src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp

Thu, 30 Apr 2009 15:07:53 -0700

author
johnc
date
Thu, 30 Apr 2009 15:07:53 -0700
changeset 1186
20c6f43950b5
parent 1112
96b229c54d1e
child 1229
315a5d70b295
permissions
-rw-r--r--

6490395: G1: Tidy up command line flags.
Summary: Change G1 flag names to be more consistent and disable some in 'product' mode.
Reviewed-by: tonyp, iveresov

     1 /*
     2  * Copyright 2001-2009 Sun Microsystems, Inc.  All Rights Reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
    20  * CA 95054 USA or visit www.sun.com if you need additional information or
    21  * have any questions.
    22  *
    23  */
    25 #include "incls/_precompiled.incl"
    26 #include "incls/_g1CollectorPolicy.cpp.incl"
    28 #define PREDICTIONS_VERBOSE 0
    30 // <NEW PREDICTION>
    32 // Different defaults for different number of GC threads
    33 // They were chosen by running GCOld and SPECjbb on debris with different
    34 //   numbers of GC threads and choosing them based on the results
    36 // all the same
    37 static double rs_length_diff_defaults[] = {
    38   0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
    39 };
    41 static double cost_per_card_ms_defaults[] = {
    42   0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
    43 };
    45 static double cost_per_scan_only_region_ms_defaults[] = {
    46   1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
    47 };
    49 // all the same
    50 static double fully_young_cards_per_entry_ratio_defaults[] = {
    51   1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
    52 };
    54 static double cost_per_entry_ms_defaults[] = {
    55   0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
    56 };
    58 static double cost_per_byte_ms_defaults[] = {
    59   0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
    60 };
    62 // these should be pretty consistent
    63 static double constant_other_time_ms_defaults[] = {
    64   5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
    65 };
    68 static double young_other_cost_per_region_ms_defaults[] = {
    69   0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
    70 };
    72 static double non_young_other_cost_per_region_ms_defaults[] = {
    73   1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
    74 };
    76 // </NEW PREDICTION>
    78 G1CollectorPolicy::G1CollectorPolicy() :
    79   _parallel_gc_threads((ParallelGCThreads > 0) ? ParallelGCThreads : 1),
    80   _n_pauses(0),
    81   _recent_CH_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
    82   _recent_G1_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
    83   _recent_evac_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
    84   _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
    85   _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)),
    86   _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
    87   _all_pause_times_ms(new NumberSeq()),
    88   _stop_world_start(0.0),
    89   _all_stop_world_times_ms(new NumberSeq()),
    90   _all_yield_times_ms(new NumberSeq()),
    92   _all_mod_union_times_ms(new NumberSeq()),
    94   _summary(new Summary()),
    95   _abandoned_summary(new AbandonedSummary()),
    97   _cur_clear_ct_time_ms(0.0),
    99   _region_num_young(0),
   100   _region_num_tenured(0),
   101   _prev_region_num_young(0),
   102   _prev_region_num_tenured(0),
   104   _aux_num(10),
   105   _all_aux_times_ms(new NumberSeq[_aux_num]),
   106   _cur_aux_start_times_ms(new double[_aux_num]),
   107   _cur_aux_times_ms(new double[_aux_num]),
   108   _cur_aux_times_set(new bool[_aux_num]),
   110   _concurrent_mark_init_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
   111   _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
   112   _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
   114   // <NEW PREDICTION>
   116   _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   117   _prev_collection_pause_end_ms(0.0),
   118   _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
   119   _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
   120   _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   121   _cost_per_scan_only_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   122   _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
   123   _partially_young_cards_per_entry_ratio_seq(
   124                                          new TruncatedSeq(TruncatedSeqLength)),
   125   _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   126   _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   127   _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   128   _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
   129   _cost_per_scan_only_region_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
   130   _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   131   _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
   132   _non_young_other_cost_per_region_ms_seq(
   133                                          new TruncatedSeq(TruncatedSeqLength)),
   135   _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
   136   _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
   137   _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
   139   _pause_time_target_ms((double) MaxGCPauseMillis),
   141   // </NEW PREDICTION>
   143   _in_young_gc_mode(false),
   144   _full_young_gcs(true),
   145   _full_young_pause_num(0),
   146   _partial_young_pause_num(0),
   148   _during_marking(false),
   149   _in_marking_window(false),
   150   _in_marking_window_im(false),
   152   _known_garbage_ratio(0.0),
   153   _known_garbage_bytes(0),
   155   _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
   156   _target_pause_time_ms(-1.0),
   158    _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)),
   160   _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)),
   161   _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)),
   163   _recent_avg_pause_time_ratio(0.0),
   164   _num_markings(0),
   165   _n_marks(0),
   166   _n_pauses_at_mark_end(0),
   168   _all_full_gc_times_ms(new NumberSeq()),
   170   _conc_refine_enabled(0),
   171   _conc_refine_zero_traversals(0),
   172   _conc_refine_max_traversals(0),
   173   _conc_refine_current_delta(G1ConcRefineInitialDelta),
   175   // G1PausesBtwnConcMark defaults to -1
   176   // so the hack is to do the cast  QQQ FIXME
   177   _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark),
   178   _n_marks_since_last_pause(0),
   179   _conc_mark_initiated(false),
   180   _should_initiate_conc_mark(false),
   181   _should_revert_to_full_young_gcs(false),
   182   _last_full_young_gc(false),
   184   _prev_collection_pause_used_at_end_bytes(0),
   186   _collection_set(NULL),
   187 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
   188 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
   189 #endif // _MSC_VER
   191   _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
   192                                                  G1YoungSurvRateNumRegionsSummary)),
   193   _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
   194                                               G1YoungSurvRateNumRegionsSummary)),
   195   // add here any more surv rate groups
   196   _recorded_survivor_regions(0),
   197   _recorded_survivor_head(NULL),
   198   _recorded_survivor_tail(NULL),
   199   _survivors_age_table(true)
   201 {
   202   _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
   203   _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
   205   _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
   206   _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
   207   _par_last_scan_only_times_ms = new double[_parallel_gc_threads];
   208   _par_last_scan_only_regions_scanned = new double[_parallel_gc_threads];
   210   _par_last_update_rs_start_times_ms = new double[_parallel_gc_threads];
   211   _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
   212   _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
   214   _par_last_scan_rs_start_times_ms = new double[_parallel_gc_threads];
   215   _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
   216   _par_last_scan_new_refs_times_ms = new double[_parallel_gc_threads];
   218   _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
   220   _par_last_termination_times_ms = new double[_parallel_gc_threads];
   222   // start conservatively
   223   _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
   225   // <NEW PREDICTION>
   227   int index;
   228   if (ParallelGCThreads == 0)
   229     index = 0;
   230   else if (ParallelGCThreads > 8)
   231     index = 7;
   232   else
   233     index = ParallelGCThreads - 1;
   235   _pending_card_diff_seq->add(0.0);
   236   _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
   237   _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
   238   _cost_per_scan_only_region_ms_seq->add(
   239                                  cost_per_scan_only_region_ms_defaults[index]);
   240   _fully_young_cards_per_entry_ratio_seq->add(
   241                             fully_young_cards_per_entry_ratio_defaults[index]);
   242   _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
   243   _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
   244   _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
   245   _young_other_cost_per_region_ms_seq->add(
   246                                young_other_cost_per_region_ms_defaults[index]);
   247   _non_young_other_cost_per_region_ms_seq->add(
   248                            non_young_other_cost_per_region_ms_defaults[index]);
   250   // </NEW PREDICTION>
   252   double time_slice  = (double) GCPauseIntervalMillis / 1000.0;
   253   double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
   254   guarantee(max_gc_time < time_slice,
   255             "Max GC time should not be greater than the time slice");
   256   _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
   257   _sigma = (double) G1ConfidencePercent / 100.0;
   259   // start conservatively (around 50ms is about right)
   260   _concurrent_mark_init_times_ms->add(0.05);
   261   _concurrent_mark_remark_times_ms->add(0.05);
   262   _concurrent_mark_cleanup_times_ms->add(0.20);
   263   _tenuring_threshold = MaxTenuringThreshold;
   265   if (G1UseSurvivorSpaces) {
   266     // if G1FixedSurvivorSpaceSize is 0 which means the size is not
   267     // fixed, then _max_survivor_regions will be calculated at
   268     // calculate_young_list_target_config during initialization
   269     _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
   270   } else {
   271     _max_survivor_regions = 0;
   272   }
   274   initialize_all();
   275 }
   277 // Increment "i", mod "len"
   278 static void inc_mod(int& i, int len) {
   279   i++; if (i == len) i = 0;
   280 }
   282 void G1CollectorPolicy::initialize_flags() {
   283   set_min_alignment(HeapRegion::GrainBytes);
   284   set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
   285   if (SurvivorRatio < 1) {
   286     vm_exit_during_initialization("Invalid survivor ratio specified");
   287   }
   288   CollectorPolicy::initialize_flags();
   289 }
   291 void G1CollectorPolicy::init() {
   292   // Set aside an initial future to_space.
   293   _g1 = G1CollectedHeap::heap();
   294   size_t regions = Universe::heap()->capacity() / HeapRegion::GrainBytes;
   296   assert(Heap_lock->owned_by_self(), "Locking discipline.");
   298   if (G1SteadyStateUsed < 50) {
   299     vm_exit_during_initialization("G1SteadyStateUsed must be at least 50%.");
   300   }
   301   if (UseConcMarkSweepGC) {
   302     vm_exit_during_initialization("-XX:+UseG1GC is incompatible with "
   303                                   "-XX:+UseConcMarkSweepGC.");
   304   }
   306   initialize_gc_policy_counters();
   308   if (G1Gen) {
   309     _in_young_gc_mode = true;
   311     if (G1YoungGenSize == 0) {
   312       set_adaptive_young_list_length(true);
   313       _young_list_fixed_length = 0;
   314     } else {
   315       set_adaptive_young_list_length(false);
   316       _young_list_fixed_length = (G1YoungGenSize / HeapRegion::GrainBytes);
   317     }
   318      _free_regions_at_end_of_collection = _g1->free_regions();
   319      _scan_only_regions_at_end_of_collection = 0;
   320      calculate_young_list_min_length();
   321      guarantee( _young_list_min_length == 0, "invariant, not enough info" );
   322      calculate_young_list_target_config();
   323    } else {
   324      _young_list_fixed_length = 0;
   325     _in_young_gc_mode = false;
   326   }
   327 }
   329 // Create the jstat counters for the policy.
   330 void G1CollectorPolicy::initialize_gc_policy_counters()
   331 {
   332   _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 2 + G1Gen);
   333 }
   335 void G1CollectorPolicy::calculate_young_list_min_length() {
   336   _young_list_min_length = 0;
   338   if (!adaptive_young_list_length())
   339     return;
   341   if (_alloc_rate_ms_seq->num() > 3) {
   342     double now_sec = os::elapsedTime();
   343     double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
   344     double alloc_rate_ms = predict_alloc_rate_ms();
   345     int min_regions = (int) ceil(alloc_rate_ms * when_ms);
   346     int current_region_num = (int) _g1->young_list_length();
   347     _young_list_min_length = min_regions + current_region_num;
   348   }
   349 }
   351 void G1CollectorPolicy::calculate_young_list_target_config() {
   352   if (adaptive_young_list_length()) {
   353     size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
   354     calculate_young_list_target_config(rs_lengths);
   355   } else {
   356     if (full_young_gcs())
   357       _young_list_target_length = _young_list_fixed_length;
   358     else
   359       _young_list_target_length = _young_list_fixed_length / 2;
   360     _young_list_target_length = MAX2(_young_list_target_length, (size_t)1);
   361     size_t so_length = calculate_optimal_so_length(_young_list_target_length);
   362     guarantee( so_length < _young_list_target_length, "invariant" );
   363     _young_list_so_prefix_length = so_length;
   364   }
   365   calculate_survivors_policy();
   366 }
   368 // This method calculate the optimal scan-only set for a fixed young
   369 // gen size. I couldn't work out how to reuse the more elaborate one,
   370 // i.e. calculate_young_list_target_config(rs_length), as the loops are
   371 // fundamentally different (the other one finds a config for different
   372 // S-O lengths, whereas here we need to do the opposite).
   373 size_t G1CollectorPolicy::calculate_optimal_so_length(
   374                                                     size_t young_list_length) {
   375   if (!G1UseScanOnlyPrefix)
   376     return 0;
   378   if (_all_pause_times_ms->num() < 3) {
   379     // we won't use a scan-only set at the beginning to allow the rest
   380     // of the predictors to warm up
   381     return 0;
   382   }
   384   if (_cost_per_scan_only_region_ms_seq->num() < 3) {
   385     // then, we'll only set the S-O set to 1 for a little bit of time,
   386     // to get enough information on the scanning cost
   387     return 1;
   388   }
   390   size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
   391   size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
   392   size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
   393   size_t scanned_cards;
   394   if (full_young_gcs())
   395     scanned_cards = predict_young_card_num(adj_rs_lengths);
   396   else
   397     scanned_cards = predict_non_young_card_num(adj_rs_lengths);
   398   double base_time_ms = predict_base_elapsed_time_ms(pending_cards,
   399                                                      scanned_cards);
   401   size_t so_length = 0;
   402   double max_gc_eff = 0.0;
   403   for (size_t i = 0; i < young_list_length; ++i) {
   404     double gc_eff = 0.0;
   405     double pause_time_ms = 0.0;
   406     predict_gc_eff(young_list_length, i, base_time_ms,
   407                    &gc_eff, &pause_time_ms);
   408     if (gc_eff > max_gc_eff) {
   409       max_gc_eff = gc_eff;
   410       so_length = i;
   411     }
   412   }
   414   // set it to 95% of the optimal to make sure we sample the "area"
   415   // around the optimal length to get up-to-date survival rate data
   416   return so_length * 950 / 1000;
   417 }
   419 // This is a really cool piece of code! It finds the best
   420 // target configuration (young length / scan-only prefix length) so
   421 // that GC efficiency is maximized and that we also meet a pause
   422 // time. It's a triple nested loop. These loops are explained below
   423 // from the inside-out :-)
   424 //
   425 // (a) The innermost loop will try to find the optimal young length
   426 // for a fixed S-O length. It uses a binary search to speed up the
   427 // process. We assume that, for a fixed S-O length, as we add more
   428 // young regions to the CSet, the GC efficiency will only go up (I'll
   429 // skip the proof). So, using a binary search to optimize this process
   430 // makes perfect sense.
