stop");
  3550     _wallclock.stop();  // to record time
  3551     jlong ret = _wallclock.milliseconds();
  3552     _wallclock.start(); // restart
  3553     return ret;
  3555 };
  3557 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
  3558                                        const char *phase,
  3559                                        bool print_cr) :
  3560   _collector(collector), _phase(phase), _print_cr(print_cr) {
  3562   if (PrintCMSStatistics != 0) {
  3563     _collector->resetYields();
  3565   if (PrintGCDetails) {
  3566     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3567     gclog_or_tty->stamp(PrintGCTimeStamps);
  3568     gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
  3569       _collector->cmsGen()->short_name(), _phase);
  3571   _collector->resetTimer();
  3572   _wallclock.start();
  3573   _collector->startTimer();
  3576 CMSPhaseAccounting::~CMSPhaseAccounting() {
  3577   assert(_wallclock.is_active(), "Wall clock should not have stopped");
  3578   _collector->stopTimer();
  3579   _wallclock.stop();
  3580   if (PrintGCDetails) {
  3581     gclog_or_tty->date_stamp(PrintGCDateStamps);
  3582     gclog_or_tty->stamp(PrintGCTimeStamps);
  3583     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
  3584                  _collector->cmsGen()->short_name(),
  3585                  _phase, _collector->timerValue(), _wallclock.seconds());
  3586     if (_print_cr) {
  3587       gclog_or_tty->cr();
  3589     if (PrintCMSStatistics != 0) {
  3590       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
  3591                     _collector->yields());
  3596 // CMS work
  3598 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
  3599 class CMSParMarkTask : public AbstractGangTask {
  3600  protected:
  3601   CMSCollector*     _collector;
  3602   int               _n_workers;
  3603   CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
  3604       AbstractGangTask(name),
  3605       _collector(collector),
  3606       _n_workers(n_workers) {}
  3607   // Work method in support of parallel rescan ... of young gen spaces
  3608   void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
  3609                              ContiguousSpace* space,
  3610                              HeapWord** chunk_array, size_t chunk_top);
  3611   void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
  3612 };
  3614 // Parallel initial mark task
  3615 class CMSParInitialMarkTask: public CMSParMarkTask {
  3616  public:
  3617   CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
  3618       CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
  3619                      collector, n_workers) {}
  3620   void work(uint worker_id);
  3621 };
  3623 // Checkpoint the roots into this generation from outside
  3624 // this generation. [Note this initial checkpoint need only
  3625 // be approximate -- we'll do a catch up phase subsequently.]
  3626 void CMSCollector::checkpointRootsInitial(bool asynch) {
  3627   assert(_collectorState == InitialMarking, "Wrong collector state");
  3628   check_correct_thread_executing();
  3629   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
  3631   save_heap_summary();
  3632   report_heap_summary(GCWhen::BeforeGC);
  3634   ReferenceProcessor* rp = ref_processor();
  3635   SpecializationStats::clear();
  3636   assert(_restart_addr == NULL, "Control point invariant");
  3637   if (asynch) {
  3638     // acquire locks for subsequent manipulations
  3639     MutexLockerEx x(bitMapLock(),
  3640                     Mutex::_no_safepoint_check_flag);
  3641     checkpointRootsInitialWork(asynch);
  3642     // enable ("weak") refs discovery
  3643     rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
  3644     _collectorState = Marking;
  3645   } else {
  3646     // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
  3647     // which recognizes if we are a CMS generation, and doesn't try to turn on
  3648     // discovery; verify that they aren't meddling.
  3649     assert(!rp->discovery_is_atomic(),
  3650            "incorrect setting of discovery predicate");
  3651     assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
  3652            "ref discovery for this generation kind");
  3653     // already have locks
  3654     checkpointRootsInitialWork(asynch);
  3655     // now enable ("weak") refs discovery
  3656     rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
  3657     _collectorState = Marking;
  3659   SpecializationStats::print();
  3662 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
  3663   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
  3664   assert(_collectorState == InitialMarking, "just checking");
  3666   // If there has not been a GC[n-1] since last GC[n] cycle completed,
  3667   // precede our marking with a collection of all
  3668   // younger generations to keep floating garbage to a minimum.
  3669   // XXX: we won't do this for now -- it's an optimization to be done later.
  3671   // already have locks
  3672   assert_lock_strong(bitMapLock());
  3673   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
  3675   // Setup the verification and class unloading state for this
  3676   // CMS collection cycle.
  3677   setup_cms_unloading_and_verification_state();
  3679   NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
  3680     PrintGCDetails && Verbose, true, _gc_timer_cm);)
  3681   if (UseAdaptiveSizePolicy) {
  3682     size_policy()->checkpoint_roots_initial_begin();
  3685   // Reset all the PLAB chunk arrays if necessary.
  3686   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
  3687     reset_survivor_plab_arrays();
  3690   ResourceMark rm;
  3691   HandleMark  hm;
  3693   FalseClosure falseClosure;
  3694   // In the case of a synchronous collection, we will elide the
  3695   // remark step, so it's important to catch all the nmethod oops
  3696   // in this step.
  3697   // The final 'true' flag to gen_process_strong_roots will ensure this.
  3698   // If 'async' is true, we can relax the nmethod tracing.
  3699   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
  3700   GenCollectedHeap* gch = GenCollectedHeap::heap();
  3702   verify_work_stacks_empty();
  3703   verify_overflow_empty();
  3705   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
  3706   // Update the saved marks which may affect the root scans.
  3707   gch->save_marks();
  3709   // weak reference processing has not started yet.
  3710   ref_processor()->set_enqueuing_is_done(false);
  3712   // Need to remember all newly created CLDs,
  3713   // so that we can guarantee that the remark finds them.
  3714   ClassLoaderDataGraph::remember_new_clds(true);
  3716   // Whenever a CLD is found, it will be claimed before proceeding to mark
  3717   // the klasses. The claimed marks need to be cleared before marking starts.
  3718   ClassLoaderDataGraph::clear_claimed_marks();
  3720   if (CMSPrintEdenSurvivorChunks) {
  3721     print_eden_and_survivor_chunk_arrays();
  3725     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  3726     if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
  3727       // The parallel version.
  3728       FlexibleWorkGang* workers = gch->workers();
  3729       assert(workers != NULL, "Need parallel worker threads.");
  3730       int n_workers = workers->active_workers();
  3731       CMSParInitialMarkTask tsk(this, n_workers);
  3732       gch->set_par_threads(n_workers);
  3733       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
  3734       if (n_workers > 1) {
  3735         GenCollectedHeap::StrongRootsScope srs(gch);
  3736         workers->run_task(&tsk);
  3737       } else {
  3738         GenCollectedHeap::StrongRootsScope srs(gch);
  3739         tsk.work(0);
  3741       gch->set_par_threads(0);
  3742     } else {
  3743       // The serial version.
  3744       CMKlassClosure klass_closure(&notOlder);
  3745       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  3746       gch->gen_process_strong_roots(_cmsGen->level(),
  3747                                     true,   // younger gens are roots
  3748                                     true,   // activate StrongRootsScope
  3749                                     false,  // not scavenging
  3750                                     SharedHeap::ScanningOption(roots_scanning_options()),
  3751                                     &notOlder,
  3752                                     true,   // walk all of code cache if (so & SO_CodeCache)
  3753                                     NULL,
  3754                                     &klass_closure);
  3758   // Clear mod-union table; it will be dirtied in the prologue of
  3759   // CMS generation per each younger generation collection.
  3761   assert(_modUnionTable.isAllClear(),
  3762        "Was cleared in most recent final checkpoint phase"
  3763        " or no bits are set in the gc_prologue before the start of the next "
  3764        "subsequent marking phase.");
  3766   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
  3768   // Save the end of the used_region of the constituent generations
  3769   // to be used to limit the extent of sweep in each generation.
  3770   save_sweep_limits();
  3771   if (UseAdaptiveSizePolicy) {
  3772     size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
  3774   verify_overflow_empty();
  3777 bool CMSCollector::markFromRoots(bool asynch) {
  3778   // we might be tempted to assert that:
  3779   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
  3780   //        "inconsistent argument?");
  3781   // However that wouldn't be right, because it's possible that
  3782   // a safepoint is indeed in progress as a younger generation
  3783   // stop-the-world GC happens even as we mark in this generation.
  3784   assert(_collectorState == Marking, "inconsistent state?");
  3785   check_correct_thread_executing();
  3786   verify_overflow_empty();
  3788   bool res;
  3789   if (asynch) {
  3791     // Start the timers for adaptive size policy for the concurrent phases
  3792     // Do it here so that the foreground MS can use the concurrent
  3793     // timer since a foreground MS might has the sweep done concurrently
  3794     // or STW.
  3795     if (UseAdaptiveSizePolicy) {
  3796       size_policy()->concurrent_marking_begin();
  3799     // Weak ref discovery note: We may be discovering weak
  3800     // refs in this generation concurrent (but interleaved) with
  3801     // weak ref discovery by a younger generation collector.
  3803     CMSTokenSyncWithLocks ts(true, bitMapLock());
  3804     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  3805     CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
  3806     res = markFromRootsWork(asynch);
  3807     if (res) {
  3808       _collectorState = Precleaning;
  3809     } else { // We failed and a foreground collection wants to take over
  3810       assert(_foregroundGCIsActive, "internal state inconsistency");
  3811       assert(_restart_addr == NULL,  "foreground will restart from scratch");
  3812       if (PrintGCDetails) {
  3813         gclog_or_tty->print_cr("bailing out to foreground collection");
  3816     if (UseAdaptiveSizePolicy) {
  3817       size_policy()->concurrent_marking_end();
  3819   } else {
  3820     assert(SafepointSynchronize::is_at_safepoint(),
  3821            "inconsistent with asynch == false");
  3822     if (UseAdaptiveSizePolicy) {
  3823       size_policy()->ms_collection_marking_begin();
  3825     // already have locks
  3826     res = markFromRootsWork(asynch);
  3827     _collectorState = FinalMarking;
  3828     if (UseAdaptiveSizePolicy) {
  3829       GenCollectedHeap* gch = GenCollectedHeap::heap();
  3830       size_policy()->ms_collection_marking_end(gch->gc_cause());
  3833   verify_overflow_empty();
  3834   return res;
  3837 bool CMSCollector::markFromRootsWork(bool asynch) {
  3838   // iterate over marked bits in bit map, doing a full scan and mark
  3839   // from these roots using the following algorithm:
  3840   // . if oop is to the right of the current scan pointer,
  3841   //   mark corresponding bit (we'll process it later)
  3842   // . else (oop is to left of current scan pointer)
  3843   //   push oop on marking stack
  3844   // . drain the marking stack
  3846   // Note that when we do a marking step we need to hold the
  3847   // bit map lock -- recall that direct allocation (by mutators)
  3848   // and promotion (by younger generation collectors) is also
  3849   // marking the bit map. [the so-called allocate live policy.]
  3850   // Because the implementation of bit map marking is not
  3851   // robust wrt simultaneous marking of bits in the same word,
  3852   // we need to make sure that there is no such interference
  3853   // between concurrent such updates.
  3855   // already have locks
  3856   assert_lock_strong(bitMapLock());
  3858   verify_work_stacks_empty();
  3859   verify_overflow_empty();
  3860   bool result = false;
  3861   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
  3862     result = do_marking_mt(asynch);
  3863   } else {
  3864     result = do_marking_st(asynch);
  3866   return result;
  3869 // Forward decl
  3870 class CMSConcMarkingTask;
  3872 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
  3873   CMSCollector*       _collector;
  3874   CMSConcMarkingTask* _task;
  3875  public:
  3876   virtual void yield();
  3878   // "n_threads" is the number of threads to be terminated.
  3879   // "queue_set" is a set of work queues of other threads.
  3880   // "collector" is the CMS collector associated with this task terminator.
  3881   // "yield" indicates whether we need the gang as a whole to yield.
  3882   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
  3883     ParallelTaskTerminator(n_threads, queue_set),
  3884     _collector(collector) { }
  3886   void set_task(CMSConcMarkingTask* task) {
  3887     _task = task;
  3889 };
  3891 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
  3892   CMSConcMarkingTask* _task;
  3893  public:
  3894   bool should_exit_termination();
  3895   void set_task(CMSConcMarkingTask* task) {
  3896     _task = task;
  3898 };
  3900 // MT Concurrent Marking Task
  3901 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
  3902   CMSCollector* _collector;
  3903   int           _n_workers;                  // requested/desired # workers
  3904   bool          _asynch;
  3905   bool          _result;
  3906   CompactibleFreeListSpace*  _cms_space;
  3907   char          _pad_front[64];   // padding to ...
  3908   HeapWord*     _global_finger;   // ... avoid sharing cache line
  3909   char          _pad_back[64];
  3910   HeapWord*     _restart_addr;
  3912   //  Exposed here for yielding support
  3913   Mutex* const _bit_map_lock;
  3915   // The per thread work queues, available here for stealing
  3916   OopTaskQueueSet*  _task_queues;
  3918   // Termination (and yielding) support
  3919   CMSConcMarkingTerminator _term;
  3920   CMSConcMarkingTerminatorTerminator _term_term;
  3922  public:
  3923   CMSConcMarkingTask(CMSCollector* collector,
  3924                  CompactibleFreeListSpace* cms_space,
  3925                  bool asynch,
  3926                  YieldingFlexibleWorkGang* workers,
  3927                  OopTaskQueueSet* task_queues):
  3928     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
  3929     _collector(collector),
  3930     _cms_space(cms_space),
  3931     _asynch(asynch), _n_workers(0), _result(true),
  3932     _task_queues(task_queues),
  3933     _term(_n_workers, task_queues, _collector),
  3934     _bit_map_lock(collector->bitMapLock())
  3936     _requested_size = _n_workers;
  3937     _term.set_task(this);
  3938     _term_term.set_task(this);
  3939     _restart_addr = _global_finger = _cms_space->bottom();
  3943   OopTaskQueueSet* task_queues()  { return _task_queues; }
  3945   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  3947   HeapWord** global_finger_addr() { return &_global_finger; }
  3949   CMSConcMarkingTerminator* terminator() { return &_term; }
  3951   virtual void set_for_termination(int active_workers) {
  3952     terminator()->reset_for_reuse(active_workers);
  3955   void work(uint worker_id);
  3956   bool should_yield() {
  3957     return    ConcurrentMarkSweepThread::should_yield()
  3958            && !_collector->foregroundGCIsActive()
  3959            && _asynch;
  3962   virtual void coordinator_yield();  // stuff done by coordinator
  3963   bool result() { return _result; }
  3965   void reset(HeapWord* ra) {
  3966     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
  3967     _restart_addr = _global_finger = ra;
  3968     _term.reset_for_reuse();
  3971   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  3972                                            OopTaskQueue* work_q);
  3974  private:
  3975   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
  3976   void do_work_steal(int i);
  3977   void bump_global_finger(HeapWord* f);
  3978 };
  3980 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
  3981   assert(_task != NULL, "Error");
  3982   return _task->yielding();
  3983   // Note that we do not need the disjunct || _task->should_yield() above
  3984   // because we want terminating threads to yield only if the task
  3985   // is already in the midst of yielding, which happens only after at least one
  3986   // thread has yielded.
  3989 void CMSConcMarkingTerminator::yield() {
  3990   if (_task->should_yield()) {
  3991     _task->yield();
  3992   } else {
  3993     ParallelTaskTerminator::yield();
  3997 ////////////////////////////////////////////////////////////////
  3998 // Concurrent Marking Algorithm Sketch
  3999 ////////////////////////////////////////////////////////////////
  4000 // Until all tasks exhausted (both spaces):
  4001 // -- claim next available chunk
  4002 // -- bump global finger via CAS
  4003 // -- find first object that starts in this chunk
  4004 //    and start scanning bitmap from that position
  4005 // -- scan marked objects for oops
  4006 // -- CAS-mark target, and if successful:
  4007 //    . if target oop is above global finger (volatile read)
  4008 //      nothing to do
  4009 //    . if target oop is in chunk and above local finger
  4010 //        then nothing to do
  4011 //    . else push on work-queue
  4012 // -- Deal with possible overflow issues:
  4013 //    . local work-queue overflow causes stuff to be pushed on
  4014 //      global (common) overflow queue
  4015 //    . always first empty local work queue
  4016 //    . then get a batch of oops from global work queue if any
  4017 //    . then do work stealing
  4018 // -- When all tasks claimed (both spaces)
  4019 //    and local work queue empty,
  4020 //    then in a loop do:
  4021 //    . check global overflow stack; steal a batch of oops and trace
  4022 //    . try to steal from other threads oif GOS is empty
  4023 //    . if neither is available, offer termination
  4024 // -- Terminate and return result
  4025 //
  4026 void CMSConcMarkingTask::work(uint worker_id) {
  4027   elapsedTimer _timer;
  4028   ResourceMark rm;
  4029   HandleMark hm;
  4031   DEBUG_ONLY(_collector->verify_overflow_empty();)
  4033   // Before we begin work, our work queue should be empty
  4034   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
  4035   // Scan the bitmap covering _cms_space, tracing through grey objects.
  4036   _timer.start();
  4037   do_scan_and_mark(worker_id, _cms_space);
  4038   _timer.stop();
  4039   if (PrintCMSStatistics != 0) {
  4040     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
  4041       worker_id, _timer.seconds());
  4042       // XXX: need xxx/xxx type of notation, two timers
  4045   // ... do work stealing
  4046   _timer.reset();
  4047   _timer.start();
  4048   do_work_steal(worker_id);
  4049   _timer.stop();
  4050   if (PrintCMSStatistics != 0) {
  4051     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
  4052       worker_id, _timer.seconds());
  4053       // XXX: need xxx/xxx type of notation, two timers
  4055   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
  4056   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
  4057   // Note that under the current task protocol, the
  4058   // following assertion is true even of the spaces
  4059   // expanded since the completion of the concurrent
  4060   // marking. XXX This will likely change under a strict
  4061   // ABORT semantics.
  4062   // After perm removal the comparison was changed to
  4063   // greater than or equal to from strictly greater than.
  4064   // Before perm removal the highest address sweep would
  4065   // have been at the end of perm gen but now is at the
  4066   // end of the tenured gen.
  4067   assert(_global_finger >=  _cms_space->end(),
  4068          "All tasks have been completed");
  4069   DEBUG_ONLY(_collector->verify_overflow_empty();)
  4072 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
  4073   HeapWord* read = _global_finger;
  4074   HeapWord* cur  = read;
  4075   while (f > read) {
  4076     cur = read;
  4077     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
  4078     if (cur == read) {
  4079       // our cas succeeded
  4080       assert(_global_finger >= f, "protocol consistency");
  4081       break;
  4086 // This is really inefficient, and should be redone by
  4087 // using (not yet available) block-read and -write interfaces to the
  4088 // stack and the work_queue. XXX FIX ME !!!
  4089 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
  4090                                                       OopTaskQueue* work_q) {
  4091   // Fast lock-free check
  4092   if (ovflw_stk->length() == 0) {
  4093     return false;
  4095   assert(work_q->size() == 0, "Shouldn't steal");
  4096   MutexLockerEx ml(ovflw_stk->par_lock(),
  4097                    Mutex::_no_safepoint_check_flag);
  4098   // Grab up to 1/4 the size of the work queue
  4099   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  4100                     (size_t)ParGCDesiredObjsFromOverflowList);
  4101   num = MIN2(num, ovflw_stk->length());
  4102   for (int i = (int) num; i > 0; i--) {
  4103     oop cur = ovflw_stk->pop();
  4104     assert(cur != NULL, "Counted wrong?");
  4105     work_q->push(cur);
  4107   return num > 0;
  4110 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
  4111   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  4112   int n_tasks = pst->n_tasks();
  4113   // We allow that there may be no tasks to do here because
  4114   // we are restarting after a stack overflow.
  4115   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
  4116   uint nth_task = 0;
  4118   HeapWord* aligned_start = sp->bottom();
  4119   if (sp->used_region().contains(_restart_addr)) {
  4120     // Align down to a card boundary for the start of 0th task
  4121     // for this space.
  4122     aligned_start =
  4123       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
  4124                                  CardTableModRefBS::card_size);
  4127   size_t chunk_size = sp->marking_task_size();
  4128   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  4129     // Having claimed the nth task in this space,
  4130     // compute the chunk that it corresponds to:
  4131     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
  4132                                aligned_start + (nth_task+1)*chunk_size);
  4133     // Try and bump the global finger via a CAS;
  4134     // note that we need to do the global finger bump
  4135     // _before_ taking the intersection below, because
  4136     // the task corresponding to that region will be
  4137     // deemed done even if the used_region() expands
  4138     // because of allocation -- as it almost certainly will
  4139     // during start-up while the threads yield in the
  4140     // closure below.
  4141     HeapWord* finger = span.end();
  4142     bump_global_finger(finger);   // atomically
  4143     // There are null tasks here corresponding to chunks
  4144     // beyond the "top" address of the space.
  4145     span = span.intersection(sp->used_region());
  4146     if (!span.is_empty()) {  // Non-null task
  4147       HeapWord* prev_obj;
  4148       assert(!span.contains(_restart_addr) || nth_task == 0,
  4149              "Inconsistency");
  4150       if (nth_task == 0) {
  4151         // For the 0th task, we'll not need to compute a block_start.
  4152         if (span.contains(_restart_addr)) {
  4153           // In the case of a restart because of stack overflow,
  4154           // we might additionally skip a chunk prefix.
  4155           prev_obj = _restart_addr;
  4156         } else {
  4157           prev_obj = span.start();
  4159       } else {
  4160         // We want to skip the first object because
  4161         // the protocol is to scan any object in its entirety
  4162         // that _starts_ in this span; a fortiori, any
  4163         // object starting in an earlier span is scanned
  4164         // as part of an earlier claimed task.
  4165         // Below we use the "careful" version of block_start
  4166         // so we do not try to navigate uninitialized objects.
  4167         prev_obj = sp->block_start_careful(span.start());
  4168         // Below we use a variant of block_size that uses the
  4169         // Printezis bits to avoid waiting for allocated
  4170         // objects to become initialized/parsable.
  4171         while (prev_obj < span.start()) {
  4172           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
  4173           if (sz > 0) {
  4174             prev_obj += sz;
  4175           } else {
  4176             // In this case we may end up doing a bit of redundant
  4177             // scanning, but that appears unavoidable, short of
  4178             // locking the free list locks; see bug 6324141.
  4179             break;
  4183       if (prev_obj < span.end()) {
  4184         MemRegion my_span = MemRegion(prev_obj, span.end());
  4185         // Do the marking work within a non-empty span --
  4186         // the last argument to the constructor indicates whether the
  4187         // iteration should be incremental with periodic yields.
  4188         Par_MarkFromRootsClosure cl(this, _collector, my_span,
  4189                                     &_collector->_markBitMap,
  4190                                     work_queue(i),
  4191                                     &_collector->_markStack,
  4192                                     _asynch);
  4193         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
  4194       } // else nothing to do for this task
  4195     }   // else nothing to do for this task
  4197   // We'd be tempted to assert here that since there are no
  4198   // more tasks left to claim in this space, the global_finger
  4199   // must exceed space->top() and a fortiori space->end(). However,
  4200   // that would not quite be correct because the bumping of
  4201   // global_finger occurs strictly after the claiming of a task,
  4202   // so by the time we reach here the global finger may not yet
  4203   // have been bumped up by the thread that claimed the last
  4204   // task.
  4205   pst->all_tasks_completed();
  4208 class Par_ConcMarkingClosure: public CMSOopClosure {
  4209  private:
  4210   CMSCollector* _collector;
  4211   CMSConcMarkingTask* _task;
  4212   MemRegion     _span;
  4213   CMSBitMap*    _bit_map;
  4214   CMSMarkStack* _overflow_stack;
  4215   OopTaskQueue* _work_queue;
  4216  protected:
  4217   DO_OOP_WORK_DEFN
  4218  public:
  4219   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
  4220                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
  4221     CMSOopClosure(collector->ref_processor()),
  4222     _collector(collector),
  4223     _task(task),
  4224     _span(collector->_span),
  4225     _work_queue(work_queue),
  4226     _bit_map(bit_map),
  4227     _overflow_stack(overflow_stack)
  4228   { }
  4229   virtual void do_oop(oop* p);
  4230   virtual void do_oop(narrowOop* p);
  4232   void trim_queue(size_t max);
  4233   void handle_stack_overflow(HeapWord* lost);
  4234   void do_yield_check() {
  4235     if (_task->should_yield()) {
  4236       _task->yield();
  4239 };
  4241 // Grey object scanning during work stealing phase --
  4242 // the salient assumption here is that any references
  4243 // that are in these stolen objects being scanned must
  4244 // already have been initialized (else they would not have
  4245 // been published), so we do not need to check for
  4246 // uninitialized objects before pushing here.
  4247 void Par_ConcMarkingClosure::do_oop(oop obj) {
  4248   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  4249   HeapWord* addr = (HeapWord*)obj;
  4250   // Check if oop points into the CMS generation
  4251   // and is not marked
  4252   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  4253     // a white object ...
  4254     // If we manage to "claim" the object, by being the
  4255     // first thread to mark it, then we push it on our
  4256     // marking stack
  4257     if (_bit_map->par_mark(addr)) {     // ... now grey
  4258       // push on work queue (grey set)
  4259       bool simulate_overflow = false;
  4260       NOT_PRODUCT(
  4261         if (CMSMarkStackOverflowALot &&
  4262             _collector->simulate_overflow()) {
  4263           // simulate a stack overflow
  4264           simulate_overflow = true;
  4267       if (simulate_overflow ||
  4268           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  4269         // stack overflow
  4270         if (PrintCMSStatistics != 0) {
  4271           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  4272                                  SIZE_FORMAT, _overflow_stack->capacity());
  4274         // We cannot assert that the overflow stack is full because
  4275         // it may have been emptied since.
  4276         assert(simulate_overflow ||
  4277                _work_queue->size() == _work_queue->max_elems(),
  4278               "Else push should have succeeded");
  4279         handle_stack_overflow(addr);
  4281     } // Else, some other thread got there first
  4282     do_yield_check();
  4286 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
  4287 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
  4289 void Par_ConcMarkingClosure::trim_queue(size_t max) {
  4290   while (_work_queue->size() > max) {
  4291     oop new_oop;
  4292     if (_work_queue->pop_local(new_oop)) {
  4293       assert(new_oop->is_oop(), "Should be an oop");
  4294       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
  4295       assert(_span.contains((HeapWord*)new_oop), "Not in span");
  4296       new_oop->oop_iterate(this);  // do_oop() above
  4297       do_yield_check();
  4302 // Upon stack overflow, we discard (part of) the stack,
  4303 // remembering the least address amongst those discarded
  4304 // in CMSCollector's _restart_address.
  4305 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
  4306   // We need to do this under a mutex to prevent other
  4307   // workers from interfering with the work done below.
  4308   MutexLockerEx ml(_overflow_stack->par_lock(),
  4309                    Mutex::_no_safepoint_check_flag);
  4310   // Remember the least grey address discarded
  4311   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  4312   _collector->lower_restart_addr(ra);
  4313   _overflow_stack->reset();  // discard stack contents
  4314   _overflow_stack->expand(); // expand the stack if possible
  4318 void CMSConcMarkingTask::do_work_steal(int i) {
  4319   OopTaskQueue* work_q = work_queue(i);
  4320   oop obj_to_scan;
  4321   CMSBitMap* bm = &(_collector->_markBitMap);
  4322   CMSMarkStack* ovflw = &(_collector->_markStack);
  4323   int* seed = _collector->hash_seed(i);
  4324   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
  4325   while (true) {
  4326     cl.trim_queue(0);
  4327     assert(work_q->size() == 0, "Should have been emptied above");
  4328     if (get_work_from_overflow_stack(ovflw, work_q)) {
  4329       // Can't assert below because the work obtained from the
  4330       // overflow stack may already have been stolen from us.
