1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/concurrentG1Refine.hpp" 27 #include "gc/g1/concurrentMark.hpp" 28 #include "gc/g1/concurrentMarkThread.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorPolicy.hpp" 31 #include "gc/g1/g1IHOPControl.hpp" 32 #include "gc/g1/g1ErgoVerbose.hpp" 33 #include "gc/g1/g1GCPhaseTimes.hpp" 34 #include "gc/g1/g1Log.hpp" 35 #include "gc/g1/heapRegion.inline.hpp" 36 #include "gc/g1/heapRegionRemSet.hpp" 37 #include "gc/shared/gcPolicyCounters.hpp" 38 #include "runtime/arguments.hpp" 39 #include "runtime/java.hpp" 40 #include "runtime/mutexLocker.hpp" 41 #include "utilities/debug.hpp" 42 43 // Different defaults for different number of GC threads 44 // They were chosen by running GCOld and SPECjbb on debris with different 45 // numbers of GC threads and choosing them based on the results 46 47 // all the same 48 static double rs_length_diff_defaults[] = { 49 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 50 }; 51 52 static double cost_per_card_ms_defaults[] = { 53 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 54 }; 55 56 // all the same 57 static double young_cards_per_entry_ratio_defaults[] = { 58 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 59 }; 60 61 static double cost_per_entry_ms_defaults[] = { 62 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 63 }; 64 65 static double cost_per_byte_ms_defaults[] = { 66 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 67 }; 68 69 // these should be pretty consistent 70 static double constant_other_time_ms_defaults[] = { 71 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 72 }; 73 74 75 static double young_other_cost_per_region_ms_defaults[] = { 76 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 77 }; 78 79 static double non_young_other_cost_per_region_ms_defaults[] = { 80 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 81 }; 82 83 G1CollectorPolicy::G1CollectorPolicy() : 84 _predictor(G1ConfidencePercent / 100.0), 85 _parallel_gc_threads(ParallelGCThreads), 86 87 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 88 _stop_world_start(0.0), 89 90 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 92 93 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 94 _prev_collection_pause_end_ms(0.0), 95 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 106 _non_young_other_cost_per_region_ms_seq( 107 new TruncatedSeq(TruncatedSeqLength)), 108 109 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 110 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 111 112 _pause_time_target_ms((double) MaxGCPauseMillis), 113 114 _recent_prev_end_times_for_all_gcs_sec( 115 new TruncatedSeq(NumPrevPausesForHeuristics)), 116 117 _recent_avg_pause_time_ratio(0.0), 118 _rs_lengths_prediction(0), 119 _max_survivor_regions(0), 120 121 _eden_used_bytes_before_gc(0), 122 _survivor_used_bytes_before_gc(0), 123 _heap_used_bytes_before_gc(0), 124 _metaspace_used_bytes_before_gc(0), 125 _eden_capacity_bytes_before_gc(0), 126 _heap_capacity_bytes_before_gc(0), 127 128 _eden_cset_region_length(0), 129 _survivor_cset_region_length(0), 130 _old_cset_region_length(0), 131 132 _collection_set(NULL), 133 _collection_set_bytes_used_before(0), 134 135 // Incremental CSet attributes 136 _inc_cset_build_state(Inactive), 137 _inc_cset_head(NULL), 138 _inc_cset_tail(NULL), 139 _inc_cset_bytes_used_before(0), 140 _inc_cset_max_finger(NULL), 141 _inc_cset_recorded_rs_lengths(0), 142 _inc_cset_recorded_rs_lengths_diffs(0), 143 _inc_cset_predicted_elapsed_time_ms(0.0), 144 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 145 146 // add here any more surv rate groups 147 _recorded_survivor_regions(0), 148 _recorded_survivor_head(NULL), 149 _recorded_survivor_tail(NULL), 150 _survivors_age_table(true), 151 152 _gc_overhead_perc(0.0), 153 154 _last_old_allocated_bytes(0), 155 _ihop_control(NULL), 156 _initial_mark_to_mixed() { 157 158 // SurvRateGroups below must be initialized after the predictor because they 159 // indirectly use it through this object passed to their constructor. 160 _short_lived_surv_rate_group = 161 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 162 _survivor_surv_rate_group = 163 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 164 165 // Set up the region size and associated fields. Given that the 166 // policy is created before the heap, we have to set this up here, 167 // so it's done as soon as possible. 168 169 // It would have been natural to pass initial_heap_byte_size() and 170 // max_heap_byte_size() to setup_heap_region_size() but those have 171 // not been set up at this point since they should be aligned with 172 // the region size. So, there is a circular dependency here. We base 173 // the region size on the heap size, but the heap size should be 174 // aligned with the region size. To get around this we use the 175 // unaligned values for the heap. 176 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 177 HeapRegionRemSet::setup_remset_size(); 178 179 G1ErgoVerbose::initialize(); 180 if (PrintAdaptiveSizePolicy) { 181 // Currently, we only use a single switch for all the heuristics. 