   431 //
   432 // (b) The middle loop will fix the S-O length before calling the
   433 // innermost one. It will vary it between two parameters, increasing
   434 // it by a given increment.
   435 //
   436 // (c) The outermost loop will call the middle loop three times.
   437 //   (1) The first time it will explore all possible S-O length values
   438 //   from 0 to as large as it can get, using a coarse increment (to
   439 //   quickly "home in" to where the optimal seems to be).
   440 //   (2) The second time it will explore the values around the optimal
   441 //   that was found by the first iteration using a fine increment.
   442 //   (3) Once the optimal config has been determined by the second
   443 //   iteration, we'll redo the calculation, but setting the S-O length
   444 //   to 95% of the optimal to make sure we sample the "area"
   445 //   around the optimal length to get up-to-date survival rate data
   446 //
   447 // Termination conditions for the iterations are several: the pause
   448 // time is over the limit, we do not have enough to-space, etc.
   450 void G1CollectorPolicy::calculate_young_list_target_config(size_t rs_lengths) {
   451   guarantee( adaptive_young_list_length(), "pre-condition" );
   453   double start_time_sec = os::elapsedTime();
   454   size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1MinReservePercent);
   455   min_reserve_perc = MIN2((size_t) 50, min_reserve_perc);
   456   size_t reserve_regions =
   457     (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0);
   459   if (full_young_gcs() && _free_regions_at_end_of_collection > 0) {
   460     // we are in fully-young mode and there are free regions in the heap
   462     double survivor_regions_evac_time =
   463         predict_survivor_regions_evac_time();
   465     size_t min_so_length = 0;
   466     size_t max_so_length = 0;
   468     if (G1UseScanOnlyPrefix) {
   469       if (_all_pause_times_ms->num() < 3) {
   470         // we won't use a scan-only set at the beginning to allow the rest
   471         // of the predictors to warm up
   472         min_so_length = 0;
   473         max_so_length = 0;
   474       } else if (_cost_per_scan_only_region_ms_seq->num() < 3) {
   475         // then, we'll only set the S-O set to 1 for a little bit of time,
   476         // to get enough information on the scanning cost
   477         min_so_length = 1;
   478         max_so_length = 1;
   479       } else if (_in_marking_window || _last_full_young_gc) {
   480         // no S-O prefix during a marking phase either, as at the end
   481         // of the marking phase we'll have to use a very small young
   482         // length target to fill up the rest of the CSet with
   483         // non-young regions and, if we have lots of scan-only regions
   484         // left-over, we will not be able to add any more non-young
   485         // regions.
   486         min_so_length = 0;
   487         max_so_length = 0;
   488       } else {
   489         // this is the common case; we'll never reach the maximum, we
   490         // one of the end conditions will fire well before that
   491         // (hopefully!)
   492         min_so_length = 0;
   493         max_so_length = _free_regions_at_end_of_collection - 1;
   494       }
   495     } else {
   496       // no S-O prefix, as the switch is not set, but we still need to
   497       // do one iteration to calculate the best young target that
   498       // meets the pause time; this way we reuse the same code instead
   499       // of replicating it
   500       min_so_length = 0;
   501       max_so_length = 0;
   502     }
   504     double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
   505     size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
   506     size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
   507     size_t scanned_cards;
   508     if (full_young_gcs())
   509       scanned_cards = predict_young_card_num(adj_rs_lengths);
   510     else
   511       scanned_cards = predict_non_young_card_num(adj_rs_lengths);
   512     // calculate this once, so that we don't have to recalculate it in
   513     // the innermost loop
   514     double base_time_ms = predict_base_elapsed_time_ms(pending_cards, scanned_cards)
   515                           + survivor_regions_evac_time;
   516     // the result
   517     size_t final_young_length = 0;
   518     size_t final_so_length = 0;
   519     double final_gc_eff = 0.0;
   520     // we'll also keep track of how many times we go into the inner loop
   521     // this is for profiling reasons
   522     size_t calculations = 0;
   524     // this determines which of the three iterations the outer loop is in
   525     typedef enum {
   526       pass_type_coarse,
   527       pass_type_fine,
   528       pass_type_final
   529     } pass_type_t;
   531     // range of the outer loop's iteration
   532     size_t from_so_length   = min_so_length;
   533     size_t to_so_length     = max_so_length;
   534     guarantee( from_so_length <= to_so_length, "invariant" );
   536     // this will keep the S-O length that's found by the second
   537     // iteration of the outer loop; we'll keep it just in case the third
   538     // iteration fails to find something
   539     size_t fine_so_length   = 0;
   541     // the increment step for the coarse (first) iteration
   542     size_t so_coarse_increments = 5;
   544     // the common case, we'll start with the coarse iteration
   545     pass_type_t pass = pass_type_coarse;
   546     size_t so_length_incr = so_coarse_increments;
   548     if (from_so_length == to_so_length) {
   549       // not point in doing the coarse iteration, we'll go directly into
   550       // the fine one (we essentially trying to find the optimal young
   551       // length for a fixed S-O length).
   552       so_length_incr = 1;
   553       pass = pass_type_final;
   554     } else if (to_so_length - from_so_length < 3 * so_coarse_increments) {
   555       // again, the range is too short so no point in foind the coarse
   556       // iteration either
   557       so_length_incr = 1;
   558       pass = pass_type_fine;
   559     }
   561     bool done = false;
   562     // this is the outermost loop
   563     while (!done) {
   564 #ifdef TRACE_CALC_YOUNG_CONFIG
   565       // leave this in for debugging, just in case
   566       gclog_or_tty->print_cr("searching between " SIZE_FORMAT " and " SIZE_FORMAT
   567                              ", incr " SIZE_FORMAT ", pass %s",
   568                              from_so_length, to_so_length, so_length_incr,
   569                              (pass == pass_type_coarse) ? "coarse" :
   570                              (pass == pass_type_fine) ? "fine" : "final");
   571 #endif // TRACE_CALC_YOUNG_CONFIG
   573       size_t so_length = from_so_length;
   574       size_t init_free_regions =
   575         MAX2((size_t)0,
   576              _free_regions_at_end_of_collection +
   577              _scan_only_regions_at_end_of_collection - reserve_regions);
   579       // this determines whether a configuration was found
   580       bool gc_eff_set = false;
   581       // this is the middle loop
   582       while (so_length <= to_so_length) {
   583         // base time, which excludes region-related time; again we
   584         // calculate it once to avoid recalculating it in the
   585         // innermost loop
   586         double base_time_with_so_ms =
   587                            base_time_ms + predict_scan_only_time_ms(so_length);
   588         // it's already over the pause target, go around
   589         if (base_time_with_so_ms > target_pause_time_ms)
   590           break;
   592         size_t starting_young_length = so_length+1;
   594         // we make sure that the short young length that makes sense
   595         // (one more than the S-O length) is feasible
   596         size_t min_young_length = starting_young_length;
   597         double min_gc_eff;
   598         bool min_ok;
   599         ++calculations;
   600         min_ok = predict_gc_eff(min_young_length, so_length,
   601                                 base_time_with_so_ms,
   602                                 init_free_regions, target_pause_time_ms,
   603                                 &min_gc_eff);
   605         if (min_ok) {
   606           // the shortest young length is indeed feasible; we'll know
   607           // set up the max young length and we'll do a binary search
   608           // between min_young_length and max_young_length
   609           size_t max_young_length = _free_regions_at_end_of_collection - 1;
   610           double max_gc_eff = 0.0;
   611           bool max_ok = false;
   613           // the innermost loop! (finally!)
   614           while (max_young_length > min_young_length) {
   615             // we'll make sure that min_young_length is always at a
   616             // feasible config
   617             guarantee( min_ok, "invariant" );
   619             ++calculations;
   620             max_ok = predict_gc_eff(max_young_length, so_length,
   621                                     base_time_with_so_ms,
   622                                     init_free_regions, target_pause_time_ms,
   623                                     &max_gc_eff);
   625             size_t diff = (max_young_length - min_young_length) / 2;
   626             if (max_ok) {
   627               min_young_length = max_young_length;
   628               min_gc_eff = max_gc_eff;
   629               min_ok = true;
   630             }
   631             max_young_length = min_young_length + diff;
   632           }
   634           // the innermost loop found a config
   635           guarantee( min_ok, "invariant" );
   636           if (min_gc_eff > final_gc_eff) {
   637             // it's the best config so far, so we'll keep it
   638             final_gc_eff = min_gc_eff;
   639             final_young_length = min_young_length;
   640             final_so_length = so_length;
   641             gc_eff_set = true;
   642           }
   643         }
   645         // incremental the fixed S-O length and go around
   646         so_length += so_length_incr;
   647       }
   649       // this is the end of the outermost loop and we need to decide
   650       // what to do during the next iteration
   651       if (pass == pass_type_coarse) {
   652         // we just did the coarse pass (first iteration)
   654         if (!gc_eff_set)
   655           // we didn't find a feasible config so we'll just bail out; of
   656           // course, it might be the case that we missed it; but I'd say
   657           // it's a bit unlikely
   658           done = true;
   659         else {
   660           // We did find a feasible config with optimal GC eff during
   661           // the first pass. So the second pass we'll only consider the
   662           // S-O lengths around that config with a fine increment.
   664           guarantee( so_length_incr == so_coarse_increments, "invariant" );
   665           guarantee( final_so_length >= min_so_length, "invariant" );
   667 #ifdef TRACE_CALC_YOUNG_CONFIG
   668           // leave this in for debugging, just in case
   669           gclog_or_tty->print_cr("  coarse pass: SO length " SIZE_FORMAT,
   670                                  final_so_length);
   671 #endif // TRACE_CALC_YOUNG_CONFIG
   673           from_so_length =
   674             (final_so_length - min_so_length > so_coarse_increments) ?
   675             final_so_length - so_coarse_increments + 1 : min_so_length;
   676           to_so_length =
   677             (max_so_length - final_so_length > so_coarse_increments) ?
   678             final_so_length + so_coarse_increments - 1 : max_so_length;
   680           pass = pass_type_fine;
   681           so_length_incr = 1;
   682         }
   683       } else if (pass == pass_type_fine) {
   684         // we just finished the second pass
   686         if (!gc_eff_set) {
   687           // we didn't find a feasible config (yes, it's possible;
   688           // notice that, sometimes, we go directly into the fine
   689           // iteration and skip the coarse one) so we bail out
   690           done = true;
   691         } else {
   692           // We did find a feasible config with optimal GC eff
   693           guarantee( so_length_incr == 1, "invariant" );
   695           if (final_so_length == 0) {
   696             // The config is of an empty S-O set, so we'll just bail out
   697             done = true;
   698           } else {
   699             // we'll go around once more, setting the S-O length to 95%
   700             // of the optimal
   701             size_t new_so_length = 950 * final_so_length / 1000;
   703 #ifdef TRACE_CALC_YOUNG_CONFIG
   704             // leave this in for debugging, just in case
   705             gclog_or_tty->print_cr("  fine pass: SO length " SIZE_FORMAT
   706                                    ", setting it to " SIZE_FORMAT,
   707                                     final_so_length, new_so_length);
   708 #endif // TRACE_CALC_YOUNG_CONFIG
   710             from_so_length = new_so_length;
   711             to_so_length = new_so_length;
   712             fine_so_length = final_so_length;
   714             pass = pass_type_final;
   715           }
   716         }
   717       } else if (pass == pass_type_final) {
   718         // we just finished the final (third) pass
   720         if (!gc_eff_set)
   721           // we didn't find a feasible config, so we'll just use the one
   722           // we found during the second pass, which we saved
   723           final_so_length = fine_so_length;
   725         // and we're done!
   726         done = true;
   727       } else {
   728         guarantee( false, "should never reach here" );
   729       }
   731       // we now go around the outermost loop
   732     }
   734     // we should have at least one region in the target young length
   735     _young_list_target_length =
   736         MAX2((size_t) 1, final_young_length + _recorded_survivor_regions);
   737     if (final_so_length >= final_young_length)
   738       // and we need to ensure that the S-O length is not greater than
   739       // the target young length (this is being a bit careful)
   740       final_so_length = 0;
   741     _young_list_so_prefix_length = final_so_length;
   742     guarantee( !_in_marking_window || !_last_full_young_gc ||
   743                _young_list_so_prefix_length == 0, "invariant" );
   745     // let's keep an eye of how long we spend on this calculation
   746     // right now, I assume that we'll print it when we need it; we
   747     // should really adde it to the breakdown of a pause
   748     double end_time_sec = os::elapsedTime();
   749     double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0;
   751 #ifdef TRACE_CALC_YOUNG_CONFIG
   752     // leave this in for debugging, just in case
   753     gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT
   754                            ", SO = " SIZE_FORMAT ", "
   755                            "elapsed %1.2lf ms, calcs: " SIZE_FORMAT " (%s%s) "
   756                            SIZE_FORMAT SIZE_FORMAT,
   757                            target_pause_time_ms,
   758                            _young_list_target_length - _young_list_so_prefix_length,
   759                            _young_list_so_prefix_length,
   760                            elapsed_time_ms,
   761                            calculations,
   762                            full_young_gcs() ? "full" : "partial",
   763                            should_initiate_conc_mark() ? " i-m" : "",
   764                            _in_marking_window,
   765                            _in_marking_window_im);
   766 #endif // TRACE_CALC_YOUNG_CONFIG
   768     if (_young_list_target_length < _young_list_min_length) {
   769       // bummer; this means that, if we do a pause when the optimal
   770       // config dictates, we'll violate the pause spacing target (the
   771       // min length was calculate based on the application's current
   772       // alloc rate);
   774       // so, we have to bite the bullet, and allocate the minimum
   775       // number. We'll violate our target, but we just can't meet it.