  4331       // assert(work_q->size() > 0, "Work from overflow stack");
  4332       continue;
  4333     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  4334       assert(obj_to_scan->is_oop(), "Should be an oop");
  4335       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
  4336       obj_to_scan->oop_iterate(&cl);
  4337     } else if (terminator()->offer_termination(&_term_term)) {
  4338       assert(work_q->size() == 0, "Impossible!");
  4339       break;
  4340     } else if (yielding() || should_yield()) {
  4341       yield();
  4346 // This is run by the CMS (coordinator) thread.
  4347 void CMSConcMarkingTask::coordinator_yield() {
  4348   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4349          "CMS thread should hold CMS token");
  4350   // First give up the locks, then yield, then re-lock
  4351   // We should probably use a constructor/destructor idiom to
  4352   // do this unlock/lock or modify the MutexUnlocker class to
  4353   // serve our purpose. XXX
  4354   assert_lock_strong(_bit_map_lock);
  4355   _bit_map_lock->unlock();
  4356   ConcurrentMarkSweepThread::desynchronize(true);
  4357   ConcurrentMarkSweepThread::acknowledge_yield_request();
  4358   _collector->stopTimer();
  4359   if (PrintCMSStatistics != 0) {
  4360     _collector->incrementYields();
  4362   _collector->icms_wait();
  4364   // It is possible for whichever thread initiated the yield request
  4365   // not to get a chance to wake up and take the bitmap lock between
  4366   // this thread releasing it and reacquiring it. So, while the
  4367   // should_yield() flag is on, let's sleep for a bit to give the
  4368   // other thread a chance to wake up. The limit imposed on the number
  4369   // of iterations is defensive, to avoid any unforseen circumstances
  4370   // putting us into an infinite loop. Since it's always been this
  4371   // (coordinator_yield()) method that was observed to cause the
  4372   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
  4373   // which is by default non-zero. For the other seven methods that
  4374   // also perform the yield operation, as are using a different
  4375   // parameter (CMSYieldSleepCount) which is by default zero. This way we
  4376   // can enable the sleeping for those methods too, if necessary.
  4377   // See 6442774.
  4378   //
  4379   // We really need to reconsider the synchronization between the GC
  4380   // thread and the yield-requesting threads in the future and we
  4381   // should really use wait/notify, which is the recommended
  4382   // way of doing this type of interaction. Additionally, we should
  4383   // consolidate the eight methods that do the yield operation and they
  4384   // are almost identical into one for better maintenability and
  4385   // readability. See 6445193.
  4386   //
  4387   // Tony 2006.06.29
  4388   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
  4389                    ConcurrentMarkSweepThread::should_yield() &&
  4390                    !CMSCollector::foregroundGCIsActive(); ++i) {
  4391     os::sleep(Thread::current(), 1, false);
  4392     ConcurrentMarkSweepThread::acknowledge_yield_request();
  4395   ConcurrentMarkSweepThread::synchronize(true);
  4396   _bit_map_lock->lock_without_safepoint_check();
  4397   _collector->startTimer();
  4400 bool CMSCollector::do_marking_mt(bool asynch) {
  4401   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
  4402   int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
  4403                                        conc_workers()->total_workers(),
  4404                                        conc_workers()->active_workers(),
  4405                                        Threads::number_of_non_daemon_threads());
  4406   conc_workers()->set_active_workers(num_workers);
  4408   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  4410   CMSConcMarkingTask tsk(this,
  4411                          cms_space,
  4412                          asynch,
  4413                          conc_workers(),
  4414                          task_queues());
  4416   // Since the actual number of workers we get may be different
  4417   // from the number we requested above, do we need to do anything different
  4418   // below? In particular, may be we need to subclass the SequantialSubTasksDone
  4419   // class?? XXX
  4420   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
  4422   // Refs discovery is already non-atomic.
  4423   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
  4424   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
  4425   conc_workers()->start_task(&tsk);
  4426   while (tsk.yielded()) {
  4427     tsk.coordinator_yield();
  4428     conc_workers()->continue_task(&tsk);
  4430   // If the task was aborted, _restart_addr will be non-NULL
  4431   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
  4432   while (_restart_addr != NULL) {
  4433     // XXX For now we do not make use of ABORTED state and have not
  4434     // yet implemented the right abort semantics (even in the original
  4435     // single-threaded CMS case). That needs some more investigation
  4436     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
  4437     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
  4438     // If _restart_addr is non-NULL, a marking stack overflow
  4439     // occurred; we need to do a fresh marking iteration from the
  4440     // indicated restart address.
  4441     if (_foregroundGCIsActive && asynch) {
  4442       // We may be running into repeated stack overflows, having
  4443       // reached the limit of the stack size, while making very
  4444       // slow forward progress. It may be best to bail out and
  4445       // let the foreground collector do its job.
  4446       // Clear _restart_addr, so that foreground GC
  4447       // works from scratch. This avoids the headache of
  4448       // a "rescan" which would otherwise be needed because
  4449       // of the dirty mod union table & card table.
  4450       _restart_addr = NULL;
  4451       return false;
  4453     // Adjust the task to restart from _restart_addr
  4454     tsk.reset(_restart_addr);
  4455     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
  4456                   _restart_addr);
  4457     _restart_addr = NULL;
  4458     // Get the workers going again
  4459     conc_workers()->start_task(&tsk);
  4460     while (tsk.yielded()) {
  4461       tsk.coordinator_yield();
  4462       conc_workers()->continue_task(&tsk);
  4465   assert(tsk.completed(), "Inconsistency");
  4466   assert(tsk.result() == true, "Inconsistency");
  4467   return true;
  4470 bool CMSCollector::do_marking_st(bool asynch) {
  4471   ResourceMark rm;
  4472   HandleMark   hm;
  4474   // Temporarily make refs discovery single threaded (non-MT)
  4475   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
  4476   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
  4477     &_markStack, CMSYield && asynch);
  4478   // the last argument to iterate indicates whether the iteration
  4479   // should be incremental with periodic yields.
  4480   _markBitMap.iterate(&markFromRootsClosure);
  4481   // If _restart_addr is non-NULL, a marking stack overflow
  4482   // occurred; we need to do a fresh iteration from the
  4483   // indicated restart address.
  4484   while (_restart_addr != NULL) {
  4485     if (_foregroundGCIsActive && asynch) {
  4486       // We may be running into repeated stack overflows, having
  4487       // reached the limit of the stack size, while making very
  4488       // slow forward progress. It may be best to bail out and
  4489       // let the foreground collector do its job.
  4490       // Clear _restart_addr, so that foreground GC
  4491       // works from scratch. This avoids the headache of
  4492       // a "rescan" which would otherwise be needed because
  4493       // of the dirty mod union table & card table.
  4494       _restart_addr = NULL;
  4495       return false;  // indicating failure to complete marking
  4497     // Deal with stack overflow:
  4498     // we restart marking from _restart_addr
  4499     HeapWord* ra = _restart_addr;
  4500     markFromRootsClosure.reset(ra);
  4501     _restart_addr = NULL;
  4502     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
  4504   return true;
  4507 void CMSCollector::preclean() {
  4508   check_correct_thread_executing();
  4509   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
  4510   verify_work_stacks_empty();
  4511   verify_overflow_empty();
  4512   _abort_preclean = false;
  4513   if (CMSPrecleaningEnabled) {
  4514     if (!CMSEdenChunksRecordAlways) {
  4515       _eden_chunk_index = 0;
  4517     size_t used = get_eden_used();
  4518     size_t capacity = get_eden_capacity();
  4519     // Don't start sampling unless we will get sufficiently
  4520     // many samples.
  4521     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
  4522                 * CMSScheduleRemarkEdenPenetration)) {
  4523       _start_sampling = true;
  4524     } else {
  4525       _start_sampling = false;
  4527     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4528     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
  4529     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
  4531   CMSTokenSync x(true); // is cms thread
  4532   if (CMSPrecleaningEnabled) {
  4533     sample_eden();
  4534     _collectorState = AbortablePreclean;
  4535   } else {
  4536     _collectorState = FinalMarking;
  4538   verify_work_stacks_empty();
  4539   verify_overflow_empty();
  4542 // Try and schedule the remark such that young gen
  4543 // occupancy is CMSScheduleRemarkEdenPenetration %.
  4544 void CMSCollector::abortable_preclean() {
  4545   check_correct_thread_executing();
  4546   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
  4547   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
  4549   // If Eden's current occupancy is below this threshold,
  4550   // immediately schedule the remark; else preclean
  4551   // past the next scavenge in an effort to
  4552   // schedule the pause as described avove. By choosing
  4553   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
  4554   // we will never do an actual abortable preclean cycle.
  4555   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
  4556     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  4557     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
  4558     // We need more smarts in the abortable preclean
  4559     // loop below to deal with cases where allocation
  4560     // in young gen is very very slow, and our precleaning
  4561     // is running a losing race against a horde of
  4562     // mutators intent on flooding us with CMS updates
  4563     // (dirty cards).
  4564     // One, admittedly dumb, strategy is to give up
  4565     // after a certain number of abortable precleaning loops
  4566     // or after a certain maximum time. We want to make
  4567     // this smarter in the next iteration.
  4568     // XXX FIX ME!!! YSR
  4569     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
  4570     while (!(should_abort_preclean() ||
  4571              ConcurrentMarkSweepThread::should_terminate())) {
  4572       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
  4573       cumworkdone += workdone;
  4574       loops++;
  4575       // Voluntarily terminate abortable preclean phase if we have
  4576       // been at it for too long.
  4577       if ((CMSMaxAbortablePrecleanLoops != 0) &&
  4578           loops >= CMSMaxAbortablePrecleanLoops) {
  4579         if (PrintGCDetails) {
  4580           gclog_or_tty->print(" CMS: abort preclean due to loops ");
  4582         break;
  4584       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
  4585         if (PrintGCDetails) {
  4586           gclog_or_tty->print(" CMS: abort preclean due to time ");
  4588         break;
  4590       // If we are doing little work each iteration, we should
  4591       // take a short break.
  4592       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
  4593         // Sleep for some time, waiting for work to accumulate
  4594         stopTimer();
  4595         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
  4596         startTimer();
  4597         waited++;
  4600     if (PrintCMSStatistics > 0) {
  4601       gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
  4602                           loops, waited, cumworkdone);
  4605   CMSTokenSync x(true); // is cms thread
  4606   if (_collectorState != Idling) {
  4607     assert(_collectorState == AbortablePreclean,
  4608            "Spontaneous state transition?");
  4609     _collectorState = FinalMarking;
  4610   } // Else, a foreground collection completed this CMS cycle.
  4611   return;
  4614 // Respond to an Eden sampling opportunity
  4615 void CMSCollector::sample_eden() {
  4616   // Make sure a young gc cannot sneak in between our
  4617   // reading and recording of a sample.
  4618   assert(Thread::current()->is_ConcurrentGC_thread(),
  4619          "Only the cms thread may collect Eden samples");
  4620   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  4621          "Should collect samples while holding CMS token");
  4622   if (!_start_sampling) {
  4623     return;
  4625   // When CMSEdenChunksRecordAlways is true, the eden chunk array
  4626   // is populated by the young generation.
  4627   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
  4628     if (_eden_chunk_index < _eden_chunk_capacity) {
  4629       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
  4630       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
  4631              "Unexpected state of Eden");
  4632       // We'd like to check that what we just sampled is an oop-start address;
  4633       // however, we cannot do that here since the object may not yet have been
  4634       // initialized. So we'll instead do the check when we _use_ this sample
  4635       // later.
  4636       if (_eden_chunk_index == 0 ||
  4637           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
  4638                          _eden_chunk_array[_eden_chunk_index-1])
  4639            >= CMSSamplingGrain)) {
  4640         _eden_chunk_index++;  // commit sample
  4644   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
  4645     size_t used = get_eden_used();
  4646     size_t capacity = get_eden_capacity();
  4647     assert(used <= capacity, "Unexpected state of Eden");
  4648     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
  4649       _abort_preclean = true;
  4655 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
  4656   assert(_collectorState == Precleaning ||
  4657          _collectorState == AbortablePreclean, "incorrect state");
  4658   ResourceMark rm;
  4659   HandleMark   hm;
  4661   // Precleaning is currently not MT but the reference processor
  4662   // may be set for MT.  Disable it temporarily here.
  4663   ReferenceProcessor* rp = ref_processor();
  4664   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
  4666   // Do one pass of scrubbing the discovered reference lists
  4667   // to remove any reference objects with strongly-reachable
  4668   // referents.
  4669   if (clean_refs) {
  4670     CMSPrecleanRefsYieldClosure yield_cl(this);
  4671     assert(rp->span().equals(_span), "Spans should be equal");
  4672     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
  4673                                    &_markStack, true /* preclean */);
  4674     CMSDrainMarkingStackClosure complete_trace(this,
  4675                                    _span, &_markBitMap, &_markStack,
  4676                                    &keep_alive, true /* preclean */);
  4678     // We don't want this step to interfere with a young
  4679     // collection because we don't want to take CPU
  4680     // or memory bandwidth away from the young GC threads
  4681     // (which may be as many as there are CPUs).
  4682     // Note that we don't need to protect ourselves from
  4683     // interference with mutators because they can't
  4684     // manipulate the discovered reference lists nor affect
  4685     // the computed reachability of the referents, the
  4686     // only properties manipulated by the precleaning
  4687     // of these reference lists.
  4688     stopTimer();
  4689     CMSTokenSyncWithLocks x(true /* is cms thread */,
  4690                             bitMapLock());
  4691     startTimer();
  4692     sample_eden();
  4694     // The following will yield to allow foreground
  4695     // collection to proceed promptly. XXX YSR:
  4696     // The code in this method may need further
  4697     // tweaking for better performance and some restructuring
  4698     // for cleaner interfaces.
  4699     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
  4700     rp->preclean_discovered_references(
  4701           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
  4702           gc_timer);
  4705   if (clean_survivor) {  // preclean the active survivor space(s)
  4706     assert(_young_gen->kind() == Generation::DefNew ||
  4707            _young_gen->kind() == Generation::ParNew ||
  4708            _young_gen->kind() == Generation::ASParNew,
  4709          "incorrect type for cast");
  4710     DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  4711     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
  4712                              &_markBitMap, &_modUnionTable,
  4713                              &_markStack, true /* precleaning phase */);
  4714     stopTimer();
  4715     CMSTokenSyncWithLocks ts(true /* is cms thread */,
  4716                              bitMapLock());
  4717     startTimer();
  4718     unsigned int before_count =
  4719       GenCollectedHeap::heap()->total_collections();
  4720     SurvivorSpacePrecleanClosure
  4721       sss_cl(this, _span, &_markBitMap, &_markStack,
  4722              &pam_cl, before_count, CMSYield);
  4723     dng->from()->object_iterate_careful(&sss_cl);
  4724     dng->to()->object_iterate_careful(&sss_cl);
  4726   MarkRefsIntoAndScanClosure
  4727     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
  4728              &_markStack, this, CMSYield,
  4729              true /* precleaning phase */);
  4730   // CAUTION: The following closure has persistent state that may need to
  4731   // be reset upon a decrease in the sequence of addresses it
  4732   // processes.
  4733   ScanMarkedObjectsAgainCarefullyClosure
  4734     smoac_cl(this, _span,
  4735       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
  4737   // Preclean dirty cards in ModUnionTable and CardTable using
  4738   // appropriate convergence criterion;
  4739   // repeat CMSPrecleanIter times unless we find that
  4740   // we are losing.
  4741   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
  4742   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
  4743          "Bad convergence multiplier");
  4744   assert(CMSPrecleanThreshold >= 100,
  4745          "Unreasonably low CMSPrecleanThreshold");
  4747   size_t numIter, cumNumCards, lastNumCards, curNumCards;
  4748   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
  4749        numIter < CMSPrecleanIter;
  4750        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
  4751     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
  4752     if (Verbose && PrintGCDetails) {
  4753       gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
  4755     // Either there are very few dirty cards, so re-mark
  4756     // pause will be small anyway, or our pre-cleaning isn't
  4757     // that much faster than the rate at which cards are being
  4758     // dirtied, so we might as well stop and re-mark since
  4759     // precleaning won't improve our re-mark time by much.
  4760     if (curNumCards <= CMSPrecleanThreshold ||
  4761         (numIter > 0 &&
  4762          (curNumCards * CMSPrecleanDenominator >
  4763          lastNumCards * CMSPrecleanNumerator))) {
  4764       numIter++;
  4765       cumNumCards += curNumCards;
  4766       break;
  4770   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
  4772   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
  4773   cumNumCards += curNumCards;
  4774   if (PrintGCDetails && PrintCMSStatistics != 0) {
  4775     gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
  4776                   curNumCards, cumNumCards, numIter);
  4778   return cumNumCards;   // as a measure of useful work done
  4781 // PRECLEANING NOTES:
  4782 // Precleaning involves:
  4783 // . reading the bits of the modUnionTable and clearing the set bits.
  4784 // . For the cards corresponding to the set bits, we scan the
  4785 //   objects on those cards. This means we need the free_list_lock
  4786 //   so that we can safely iterate over the CMS space when scanning
  4787 //   for oops.
  4788 // . When we scan the objects, we'll be both reading and setting
  4789 //   marks in the marking bit map, so we'll need the marking bit map.
  4790 // . For protecting _collector_state transitions, we take the CGC_lock.
  4791 //   Note that any races in the reading of of card table entries by the
  4792 //   CMS thread on the one hand and the clearing of those entries by the
  4793 //   VM thread or the setting of those entries by the mutator threads on the
  4794 //   other are quite benign. However, for efficiency it makes sense to keep
  4795 //   the VM thread from racing with the CMS thread while the latter is
  4796 //   dirty card info to the modUnionTable. We therefore also use the
  4797 //   CGC_lock to protect the reading of the card table and the mod union
  4798 //   table by the CM thread.
  4799 // . We run concurrently with mutator updates, so scanning
  4800 //   needs to be done carefully  -- we should not try to scan
  4801 //   potentially uninitialized objects.
  4802 //
  4803 // Locking strategy: While holding the CGC_lock, we scan over and
  4804 // reset a maximal dirty range of the mod union / card tables, then lock
  4805 // the free_list_lock and bitmap lock to do a full marking, then
  4806 // release these locks; and repeat the cycle. This allows for a
  4807 // certain amount of fairness in the sharing of these locks between
  4808 // the CMS collector on the one hand, and the VM thread and the
  4809 // mutators on the other.
  4811 // NOTE: preclean_mod_union_table() and preclean_card_table()
  4812 // further below are largely identical; if you need to modify
  4813 // one of these methods, please check the other method too.
  4815 size_t CMSCollector::preclean_mod_union_table(
  4816   ConcurrentMarkSweepGeneration* gen,
  4817   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4818   verify_work_stacks_empty();
  4819   verify_overflow_empty();
  4821   // strategy: starting with the first card, accumulate contiguous
  4822   // ranges of dirty cards; clear these cards, then scan the region
  4823   // covered by these cards.
  4825   // Since all of the MUT is committed ahead, we can just use
  4826   // that, in case the generations expand while we are precleaning.
  4827   // It might also be fine to just use the committed part of the
  4828   // generation, but we might potentially miss cards when the
  4829   // generation is rapidly expanding while we are in the midst
  4830   // of precleaning.
  4831   HeapWord* startAddr = gen->reserved().start();
  4832   HeapWord* endAddr   = gen->reserved().end();
  4834   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4836   size_t numDirtyCards, cumNumDirtyCards;
  4837   HeapWord *nextAddr, *lastAddr;
  4838   for (cumNumDirtyCards = numDirtyCards = 0,
  4839        nextAddr = lastAddr = startAddr;
  4840        nextAddr < endAddr;
  4841        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4843     ResourceMark rm;
  4844     HandleMark   hm;
  4846     MemRegion dirtyRegion;
  4848       stopTimer();
  4849       // Potential yield point
  4850       CMSTokenSync ts(true);
  4851       startTimer();
  4852       sample_eden();
  4853       // Get dirty region starting at nextOffset (inclusive),
  4854       // simultaneously clearing it.
  4855       dirtyRegion =
  4856         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
  4857       assert(dirtyRegion.start() >= nextAddr,
  4858              "returned region inconsistent?");
  4860     // Remember where the next search should begin.
  4861     // The returned region (if non-empty) is a right open interval,
  4862     // so lastOffset is obtained from the right end of that
  4863     // interval.
  4864     lastAddr = dirtyRegion.end();
  4865     // Should do something more transparent and less hacky XXX
  4866     numDirtyCards =
  4867       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
  4869     // We'll scan the cards in the dirty region (with periodic
  4870     // yields for foreground GC as needed).
  4871     if (!dirtyRegion.is_empty()) {
  4872       assert(numDirtyCards > 0, "consistency check");
  4873       HeapWord* stop_point = NULL;
  4874       stopTimer();
  4875       // Potential yield point
  4876       CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
  4877                                bitMapLock());
  4878       startTimer();
  4880         verify_work_stacks_empty();
  4881         verify_overflow_empty();
  4882         sample_eden();
  4883         stop_point =
  4884           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4886       if (stop_point != NULL) {
  4887         // The careful iteration stopped early either because it found an
  4888         // uninitialized object, or because we were in the midst of an
  4889         // "abortable preclean", which should now be aborted. Redirty
  4890         // the bits corresponding to the partially-scanned or unscanned
  4891         // cards. We'll either restart at the next block boundary or
  4892         // abort the preclean.
  4893         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
  4894                "Should only be AbortablePreclean.");
  4895         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
  4896         if (should_abort_preclean()) {
  4897           break; // out of preclean loop
  4898         } else {
  4899           // Compute the next address at which preclean should pick up;
  4900           // might need bitMapLock in order to read P-bits.
  4901           lastAddr = next_card_start_after_block(stop_point);
  4904     } else {
  4905       assert(lastAddr == endAddr, "consistency check");
  4906       assert(numDirtyCards == 0, "consistency check");
  4907       break;
  4910   verify_work_stacks_empty();
  4911   verify_overflow_empty();
  4912   return cumNumDirtyCards;
  4915 // NOTE: preclean_mod_union_table() above and preclean_card_table()
  4916 // below are largely identical; if you need to modify
  4917 // one of these methods, please check the other method too.
  4919 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
  4920   ScanMarkedObjectsAgainCarefullyClosure* cl) {
  4921   // strategy: it's similar to precleamModUnionTable above, in that
  4922   // we accumulate contiguous ranges of dirty cards, mark these cards
  4923   // precleaned, then scan the region covered by these cards.
  4924   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
  4925   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
  4927   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
  4929   size_t numDirtyCards, cumNumDirtyCards;
  4930   HeapWord *lastAddr, *nextAddr;
  4932   for (cumNumDirtyCards = numDirtyCards = 0,
  4933        nextAddr = lastAddr = startAddr;
  4934        nextAddr < endAddr;
  4935        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
  4937     ResourceMark rm;
  4938     HandleMark   hm;
  4940     MemRegion dirtyRegion;
  4942       // See comments in "Precleaning notes" above on why we
  4943       // do this locking. XXX Could the locking overheads be
  4944       // too high when dirty cards are sparse? [I don't think so.]
  4945       stopTimer();
  4946       CMSTokenSync x(true); // is cms thread
  4947       startTimer();
  4948       sample_eden();
  4949       // Get and clear dirty region from card table
  4950       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
  4951                                     MemRegion(nextAddr, endAddr),
  4952                                     true,
  4953                                     CardTableModRefBS::precleaned_card_val());
  4955       assert(dirtyRegion.start() >= nextAddr,
  4956              "returned region inconsistent?");
  4958     lastAddr = dirtyRegion.end();
  4959     numDirtyCards =
  4960       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
  4962     if (!dirtyRegion.is_empty()) {
  4963       stopTimer();
  4964       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
  4965       startTimer();
  4966       sample_eden();
  4967       verify_work_stacks_empty();
  4968       verify_overflow_empty();
  4969       HeapWord* stop_point =
  4970         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
  4971       if (stop_point != NULL) {
  4972         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
  4973                "Should only be AbortablePreclean.");
  4974         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
  4975         if (should_abort_preclean()) {
  4976           break; // out of preclean loop
  4977         } else {
  4978           // Compute the next address at which preclean should pick up.
  4979           lastAddr = next_card_start_after_block(stop_point);
  4982     } else {
  4983       break;
  4986   verify_work_stacks_empty();
  4987   verify_overflow_empty();
  4988   return cumNumDirtyCards;
  4991 class PrecleanKlassClosure : public KlassClosure {
  4992   CMKlassClosure _cm_klass_closure;
  4993  public:
  4994   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
  4995   void do_klass(Klass* k) {
  4996     if (k->has_accumulated_modified_oops()) {
  4997       k->clear_accumulated_modified_oops();
  4999       _cm_klass_closure.do_klass(k);
  5002 };
  5004 // The freelist lock is needed to prevent asserts, is it really needed?
  5005 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
  5007   cl->set_freelistLock(freelistLock);
  5009   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
  5011   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
  5012   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
  5013   PrecleanKlassClosure preclean_klass_closure(cl);
  5014   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
  5016   verify_work_stacks_empty();
  5017   verify_overflow_empty();
  5020 void CMSCollector::checkpointRootsFinal(bool asynch,
  5021   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  5022   assert(_collectorState == FinalMarking, "incorrect state transition?");
  5023   check_correct_thread_executing();
  5024   // world is stopped at this checkpoint
  5025   assert(SafepointSynchronize::is_at_safepoint(),
  5026          "world should be stopped");
  5027   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
  5029   verify_work_stacks_empty();
  5030   verify_overflow_empty();
  5032   SpecializationStats::clear();
  5033   if (PrintGCDetails) {
  5034     gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
  5035                         _young_gen->used() / K,
  5036                         _young_gen->capacity() / K);
  5038   if (asynch) {
  5039     if (CMSScavengeBeforeRemark) {
  5040       GenCollectedHeap* gch = GenCollectedHeap::heap();
  5041       // Temporarily set flag to false, GCH->do_collection will
  5042       // expect it to be false and set to true
  5043       FlagSetting fl(gch->_is_gc_active, false);
  5044       NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
  5045         PrintGCDetails && Verbose, true, _gc_timer_cm);)
  5046       int level = _cmsGen->level() - 1;
  5047       if (level >= 0) {
  5048         gch->do_collection(true,        // full (i.e. force, see below)
  5049                            false,       // !clear_all_soft_refs
  5050                            0,           // size
  5051                            false,       // is_tlab
  5052                            level        // max_level
  5053                           );
  5056     FreelistLocker x(this);
  5057     MutexLockerEx y(bitMapLock(),
  5058                     Mutex::_no_safepoint_check_flag);
  5059     assert(!init_mark_was_synchronous, "but that's impossible!");
  5060     checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
  5061   } else {
  5062     // already have all the locks
  5063     checkpointRootsFinalWork(asynch, clear_all_soft_refs,
  5064                              init_mark_was_synchronous);
  5066   verify_work_stacks_empty();
  5067   verify_overflow_empty();
  5068   SpecializationStats::print();
  5071 void CMSCollector::checkpointRootsFinalWork(bool asynch,
  5072   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
  5074   NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
  5076   assert(haveFreelistLocks(), "must have free list locks");
  5077   assert_lock_strong(bitMapLock());
  5079   if (UseAdaptiveSizePolicy) {
  5080     size_policy()->checkpoint_roots_final_begin();
  5083   ResourceMark rm;
  5084   HandleMark   hm;
  5086   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5088   if (should_unload_classes()) {
  5089     CodeCache::gc_prologue();
  5091   assert(haveFreelistLocks(), "must have free list locks");
  5092   assert_lock_strong(bitMapLock());
  5094   if (!init_mark_was_synchronous) {
  5095     // We might assume that we need not fill TLAB's when
  5096     // CMSScavengeBeforeRemark is set, because we may have just done
  5097     // a scavenge which would have filled all TLAB's -- and besides
  5098     // Eden would be empty. This however may not always be the case --
  5099     // for instance although we asked for a scavenge, it may not have
  5100     // happened because of a JNI critical section. We probably need
  5101     // a policy for deciding whether we can in that case wait until
  5102     // the critical section releases and then do the remark following
  5103     // the scavenge, and skip it here. In the absence of that policy,
  5104     // or of an indication of whether the scavenge did indeed occur,
  5105     // we cannot rely on TLAB's having been filled and must do
  5106     // so here just in case a scavenge did not happen.
