src/share/vm/gc_implementation/parNew/parNewGeneration.hpp

Mon, 26 Jan 2009 12:47:21 -0800

author
ysr
date
Mon, 26 Jan 2009 12:47:21 -0800
changeset 969
5cfd8d19e546
parent 631
d1605aabd0a1
child 1014
0fbdb4381b99
permissions
-rw-r--r--

6786503: Overflow list performance can be improved
Summary: Avoid overflow list walk in CMS & ParNew when it is unnecessary. Fix a couple of correctness issues, including a C-heap leak, in ParNew at the intersection of promotion failure, work queue overflow and object array chunking. Add stress testing option and related assertion checking.
Reviewed-by: jmasa

     1 /*
     2  * Copyright 2001-2008 Sun Microsystems, Inc.  All Rights Reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
    20  * CA 95054 USA or visit www.sun.com if you need additional information or
    21  * have any questions.
    22  *
    23  */
    25 class ChunkArray;
    26 class ParScanWithoutBarrierClosure;
    27 class ParScanWithBarrierClosure;
    28 class ParRootScanWithoutBarrierClosure;
    29 class ParRootScanWithBarrierTwoGensClosure;
    30 class ParEvacuateFollowersClosure;
    32 // It would be better if these types could be kept local to the .cpp file,
    33 // but they must be here to allow ParScanClosure::do_oop_work to be defined
    34 // in genOopClosures.inline.hpp.
    36 typedef OopTaskQueue    ObjToScanQueue;
    37 typedef OopTaskQueueSet ObjToScanQueueSet;
    39 // Enable this to get push/pop/steal stats.
    40 const int PAR_STATS_ENABLED = 0;
    42 class ParKeepAliveClosure: public DefNewGeneration::KeepAliveClosure {
    43  private:
    44   ParScanWeakRefClosure* _par_cl;
    45  protected:
    46   template <class T> void do_oop_work(T* p);
    47  public:
    48   ParKeepAliveClosure(ParScanWeakRefClosure* cl);
    49   virtual void do_oop(oop* p);
    50   virtual void do_oop(narrowOop* p);
    51 };
    53 // The state needed by thread performing parallel young-gen collection.
    54 class ParScanThreadState {
    55   friend class ParScanThreadStateSet;
    56  private:
    57   ObjToScanQueue *_work_queue;
    59   ParGCAllocBuffer _to_space_alloc_buffer;
    61   ParScanWithoutBarrierClosure         _to_space_closure; // scan_without_gc_barrier
    62   ParScanWithBarrierClosure            _old_gen_closure; // scan_with_gc_barrier
    63   ParRootScanWithoutBarrierClosure     _to_space_root_closure; // scan_root_without_gc_barrier
    64   // One of these two will be passed to process_strong_roots, which will
    65   // set its generation.  The first is for two-gen configs where the
    66   // old gen collects the perm gen; the second is for arbitrary configs.
    67   // The second isn't used right now (it used to be used for the train, an
    68   // incremental collector) but the declaration has been left as a reminder.
    69   ParRootScanWithBarrierTwoGensClosure _older_gen_closure;
    70   // This closure will always be bound to the old gen; it will be used
    71   // in evacuate_followers.
    72   ParRootScanWithBarrierTwoGensClosure _old_gen_root_closure; // scan_old_root_with_gc_barrier
    73   ParEvacuateFollowersClosure          _evacuate_followers;
    74   DefNewGeneration::IsAliveClosure     _is_alive_closure;
    75   ParScanWeakRefClosure                _scan_weak_ref_closure;
    76   ParKeepAliveClosure                  _keep_alive_closure;
    79   Space* _to_space;
    80   Space* to_space() { return _to_space; }
    82   Generation* _old_gen;
    83   Generation* old_gen() { return _old_gen; }
    85   HeapWord *_young_old_boundary;
    87   int _hash_seed;
    88   int _thread_num;
    89   ageTable _ageTable;
    91   bool _to_space_full;
    93   int _pushes, _pops, _steals, _steal_attempts, _term_attempts;
    94   int _overflow_pushes, _overflow_refills, _overflow_refill_objs;
    96   // Timing numbers.