   777       size_t so_length = 0;
   778       // a note further up explains why we do not want an S-O length
   779       // during marking
   780       if (!_in_marking_window && !_last_full_young_gc)
   781         // but we can still try to see whether we can find an optimal
   782         // S-O length
   783         so_length = calculate_optimal_so_length(_young_list_min_length);
   785 #ifdef TRACE_CALC_YOUNG_CONFIG
   786       // leave this in for debugging, just in case
   787       gclog_or_tty->print_cr("adjusted target length from "
   788                              SIZE_FORMAT " to " SIZE_FORMAT
   789                              ", SO " SIZE_FORMAT,
   790                              _young_list_target_length, _young_list_min_length,
   791                              so_length);
   792 #endif // TRACE_CALC_YOUNG_CONFIG
   794       _young_list_target_length =
   795         MAX2(_young_list_min_length, (size_t)1);
   796       _young_list_so_prefix_length = so_length;
   797     }
   798   } else {
   799     // we are in a partially-young mode or we've run out of regions (due
   800     // to evacuation failure)
   802 #ifdef TRACE_CALC_YOUNG_CONFIG
   803     // leave this in for debugging, just in case
   804     gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT
   805                            ", SO " SIZE_FORMAT,
   806                            _young_list_min_length, 0);
   807 #endif // TRACE_CALC_YOUNG_CONFIG
   809     // we'll do the pause as soon as possible and with no S-O prefix
   810     // (see above for the reasons behind the latter)
   811     _young_list_target_length =
   812       MAX2(_young_list_min_length, (size_t) 1);
   813     _young_list_so_prefix_length = 0;
   814   }
   816   _rs_lengths_prediction = rs_lengths;
   817 }
   819 // This is used by: calculate_optimal_so_length(length). It returns
   820 // the GC eff and predicted pause time for a particular config
   821 void
   822 G1CollectorPolicy::predict_gc_eff(size_t young_length,
   823                                   size_t so_length,
   824                                   double base_time_ms,
   825                                   double* ret_gc_eff,
   826                                   double* ret_pause_time_ms) {
   827   double so_time_ms = predict_scan_only_time_ms(so_length);
   828   double accum_surv_rate_adj = 0.0;
   829   if (so_length > 0)
   830     accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
   831   double accum_surv_rate =
   832     accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
   833   size_t bytes_to_copy =
   834     (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
   835   double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
   836   double young_other_time_ms =
   837                        predict_young_other_time_ms(young_length - so_length);
   838   double pause_time_ms =
   839                 base_time_ms + so_time_ms + copy_time_ms + young_other_time_ms;
   840   size_t reclaimed_bytes =
   841     (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
   842   double gc_eff = (double) reclaimed_bytes / pause_time_ms;
   844   *ret_gc_eff = gc_eff;
   845   *ret_pause_time_ms = pause_time_ms;
   846 }
   848 // This is used by: calculate_young_list_target_config(rs_length). It
   849 // returns the GC eff of a particular config. It returns false if that
   850 // config violates any of the end conditions of the search in the
   851 // calling method, or true upon success. The end conditions were put
   852 // here since it's called twice and it was best not to replicate them
   853 // in the caller. Also, passing the parameteres avoids having to
   854 // recalculate them in the innermost loop.
   855 bool
   856 G1CollectorPolicy::predict_gc_eff(size_t young_length,
   857                                   size_t so_length,
   858                                   double base_time_with_so_ms,
   859                                   size_t init_free_regions,
   860                                   double target_pause_time_ms,
   861                                   double* ret_gc_eff) {
   862   *ret_gc_eff = 0.0;
   864   if (young_length >= init_free_regions)
   865     // end condition 1: not enough space for the young regions
   866     return false;
   868   double accum_surv_rate_adj = 0.0;
   869   if (so_length > 0)
   870     accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
   871   double accum_surv_rate =
   872     accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
   873   size_t bytes_to_copy =
   874     (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
   875   double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
   876   double young_other_time_ms =
   877                        predict_young_other_time_ms(young_length - so_length);
   878   double pause_time_ms =
   879                    base_time_with_so_ms + copy_time_ms + young_other_time_ms;
   881   if (pause_time_ms > target_pause_time_ms)
   882     // end condition 2: over the target pause time
   883     return false;
   885   size_t reclaimed_bytes =
   886     (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
   887   size_t free_bytes =
   888                  (init_free_regions - young_length) * HeapRegion::GrainBytes;
   890   if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes)
   891     // end condition 3: out of to-space (conservatively)
   892     return false;
   894   // success!
   895   double gc_eff = (double) reclaimed_bytes / pause_time_ms;
   896   *ret_gc_eff = gc_eff;
   898   return true;
   899 }
   901 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
   902   double survivor_regions_evac_time = 0.0;
   903   for (HeapRegion * r = _recorded_survivor_head;
   904        r != NULL && r != _recorded_survivor_tail->get_next_young_region();
   905        r = r->get_next_young_region()) {
   906     survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
   907   }
   908   return survivor_regions_evac_time;
   909 }
   911 void G1CollectorPolicy::check_prediction_validity() {
   912   guarantee( adaptive_young_list_length(), "should not call this otherwise" );
   914   size_t rs_lengths = _g1->young_list_sampled_rs_lengths();
   915   if (rs_lengths > _rs_lengths_prediction) {
   916     // add 10% to avoid having to recalculate often
   917     size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
   918     calculate_young_list_target_config(rs_lengths_prediction);
   919   }
   920 }
   922 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
   923                                                bool is_tlab,
   924                                                bool* gc_overhead_limit_was_exceeded) {
   925   guarantee(false, "Not using this policy feature yet.");
   926   return NULL;
   927 }
   929 // This method controls how a collector handles one or more
   930 // of its generations being fully allocated.
   931 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
   932                                                        bool is_tlab) {
   933   guarantee(false, "Not using this policy feature yet.");
   934   return NULL;
   935 }
   938 #ifndef PRODUCT
   939 bool G1CollectorPolicy::verify_young_ages() {
   940   HeapRegion* head = _g1->young_list_first_region();
   941   return
   942     verify_young_ages(head, _short_lived_surv_rate_group);
   943   // also call verify_young_ages on any additional surv rate groups
   944 }
   946 bool
   947 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
   948                                      SurvRateGroup *surv_rate_group) {
   949   guarantee( surv_rate_group != NULL, "pre-condition" );
   951   const char* name = surv_rate_group->name();
   952   bool ret = true;
   953   int prev_age = -1;
   955   for (HeapRegion* curr = head;
   956        curr != NULL;
   957        curr = curr->get_next_young_region()) {
   958     SurvRateGroup* group = curr->surv_rate_group();
   959     if (group == NULL && !curr->is_survivor()) {
   960       gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
   961       ret = false;
   962     }
   964     if (surv_rate_group == group) {
   965       int age = curr->age_in_surv_rate_group();
   967       if (age < 0) {
   968         gclog_or_tty->print_cr("## %s: encountered negative age", name);
   969         ret = false;
   970       }
   972       if (age <= prev_age) {
   973         gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
   974                                "(%d, %d)", name, age, prev_age);
   975         ret = false;
   976       }
   977       prev_age = age;
   978     }
   979   }
   981   return ret;
   982 }
   983 #endif // PRODUCT
   985 void G1CollectorPolicy::record_full_collection_start() {
   986   _cur_collection_start_sec = os::elapsedTime();
   987   // Release the future to-space so that it is available for compaction into.
   988   _g1->set_full_collection();
   989 }
   991 void G1CollectorPolicy::record_full_collection_end() {
   992   // Consider this like a collection pause for the purposes of allocation
   993   // since last pause.
   994   double end_sec = os::elapsedTime();
   995   double full_gc_time_sec = end_sec - _cur_collection_start_sec;
   996   double full_gc_time_ms = full_gc_time_sec * 1000.0;
   998   checkpoint_conc_overhead();
  1000   _all_full_gc_times_ms->add(full_gc_time_ms);
  1002   update_recent_gc_times(end_sec, full_gc_time_ms);
  1004   _g1->clear_full_collection();
  1006   // "Nuke" the heuristics that control the fully/partially young GC
  1007   // transitions and make sure we start with fully young GCs after the
  1008   // Full GC.
  1009   set_full_young_gcs(true);
  1010   _last_full_young_gc = false;
  1011   _should_revert_to_full_young_gcs = false;
  1012   _should_initiate_conc_mark = false;
  1013   _known_garbage_bytes = 0;
  1014   _known_garbage_ratio = 0.0;
  1015   _in_marking_window = false;
  1016   _in_marking_window_im = false;
  1018   _short_lived_surv_rate_group->record_scan_only_prefix(0);
  1019   _short_lived_surv_rate_group->start_adding_regions();
  1020   // also call this on any additional surv rate groups
  1022   record_survivor_regions(0, NULL, NULL);
  1024   _prev_region_num_young   = _region_num_young;
  1025   _prev_region_num_tenured = _region_num_tenured;
  1027   _free_regions_at_end_of_collection = _g1->free_regions();
  1028   _scan_only_regions_at_end_of_collection = 0;
  1029   // Reset survivors SurvRateGroup.
  1030   _survivor_surv_rate_group->reset();
  1031   calculate_young_list_min_length();
  1032   calculate_young_list_target_config();
  1035 void G1CollectorPolicy::record_before_bytes(size_t bytes) {
  1036   _bytes_in_to_space_before_gc += bytes;
  1039 void G1CollectorPolicy::record_after_bytes(size_t bytes) {
  1040   _bytes_in_to_space_after_gc += bytes;
  1043 void G1CollectorPolicy::record_stop_world_start() {
  1044   _stop_world_start = os::elapsedTime();
  1047 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
  1048                                                       size_t start_used) {
  1049   if (PrintGCDetails) {
  1050     gclog_or_tty->stamp(PrintGCTimeStamps);
  1051     gclog_or_tty->print("[GC pause");
  1052     if (in_young_gc_mode())
  1053       gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
  1056   assert(_g1->used_regions() == _g1->recalculate_used_regions(),
  1057          "sanity");
  1058   assert(_g1->used() == _g1->recalculate_used(), "sanity");
  1060   double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
  1061   _all_stop_world_times_ms->add(s_w_t_ms);
  1062   _stop_world_start = 0.0;
  1064   _cur_collection_start_sec = start_time_sec;
  1065   _cur_collection_pause_used_at_start_bytes = start_used;
  1066   _cur_collection_pause_used_regions_at_start = _g1->used_regions();
  1067   _pending_cards = _g1->pending_card_num();
  1068   _max_pending_cards = _g1->max_pending_card_num();
  1070   _bytes_in_to_space_before_gc = 0;
  1071   _bytes_in_to_space_after_gc = 0;
  1072   _bytes_in_collection_set_before_gc = 0;
  1074 #ifdef DEBUG
  1075   // initialise these to something well known so that we can spot
  1076   // if they are not set properly
  1078   for (int i = 0; i < _parallel_gc_threads; ++i) {
  1079     _par_last_ext_root_scan_times_ms[i] = -666.0;
  1080     _par_last_mark_stack_scan_times_ms[i] = -666.0;
  1081     _par_last_scan_only_times_ms[i] = -666.0;
  1082     _par_last_scan_only_regions_scanned[i] = -666.0;
  1083     _par_last_update_rs_start_times_ms[i] = -666.0;
  1084     _par_last_update_rs_times_ms[i] = -666.0;
  1085     _par_last_update_rs_processed_buffers[i] = -666.0;
  1086     _par_last_scan_rs_start_times_ms[i] = -666.0;
  1087     _par_last_scan_rs_times_ms[i] = -666.0;
  1088     _par_last_scan_new_refs_times_ms[i] = -666.0;
  1089     _par_last_obj_copy_times_ms[i] = -666.0;
  1090     _par_last_termination_times_ms[i] = -666.0;
  1092 #endif
  1094   for (int i = 0; i < _aux_num; ++i) {
  1095     _cur_aux_times_ms[i] = 0.0;
  1096     _cur_aux_times_set[i] = false;
  1099   _satb_drain_time_set = false;
  1100   _last_satb_drain_processed_buffers = -1;
  1102   if (in_young_gc_mode())
  1103     _last_young_gc_full = false;
  1106   // do that for any other surv rate groups
  1107   _short_lived_surv_rate_group->stop_adding_regions();
  1108   size_t short_lived_so_length = _young_list_so_prefix_length;
  1109   _short_lived_surv_rate_group->record_scan_only_prefix(short_lived_so_length);
  1110   tag_scan_only(short_lived_so_length);
  1112   if (G1UseSurvivorSpaces) {
  1113     _survivors_age_table.clear();
  1116   assert( verify_young_ages(), "region age verification" );
  1119 void G1CollectorPolicy::tag_scan_only(size_t short_lived_scan_only_length) {
  1120   // done in a way that it can be extended for other surv rate groups too...