  5107     gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
  5108     // Update the saved marks which may affect the root scans.
  5109     gch->save_marks();
  5111     if (CMSPrintEdenSurvivorChunks) {
  5112       print_eden_and_survivor_chunk_arrays();
  5116       COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
  5118       // Note on the role of the mod union table:
  5119       // Since the marker in "markFromRoots" marks concurrently with
  5120       // mutators, it is possible for some reachable objects not to have been
  5121       // scanned. For instance, an only reference to an object A was
  5122       // placed in object B after the marker scanned B. Unless B is rescanned,
  5123       // A would be collected. Such updates to references in marked objects
  5124       // are detected via the mod union table which is the set of all cards
  5125       // dirtied since the first checkpoint in this GC cycle and prior to
  5126       // the most recent young generation GC, minus those cleaned up by the
  5127       // concurrent precleaning.
  5128       if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
  5129         GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
  5130         do_remark_parallel();
  5131       } else {
  5132         GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
  5133                     _gc_timer_cm);
  5134         do_remark_non_parallel();
  5137   } else {
  5138     assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
  5139     // The initial mark was stop-world, so there's no rescanning to
  5140     // do; go straight on to the next step below.
  5142   verify_work_stacks_empty();
  5143   verify_overflow_empty();
  5146     NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
  5147     refProcessingWork(asynch, clear_all_soft_refs);
  5149   verify_work_stacks_empty();
  5150   verify_overflow_empty();
  5152   if (should_unload_classes()) {
  5153     CodeCache::gc_epilogue();
  5155   JvmtiExport::gc_epilogue();
  5157   // If we encountered any (marking stack / work queue) overflow
  5158   // events during the current CMS cycle, take appropriate
  5159   // remedial measures, where possible, so as to try and avoid
  5160   // recurrence of that condition.
  5161   assert(_markStack.isEmpty(), "No grey objects");
  5162   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
  5163                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
  5164   if (ser_ovflw > 0) {
  5165     if (PrintCMSStatistics != 0) {
  5166       gclog_or_tty->print_cr("Marking stack overflow (benign) "
  5167         "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
  5168         ", kac_preclean="SIZE_FORMAT")",
  5169         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
  5170         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
  5172     _markStack.expand();
  5173     _ser_pmc_remark_ovflw = 0;
  5174     _ser_pmc_preclean_ovflw = 0;
  5175     _ser_kac_preclean_ovflw = 0;
  5176     _ser_kac_ovflw = 0;
  5178   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
  5179     if (PrintCMSStatistics != 0) {
  5180       gclog_or_tty->print_cr("Work queue overflow (benign) "
  5181         "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
  5182         _par_pmc_remark_ovflw, _par_kac_ovflw);
  5184     _par_pmc_remark_ovflw = 0;
  5185     _par_kac_ovflw = 0;
  5187   if (PrintCMSStatistics != 0) {
  5188      if (_markStack._hit_limit > 0) {
  5189        gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
  5190                               _markStack._hit_limit);
  5192      if (_markStack._failed_double > 0) {
  5193        gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
  5194                               " current capacity "SIZE_FORMAT,
  5195                               _markStack._failed_double,
  5196                               _markStack.capacity());
  5199   _markStack._hit_limit = 0;
  5200   _markStack._failed_double = 0;
  5202   if ((VerifyAfterGC || VerifyDuringGC) &&
  5203       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  5204     verify_after_remark();
  5207   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
  5209   // Change under the freelistLocks.
  5210   _collectorState = Sweeping;
  5211   // Call isAllClear() under bitMapLock
  5212   assert(_modUnionTable.isAllClear(),
  5213       "Should be clear by end of the final marking");
  5214   assert(_ct->klass_rem_set()->mod_union_is_clear(),
  5215       "Should be clear by end of the final marking");
  5216   if (UseAdaptiveSizePolicy) {
  5217     size_policy()->checkpoint_roots_final_end(gch->gc_cause());
  5221 void CMSParInitialMarkTask::work(uint worker_id) {
  5222   elapsedTimer _timer;
  5223   ResourceMark rm;
  5224   HandleMark   hm;
  5226   // ---------- scan from roots --------------
  5227   _timer.start();
  5228   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5229   Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
  5230   CMKlassClosure klass_closure(&par_mri_cl);
  5232   // ---------- young gen roots --------------
  5234     work_on_young_gen_roots(worker_id, &par_mri_cl);
  5235     _timer.stop();
  5236     if (PrintCMSStatistics != 0) {
  5237       gclog_or_tty->print_cr(
  5238         "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
  5239         worker_id, _timer.seconds());
  5243   // ---------- remaining roots --------------
  5244   _timer.reset();
  5245   _timer.start();
  5246   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
  5247                                 false,     // yg was scanned above
  5248                                 false,     // this is parallel code
  5249                                 false,     // not scavenging
  5250                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
  5251                                 &par_mri_cl,
  5252                                 true,   // walk all of code cache if (so & SO_CodeCache)
  5253                                 NULL,
  5254                                 &klass_closure);
  5255   assert(_collector->should_unload_classes()
  5256          || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
  5257          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
  5258   _timer.stop();
  5259   if (PrintCMSStatistics != 0) {
  5260     gclog_or_tty->print_cr(
  5261       "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
  5262       worker_id, _timer.seconds());
  5266 // Parallel remark task
  5267 class CMSParRemarkTask: public CMSParMarkTask {
  5268   CompactibleFreeListSpace* _cms_space;
  5270   // The per-thread work queues, available here for stealing.
  5271   OopTaskQueueSet*       _task_queues;
  5272   ParallelTaskTerminator _term;
  5274  public:
  5275   // A value of 0 passed to n_workers will cause the number of
  5276   // workers to be taken from the active workers in the work gang.
  5277   CMSParRemarkTask(CMSCollector* collector,
  5278                    CompactibleFreeListSpace* cms_space,
  5279                    int n_workers, FlexibleWorkGang* workers,
  5280                    OopTaskQueueSet* task_queues):
  5281     CMSParMarkTask("Rescan roots and grey objects in parallel",
  5282                    collector, n_workers),
  5283     _cms_space(cms_space),
  5284     _task_queues(task_queues),
  5285     _term(n_workers, task_queues) { }
  5287   OopTaskQueueSet* task_queues() { return _task_queues; }
  5289   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  5291   ParallelTaskTerminator* terminator() { return &_term; }
  5292   int n_workers() { return _n_workers; }
  5294   void work(uint worker_id);
  5296  private:
  5297   // ... of  dirty cards in old space
  5298   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
  5299                                   Par_MarkRefsIntoAndScanClosure* cl);
  5301   // ... work stealing for the above
  5302   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
  5303 };
  5305 class RemarkKlassClosure : public KlassClosure {
  5306   CMKlassClosure _cm_klass_closure;
  5307  public:
  5308   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
  5309   void do_klass(Klass* k) {
  5310     // Check if we have modified any oops in the Klass during the concurrent marking.
  5311     if (k->has_accumulated_modified_oops()) {
  5312       k->clear_accumulated_modified_oops();
  5314       // We could have transfered the current modified marks to the accumulated marks,
  5315       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
  5316     } else if (k->has_modified_oops()) {
  5317       // Don't clear anything, this info is needed by the next young collection.
  5318     } else {
  5319       // No modified oops in the Klass.
  5320       return;
  5323     // The klass has modified fields, need to scan the klass.
  5324     _cm_klass_closure.do_klass(k);
  5326 };
  5328 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
  5329   DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
  5330   EdenSpace* eden_space = dng->eden();
  5331   ContiguousSpace* from_space = dng->from();
  5332   ContiguousSpace* to_space   = dng->to();
  5334   HeapWord** eca = _collector->_eden_chunk_array;
  5335   size_t     ect = _collector->_eden_chunk_index;
  5336   HeapWord** sca = _collector->_survivor_chunk_array;
  5337   size_t     sct = _collector->_survivor_chunk_index;
  5339   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
  5340   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
  5342   do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
  5343   do_young_space_rescan(worker_id, cl, from_space, sca, sct);
  5344   do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
  5347 // work_queue(i) is passed to the closure
  5348 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
  5349 // also is passed to do_dirty_card_rescan_tasks() and to
  5350 // do_work_steal() to select the i-th task_queue.
  5352 void CMSParRemarkTask::work(uint worker_id) {
  5353   elapsedTimer _timer;
  5354   ResourceMark rm;
  5355   HandleMark   hm;
  5357   // ---------- rescan from roots --------------
  5358   _timer.start();
  5359   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5360   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
  5361     _collector->_span, _collector->ref_processor(),
  5362     &(_collector->_markBitMap),
  5363     work_queue(worker_id));
  5365   // Rescan young gen roots first since these are likely
  5366   // coarsely partitioned and may, on that account, constitute
  5367   // the critical path; thus, it's best to start off that
  5368   // work first.
  5369   // ---------- young gen roots --------------
  5371     work_on_young_gen_roots(worker_id, &par_mrias_cl);
  5372     _timer.stop();
  5373     if (PrintCMSStatistics != 0) {
  5374       gclog_or_tty->print_cr(
  5375         "Finished young gen rescan work in %dth thread: %3.3f sec",
  5376         worker_id, _timer.seconds());
  5380   // ---------- remaining roots --------------
  5381   _timer.reset();
  5382   _timer.start();
  5383   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
  5384                                 false,     // yg was scanned above
  5385                                 false,     // this is parallel code
  5386                                 false,     // not scavenging
  5387                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
  5388                                 &par_mrias_cl,
  5389                                 true,   // walk all of code cache if (so & SO_CodeCache)
  5390                                 NULL,
  5391                                 NULL);     // The dirty klasses will be handled below
  5392   assert(_collector->should_unload_classes()
  5393          || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
  5394          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
  5395   _timer.stop();
  5396   if (PrintCMSStatistics != 0) {
  5397     gclog_or_tty->print_cr(
  5398       "Finished remaining root rescan work in %dth thread: %3.3f sec",
  5399       worker_id, _timer.seconds());
  5402   // ---------- unhandled CLD scanning ----------
  5403   if (worker_id == 0) { // Single threaded at the moment.
  5404     _timer.reset();
  5405     _timer.start();
  5407     // Scan all new class loader data objects and new dependencies that were
  5408     // introduced during concurrent marking.
  5409     ResourceMark rm;
  5410     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
  5411     for (int i = 0; i < array->length(); i++) {
  5412       par_mrias_cl.do_class_loader_data(array->at(i));
  5415     // We don't need to keep track of new CLDs anymore.
  5416     ClassLoaderDataGraph::remember_new_clds(false);
  5418     _timer.stop();
  5419     if (PrintCMSStatistics != 0) {
  5420       gclog_or_tty->print_cr(
  5421           "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
  5422           worker_id, _timer.seconds());
  5426   // ---------- dirty klass scanning ----------
  5427   if (worker_id == 0) { // Single threaded at the moment.
  5428     _timer.reset();
  5429     _timer.start();
  5431     // Scan all classes that was dirtied during the concurrent marking phase.
  5432     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
  5433     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
  5435     _timer.stop();
  5436     if (PrintCMSStatistics != 0) {
  5437       gclog_or_tty->print_cr(
  5438           "Finished dirty klass scanning work in %dth thread: %3.3f sec",
  5439           worker_id, _timer.seconds());
  5443   // We might have added oops to ClassLoaderData::_handles during the
  5444   // concurrent marking phase. These oops point to newly allocated objects
  5445   // that are guaranteed to be kept alive. Either by the direct allocation
  5446   // code, or when the young collector processes the strong roots. Hence,
  5447   // we don't have to revisit the _handles block during the remark phase.
  5449   // ---------- rescan dirty cards ------------
  5450   _timer.reset();
  5451   _timer.start();
  5453   // Do the rescan tasks for each of the two spaces
  5454   // (cms_space) in turn.
  5455   // "worker_id" is passed to select the task_queue for "worker_id"
  5456   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
  5457   _timer.stop();
  5458   if (PrintCMSStatistics != 0) {
  5459     gclog_or_tty->print_cr(
  5460       "Finished dirty card rescan work in %dth thread: %3.3f sec",
  5461       worker_id, _timer.seconds());
  5464   // ---------- steal work from other threads ...
  5465   // ---------- ... and drain overflow list.
  5466   _timer.reset();
  5467   _timer.start();
  5468   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
  5469   _timer.stop();
  5470   if (PrintCMSStatistics != 0) {
  5471     gclog_or_tty->print_cr(
  5472       "Finished work stealing in %dth thread: %3.3f sec",
  5473       worker_id, _timer.seconds());
  5477 // Note that parameter "i" is not used.
  5478 void
  5479 CMSParMarkTask::do_young_space_rescan(uint worker_id,
  5480   OopsInGenClosure* cl, ContiguousSpace* space,
  5481   HeapWord** chunk_array, size_t chunk_top) {
  5482   // Until all tasks completed:
  5483   // . claim an unclaimed task
  5484   // . compute region boundaries corresponding to task claimed
  5485   //   using chunk_array
  5486   // . par_oop_iterate(cl) over that region
  5488   ResourceMark rm;
  5489   HandleMark   hm;
  5491   SequentialSubTasksDone* pst = space->par_seq_tasks();
  5493   uint nth_task = 0;
  5494   uint n_tasks  = pst->n_tasks();
  5496   if (n_tasks > 0) {
  5497     assert(pst->valid(), "Uninitialized use?");
  5498     HeapWord *start, *end;
  5499     while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5500       // We claimed task # nth_task; compute its boundaries.
  5501       if (chunk_top == 0) {  // no samples were taken
  5502         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
  5503         start = space->bottom();
  5504         end   = space->top();
  5505       } else if (nth_task == 0) {
  5506         start = space->bottom();
  5507         end   = chunk_array[nth_task];
  5508       } else if (nth_task < (uint)chunk_top) {
  5509         assert(nth_task >= 1, "Control point invariant");
  5510         start = chunk_array[nth_task - 1];
  5511         end   = chunk_array[nth_task];
  5512       } else {
  5513         assert(nth_task == (uint)chunk_top, "Control point invariant");
  5514         start = chunk_array[chunk_top - 1];
  5515         end   = space->top();
  5517       MemRegion mr(start, end);
  5518       // Verify that mr is in space
  5519       assert(mr.is_empty() || space->used_region().contains(mr),
  5520              "Should be in space");
  5521       // Verify that "start" is an object boundary
  5522       assert(mr.is_empty() || oop(mr.start())->is_oop(),
  5523              "Should be an oop");
  5524       space->par_oop_iterate(mr, cl);
  5526     pst->all_tasks_completed();
  5530 void
  5531 CMSParRemarkTask::do_dirty_card_rescan_tasks(
  5532   CompactibleFreeListSpace* sp, int i,
  5533   Par_MarkRefsIntoAndScanClosure* cl) {
  5534   // Until all tasks completed:
  5535   // . claim an unclaimed task
  5536   // . compute region boundaries corresponding to task claimed
  5537   // . transfer dirty bits ct->mut for that region
  5538   // . apply rescanclosure to dirty mut bits for that region
  5540   ResourceMark rm;
  5541   HandleMark   hm;
  5543   OopTaskQueue* work_q = work_queue(i);
  5544   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
  5545   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
  5546   // CAUTION: This closure has state that persists across calls to
  5547   // the work method dirty_range_iterate_clear() in that it has
  5548   // imbedded in it a (subtype of) UpwardsObjectClosure. The
  5549   // use of that state in the imbedded UpwardsObjectClosure instance
  5550   // assumes that the cards are always iterated (even if in parallel
  5551   // by several threads) in monotonically increasing order per each
  5552   // thread. This is true of the implementation below which picks
  5553   // card ranges (chunks) in monotonically increasing order globally
  5554   // and, a-fortiori, in monotonically increasing order per thread
  5555   // (the latter order being a subsequence of the former).
  5556   // If the work code below is ever reorganized into a more chaotic
  5557   // work-partitioning form than the current "sequential tasks"
  5558   // paradigm, the use of that persistent state will have to be
  5559   // revisited and modified appropriately. See also related
  5560   // bug 4756801 work on which should examine this code to make
  5561   // sure that the changes there do not run counter to the
  5562   // assumptions made here and necessary for correctness and
  5563   // efficiency. Note also that this code might yield inefficient
  5564   // behaviour in the case of very large objects that span one or
  5565   // more work chunks. Such objects would potentially be scanned
  5566   // several times redundantly. Work on 4756801 should try and
  5567   // address that performance anomaly if at all possible. XXX
  5568   MemRegion  full_span  = _collector->_span;
  5569   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
  5570   MarkFromDirtyCardsClosure
  5571     greyRescanClosure(_collector, full_span, // entire span of interest
  5572                       sp, bm, work_q, cl);
  5574   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
  5575   assert(pst->valid(), "Uninitialized use?");
  5576   uint nth_task = 0;
  5577   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
  5578   MemRegion span = sp->used_region();
  5579   HeapWord* start_addr = span.start();
  5580   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
  5581                                            alignment);
  5582   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
  5583   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
  5584          start_addr, "Check alignment");
  5585   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
  5586          chunk_size, "Check alignment");
  5588   while (!pst->is_task_claimed(/* reference */ nth_task)) {
  5589     // Having claimed the nth_task, compute corresponding mem-region,
  5590     // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
  5591     // The alignment restriction ensures that we do not need any
  5592     // synchronization with other gang-workers while setting or
  5593     // clearing bits in thus chunk of the MUT.
  5594     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
  5595                                     start_addr + (nth_task+1)*chunk_size);
  5596     // The last chunk's end might be way beyond end of the
  5597     // used region. In that case pull back appropriately.
  5598     if (this_span.end() > end_addr) {
  5599       this_span.set_end(end_addr);
  5600       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
  5602     // Iterate over the dirty cards covering this chunk, marking them
  5603     // precleaned, and setting the corresponding bits in the mod union
  5604     // table. Since we have been careful to partition at Card and MUT-word
  5605     // boundaries no synchronization is needed between parallel threads.
  5606     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
  5607                                                  &modUnionClosure);
  5609     // Having transferred these marks into the modUnionTable,
  5610     // rescan the marked objects on the dirty cards in the modUnionTable.
  5611     // Even if this is at a synchronous collection, the initial marking
  5612     // may have been done during an asynchronous collection so there
  5613     // may be dirty bits in the mod-union table.
  5614     _collector->_modUnionTable.dirty_range_iterate_clear(
  5615                   this_span, &greyRescanClosure);
  5616     _collector->_modUnionTable.verifyNoOneBitsInRange(
  5617                                  this_span.start(),
  5618                                  this_span.end());
  5620   pst->all_tasks_completed();  // declare that i am done
  5623 // . see if we can share work_queues with ParNew? XXX
  5624 void
  5625 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
  5626                                 int* seed) {
  5627   OopTaskQueue* work_q = work_queue(i);
  5628   NOT_PRODUCT(int num_steals = 0;)
  5629   oop obj_to_scan;
  5630   CMSBitMap* bm = &(_collector->_markBitMap);
  5632   while (true) {
  5633     // Completely finish any left over work from (an) earlier round(s)
  5634     cl->trim_queue(0);
  5635     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  5636                                          (size_t)ParGCDesiredObjsFromOverflowList);
  5637     // Now check if there's any work in the overflow list
  5638     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
  5639     // only affects the number of attempts made to get work from the
  5640     // overflow list and does not affect the number of workers.  Just
  5641     // pass ParallelGCThreads so this behavior is unchanged.
  5642     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  5643                                                 work_q,
  5644                                                 ParallelGCThreads)) {
  5645       // found something in global overflow list;
  5646       // not yet ready to go stealing work from others.
  5647       // We'd like to assert(work_q->size() != 0, ...)
  5648       // because we just took work from the overflow list,
  5649       // but of course we can't since all of that could have
  5650       // been already stolen from us.
  5651       // "He giveth and He taketh away."
  5652       continue;
  5654     // Verify that we have no work before we resort to stealing
  5655     assert(work_q->size() == 0, "Have work, shouldn't steal");
  5656     // Try to steal from other queues that have work
  5657     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  5658       NOT_PRODUCT(num_steals++;)
  5659       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  5660       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  5661       // Do scanning work
  5662       obj_to_scan->oop_iterate(cl);
  5663       // Loop around, finish this work, and try to steal some more
  5664     } else if (terminator()->offer_termination()) {
  5665         break;  // nirvana from the infinite cycle
  5668   NOT_PRODUCT(
  5669     if (PrintCMSStatistics != 0) {
  5670       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  5673   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
  5674          "Else our work is not yet done");
  5677 // Record object boundaries in _eden_chunk_array by sampling the eden
  5678 // top in the slow-path eden object allocation code path and record
  5679 // the boundaries, if CMSEdenChunksRecordAlways is true. If
  5680 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
  5681 // sampling in sample_eden() that activates during the part of the
  5682 // preclean phase.
  5683 void CMSCollector::sample_eden_chunk() {
  5684   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
  5685     if (_eden_chunk_lock->try_lock()) {
  5686       // Record a sample. This is the critical section. The contents
  5687       // of the _eden_chunk_array have to be non-decreasing in the
  5688       // address order.
  5689       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
  5690       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
  5691              "Unexpected state of Eden");
  5692       if (_eden_chunk_index == 0 ||
  5693           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
  5694            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
  5695                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
  5696         _eden_chunk_index++;  // commit sample
  5698       _eden_chunk_lock->unlock();
  5703 // Return a thread-local PLAB recording array, as appropriate.
  5704 void* CMSCollector::get_data_recorder(int thr_num) {
  5705   if (_survivor_plab_array != NULL &&
  5706       (CMSPLABRecordAlways ||
  5707        (_collectorState > Marking && _collectorState < FinalMarking))) {
  5708     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
  5709     ChunkArray* ca = &_survivor_plab_array[thr_num];
  5710     ca->reset();   // clear it so that fresh data is recorded
  5711     return (void*) ca;
  5712   } else {
  5713     return NULL;
  5717 // Reset all the thread-local PLAB recording arrays
  5718 void CMSCollector::reset_survivor_plab_arrays() {
  5719   for (uint i = 0; i < ParallelGCThreads; i++) {
  5720     _survivor_plab_array[i].reset();
  5724 // Merge the per-thread plab arrays into the global survivor chunk
  5725 // array which will provide the partitioning of the survivor space
  5726 // for CMS initial scan and rescan.
  5727 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
  5728                                               int no_of_gc_threads) {
  5729   assert(_survivor_plab_array  != NULL, "Error");
  5730   assert(_survivor_chunk_array != NULL, "Error");
  5731   assert(_collectorState == FinalMarking ||
  5732          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
  5733   for (int j = 0; j < no_of_gc_threads; j++) {
  5734     _cursor[j] = 0;
  5736   HeapWord* top = surv->top();
  5737   size_t i;
  5738   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
  5739     HeapWord* min_val = top;          // Higher than any PLAB address
  5740     uint      min_tid = 0;            // position of min_val this round
  5741     for (int j = 0; j < no_of_gc_threads; j++) {
  5742       ChunkArray* cur_sca = &_survivor_plab_array[j];
  5743       if (_cursor[j] == cur_sca->end()) {
  5744         continue;
  5746       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
  5747       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
  5748       assert(surv->used_region().contains(cur_val), "Out of bounds value");
  5749       if (cur_val < min_val) {
  5750         min_tid = j;
  5751         min_val = cur_val;
  5752       } else {
  5753         assert(cur_val < top, "All recorded addresses should be less");
  5756     // At this point min_val and min_tid are respectively
  5757     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
  5758     // and the thread (j) that witnesses that address.
  5759     // We record this address in the _survivor_chunk_array[i]
  5760     // and increment _cursor[min_tid] prior to the next round i.
  5761     if (min_val == top) {
  5762       break;
  5764     _survivor_chunk_array[i] = min_val;
  5765     _cursor[min_tid]++;
  5767   // We are all done; record the size of the _survivor_chunk_array
  5768   _survivor_chunk_index = i; // exclusive: [0, i)
  5769   if (PrintCMSStatistics > 0) {
  5770     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
  5772   // Verify that we used up all the recorded entries
  5773   #ifdef ASSERT
  5774     size_t total = 0;
  5775     for (int j = 0; j < no_of_gc_threads; j++) {
  5776       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
  5777       total += _cursor[j];
  5779     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
  5780     // Check that the merged array is in sorted order
  5781     if (total > 0) {
  5782       for (size_t i = 0; i < total - 1; i++) {
  5783         if (PrintCMSStatistics > 0) {
  5784           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
  5785                               i, _survivor_chunk_array[i]);
  5787         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
  5788                "Not sorted");
  5791   #endif // ASSERT
  5794 // Set up the space's par_seq_tasks structure for work claiming
  5795 // for parallel initial scan and rescan of young gen.
  5796 // See ParRescanTask where this is currently used.
  5797 void
  5798 CMSCollector::
  5799 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
  5800   assert(n_threads > 0, "Unexpected n_threads argument");
  5801   DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
  5803   // Eden space
  5804   if (!dng->eden()->is_empty()) {
  5805     SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
  5806     assert(!pst->valid(), "Clobbering existing data?");
  5807     // Each valid entry in [0, _eden_chunk_index) represents a task.
  5808     size_t n_tasks = _eden_chunk_index + 1;
  5809     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
  5810     // Sets the condition for completion of the subtask (how many threads
  5811     // need to finish in order to be done).
  5812     pst->set_n_threads(n_threads);
  5813     pst->set_n_tasks((int)n_tasks);
  5816   // Merge the survivor plab arrays into _survivor_chunk_array
  5817   if (_survivor_plab_array != NULL) {
  5818     merge_survivor_plab_arrays(dng->from(), n_threads);
  5819   } else {
  5820     assert(_survivor_chunk_index == 0, "Error");
  5823   // To space
  5825     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
  5826     assert(!pst->valid(), "Clobbering existing data?");
  5827     // Sets the condition for completion of the subtask (how many threads
  5828     // need to finish in order to be done).
  5829     pst->set_n_threads(n_threads);
  5830     pst->set_n_tasks(1);
  5831     assert(pst->valid(), "Error");
  5834   // From space
  5836     SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
  5837     assert(!pst->valid(), "Clobbering existing data?");
  5838     size_t n_tasks = _survivor_chunk_index + 1;
  5839     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
  5840     // Sets the condition for completion of the subtask (how many threads
  5841     // need to finish in order to be done).