182 G1ErgoVerbose::set_enabled(true); 183 // Given that we don't currently have a verboseness level 184 // parameter, we'll hardcode this to high. This can be easily 185 // changed in the future. 186 G1ErgoVerbose::set_level(ErgoHigh); 187 } else { 188 G1ErgoVerbose::set_enabled(false); 189 } 190 191 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 192 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 193 194 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 195 196 int index = MIN2(_parallel_gc_threads - 1, 7); 197 198 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 199 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 200 _cost_scan_hcc_seq->add(0.0); 201 _young_cards_per_entry_ratio_seq->add( 202 young_cards_per_entry_ratio_defaults[index]); 203 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 204 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 205 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 206 _young_other_cost_per_region_ms_seq->add( 207 young_other_cost_per_region_ms_defaults[index]); 208 _non_young_other_cost_per_region_ms_seq->add( 209 non_young_other_cost_per_region_ms_defaults[index]); 210 211 // Below, we might need to calculate the pause time target based on 212 // the pause interval. When we do so we are going to give G1 maximum 213 // flexibility and allow it to do pauses when it needs to. So, we'll 214 // arrange that the pause interval to be pause time target + 1 to 215 // ensure that a) the pause time target is maximized with respect to 216 // the pause interval and b) we maintain the invariant that pause 217 // time target < pause interval. If the user does not want this 218 // maximum flexibility, they will have to set the pause interval 219 // explicitly. 220 221 // First make sure that, if either parameter is set, its value is 222 // reasonable. 223 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 224 if (MaxGCPauseMillis < 1) { 225 vm_exit_during_initialization("MaxGCPauseMillis should be " 226 "greater than 0"); 227 } 228 } 229 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 230 if (GCPauseIntervalMillis < 1) { 231 vm_exit_during_initialization("GCPauseIntervalMillis should be " 232 "greater than 0"); 233 } 234 } 235 236 // Then, if the pause time target parameter was not set, set it to 237 // the default value. 238 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 239 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 240 // The default pause time target in G1 is 200ms 241 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 242 } else { 243 // We do not allow the pause interval to be set without the 244 // pause time target 245 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 246 "without setting MaxGCPauseMillis"); 247 } 248 } 249 250 // Then, if the interval parameter was not set, set it according to 251 // the pause time target (this will also deal with the case when the 252 // pause time target is the default value). 253 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 254 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 255 } 256 257 // Finally, make sure that the two parameters are consistent. 258 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 259 char buffer[256]; 260 jio_snprintf(buffer, 256, 261 "MaxGCPauseMillis (%u) should be less than " 262 "GCPauseIntervalMillis (%u)", 263 MaxGCPauseMillis, GCPauseIntervalMillis); 264 vm_exit_during_initialization(buffer); 265 } 266 267 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 268 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 269 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 270 271 // start conservatively (around 50ms is about right) 272 _concurrent_mark_remark_times_ms->add(0.05); 273 _concurrent_mark_cleanup_times_ms->add(0.20); 274 _tenuring_threshold = MaxTenuringThreshold; 275 276 assert(GCTimeRatio > 0, 277 "we should have set it to a default value set_g1_gc_flags() " 278 "if a user set it to 0"); 279 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 280 281 uintx reserve_perc = G1ReservePercent; 282 // Put an artificial ceiling on this so that it's not set to a silly value. 283 if (reserve_perc > 50) { 284 reserve_perc = 50; 285 warning("G1ReservePercent is set to a value that is too large, " 286 "it's been updated to " UINTX_FORMAT, reserve_perc); 287 } 288 _reserve_factor = (double) reserve_perc / 100.0; 289 // This will be set when the heap is expanded 290 // for the first time during initialization. 