  1122   HeapRegion* head = _g1->young_list_first_region();
  1123   bool finished_short_lived = (short_lived_scan_only_length == 0);
  1125   if (finished_short_lived)
  1126     return;
  1128   for (HeapRegion* curr = head;
  1129        curr != NULL;
  1130        curr = curr->get_next_young_region()) {
  1131     SurvRateGroup* surv_rate_group = curr->surv_rate_group();
  1132     int age = curr->age_in_surv_rate_group();
  1134     if (surv_rate_group == _short_lived_surv_rate_group) {
  1135       if ((size_t)age < short_lived_scan_only_length)
  1136         curr->set_scan_only();
  1137       else
  1138         finished_short_lived = true;
  1142     if (finished_short_lived)
  1143       return;
  1146   guarantee( false, "we should never reach here" );
  1149 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
  1150   _mark_closure_time_ms = mark_closure_time_ms;
  1153 void G1CollectorPolicy::record_concurrent_mark_init_start() {
  1154   _mark_init_start_sec = os::elapsedTime();
  1155   guarantee(!in_young_gc_mode(), "should not do be here in young GC mode");
  1158 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double
  1159                                                    mark_init_elapsed_time_ms) {
  1160   _during_marking = true;
  1161   _should_initiate_conc_mark = false;
  1162   _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
  1165 void G1CollectorPolicy::record_concurrent_mark_init_end() {
  1166   double end_time_sec = os::elapsedTime();
  1167   double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0;
  1168   _concurrent_mark_init_times_ms->add(elapsed_time_ms);
  1169   checkpoint_conc_overhead();
  1170   record_concurrent_mark_init_end_pre(elapsed_time_ms);
  1172   _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true);
  1175 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
  1176   _mark_remark_start_sec = os::elapsedTime();
  1177   _during_marking = false;
  1180 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
  1181   double end_time_sec = os::elapsedTime();
  1182   double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
  1183   checkpoint_conc_overhead();
  1184   _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
  1185   _cur_mark_stop_world_time_ms += elapsed_time_ms;
  1186   _prev_collection_pause_end_ms += elapsed_time_ms;
  1188   _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
  1191 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
  1192   _mark_cleanup_start_sec = os::elapsedTime();
  1195 void
  1196 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes,
  1197                                                       size_t max_live_bytes) {
  1198   record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
  1199   record_concurrent_mark_cleanup_end_work2();
  1202 void
  1203 G1CollectorPolicy::
  1204 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
  1205                                          size_t max_live_bytes) {
  1206   if (_n_marks < 2) _n_marks++;
  1207   if (G1PolicyVerbose > 0)
  1208     gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB "
  1209                            " (of " SIZE_FORMAT " MB heap).",
  1210                            max_live_bytes/M, _g1->capacity()/M);
  1213 // The important thing about this is that it includes "os::elapsedTime".
  1214 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() {
  1215   checkpoint_conc_overhead();
  1216   double end_time_sec = os::elapsedTime();
  1217   double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0;
  1218   _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
  1219   _cur_mark_stop_world_time_ms += elapsed_time_ms;
  1220   _prev_collection_pause_end_ms += elapsed_time_ms;
  1222   _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true);
  1224   _num_markings++;
  1226   // We did a marking, so reset the "since_last_mark" variables.
  1227   double considerConcMarkCost = 1.0;
  1228   // If there are available processors, concurrent activity is free...
  1229   if (Threads::number_of_non_daemon_threads() * 2 <
  1230       os::active_processor_count()) {
  1231     considerConcMarkCost = 0.0;
  1233   _n_pauses_at_mark_end = _n_pauses;
  1234   _n_marks_since_last_pause++;
  1235   _conc_mark_initiated = false;
  1238 void
  1239 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
  1240   if (in_young_gc_mode()) {
  1241     _should_revert_to_full_young_gcs = false;
  1242     _last_full_young_gc = true;
  1243     _in_marking_window = false;
  1244     if (adaptive_young_list_length())
  1245       calculate_young_list_target_config();
  1249 void G1CollectorPolicy::record_concurrent_pause() {
  1250   if (_stop_world_start > 0.0) {
  1251     double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
  1252     _all_yield_times_ms->add(yield_ms);
  1256 void G1CollectorPolicy::record_concurrent_pause_end() {
  1259 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() {
  1260   _cur_CH_strong_roots_end_sec = os::elapsedTime();
  1261   _cur_CH_strong_roots_dur_ms =
  1262     (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0;
  1265 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() {
  1266   _cur_G1_strong_roots_end_sec = os::elapsedTime();
  1267   _cur_G1_strong_roots_dur_ms =
  1268     (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0;
  1271 template<class T>
  1272 T sum_of(T* sum_arr, int start, int n, int N) {
  1273   T sum = (T)0;
  1274   for (int i = 0; i < n; i++) {
  1275     int j = (start + i) % N;
  1276     sum += sum_arr[j];
  1278   return sum;
  1281 void G1CollectorPolicy::print_par_stats (int level,
  1282                                          const char* str,
  1283                                          double* data,
  1284                                          bool summary) {
  1285   double min = data[0], max = data[0];
  1286   double total = 0.0;
  1287   int j;
  1288   for (j = 0; j < level; ++j)
  1289     gclog_or_tty->print("   ");
  1290   gclog_or_tty->print("[%s (ms):", str);
  1291   for (uint i = 0; i < ParallelGCThreads; ++i) {
  1292     double val = data[i];
  1293     if (val < min)
  1294       min = val;
  1295     if (val > max)
  1296       max = val;
  1297     total += val;
  1298     gclog_or_tty->print("  %3.1lf", val);
  1300   if (summary) {
  1301     gclog_or_tty->print_cr("");
  1302     double avg = total / (double) ParallelGCThreads;
  1303     gclog_or_tty->print(" ");
  1304     for (j = 0; j < level; ++j)
  1305       gclog_or_tty->print("   ");
  1306     gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf",
  1307                         avg, min, max);
  1309   gclog_or_tty->print_cr("]");
  1312 void G1CollectorPolicy::print_par_buffers (int level,
  1313                                          const char* str,
  1314                                          double* data,
  1315                                          bool summary) {
  1316   double min = data[0], max = data[0];
  1317   double total = 0.0;
  1318   int j;
  1319   for (j = 0; j < level; ++j)
  1320     gclog_or_tty->print("   ");
  1321   gclog_or_tty->print("[%s :", str);
  1322   for (uint i = 0; i < ParallelGCThreads; ++i) {
  1323     double val = data[i];
  1324     if (val < min)
  1325       min = val;
  1326     if (val > max)
  1327       max = val;
  1328     total += val;
  1329     gclog_or_tty->print(" %d", (int) val);
  1331   if (summary) {
  1332     gclog_or_tty->print_cr("");
  1333     double avg = total / (double) ParallelGCThreads;
  1334     gclog_or_tty->print(" ");
  1335     for (j = 0; j < level; ++j)
  1336       gclog_or_tty->print("   ");
  1337     gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d",
  1338                (int)total, (int)avg, (int)min, (int)max);
  1340   gclog_or_tty->print_cr("]");
  1343 void G1CollectorPolicy::print_stats (int level,
  1344                                      const char* str,
  1345                                      double value) {
  1346   for (int j = 0; j < level; ++j)
  1347     gclog_or_tty->print("   ");
  1348   gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value);
  1351 void G1CollectorPolicy::print_stats (int level,
  1352                                      const char* str,
  1353                                      int value) {
  1354   for (int j = 0; j < level; ++j)
  1355     gclog_or_tty->print("   ");
  1356   gclog_or_tty->print_cr("[%s: %d]", str, value);
  1359 double G1CollectorPolicy::avg_value (double* data) {
  1360   if (ParallelGCThreads > 0) {
  1361     double ret = 0.0;
  1362     for (uint i = 0; i < ParallelGCThreads; ++i)
  1363       ret += data[i];
  1364     return ret / (double) ParallelGCThreads;
  1365   } else {
  1366     return data[0];
  1370 double G1CollectorPolicy::max_value (double* data) {
  1371   if (ParallelGCThreads > 0) {
  1372     double ret = data[0];
  1373     for (uint i = 1; i < ParallelGCThreads; ++i)
  1374       if (data[i] > ret)
  1375         ret = data[i];
  1376     return ret;
  1377   } else {
  1378     return data[0];
  1382 double G1CollectorPolicy::sum_of_values (double* data) {
  1383   if (ParallelGCThreads > 0) {
  1384     double sum = 0.0;
  1385     for (uint i = 0; i < ParallelGCThreads; i++)
  1386       sum += data[i];
  1387     return sum;
  1388   } else {
  1389     return data[0];
  1393 double G1CollectorPolicy::max_sum (double* data1,
  1394                                    double* data2) {
  1395   double ret = data1[0] + data2[0];
  1397   if (ParallelGCThreads > 0) {
  1398     for (uint i = 1; i < ParallelGCThreads; ++i) {
  1399       double data = data1[i] + data2[i];
  1400       if (data > ret)
  1401         ret = data;
  1404   return ret;
  1407 // Anything below that is considered to be zero
  1408 #define MIN_TIMER_GRANULARITY 0.0000001
  1410 void G1CollectorPolicy::record_collection_pause_end(bool abandoned) {
  1411   double end_time_sec = os::elapsedTime();
  1412   double elapsed_ms = _last_pause_time_ms;
  1413   bool parallel = ParallelGCThreads > 0;
  1414   double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0;
  1415   size_t rs_size =
  1416     _cur_collection_pause_used_regions_at_start - collection_set_size();
  1417   size_t cur_used_bytes = _g1->used();
  1418   assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
  1419   bool last_pause_included_initial_mark = false;
  1420   bool update_stats = !abandoned && !_g1->evacuation_failed();
  1422 #ifndef PRODUCT
  1423   if (G1YoungSurvRateVerbose) {
  1424     gclog_or_tty->print_cr("");
  1425     _short_lived_surv_rate_group->print();
  1426     // do that for any other surv rate groups too
  1428 #endif // PRODUCT
  1430   checkpoint_conc_overhead();
  1432   if (in_young_gc_mode()) {
  1433     last_pause_included_initial_mark = _should_initiate_conc_mark;
  1434     if (last_pause_included_initial_mark)
  1435       record_concurrent_mark_init_end_pre(0.0);
  1437     size_t min_used_targ =
  1438       (_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta);
  1440     if (cur_used_bytes > min_used_targ) {
  1441       if (cur_used_bytes <= _prev_collection_pause_used_at_end_bytes) {
  1442       } else if (!_g1->mark_in_progress() && !_last_full_young_gc) {
  1443         _should_initiate_conc_mark = true;
  1447     _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
  1450   _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
  1451                           end_time_sec, false);
  1453   guarantee(_cur_collection_pause_used_regions_at_start >=
  1454             collection_set_size(),
  1455             "Negative RS size?");
  1457   // This assert is exempted when we're doing parallel collection pauses,
  1458   // because the fragmentation caused by the parallel GC allocation buffers
  1459   // can lead to more memory being used during collection than was used
  1460   // before. Best leave this out until the fragmentation problem is fixed.
  1461   // Pauses in which evacuation failed can also lead to negative
  1462   // collections, since no space is reclaimed from a region containing an
  1463   // object whose evacuation failed.
  1464   // Further, we're now always doing parallel collection.  But I'm still
  1465   // leaving this here as a placeholder for a more precise assertion later.
  1466   // (DLD, 10/05.)
  1467   assert((true || parallel) // Always using GC LABs now.
  1468          || _g1->evacuation_failed()
  1469          || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
  1470          "Negative collection");
  1472   size_t freed_bytes =
  1473     _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
  1474   size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
  1475   double survival_fraction =
  1476     (double)surviving_bytes/
  1477     (double)_collection_set_bytes_used_before;
  1479   _n_pauses++;
  1481   if (update_stats) {
  1482     _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms);
  1483     _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms);
  1484     _recent_evac_times_ms->add(evac_ms);
  1485     _recent_pause_times_ms->add(elapsed_ms);
  1487     _recent_rs_sizes->add(rs_size);
  1489     // We exempt parallel collection from this check because Alloc Buffer
  1490     // fragmentation can produce negative collections.  Same with evac
  1491     // failure.
  1492     // Further, we're now always doing parallel collection.  But I'm still
  1493     // leaving this here as a placeholder for a more precise assertion later.
  1494     // (DLD, 10/05.
  1495     assert((true || parallel)
  1496            || _g1->evacuation_failed()
  1497            || surviving_bytes <= _collection_set_bytes_used_before,
  1498            "Or else negative collection!");
  1499     _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
  1500     _recent_CS_bytes_surviving->add(surviving_bytes);
  1502     // this is where we update the allocation rate of the application
  1503     double app_time_ms =
  1504       (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
  1505     if (app_time_ms < MIN_TIMER_GRANULARITY) {
  1506       // This usually happens due to the timer not having the required
  1507       // granularity. Some Linuxes are the usual culprits.
  1508       // We'll just set it to something (arbitrarily) small.