  5842     pst->set_n_threads(n_threads);
  5843     pst->set_n_tasks((int)n_tasks);
  5844     assert(pst->valid(), "Error");
  5848 // Parallel version of remark
  5849 void CMSCollector::do_remark_parallel() {
  5850   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5851   FlexibleWorkGang* workers = gch->workers();
  5852   assert(workers != NULL, "Need parallel worker threads.");
  5853   // Choose to use the number of GC workers most recently set
  5854   // into "active_workers".  If active_workers is not set, set it
  5855   // to ParallelGCThreads.
  5856   int n_workers = workers->active_workers();
  5857   if (n_workers == 0) {
  5858     assert(n_workers > 0, "Should have been set during scavenge");
  5859     n_workers = ParallelGCThreads;
  5860     workers->set_active_workers(n_workers);
  5862   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
  5864   CMSParRemarkTask tsk(this,
  5865     cms_space,
  5866     n_workers, workers, task_queues());
  5868   // Set up for parallel process_strong_roots work.
  5869   gch->set_par_threads(n_workers);
  5870   // We won't be iterating over the cards in the card table updating
  5871   // the younger_gen cards, so we shouldn't call the following else
  5872   // the verification code as well as subsequent younger_refs_iterate
  5873   // code would get confused. XXX
  5874   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
  5876   // The young gen rescan work will not be done as part of
  5877   // process_strong_roots (which currently doesn't knw how to
  5878   // parallelize such a scan), but rather will be broken up into
  5879   // a set of parallel tasks (via the sampling that the [abortable]
  5880   // preclean phase did of EdenSpace, plus the [two] tasks of
  5881   // scanning the [two] survivor spaces. Further fine-grain
  5882   // parallelization of the scanning of the survivor spaces
  5883   // themselves, and of precleaning of the younger gen itself
  5884   // is deferred to the future.
  5885   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
  5887   // The dirty card rescan work is broken up into a "sequence"
  5888   // of parallel tasks (per constituent space) that are dynamically
  5889   // claimed by the parallel threads.
  5890   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
  5892   // It turns out that even when we're using 1 thread, doing the work in a
  5893   // separate thread causes wide variance in run times.  We can't help this
  5894   // in the multi-threaded case, but we special-case n=1 here to get
  5895   // repeatable measurements of the 1-thread overhead of the parallel code.
  5896   if (n_workers > 1) {
  5897     // Make refs discovery MT-safe, if it isn't already: it may not
  5898     // necessarily be so, since it's possible that we are doing
  5899     // ST marking.
  5900     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
  5901     GenCollectedHeap::StrongRootsScope srs(gch);
  5902     workers->run_task(&tsk);
  5903   } else {
  5904     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
  5905     GenCollectedHeap::StrongRootsScope srs(gch);
  5906     tsk.work(0);
  5909   gch->set_par_threads(0);  // 0 ==> non-parallel.
  5910   // restore, single-threaded for now, any preserved marks
  5911   // as a result of work_q overflow
  5912   restore_preserved_marks_if_any();
  5915 // Non-parallel version of remark
  5916 void CMSCollector::do_remark_non_parallel() {
  5917   ResourceMark rm;
  5918   HandleMark   hm;
  5919   GenCollectedHeap* gch = GenCollectedHeap::heap();
  5920   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
  5922   MarkRefsIntoAndScanClosure
  5923     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
  5924              &_markStack, this,
  5925              false /* should_yield */, false /* not precleaning */);
  5926   MarkFromDirtyCardsClosure
  5927     markFromDirtyCardsClosure(this, _span,
  5928                               NULL,  // space is set further below
  5929                               &_markBitMap, &_markStack, &mrias_cl);
  5931     GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
  5932     // Iterate over the dirty cards, setting the corresponding bits in the
  5933     // mod union table.
  5935       ModUnionClosure modUnionClosure(&_modUnionTable);
  5936       _ct->ct_bs()->dirty_card_iterate(
  5937                       _cmsGen->used_region(),
  5938                       &modUnionClosure);
  5940     // Having transferred these marks into the modUnionTable, we just need
  5941     // to rescan the marked objects on the dirty cards in the modUnionTable.
  5942     // The initial marking may have been done during an asynchronous
  5943     // collection so there may be dirty bits in the mod-union table.
  5944     const int alignment =
  5945       CardTableModRefBS::card_size * BitsPerWord;
  5947       // ... First handle dirty cards in CMS gen
  5948       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
  5949       MemRegion ur = _cmsGen->used_region();
  5950       HeapWord* lb = ur.start();
  5951       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
  5952       MemRegion cms_span(lb, ub);
  5953       _modUnionTable.dirty_range_iterate_clear(cms_span,
  5954                                                &markFromDirtyCardsClosure);
  5955       verify_work_stacks_empty();
  5956       if (PrintCMSStatistics != 0) {
  5957         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
  5958           markFromDirtyCardsClosure.num_dirty_cards());
  5962   if (VerifyDuringGC &&
  5963       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
  5964     HandleMark hm;  // Discard invalid handles created during verification
  5965     Universe::verify();
  5968     GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
  5970     verify_work_stacks_empty();
  5972     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
  5973     GenCollectedHeap::StrongRootsScope srs(gch);
  5974     gch->gen_process_strong_roots(_cmsGen->level(),
  5975                                   true,  // younger gens as roots
  5976                                   false, // use the local StrongRootsScope
  5977                                   false, // not scavenging
  5978                                   SharedHeap::ScanningOption(roots_scanning_options()),
  5979                                   &mrias_cl,
  5980                                   true,   // walk code active on stacks
  5981                                   NULL,
  5982                                   NULL);  // The dirty klasses will be handled below
  5984     assert(should_unload_classes()
  5985            || (roots_scanning_options() & SharedHeap::SO_CodeCache),
  5986            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
  5990     GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
  5992     verify_work_stacks_empty();
  5994     // Scan all class loader data objects that might have been introduced
  5995     // during concurrent marking.
  5996     ResourceMark rm;
  5997     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
  5998     for (int i = 0; i < array->length(); i++) {
  5999       mrias_cl.do_class_loader_data(array->at(i));
  6002     // We don't need to keep track of new CLDs anymore.
  6003     ClassLoaderDataGraph::remember_new_clds(false);
  6005     verify_work_stacks_empty();
  6009     GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
  6011     verify_work_stacks_empty();
  6013     RemarkKlassClosure remark_klass_closure(&mrias_cl);
  6014     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
  6016     verify_work_stacks_empty();
  6019   // We might have added oops to ClassLoaderData::_handles during the
  6020   // concurrent marking phase. These oops point to newly allocated objects
  6021   // that are guaranteed to be kept alive. Either by the direct allocation
  6022   // code, or when the young collector processes the strong roots. Hence,
  6023   // we don't have to revisit the _handles block during the remark phase.
  6025   verify_work_stacks_empty();
  6026   // Restore evacuated mark words, if any, used for overflow list links
  6027   if (!CMSOverflowEarlyRestoration) {
  6028     restore_preserved_marks_if_any();
  6030   verify_overflow_empty();
  6033 ////////////////////////////////////////////////////////
  6034 // Parallel Reference Processing Task Proxy Class
  6035 ////////////////////////////////////////////////////////
  6036 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
  6037   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
  6038   CMSCollector*          _collector;
  6039   CMSBitMap*             _mark_bit_map;
  6040   const MemRegion        _span;
  6041   ProcessTask&           _task;
  6043 public:
  6044   CMSRefProcTaskProxy(ProcessTask&     task,
  6045                       CMSCollector*    collector,
  6046                       const MemRegion& span,
  6047                       CMSBitMap*       mark_bit_map,
  6048                       AbstractWorkGang* workers,
  6049                       OopTaskQueueSet* task_queues):
  6050     // XXX Should superclass AGTWOQ also know about AWG since it knows
  6051     // about the task_queues used by the AWG? Then it could initialize
  6052     // the terminator() object. See 6984287. The set_for_termination()
  6053     // below is a temporary band-aid for the regression in 6984287.
  6054     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
  6055       task_queues),
  6056     _task(task),
  6057     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
  6059     assert(_collector->_span.equals(_span) && !_span.is_empty(),
  6060            "Inconsistency in _span");
  6061     set_for_termination(workers->active_workers());
  6064   OopTaskQueueSet* task_queues() { return queues(); }
  6066   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
  6068   void do_work_steal(int i,
  6069                      CMSParDrainMarkingStackClosure* drain,
  6070                      CMSParKeepAliveClosure* keep_alive,
  6071                      int* seed);
  6073   virtual void work(uint worker_id);
  6074 };
  6076 void CMSRefProcTaskProxy::work(uint worker_id) {
  6077   assert(_collector->_span.equals(_span), "Inconsistency in _span");
  6078   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
  6079                                         _mark_bit_map,
  6080                                         work_queue(worker_id));
  6081   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
  6082                                                  _mark_bit_map,
  6083                                                  work_queue(worker_id));
  6084   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
  6085   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
  6086   if (_task.marks_oops_alive()) {
  6087     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
  6088                   _collector->hash_seed(worker_id));
  6090   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
  6091   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
  6094 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
  6095   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
  6096   EnqueueTask& _task;
  6098 public:
  6099   CMSRefEnqueueTaskProxy(EnqueueTask& task)
  6100     : AbstractGangTask("Enqueue reference objects in parallel"),
  6101       _task(task)
  6102   { }
  6104   virtual void work(uint worker_id)
  6106     _task.work(worker_id);
  6108 };
  6110 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
  6111   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
  6112    _span(span),
  6113    _bit_map(bit_map),
  6114    _work_queue(work_queue),
  6115    _mark_and_push(collector, span, bit_map, work_queue),
  6116    _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  6117                         (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
  6118 { }
  6120 // . see if we can share work_queues with ParNew? XXX
  6121 void CMSRefProcTaskProxy::do_work_steal(int i,
  6122   CMSParDrainMarkingStackClosure* drain,
  6123   CMSParKeepAliveClosure* keep_alive,
  6124   int* seed) {
  6125   OopTaskQueue* work_q = work_queue(i);
  6126   NOT_PRODUCT(int num_steals = 0;)
  6127   oop obj_to_scan;
  6129   while (true) {
  6130     // Completely finish any left over work from (an) earlier round(s)
  6131     drain->trim_queue(0);
  6132     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
  6133                                          (size_t)ParGCDesiredObjsFromOverflowList);
  6134     // Now check if there's any work in the overflow list
  6135     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
  6136     // only affects the number of attempts made to get work from the
  6137     // overflow list and does not affect the number of workers.  Just
  6138     // pass ParallelGCThreads so this behavior is unchanged.
  6139     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
  6140                                                 work_q,
  6141                                                 ParallelGCThreads)) {
  6142       // Found something in global overflow list;
  6143       // not yet ready to go stealing work from others.
  6144       // We'd like to assert(work_q->size() != 0, ...)
  6145       // because we just took work from the overflow list,
  6146       // but of course we can't, since all of that might have
  6147       // been already stolen from us.
  6148       continue;
  6150     // Verify that we have no work before we resort to stealing
  6151     assert(work_q->size() == 0, "Have work, shouldn't steal");
  6152     // Try to steal from other queues that have work
  6153     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
  6154       NOT_PRODUCT(num_steals++;)
  6155       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
  6156       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
  6157       // Do scanning work
  6158       obj_to_scan->oop_iterate(keep_alive);
  6159       // Loop around, finish this work, and try to steal some more
  6160     } else if (terminator()->offer_termination()) {
  6161       break;  // nirvana from the infinite cycle
  6164   NOT_PRODUCT(
  6165     if (PrintCMSStatistics != 0) {
  6166       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
  6171 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
  6173   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6174   FlexibleWorkGang* workers = gch->workers();
  6175   assert(workers != NULL, "Need parallel worker threads.");
  6176   CMSRefProcTaskProxy rp_task(task, &_collector,
  6177                               _collector.ref_processor()->span(),
  6178                               _collector.markBitMap(),
  6179                               workers, _collector.task_queues());
  6180   workers->run_task(&rp_task);
  6183 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
  6186   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6187   FlexibleWorkGang* workers = gch->workers();
  6188   assert(workers != NULL, "Need parallel worker threads.");
  6189   CMSRefEnqueueTaskProxy enq_task(task);
  6190   workers->run_task(&enq_task);
  6193 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
  6195   ResourceMark rm;
  6196   HandleMark   hm;
  6198   ReferenceProcessor* rp = ref_processor();
  6199   assert(rp->span().equals(_span), "Spans should be equal");
  6200   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
  6201   // Process weak references.
  6202   rp->setup_policy(clear_all_soft_refs);
  6203   verify_work_stacks_empty();
  6205   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
  6206                                           &_markStack, false /* !preclean */);
  6207   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
  6208                                 _span, &_markBitMap, &_markStack,
  6209                                 &cmsKeepAliveClosure, false /* !preclean */);
  6211     GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
  6213     ReferenceProcessorStats stats;
  6214     if (rp->processing_is_mt()) {
  6215       // Set the degree of MT here.  If the discovery is done MT, there
  6216       // may have been a different number of threads doing the discovery
  6217       // and a different number of discovered lists may have Ref objects.
  6218       // That is OK as long as the Reference lists are balanced (see
  6219       // balance_all_queues() and balance_queues()).
  6220       GenCollectedHeap* gch = GenCollectedHeap::heap();
  6221       int active_workers = ParallelGCThreads;
  6222       FlexibleWorkGang* workers = gch->workers();
  6223       if (workers != NULL) {
  6224         active_workers = workers->active_workers();
  6225         // The expectation is that active_workers will have already
  6226         // been set to a reasonable value.  If it has not been set,
  6227         // investigate.
  6228         assert(active_workers > 0, "Should have been set during scavenge");
  6230       rp->set_active_mt_degree(active_workers);
  6231       CMSRefProcTaskExecutor task_executor(*this);
  6232       stats = rp->process_discovered_references(&_is_alive_closure,
  6233                                         &cmsKeepAliveClosure,
  6234                                         &cmsDrainMarkingStackClosure,
  6235                                         &task_executor,
  6236                                         _gc_timer_cm);
  6237     } else {
  6238       stats = rp->process_discovered_references(&_is_alive_closure,
  6239                                         &cmsKeepAliveClosure,
  6240                                         &cmsDrainMarkingStackClosure,
  6241                                         NULL,
  6242                                         _gc_timer_cm);
  6244     _gc_tracer_cm->report_gc_reference_stats(stats);
  6248   // This is the point where the entire marking should have completed.
  6249   verify_work_stacks_empty();
  6251   if (should_unload_classes()) {
  6253       GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
  6255       // Unload classes and purge the SystemDictionary.
  6256       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
  6258       // Unload nmethods.
  6259       CodeCache::do_unloading(&_is_alive_closure, purged_class);
  6261       // Prune dead klasses from subklass/sibling/implementor lists.
  6262       Klass::clean_weak_klass_links(&_is_alive_closure);
  6266       GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
  6267       // Clean up unreferenced symbols in symbol table.
  6268       SymbolTable::unlink();
  6272   // CMS doesn't use the StringTable as hard roots when class unloading is turned off.
  6273   // Need to check if we really scanned the StringTable.
  6274   if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) {
  6275     GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
  6276     // Delete entries for dead interned strings.
  6277     StringTable::unlink(&_is_alive_closure);
  6280   // Restore any preserved marks as a result of mark stack or
  6281   // work queue overflow
  6282   restore_preserved_marks_if_any();  // done single-threaded for now
  6284   rp->set_enqueuing_is_done(true);
  6285   if (rp->processing_is_mt()) {
  6286     rp->balance_all_queues();
  6287     CMSRefProcTaskExecutor task_executor(*this);
  6288     rp->enqueue_discovered_references(&task_executor);
  6289   } else {
  6290     rp->enqueue_discovered_references(NULL);
  6292   rp->verify_no_references_recorded();
  6293   assert(!rp->discovery_enabled(), "should have been disabled");
  6296 #ifndef PRODUCT
  6297 void CMSCollector::check_correct_thread_executing() {
  6298   Thread* t = Thread::current();
  6299   // Only the VM thread or the CMS thread should be here.
  6300   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
  6301          "Unexpected thread type");
  6302   // If this is the vm thread, the foreground process
  6303   // should not be waiting.  Note that _foregroundGCIsActive is
  6304   // true while the foreground collector is waiting.
  6305   if (_foregroundGCShouldWait) {
  6306     // We cannot be the VM thread
  6307     assert(t->is_ConcurrentGC_thread(),
  6308            "Should be CMS thread");
  6309   } else {
  6310     // We can be the CMS thread only if we are in a stop-world
  6311     // phase of CMS collection.
  6312     if (t->is_ConcurrentGC_thread()) {
  6313       assert(_collectorState == InitialMarking ||
  6314              _collectorState == FinalMarking,
  6315              "Should be a stop-world phase");
  6316       // The CMS thread should be holding the CMS_token.
  6317       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6318              "Potential interference with concurrently "
  6319              "executing VM thread");
  6323 #endif
  6325 void CMSCollector::sweep(bool asynch) {
  6326   assert(_collectorState == Sweeping, "just checking");
  6327   check_correct_thread_executing();
  6328   verify_work_stacks_empty();
  6329   verify_overflow_empty();
  6330   increment_sweep_count();
  6331   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
  6333   _inter_sweep_timer.stop();
  6334   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
  6335   size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
  6337   assert(!_intra_sweep_timer.is_active(), "Should not be active");
  6338   _intra_sweep_timer.reset();
  6339   _intra_sweep_timer.start();
  6340   if (asynch) {
  6341     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6342     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
  6343     // First sweep the old gen
  6345       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
  6346                                bitMapLock());
  6347       sweepWork(_cmsGen, asynch);
  6350     // Update Universe::_heap_*_at_gc figures.
  6351     // We need all the free list locks to make the abstract state
  6352     // transition from Sweeping to Resetting. See detailed note
  6353     // further below.
  6355       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
  6356       // Update heap occupancy information which is used as
  6357       // input to soft ref clearing policy at the next gc.
  6358       Universe::update_heap_info_at_gc();
  6359       _collectorState = Resizing;
  6361   } else {
  6362     // already have needed locks
  6363     sweepWork(_cmsGen,  asynch);
  6364     // Update heap occupancy information which is used as
  6365     // input to soft ref clearing policy at the next gc.
  6366     Universe::update_heap_info_at_gc();
  6367     _collectorState = Resizing;
  6369   verify_work_stacks_empty();
  6370   verify_overflow_empty();
  6372   if (should_unload_classes()) {
  6373     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
  6374     // requires that the virtual spaces are stable and not deleted.
  6375     ClassLoaderDataGraph::set_should_purge(true);
  6378   _intra_sweep_timer.stop();
  6379   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
  6381   _inter_sweep_timer.reset();
  6382   _inter_sweep_timer.start();
  6384   // We need to use a monotonically non-deccreasing time in ms
  6385   // or we will see time-warp warnings and os::javaTimeMillis()
  6386   // does not guarantee monotonicity.
  6387   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
  6388   update_time_of_last_gc(now);
  6390   // NOTE on abstract state transitions:
  6391   // Mutators allocate-live and/or mark the mod-union table dirty
  6392   // based on the state of the collection.  The former is done in
  6393   // the interval [Marking, Sweeping] and the latter in the interval
  6394   // [Marking, Sweeping).  Thus the transitions into the Marking state
  6395   // and out of the Sweeping state must be synchronously visible
  6396   // globally to the mutators.
  6397   // The transition into the Marking state happens with the world
  6398   // stopped so the mutators will globally see it.  Sweeping is
  6399   // done asynchronously by the background collector so the transition
  6400   // from the Sweeping state to the Resizing state must be done
  6401   // under the freelistLock (as is the check for whether to
  6402   // allocate-live and whether to dirty the mod-union table).
  6403   assert(_collectorState == Resizing, "Change of collector state to"
  6404     " Resizing must be done under the freelistLocks (plural)");
  6406   // Now that sweeping has been completed, we clear
  6407   // the incremental_collection_failed flag,
  6408   // thus inviting a younger gen collection to promote into
  6409   // this generation. If such a promotion may still fail,
  6410   // the flag will be set again when a young collection is
  6411   // attempted.
  6412   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6413   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
  6414   gch->update_full_collections_completed(_collection_count_start);
  6417 // FIX ME!!! Looks like this belongs in CFLSpace, with
  6418 // CMSGen merely delegating to it.
  6419 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
  6420   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
  6421   HeapWord*  minAddr        = _cmsSpace->bottom();
  6422   HeapWord*  largestAddr    =
  6423     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
  6424   if (largestAddr == NULL) {
  6425     // The dictionary appears to be empty.  In this case
  6426     // try to coalesce at the end of the heap.
  6427     largestAddr = _cmsSpace->end();
  6429   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
  6430   size_t nearLargestOffset =
  6431     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
  6432   if (PrintFLSStatistics != 0) {
  6433     gclog_or_tty->print_cr(
  6434       "CMS: Large Block: " PTR_FORMAT ";"
  6435       " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
  6436       largestAddr,
  6437       _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
  6439   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
  6442 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
  6443   return addr >= _cmsSpace->nearLargestChunk();
  6446 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
  6447   return _cmsSpace->find_chunk_at_end();
  6450 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
  6451                                                     bool full) {
  6452   // The next lower level has been collected.  Gather any statistics
  6453   // that are of interest at this point.
  6454   if (!full && (current_level + 1) == level()) {
  6455     // Gather statistics on the young generation collection.
  6456     collector()->stats().record_gc0_end(used());
  6460 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
  6461   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6462   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
  6463     "Wrong type of heap");
  6464   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
  6465     gch->gen_policy()->size_policy();
  6466   assert(sp->is_gc_cms_adaptive_size_policy(),
  6467     "Wrong type of size policy");
  6468   return sp;
  6471 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
  6472   if (PrintGCDetails && Verbose) {
  6473     gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
  6475   _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
  6476   _debug_collection_type =
  6477     (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
  6478   if (PrintGCDetails && Verbose) {
  6479     gclog_or_tty->print_cr("to %d ", _debug_collection_type);
  6483 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
  6484   bool asynch) {
  6485   // We iterate over the space(s) underlying this generation,
  6486   // checking the mark bit map to see if the bits corresponding
  6487   // to specific blocks are marked or not. Blocks that are
  6488   // marked are live and are not swept up. All remaining blocks
  6489   // are swept up, with coalescing on-the-fly as we sweep up
  6490   // contiguous free and/or garbage blocks:
  6491   // We need to ensure that the sweeper synchronizes with allocators
  6492   // and stop-the-world collectors. In particular, the following
  6493   // locks are used:
  6494   // . CMS token: if this is held, a stop the world collection cannot occur
  6495   // . freelistLock: if this is held no allocation can occur from this
  6496   //                 generation by another thread
  6497   // . bitMapLock: if this is held, no other thread can access or update
  6498   //
  6500   // Note that we need to hold the freelistLock if we use
  6501   // block iterate below; else the iterator might go awry if
  6502   // a mutator (or promotion) causes block contents to change
  6503   // (for instance if the allocator divvies up a block).
  6504   // If we hold the free list lock, for all practical purposes
  6505   // young generation GC's can't occur (they'll usually need to
  6506   // promote), so we might as well prevent all young generation
  6507   // GC's while we do a sweeping step. For the same reason, we might
  6508   // as well take the bit map lock for the entire duration
  6510   // check that we hold the requisite locks
  6511   assert(have_cms_token(), "Should hold cms token");
  6512   assert(   (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
  6513          || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
  6514         "Should possess CMS token to sweep");
  6515   assert_lock_strong(gen->freelistLock());
  6516   assert_lock_strong(bitMapLock());
  6518   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
  6519   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
  6520   gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
  6521                                       _inter_sweep_estimate.padded_average(),
  6522                                       _intra_sweep_estimate.padded_average());
  6523   gen->setNearLargestChunk();
  6526     SweepClosure sweepClosure(this, gen, &_markBitMap,
  6527                             CMSYield && asynch);
  6528     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
  6529     // We need to free-up/coalesce garbage/blocks from a
  6530     // co-terminal free run. This is done in the SweepClosure
  6531     // destructor; so, do not remove this scope, else the
  6532     // end-of-sweep-census below will be off by a little bit.
  6534   gen->cmsSpace()->sweep_completed();
  6535   gen->cmsSpace()->endSweepFLCensus(sweep_count());
  6536   if (should_unload_classes()) {                // unloaded classes this cycle,
  6537     _concurrent_cycles_since_last_unload = 0;   // ... reset count
  6538   } else {                                      // did not unload classes,
  6539     _concurrent_cycles_since_last_unload++;     // ... increment count
  6543 // Reset CMS data structures (for now just the marking bit map)
  6544 // preparatory for the next cycle.
  6545 void CMSCollector::reset(bool asynch) {
  6546   GenCollectedHeap* gch = GenCollectedHeap::heap();
  6547   CMSAdaptiveSizePolicy* sp = size_policy();
  6548   AdaptiveSizePolicyOutput(sp, gch->total_collections());
  6549   if (asynch) {
  6550     CMSTokenSyncWithLocks ts(true, bitMapLock());
  6552     // If the state is not "Resetting", the foreground  thread
  6553     // has done a collection and the resetting.
  6554     if (_collectorState != Resetting) {
  6555       assert(_collectorState == Idling, "The state should only change"
  6556         " because the foreground collector has finished the collection");
  6557       return;
  6560     // Clear the mark bitmap (no grey objects to start with)
  6561     // for the next cycle.
  6562     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6563     CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
  6565     HeapWord* curAddr = _markBitMap.startWord();
  6566     while (curAddr < _markBitMap.endWord()) {
  6567       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
  6568       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
  6569       _markBitMap.clear_large_range(chunk);
  6570       if (ConcurrentMarkSweepThread::should_yield() &&
  6571           !foregroundGCIsActive() &&
  6572           CMSYield) {
  6573         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  6574                "CMS thread should hold CMS token");
  6575         assert_lock_strong(bitMapLock());
  6576         bitMapLock()->unlock();
  6577         ConcurrentMarkSweepThread::desynchronize(true);
  6578         ConcurrentMarkSweepThread::acknowledge_yield_request();
  6579         stopTimer();
  6580         if (PrintCMSStatistics != 0) {
  6581           incrementYields();
  6583         icms_wait();
  6585         // See the comment in coordinator_yield()
  6586         for (unsigned i = 0; i < CMSYieldSleepCount &&
  6587                          ConcurrentMarkSweepThread::should_yield() &&
  6588                          !CMSCollector::foregroundGCIsActive(); ++i) {
  6589           os::sleep(Thread::current(), 1, false);
  6590           ConcurrentMarkSweepThread::acknowledge_yield_request();
  6593         ConcurrentMarkSweepThread::synchronize(true);
  6594         bitMapLock()->lock_without_safepoint_check();
  6595         startTimer();
  6597       curAddr = chunk.end();
  6599     // A successful mostly concurrent collection has been done.
  6600     // Because only the full (i.e., concurrent mode failure) collections
  6601     // are being measured for gc overhead limits, clean the "near" flag
  6602     // and count.
  6603     sp->reset_gc_overhead_limit_count();
  6604     _collectorState = Idling;
  6605   } else {
  6606     // already have the lock
  6607     assert(_collectorState == Resetting, "just checking");
  6608     assert_lock_strong(bitMapLock());
  6609     _markBitMap.clear_all();
  6610     _collectorState = Idling;
  6613   // Stop incremental mode after a cycle completes, so that any future cycles
  6614   // are triggered by allocation.
