src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.cpp

Thu, 22 Jul 2010 10:27:41 -0400

author
tonyp
date
Thu, 22 Jul 2010 10:27:41 -0400
changeset 2061
9d7a8ab3736b
parent 2020
a93a9eda13f7
child 2191
894b1d7c7e01
permissions
-rw-r--r--

6962589: remove breadth first scanning code from parallel gc
Summary: Remove the breadth-first copying order from ParallelScavenge and use depth-first by default.
Reviewed-by: jcoomes, ysr, johnc

     1 /*
     2  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    25 #include "incls/_precompiled.incl"
    26 #include "incls/_psPromotionManager.cpp.incl"
    28 PSPromotionManager**         PSPromotionManager::_manager_array = NULL;
    29 OopStarTaskQueueSet*         PSPromotionManager::_stack_array_depth = NULL;
    30 PSOldGen*                    PSPromotionManager::_old_gen = NULL;
    31 MutableSpace*                PSPromotionManager::_young_space = NULL;
    33 void PSPromotionManager::initialize() {
    34   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
    35   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
    37   _old_gen = heap->old_gen();
    38   _young_space = heap->young_gen()->to_space();
    40   assert(_manager_array == NULL, "Attempt to initialize twice");
    41   _manager_array = NEW_C_HEAP_ARRAY(PSPromotionManager*, ParallelGCThreads+1 );
    42   guarantee(_manager_array != NULL, "Could not initialize promotion manager");
    44   _stack_array_depth = new OopStarTaskQueueSet(ParallelGCThreads);
    45   guarantee(_stack_array_depth != NULL, "Cound not initialize promotion manager");
    47   // Create and register the PSPromotionManager(s) for the worker threads.
    48   for(uint i=0; i<ParallelGCThreads; i++) {
    49     _manager_array[i] = new PSPromotionManager();
    50     guarantee(_manager_array[i] != NULL, "Could not create PSPromotionManager");
    51     stack_array_depth()->register_queue(i, _manager_array[i]->claimed_stack_depth());
    52   }
    54   // The VMThread gets its own PSPromotionManager, which is not available
    55   // for work stealing.
    56   _manager_array[ParallelGCThreads] = new PSPromotionManager();
    57   guarantee(_manager_array[ParallelGCThreads] != NULL, "Could not create PSPromotionManager");
    58 }
    60 PSPromotionManager* PSPromotionManager::gc_thread_promotion_manager(int index) {
    61   assert(index >= 0 && index < (int)ParallelGCThreads, "index out of range");
    62   assert(_manager_array != NULL, "Sanity");
    63   return _manager_array[index];
    64 }
    66 PSPromotionManager* PSPromotionManager::vm_thread_promotion_manager() {
    67   assert(_manager_array != NULL, "Sanity");
    68   return _manager_array[ParallelGCThreads];
    69 }
    71 void PSPromotionManager::pre_scavenge() {
    72   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
    73   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
    75   _young_space = heap->young_gen()->to_space();
    77   for(uint i=0; i<ParallelGCThreads+1; i++) {
    78     manager_array(i)->reset();
    79   }
    80 }
    82 void PSPromotionManager::post_scavenge() {
    83   TASKQUEUE_STATS_ONLY(if (PrintGCDetails && ParallelGCVerbose) print_stats());
    84   for (uint i = 0; i < ParallelGCThreads + 1; i++) {
    85     PSPromotionManager* manager = manager_array(i);
    86     assert(manager->claimed_stack_depth()->is_empty(), "should be empty");
    87     manager->flush_labs();
    88   }
    89 }
    91 #if TASKQUEUE_STATS
    92 void
    93 PSPromotionManager::print_taskqueue_stats(uint i) const {
    94   tty->print("%3u ", i);
    95   _claimed_stack_depth.stats.print();
    96   tty->cr();
    97 }
    99 void
   100 PSPromotionManager::print_local_stats(uint i) const {
   101   #define FMT " " SIZE_FORMAT_W(10)
   102   tty->print_cr("%3u" FMT FMT FMT FMT, i, _masked_pushes, _masked_steals,
   103                 _arrays_chunked, _array_chunks_processed);
   104   #undef FMT
   105 }
   107 static const char* const pm_stats_hdr[] = {
   108   "    --------masked-------     arrays      array",
   109   "thr       push      steal    chunked     chunks",
   110   "--- ---------- ---------- ---------- ----------"
   111 };
   113 void
   114 PSPromotionManager::print_stats() {
   115   tty->print_cr("== GC Tasks Stats, GC %3d",
   116                 Universe::heap()->total_collections());
   118   tty->print("thr "); TaskQueueStats::print_header(1); tty->cr();
   119   tty->print("--- "); TaskQueueStats::print_header(2); tty->cr();
   120   for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
   121     manager_array(i)->print_taskqueue_stats(i);
   122   }
   124   const uint hlines = sizeof(pm_stats_hdr) / sizeof(pm_stats_hdr[0]);
   125   for (uint i = 0; i < hlines; ++i) tty->print_cr(pm_stats_hdr[i]);
   126   for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
   127     manager_array(i)->print_local_stats(i);
   128   }
   129 }
   131 void
   132 PSPromotionManager::reset_stats() {
   133   claimed_stack_depth()->stats.reset();
   134   _masked_pushes = _masked_steals = 0;
   135   _arrays_chunked = _array_chunks_processed = 0;
   136 }
   137 #endif // TASKQUEUE_STATS
   139 PSPromotionManager::PSPromotionManager() {
   140   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
   141   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
   143   // We set the old lab's start array.