  1509       app_time_ms = 1.0;
  1511     size_t regions_allocated =
  1512       (_region_num_young - _prev_region_num_young) +
  1513       (_region_num_tenured - _prev_region_num_tenured);
  1514     double alloc_rate_ms = (double) regions_allocated / app_time_ms;
  1515     _alloc_rate_ms_seq->add(alloc_rate_ms);
  1516     _prev_region_num_young   = _region_num_young;
  1517     _prev_region_num_tenured = _region_num_tenured;
  1519     double interval_ms =
  1520       (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
  1521     update_recent_gc_times(end_time_sec, elapsed_ms);
  1522     _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
  1523     assert(recent_avg_pause_time_ratio() < 1.00, "All GC?");
  1526   if (G1PolicyVerbose > 1) {
  1527     gclog_or_tty->print_cr("   Recording collection pause(%d)", _n_pauses);
  1530   PauseSummary* summary;
  1531   if (abandoned) {
  1532     summary = _abandoned_summary;
  1533   } else {
  1534     summary = _summary;
  1537   double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
  1538   double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
  1539   double scan_only_time = avg_value(_par_last_scan_only_times_ms);
  1540   double scan_only_regions_scanned =
  1541     sum_of_values(_par_last_scan_only_regions_scanned);
  1542   double update_rs_time = avg_value(_par_last_update_rs_times_ms);
  1543   double update_rs_processed_buffers =
  1544     sum_of_values(_par_last_update_rs_processed_buffers);
  1545   double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
  1546   double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
  1547   double termination_time = avg_value(_par_last_termination_times_ms);
  1549   double parallel_other_time = _cur_collection_par_time_ms -
  1550     (update_rs_time + ext_root_scan_time + mark_stack_scan_time +
  1551      scan_only_time + scan_rs_time + obj_copy_time + termination_time);
  1552   if (update_stats) {
  1553     MainBodySummary* body_summary = summary->main_body_summary();
  1554     guarantee(body_summary != NULL, "should not be null!");
  1556     if (_satb_drain_time_set)
  1557       body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
  1558     else
  1559       body_summary->record_satb_drain_time_ms(0.0);
  1560     body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
  1561     body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
  1562     body_summary->record_scan_only_time_ms(scan_only_time);
  1563     body_summary->record_update_rs_time_ms(update_rs_time);
  1564     body_summary->record_scan_rs_time_ms(scan_rs_time);
  1565     body_summary->record_obj_copy_time_ms(obj_copy_time);
  1566     if (parallel) {
  1567       body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
  1568       body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
  1569       body_summary->record_termination_time_ms(termination_time);
  1570       body_summary->record_parallel_other_time_ms(parallel_other_time);
  1572     body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
  1575   if (G1PolicyVerbose > 1) {
  1576     gclog_or_tty->print_cr("      ET: %10.6f ms           (avg: %10.6f ms)\n"
  1577                            "        CH Strong: %10.6f ms    (avg: %10.6f ms)\n"
  1578                            "        G1 Strong: %10.6f ms    (avg: %10.6f ms)\n"
  1579                            "        Evac:      %10.6f ms    (avg: %10.6f ms)\n"
  1580                            "       ET-RS:  %10.6f ms      (avg: %10.6f ms)\n"
  1581                            "      |RS|: " SIZE_FORMAT,
  1582                            elapsed_ms, recent_avg_time_for_pauses_ms(),
  1583                            _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(),
  1584                            _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(),
  1585                            evac_ms, recent_avg_time_for_evac_ms(),
  1586                            scan_rs_time,
  1587                            recent_avg_time_for_pauses_ms() -
  1588                            recent_avg_time_for_G1_strong_ms(),
  1589                            rs_size);
  1591     gclog_or_tty->print_cr("       Used at start: " SIZE_FORMAT"K"
  1592                            "       At end " SIZE_FORMAT "K\n"
  1593                            "       garbage      : " SIZE_FORMAT "K"
  1594                            "       of     " SIZE_FORMAT "K\n"
  1595                            "       survival     : %6.2f%%  (%6.2f%% avg)",
  1596                            _cur_collection_pause_used_at_start_bytes/K,
  1597                            _g1->used()/K, freed_bytes/K,
  1598                            _collection_set_bytes_used_before/K,
  1599                            survival_fraction*100.0,
  1600                            recent_avg_survival_fraction()*100.0);
  1601     gclog_or_tty->print_cr("       Recent %% gc pause time: %6.2f",
  1602                            recent_avg_pause_time_ratio() * 100.0);
  1605   double other_time_ms = elapsed_ms;
  1607   if (!abandoned) {
  1608     if (_satb_drain_time_set)
  1609       other_time_ms -= _cur_satb_drain_time_ms;
  1611     if (parallel)
  1612       other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms;
  1613     else
  1614       other_time_ms -=
  1615         update_rs_time +
  1616         ext_root_scan_time + mark_stack_scan_time + scan_only_time +
  1617         scan_rs_time + obj_copy_time;
  1620   if (PrintGCDetails) {
  1621     gclog_or_tty->print_cr("%s%s, %1.8lf secs]",
  1622                            abandoned ? " (abandoned)" : "",
  1623                            (last_pause_included_initial_mark) ? " (initial-mark)" : "",
  1624                            elapsed_ms / 1000.0);
  1626     if (!abandoned) {
  1627       if (_satb_drain_time_set) {
  1628         print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
  1630       if (_last_satb_drain_processed_buffers >= 0) {
  1631         print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
  1633       if (parallel) {
  1634         print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
  1635         print_par_stats(2, "Update RS (Start)", _par_last_update_rs_start_times_ms, false);
  1636         print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
  1637         if (G1RSBarrierUseQueue)
  1638           print_par_buffers(3, "Processed Buffers",
  1639                             _par_last_update_rs_processed_buffers, true);
  1640         print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
  1641         print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
  1642         print_par_stats(2, "Scan-Only Scanning", _par_last_scan_only_times_ms);
  1643         print_par_buffers(3, "Scan-Only Regions",
  1644                           _par_last_scan_only_regions_scanned, true);
  1645         print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
  1646         print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
  1647         print_par_stats(2, "Termination", _par_last_termination_times_ms);
  1648         print_stats(2, "Other", parallel_other_time);
  1649         print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
  1650       } else {
  1651         print_stats(1, "Update RS", update_rs_time);
  1652         if (G1RSBarrierUseQueue)
  1653           print_stats(2, "Processed Buffers",
  1654                       (int)update_rs_processed_buffers);
  1655         print_stats(1, "Ext Root Scanning", ext_root_scan_time);
  1656         print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
  1657         print_stats(1, "Scan-Only Scanning", scan_only_time);
  1658         print_stats(1, "Scan RS", scan_rs_time);
  1659         print_stats(1, "Object Copying", obj_copy_time);
  1662     print_stats(1, "Other", other_time_ms);
  1663     for (int i = 0; i < _aux_num; ++i) {
  1664       if (_cur_aux_times_set[i]) {
  1665         char buffer[96];
  1666         sprintf(buffer, "Aux%d", i);
  1667         print_stats(1, buffer, _cur_aux_times_ms[i]);
  1671   if (PrintGCDetails)
  1672     gclog_or_tty->print("   [");
  1673   if (PrintGC || PrintGCDetails)
  1674     _g1->print_size_transition(gclog_or_tty,
  1675                                _cur_collection_pause_used_at_start_bytes,
  1676                                _g1->used(), _g1->capacity());
  1677   if (PrintGCDetails)
  1678     gclog_or_tty->print_cr("]");
  1680   _all_pause_times_ms->add(elapsed_ms);
  1681   if (update_stats) {
  1682     summary->record_total_time_ms(elapsed_ms);
  1683     summary->record_other_time_ms(other_time_ms);
  1685   for (int i = 0; i < _aux_num; ++i)
  1686     if (_cur_aux_times_set[i])
  1687       _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
  1689   // Reset marks-between-pauses counter.
  1690   _n_marks_since_last_pause = 0;
  1692   // Update the efficiency-since-mark vars.
  1693   double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
  1694   if (elapsed_ms < MIN_TIMER_GRANULARITY) {
  1695     // This usually happens due to the timer not having the required
  1696     // granularity. Some Linuxes are the usual culprits.
  1697     // We'll just set it to something (arbitrarily) small.
  1698     proc_ms = 1.0;
  1700   double cur_efficiency = (double) freed_bytes / proc_ms;
  1702   bool new_in_marking_window = _in_marking_window;
  1703   bool new_in_marking_window_im = false;
  1704   if (_should_initiate_conc_mark) {
  1705     new_in_marking_window = true;
  1706     new_in_marking_window_im = true;
  1709   if (in_young_gc_mode()) {
  1710     if (_last_full_young_gc) {
  1711       set_full_young_gcs(false);
  1712       _last_full_young_gc = false;
  1715     if ( !_last_young_gc_full ) {
  1716       if ( _should_revert_to_full_young_gcs ||
  1717            _known_garbage_ratio < 0.05 ||
  1718            (adaptive_young_list_length() &&
  1719            (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) {
  1720         set_full_young_gcs(true);
  1723     _should_revert_to_full_young_gcs = false;
  1725     if (_last_young_gc_full && !_during_marking)
  1726       _young_gc_eff_seq->add(cur_efficiency);
  1729   _short_lived_surv_rate_group->start_adding_regions();
  1730   // do that for any other surv rate groupsx
  1732   // <NEW PREDICTION>
  1734   if (update_stats) {
  1735     double pause_time_ms = elapsed_ms;
  1737     size_t diff = 0;
  1738     if (_max_pending_cards >= _pending_cards)
  1739       diff = _max_pending_cards - _pending_cards;
  1740     _pending_card_diff_seq->add((double) diff);
  1742     double cost_per_card_ms = 0.0;
  1743     if (_pending_cards > 0) {
  1744       cost_per_card_ms = update_rs_time / (double) _pending_cards;
  1745       _cost_per_card_ms_seq->add(cost_per_card_ms);
  1748     double cost_per_scan_only_region_ms = 0.0;
  1749     if (scan_only_regions_scanned > 0.0) {
  1750       cost_per_scan_only_region_ms =
  1751         scan_only_time / scan_only_regions_scanned;
  1752       if (_in_marking_window_im)
  1753         _cost_per_scan_only_region_ms_during_cm_seq->add(cost_per_scan_only_region_ms);
  1754       else
  1755         _cost_per_scan_only_region_ms_seq->add(cost_per_scan_only_region_ms);
  1758     size_t cards_scanned = _g1->cards_scanned();
  1760     double cost_per_entry_ms = 0.0;
  1761     if (cards_scanned > 10) {
  1762       cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
  1763       if (_last_young_gc_full)
  1764         _cost_per_entry_ms_seq->add(cost_per_entry_ms);
  1765       else
  1766         _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
  1769     if (_max_rs_lengths > 0) {
  1770       double cards_per_entry_ratio =
  1771         (double) cards_scanned / (double) _max_rs_lengths;
  1772       if (_last_young_gc_full)
  1773         _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
  1774       else
  1775         _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
  1778     size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
  1779     if (rs_length_diff >= 0)
  1780       _rs_length_diff_seq->add((double) rs_length_diff);
  1782     size_t copied_bytes = surviving_bytes;
  1783     double cost_per_byte_ms = 0.0;
  1784     if (copied_bytes > 0) {
  1785       cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
  1786       if (_in_marking_window)
  1787         _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
  1788       else
  1789         _cost_per_byte_ms_seq->add(cost_per_byte_ms);
  1792     double all_other_time_ms = pause_time_ms -
  1793       (update_rs_time + scan_only_time + scan_rs_time + obj_copy_time +
  1794        _mark_closure_time_ms + termination_time);
  1796     double young_other_time_ms = 0.0;
  1797     if (_recorded_young_regions > 0) {
  1798       young_other_time_ms =
  1799         _recorded_young_cset_choice_time_ms +
  1800         _recorded_young_free_cset_time_ms;
  1801       _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
  1802                                              (double) _recorded_young_regions);
  1804     double non_young_other_time_ms = 0.0;
  1805     if (_recorded_non_young_regions > 0) {
  1806       non_young_other_time_ms =
  1807         _recorded_non_young_cset_choice_time_ms +
  1808         _recorded_non_young_free_cset_time_ms;
  1810       _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
  1811                                          (double) _recorded_non_young_regions);
  1814     double constant_other_time_ms = all_other_time_ms -
  1815       (young_other_time_ms + non_young_other_time_ms);
  1816     _constant_other_time_ms_seq->add(constant_other_time_ms);
  1818     double survival_ratio = 0.0;
  1819     if (_bytes_in_collection_set_before_gc > 0) {
  1820       survival_ratio = (double) bytes_in_to_space_during_gc() /
  1821         (double) _bytes_in_collection_set_before_gc;
  1824     _pending_cards_seq->add((double) _pending_cards);
  1825     _scanned_cards_seq->add((double) cards_scanned);
  1826     _rs_lengths_seq->add((double) _max_rs_lengths);
  1828     double expensive_region_limit_ms =
  1829       (double) MaxGCPauseMillis - predict_constant_other_time_ms();
  1830     if (expensive_region_limit_ms < 0.0) {
  1831       // this means that the other time was predicted to be longer than
  1832       // than the max pause time
  1833       expensive_region_limit_ms = (double) MaxGCPauseMillis;
  1835     _expensive_region_limit_ms = expensive_region_limit_ms;
  1837     if (PREDICTIONS_VERBOSE) {
  1838       gclog_or_tty->print_cr("");
  1839       gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
  1840                     "REGIONS %d %d %d %d "
  1841                     "PENDING_CARDS %d %d "
  1842                     "CARDS_SCANNED %d %d "