  6615   stop_icms();
  6617   NOT_PRODUCT(
  6618     if (RotateCMSCollectionTypes) {
  6619       _cmsGen->rotate_debug_collection_type();
  6623   register_gc_end();
  6626 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
  6627   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
  6628   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
  6629   GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
  6630   TraceCollectorStats tcs(counters());
  6632   switch (op) {
  6633     case CMS_op_checkpointRootsInitial: {
  6634       SvcGCMarker sgcm(SvcGCMarker::OTHER);
  6635       checkpointRootsInitial(true);       // asynch
  6636       if (PrintGC) {
  6637         _cmsGen->printOccupancy("initial-mark");
  6639       break;
  6641     case CMS_op_checkpointRootsFinal: {
  6642       SvcGCMarker sgcm(SvcGCMarker::OTHER);
  6643       checkpointRootsFinal(true,    // asynch
  6644                            false,   // !clear_all_soft_refs
  6645                            false);  // !init_mark_was_synchronous
  6646       if (PrintGC) {
  6647         _cmsGen->printOccupancy("remark");
  6649       break;
  6651     default:
  6652       fatal("No such CMS_op");
  6656 #ifndef PRODUCT
  6657 size_t const CMSCollector::skip_header_HeapWords() {
  6658   return FreeChunk::header_size();
  6661 // Try and collect here conditions that should hold when
  6662 // CMS thread is exiting. The idea is that the foreground GC
  6663 // thread should not be blocked if it wants to terminate
  6664 // the CMS thread and yet continue to run the VM for a while
  6665 // after that.
  6666 void CMSCollector::verify_ok_to_terminate() const {
  6667   assert(Thread::current()->is_ConcurrentGC_thread(),
  6668          "should be called by CMS thread");
  6669   assert(!_foregroundGCShouldWait, "should be false");
  6670   // We could check here that all the various low-level locks
  6671   // are not held by the CMS thread, but that is overkill; see
  6672   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
  6673   // is checked.
  6675 #endif
  6677 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
  6678    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
  6679           "missing Printezis mark?");
  6680   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6681   size_t size = pointer_delta(nextOneAddr + 1, addr);
  6682   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6683          "alignment problem");
  6684   assert(size >= 3, "Necessary for Printezis marks to work");
  6685   return size;
  6688 // A variant of the above (block_size_using_printezis_bits()) except
  6689 // that we return 0 if the P-bits are not yet set.
  6690 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
  6691   if (_markBitMap.isMarked(addr + 1)) {
  6692     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
  6693     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
  6694     size_t size = pointer_delta(nextOneAddr + 1, addr);
  6695     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  6696            "alignment problem");
  6697     assert(size >= 3, "Necessary for Printezis marks to work");
  6698     return size;
  6700   return 0;
  6703 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
  6704   size_t sz = 0;
  6705   oop p = (oop)addr;
  6706   if (p->klass_or_null() != NULL) {
  6707     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
  6708   } else {
  6709     sz = block_size_using_printezis_bits(addr);
  6711   assert(sz > 0, "size must be nonzero");
  6712   HeapWord* next_block = addr + sz;
  6713   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
  6714                                              CardTableModRefBS::card_size);
  6715   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
  6716          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
  6717          "must be different cards");
  6718   return next_card;
  6722 // CMS Bit Map Wrapper /////////////////////////////////////////
  6724 // Construct a CMS bit map infrastructure, but don't create the
  6725 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
  6726 // further below.
  6727 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
  6728   _bm(),
  6729   _shifter(shifter),
  6730   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
  6732   _bmStartWord = 0;
  6733   _bmWordSize  = 0;
  6736 bool CMSBitMap::allocate(MemRegion mr) {
  6737   _bmStartWord = mr.start();
  6738   _bmWordSize  = mr.word_size();
  6739   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
  6740                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
  6741   if (!brs.is_reserved()) {
  6742     warning("CMS bit map allocation failure");
  6743     return false;
  6745   // For now we'll just commit all of the bit map up fromt.
  6746   // Later on we'll try to be more parsimonious with swap.
  6747   if (!_virtual_space.initialize(brs, brs.size())) {
  6748     warning("CMS bit map backing store failure");
  6749     return false;
  6751   assert(_virtual_space.committed_size() == brs.size(),
  6752          "didn't reserve backing store for all of CMS bit map?");
  6753   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
  6754   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
  6755          _bmWordSize, "inconsistency in bit map sizing");
  6756   _bm.set_size(_bmWordSize >> _shifter);
  6758   // bm.clear(); // can we rely on getting zero'd memory? verify below
  6759   assert(isAllClear(),
  6760          "Expected zero'd memory from ReservedSpace constructor");
  6761   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
  6762          "consistency check");
  6763   return true;
  6766 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
  6767   HeapWord *next_addr, *end_addr, *last_addr;
  6768   assert_locked();
  6769   assert(covers(mr), "out-of-range error");
  6770   // XXX assert that start and end are appropriately aligned
  6771   for (next_addr = mr.start(), end_addr = mr.end();
  6772        next_addr < end_addr; next_addr = last_addr) {
  6773     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
  6774     last_addr = dirty_region.end();
  6775     if (!dirty_region.is_empty()) {
  6776       cl->do_MemRegion(dirty_region);
  6777     } else {
  6778       assert(last_addr == end_addr, "program logic");
  6779       return;
  6784 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
  6785   _bm.print_on_error(st, prefix);
  6788 #ifndef PRODUCT
  6789 void CMSBitMap::assert_locked() const {
  6790   CMSLockVerifier::assert_locked(lock());
  6793 bool CMSBitMap::covers(MemRegion mr) const {
  6794   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  6795   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
  6796          "size inconsistency");
  6797   return (mr.start() >= _bmStartWord) &&
  6798          (mr.end()   <= endWord());
  6801 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
  6802     return (start >= _bmStartWord && (start + size) <= endWord());
  6805 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
  6806   // verify that there are no 1 bits in the interval [left, right)
  6807   FalseBitMapClosure falseBitMapClosure;
  6808   iterate(&falseBitMapClosure, left, right);
  6811 void CMSBitMap::region_invariant(MemRegion mr)
  6813   assert_locked();
  6814   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
  6815   assert(!mr.is_empty(), "unexpected empty region");
  6816   assert(covers(mr), "mr should be covered by bit map");
  6817   // convert address range into offset range
  6818   size_t start_ofs = heapWordToOffset(mr.start());
  6819   // Make sure that end() is appropriately aligned
  6820   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
  6821                         (1 << (_shifter+LogHeapWordSize))),
  6822          "Misaligned mr.end()");
  6823   size_t end_ofs   = heapWordToOffset(mr.end());
  6824   assert(end_ofs > start_ofs, "Should mark at least one bit");
  6827 #endif
  6829 bool CMSMarkStack::allocate(size_t size) {
  6830   // allocate a stack of the requisite depth
  6831   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6832                    size * sizeof(oop)));
  6833   if (!rs.is_reserved()) {
  6834     warning("CMSMarkStack allocation failure");
  6835     return false;
  6837   if (!_virtual_space.initialize(rs, rs.size())) {
  6838     warning("CMSMarkStack backing store failure");
  6839     return false;
  6841   assert(_virtual_space.committed_size() == rs.size(),
  6842          "didn't reserve backing store for all of CMS stack?");
  6843   _base = (oop*)(_virtual_space.low());
  6844   _index = 0;
  6845   _capacity = size;
  6846   NOT_PRODUCT(_max_depth = 0);
  6847   return true;
  6850 // XXX FIX ME !!! In the MT case we come in here holding a
  6851 // leaf lock. For printing we need to take a further lock
  6852 // which has lower rank. We need to recallibrate the two
  6853 // lock-ranks involved in order to be able to rpint the
  6854 // messages below. (Or defer the printing to the caller.
  6855 // For now we take the expedient path of just disabling the
  6856 // messages for the problematic case.)
  6857 void CMSMarkStack::expand() {
  6858   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
  6859   if (_capacity == MarkStackSizeMax) {
  6860     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6861       // We print a warning message only once per CMS cycle.
  6862       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
  6864     return;
  6866   // Double capacity if possible
  6867   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
  6868   // Do not give up existing stack until we have managed to
  6869   // get the double capacity that we desired.
  6870   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
  6871                    new_capacity * sizeof(oop)));
  6872   if (rs.is_reserved()) {
  6873     // Release the backing store associated with old stack
  6874     _virtual_space.release();
  6875     // Reinitialize virtual space for new stack
  6876     if (!_virtual_space.initialize(rs, rs.size())) {
  6877       fatal("Not enough swap for expanded marking stack");
  6879     _base = (oop*)(_virtual_space.low());
  6880     _index = 0;
  6881     _capacity = new_capacity;
  6882   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
  6883     // Failed to double capacity, continue;
  6884     // we print a detail message only once per CMS cycle.
  6885     gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
  6886             SIZE_FORMAT"K",
  6887             _capacity / K, new_capacity / K);
  6892 // Closures
  6893 // XXX: there seems to be a lot of code  duplication here;
  6894 // should refactor and consolidate common code.
  6896 // This closure is used to mark refs into the CMS generation in
  6897 // the CMS bit map. Called at the first checkpoint. This closure
  6898 // assumes that we do not need to re-mark dirty cards; if the CMS
  6899 // generation on which this is used is not an oldest
  6900 // generation then this will lose younger_gen cards!
  6902 MarkRefsIntoClosure::MarkRefsIntoClosure(
  6903   MemRegion span, CMSBitMap* bitMap):
  6904     _span(span),
  6905     _bitMap(bitMap)
  6907     assert(_ref_processor == NULL, "deliberately left NULL");
  6908     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
  6911 void MarkRefsIntoClosure::do_oop(oop obj) {
  6912   // if p points into _span, then mark corresponding bit in _markBitMap
  6913   assert(obj->is_oop(), "expected an oop");
  6914   HeapWord* addr = (HeapWord*)obj;
  6915   if (_span.contains(addr)) {
  6916     // this should be made more efficient
  6917     _bitMap->mark(addr);
  6921 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
  6922 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
  6924 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
  6925   MemRegion span, CMSBitMap* bitMap):
  6926     _span(span),
  6927     _bitMap(bitMap)
  6929     assert(_ref_processor == NULL, "deliberately left NULL");
  6930     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
  6933 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
  6934   // if p points into _span, then mark corresponding bit in _markBitMap
  6935   assert(obj->is_oop(), "expected an oop");
  6936   HeapWord* addr = (HeapWord*)obj;
  6937   if (_span.contains(addr)) {
  6938     // this should be made more efficient
  6939     _bitMap->par_mark(addr);
  6943 void Par_MarkRefsIntoClosure::do_oop(oop* p)       { Par_MarkRefsIntoClosure::do_oop_work(p); }
  6944 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
  6946 // A variant of the above, used for CMS marking verification.
  6947 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
  6948   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
  6949     _span(span),
  6950     _verification_bm(verification_bm),
  6951     _cms_bm(cms_bm)
  6953     assert(_ref_processor == NULL, "deliberately left NULL");
  6954     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
  6957 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
  6958   // if p points into _span, then mark corresponding bit in _markBitMap
  6959   assert(obj->is_oop(), "expected an oop");
  6960   HeapWord* addr = (HeapWord*)obj;
  6961   if (_span.contains(addr)) {
  6962     _verification_bm->mark(addr);
  6963     if (!_cms_bm->isMarked(addr)) {
  6964       oop(addr)->print();
  6965       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
  6966       fatal("... aborting");
  6971 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6972 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
  6974 //////////////////////////////////////////////////
  6975 // MarkRefsIntoAndScanClosure
  6976 //////////////////////////////////////////////////
  6978 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
  6979                                                        ReferenceProcessor* rp,
  6980                                                        CMSBitMap* bit_map,
  6981                                                        CMSBitMap* mod_union_table,
  6982                                                        CMSMarkStack*  mark_stack,
  6983                                                        CMSCollector* collector,
  6984                                                        bool should_yield,
  6985                                                        bool concurrent_precleaning):
  6986   _collector(collector),
  6987   _span(span),
  6988   _bit_map(bit_map),
  6989   _mark_stack(mark_stack),
  6990   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
  6991                       mark_stack, concurrent_precleaning),
  6992   _yield(should_yield),
  6993   _concurrent_precleaning(concurrent_precleaning),
  6994   _freelistLock(NULL)
  6996   _ref_processor = rp;
  6997   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7000 // This closure is used to mark refs into the CMS generation at the
  7001 // second (final) checkpoint, and to scan and transitively follow
  7002 // the unmarked oops. It is also used during the concurrent precleaning
  7003 // phase while scanning objects on dirty cards in the CMS generation.
  7004 // The marks are made in the marking bit map and the marking stack is
  7005 // used for keeping the (newly) grey objects during the scan.
  7006 // The parallel version (Par_...) appears further below.
  7007 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  7008   if (obj != NULL) {
  7009     assert(obj->is_oop(), "expected an oop");
  7010     HeapWord* addr = (HeapWord*)obj;
  7011     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  7012     assert(_collector->overflow_list_is_empty(),
  7013            "overflow list should be empty");
  7014     if (_span.contains(addr) &&
  7015         !_bit_map->isMarked(addr)) {
  7016       // mark bit map (object is now grey)
  7017       _bit_map->mark(addr);
  7018       // push on marking stack (stack should be empty), and drain the
  7019       // stack by applying this closure to the oops in the oops popped
  7020       // from the stack (i.e. blacken the grey objects)
  7021       bool res = _mark_stack->push(obj);
  7022       assert(res, "Should have space to push on empty stack");
  7023       do {
  7024         oop new_oop = _mark_stack->pop();
  7025         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  7026         assert(_bit_map->isMarked((HeapWord*)new_oop),
  7027                "only grey objects on this stack");
  7028         // iterate over the oops in this oop, marking and pushing
  7029         // the ones in CMS heap (i.e. in _span).
  7030         new_oop->oop_iterate(&_pushAndMarkClosure);
  7031         // check if it's time to yield
  7032         do_yield_check();
  7033       } while (!_mark_stack->isEmpty() ||
  7034                (!_concurrent_precleaning && take_from_overflow_list()));
  7035         // if marking stack is empty, and we are not doing this
  7036         // during precleaning, then check the overflow list
  7038     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  7039     assert(_collector->overflow_list_is_empty(),
  7040            "overflow list was drained above");
  7041     // We could restore evacuated mark words, if any, used for
  7042     // overflow list links here because the overflow list is
  7043     // provably empty here. That would reduce the maximum
  7044     // size requirements for preserved_{oop,mark}_stack.
  7045     // But we'll just postpone it until we are all done
  7046     // so we can just stream through.
  7047     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
  7048       _collector->restore_preserved_marks_if_any();
  7049       assert(_collector->no_preserved_marks(), "No preserved marks");
  7051     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
  7052            "All preserved marks should have been restored above");
  7056 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  7057 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
  7059 void MarkRefsIntoAndScanClosure::do_yield_work() {
  7060   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7061          "CMS thread should hold CMS token");
  7062   assert_lock_strong(_freelistLock);
  7063   assert_lock_strong(_bit_map->lock());
  7064   // relinquish the free_list_lock and bitMaplock()
  7065   _bit_map->lock()->unlock();
  7066   _freelistLock->unlock();
  7067   ConcurrentMarkSweepThread::desynchronize(true);
  7068   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7069   _collector->stopTimer();
  7070   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7071   if (PrintCMSStatistics != 0) {
  7072     _collector->incrementYields();
  7074   _collector->icms_wait();
  7076   // See the comment in coordinator_yield()
  7077   for (unsigned i = 0;
  7078        i < CMSYieldSleepCount &&
  7079        ConcurrentMarkSweepThread::should_yield() &&
  7080        !CMSCollector::foregroundGCIsActive();
  7081        ++i) {
  7082     os::sleep(Thread::current(), 1, false);
  7083     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7086   ConcurrentMarkSweepThread::synchronize(true);
  7087   _freelistLock->lock_without_safepoint_check();
  7088   _bit_map->lock()->lock_without_safepoint_check();
  7089   _collector->startTimer();
  7092 ///////////////////////////////////////////////////////////
  7093 // Par_MarkRefsIntoAndScanClosure: a parallel version of
  7094 //                                 MarkRefsIntoAndScanClosure
  7095 ///////////////////////////////////////////////////////////
  7096 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
  7097   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
  7098   CMSBitMap* bit_map, OopTaskQueue* work_queue):
  7099   _span(span),
  7100   _bit_map(bit_map),
  7101   _work_queue(work_queue),
  7102   _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
  7103                        (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
  7104   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
  7106   _ref_processor = rp;
  7107   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7110 // This closure is used to mark refs into the CMS generation at the
  7111 // second (final) checkpoint, and to scan and transitively follow
  7112 // the unmarked oops. The marks are made in the marking bit map and
  7113 // the work_queue is used for keeping the (newly) grey objects during
  7114 // the scan phase whence they are also available for stealing by parallel
  7115 // threads. Since the marking bit map is shared, updates are
  7116 // synchronized (via CAS).
  7117 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
  7118   if (obj != NULL) {
  7119     // Ignore mark word because this could be an already marked oop
  7120     // that may be chained at the end of the overflow list.
  7121     assert(obj->is_oop(true), "expected an oop");
  7122     HeapWord* addr = (HeapWord*)obj;
  7123     if (_span.contains(addr) &&
  7124         !_bit_map->isMarked(addr)) {
  7125       // mark bit map (object will become grey):
  7126       // It is possible for several threads to be
  7127       // trying to "claim" this object concurrently;
  7128       // the unique thread that succeeds in marking the
  7129       // object first will do the subsequent push on
  7130       // to the work queue (or overflow list).
  7131       if (_bit_map->par_mark(addr)) {
  7132         // push on work_queue (which may not be empty), and trim the
  7133         // queue to an appropriate length by applying this closure to
  7134         // the oops in the oops popped from the stack (i.e. blacken the
  7135         // grey objects)
  7136         bool res = _work_queue->push(obj);
  7137         assert(res, "Low water mark should be less than capacity?");
  7138         trim_queue(_low_water_mark);
  7139       } // Else, another thread claimed the object
  7144 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  7145 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
  7147 // This closure is used to rescan the marked objects on the dirty cards
  7148 // in the mod union table and the card table proper.
  7149 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
  7150   oop p, MemRegion mr) {
  7152   size_t size = 0;
  7153   HeapWord* addr = (HeapWord*)p;
  7154   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  7155   assert(_span.contains(addr), "we are scanning the CMS generation");
  7156   // check if it's time to yield
  7157   if (do_yield_check()) {
  7158     // We yielded for some foreground stop-world work,
  7159     // and we have been asked to abort this ongoing preclean cycle.
  7160     return 0;
  7162   if (_bitMap->isMarked(addr)) {
  7163     // it's marked; is it potentially uninitialized?
  7164     if (p->klass_or_null() != NULL) {
  7165         // an initialized object; ignore mark word in verification below
  7166         // since we are running concurrent with mutators
  7167         assert(p->is_oop(true), "should be an oop");
  7168         if (p->is_objArray()) {
  7169           // objArrays are precisely marked; restrict scanning
  7170           // to dirty cards only.
  7171           size = CompactibleFreeListSpace::adjustObjectSize(
  7172                    p->oop_iterate(_scanningClosure, mr));
  7173         } else {
  7174           // A non-array may have been imprecisely marked; we need
  7175           // to scan object in its entirety.
  7176           size = CompactibleFreeListSpace::adjustObjectSize(
  7177                    p->oop_iterate(_scanningClosure));
  7179         #ifdef ASSERT
  7180           size_t direct_size =
  7181             CompactibleFreeListSpace::adjustObjectSize(p->size());
  7182           assert(size == direct_size, "Inconsistency in size");
  7183           assert(size >= 3, "Necessary for Printezis marks to work");
  7184           if (!_bitMap->isMarked(addr+1)) {
  7185             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
  7186           } else {
  7187             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
  7188             assert(_bitMap->isMarked(addr+size-1),
  7189                    "inconsistent Printezis mark");
  7191         #endif // ASSERT
  7192     } else {
  7193       // an unitialized object
  7194       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
  7195       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  7196       size = pointer_delta(nextOneAddr + 1, addr);
  7197       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  7198              "alignment problem");
  7199       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
  7200       // will dirty the card when the klass pointer is installed in the
  7201       // object (signalling the completion of initialization).
  7203   } else {
  7204     // Either a not yet marked object or an uninitialized object
  7205     if (p->klass_or_null() == NULL) {
  7206       // An uninitialized object, skip to the next card, since
  7207       // we may not be able to read its P-bits yet.
  7208       assert(size == 0, "Initial value");
  7209     } else {
  7210       // An object not (yet) reached by marking: we merely need to
  7211       // compute its size so as to go look at the next block.
  7212       assert(p->is_oop(true), "should be an oop");
  7213       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
  7216   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  7217   return size;
  7220 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
  7221   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7222          "CMS thread should hold CMS token");
  7223   assert_lock_strong(_freelistLock);
  7224   assert_lock_strong(_bitMap->lock());
  7225   // relinquish the free_list_lock and bitMaplock()
  7226   _bitMap->lock()->unlock();
  7227   _freelistLock->unlock();
  7228   ConcurrentMarkSweepThread::desynchronize(true);
  7229   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7230   _collector->stopTimer();
  7231   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7232   if (PrintCMSStatistics != 0) {
  7233     _collector->incrementYields();
  7235   _collector->icms_wait();
  7237   // See the comment in coordinator_yield()
  7238   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7239                    ConcurrentMarkSweepThread::should_yield() &&
  7240                    !CMSCollector::foregroundGCIsActive(); ++i) {
  7241     os::sleep(Thread::current(), 1, false);
  7242     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7245   ConcurrentMarkSweepThread::synchronize(true);
  7246   _freelistLock->lock_without_safepoint_check();
  7247   _bitMap->lock()->lock_without_safepoint_check();
  7248   _collector->startTimer();
  7252 //////////////////////////////////////////////////////////////////
  7253 // SurvivorSpacePrecleanClosure
  7254 //////////////////////////////////////////////////////////////////
  7255 // This (single-threaded) closure is used to preclean the oops in
  7256 // the survivor spaces.
  7257 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
  7259   HeapWord* addr = (HeapWord*)p;
  7260   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
  7261   assert(!_span.contains(addr), "we are scanning the survivor spaces");
  7262   assert(p->klass_or_null() != NULL, "object should be initializd");
  7263   // an initialized object; ignore mark word in verification below
  7264   // since we are running concurrent with mutators
  7265   assert(p->is_oop(true), "should be an oop");
  7266   // Note that we do not yield while we iterate over
  7267   // the interior oops of p, pushing the relevant ones
  7268   // on our marking stack.
  7269   size_t size = p->oop_iterate(_scanning_closure);
  7270   do_yield_check();
  7271   // Observe that below, we do not abandon the preclean
  7272   // phase as soon as we should; rather we empty the
  7273   // marking stack before returning. This is to satisfy
  7274   // some existing assertions. In general, it may be a
  7275   // good idea to abort immediately and complete the marking
  7276   // from the grey objects at a later time.
  7277   while (!_mark_stack->isEmpty()) {
  7278     oop new_oop = _mark_stack->pop();
  7279     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  7280     assert(_bit_map->isMarked((HeapWord*)new_oop),
  7281            "only grey objects on this stack");
  7282     // iterate over the oops in this oop, marking and pushing
  7283     // the ones in CMS heap (i.e. in _span).
  7284     new_oop->oop_iterate(_scanning_closure);
  7285     // check if it's time to yield
  7286     do_yield_check();
  7288   unsigned int after_count =
  7289     GenCollectedHeap::heap()->total_collections();
  7290   bool abort = (_before_count != after_count) ||
  7291                _collector->should_abort_preclean();
  7292   return abort ? 0 : size;
  7295 void SurvivorSpacePrecleanClosure::do_yield_work() {
  7296   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7297          "CMS thread should hold CMS token");
  7298   assert_lock_strong(_bit_map->lock());
  7299   // Relinquish the bit map lock
  7300   _bit_map->lock()->unlock();
  7301   ConcurrentMarkSweepThread::desynchronize(true);
  7302   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7303   _collector->stopTimer();
  7304   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7305   if (PrintCMSStatistics != 0) {
  7306     _collector->incrementYields();
  7308   _collector->icms_wait();
  7310   // See the comment in coordinator_yield()
  7311   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7312                        ConcurrentMarkSweepThread::should_yield() &&
  7313                        !CMSCollector::foregroundGCIsActive(); ++i) {
  7314     os::sleep(Thread::current(), 1, false);
  7315     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7318   ConcurrentMarkSweepThread::synchronize(true);
  7319   _bit_map->lock()->lock_without_safepoint_check();
  7320   _collector->startTimer();
  7323 // This closure is used to rescan the marked objects on the dirty cards
  7324 // in the mod union table and the card table proper. In the parallel
  7325 // case, although the bitMap is shared, we do a single read so the
  7326 // isMarked() query is "safe".
  7327 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
  7328   // Ignore mark word because we are running concurrent with mutators
  7329   assert(p->is_oop_or_null(true), "expected an oop or null");
  7330   HeapWord* addr = (HeapWord*)p;
  7331   assert(_span.contains(addr), "we are scanning the CMS generation");
  7332   bool is_obj_array = false;
  7333   #ifdef ASSERT
  7334     if (!_parallel) {
  7335       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
  7336       assert(_collector->overflow_list_is_empty(),
  7337              "overflow list should be empty");
  7340   #endif // ASSERT
  7341   if (_bit_map->isMarked(addr)) {
  7342     // Obj arrays are precisely marked, non-arrays are not;
  7343     // so we scan objArrays precisely and non-arrays in their
  7344     // entirety.
  7345     if (p->is_objArray()) {
  7346       is_obj_array = true;
  7347       if (_parallel) {
  7348         p->oop_iterate(_par_scan_closure, mr);
  7349       } else {
  7350         p->oop_iterate(_scan_closure, mr);
  7352     } else {
  7353       if (_parallel) {
  7354         p->oop_iterate(_par_scan_closure);
  7355       } else {
  7356         p->oop_iterate(_scan_closure);
  7360   #ifdef ASSERT
  7361     if (!_parallel) {
  7362       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
  7363       assert(_collector->overflow_list_is_empty(),
  7364              "overflow list should be empty");
  7367   #endif // ASSERT
  7368   return is_obj_array;
  7371 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
  7372                         MemRegion span,
  7373                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
  7374                         bool should_yield, bool verifying):
  7375   _collector(collector),
  7376   _span(span),
  7377   _bitMap(bitMap),
  7378   _mut(&collector->_modUnionTable),
  7379   _markStack(markStack),
  7380   _yield(should_yield),
  7381   _skipBits(0)
  7383   assert(_markStack->isEmpty(), "stack should be empty");
  7384   _finger = _bitMap->startWord();
  7385   _threshold = _finger;
  7386   assert(_collector->_restart_addr == NULL, "Sanity check");
  7387   assert(_span.contains(_finger), "Out of bounds _finger?");
  7388   DEBUG_ONLY(_verifying = verifying;)
  7391 void MarkFromRootsClosure::reset(HeapWord* addr) {
  7392   assert(_markStack->isEmpty(), "would cause duplicates on stack");
  7393   assert(_span.contains(addr), "Out of bounds _finger?");
  7394   _finger = addr;
  7395   _threshold = (HeapWord*)round_to(
  7396                  (intptr_t)_finger, CardTableModRefBS::card_size);
  7399 // Should revisit to see if this should be restructured for
  7400 // greater efficiency.