    97   double _start;
    98   double _start_strong_roots;
    99   double _strong_roots_time;
   100   double _start_term;
   101   double _term_time;
   103   // Helper for trim_queues. Scans subset of an array and makes
   104   // remainder available for work stealing.
   105   void scan_partial_array_and_push_remainder(oop obj);
   107   // In support of CMS' parallel rescan of survivor space.
   108   ChunkArray* _survivor_chunk_array;
   109   ChunkArray* survivor_chunk_array() { return _survivor_chunk_array; }
   111   void record_survivor_plab(HeapWord* plab_start, size_t plab_word_size);
   113   ParScanThreadState(Space* to_space_, ParNewGeneration* gen_,
   114                      Generation* old_gen_, int thread_num_,
   115                      ObjToScanQueueSet* work_queue_set_, size_t desired_plab_sz_,
   116                      ParallelTaskTerminator& term_);
   118  public:
   119   ageTable* age_table() {return &_ageTable;}
   121   ObjToScanQueue* work_queue() { return _work_queue; }
   123   ParGCAllocBuffer* to_space_alloc_buffer() {
   124     return &_to_space_alloc_buffer;
   125   }
   127   ParEvacuateFollowersClosure&      evacuate_followers_closure() { return _evacuate_followers; }
   128   DefNewGeneration::IsAliveClosure& is_alive_closure() { return _is_alive_closure; }
   129   ParScanWeakRefClosure&            scan_weak_ref_closure() { return _scan_weak_ref_closure; }
   130   ParKeepAliveClosure&              keep_alive_closure() { return _keep_alive_closure; }
   131   ParScanClosure&                   older_gen_closure() { return _older_gen_closure; }
   132   ParRootScanWithoutBarrierClosure& to_space_root_closure() { return _to_space_root_closure; };
   134   // Decrease queue size below "max_size".
   135   void trim_queues(int max_size);
   137   // Is new_obj a candidate for scan_partial_array_and_push_remainder method.
   138   inline bool should_be_partially_scanned(oop new_obj, oop old_obj) const;
   140   int* hash_seed()  { return &_hash_seed; }
   141   int  thread_num() { return _thread_num; }
   143   // Allocate a to-space block of size "sz", or else return NULL.
   144   HeapWord* alloc_in_to_space_slow(size_t word_sz);
   146   HeapWord* alloc_in_to_space(size_t word_sz) {
   147     HeapWord* obj = to_space_alloc_buffer()->allocate(word_sz);
   148     if (obj != NULL) return obj;
   149     else return alloc_in_to_space_slow(word_sz);
   150   }
   152   HeapWord* young_old_boundary() { return _young_old_boundary; }
   154   void set_young_old_boundary(HeapWord *boundary) {
   155     _young_old_boundary = boundary;
   156   }
   158   // Undo the most recent allocation ("obj", of "word_sz").
   159   void undo_alloc_in_to_space(HeapWord* obj, size_t word_sz);
   161   int pushes() { return _pushes; }
   162   int pops()   { return _pops; }
   163   int steals() { return _steals; }
   164   int steal_attempts() { return _steal_attempts; }
   165   int term_attempts()  { return _term_attempts; }
   166   int overflow_pushes() { return _overflow_pushes; }
   167   int overflow_refills() { return _overflow_refills; }
   168   int overflow_refill_objs() { return _overflow_refill_objs; }
   170   void note_push()  { if (PAR_STATS_ENABLED) _pushes++; }
   171   void note_pop()   { if (PAR_STATS_ENABLED) _pops++; }
   172   void note_steal() { if (PAR_STATS_ENABLED) _steals++; }
   173   void note_steal_attempt() { if (PAR_STATS_ENABLED) _steal_attempts++; }
   174   void note_term_attempt()  { if (PAR_STATS_ENABLED) _term_attempts++; }
   175   void note_overflow_push() { if (PAR_STATS_ENABLED) _overflow_pushes++; }
   176   void note_overflow_refill(int objs) {
   177     if (PAR_STATS_ENABLED) {
   178       _overflow_refills++;
   179       _overflow_refill_objs += objs;
   180     }
   181   }
   183   void start_strong_roots() {
   184     _start_strong_roots = os::elapsedTime();
   185   }
   186   void end_strong_roots() {
   187     _strong_roots_time += (os::elapsedTime() - _start_strong_roots);