  1843                     "RS_LENGTHS %d %d "
  1844                     "SCAN_ONLY_SCAN %1.6lf %1.6lf "
  1845                     "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
  1846                     "SURVIVAL_RATIO %1.6lf %1.6lf "
  1847                     "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
  1848                     "OTHER_YOUNG %1.6lf %1.6lf "
  1849                     "OTHER_NON_YOUNG %1.6lf %1.6lf "
  1850                     "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
  1851                     "ELAPSED %1.6lf %1.6lf ",
  1852                     _cur_collection_start_sec,
  1853                     (!_last_young_gc_full) ? 2 :
  1854                     (last_pause_included_initial_mark) ? 1 : 0,
  1855                     _recorded_region_num,
  1856                     _recorded_young_regions,
  1857                     _recorded_scan_only_regions,
  1858                     _recorded_non_young_regions,
  1859                     _predicted_pending_cards, _pending_cards,
  1860                     _predicted_cards_scanned, cards_scanned,
  1861                     _predicted_rs_lengths, _max_rs_lengths,
  1862                     _predicted_scan_only_scan_time_ms, scan_only_time,
  1863                     _predicted_rs_update_time_ms, update_rs_time,
  1864                     _predicted_rs_scan_time_ms, scan_rs_time,
  1865                     _predicted_survival_ratio, survival_ratio,
  1866                     _predicted_object_copy_time_ms, obj_copy_time,
  1867                     _predicted_constant_other_time_ms, constant_other_time_ms,
  1868                     _predicted_young_other_time_ms, young_other_time_ms,
  1869                     _predicted_non_young_other_time_ms,
  1870                     non_young_other_time_ms,
  1871                     _vtime_diff_ms, termination_time,
  1872                     _predicted_pause_time_ms, elapsed_ms);
  1875     if (G1PolicyVerbose > 0) {
  1876       gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
  1877                     _predicted_pause_time_ms,
  1878                     (_within_target) ? "within" : "outside",
  1879                     elapsed_ms);
  1884   _in_marking_window = new_in_marking_window;
  1885   _in_marking_window_im = new_in_marking_window_im;
  1886   _free_regions_at_end_of_collection = _g1->free_regions();
  1887   _scan_only_regions_at_end_of_collection = _g1->young_list_length();
  1888   calculate_young_list_min_length();
  1889   calculate_young_list_target_config();
  1891   // </NEW PREDICTION>
  1893   _target_pause_time_ms = -1.0;
  1896 // <NEW PREDICTION>
  1898 double
  1899 G1CollectorPolicy::
  1900 predict_young_collection_elapsed_time_ms(size_t adjustment) {
  1901   guarantee( adjustment == 0 || adjustment == 1, "invariant" );
  1903   G1CollectedHeap* g1h = G1CollectedHeap::heap();
  1904   size_t young_num = g1h->young_list_length();
  1905   if (young_num == 0)
  1906     return 0.0;
  1908   young_num += adjustment;
  1909   size_t pending_cards = predict_pending_cards();
  1910   size_t rs_lengths = g1h->young_list_sampled_rs_lengths() +
  1911                       predict_rs_length_diff();
  1912   size_t card_num;
  1913   if (full_young_gcs())
  1914     card_num = predict_young_card_num(rs_lengths);
  1915   else
  1916     card_num = predict_non_young_card_num(rs_lengths);
  1917   size_t young_byte_size = young_num * HeapRegion::GrainBytes;
  1918   double accum_yg_surv_rate =
  1919     _short_lived_surv_rate_group->accum_surv_rate(adjustment);
  1921   size_t bytes_to_copy =
  1922     (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
  1924   return
  1925     predict_rs_update_time_ms(pending_cards) +
  1926     predict_rs_scan_time_ms(card_num) +
  1927     predict_object_copy_time_ms(bytes_to_copy) +
  1928     predict_young_other_time_ms(young_num) +
  1929     predict_constant_other_time_ms();
  1932 double
  1933 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
  1934   size_t rs_length = predict_rs_length_diff();
  1935   size_t card_num;
  1936   if (full_young_gcs())
  1937     card_num = predict_young_card_num(rs_length);
  1938   else
  1939     card_num = predict_non_young_card_num(rs_length);
  1940   return predict_base_elapsed_time_ms(pending_cards, card_num);
  1943 double
  1944 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
  1945                                                 size_t scanned_cards) {
  1946   return
  1947     predict_rs_update_time_ms(pending_cards) +
  1948     predict_rs_scan_time_ms(scanned_cards) +
  1949     predict_constant_other_time_ms();
  1952 double
  1953 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
  1954                                                   bool young) {
  1955   size_t rs_length = hr->rem_set()->occupied();
  1956   size_t card_num;
  1957   if (full_young_gcs())
  1958     card_num = predict_young_card_num(rs_length);
  1959   else
  1960     card_num = predict_non_young_card_num(rs_length);
  1961   size_t bytes_to_copy = predict_bytes_to_copy(hr);
  1963   double region_elapsed_time_ms =
  1964     predict_rs_scan_time_ms(card_num) +
  1965     predict_object_copy_time_ms(bytes_to_copy);
  1967   if (young)
  1968     region_elapsed_time_ms += predict_young_other_time_ms(1);
  1969   else
  1970     region_elapsed_time_ms += predict_non_young_other_time_ms(1);
  1972   return region_elapsed_time_ms;
  1975 size_t
  1976 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
  1977   size_t bytes_to_copy;
  1978   if (hr->is_marked())
  1979     bytes_to_copy = hr->max_live_bytes();
  1980   else {
  1981     guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
  1982                "invariant" );
  1983     int age = hr->age_in_surv_rate_group();
  1984     double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
  1985     bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
  1988   return bytes_to_copy;
  1991 void
  1992 G1CollectorPolicy::start_recording_regions() {
  1993   _recorded_rs_lengths            = 0;
  1994   _recorded_scan_only_regions     = 0;
  1995   _recorded_young_regions         = 0;
  1996   _recorded_non_young_regions     = 0;
  1998 #if PREDICTIONS_VERBOSE
  1999   _predicted_rs_lengths           = 0;
  2000   _predicted_cards_scanned        = 0;
  2002   _recorded_marked_bytes          = 0;
  2003   _recorded_young_bytes           = 0;
  2004   _predicted_bytes_to_copy        = 0;
  2005 #endif // PREDICTIONS_VERBOSE
  2008 void
  2009 G1CollectorPolicy::record_cset_region(HeapRegion* hr, bool young) {
  2010   if (young) {
  2011     ++_recorded_young_regions;
  2012   } else {
  2013     ++_recorded_non_young_regions;
  2015 #if PREDICTIONS_VERBOSE
  2016   if (young) {
  2017     _recorded_young_bytes += hr->used();
  2018   } else {
  2019     _recorded_marked_bytes += hr->max_live_bytes();
  2021   _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
  2022 #endif // PREDICTIONS_VERBOSE
  2024   size_t rs_length = hr->rem_set()->occupied();
  2025   _recorded_rs_lengths += rs_length;
  2028 void
  2029 G1CollectorPolicy::record_scan_only_regions(size_t scan_only_length) {
  2030   _recorded_scan_only_regions = scan_only_length;
  2033 void
  2034 G1CollectorPolicy::end_recording_regions() {
  2035 #if PREDICTIONS_VERBOSE
  2036   _predicted_pending_cards = predict_pending_cards();
  2037   _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
  2038   if (full_young_gcs())
  2039     _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
  2040   else
  2041     _predicted_cards_scanned +=
  2042       predict_non_young_card_num(_predicted_rs_lengths);
  2043   _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
  2045   _predicted_scan_only_scan_time_ms =
  2046     predict_scan_only_time_ms(_recorded_scan_only_regions);
  2047   _predicted_rs_update_time_ms =
  2048     predict_rs_update_time_ms(_g1->pending_card_num());
  2049   _predicted_rs_scan_time_ms =
  2050     predict_rs_scan_time_ms(_predicted_cards_scanned);
  2051   _predicted_object_copy_time_ms =
  2052     predict_object_copy_time_ms(_predicted_bytes_to_copy);
  2053   _predicted_constant_other_time_ms =
  2054     predict_constant_other_time_ms();
  2055   _predicted_young_other_time_ms =
  2056     predict_young_other_time_ms(_recorded_young_regions);
  2057   _predicted_non_young_other_time_ms =
  2058     predict_non_young_other_time_ms(_recorded_non_young_regions);
  2060   _predicted_pause_time_ms =
  2061     _predicted_scan_only_scan_time_ms +
  2062     _predicted_rs_update_time_ms +
  2063     _predicted_rs_scan_time_ms +
  2064     _predicted_object_copy_time_ms +
  2065     _predicted_constant_other_time_ms +
  2066     _predicted_young_other_time_ms +
  2067     _predicted_non_young_other_time_ms;
  2068 #endif // PREDICTIONS_VERBOSE
  2071 void G1CollectorPolicy::check_if_region_is_too_expensive(double
  2072                                                            predicted_time_ms) {
  2073   // I don't think we need to do this when in young GC mode since
  2074   // marking will be initiated next time we hit the soft limit anyway...
  2075   if (predicted_time_ms > _expensive_region_limit_ms) {
  2076     if (!in_young_gc_mode()) {
  2077         set_full_young_gcs(true);
  2078       _should_initiate_conc_mark = true;
  2079     } else
  2080       // no point in doing another partial one
  2081       _should_revert_to_full_young_gcs = true;
  2085 // </NEW PREDICTION>
  2088 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
  2089                                                double elapsed_ms) {
  2090   _recent_gc_times_ms->add(elapsed_ms);
  2091   _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
  2092   _prev_collection_pause_end_ms = end_time_sec * 1000.0;
  2095 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
  2096   if (_recent_pause_times_ms->num() == 0) return (double) MaxGCPauseMillis;
  2097   else return _recent_pause_times_ms->avg();
  2100 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() {
  2101   if (_recent_CH_strong_roots_times_ms->num() == 0)
  2102     return (double)MaxGCPauseMillis/3.0;
  2103   else return _recent_CH_strong_roots_times_ms->avg();
  2106 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() {
  2107   if (_recent_G1_strong_roots_times_ms->num() == 0)
  2108     return (double)MaxGCPauseMillis/3.0;
  2109   else return _recent_G1_strong_roots_times_ms->avg();
  2112 double G1CollectorPolicy::recent_avg_time_for_evac_ms() {
  2113   if (_recent_evac_times_ms->num() == 0) return (double)MaxGCPauseMillis/3.0;
  2114   else return _recent_evac_times_ms->avg();
  2117 int G1CollectorPolicy::number_of_recent_gcs() {
  2118   assert(_recent_CH_strong_roots_times_ms->num() ==
  2119          _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync");
  2120   assert(_recent_G1_strong_roots_times_ms->num() ==
  2121          _recent_evac_times_ms->num(), "Sequence out of sync");
  2122   assert(_recent_evac_times_ms->num() ==
  2123          _recent_pause_times_ms->num(), "Sequence out of sync");
  2124   assert(_recent_pause_times_ms->num() ==
  2125          _recent_CS_bytes_used_before->num(), "Sequence out of sync");
  2126   assert(_recent_CS_bytes_used_before->num() ==
  2127          _recent_CS_bytes_surviving->num(), "Sequence out of sync");
  2128   return _recent_pause_times_ms->num();
  2131 double G1CollectorPolicy::recent_avg_survival_fraction() {
  2132   return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
  2133                                            _recent_CS_bytes_used_before);
  2136 double G1CollectorPolicy::last_survival_fraction() {
  2137   return last_survival_fraction_work(_recent_CS_bytes_surviving,
  2138                                      _recent_CS_bytes_used_before);
  2141 double
  2142 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
  2143                                                      TruncatedSeq* before) {
  2144   assert(surviving->num() == before->num(), "Sequence out of sync");
  2145   if (before->sum() > 0.0) {
  2146       double recent_survival_rate = surviving->sum() / before->sum();
  2147       // We exempt parallel collection from this check because Alloc Buffer
  2148       // fragmentation can produce negative collections.
  2149       // Further, we're now always doing parallel collection.  But I'm still
  2150       // leaving this here as a placeholder for a more precise assertion later.
  2151       // (DLD, 10/05.)
  2152       assert((true || ParallelGCThreads > 0) ||
  2153              _g1->evacuation_failed() ||
  2154              recent_survival_rate <= 1.0, "Or bad frac");
  2155       return recent_survival_rate;
  2156   } else {
  2157     return 1.0; // Be conservative.
  2161 double
  2162 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
  2163                                                TruncatedSeq* before) {
  2164   assert(surviving->num() == before->num(), "Sequence out of sync");
  2165   if (surviving->num() > 0 && before->last() > 0.0) {
  2166     double last_survival_rate = surviving->last() / before->last();
  2167     // We exempt parallel collection from this check because Alloc Buffer
  2168     // fragmentation can produce negative collections.
  2169     // Further, we're now always doing parallel collection.  But I'm still
  2170     // leaving this here as a placeholder for a more precise assertion later.
  2171     // (DLD, 10/05.)
  2172     assert((true || ParallelGCThreads > 0) ||
  2173            last_survival_rate <= 1.0, "Or bad frac");
  2174     return last_survival_rate;
  2175   } else {
  2176     return 1.0;
  2180 static const int survival_min_obs = 5;
  2181 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
  2182 static const double min_survival_rate = 0.1;
  2184 double
  2185 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
  2186                                                            double latest) {
  2187   double res = avg;
  2188   if (number_of_recent_gcs() < survival_min_obs) {
  2189     res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
  2191   res = MAX2(res, latest);
  2192   res = MAX2(res, min_survival_rate);
  2193   // In the parallel case, LAB fragmentation can produce "negative
  2194   // collections"; so can evac failure.  Cap at 1.0
  2195   res = MIN2(res, 1.0);
  2196   return res;
  2199 size_t G1CollectorPolicy::expansion_amount() {
  2200   if ((int)(recent_avg_pause_time_ratio() * 100.0) > G1GCPercent) {
  2201     // We will double the existing space, or take
  2202     // G1ExpandByPercentOfAvailable % of the available expansion
  2203     // space, whichever is smaller, bounded below by a minimum
  2204     // expansion (unless that's all that's left.)