   144   _old_lab.set_start_array(old_gen()->start_array());
   146   uint queue_size;
   147   claimed_stack_depth()->initialize();
   148   queue_size = claimed_stack_depth()->max_elems();
   150   _totally_drain = (ParallelGCThreads == 1) || (GCDrainStackTargetSize == 0);
   151   if (_totally_drain) {
   152     _target_stack_size = 0;
   153   } else {
   154     // don't let the target stack size to be more than 1/4 of the entries
   155     _target_stack_size = (uint) MIN2((uint) GCDrainStackTargetSize,
   156                                      (uint) (queue_size / 4));
   157   }
   159   _array_chunk_size = ParGCArrayScanChunk;
   160   // let's choose 1.5x the chunk size
   161   _min_array_size_for_chunking = 3 * _array_chunk_size / 2;
   163   reset();
   164 }
   166 void PSPromotionManager::reset() {
   167   assert(stacks_empty(), "reset of non-empty stack");
   169   // We need to get an assert in here to make sure the labs are always flushed.
   171   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
   172   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
   174   // Do not prefill the LAB's, save heap wastage!
   175   HeapWord* lab_base = young_space()->top();
   176   _young_lab.initialize(MemRegion(lab_base, (size_t)0));
   177   _young_gen_is_full = false;
   179   lab_base = old_gen()->object_space()->top();
   180   _old_lab.initialize(MemRegion(lab_base, (size_t)0));
   181   _old_gen_is_full = false;
   183   TASKQUEUE_STATS_ONLY(reset_stats());
   184 }
   187 void PSPromotionManager::drain_stacks_depth(bool totally_drain) {
   188   assert(claimed_stack_depth()->overflow_stack() != NULL, "invariant");
   189   totally_drain = totally_drain || _totally_drain;
   191 #ifdef ASSERT
   192   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
   193   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
   194   MutableSpace* to_space = heap->young_gen()->to_space();
   195   MutableSpace* old_space = heap->old_gen()->object_space();
   196   MutableSpace* perm_space = heap->perm_gen()->object_space();
   197 #endif /* ASSERT */
   199   OopStarTaskQueue* const tq = claimed_stack_depth();
   200   do {
   201     StarTask p;
   203     // Drain overflow stack first, so other threads can steal from
   204     // claimed stack while we work.
   205     while (tq->pop_overflow(p)) {
   206       process_popped_location_depth(p);
   207     }
   209     if (totally_drain) {
   210       while (tq->pop_local(p)) {
   211         process_popped_location_depth(p);
   212       }
   213     } else {
   214       while (tq->size() > _target_stack_size && tq->pop_local(p)) {
   215         process_popped_location_depth(p);
   216       }
   217     }
   218   } while (totally_drain && !tq->taskqueue_empty() || !tq->overflow_empty());
   220   assert(!totally_drain || tq->taskqueue_empty(), "Sanity");
   221   assert(totally_drain || tq->size() <= _target_stack_size, "Sanity");
   222   assert(tq->overflow_empty(), "Sanity");
   223 }
   225 void PSPromotionManager::flush_labs() {
   226   assert(stacks_empty(), "Attempt to flush lab with live stack");
   228   // If either promotion lab fills up, we can flush the
   229   // lab but not refill it, so check first.
   230   assert(!_young_lab.is_flushed() || _young_gen_is_full, "Sanity");
   231   if (!_young_lab.is_flushed())
   232     _young_lab.flush();
   234   assert(!_old_lab.is_flushed() || _old_gen_is_full, "Sanity");
   235   if (!_old_lab.is_flushed())
   236     _old_lab.flush();
   238   // Let PSScavenge know if we overflowed
   239   if (_young_gen_is_full) {
   240     PSScavenge::set_survivor_overflow(true);
   241   }
   242 }
   244 //
   245 // This method is pretty bulky. It would be nice to split it up
   246 // into smaller submethods, but we need to be careful not to hurt
   247 // performance.