   188   }
   189   double strong_roots_time() { return _strong_roots_time; }
   190   void start_term_time() {
   191     note_term_attempt();
   192     _start_term = os::elapsedTime();
   193   }
   194   void end_term_time() {
   195     _term_time += (os::elapsedTime() - _start_term);
   196   }
   197   double term_time() { return _term_time; }
   199   double elapsed() {
   200     return os::elapsedTime() - _start;
   201   }
   202 };
   204 class ParNewGenTask: public AbstractGangTask {
   205  private:
   206   ParNewGeneration*            _gen;
   207   Generation*                  _next_gen;
   208   HeapWord*                    _young_old_boundary;
   209   class ParScanThreadStateSet* _state_set;
   211 public:
   212   ParNewGenTask(ParNewGeneration*      gen,
   213                 Generation*            next_gen,
   214                 HeapWord*              young_old_boundary,
   215                 ParScanThreadStateSet* state_set);
   217   HeapWord* young_old_boundary() { return _young_old_boundary; }
   219   void work(int i);
   220 };
   222 class KeepAliveClosure: public DefNewGeneration::KeepAliveClosure {
   223  protected:
   224   template <class T> void do_oop_work(T* p);
   225  public:
   226   KeepAliveClosure(ScanWeakRefClosure* cl);
   227   virtual void do_oop(oop* p);
   228   virtual void do_oop(narrowOop* p);
   229 };
   231 class EvacuateFollowersClosureGeneral: public VoidClosure {
   232  private:
   233   GenCollectedHeap* _gch;
   234   int               _level;
   235   OopsInGenClosure* _scan_cur_or_nonheap;
   236   OopsInGenClosure* _scan_older;
   237  public:
   238   EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
   239                                   OopsInGenClosure* cur,
   240                                   OopsInGenClosure* older);
   241   virtual void do_void();
   242 };
   244 // Closure for scanning ParNewGeneration.
   245 // Same as ScanClosure, except does parallel GC barrier.
   246 class ScanClosureWithParBarrier: public ScanClosure {
   247  protected:
   248   template <class T> void do_oop_work(T* p);
   249  public:
   250   ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier);
   251   virtual void do_oop(oop* p);
   252   virtual void do_oop(narrowOop* p);
   253 };
   255 // Implements AbstractRefProcTaskExecutor for ParNew.
   256 class ParNewRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
   257  private:
   258   ParNewGeneration&      _generation;
   259   ParScanThreadStateSet& _state_set;
   260  public:
   261   ParNewRefProcTaskExecutor(ParNewGeneration& generation,
   262                             ParScanThreadStateSet& state_set)
   263     : _generation(generation), _state_set(state_set)
   264   { }
   266   // Executes a task using worker threads.
   267   virtual void execute(ProcessTask& task);
   268   virtual void execute(EnqueueTask& task);
   269   // Switch to single threaded mode.
   270   virtual void set_single_threaded_mode();
   271 };
   274 // A Generation that does parallel young-gen collection.
   276 class ParNewGeneration: public DefNewGeneration {
   277   friend class ParNewGenTask;
   278   friend class ParNewRefProcTask;
   279   friend class ParNewRefProcTaskExecutor;
   280   friend class ParScanThreadStateSet;
   281   friend class ParEvacuateFollowersClosure;
   283  private:
   284   // XXX use a global constant instead of 64!
   285   struct ObjToScanQueuePadded {
   286         ObjToScanQueue work_queue;
   287         char pad[64 - sizeof(ObjToScanQueue)];  // prevent false sharing
   288   };
   290   // The per-thread work queues, available here for stealing.
   291   ObjToScanQueueSet* _task_queues;
   293   // Desired size of survivor space plab's
   294   PLABStats _plab_stats;
   296   // A list of from-space images of to-be-scanned objects, threaded through
   297   // klass-pointers (klass information already copied to the forwarded
   298   // image.)  Manipulated with CAS.