291 _reserve_regions = 0; 292 293 _collectionSetChooser = new CollectionSetChooser(); 294 } 295 296 G1CollectorPolicy::~G1CollectorPolicy() { 297 if (_ihop_control != NULL) { 298 delete _ihop_control; 299 } 300 } 301 302 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 303 return _predictor.get_new_prediction(seq); 304 } 305 306 void G1CollectorPolicy::initialize_alignments() { 307 _space_alignment = HeapRegion::GrainBytes; 308 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 309 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 310 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 311 } 312 313 void G1CollectorPolicy::initialize_flags() { 314 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 315 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 316 } 317 318 if (SurvivorRatio < 1) { 319 vm_exit_during_initialization("Invalid survivor ratio specified"); 320 } 321 CollectorPolicy::initialize_flags(); 322 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 323 } 324 325 void G1CollectorPolicy::post_heap_initialize() { 326 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 327 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 328 if (max_young_size != MaxNewSize) { 329 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 330 } 331 332 _ihop_control = create_ihop_control(); 333 } 334 335 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 336 337 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 338 _min_desired_young_length(0), _max_desired_young_length(0) { 339 if (FLAG_IS_CMDLINE(NewRatio)) { 340 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 341 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 342 } else { 343 _sizer_kind = SizerNewRatio; 344 _adaptive_size = false; 345 return; 346 } 347 } 348 349 if (NewSize > MaxNewSize) { 350 if (FLAG_IS_CMDLINE(MaxNewSize)) { 351 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 352 "A new max generation size of " SIZE_FORMAT "k will be used.", 353 NewSize/K, MaxNewSize/K, NewSize/K); 354 } 355 MaxNewSize = NewSize; 356 } 357 358 if (FLAG_IS_CMDLINE(NewSize)) { 359 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 360 1U); 361 if (FLAG_IS_CMDLINE(MaxNewSize)) { 362 _max_desired_young_length = 363 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 364 1U); 365 _sizer_kind = SizerMaxAndNewSize; 366 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 367 } else { 368 _sizer_kind = SizerNewSizeOnly; 369 } 370 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 371 _max_desired_young_length = 372 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 373 1U); 374 _sizer_kind = SizerMaxNewSizeOnly; 375 } 376 } 377 378 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 379 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 380 return MAX2(1U, default_value); 381 } 382 383 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 384 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 385 return MAX2(1U, default_value); 386 } 387 388 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 389 assert(number_of_heap_regions > 0, "Heap must be initialized"); 390 391 switch (_sizer_kind) { 392 case SizerDefaults: 393 *min_young_length = calculate_default_min_length(number_of_heap_regions); 394 *max_young_length = calculate_default_max_length(number_of_heap_regions); 395 break; 396 case SizerNewSizeOnly: 397 *max_young_length = calculate_default_max_length(number_of_heap_regions); 398 *max_young_length = MAX2(*min_young_length, *max_young_length); 399 break; 400 case SizerMaxNewSizeOnly: 401 *min_young_length = calculate_default_min_length(number_of_heap_regions); 402 *min_young_length = MIN2(*min_young_length, *max_young_length); 403 break; 404 case SizerMaxAndNewSize: 405 // Do nothing. Values set on the command line, don't update them at runtime. 406 break; 407 case SizerNewRatio: 408 *min_young_length = number_of_heap_regions / (NewRatio + 1); 409 *max_young_length = *min_young_length; 410 break; 411 default: 412 ShouldNotReachHere(); 413 } 414 415 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 416 } 417 418 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 419 // We need to pass the desired values because recalculation may not update these 420 // values in some cases. 421 uint temp = _min_desired_young_length; 422 uint result = _max_desired_young_length; 423 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 424 return result; 425 } 426 427 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 428 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 429 &_max_desired_young_length); 430 } 431 432 void G1CollectorPolicy::init() { 433 // Set aside an initial future to_space. 