  7401 bool MarkFromRootsClosure::do_bit(size_t offset) {
  7402   if (_skipBits > 0) {
  7403     _skipBits--;
  7404     return true;
  7406   // convert offset into a HeapWord*
  7407   HeapWord* addr = _bitMap->startWord() + offset;
  7408   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
  7409          "address out of range");
  7410   assert(_bitMap->isMarked(addr), "tautology");
  7411   if (_bitMap->isMarked(addr+1)) {
  7412     // this is an allocated but not yet initialized object
  7413     assert(_skipBits == 0, "tautology");
  7414     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
  7415     oop p = oop(addr);
  7416     if (p->klass_or_null() == NULL) {
  7417       DEBUG_ONLY(if (!_verifying) {)
  7418         // We re-dirty the cards on which this object lies and increase
  7419         // the _threshold so that we'll come back to scan this object
  7420         // during the preclean or remark phase. (CMSCleanOnEnter)
  7421         if (CMSCleanOnEnter) {
  7422           size_t sz = _collector->block_size_using_printezis_bits(addr);
  7423           HeapWord* end_card_addr   = (HeapWord*)round_to(
  7424                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  7425           MemRegion redirty_range = MemRegion(addr, end_card_addr);
  7426           assert(!redirty_range.is_empty(), "Arithmetical tautology");
  7427           // Bump _threshold to end_card_addr; note that
  7428           // _threshold cannot possibly exceed end_card_addr, anyhow.
  7429           // This prevents future clearing of the card as the scan proceeds
  7430           // to the right.
  7431           assert(_threshold <= end_card_addr,
  7432                  "Because we are just scanning into this object");
  7433           if (_threshold < end_card_addr) {
  7434             _threshold = end_card_addr;
  7436           if (p->klass_or_null() != NULL) {
  7437             // Redirty the range of cards...
  7438             _mut->mark_range(redirty_range);
  7439           } // ...else the setting of klass will dirty the card anyway.
  7441       DEBUG_ONLY(})
  7442       return true;
  7445   scanOopsInOop(addr);
  7446   return true;
  7449 // We take a break if we've been at this for a while,
  7450 // so as to avoid monopolizing the locks involved.
  7451 void MarkFromRootsClosure::do_yield_work() {
  7452   // First give up the locks, then yield, then re-lock
  7453   // We should probably use a constructor/destructor idiom to
  7454   // do this unlock/lock or modify the MutexUnlocker class to
  7455   // serve our purpose. XXX
  7456   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  7457          "CMS thread should hold CMS token");
  7458   assert_lock_strong(_bitMap->lock());
  7459   _bitMap->lock()->unlock();
  7460   ConcurrentMarkSweepThread::desynchronize(true);
  7461   ConcurrentMarkSweepThread::acknowledge_yield_request();
  7462   _collector->stopTimer();
  7463   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  7464   if (PrintCMSStatistics != 0) {
  7465     _collector->incrementYields();
  7467   _collector->icms_wait();
  7469   // See the comment in coordinator_yield()
  7470   for (unsigned i = 0; i < CMSYieldSleepCount &&
  7471                        ConcurrentMarkSweepThread::should_yield() &&
  7472                        !CMSCollector::foregroundGCIsActive(); ++i) {
  7473     os::sleep(Thread::current(), 1, false);
  7474     ConcurrentMarkSweepThread::acknowledge_yield_request();
  7477   ConcurrentMarkSweepThread::synchronize(true);
  7478   _bitMap->lock()->lock_without_safepoint_check();
  7479   _collector->startTimer();
  7482 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
  7483   assert(_bitMap->isMarked(ptr), "expected bit to be set");
  7484   assert(_markStack->isEmpty(),
  7485          "should drain stack to limit stack usage");
  7486   // convert ptr to an oop preparatory to scanning
  7487   oop obj = oop(ptr);
  7488   // Ignore mark word in verification below, since we
  7489   // may be running concurrent with mutators.
  7490   assert(obj->is_oop(true), "should be an oop");
  7491   assert(_finger <= ptr, "_finger runneth ahead");
  7492   // advance the finger to right end of this object
  7493   _finger = ptr + obj->size();
  7494   assert(_finger > ptr, "we just incremented it above");
  7495   // On large heaps, it may take us some time to get through
  7496   // the marking phase (especially if running iCMS). During
  7497   // this time it's possible that a lot of mutations have
  7498   // accumulated in the card table and the mod union table --
  7499   // these mutation records are redundant until we have
  7500   // actually traced into the corresponding card.
  7501   // Here, we check whether advancing the finger would make
  7502   // us cross into a new card, and if so clear corresponding
  7503   // cards in the MUT (preclean them in the card-table in the
  7504   // future).
  7506   DEBUG_ONLY(if (!_verifying) {)
  7507     // The clean-on-enter optimization is disabled by default,
  7508     // until we fix 6178663.
  7509     if (CMSCleanOnEnter && (_finger > _threshold)) {
  7510       // [_threshold, _finger) represents the interval
  7511       // of cards to be cleared  in MUT (or precleaned in card table).
  7512       // The set of cards to be cleared is all those that overlap
  7513       // with the interval [_threshold, _finger); note that
  7514       // _threshold is always kept card-aligned but _finger isn't
  7515       // always card-aligned.
  7516       HeapWord* old_threshold = _threshold;
  7517       assert(old_threshold == (HeapWord*)round_to(
  7518               (intptr_t)old_threshold, CardTableModRefBS::card_size),
  7519              "_threshold should always be card-aligned");
  7520       _threshold = (HeapWord*)round_to(
  7521                      (intptr_t)_finger, CardTableModRefBS::card_size);
  7522       MemRegion mr(old_threshold, _threshold);
  7523       assert(!mr.is_empty(), "Control point invariant");
  7524       assert(_span.contains(mr), "Should clear within span");
  7525       _mut->clear_range(mr);
  7527   DEBUG_ONLY(})
  7528   // Note: the finger doesn't advance while we drain
  7529   // the stack below.
  7530   PushOrMarkClosure pushOrMarkClosure(_collector,
  7531                                       _span, _bitMap, _markStack,
  7532                                       _finger, this);
  7533   bool res = _markStack->push(obj);
  7534   assert(res, "Empty non-zero size stack should have space for single push");
  7535   while (!_markStack->isEmpty()) {
  7536     oop new_oop = _markStack->pop();
  7537     // Skip verifying header mark word below because we are
  7538     // running concurrent with mutators.
  7539     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7540     // now scan this oop's oops
  7541     new_oop->oop_iterate(&pushOrMarkClosure);
  7542     do_yield_check();
  7544   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
  7547 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
  7548                        CMSCollector* collector, MemRegion span,
  7549                        CMSBitMap* bit_map,
  7550                        OopTaskQueue* work_queue,
  7551                        CMSMarkStack*  overflow_stack,
  7552                        bool should_yield):
  7553   _collector(collector),
  7554   _whole_span(collector->_span),
  7555   _span(span),
  7556   _bit_map(bit_map),
  7557   _mut(&collector->_modUnionTable),
  7558   _work_queue(work_queue),
  7559   _overflow_stack(overflow_stack),
  7560   _yield(should_yield),
  7561   _skip_bits(0),
  7562   _task(task)
  7564   assert(_work_queue->size() == 0, "work_queue should be empty");
  7565   _finger = span.start();
  7566   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
  7567   assert(_span.contains(_finger), "Out of bounds _finger?");
  7570 // Should revisit to see if this should be restructured for
  7571 // greater efficiency.
  7572 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
  7573   if (_skip_bits > 0) {
  7574     _skip_bits--;
  7575     return true;
  7577   // convert offset into a HeapWord*
  7578   HeapWord* addr = _bit_map->startWord() + offset;
  7579   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
  7580          "address out of range");
  7581   assert(_bit_map->isMarked(addr), "tautology");
  7582   if (_bit_map->isMarked(addr+1)) {
  7583     // this is an allocated object that might not yet be initialized
  7584     assert(_skip_bits == 0, "tautology");
  7585     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
  7586     oop p = oop(addr);
  7587     if (p->klass_or_null() == NULL) {
  7588       // in the case of Clean-on-Enter optimization, redirty card
  7589       // and avoid clearing card by increasing  the threshold.
  7590       return true;
  7593   scan_oops_in_oop(addr);
  7594   return true;
  7597 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
  7598   assert(_bit_map->isMarked(ptr), "expected bit to be set");
  7599   // Should we assert that our work queue is empty or
  7600   // below some drain limit?
  7601   assert(_work_queue->size() == 0,
  7602          "should drain stack to limit stack usage");
  7603   // convert ptr to an oop preparatory to scanning
  7604   oop obj = oop(ptr);
  7605   // Ignore mark word in verification below, since we
  7606   // may be running concurrent with mutators.
  7607   assert(obj->is_oop(true), "should be an oop");
  7608   assert(_finger <= ptr, "_finger runneth ahead");
  7609   // advance the finger to right end of this object
  7610   _finger = ptr + obj->size();
  7611   assert(_finger > ptr, "we just incremented it above");
  7612   // On large heaps, it may take us some time to get through
  7613   // the marking phase (especially if running iCMS). During
  7614   // this time it's possible that a lot of mutations have
  7615   // accumulated in the card table and the mod union table --
  7616   // these mutation records are redundant until we have
  7617   // actually traced into the corresponding card.
  7618   // Here, we check whether advancing the finger would make
  7619   // us cross into a new card, and if so clear corresponding
  7620   // cards in the MUT (preclean them in the card-table in the
  7621   // future).
  7623   // The clean-on-enter optimization is disabled by default,
  7624   // until we fix 6178663.
  7625   if (CMSCleanOnEnter && (_finger > _threshold)) {
  7626     // [_threshold, _finger) represents the interval
  7627     // of cards to be cleared  in MUT (or precleaned in card table).
  7628     // The set of cards to be cleared is all those that overlap
  7629     // with the interval [_threshold, _finger); note that
  7630     // _threshold is always kept card-aligned but _finger isn't
  7631     // always card-aligned.
  7632     HeapWord* old_threshold = _threshold;
  7633     assert(old_threshold == (HeapWord*)round_to(
  7634             (intptr_t)old_threshold, CardTableModRefBS::card_size),
  7635            "_threshold should always be card-aligned");
  7636     _threshold = (HeapWord*)round_to(
  7637                    (intptr_t)_finger, CardTableModRefBS::card_size);
  7638     MemRegion mr(old_threshold, _threshold);
  7639     assert(!mr.is_empty(), "Control point invariant");
  7640     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
  7641     _mut->clear_range(mr);
  7644   // Note: the local finger doesn't advance while we drain
  7645   // the stack below, but the global finger sure can and will.
  7646   HeapWord** gfa = _task->global_finger_addr();
  7647   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
  7648                                       _span, _bit_map,
  7649                                       _work_queue,
  7650                                       _overflow_stack,
  7651                                       _finger,
  7652                                       gfa, this);
  7653   bool res = _work_queue->push(obj);   // overflow could occur here
  7654   assert(res, "Will hold once we use workqueues");
  7655   while (true) {
  7656     oop new_oop;
  7657     if (!_work_queue->pop_local(new_oop)) {
  7658       // We emptied our work_queue; check if there's stuff that can
  7659       // be gotten from the overflow stack.
  7660       if (CMSConcMarkingTask::get_work_from_overflow_stack(
  7661             _overflow_stack, _work_queue)) {
  7662         do_yield_check();
  7663         continue;
  7664       } else {  // done
  7665         break;
  7668     // Skip verifying header mark word below because we are
  7669     // running concurrent with mutators.
  7670     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
  7671     // now scan this oop's oops
  7672     new_oop->oop_iterate(&pushOrMarkClosure);
  7673     do_yield_check();
  7675   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
  7678 // Yield in response to a request from VM Thread or
  7679 // from mutators.
  7680 void Par_MarkFromRootsClosure::do_yield_work() {
  7681   assert(_task != NULL, "sanity");
  7682   _task->yield();
  7685 // A variant of the above used for verifying CMS marking work.
  7686 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
  7687                         MemRegion span,
  7688                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7689                         CMSMarkStack*  mark_stack):
  7690   _collector(collector),
  7691   _span(span),
  7692   _verification_bm(verification_bm),
  7693   _cms_bm(cms_bm),
  7694   _mark_stack(mark_stack),
  7695   _pam_verify_closure(collector, span, verification_bm, cms_bm,
  7696                       mark_stack)
  7698   assert(_mark_stack->isEmpty(), "stack should be empty");
  7699   _finger = _verification_bm->startWord();
  7700   assert(_collector->_restart_addr == NULL, "Sanity check");
  7701   assert(_span.contains(_finger), "Out of bounds _finger?");
  7704 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
  7705   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
  7706   assert(_span.contains(addr), "Out of bounds _finger?");
  7707   _finger = addr;
  7710 // Should revisit to see if this should be restructured for
  7711 // greater efficiency.
  7712 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
  7713   // convert offset into a HeapWord*
  7714   HeapWord* addr = _verification_bm->startWord() + offset;
  7715   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
  7716          "address out of range");
  7717   assert(_verification_bm->isMarked(addr), "tautology");
  7718   assert(_cms_bm->isMarked(addr), "tautology");
  7720   assert(_mark_stack->isEmpty(),
  7721          "should drain stack to limit stack usage");
  7722   // convert addr to an oop preparatory to scanning
  7723   oop obj = oop(addr);
  7724   assert(obj->is_oop(), "should be an oop");
  7725   assert(_finger <= addr, "_finger runneth ahead");
  7726   // advance the finger to right end of this object
  7727   _finger = addr + obj->size();
  7728   assert(_finger > addr, "we just incremented it above");
  7729   // Note: the finger doesn't advance while we drain
  7730   // the stack below.
  7731   bool res = _mark_stack->push(obj);
  7732   assert(res, "Empty non-zero size stack should have space for single push");
  7733   while (!_mark_stack->isEmpty()) {
  7734     oop new_oop = _mark_stack->pop();
  7735     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
  7736     // now scan this oop's oops
  7737     new_oop->oop_iterate(&_pam_verify_closure);
  7739   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
  7740   return true;
  7743 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
  7744   CMSCollector* collector, MemRegion span,
  7745   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
  7746   CMSMarkStack*  mark_stack):
  7747   CMSOopClosure(collector->ref_processor()),
  7748   _collector(collector),
  7749   _span(span),
  7750   _verification_bm(verification_bm),
  7751   _cms_bm(cms_bm),
  7752   _mark_stack(mark_stack)
  7753 { }
  7755 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
  7756 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
  7758 // Upon stack overflow, we discard (part of) the stack,
  7759 // remembering the least address amongst those discarded
  7760 // in CMSCollector's _restart_address.
  7761 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
  7762   // Remember the least grey address discarded
  7763   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
  7764   _collector->lower_restart_addr(ra);
  7765   _mark_stack->reset();  // discard stack contents
  7766   _mark_stack->expand(); // expand the stack if possible
  7769 void PushAndMarkVerifyClosure::do_oop(oop obj) {
  7770   assert(obj->is_oop_or_null(), "expected an oop or NULL");
  7771   HeapWord* addr = (HeapWord*)obj;
  7772   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
  7773     // Oop lies in _span and isn't yet grey or black
  7774     _verification_bm->mark(addr);            // now grey
  7775     if (!_cms_bm->isMarked(addr)) {
  7776       oop(addr)->print();
  7777       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
  7778                              addr);
  7779       fatal("... aborting");
  7782     if (!_mark_stack->push(obj)) { // stack overflow
  7783       if (PrintCMSStatistics != 0) {
  7784         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7785                                SIZE_FORMAT, _mark_stack->capacity());
  7787       assert(_mark_stack->isFull(), "Else push should have succeeded");
  7788       handle_stack_overflow(addr);
  7790     // anything including and to the right of _finger
  7791     // will be scanned as we iterate over the remainder of the
  7792     // bit map
  7796 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
  7797                      MemRegion span,
  7798                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
  7799                      HeapWord* finger, MarkFromRootsClosure* parent) :
  7800   CMSOopClosure(collector->ref_processor()),
  7801   _collector(collector),
  7802   _span(span),
  7803   _bitMap(bitMap),
  7804   _markStack(markStack),
  7805   _finger(finger),
  7806   _parent(parent)
  7807 { }
  7809 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
  7810                      MemRegion span,
  7811                      CMSBitMap* bit_map,
  7812                      OopTaskQueue* work_queue,
  7813                      CMSMarkStack*  overflow_stack,
  7814                      HeapWord* finger,
  7815                      HeapWord** global_finger_addr,
  7816                      Par_MarkFromRootsClosure* parent) :
  7817   CMSOopClosure(collector->ref_processor()),
  7818   _collector(collector),
  7819   _whole_span(collector->_span),
  7820   _span(span),
  7821   _bit_map(bit_map),
  7822   _work_queue(work_queue),
  7823   _overflow_stack(overflow_stack),
  7824   _finger(finger),
  7825   _global_finger_addr(global_finger_addr),
  7826   _parent(parent)
  7827 { }
  7829 // Assumes thread-safe access by callers, who are
  7830 // responsible for mutual exclusion.
  7831 void CMSCollector::lower_restart_addr(HeapWord* low) {
  7832   assert(_span.contains(low), "Out of bounds addr");
  7833   if (_restart_addr == NULL) {
  7834     _restart_addr = low;
  7835   } else {
  7836     _restart_addr = MIN2(_restart_addr, low);
  7840 // Upon stack overflow, we discard (part of) the stack,
  7841 // remembering the least address amongst those discarded
  7842 // in CMSCollector's _restart_address.
  7843 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7844   // Remember the least grey address discarded
  7845   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
  7846   _collector->lower_restart_addr(ra);
  7847   _markStack->reset();  // discard stack contents
  7848   _markStack->expand(); // expand the stack if possible
  7851 // Upon stack overflow, we discard (part of) the stack,
  7852 // remembering the least address amongst those discarded
  7853 // in CMSCollector's _restart_address.
  7854 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
  7855   // We need to do this under a mutex to prevent other
  7856   // workers from interfering with the work done below.
  7857   MutexLockerEx ml(_overflow_stack->par_lock(),
  7858                    Mutex::_no_safepoint_check_flag);
  7859   // Remember the least grey address discarded
  7860   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
  7861   _collector->lower_restart_addr(ra);
  7862   _overflow_stack->reset();  // discard stack contents
  7863   _overflow_stack->expand(); // expand the stack if possible
  7866 void CMKlassClosure::do_klass(Klass* k) {
  7867   assert(_oop_closure != NULL, "Not initialized?");
  7868   k->oops_do(_oop_closure);
  7871 void PushOrMarkClosure::do_oop(oop obj) {
  7872   // Ignore mark word because we are running concurrent with mutators.
  7873   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7874   HeapWord* addr = (HeapWord*)obj;
  7875   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
  7876     // Oop lies in _span and isn't yet grey or black
  7877     _bitMap->mark(addr);            // now grey
  7878     if (addr < _finger) {
  7879       // the bit map iteration has already either passed, or
  7880       // sampled, this bit in the bit map; we'll need to
  7881       // use the marking stack to scan this oop's oops.
  7882       bool simulate_overflow = false;
  7883       NOT_PRODUCT(
  7884         if (CMSMarkStackOverflowALot &&
  7885             _collector->simulate_overflow()) {
  7886           // simulate a stack overflow
  7887           simulate_overflow = true;
  7890       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
  7891         if (PrintCMSStatistics != 0) {
  7892           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7893                                  SIZE_FORMAT, _markStack->capacity());
  7895         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
  7896         handle_stack_overflow(addr);
  7899     // anything including and to the right of _finger
  7900     // will be scanned as we iterate over the remainder of the
  7901     // bit map
  7902     do_yield_check();
  7906 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
  7907 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
  7909 void Par_PushOrMarkClosure::do_oop(oop obj) {
  7910   // Ignore mark word because we are running concurrent with mutators.
  7911   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
  7912   HeapWord* addr = (HeapWord*)obj;
  7913   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7914     // Oop lies in _span and isn't yet grey or black
  7915     // We read the global_finger (volatile read) strictly after marking oop
  7916     bool res = _bit_map->par_mark(addr);    // now grey
  7917     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
  7918     // Should we push this marked oop on our stack?
  7919     // -- if someone else marked it, nothing to do
  7920     // -- if target oop is above global finger nothing to do
  7921     // -- if target oop is in chunk and above local finger
  7922     //      then nothing to do
  7923     // -- else push on work queue
  7924     if (   !res       // someone else marked it, they will deal with it
  7925         || (addr >= *gfa)  // will be scanned in a later task
  7926         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
  7927       return;
  7929     // the bit map iteration has already either passed, or
  7930     // sampled, this bit in the bit map; we'll need to
  7931     // use the marking stack to scan this oop's oops.
  7932     bool simulate_overflow = false;
  7933     NOT_PRODUCT(
  7934       if (CMSMarkStackOverflowALot &&
  7935           _collector->simulate_overflow()) {
  7936         // simulate a stack overflow
  7937         simulate_overflow = true;
  7940     if (simulate_overflow ||
  7941         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
  7942       // stack overflow
  7943       if (PrintCMSStatistics != 0) {
  7944         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
  7945                                SIZE_FORMAT, _overflow_stack->capacity());
  7947       // We cannot assert that the overflow stack is full because
  7948       // it may have been emptied since.
  7949       assert(simulate_overflow ||
  7950              _work_queue->size() == _work_queue->max_elems(),
  7951             "Else push should have succeeded");
  7952       handle_stack_overflow(addr);
  7954     do_yield_check();
  7958 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
  7959 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
  7961 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
  7962                                        MemRegion span,
  7963                                        ReferenceProcessor* rp,
  7964                                        CMSBitMap* bit_map,
  7965                                        CMSBitMap* mod_union_table,
  7966                                        CMSMarkStack*  mark_stack,
  7967                                        bool           concurrent_precleaning):
  7968   CMSOopClosure(rp),
  7969   _collector(collector),
  7970   _span(span),
  7971   _bit_map(bit_map),
  7972   _mod_union_table(mod_union_table),
  7973   _mark_stack(mark_stack),
  7974   _concurrent_precleaning(concurrent_precleaning)
  7976   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  7979 // Grey object rescan during pre-cleaning and second checkpoint phases --
  7980 // the non-parallel version (the parallel version appears further below.)
  7981 void PushAndMarkClosure::do_oop(oop obj) {
  7982   // Ignore mark word verification. If during concurrent precleaning,
  7983   // the object monitor may be locked. If during the checkpoint
  7984   // phases, the object may already have been reached by a  different
  7985   // path and may be at the end of the global overflow list (so
  7986   // the mark word may be NULL).
  7987   assert(obj->is_oop_or_null(true /* ignore mark word */),
  7988          "expected an oop or NULL");
  7989   HeapWord* addr = (HeapWord*)obj;
  7990   // Check if oop points into the CMS generation
  7991   // and is not marked
  7992   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  7993     // a white object ...
  7994     _bit_map->mark(addr);         // ... now grey
  7995     // push on the marking stack (grey set)
  7996     bool simulate_overflow = false;
  7997     NOT_PRODUCT(
  7998       if (CMSMarkStackOverflowALot &&
  7999           _collector->simulate_overflow()) {
  8000         // simulate a stack overflow
  8001         simulate_overflow = true;
  8004     if (simulate_overflow || !_mark_stack->push(obj)) {
  8005       if (_concurrent_precleaning) {
  8006          // During precleaning we can just dirty the appropriate card(s)
  8007          // in the mod union table, thus ensuring that the object remains
  8008          // in the grey set  and continue. In the case of object arrays
  8009          // we need to dirty all of the cards that the object spans,
  8010          // since the rescan of object arrays will be limited to the
  8011          // dirty cards.
  8012          // Note that no one can be intefering with us in this action
  8013          // of dirtying the mod union table, so no locking or atomics
  8014          // are required.
  8015          if (obj->is_objArray()) {
  8016            size_t sz = obj->size();
  8017            HeapWord* end_card_addr = (HeapWord*)round_to(
  8018                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
  8019            MemRegion redirty_range = MemRegion(addr, end_card_addr);
  8020            assert(!redirty_range.is_empty(), "Arithmetical tautology");
  8021            _mod_union_table->mark_range(redirty_range);
  8022          } else {
  8023            _mod_union_table->mark(addr);
  8025          _collector->_ser_pmc_preclean_ovflw++;
  8026       } else {
  8027          // During the remark phase, we need to remember this oop
  8028          // in the overflow list.
  8029          _collector->push_on_overflow_list(obj);
  8030          _collector->_ser_pmc_remark_ovflw++;
  8036 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
  8037                                                MemRegion span,
  8038                                                ReferenceProcessor* rp,
  8039                                                CMSBitMap* bit_map,
  8040                                                OopTaskQueue* work_queue):
  8041   CMSOopClosure(rp),
  8042   _collector(collector),
  8043   _span(span),
  8044   _bit_map(bit_map),
  8045   _work_queue(work_queue)
  8047   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
  8050 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
  8051 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
  8053 // Grey object rescan during second checkpoint phase --
  8054 // the parallel version.
  8055 void Par_PushAndMarkClosure::do_oop(oop obj) {
  8056   // In the assert below, we ignore the mark word because
  8057   // this oop may point to an already visited object that is
  8058   // on the overflow stack (in which case the mark word has
  8059   // been hijacked for chaining into the overflow stack --
  8060   // if this is the last object in the overflow stack then
  8061   // its mark word will be NULL). Because this object may
  8062   // have been subsequently popped off the global overflow
  8063   // stack, and the mark word possibly restored to the prototypical
  8064   // value, by the time we get to examined this failing assert in
  8065   // the debugger, is_oop_or_null(false) may subsequently start
  8066   // to hold.
  8067   assert(obj->is_oop_or_null(true),
  8068          "expected an oop or NULL");
  8069   HeapWord* addr = (HeapWord*)obj;
  8070   // Check if oop points into the CMS generation
  8071   // and is not marked
  8072   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
  8073     // a white object ...
  8074     // If we manage to "claim" the object, by being the
  8075     // first thread to mark it, then we push it on our
  8076     // marking stack
  8077     if (_bit_map->par_mark(addr)) {     // ... now grey
  8078       // push on work queue (grey set)
  8079       bool simulate_overflow = false;
  8080       NOT_PRODUCT(
  8081         if (CMSMarkStackOverflowALot &&
  8082             _collector->par_simulate_overflow()) {
  8083           // simulate a stack overflow
  8084           simulate_overflow = true;
  8087       if (simulate_overflow || !_work_queue->push(obj)) {
  8088         _collector->par_push_on_overflow_list(obj);
  8089         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
  8091     } // Else, some other thread got there first
  8095 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
  8096 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
  8098 void CMSPrecleanRefsYieldClosure::do_yield_work() {
  8099   Mutex* bml = _collector->bitMapLock();
  8100   assert_lock_strong(bml);
  8101   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  8102          "CMS thread should hold CMS token");
  8104   bml->unlock();
  8105   ConcurrentMarkSweepThread::desynchronize(true);
  8107   ConcurrentMarkSweepThread::acknowledge_yield_request();
  8109   _collector->stopTimer();
  8110   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  8111   if (PrintCMSStatistics != 0) {
  8112     _collector->incrementYields();
  8114   _collector->icms_wait();
  8116   // See the comment in coordinator_yield()
  8117   for (unsigned i = 0; i < CMSYieldSleepCount &&
  8118                        ConcurrentMarkSweepThread::should_yield() &&
  8119                        !CMSCollector::foregroundGCIsActive(); ++i) {
  8120     os::sleep(Thread::current(), 1, false);
  8121     ConcurrentMarkSweepThread::acknowledge_yield_request();
  8124   ConcurrentMarkSweepThread::synchronize(true);
  8125   bml->lock();
  8127   _collector->startTimer();
  8130 bool CMSPrecleanRefsYieldClosure::should_return() {
  8131   if (ConcurrentMarkSweepThread::should_yield()) {
  8132     do_yield_work();
  8134   return _collector->foregroundGCIsActive();
  8137 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
  8138   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
  8139          "mr should be aligned to start at a card boundary");
  8140   // We'd like to assert:
  8141   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
  8142   //        "mr should be a range of cards");
  8143   // However, that would be too strong in one case -- the last
  8144   // partition ends at _unallocated_block which, in general, can be
  8145   // an arbitrary boundary, not necessarily card aligned.