   248 //
   250 oop PSPromotionManager::copy_to_survivor_space(oop o) {
   251   assert(PSScavenge::should_scavenge(&o), "Sanity");
   253   oop new_obj = NULL;
   255   // NOTE! We must be very careful with any methods that access the mark
   256   // in o. There may be multiple threads racing on it, and it may be forwarded
   257   // at any time. Do not use oop methods for accessing the mark!
   258   markOop test_mark = o->mark();
   260   // The same test as "o->is_forwarded()"
   261   if (!test_mark->is_marked()) {
   262     bool new_obj_is_tenured = false;
   263     size_t new_obj_size = o->size();
   265     // Find the objects age, MT safe.
   266     int age = (test_mark->has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
   267       test_mark->displaced_mark_helper()->age() : test_mark->age();
   269     // Try allocating obj in to-space (unless too old)
   270     if (age < PSScavenge::tenuring_threshold()) {
   271       new_obj = (oop) _young_lab.allocate(new_obj_size);
   272       if (new_obj == NULL && !_young_gen_is_full) {
   273         // Do we allocate directly, or flush and refill?
   274         if (new_obj_size > (YoungPLABSize / 2)) {
   275           // Allocate this object directly
   276           new_obj = (oop)young_space()->cas_allocate(new_obj_size);
   277         } else {
   278           // Flush and fill
   279           _young_lab.flush();
   281           HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
   282           if (lab_base != NULL) {
   283             _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
   284             // Try the young lab allocation again.
   285             new_obj = (oop) _young_lab.allocate(new_obj_size);
   286           } else {
   287             _young_gen_is_full = true;
   288           }
   289         }
   290       }
   291     }
   293     // Otherwise try allocating obj tenured
   294     if (new_obj == NULL) {
   295 #ifndef PRODUCT
   296       if (Universe::heap()->promotion_should_fail()) {
   297         return oop_promotion_failed(o, test_mark);
   298       }
   299 #endif  // #ifndef PRODUCT
   301       new_obj = (oop) _old_lab.allocate(new_obj_size);
   302       new_obj_is_tenured = true;
   304       if (new_obj == NULL) {
   305         if (!_old_gen_is_full) {
   306           // Do we allocate directly, or flush and refill?
   307           if (new_obj_size > (OldPLABSize / 2)) {
   308             // Allocate this object directly
   309             new_obj = (oop)old_gen()->cas_allocate(new_obj_size);
   310           } else {
   311             // Flush and fill
   312             _old_lab.flush();
   314             HeapWord* lab_base = old_gen()->cas_allocate(OldPLABSize);
   315             if(lab_base != NULL) {
   316               _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
   317               // Try the old lab allocation again.
   318               new_obj = (oop) _old_lab.allocate(new_obj_size);
   319             }
   320           }
   321         }
   323         // This is the promotion failed test, and code handling.
   324         // The code belongs here for two reasons. It is slightly
   325         // different thatn the code below, and cannot share the
   326         // CAS testing code. Keeping the code here also minimizes
   327         // the impact on the common case fast path code.
   329         if (new_obj == NULL) {
   330           _old_gen_is_full = true;
   331           return oop_promotion_failed(o, test_mark);
   332         }
   333       }
   334     }
   336     assert(new_obj != NULL, "allocation should have succeeded");
   338     // Copy obj
   339     Copy::aligned_disjoint_words((HeapWord*)o, (HeapWord*)new_obj, new_obj_size);
   341     // Now we have to CAS in the header.
   342     if (o->cas_forward_to(new_obj, test_mark)) {
   343       // We won any races, we "own" this object.
   344       assert(new_obj == o->forwardee(), "Sanity");
   346       // Increment age if obj still in new generation. Now that
   347       // we're dealing with a markOop that cannot change, it is
   348       // okay to use the non mt safe oop methods.
   349       if (!new_obj_is_tenured) {
   350         new_obj->incr_age();
   351         assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
   352       }
   354       // Do the size comparison first with new_obj_size, which we
   355       // already have. Hopefully, only a few objects are larger than
   356       // _min_array_size_for_chunking, and most of them will be arrays.
   357       // So, the is->objArray() test would be very infrequent.