434 _g1 = G1CollectedHeap::heap(); 435 436 assert(Heap_lock->owned_by_self(), "Locking discipline."); 437 438 initialize_gc_policy_counters(); 439 440 if (adaptive_young_list_length()) { 441 _young_list_fixed_length = 0; 442 } else { 443 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 444 } 445 _free_regions_at_end_of_collection = _g1->num_free_regions(); 446 447 update_young_list_max_and_target_length(); 448 // We may immediately start allocating regions and placing them on the 449 // collection set list. Initialize the per-collection set info 450 start_incremental_cset_building(); 451 } 452 453 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 454 phase_times()->note_gc_start(num_active_workers); 455 } 456 457 // Create the jstat counters for the policy. 458 void G1CollectorPolicy::initialize_gc_policy_counters() { 459 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 460 } 461 462 bool G1CollectorPolicy::predict_will_fit(uint young_length, 463 double base_time_ms, 464 uint base_free_regions, 465 double target_pause_time_ms) const { 466 if (young_length >= base_free_regions) { 467 // end condition 1: not enough space for the young regions 468 return false; 469 } 470 471 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 472 size_t bytes_to_copy = 473 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 474 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 475 double young_other_time_ms = predict_young_other_time_ms(young_length); 476 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 477 if (pause_time_ms > target_pause_time_ms) { 478 // end condition 2: prediction is over the target pause time 479 return false; 480 } 481 482 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 483 if ((2.0 /* magic */ * _predictor.sigma()) * bytes_to_copy > free_bytes) { 484 // end condition 3: out-of-space (conservatively!) 485 return false; 486 } 487 488 // success! 489 return true; 490 } 491 492 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 493 // re-calculate the necessary reserve 494 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 495 // We use ceiling so that if reserve_regions_d is > 0.0 (but 496 // smaller than 1.0) we'll get 1. 497 _reserve_regions = (uint) ceil(reserve_regions_d); 498 499 _young_gen_sizer->heap_size_changed(new_number_of_regions); 500 } 501 502 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 503 uint base_min_length) const { 504 uint desired_min_length = 0; 505 if (adaptive_young_list_length()) { 506 if (_alloc_rate_ms_seq->num() > 3) { 507 double now_sec = os::elapsedTime(); 508 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 509 double alloc_rate_ms = predict_alloc_rate_ms(); 510 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 511 } else { 512 // otherwise we don't have enough info to make the prediction 513 } 514 } 515 desired_min_length += base_min_length; 516 // make sure we don't go below any user-defined minimum bound 517 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 518 } 519 520 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 521 // Here, we might want to also take into account any additional 522 // constraints (i.e., user-defined minimum bound). Currently, we 523 // effectively don't set this bound. 524 return _young_gen_sizer->max_desired_young_length(); 525 } 526 527 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t* unbounded_target_length) { 528 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq), unbounded_target_length); 529 } 530 531 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths, size_t* unbounded_target_length) { 532 update_young_list_target_length(rs_lengths, unbounded_target_length); 533 update_max_gc_locker_expansion(); 534 } 535 536 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths, size_t* unbounded_target_length) { 537 _young_list_target_length = bounded_young_list_target_length(rs_lengths, unbounded_target_length); 538 } 539 540 void G1CollectorPolicy::update_young_list_target_length() { 541 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 542 } 543 544 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths, size_t* unbounded_target_length) const { 545 // Calculate the absolute and desired min bounds. 546 547 // This is how many young regions we already have (currently: the survivors). 548 uint base_min_length = recorded_survivor_regions(); 549 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 550 // This is the absolute minimum young length. Ensure that we 551 // will at least have one eden region available for allocation. 552 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 553 // If we shrank the young list target it should not shrink below the current size. 554 desired_min_length = MAX2(desired_min_length, absolute_min_length); 555 // Calculate the absolute and desired max bounds. 556 557 uint desired_max_length = calculate_young_list_desired_max_length(); 558 559 uint young_list_target_length = 0; 560 if (adaptive_young_list_length()) { 561 if (collector_state()->gcs_are_young()) { 562 young_list_target_length = 563 calculate_young_list_target_length(rs_lengths, 564 base_min_length, 565 desired_min_length, 566 desired_max_length); 567 } else { 568 // Don't calculate anything and let the code below bound it to 569 // the desired_min_length, i.e., do the next GC as soon as 570 // possible to maximize how many old regions we can add to it. 571 } 572 } else { 573 // The user asked for a fixed young gen so we'll fix the young gen 574 // whether the next GC is young or mixed. 