   299   oop _overflow_list;
   300   NOT_PRODUCT(ssize_t _num_par_pushes;)
   302   // If true, older generation does not support promotion undo, so avoid.
   303   static bool _avoid_promotion_undo;
   305   // This closure is used by the reference processor to filter out
   306   // references to live referent.
   307   DefNewGeneration::IsAliveClosure _is_alive_closure;
   309   static oop real_forwardee_slow(oop obj);
   310   static void waste_some_time();
   312   // Preserve the mark of "obj", if necessary, in preparation for its mark
   313   // word being overwritten with a self-forwarding-pointer.
   314   void preserve_mark_if_necessary(oop obj, markOop m);
   316  protected:
   318   bool _survivor_overflow;
   320   bool avoid_promotion_undo() { return _avoid_promotion_undo; }
   321   void set_avoid_promotion_undo(bool v) { _avoid_promotion_undo = v; }
   323   bool survivor_overflow() { return _survivor_overflow; }
   324   void set_survivor_overflow(bool v) { _survivor_overflow = v; }
   326   // Adjust the tenuring threshold.  See the implementation for
   327   // the details of the policy.
   328   virtual void adjust_desired_tenuring_threshold();
   330  public:
   331   ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level);
   333   ~ParNewGeneration() {
   334     for (uint i = 0; i < ParallelGCThreads; i++)
   335         delete _task_queues->queue(i);
   337     delete _task_queues;
   338   }
   340   virtual void ref_processor_init();
   341   virtual Generation::Name kind()        { return Generation::ParNew; }
   342   virtual const char* name() const;
   343   virtual const char* short_name() const { return "ParNew"; }
   345   // override
   346   virtual bool refs_discovery_is_mt()     const {
   347     assert(UseParNewGC, "ParNewGeneration only when UseParNewGC");
   348     return ParallelGCThreads > 1;
   349   }
   351   // Make the collection virtual.
   352   virtual void collect(bool   full,
   353                        bool   clear_all_soft_refs,
   354                        size_t size,
   355                        bool   is_tlab);
   357   // This needs to be visible to the closure function.
   358   // "obj" is the object to be copied, "m" is a recent value of its mark
   359   // that must not contain a forwarding pointer (though one might be
   360   // inserted in "obj"s mark word by a parallel thread).
   361   inline oop copy_to_survivor_space(ParScanThreadState* par_scan_state,
   362                              oop obj, size_t obj_sz, markOop m) {
   363     if (_avoid_promotion_undo) {
   364        return copy_to_survivor_space_avoiding_promotion_undo(par_scan_state,
   365                                                              obj, obj_sz, m);
   366     }
   368     return copy_to_survivor_space_with_undo(par_scan_state, obj, obj_sz, m);
   369   }
   371   oop copy_to_survivor_space_avoiding_promotion_undo(ParScanThreadState* par_scan_state,
   372                              oop obj, size_t obj_sz, markOop m);
   374   oop copy_to_survivor_space_with_undo(ParScanThreadState* par_scan_state,
   375                              oop obj, size_t obj_sz, markOop m);
   377   // in support of testing overflow code
   378   NOT_PRODUCT(int _overflow_counter;)
   379   NOT_PRODUCT(bool should_simulate_overflow();)
   381   // Push the given (from-space) object on the global overflow list.
   382   void push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state);
   384   // If the global overflow list is non-empty, move some tasks from it
   385   // onto "work_q" (which must be empty).  No more than 1/4 of the
   386   // max_elems of "work_q" are moved.
   387   bool take_from_overflow_list(ParScanThreadState* par_scan_state);
   389   // The task queues to be used by parallel GC threads.
   390   ObjToScanQueueSet* task_queues() {
   391     return _task_queues;
   392   }
   394   PLABStats* plab_stats() {
   395     return &_plab_stats;
   396   }
   398   size_t desired_plab_sz() {
   399     return _plab_stats.desired_plab_sz();
   400   }
   402   static oop real_forwardee(oop obj);
   404   DEBUG_ONLY(static bool is_legal_forward_ptr(oop p);)
   405 };

mercurial