   358       if (new_obj_size > _min_array_size_for_chunking &&
   359           new_obj->is_objArray() &&
   360           PSChunkLargeArrays) {
   361         // we'll chunk it
   362         oop* const masked_o = mask_chunked_array_oop(o);
   363         push_depth(masked_o);
   364         TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_masked_pushes);
   365       } else {
   366         // we'll just push its contents
   367         new_obj->push_contents(this);
   368       }
   369     }  else {
   370       // We lost, someone else "owns" this object
   371       guarantee(o->is_forwarded(), "Object must be forwarded if the cas failed.");
   373       // Try to deallocate the space.  If it was directly allocated we cannot
   374       // deallocate it, so we have to test.  If the deallocation fails,
   375       // overwrite with a filler object.
   376       if (new_obj_is_tenured) {
   377         if (!_old_lab.unallocate_object(new_obj)) {
   378           CollectedHeap::fill_with_object((HeapWord*) new_obj, new_obj_size);
   379         }
   380       } else if (!_young_lab.unallocate_object(new_obj)) {
   381         CollectedHeap::fill_with_object((HeapWord*) new_obj, new_obj_size);
   382       }
   384       // don't update this before the unallocation!
   385       new_obj = o->forwardee();
   386     }
   387   } else {
   388     assert(o->is_forwarded(), "Sanity");
   389     new_obj = o->forwardee();
   390   }
   392 #ifdef DEBUG
   393   // This code must come after the CAS test, or it will print incorrect
   394   // information.
   395   if (TraceScavenge) {
   396     gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (" SIZE_FORMAT ")}",
   397        PSScavenge::should_scavenge(&new_obj) ? "copying" : "tenuring",
   398        new_obj->blueprint()->internal_name(), o, new_obj, new_obj->size());
   399   }
   400 #endif
   402   return new_obj;
   403 }
   405 template <class T> void PSPromotionManager::process_array_chunk_work(
   406                                                  oop obj,
   407                                                  int start, int end) {
   408   assert(start < end, "invariant");
   409   T* const base      = (T*)objArrayOop(obj)->base();
   410   T* p               = base + start;
   411   T* const chunk_end = base + end;
   412   while (p < chunk_end) {
   413     if (PSScavenge::should_scavenge(p)) {
   414       claim_or_forward_depth(p);
   415     }
   416     ++p;
   417   }
   418 }
   420 void PSPromotionManager::process_array_chunk(oop old) {
   421   assert(PSChunkLargeArrays, "invariant");
   422   assert(old->is_objArray(), "invariant");
   423   assert(old->is_forwarded(), "invariant");
   425   TASKQUEUE_STATS_ONLY(++_array_chunks_processed);
   427   oop const obj = old->forwardee();
   429   int start;
   430   int const end = arrayOop(old)->length();
   431   if (end > (int) _min_array_size_for_chunking) {
   432     // we'll chunk more
   433     start = end - _array_chunk_size;
   434     assert(start > 0, "invariant");
   435     arrayOop(old)->set_length(start);
   436     push_depth(mask_chunked_array_oop(old));
   437     TASKQUEUE_STATS_ONLY(++_masked_pushes);
   438   } else {
   439     // this is the final chunk for this array
   440     start = 0;
   441     int const actual_length = arrayOop(obj)->length();
   442     arrayOop(old)->set_length(actual_length);
   443   }
   445   if (UseCompressedOops) {
   446     process_array_chunk_work<narrowOop>(obj, start, end);
   447   } else {
   448     process_array_chunk_work<oop>(obj, start, end);
   449   }
   450 }
   452 oop PSPromotionManager::oop_promotion_failed(oop obj, markOop obj_mark) {
   453   assert(_old_gen_is_full || PromotionFailureALot, "Sanity");
   455   // Attempt to CAS in the header.
   456   // This tests if the header is still the same as when
   457   // this started.  If it is the same (i.e., no forwarding
   458   // pointer has been installed), then this thread owns
   459   // it.
   460   if (obj->cas_forward_to(obj, obj_mark)) {
   461     // We won any races, we "own" this object.
   462     assert(obj == obj->forwardee(), "Sanity");
   464     obj->push_contents(this);
   466     // Save the mark if needed
   467     PSScavenge::oop_promotion_failed(obj, obj_mark);
   468   }  else {
   469     // We lost, someone else "owns" this object
   470     guarantee(obj->is_forwarded(), "Object must be forwarded if the cas failed.");
   472     // No unallocation to worry about.
   473     obj = obj->forwardee();
   474   }
   476 #ifdef DEBUG
   477   if (TraceScavenge) {
   478     gclog_or_tty->print_cr("{%s %s 0x%x (%d)}",
   479                            "promotion-failure",
   480                            obj->blueprint()->internal_name(),
   481                            obj, obj->size());
   483   }
   484 #endif
   486   return obj;
   487 }

mercurial