  8146   if (PrintCMSStatistics != 0) {
  8147     _num_dirty_cards +=
  8148          mr.word_size()/CardTableModRefBS::card_size_in_words;
  8150   _space->object_iterate_mem(mr, &_scan_cl);
  8153 SweepClosure::SweepClosure(CMSCollector* collector,
  8154                            ConcurrentMarkSweepGeneration* g,
  8155                            CMSBitMap* bitMap, bool should_yield) :
  8156   _collector(collector),
  8157   _g(g),
  8158   _sp(g->cmsSpace()),
  8159   _limit(_sp->sweep_limit()),
  8160   _freelistLock(_sp->freelistLock()),
  8161   _bitMap(bitMap),
  8162   _yield(should_yield),
  8163   _inFreeRange(false),           // No free range at beginning of sweep
  8164   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
  8165   _lastFreeRangeCoalesced(false),
  8166   _freeFinger(g->used_region().start())
  8168   NOT_PRODUCT(
  8169     _numObjectsFreed = 0;
  8170     _numWordsFreed   = 0;
  8171     _numObjectsLive = 0;
  8172     _numWordsLive = 0;
  8173     _numObjectsAlreadyFree = 0;
  8174     _numWordsAlreadyFree = 0;
  8175     _last_fc = NULL;
  8177     _sp->initializeIndexedFreeListArrayReturnedBytes();
  8178     _sp->dictionary()->initialize_dict_returned_bytes();
  8180   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  8181          "sweep _limit out of bounds");
  8182   if (CMSTraceSweeper) {
  8183     gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
  8184                         _limit);
  8188 void SweepClosure::print_on(outputStream* st) const {
  8189   tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
  8190                 _sp->bottom(), _sp->end());
  8191   tty->print_cr("_limit = " PTR_FORMAT, _limit);
  8192   tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
  8193   NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
  8194   tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
  8195                 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
  8198 #ifndef PRODUCT
  8199 // Assertion checking only:  no useful work in product mode --
  8200 // however, if any of the flags below become product flags,
  8201 // you may need to review this code to see if it needs to be
  8202 // enabled in product mode.
  8203 SweepClosure::~SweepClosure() {
  8204   assert_lock_strong(_freelistLock);
  8205   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  8206          "sweep _limit out of bounds");
  8207   if (inFreeRange()) {
  8208     warning("inFreeRange() should have been reset; dumping state of SweepClosure");
  8209     print();
  8210     ShouldNotReachHere();
  8212   if (Verbose && PrintGC) {
  8213     gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
  8214                         _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
  8215     gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects,  "
  8216                            SIZE_FORMAT" bytes  "
  8217       "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
  8218       _numObjectsLive, _numWordsLive*sizeof(HeapWord),
  8219       _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
  8220     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
  8221                         * sizeof(HeapWord);
  8222     gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
  8224     if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
  8225       size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
  8226       size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
  8227       size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
  8228       gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
  8229       gclog_or_tty->print("   Indexed List Returned "SIZE_FORMAT" bytes",
  8230         indexListReturnedBytes);
  8231       gclog_or_tty->print_cr("        Dictionary Returned "SIZE_FORMAT" bytes",
  8232         dict_returned_bytes);
  8235   if (CMSTraceSweeper) {
  8236     gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
  8237                            _limit);
  8240 #endif  // PRODUCT
  8242 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
  8243     bool freeRangeInFreeLists) {
  8244   if (CMSTraceSweeper) {
  8245     gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
  8246                freeFinger, freeRangeInFreeLists);
  8248   assert(!inFreeRange(), "Trampling existing free range");
  8249   set_inFreeRange(true);
  8250   set_lastFreeRangeCoalesced(false);
  8252   set_freeFinger(freeFinger);
  8253   set_freeRangeInFreeLists(freeRangeInFreeLists);
  8254   if (CMSTestInFreeList) {
  8255     if (freeRangeInFreeLists) {
  8256       FreeChunk* fc = (FreeChunk*) freeFinger;
  8257       assert(fc->is_free(), "A chunk on the free list should be free.");
  8258       assert(fc->size() > 0, "Free range should have a size");
  8259       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
  8264 // Note that the sweeper runs concurrently with mutators. Thus,
  8265 // it is possible for direct allocation in this generation to happen
  8266 // in the middle of the sweep. Note that the sweeper also coalesces
  8267 // contiguous free blocks. Thus, unless the sweeper and the allocator
  8268 // synchronize appropriately freshly allocated blocks may get swept up.
  8269 // This is accomplished by the sweeper locking the free lists while
  8270 // it is sweeping. Thus blocks that are determined to be free are
  8271 // indeed free. There is however one additional complication:
  8272 // blocks that have been allocated since the final checkpoint and
  8273 // mark, will not have been marked and so would be treated as
  8274 // unreachable and swept up. To prevent this, the allocator marks
  8275 // the bit map when allocating during the sweep phase. This leads,
  8276 // however, to a further complication -- objects may have been allocated
  8277 // but not yet initialized -- in the sense that the header isn't yet
  8278 // installed. The sweeper can not then determine the size of the block
  8279 // in order to skip over it. To deal with this case, we use a technique
  8280 // (due to Printezis) to encode such uninitialized block sizes in the
  8281 // bit map. Since the bit map uses a bit per every HeapWord, but the
  8282 // CMS generation has a minimum object size of 3 HeapWords, it follows
  8283 // that "normal marks" won't be adjacent in the bit map (there will
  8284 // always be at least two 0 bits between successive 1 bits). We make use
  8285 // of these "unused" bits to represent uninitialized blocks -- the bit
  8286 // corresponding to the start of the uninitialized object and the next
  8287 // bit are both set. Finally, a 1 bit marks the end of the object that
  8288 // started with the two consecutive 1 bits to indicate its potentially
  8289 // uninitialized state.
  8291 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
  8292   FreeChunk* fc = (FreeChunk*)addr;
  8293   size_t res;
  8295   // Check if we are done sweeping. Below we check "addr >= _limit" rather
  8296   // than "addr == _limit" because although _limit was a block boundary when
  8297   // we started the sweep, it may no longer be one because heap expansion
  8298   // may have caused us to coalesce the block ending at the address _limit
  8299   // with a newly expanded chunk (this happens when _limit was set to the
  8300   // previous _end of the space), so we may have stepped past _limit:
  8301   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
  8302   if (addr >= _limit) { // we have swept up to or past the limit: finish up
  8303     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
  8304            "sweep _limit out of bounds");
  8305     assert(addr < _sp->end(), "addr out of bounds");
  8306     // Flush any free range we might be holding as a single
  8307     // coalesced chunk to the appropriate free list.
  8308     if (inFreeRange()) {
  8309       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
  8310              err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
  8311       flush_cur_free_chunk(freeFinger(),
  8312                            pointer_delta(addr, freeFinger()));
  8313       if (CMSTraceSweeper) {
  8314         gclog_or_tty->print("Sweep: last chunk: ");
  8315         gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
  8316                    "[coalesced:"SIZE_FORMAT"]\n",
  8317                    freeFinger(), pointer_delta(addr, freeFinger()),
  8318                    lastFreeRangeCoalesced());
  8322     // help the iterator loop finish
  8323     return pointer_delta(_sp->end(), addr);
  8326   assert(addr < _limit, "sweep invariant");
  8327   // check if we should yield
  8328   do_yield_check(addr);
  8329   if (fc->is_free()) {
  8330     // Chunk that is already free
  8331     res = fc->size();
  8332     do_already_free_chunk(fc);
  8333     debug_only(_sp->verifyFreeLists());
  8334     // If we flush the chunk at hand in lookahead_and_flush()
  8335     // and it's coalesced with a preceding chunk, then the
  8336     // process of "mangling" the payload of the coalesced block
  8337     // will cause erasure of the size information from the
  8338     // (erstwhile) header of all the coalesced blocks but the
  8339     // first, so the first disjunct in the assert will not hold
  8340     // in that specific case (in which case the second disjunct
  8341     // will hold).
  8342     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
  8343            "Otherwise the size info doesn't change at this step");
  8344     NOT_PRODUCT(
  8345       _numObjectsAlreadyFree++;
  8346       _numWordsAlreadyFree += res;
  8348     NOT_PRODUCT(_last_fc = fc;)
  8349   } else if (!_bitMap->isMarked(addr)) {
  8350     // Chunk is fresh garbage
  8351     res = do_garbage_chunk(fc);
  8352     debug_only(_sp->verifyFreeLists());
  8353     NOT_PRODUCT(
  8354       _numObjectsFreed++;
  8355       _numWordsFreed += res;
  8357   } else {
  8358     // Chunk that is alive.
  8359     res = do_live_chunk(fc);
  8360     debug_only(_sp->verifyFreeLists());
  8361     NOT_PRODUCT(
  8362         _numObjectsLive++;
  8363         _numWordsLive += res;
  8366   return res;
  8369 // For the smart allocation, record following
  8370 //  split deaths - a free chunk is removed from its free list because
  8371 //      it is being split into two or more chunks.
  8372 //  split birth - a free chunk is being added to its free list because
  8373 //      a larger free chunk has been split and resulted in this free chunk.
  8374 //  coal death - a free chunk is being removed from its free list because
  8375 //      it is being coalesced into a large free chunk.
  8376 //  coal birth - a free chunk is being added to its free list because
  8377 //      it was created when two or more free chunks where coalesced into
  8378 //      this free chunk.
  8379 //
  8380 // These statistics are used to determine the desired number of free
  8381 // chunks of a given size.  The desired number is chosen to be relative
  8382 // to the end of a CMS sweep.  The desired number at the end of a sweep
  8383 // is the
  8384 //      count-at-end-of-previous-sweep (an amount that was enough)
  8385 //              - count-at-beginning-of-current-sweep  (the excess)
  8386 //              + split-births  (gains in this size during interval)
  8387 //              - split-deaths  (demands on this size during interval)
  8388 // where the interval is from the end of one sweep to the end of the
  8389 // next.
  8390 //
  8391 // When sweeping the sweeper maintains an accumulated chunk which is
  8392 // the chunk that is made up of chunks that have been coalesced.  That
  8393 // will be termed the left-hand chunk.  A new chunk of garbage that
  8394 // is being considered for coalescing will be referred to as the
  8395 // right-hand chunk.
  8396 //
  8397 // When making a decision on whether to coalesce a right-hand chunk with
  8398 // the current left-hand chunk, the current count vs. the desired count
  8399 // of the left-hand chunk is considered.  Also if the right-hand chunk
  8400 // is near the large chunk at the end of the heap (see
  8401 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
  8402 // left-hand chunk is coalesced.
  8403 //
  8404 // When making a decision about whether to split a chunk, the desired count
  8405 // vs. the current count of the candidate to be split is also considered.
  8406 // If the candidate is underpopulated (currently fewer chunks than desired)
  8407 // a chunk of an overpopulated (currently more chunks than desired) size may
  8408 // be chosen.  The "hint" associated with a free list, if non-null, points
  8409 // to a free list which may be overpopulated.
  8410 //
  8412 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
  8413   const size_t size = fc->size();
  8414   // Chunks that cannot be coalesced are not in the
  8415   // free lists.
  8416   if (CMSTestInFreeList && !fc->cantCoalesce()) {
  8417     assert(_sp->verify_chunk_in_free_list(fc),
  8418       "free chunk should be in free lists");
  8420   // a chunk that is already free, should not have been
  8421   // marked in the bit map
  8422   HeapWord* const addr = (HeapWord*) fc;
  8423   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
  8424   // Verify that the bit map has no bits marked between
  8425   // addr and purported end of this block.
  8426   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8428   // Some chunks cannot be coalesced under any circumstances.
  8429   // See the definition of cantCoalesce().
  8430   if (!fc->cantCoalesce()) {
  8431     // This chunk can potentially be coalesced.
  8432     if (_sp->adaptive_freelists()) {
  8433       // All the work is done in
  8434       do_post_free_or_garbage_chunk(fc, size);
  8435     } else {  // Not adaptive free lists
  8436       // this is a free chunk that can potentially be coalesced by the sweeper;
  8437       if (!inFreeRange()) {
  8438         // if the next chunk is a free block that can't be coalesced
  8439         // it doesn't make sense to remove this chunk from the free lists
  8440         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
  8441         assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
  8442         if ((HeapWord*)nextChunk < _sp->end() &&     // There is another free chunk to the right ...
  8443             nextChunk->is_free()               &&     // ... which is free...
  8444             nextChunk->cantCoalesce()) {             // ... but can't be coalesced
  8445           // nothing to do
  8446         } else {
  8447           // Potentially the start of a new free range:
  8448           // Don't eagerly remove it from the free lists.
  8449           // No need to remove it if it will just be put
  8450           // back again.  (Also from a pragmatic point of view
  8451           // if it is a free block in a region that is beyond
  8452           // any allocated blocks, an assertion will fail)
  8453           // Remember the start of a free run.
  8454           initialize_free_range(addr, true);
  8455           // end - can coalesce with next chunk
  8457       } else {
  8458         // the midst of a free range, we are coalescing
  8459         print_free_block_coalesced(fc);
  8460         if (CMSTraceSweeper) {
  8461           gclog_or_tty->print("  -- pick up free block 0x%x (%d)\n", fc, size);
  8463         // remove it from the free lists
  8464         _sp->removeFreeChunkFromFreeLists(fc);
  8465         set_lastFreeRangeCoalesced(true);
  8466         // If the chunk is being coalesced and the current free range is
  8467         // in the free lists, remove the current free range so that it
  8468         // will be returned to the free lists in its entirety - all
  8469         // the coalesced pieces included.
  8470         if (freeRangeInFreeLists()) {
  8471           FreeChunk* ffc = (FreeChunk*) freeFinger();
  8472           assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8473             "Size of free range is inconsistent with chunk size.");
  8474           if (CMSTestInFreeList) {
  8475             assert(_sp->verify_chunk_in_free_list(ffc),
  8476               "free range is not in free lists");
  8478           _sp->removeFreeChunkFromFreeLists(ffc);
  8479           set_freeRangeInFreeLists(false);
  8483     // Note that if the chunk is not coalescable (the else arm
  8484     // below), we unconditionally flush, without needing to do
  8485     // a "lookahead," as we do below.
  8486     if (inFreeRange()) lookahead_and_flush(fc, size);
  8487   } else {
  8488     // Code path common to both original and adaptive free lists.
  8490     // cant coalesce with previous block; this should be treated
  8491     // as the end of a free run if any
  8492     if (inFreeRange()) {
  8493       // we kicked some butt; time to pick up the garbage
  8494       assert(freeFinger() < addr, "freeFinger points too high");
  8495       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8497     // else, nothing to do, just continue
  8501 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
  8502   // This is a chunk of garbage.  It is not in any free list.
  8503   // Add it to a free list or let it possibly be coalesced into
  8504   // a larger chunk.
  8505   HeapWord* const addr = (HeapWord*) fc;
  8506   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  8508   if (_sp->adaptive_freelists()) {
  8509     // Verify that the bit map has no bits marked between
  8510     // addr and purported end of just dead object.
  8511     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8513     do_post_free_or_garbage_chunk(fc, size);
  8514   } else {
  8515     if (!inFreeRange()) {
  8516       // start of a new free range
  8517       assert(size > 0, "A free range should have a size");
  8518       initialize_free_range(addr, false);
  8519     } else {
  8520       // this will be swept up when we hit the end of the
  8521       // free range
  8522       if (CMSTraceSweeper) {
  8523         gclog_or_tty->print("  -- pick up garbage 0x%x (%d) \n", fc, size);
  8525       // If the chunk is being coalesced and the current free range is
  8526       // in the free lists, remove the current free range so that it
  8527       // will be returned to the free lists in its entirety - all
  8528       // the coalesced pieces included.
  8529       if (freeRangeInFreeLists()) {
  8530         FreeChunk* ffc = (FreeChunk*)freeFinger();
  8531         assert(ffc->size() == pointer_delta(addr, freeFinger()),
  8532           "Size of free range is inconsistent with chunk size.");
  8533         if (CMSTestInFreeList) {
  8534           assert(_sp->verify_chunk_in_free_list(ffc),
  8535             "free range is not in free lists");
  8537         _sp->removeFreeChunkFromFreeLists(ffc);
  8538         set_freeRangeInFreeLists(false);
  8540       set_lastFreeRangeCoalesced(true);
  8542     // this will be swept up when we hit the end of the free range
  8544     // Verify that the bit map has no bits marked between
  8545     // addr and purported end of just dead object.
  8546     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
  8548   assert(_limit >= addr + size,
  8549          "A freshly garbage chunk can't possibly straddle over _limit");
  8550   if (inFreeRange()) lookahead_and_flush(fc, size);
  8551   return size;
  8554 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
  8555   HeapWord* addr = (HeapWord*) fc;
  8556   // The sweeper has just found a live object. Return any accumulated
  8557   // left hand chunk to the free lists.
  8558   if (inFreeRange()) {
  8559     assert(freeFinger() < addr, "freeFinger points too high");
  8560     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8563   // This object is live: we'd normally expect this to be
  8564   // an oop, and like to assert the following:
  8565   // assert(oop(addr)->is_oop(), "live block should be an oop");
  8566   // However, as we commented above, this may be an object whose
  8567   // header hasn't yet been initialized.
  8568   size_t size;
  8569   assert(_bitMap->isMarked(addr), "Tautology for this control point");
  8570   if (_bitMap->isMarked(addr + 1)) {
  8571     // Determine the size from the bit map, rather than trying to
  8572     // compute it from the object header.
  8573     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
  8574     size = pointer_delta(nextOneAddr + 1, addr);
  8575     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
  8576            "alignment problem");
  8578 #ifdef ASSERT
  8579       if (oop(addr)->klass_or_null() != NULL) {
  8580         // Ignore mark word because we are running concurrent with mutators
  8581         assert(oop(addr)->is_oop(true), "live block should be an oop");
  8582         assert(size ==
  8583                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
  8584                "P-mark and computed size do not agree");
  8586 #endif
  8588   } else {
  8589     // This should be an initialized object that's alive.
  8590     assert(oop(addr)->klass_or_null() != NULL,
  8591            "Should be an initialized object");
  8592     // Ignore mark word because we are running concurrent with mutators
  8593     assert(oop(addr)->is_oop(true), "live block should be an oop");
  8594     // Verify that the bit map has no bits marked between
  8595     // addr and purported end of this block.
  8596     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
  8597     assert(size >= 3, "Necessary for Printezis marks to work");
  8598     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
  8599     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
  8601   return size;
  8604 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
  8605                                                  size_t chunkSize) {
  8606   // do_post_free_or_garbage_chunk() should only be called in the case
  8607   // of the adaptive free list allocator.
  8608   const bool fcInFreeLists = fc->is_free();
  8609   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
  8610   assert((HeapWord*)fc <= _limit, "sweep invariant");
  8611   if (CMSTestInFreeList && fcInFreeLists) {
  8612     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
  8615   if (CMSTraceSweeper) {
  8616     gclog_or_tty->print_cr("  -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
  8619   HeapWord* const fc_addr = (HeapWord*) fc;
  8621   bool coalesce;
  8622   const size_t left  = pointer_delta(fc_addr, freeFinger());
  8623   const size_t right = chunkSize;
  8624   switch (FLSCoalescePolicy) {
  8625     // numeric value forms a coalition aggressiveness metric
  8626     case 0:  { // never coalesce
  8627       coalesce = false;
  8628       break;
  8630     case 1: { // coalesce if left & right chunks on overpopulated lists
  8631       coalesce = _sp->coalOverPopulated(left) &&
  8632                  _sp->coalOverPopulated(right);
  8633       break;
  8635     case 2: { // coalesce if left chunk on overpopulated list (default)
  8636       coalesce = _sp->coalOverPopulated(left);
  8637       break;
  8639     case 3: { // coalesce if left OR right chunk on overpopulated list
  8640       coalesce = _sp->coalOverPopulated(left) ||
  8641                  _sp->coalOverPopulated(right);
  8642       break;
  8644     case 4: { // always coalesce
  8645       coalesce = true;
  8646       break;
  8648     default:
  8649      ShouldNotReachHere();
  8652   // Should the current free range be coalesced?
  8653   // If the chunk is in a free range and either we decided to coalesce above
  8654   // or the chunk is near the large block at the end of the heap
  8655   // (isNearLargestChunk() returns true), then coalesce this chunk.
  8656   const bool doCoalesce = inFreeRange()
  8657                           && (coalesce || _g->isNearLargestChunk(fc_addr));
  8658   if (doCoalesce) {
  8659     // Coalesce the current free range on the left with the new
  8660     // chunk on the right.  If either is on a free list,
  8661     // it must be removed from the list and stashed in the closure.
  8662     if (freeRangeInFreeLists()) {
  8663       FreeChunk* const ffc = (FreeChunk*)freeFinger();
  8664       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
  8665         "Size of free range is inconsistent with chunk size.");
  8666       if (CMSTestInFreeList) {
  8667         assert(_sp->verify_chunk_in_free_list(ffc),
  8668           "Chunk is not in free lists");
  8670       _sp->coalDeath(ffc->size());
  8671       _sp->removeFreeChunkFromFreeLists(ffc);
  8672       set_freeRangeInFreeLists(false);
  8674     if (fcInFreeLists) {
  8675       _sp->coalDeath(chunkSize);
  8676       assert(fc->size() == chunkSize,
  8677         "The chunk has the wrong size or is not in the free lists");
  8678       _sp->removeFreeChunkFromFreeLists(fc);
  8680     set_lastFreeRangeCoalesced(true);
  8681     print_free_block_coalesced(fc);
  8682   } else {  // not in a free range and/or should not coalesce
  8683     // Return the current free range and start a new one.
  8684     if (inFreeRange()) {
  8685       // In a free range but cannot coalesce with the right hand chunk.
  8686       // Put the current free range into the free lists.
  8687       flush_cur_free_chunk(freeFinger(),
  8688                            pointer_delta(fc_addr, freeFinger()));
  8690     // Set up for new free range.  Pass along whether the right hand
  8691     // chunk is in the free lists.
  8692     initialize_free_range((HeapWord*)fc, fcInFreeLists);
  8696 // Lookahead flush:
  8697 // If we are tracking a free range, and this is the last chunk that
  8698 // we'll look at because its end crosses past _limit, we'll preemptively
  8699 // flush it along with any free range we may be holding on to. Note that
  8700 // this can be the case only for an already free or freshly garbage
  8701 // chunk. If this block is an object, it can never straddle
  8702 // over _limit. The "straddling" occurs when _limit is set at
  8703 // the previous end of the space when this cycle started, and
  8704 // a subsequent heap expansion caused the previously co-terminal
  8705 // free block to be coalesced with the newly expanded portion,
  8706 // thus rendering _limit a non-block-boundary making it dangerous
  8707 // for the sweeper to step over and examine.
  8708 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
  8709   assert(inFreeRange(), "Should only be called if currently in a free range.");
  8710   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
  8711   assert(_sp->used_region().contains(eob - 1),
  8712          err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
  8713                  " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
  8714                  " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
  8715                  eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
  8716   if (eob >= _limit) {
  8717     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
  8718     if (CMSTraceSweeper) {
  8719       gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
  8720                              "[" PTR_FORMAT "," PTR_FORMAT ") in space "
  8721                              "[" PTR_FORMAT "," PTR_FORMAT ")",
  8722                              _limit, fc, eob, _sp->bottom(), _sp->end());
  8724     // Return the storage we are tracking back into the free lists.
  8725     if (CMSTraceSweeper) {
  8726       gclog_or_tty->print_cr("Flushing ... ");
  8728     assert(freeFinger() < eob, "Error");
  8729     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
  8733 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
  8734   assert(inFreeRange(), "Should only be called if currently in a free range.");
  8735   assert(size > 0,
  8736     "A zero sized chunk cannot be added to the free lists.");
  8737   if (!freeRangeInFreeLists()) {
  8738     if (CMSTestInFreeList) {
  8739       FreeChunk* fc = (FreeChunk*) chunk;
  8740       fc->set_size(size);
  8741       assert(!_sp->verify_chunk_in_free_list(fc),
  8742         "chunk should not be in free lists yet");
  8744     if (CMSTraceSweeper) {
  8745       gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
  8746                     chunk, size);
  8748     // A new free range is going to be starting.  The current
  8749     // free range has not been added to the free lists yet or
  8750     // was removed so add it back.
  8751     // If the current free range was coalesced, then the death
  8752     // of the free range was recorded.  Record a birth now.
  8753     if (lastFreeRangeCoalesced()) {
  8754       _sp->coalBirth(size);
  8756     _sp->addChunkAndRepairOffsetTable(chunk, size,
  8757             lastFreeRangeCoalesced());
  8758   } else if (CMSTraceSweeper) {
  8759     gclog_or_tty->print_cr("Already in free list: nothing to flush");
  8761   set_inFreeRange(false);
  8762   set_freeRangeInFreeLists(false);
  8765 // We take a break if we've been at this for a while,
  8766 // so as to avoid monopolizing the locks involved.
  8767 void SweepClosure::do_yield_work(HeapWord* addr) {
  8768   // Return current free chunk being used for coalescing (if any)
  8769   // to the appropriate freelist.  After yielding, the next
  8770   // free block encountered will start a coalescing range of
  8771   // free blocks.  If the next free block is adjacent to the
  8772   // chunk just flushed, they will need to wait for the next
  8773   // sweep to be coalesced.
  8774   if (inFreeRange()) {
  8775     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
  8778   // First give up the locks, then yield, then re-lock.
  8779   // We should probably use a constructor/destructor idiom to
  8780   // do this unlock/lock or modify the MutexUnlocker class to
  8781   // serve our purpose. XXX
  8782   assert_lock_strong(_bitMap->lock());
  8783   assert_lock_strong(_freelistLock);
  8784   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
  8785          "CMS thread should hold CMS token");
  8786   _bitMap->lock()->unlock();
  8787   _freelistLock->unlock();
  8788   ConcurrentMarkSweepThread::desynchronize(true);
  8789   ConcurrentMarkSweepThread::acknowledge_yield_request();
  8790   _collector->stopTimer();
  8791   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
  8792   if (PrintCMSStatistics != 0) {
  8793     _collector->incrementYields();
  8795   _collector->icms_wait();
  8797   // See the comment in coordinator_yield()
  8798   for (unsigned i = 0; i < CMSYieldSleepCount &&
  8799                        ConcurrentMarkSweepThread::should_yield() &&
  8800                        !CMSCollector::foregroundGCIsActive(); ++i) {
  8801     os::sleep(Thread::current(), 1, false);
  8802     ConcurrentMarkSweepThread::acknowledge_yield_request();
  8805   ConcurrentMarkSweepThread::synchronize(true);
  8806   _freelistLock->lock();
  8807   _bitMap->lock()->lock_without_safepoint_check();
  8808   _collector->startTimer();
  8811 #ifndef PRODUCT
  8812 // This is actually very useful in a product build if it can
  8813 // be called from the debugger.  Compile it into the product
  8814 // as needed.