  2205     const size_t min_expand_bytes = 1*M;
  2206     size_t reserved_bytes = _g1->g1_reserved_obj_bytes();
  2207     size_t committed_bytes = _g1->capacity();
  2208     size_t uncommitted_bytes = reserved_bytes - committed_bytes;
  2209     size_t expand_bytes;
  2210     size_t expand_bytes_via_pct =
  2211       uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
  2212     expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
  2213     expand_bytes = MAX2(expand_bytes, min_expand_bytes);
  2214     expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
  2215     if (G1PolicyVerbose > 1) {
  2216       gclog_or_tty->print("Decided to expand: ratio = %5.2f, "
  2217                  "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n"
  2218                  "                   Answer = %d.\n",
  2219                  recent_avg_pause_time_ratio(),
  2220                  byte_size_in_proper_unit(committed_bytes),
  2221                  proper_unit_for_byte_size(committed_bytes),
  2222                  byte_size_in_proper_unit(uncommitted_bytes),
  2223                  proper_unit_for_byte_size(uncommitted_bytes),
  2224                  byte_size_in_proper_unit(expand_bytes_via_pct),
  2225                  proper_unit_for_byte_size(expand_bytes_via_pct),
  2226                  byte_size_in_proper_unit(expand_bytes),
  2227                  proper_unit_for_byte_size(expand_bytes));
  2229     return expand_bytes;
  2230   } else {
  2231     return 0;
  2235 void G1CollectorPolicy::note_start_of_mark_thread() {
  2236   _mark_thread_startup_sec = os::elapsedTime();
  2239 class CountCSClosure: public HeapRegionClosure {
  2240   G1CollectorPolicy* _g1_policy;
  2241 public:
  2242   CountCSClosure(G1CollectorPolicy* g1_policy) :
  2243     _g1_policy(g1_policy) {}
  2244   bool doHeapRegion(HeapRegion* r) {
  2245     _g1_policy->_bytes_in_collection_set_before_gc += r->used();
  2246     return false;
  2248 };
  2250 void G1CollectorPolicy::count_CS_bytes_used() {
  2251   CountCSClosure cs_closure(this);
  2252   _g1->collection_set_iterate(&cs_closure);
  2255 static void print_indent(int level) {
  2256   for (int j = 0; j < level+1; ++j)
  2257     gclog_or_tty->print("   ");
  2260 void G1CollectorPolicy::print_summary (int level,
  2261                                        const char* str,
  2262                                        NumberSeq* seq) const {
  2263   double sum = seq->sum();
  2264   print_indent(level);
  2265   gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
  2266                 str, sum / 1000.0, seq->avg());
  2269 void G1CollectorPolicy::print_summary_sd (int level,
  2270                                           const char* str,
  2271                                           NumberSeq* seq) const {
  2272   print_summary(level, str, seq);
  2273   print_indent(level + 5);
  2274   gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
  2275                 seq->num(), seq->sd(), seq->maximum());
  2278 void G1CollectorPolicy::check_other_times(int level,
  2279                                         NumberSeq* other_times_ms,
  2280                                         NumberSeq* calc_other_times_ms) const {
  2281   bool should_print = false;
  2283   double max_sum = MAX2(fabs(other_times_ms->sum()),
  2284                         fabs(calc_other_times_ms->sum()));
  2285   double min_sum = MIN2(fabs(other_times_ms->sum()),
  2286                         fabs(calc_other_times_ms->sum()));
  2287   double sum_ratio = max_sum / min_sum;
  2288   if (sum_ratio > 1.1) {
  2289     should_print = true;
  2290     print_indent(level + 1);
  2291     gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
  2294   double max_avg = MAX2(fabs(other_times_ms->avg()),
  2295                         fabs(calc_other_times_ms->avg()));
  2296   double min_avg = MIN2(fabs(other_times_ms->avg()),
  2297                         fabs(calc_other_times_ms->avg()));
  2298   double avg_ratio = max_avg / min_avg;
  2299   if (avg_ratio > 1.1) {
  2300     should_print = true;
  2301     print_indent(level + 1);
  2302     gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
  2305   if (other_times_ms->sum() < -0.01) {
  2306     print_indent(level + 1);
  2307     gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
  2310   if (other_times_ms->avg() < -0.01) {
  2311     print_indent(level + 1);
  2312     gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
  2315   if (calc_other_times_ms->sum() < -0.01) {
  2316     should_print = true;
  2317     print_indent(level + 1);
  2318     gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
  2321   if (calc_other_times_ms->avg() < -0.01) {
  2322     should_print = true;
  2323     print_indent(level + 1);
  2324     gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
  2327   if (should_print)
  2328     print_summary(level, "Other(Calc)", calc_other_times_ms);
  2331 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
  2332   bool parallel = ParallelGCThreads > 0;
  2333   MainBodySummary*    body_summary = summary->main_body_summary();
  2334   if (summary->get_total_seq()->num() > 0) {
  2335     print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
  2336     if (body_summary != NULL) {
  2337       print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
  2338       if (parallel) {
  2339         print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
  2340         print_summary(2, "Update RS", body_summary->get_update_rs_seq());
  2341         print_summary(2, "Ext Root Scanning",
  2342                       body_summary->get_ext_root_scan_seq());
  2343         print_summary(2, "Mark Stack Scanning",
  2344                       body_summary->get_mark_stack_scan_seq());
  2345         print_summary(2, "Scan-Only Scanning",
  2346                       body_summary->get_scan_only_seq());
  2347         print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
  2348         print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
  2349         print_summary(2, "Termination", body_summary->get_termination_seq());
  2350         print_summary(2, "Other", body_summary->get_parallel_other_seq());
  2352           NumberSeq* other_parts[] = {
  2353             body_summary->get_update_rs_seq(),
  2354             body_summary->get_ext_root_scan_seq(),
  2355             body_summary->get_mark_stack_scan_seq(),
  2356             body_summary->get_scan_only_seq(),
  2357             body_summary->get_scan_rs_seq(),
  2358             body_summary->get_obj_copy_seq(),
  2359             body_summary->get_termination_seq()
  2360           };
  2361           NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
  2362                                         7, other_parts);
  2363           check_other_times(2, body_summary->get_parallel_other_seq(),
  2364                             &calc_other_times_ms);
  2366         print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
  2367         print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
  2368       } else {
  2369         print_summary(1, "Update RS", body_summary->get_update_rs_seq());
  2370         print_summary(1, "Ext Root Scanning",
  2371                       body_summary->get_ext_root_scan_seq());
  2372         print_summary(1, "Mark Stack Scanning",
  2373                       body_summary->get_mark_stack_scan_seq());
  2374         print_summary(1, "Scan-Only Scanning",
  2375                       body_summary->get_scan_only_seq());
  2376         print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
  2377         print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
  2380     print_summary(1, "Other", summary->get_other_seq());
  2382       NumberSeq calc_other_times_ms;
  2383       if (body_summary != NULL) {
  2384         // not abandoned
  2385         if (parallel) {
  2386           // parallel
  2387           NumberSeq* other_parts[] = {
  2388             body_summary->get_satb_drain_seq(),
  2389             body_summary->get_parallel_seq(),
  2390             body_summary->get_clear_ct_seq()
  2391           };
  2392           calc_other_times_ms = NumberSeq(summary->get_total_seq(),
  2393                                           3, other_parts);
  2394         } else {
  2395           // serial
  2396           NumberSeq* other_parts[] = {
  2397             body_summary->get_satb_drain_seq(),
  2398             body_summary->get_update_rs_seq(),
  2399             body_summary->get_ext_root_scan_seq(),
  2400             body_summary->get_mark_stack_scan_seq(),
  2401             body_summary->get_scan_only_seq(),
  2402             body_summary->get_scan_rs_seq(),
  2403             body_summary->get_obj_copy_seq()
  2404           };
  2405           calc_other_times_ms = NumberSeq(summary->get_total_seq(),
  2406                                           7, other_parts);
  2408       } else {
  2409         // abandoned
  2410         calc_other_times_ms = NumberSeq();
  2412       check_other_times(1,  summary->get_other_seq(), &calc_other_times_ms);
  2414   } else {
  2415     print_indent(0);
  2416     gclog_or_tty->print_cr("none");
  2418   gclog_or_tty->print_cr("");
  2421 void
  2422 G1CollectorPolicy::print_abandoned_summary(PauseSummary* summary) const {
  2423   bool printed = false;
  2424   if (summary->get_total_seq()->num() > 0) {
  2425     printed = true;
  2426     print_summary(summary);
  2428   if (!printed) {
  2429     print_indent(0);
  2430     gclog_or_tty->print_cr("none");
  2431     gclog_or_tty->print_cr("");
  2435 void G1CollectorPolicy::print_tracing_info() const {
  2436   if (TraceGen0Time) {
  2437     gclog_or_tty->print_cr("ALL PAUSES");
  2438     print_summary_sd(0, "Total", _all_pause_times_ms);
  2439     gclog_or_tty->print_cr("");
  2440     gclog_or_tty->print_cr("");
  2441     gclog_or_tty->print_cr("   Full Young GC Pauses:    %8d", _full_young_pause_num);
  2442     gclog_or_tty->print_cr("   Partial Young GC Pauses: %8d", _partial_young_pause_num);
  2443     gclog_or_tty->print_cr("");
  2445     gclog_or_tty->print_cr("EVACUATION PAUSES");
  2446     print_summary(_summary);
  2448     gclog_or_tty->print_cr("ABANDONED PAUSES");
  2449     print_abandoned_summary(_abandoned_summary);
  2451     gclog_or_tty->print_cr("MISC");
  2452     print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
  2453     print_summary_sd(0, "Yields", _all_yield_times_ms);
  2454     for (int i = 0; i < _aux_num; ++i) {
  2455       if (_all_aux_times_ms[i].num() > 0) {
  2456         char buffer[96];
  2457         sprintf(buffer, "Aux%d", i);
  2458         print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
  2462     size_t all_region_num = _region_num_young + _region_num_tenured;
  2463     gclog_or_tty->print_cr("   New Regions %8d, Young %8d (%6.2lf%%), "
  2464                "Tenured %8d (%6.2lf%%)",
  2465                all_region_num,
  2466                _region_num_young,
  2467                (double) _region_num_young / (double) all_region_num * 100.0,
  2468                _region_num_tenured,
  2469                (double) _region_num_tenured / (double) all_region_num * 100.0);
  2471     if (!G1RSBarrierUseQueue) {
  2472       gclog_or_tty->print_cr("Of %d times conc refinement was enabled, %d (%7.2f%%) "
  2473                     "did zero traversals.",
  2474                     _conc_refine_enabled, _conc_refine_zero_traversals,
  2475                     _conc_refine_enabled > 0 ?
  2476                     100.0 * (float)_conc_refine_zero_traversals/
  2477                     (float)_conc_refine_enabled : 0.0);
  2478       gclog_or_tty->print_cr("  Max # of traversals = %d.",
  2479                     _conc_refine_max_traversals);
  2480       gclog_or_tty->print_cr("");
  2483   if (TraceGen1Time) {
  2484     if (_all_full_gc_times_ms->num() > 0) {
  2485       gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
  2486                  _all_full_gc_times_ms->num(),
  2487                  _all_full_gc_times_ms->sum() / 1000.0);
  2488       gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
  2489       gclog_or_tty->print_cr("                     [std. dev = %8.2f ms, max = %8.2f ms]",
  2490                     _all_full_gc_times_ms->sd(),
  2491                     _all_full_gc_times_ms->maximum());
  2496 void G1CollectorPolicy::print_yg_surv_rate_info() const {
  2497 #ifndef PRODUCT
  2498   _short_lived_surv_rate_group->print_surv_rate_summary();
  2499   // add this call for any other surv rate groups
  2500 #endif // PRODUCT
  2503 void G1CollectorPolicy::update_conc_refine_data() {
  2504   unsigned traversals = _g1->concurrent_g1_refine()->disable();
  2505   if (traversals == 0) _conc_refine_zero_traversals++;
  2506   _conc_refine_max_traversals = MAX2(_conc_refine_max_traversals,
  2507                                      (size_t)traversals);
  2509   if (G1PolicyVerbose > 1)
  2510     gclog_or_tty->print_cr("Did a CR traversal series: %d traversals.", traversals);
  2511   double multiplier = 1.0;
  2512   if (traversals == 0) {
  2513     multiplier = 4.0;
  2514   } else if (traversals > (size_t)G1ConcRefineTargTraversals) {
  2515     multiplier = 1.0/1.5;
  2516   } else if (traversals < (size_t)G1ConcRefineTargTraversals) {
  2517     multiplier = 1.5;
  2519   if (G1PolicyVerbose > 1) {
  2520     gclog_or_tty->print_cr("  Multiplier = %7.2f.", multiplier);
  2521     gclog_or_tty->print("  Delta went from %d regions to ",
  2522                _conc_refine_current_delta);
  2524   _conc_refine_current_delta =
  2525     MIN2(_g1->n_regions(),
  2526          (size_t)(_conc_refine_current_delta * multiplier));
  2527   _conc_refine_current_delta =
  2528     MAX2(_conc_refine_current_delta, (size_t)1);
  2529   if (G1PolicyVerbose > 1) {
  2530     gclog_or_tty->print_cr("%d regions.", _conc_refine_current_delta);
  2532   _conc_refine_enabled++;
  2535 bool
  2536 G1CollectorPolicy::should_add_next_region_to_young_list() {
  2537   assert(in_young_gc_mode(), "should be in young GC mode");
  2538   bool ret;
  2539   size_t young_list_length = _g1->young_list_length();
  2540   size_t young_list_max_length = _young_list_target_length;
  2541   if (G1FixedEdenSize) {
  2542     young_list_max_length -= _max_survivor_regions;
  2544   if (young_list_length < young_list_max_length) {
  2545     ret = true;
  2546     ++_region_num_young;
  2547   } else {
  2548     ret = false;
  2549     ++_region_num_tenured;
  2552   return ret;
  2555 #ifndef PRODUCT
  2556 // for debugging, bit of a hack...
  2557 static char*
  2558 region_num_to_mbs(int length) {
  2559   static char buffer[64];
  2560   double bytes = (double) (length * HeapRegion::GrainBytes);
  2561   double mbs = bytes / (double) (1024 * 1024);
  2562   sprintf(buffer, "%7.2lfMB", mbs);
  2563   return buffer;
  2565 #endif // PRODUCT
  2567 void
  2568 G1CollectorPolicy::checkpoint_conc_overhead() {
  2569   double conc_overhead = 0.0;
  2570   if (G1AccountConcurrentOverhead)
  2571     conc_overhead = COTracker::totalPredConcOverhead();
  2572   _mmu_tracker->update_conc_overhead(conc_overhead);
  2573 #if 0
  2574   gclog_or_tty->print(" CO %1.4lf TARGET %1.4lf",
  2575              conc_overhead, _mmu_tracker->max_gc_time());
  2576 #endif
  2580 size_t G1CollectorPolicy::max_regions(int purpose) {
  2581   switch (purpose) {
  2582     case GCAllocForSurvived:
  2583       return _max_survivor_regions;
  2584     case GCAllocForTenured:
  2585       return REGIONS_UNLIMITED;
  2586     default:
  2587       ShouldNotReachHere();
  2588       return REGIONS_UNLIMITED;
  2589   };
  2592 // Calculates survivor space parameters.
  2593 void G1CollectorPolicy::calculate_survivors_policy()
  2595   if (!G1UseSurvivorSpaces) {
  2596     return;
  2598   if (G1FixedSurvivorSpaceSize == 0) {
  2599     _max_survivor_regions = _young_list_target_length / SurvivorRatio;
  2600   } else {
  2601     _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
  2604   if (G1FixedTenuringThreshold) {
  2605     _tenuring_threshold = MaxTenuringThreshold;
  2606   } else {
  2607     _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
  2608         HeapRegion::GrainWords * _max_survivor_regions);
  2612 bool
  2613 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t
  2614                                                                word_size) {
  2615   assert(_g1->regions_accounted_for(), "Region leakage!");
  2616   // Initiate a pause when we reach the steady-state "used" target.