  8815 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
  8816   return debug_cms_space->verify_chunk_in_free_list(fc);
  8818 #endif
  8820 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
  8821   if (CMSTraceSweeper) {
  8822     gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
  8823                            fc, fc->size());
  8827 // CMSIsAliveClosure
  8828 bool CMSIsAliveClosure::do_object_b(oop obj) {
  8829   HeapWord* addr = (HeapWord*)obj;
  8830   return addr != NULL &&
  8831          (!_span.contains(addr) || _bit_map->isMarked(addr));
  8835 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
  8836                       MemRegion span,
  8837                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
  8838                       bool cpc):
  8839   _collector(collector),
  8840   _span(span),
  8841   _bit_map(bit_map),
  8842   _mark_stack(mark_stack),
  8843   _concurrent_precleaning(cpc) {
  8844   assert(!_span.is_empty(), "Empty span could spell trouble");
  8848 // CMSKeepAliveClosure: the serial version
  8849 void CMSKeepAliveClosure::do_oop(oop obj) {
  8850   HeapWord* addr = (HeapWord*)obj;
  8851   if (_span.contains(addr) &&
  8852       !_bit_map->isMarked(addr)) {
  8853     _bit_map->mark(addr);
  8854     bool simulate_overflow = false;
  8855     NOT_PRODUCT(
  8856       if (CMSMarkStackOverflowALot &&
  8857           _collector->simulate_overflow()) {
  8858         // simulate a stack overflow
  8859         simulate_overflow = true;
  8862     if (simulate_overflow || !_mark_stack->push(obj)) {
  8863       if (_concurrent_precleaning) {
  8864         // We dirty the overflown object and let the remark
  8865         // phase deal with it.
  8866         assert(_collector->overflow_list_is_empty(), "Error");
  8867         // In the case of object arrays, we need to dirty all of
  8868         // the cards that the object spans. No locking or atomics
  8869         // are needed since no one else can be mutating the mod union
  8870         // table.
  8871         if (obj->is_objArray()) {
  8872           size_t sz = obj->size();
  8873           HeapWord* end_card_addr =
  8874             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
  8875           MemRegion redirty_range = MemRegion(addr, end_card_addr);
  8876           assert(!redirty_range.is_empty(), "Arithmetical tautology");
  8877           _collector->_modUnionTable.mark_range(redirty_range);
  8878         } else {
  8879           _collector->_modUnionTable.mark(addr);
  8881         _collector->_ser_kac_preclean_ovflw++;
  8882       } else {
  8883         _collector->push_on_overflow_list(obj);
  8884         _collector->_ser_kac_ovflw++;
  8890 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
  8891 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
  8893 // CMSParKeepAliveClosure: a parallel version of the above.
  8894 // The work queues are private to each closure (thread),
  8895 // but (may be) available for stealing by other threads.
  8896 void CMSParKeepAliveClosure::do_oop(oop obj) {
  8897   HeapWord* addr = (HeapWord*)obj;
  8898   if (_span.contains(addr) &&
  8899       !_bit_map->isMarked(addr)) {
  8900     // In general, during recursive tracing, several threads
  8901     // may be concurrently getting here; the first one to
  8902     // "tag" it, claims it.
  8903     if (_bit_map->par_mark(addr)) {
  8904       bool res = _work_queue->push(obj);
  8905       assert(res, "Low water mark should be much less than capacity");
  8906       // Do a recursive trim in the hope that this will keep
  8907       // stack usage lower, but leave some oops for potential stealers
  8908       trim_queue(_low_water_mark);
  8909     } // Else, another thread got there first
  8913 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
  8914 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
  8916 void CMSParKeepAliveClosure::trim_queue(uint max) {
  8917   while (_work_queue->size() > max) {
  8918     oop new_oop;
  8919     if (_work_queue->pop_local(new_oop)) {
  8920       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
  8921       assert(_bit_map->isMarked((HeapWord*)new_oop),
  8922              "no white objects on this stack!");
  8923       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  8924       // iterate over the oops in this oop, marking and pushing
  8925       // the ones in CMS heap (i.e. in _span).
  8926       new_oop->oop_iterate(&_mark_and_push);
  8931 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
  8932                                 CMSCollector* collector,
  8933                                 MemRegion span, CMSBitMap* bit_map,
  8934                                 OopTaskQueue* work_queue):
  8935   _collector(collector),
  8936   _span(span),
  8937   _bit_map(bit_map),
  8938   _work_queue(work_queue) { }
  8940 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
  8941   HeapWord* addr = (HeapWord*)obj;
  8942   if (_span.contains(addr) &&
  8943       !_bit_map->isMarked(addr)) {
  8944     if (_bit_map->par_mark(addr)) {
  8945       bool simulate_overflow = false;
  8946       NOT_PRODUCT(
  8947         if (CMSMarkStackOverflowALot &&
  8948             _collector->par_simulate_overflow()) {
  8949           // simulate a stack overflow
  8950           simulate_overflow = true;
  8953       if (simulate_overflow || !_work_queue->push(obj)) {
  8954         _collector->par_push_on_overflow_list(obj);
  8955         _collector->_par_kac_ovflw++;
  8957     } // Else another thread got there already
  8961 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8962 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
  8964 //////////////////////////////////////////////////////////////////
  8965 //  CMSExpansionCause                /////////////////////////////
  8966 //////////////////////////////////////////////////////////////////
  8967 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
  8968   switch (cause) {
  8969     case _no_expansion:
  8970       return "No expansion";
  8971     case _satisfy_free_ratio:
  8972       return "Free ratio";
  8973     case _satisfy_promotion:
  8974       return "Satisfy promotion";
  8975     case _satisfy_allocation:
  8976       return "allocation";
  8977     case _allocate_par_lab:
  8978       return "Par LAB";
  8979     case _allocate_par_spooling_space:
  8980       return "Par Spooling Space";
  8981     case _adaptive_size_policy:
  8982       return "Ergonomics";
  8983     default:
  8984       return "unknown";
  8988 void CMSDrainMarkingStackClosure::do_void() {
  8989   // the max number to take from overflow list at a time
  8990   const size_t num = _mark_stack->capacity()/4;
  8991   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
  8992          "Overflow list should be NULL during concurrent phases");
  8993   while (!_mark_stack->isEmpty() ||
  8994          // if stack is empty, check the overflow list
  8995          _collector->take_from_overflow_list(num, _mark_stack)) {
  8996     oop obj = _mark_stack->pop();
  8997     HeapWord* addr = (HeapWord*)obj;
  8998     assert(_span.contains(addr), "Should be within span");
  8999     assert(_bit_map->isMarked(addr), "Should be marked");
  9000     assert(obj->is_oop(), "Should be an oop");
  9001     obj->oop_iterate(_keep_alive);
  9005 void CMSParDrainMarkingStackClosure::do_void() {
  9006   // drain queue
  9007   trim_queue(0);
  9010 // Trim our work_queue so its length is below max at return
  9011 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
  9012   while (_work_queue->size() > max) {
  9013     oop new_oop;
  9014     if (_work_queue->pop_local(new_oop)) {
  9015       assert(new_oop->is_oop(), "Expected an oop");
  9016       assert(_bit_map->isMarked((HeapWord*)new_oop),
  9017              "no white objects on this stack!");
  9018       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
  9019       // iterate over the oops in this oop, marking and pushing
  9020       // the ones in CMS heap (i.e. in _span).
  9021       new_oop->oop_iterate(&_mark_and_push);
  9026 ////////////////////////////////////////////////////////////////////
  9027 // Support for Marking Stack Overflow list handling and related code
  9028 ////////////////////////////////////////////////////////////////////
  9029 // Much of the following code is similar in shape and spirit to the
  9030 // code used in ParNewGC. We should try and share that code
  9031 // as much as possible in the future.
  9033 #ifndef PRODUCT
  9034 // Debugging support for CMSStackOverflowALot
  9036 // It's OK to call this multi-threaded;  the worst thing
  9037 // that can happen is that we'll get a bunch of closely
  9038 // spaced simulated oveflows, but that's OK, in fact
  9039 // probably good as it would exercise the overflow code
  9040 // under contention.
  9041 bool CMSCollector::simulate_overflow() {
  9042   if (_overflow_counter-- <= 0) { // just being defensive
  9043     _overflow_counter = CMSMarkStackOverflowInterval;
  9044     return true;
  9045   } else {
  9046     return false;
  9050 bool CMSCollector::par_simulate_overflow() {
  9051   return simulate_overflow();
  9053 #endif
  9055 // Single-threaded
  9056 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
  9057   assert(stack->isEmpty(), "Expected precondition");
  9058   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
  9059   size_t i = num;
  9060   oop  cur = _overflow_list;
  9061   const markOop proto = markOopDesc::prototype();
  9062   NOT_PRODUCT(ssize_t n = 0;)
  9063   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
  9064     next = oop(cur->mark());
  9065     cur->set_mark(proto);   // until proven otherwise
  9066     assert(cur->is_oop(), "Should be an oop");
  9067     bool res = stack->push(cur);
  9068     assert(res, "Bit off more than can chew?");
  9069     NOT_PRODUCT(n++;)
  9071   _overflow_list = cur;
  9072 #ifndef PRODUCT
  9073   assert(_num_par_pushes >= n, "Too many pops?");
  9074   _num_par_pushes -=n;
  9075 #endif
  9076   return !stack->isEmpty();
  9079 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
  9080 // (MT-safe) Get a prefix of at most "num" from the list.
  9081 // The overflow list is chained through the mark word of
  9082 // each object in the list. We fetch the entire list,
  9083 // break off a prefix of the right size and return the
  9084 // remainder. If other threads try to take objects from
  9085 // the overflow list at that time, they will wait for
  9086 // some time to see if data becomes available. If (and
  9087 // only if) another thread places one or more object(s)
  9088 // on the global list before we have returned the suffix
  9089 // to the global list, we will walk down our local list
  9090 // to find its end and append the global list to
  9091 // our suffix before returning it. This suffix walk can
  9092 // prove to be expensive (quadratic in the amount of traffic)
  9093 // when there are many objects in the overflow list and
  9094 // there is much producer-consumer contention on the list.
  9095 // *NOTE*: The overflow list manipulation code here and
  9096 // in ParNewGeneration:: are very similar in shape,
  9097 // except that in the ParNew case we use the old (from/eden)
  9098 // copy of the object to thread the list via its klass word.
  9099 // Because of the common code, if you make any changes in
  9100 // the code below, please check the ParNew version to see if
  9101 // similar changes might be needed.
  9102 // CR 6797058 has been filed to consolidate the common code.
  9103 bool CMSCollector::par_take_from_overflow_list(size_t num,
  9104                                                OopTaskQueue* work_q,
  9105                                                int no_of_gc_threads) {
  9106   assert(work_q->size() == 0, "First empty local work queue");
  9107   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
  9108   if (_overflow_list == NULL) {
  9109     return false;
  9111   // Grab the entire list; we'll put back a suffix
  9112   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
  9113   Thread* tid = Thread::current();
  9114   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
  9115   // set to ParallelGCThreads.
  9116   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
  9117   size_t sleep_time_millis = MAX2((size_t)1, num/100);
  9118   // If the list is busy, we spin for a short while,
  9119   // sleeping between attempts to get the list.
  9120   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
  9121     os::sleep(tid, sleep_time_millis, false);
  9122     if (_overflow_list == NULL) {
  9123       // Nothing left to take
  9124       return false;
  9125     } else if (_overflow_list != BUSY) {
  9126       // Try and grab the prefix
  9127       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
  9130   // If the list was found to be empty, or we spun long
  9131   // enough, we give up and return empty-handed. If we leave
  9132   // the list in the BUSY state below, it must be the case that
  9133   // some other thread holds the overflow list and will set it
  9134   // to a non-BUSY state in the future.
  9135   if (prefix == NULL || prefix == BUSY) {
  9136      // Nothing to take or waited long enough
  9137      if (prefix == NULL) {
  9138        // Write back the NULL in case we overwrote it with BUSY above
  9139        // and it is still the same value.
  9140        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
  9142      return false;
  9144   assert(prefix != NULL && prefix != BUSY, "Error");
  9145   size_t i = num;
  9146   oop cur = prefix;
  9147   // Walk down the first "num" objects, unless we reach the end.
  9148   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
  9149   if (cur->mark() == NULL) {
  9150     // We have "num" or fewer elements in the list, so there
  9151     // is nothing to return to the global list.
  9152     // Write back the NULL in lieu of the BUSY we wrote
  9153     // above, if it is still the same value.
  9154     if (_overflow_list == BUSY) {
  9155       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
  9157   } else {
  9158     // Chop off the suffix and rerturn it to the global list.
  9159     assert(cur->mark() != BUSY, "Error");
  9160     oop suffix_head = cur->mark(); // suffix will be put back on global list
  9161     cur->set_mark(NULL);           // break off suffix
  9162     // It's possible that the list is still in the empty(busy) state
  9163     // we left it in a short while ago; in that case we may be
  9164     // able to place back the suffix without incurring the cost
  9165     // of a walk down the list.
  9166     oop observed_overflow_list = _overflow_list;
  9167     oop cur_overflow_list = observed_overflow_list;
  9168     bool attached = false;
  9169     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
  9170       observed_overflow_list =
  9171         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
  9172       if (cur_overflow_list == observed_overflow_list) {
  9173         attached = true;
  9174         break;
  9175       } else cur_overflow_list = observed_overflow_list;
  9177     if (!attached) {
  9178       // Too bad, someone else sneaked in (at least) an element; we'll need
  9179       // to do a splice. Find tail of suffix so we can prepend suffix to global
  9180       // list.
  9181       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
  9182       oop suffix_tail = cur;
  9183       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
  9184              "Tautology");
  9185       observed_overflow_list = _overflow_list;
  9186       do {
  9187         cur_overflow_list = observed_overflow_list;
  9188         if (cur_overflow_list != BUSY) {
  9189           // Do the splice ...
  9190           suffix_tail->set_mark(markOop(cur_overflow_list));
  9191         } else { // cur_overflow_list == BUSY
  9192           suffix_tail->set_mark(NULL);
  9194         // ... and try to place spliced list back on overflow_list ...
  9195         observed_overflow_list =
  9196           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
  9197       } while (cur_overflow_list != observed_overflow_list);
  9198       // ... until we have succeeded in doing so.
  9202   // Push the prefix elements on work_q
  9203   assert(prefix != NULL, "control point invariant");
  9204   const markOop proto = markOopDesc::prototype();
  9205   oop next;
  9206   NOT_PRODUCT(ssize_t n = 0;)
  9207   for (cur = prefix; cur != NULL; cur = next) {
  9208     next = oop(cur->mark());
  9209     cur->set_mark(proto);   // until proven otherwise
  9210     assert(cur->is_oop(), "Should be an oop");
  9211     bool res = work_q->push(cur);
  9212     assert(res, "Bit off more than we can chew?");
  9213     NOT_PRODUCT(n++;)
  9215 #ifndef PRODUCT
  9216   assert(_num_par_pushes >= n, "Too many pops?");
  9217   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
  9218 #endif
  9219   return true;
  9222 // Single-threaded
  9223 void CMSCollector::push_on_overflow_list(oop p) {
  9224   NOT_PRODUCT(_num_par_pushes++;)
  9225   assert(p->is_oop(), "Not an oop");
  9226   preserve_mark_if_necessary(p);
  9227   p->set_mark((markOop)_overflow_list);
  9228   _overflow_list = p;
  9231 // Multi-threaded; use CAS to prepend to overflow list
  9232 void CMSCollector::par_push_on_overflow_list(oop p) {
  9233   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
  9234   assert(p->is_oop(), "Not an oop");
  9235   par_preserve_mark_if_necessary(p);
  9236   oop observed_overflow_list = _overflow_list;
  9237   oop cur_overflow_list;
  9238   do {
  9239     cur_overflow_list = observed_overflow_list;
  9240     if (cur_overflow_list != BUSY) {
  9241       p->set_mark(markOop(cur_overflow_list));
  9242     } else {
  9243       p->set_mark(NULL);
  9245     observed_overflow_list =
  9246       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
  9247   } while (cur_overflow_list != observed_overflow_list);
  9249 #undef BUSY
  9251 // Single threaded
  9252 // General Note on GrowableArray: pushes may silently fail
  9253 // because we are (temporarily) out of C-heap for expanding
  9254 // the stack. The problem is quite ubiquitous and affects
  9255 // a lot of code in the JVM. The prudent thing for GrowableArray
  9256 // to do (for now) is to exit with an error. However, that may
  9257 // be too draconian in some cases because the caller may be
  9258 // able to recover without much harm. For such cases, we
  9259 // should probably introduce a "soft_push" method which returns
  9260 // an indication of success or failure with the assumption that
  9261 // the caller may be able to recover from a failure; code in
  9262 // the VM can then be changed, incrementally, to deal with such
  9263 // failures where possible, thus, incrementally hardening the VM
  9264 // in such low resource situations.
  9265 void CMSCollector::preserve_mark_work(oop p, markOop m) {
  9266   _preserved_oop_stack.push(p);
  9267   _preserved_mark_stack.push(m);
  9268   assert(m == p->mark(), "Mark word changed");
  9269   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
  9270          "bijection");
  9273 // Single threaded
  9274 void CMSCollector::preserve_mark_if_necessary(oop p) {
  9275   markOop m = p->mark();
  9276   if (m->must_be_preserved(p)) {
  9277     preserve_mark_work(p, m);
  9281 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
  9282   markOop m = p->mark();
  9283   if (m->must_be_preserved(p)) {
  9284     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
  9285     // Even though we read the mark word without holding
  9286     // the lock, we are assured that it will not change
  9287     // because we "own" this oop, so no other thread can
  9288     // be trying to push it on the overflow list; see
  9289     // the assertion in preserve_mark_work() that checks
  9290     // that m == p->mark().
  9291     preserve_mark_work(p, m);
  9295 // We should be able to do this multi-threaded,
  9296 // a chunk of stack being a task (this is
  9297 // correct because each oop only ever appears
  9298 // once in the overflow list. However, it's
  9299 // not very easy to completely overlap this with
  9300 // other operations, so will generally not be done
  9301 // until all work's been completed. Because we
  9302 // expect the preserved oop stack (set) to be small,
  9303 // it's probably fine to do this single-threaded.
  9304 // We can explore cleverer concurrent/overlapped/parallel
  9305 // processing of preserved marks if we feel the
  9306 // need for this in the future. Stack overflow should
  9307 // be so rare in practice and, when it happens, its
  9308 // effect on performance so great that this will
  9309 // likely just be in the noise anyway.
  9310 void CMSCollector::restore_preserved_marks_if_any() {
  9311   assert(SafepointSynchronize::is_at_safepoint(),
  9312          "world should be stopped");
  9313   assert(Thread::current()->is_ConcurrentGC_thread() ||
  9314          Thread::current()->is_VM_thread(),
  9315          "should be single-threaded");
  9316   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
  9317          "bijection");
  9319   while (!_preserved_oop_stack.is_empty()) {
  9320     oop p = _preserved_oop_stack.pop();
  9321     assert(p->is_oop(), "Should be an oop");
  9322     assert(_span.contains(p), "oop should be in _span");
  9323     assert(p->mark() == markOopDesc::prototype(),
  9324            "Set when taken from overflow list");
  9325     markOop m = _preserved_mark_stack.pop();
  9326     p->set_mark(m);
  9328   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
  9329          "stacks were cleared above");
  9332 #ifndef PRODUCT
  9333 bool CMSCollector::no_preserved_marks() const {
  9334   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
  9336 #endif
  9338 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
  9340   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
  9341   CMSAdaptiveSizePolicy* size_policy =
  9342     (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
  9343   assert(size_policy->is_gc_cms_adaptive_size_policy(),
  9344     "Wrong type for size policy");
  9345   return size_policy;
  9348 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
  9349                                            size_t desired_promo_size) {
  9350   if (cur_promo_size < desired_promo_size) {
  9351     size_t expand_bytes = desired_promo_size - cur_promo_size;
  9352     if (PrintAdaptiveSizePolicy && Verbose) {
  9353       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  9354         "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
  9355         expand_bytes);
  9357     expand(expand_bytes,
  9358            MinHeapDeltaBytes,
  9359            CMSExpansionCause::_adaptive_size_policy);
  9360   } else if (desired_promo_size < cur_promo_size) {
  9361     size_t shrink_bytes = cur_promo_size - desired_promo_size;
  9362     if (PrintAdaptiveSizePolicy && Verbose) {
  9363       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
  9364         "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
  9365         shrink_bytes);
  9367     shrink(shrink_bytes);
  9371 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
  9372   GenCollectedHeap* gch = GenCollectedHeap::heap();
  9373   CMSGCAdaptivePolicyCounters* counters =
  9374     (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
  9375   assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
  9376     "Wrong kind of counters");
  9377   return counters;
  9381 void ASConcurrentMarkSweepGeneration::update_counters() {
  9382   if (UsePerfData) {
  9383     _space_counters->update_all();
  9384     _gen_counters->update_all();
  9385     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  9386     GenCollectedHeap* gch = GenCollectedHeap::heap();
  9387     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  9388     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  9389       "Wrong gc statistics type");
  9390     counters->update_counters(gc_stats_l);
  9394 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
  9395   if (UsePerfData) {
  9396     _space_counters->update_used(used);
  9397     _space_counters->update_capacity();
  9398     _gen_counters->update_all();
  9400     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
  9401     GenCollectedHeap* gch = GenCollectedHeap::heap();
  9402     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
  9403     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
  9404       "Wrong gc statistics type");
  9405     counters->update_counters(gc_stats_l);
  9409 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
  9410   assert_locked_or_safepoint(Heap_lock);
  9411   assert_lock_strong(freelistLock());
  9412   HeapWord* old_end = _cmsSpace->end();
  9413   HeapWord* unallocated_start = _cmsSpace->unallocated_block();
  9414   assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
  9415   FreeChunk* chunk_at_end = find_chunk_at_end();
  9416   if (chunk_at_end == NULL) {
  9417     // No room to shrink
  9418     if (PrintGCDetails && Verbose) {
  9419       gclog_or_tty->print_cr("No room to shrink: old_end  "
  9420         PTR_FORMAT "  unallocated_start  " PTR_FORMAT
  9421         " chunk_at_end  " PTR_FORMAT,
  9422         old_end, unallocated_start, chunk_at_end);
  9424     return;
  9425   } else {
  9427     // Find the chunk at the end of the space and determine
  9428     // how much it can be shrunk.
  9429     size_t shrinkable_size_in_bytes = chunk_at_end->size();
  9430     size_t aligned_shrinkable_size_in_bytes =
  9431       align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
  9432     assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
  9433       "Inconsistent chunk at end of space");
  9434     size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
  9435     size_t word_size_before = heap_word_size(_virtual_space.committed_size());
  9437     // Shrink the underlying space
  9438     _virtual_space.shrink_by(bytes);
  9439     if (PrintGCDetails && Verbose) {
  9440       gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
  9441         " desired_bytes " SIZE_FORMAT
  9442         " shrinkable_size_in_bytes " SIZE_FORMAT
  9443         " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
  9444         "  bytes  " SIZE_FORMAT,
  9445         desired_bytes, shrinkable_size_in_bytes,
  9446         aligned_shrinkable_size_in_bytes, bytes);
  9447       gclog_or_tty->print_cr("          old_end  " SIZE_FORMAT
  9448         "  unallocated_start  " SIZE_FORMAT,
  9449         old_end, unallocated_start);
  9452     // If the space did shrink (shrinking is not guaranteed),
  9453     // shrink the chunk at the end by the appropriate amount.
  9454     if (((HeapWord*)_virtual_space.high()) < old_end) {
  9455       size_t new_word_size =
  9456         heap_word_size(_virtual_space.committed_size());
  9458       // Have to remove the chunk from the dictionary because it is changing
  9459       // size and might be someplace elsewhere in the dictionary.
  9461       // Get the chunk at end, shrink it, and put it
  9462       // back.
  9463       _cmsSpace->removeChunkFromDictionary(chunk_at_end);
  9464       size_t word_size_change = word_size_before - new_word_size;
  9465       size_t chunk_at_end_old_size = chunk_at_end->size();
  9466       assert(chunk_at_end_old_size >= word_size_change,
  9467         "Shrink is too large");
  9468       chunk_at_end->set_size(chunk_at_end_old_size -
  9469                           word_size_change);
  9470       _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
  9471         word_size_change);
  9473       _cmsSpace->returnChunkToDictionary(chunk_at_end);
  9475       MemRegion mr(_cmsSpace->bottom(), new_word_size);
  9476       _bts->resize(new_word_size);  // resize the block offset shared array
  9477       Universe::heap()->barrier_set()->resize_covered_region(mr);
  9478       _cmsSpace->assert_locked();
  9479       _cmsSpace->set_end((HeapWord*)_virtual_space.high());
  9481       NOT_PRODUCT(_cmsSpace->dictionary()->verify());
  9483       // update the space and generation capacity counters
  9484       if (UsePerfData) {
  9485         _space_counters->update_capacity();
  9486         _gen_counters->update_all();
  9489       if (Verbose && PrintGCDetails) {
  9490         size_t new_mem_size = _virtual_space.committed_size();
  9491         size_t old_mem_size = new_mem_size + bytes;
  9492         gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
  9493                       name(), old_mem_size/K, bytes/K, new_mem_size/K);
  9497     assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
  9498       "Inconsistency at end of space");
  9499     assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
  9500       "Shrinking is inconsistent");
  9501     return;
  9504 // Transfer some number of overflown objects to usual marking
  9505 // stack. Return true if some objects were transferred.
  9506 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
  9507   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
  9508                     (size_t)ParGCDesiredObjsFromOverflowList);
  9510   bool res = _collector->take_from_overflow_list(num, _mark_stack);
  9511   assert(_collector->overflow_list_is_empty() || res,
  9512          "If list is not empty, we should have taken something");
  9513   assert(!res || !_mark_stack->isEmpty(),
  9514          "If we took something, it should now be on our stack");
  9515   return res;
  9518 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
  9519   size_t res = _sp->block_size_no_stall(addr, _collector);
  9520   if (_sp->block_is_obj(addr)) {
  9521     if (_live_bit_map->isMarked(addr)) {
  9522       // It can't have been dead in a previous cycle
  9523       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
  9524     } else {
  9525       _dead_bit_map->mark(addr);      // mark the dead object
  9528   // Could be 0, if the block size could not be computed without stalling.
  9529   return res;
  9532 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
  9534   switch (phase) {
  9535     case CMSCollector::InitialMarking:
  9536       initialize(true  /* fullGC */ ,
  9537                  cause /* cause of the GC */,
  9538                  true  /* recordGCBeginTime */,
  9539                  true  /* recordPreGCUsage */,
  9540                  false /* recordPeakUsage */,
  9541                  false /* recordPostGCusage */,
  9542                  true  /* recordAccumulatedGCTime */,
  9543                  false /* recordGCEndTime */,
  9544                  false /* countCollection */  );
  9545       break;
  9547     case CMSCollector::FinalMarking:
  9548       initialize(true  /* fullGC */ ,
  9549                  cause /* cause of the GC */,
  9550                  false /* recordGCBeginTime */,
  9551                  false /* recordPreGCUsage */,
  9552                  false /* recordPeakUsage */,
  9553                  false /* recordPostGCusage */,
  9554                  true  /* recordAccumulatedGCTime */,
  9555                  false /* recordGCEndTime */,
  9556                  false /* countCollection */  );
  9557       break;
  9559     case CMSCollector::Sweeping:
  9560       initialize(true  /* fullGC */ ,
  9561                  cause /* cause of the GC */,
  9562                  false /* recordGCBeginTime */,
  9563                  false /* recordPreGCUsage */,
  9564                  true  /* recordPeakUsage */,
  9565                  true  /* recordPostGCusage */,
  9566                  false /* recordAccumulatedGCTime */,
  9567                  true  /* recordGCEndTime */,
  9568                  true  /* countCollection */  );
  9569       break;
  9571     default:
  9572       ShouldNotReachHere();

mercurial