  2617   size_t used_hard = (_g1->capacity() / 100) * G1SteadyStateUsed;
  2618   size_t used_soft =
  2619    MAX2((_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta),
  2620         used_hard/2);
  2621   size_t used = _g1->used();
  2623   double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
  2625   size_t young_list_length = _g1->young_list_length();
  2626   size_t young_list_max_length = _young_list_target_length;
  2627   if (G1FixedEdenSize) {
  2628     young_list_max_length -= _max_survivor_regions;
  2630   bool reached_target_length = young_list_length >= young_list_max_length;
  2632   if (in_young_gc_mode()) {
  2633     if (reached_target_length) {
  2634       assert( young_list_length > 0 && _g1->young_list_length() > 0,
  2635               "invariant" );
  2636       _target_pause_time_ms = max_pause_time_ms;
  2637       return true;
  2639   } else {
  2640     guarantee( false, "should not reach here" );
  2643   return false;
  2646 #ifndef PRODUCT
  2647 class HRSortIndexIsOKClosure: public HeapRegionClosure {
  2648   CollectionSetChooser* _chooser;
  2649 public:
  2650   HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
  2651     _chooser(chooser) {}
  2653   bool doHeapRegion(HeapRegion* r) {
  2654     if (!r->continuesHumongous()) {
  2655       assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
  2657     return false;
  2659 };
  2661 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() {
  2662   HRSortIndexIsOKClosure cl(_collectionSetChooser);
  2663   _g1->heap_region_iterate(&cl);
  2664   return true;
  2666 #endif
  2668 void
  2669 G1CollectorPolicy_BestRegionsFirst::
  2670 record_collection_pause_start(double start_time_sec, size_t start_used) {
  2671   G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used);
  2674 class NextNonCSElemFinder: public HeapRegionClosure {
  2675   HeapRegion* _res;
  2676 public:
  2677   NextNonCSElemFinder(): _res(NULL) {}
  2678   bool doHeapRegion(HeapRegion* r) {
  2679     if (!r->in_collection_set()) {
  2680       _res = r;
  2681       return true;
  2682     } else {
  2683       return false;
  2686   HeapRegion* res() { return _res; }
  2687 };
  2689 class KnownGarbageClosure: public HeapRegionClosure {
  2690   CollectionSetChooser* _hrSorted;
  2692 public:
  2693   KnownGarbageClosure(CollectionSetChooser* hrSorted) :
  2694     _hrSorted(hrSorted)
  2695   {}
  2697   bool doHeapRegion(HeapRegion* r) {
  2698     // We only include humongous regions in collection
  2699     // sets when concurrent mark shows that their contained object is
  2700     // unreachable.
  2702     // Do we have any marking information for this region?
  2703     if (r->is_marked()) {
  2704       // We don't include humongous regions in collection
  2705       // sets because we collect them immediately at the end of a marking
  2706       // cycle.  We also don't include young regions because we *must*
  2707       // include them in the next collection pause.
  2708       if (!r->isHumongous() && !r->is_young()) {
  2709         _hrSorted->addMarkedHeapRegion(r);
  2712     return false;
  2714 };
  2716 class ParKnownGarbageHRClosure: public HeapRegionClosure {
  2717   CollectionSetChooser* _hrSorted;
  2718   jint _marked_regions_added;
  2719   jint _chunk_size;
  2720   jint _cur_chunk_idx;
  2721   jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
  2722   int _worker;
  2723   int _invokes;
  2725   void get_new_chunk() {
  2726     _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
  2727     _cur_chunk_end = _cur_chunk_idx + _chunk_size;
  2729   void add_region(HeapRegion* r) {
  2730     if (_cur_chunk_idx == _cur_chunk_end) {
  2731       get_new_chunk();
  2733     assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
  2734     _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
  2735     _marked_regions_added++;
  2736     _cur_chunk_idx++;
  2739 public:
  2740   ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
  2741                            jint chunk_size,
  2742                            int worker) :
  2743     _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
  2744     _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
  2745     _invokes(0)
  2746   {}
  2748   bool doHeapRegion(HeapRegion* r) {
  2749     // We only include humongous regions in collection
  2750     // sets when concurrent mark shows that their contained object is
  2751     // unreachable.
  2752     _invokes++;
  2754     // Do we have any marking information for this region?
  2755     if (r->is_marked()) {
  2756       // We don't include humongous regions in collection
  2757       // sets because we collect them immediately at the end of a marking
  2758       // cycle.
  2759       // We also do not include young regions in collection sets
  2760       if (!r->isHumongous() && !r->is_young()) {
  2761         add_region(r);
  2764     return false;
  2766   jint marked_regions_added() { return _marked_regions_added; }
  2767   int invokes() { return _invokes; }
  2768 };
  2770 class ParKnownGarbageTask: public AbstractGangTask {
  2771   CollectionSetChooser* _hrSorted;
  2772   jint _chunk_size;
  2773   G1CollectedHeap* _g1;
  2774 public:
  2775   ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
  2776     AbstractGangTask("ParKnownGarbageTask"),
  2777     _hrSorted(hrSorted), _chunk_size(chunk_size),
  2778     _g1(G1CollectedHeap::heap())
  2779   {}
  2781   void work(int i) {
  2782     ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
  2783     // Back to zero for the claim value.
  2784     _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
  2785                                          HeapRegion::InitialClaimValue);
  2786     jint regions_added = parKnownGarbageCl.marked_regions_added();
  2787     _hrSorted->incNumMarkedHeapRegions(regions_added);
  2788     if (G1PrintParCleanupStats) {
  2789       gclog_or_tty->print("     Thread %d called %d times, added %d regions to list.\n",
  2790                  i, parKnownGarbageCl.invokes(), regions_added);
  2793 };
  2795 void
  2796 G1CollectorPolicy_BestRegionsFirst::
  2797 record_concurrent_mark_cleanup_end(size_t freed_bytes,
  2798                                    size_t max_live_bytes) {
  2799   double start;
  2800   if (G1PrintParCleanupStats) start = os::elapsedTime();
  2801   record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
  2803   _collectionSetChooser->clearMarkedHeapRegions();
  2804   double clear_marked_end;
  2805   if (G1PrintParCleanupStats) {
  2806     clear_marked_end = os::elapsedTime();
  2807     gclog_or_tty->print_cr("  clear marked regions + work1: %8.3f ms.",
  2808                   (clear_marked_end - start)*1000.0);
  2810   if (ParallelGCThreads > 0) {
  2811     const size_t OverpartitionFactor = 4;
  2812     const size_t MinChunkSize = 8;
  2813     const size_t ChunkSize =
  2814       MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
  2815            MinChunkSize);
  2816     _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
  2817                                                              ChunkSize);
  2818     ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
  2819                                             (int) ChunkSize);
  2820     _g1->workers()->run_task(&parKnownGarbageTask);
  2822     assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
  2823            "sanity check");
  2824   } else {
  2825     KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
  2826     _g1->heap_region_iterate(&knownGarbagecl);
  2828   double known_garbage_end;
  2829   if (G1PrintParCleanupStats) {
  2830     known_garbage_end = os::elapsedTime();
  2831     gclog_or_tty->print_cr("  compute known garbage: %8.3f ms.",
  2832                   (known_garbage_end - clear_marked_end)*1000.0);
  2834   _collectionSetChooser->sortMarkedHeapRegions();
  2835   double sort_end;
  2836   if (G1PrintParCleanupStats) {
  2837     sort_end = os::elapsedTime();
  2838     gclog_or_tty->print_cr("  sorting: %8.3f ms.",
  2839                   (sort_end - known_garbage_end)*1000.0);
  2842   record_concurrent_mark_cleanup_end_work2();
  2843   double work2_end;
  2844   if (G1PrintParCleanupStats) {
  2845     work2_end = os::elapsedTime();
  2846     gclog_or_tty->print_cr("  work2: %8.3f ms.",
  2847                   (work2_end - sort_end)*1000.0);
  2851 // Add the heap region to the collection set and return the conservative
  2852 // estimate of the number of live bytes.
  2853 void G1CollectorPolicy::
  2854 add_to_collection_set(HeapRegion* hr) {
  2855   if (G1PrintRegions) {
  2856     gclog_or_tty->print_cr("added region to cset %d:["PTR_FORMAT", "PTR_FORMAT"], "
  2857                   "top "PTR_FORMAT", young %s",
  2858                   hr->hrs_index(), hr->bottom(), hr->end(),
  2859                   hr->top(), (hr->is_young()) ? "YES" : "NO");
  2862   if (_g1->mark_in_progress())
  2863     _g1->concurrent_mark()->registerCSetRegion(hr);
  2865   assert(!hr->in_collection_set(),
  2866               "should not already be in the CSet");
  2867   hr->set_in_collection_set(true);
  2868   hr->set_next_in_collection_set(_collection_set);
  2869   _collection_set = hr;
  2870   _collection_set_size++;
  2871   _collection_set_bytes_used_before += hr->used();
  2872   _g1->register_region_with_in_cset_fast_test(hr);
  2875 void
  2876 G1CollectorPolicy_BestRegionsFirst::
  2877 choose_collection_set() {
  2878   double non_young_start_time_sec;
  2879   start_recording_regions();
  2881   guarantee(_target_pause_time_ms > -1.0,
  2882             "_target_pause_time_ms should have been set!");
  2883   assert(_collection_set == NULL, "Precondition");
  2885   double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
  2886   double predicted_pause_time_ms = base_time_ms;
  2888   double target_time_ms = _target_pause_time_ms;
  2889   double time_remaining_ms = target_time_ms - base_time_ms;
  2891   // the 10% and 50% values are arbitrary...
  2892   if (time_remaining_ms < 0.10*target_time_ms) {
  2893     time_remaining_ms = 0.50 * target_time_ms;
  2894     _within_target = false;
  2895   } else {
  2896     _within_target = true;
  2899   // We figure out the number of bytes available for future to-space.
  2900   // For new regions without marking information, we must assume the
  2901   // worst-case of complete survival.  If we have marking information for a
  2902   // region, we can bound the amount of live data.  We can add a number of
  2903   // such regions, as long as the sum of the live data bounds does not
  2904   // exceed the available evacuation space.
  2905   size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes;
  2907   size_t expansion_bytes =
  2908     _g1->expansion_regions() * HeapRegion::GrainBytes;
  2910   _collection_set_bytes_used_before = 0;
  2911   _collection_set_size = 0;
  2913   // Adjust for expansion and slop.
  2914   max_live_bytes = max_live_bytes + expansion_bytes;
  2916   assert(_g1->regions_accounted_for(), "Region leakage!");
  2918   HeapRegion* hr;
  2919   if (in_young_gc_mode()) {
  2920     double young_start_time_sec = os::elapsedTime();
  2922     if (G1PolicyVerbose > 0) {
  2923       gclog_or_tty->print_cr("Adding %d young regions to the CSet",
  2924                     _g1->young_list_length());
  2926     _young_cset_length  = 0;
  2927     _last_young_gc_full = full_young_gcs() ? true : false;
  2928     if (_last_young_gc_full)
  2929       ++_full_young_pause_num;
  2930     else
  2931       ++_partial_young_pause_num;
  2932     hr = _g1->pop_region_from_young_list();
  2933     while (hr != NULL) {
  2935       assert( hr->young_index_in_cset() == -1, "invariant" );
  2936       assert( hr->age_in_surv_rate_group() != -1, "invariant" );
  2937       hr->set_young_index_in_cset((int) _young_cset_length);
  2939       ++_young_cset_length;
  2940       double predicted_time_ms = predict_region_elapsed_time_ms(hr, true);
  2941       time_remaining_ms -= predicted_time_ms;
  2942       predicted_pause_time_ms += predicted_time_ms;
  2943       assert(!hr->in_collection_set(), "invariant");
  2944       add_to_collection_set(hr);
  2945       record_cset_region(hr, true);
  2946       max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
  2947       if (G1PolicyVerbose > 0) {
  2948         gclog_or_tty->print_cr("  Added [" PTR_FORMAT ", " PTR_FORMAT") to CS.",
  2949                       hr->bottom(), hr->end());
  2950         gclog_or_tty->print_cr("    (" SIZE_FORMAT " KB left in heap.)",
  2951                       max_live_bytes/K);
  2953       hr = _g1->pop_region_from_young_list();
  2956     record_scan_only_regions(_g1->young_list_scan_only_length());
  2958     double young_end_time_sec = os::elapsedTime();
  2959     _recorded_young_cset_choice_time_ms =
  2960       (young_end_time_sec - young_start_time_sec) * 1000.0;
  2962     non_young_start_time_sec = os::elapsedTime();
  2964     if (_young_cset_length > 0 && _last_young_gc_full) {
  2965       // don't bother adding more regions...
  2966       goto choose_collection_set_end;
  2970   if (!in_young_gc_mode() || !full_young_gcs()) {
  2971     bool should_continue = true;
  2972     NumberSeq seq;
  2973     double avg_prediction = 100000000000000000.0; // something very large
  2974     do {
  2975       hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
  2976                                                       avg_prediction);
  2977       if (hr != NULL) {
  2978         double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
  2979         time_remaining_ms -= predicted_time_ms;
  2980         predicted_pause_time_ms += predicted_time_ms;
  2981         add_to_collection_set(hr);
  2982         record_cset_region(hr, false);
  2983         max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
  2984         if (G1PolicyVerbose > 0) {
  2985           gclog_or_tty->print_cr("    (" SIZE_FORMAT " KB left in heap.)",
  2986                         max_live_bytes/K);
  2988         seq.add(predicted_time_ms);
  2989         avg_prediction = seq.avg() + seq.sd();
  2991       should_continue =
  2992         ( hr != NULL) &&
  2993         ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0
  2994           : _collection_set_size < _young_list_fixed_length );
  2995     } while (should_continue);
  2997     if (!adaptive_young_list_length() &&
  2998         _collection_set_size < _young_list_fixed_length)
  2999       _should_revert_to_full_young_gcs  = true;
  3002 choose_collection_set_end:
  3003   count_CS_bytes_used();
  3005   end_recording_regions();
  3007   double non_young_end_time_sec = os::elapsedTime();
  3008   _recorded_non_young_cset_choice_time_ms =
  3009     (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
  3012 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() {
  3013   G1CollectorPolicy::record_full_collection_end();
  3014   _collectionSetChooser->updateAfterFullCollection();
  3017 void G1CollectorPolicy_BestRegionsFirst::
  3018 expand_if_possible(size_t numRegions) {
  3019   size_t expansion_bytes = numRegions * HeapRegion::GrainBytes;
  3020   _g1->expand(expansion_bytes);
  3023 void G1CollectorPolicy_BestRegionsFirst::
  3024 record_collection_pause_end(bool abandoned) {
  3025   G1CollectorPolicy::record_collection_pause_end(abandoned);
  3026   assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
  3029 // Local Variables: ***
  3030 // c-indentation-style: gnu ***
  3031 // End: ***

mercurial