575 young_list_target_length = _young_list_fixed_length; 576 } 577 578 if (unbounded_target_length != NULL) { 579 *unbounded_target_length = young_list_target_length; 580 } 581 582 // We will try our best not to "eat" into the reserve. 583 uint absolute_max_length = 0; 584 if (_free_regions_at_end_of_collection > _reserve_regions) { 585 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 586 } 587 if (desired_max_length > absolute_max_length) { 588 desired_max_length = absolute_max_length; 589 } 590 591 // Make sure we don't go over the desired max length, nor under the 592 // desired min length. In case they clash, desired_min_length wins 593 // which is why that test is second. 594 if (young_list_target_length > desired_max_length) { 595 young_list_target_length = desired_max_length; 596 } 597 if (young_list_target_length < desired_min_length) { 598 young_list_target_length = desired_min_length; 599 } 600 601 assert(young_list_target_length > recorded_survivor_regions(), 602 "we should be able to allocate at least one eden region"); 603 assert(young_list_target_length >= absolute_min_length, "post-condition"); 604 605 return young_list_target_length; 606 } 607 608 uint 609 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 610 uint base_min_length, 611 uint desired_min_length, 612 uint desired_max_length) const { 613 assert(adaptive_young_list_length(), "pre-condition"); 614 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 615 616 // In case some edge-condition makes the desired max length too small... 617 if (desired_max_length <= desired_min_length) { 618 return desired_min_length; 619 } 620 621 // We'll adjust min_young_length and max_young_length not to include 622 // the already allocated young regions (i.e., so they reflect the 623 // min and max eden regions we'll allocate). The base_min_length 624 // will be reflected in the predictions by the 625 // survivor_regions_evac_time prediction. 626 assert(desired_min_length > base_min_length, "invariant"); 627 uint min_young_length = desired_min_length - base_min_length; 628 assert(desired_max_length > base_min_length, "invariant"); 629 uint max_young_length = desired_max_length - base_min_length; 630 631 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 632 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 633 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 634 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 635 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 636 double base_time_ms = 637 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 638 survivor_regions_evac_time; 639 uint available_free_regions = _free_regions_at_end_of_collection; 640 uint base_free_regions = 0; 641 if (available_free_regions > _reserve_regions) { 642 base_free_regions = available_free_regions - _reserve_regions; 643 } 644 645 // Here, we will make sure that the shortest young length that 646 // makes sense fits within the target pause time. 647 648 if (predict_will_fit(min_young_length, base_time_ms, 649 base_free_regions, target_pause_time_ms)) { 650 // The shortest young length will fit into the target pause time; 651 // we'll now check whether the absolute maximum number of young 652 // regions will fit in the target pause time. If not, we'll do 653 // a binary search between min_young_length and max_young_length. 654 if (predict_will_fit(max_young_length, base_time_ms, 655 base_free_regions, target_pause_time_ms)) { 656 // The maximum young length will fit into the target pause time. 657 // We are done so set min young length to the maximum length (as 658 // the result is assumed to be returned in min_young_length). 659 min_young_length = max_young_length; 660 } else { 661 // The maximum possible number of young regions will not fit within 662 // the target pause time so we'll search for the optimal 663 // length. The loop invariants are: 664 // 665 // min_young_length < max_young_length 666 // min_young_length is known to fit into the target pause time 667 // max_young_length is known not to fit into the target pause time 668 // 669 // Going into the loop we know the above hold as we've just 670 // checked them. Every time around the loop we check whether 671 // the middle value between min_young_length and 672 // max_young_length fits into the target pause time. If it 673 // does, it becomes the new min. If it doesn't, it becomes 674 // the new max. This way we maintain the loop invariants. 675 676 assert(min_young_length < max_young_length, "invariant"); 677 uint diff = (max_young_length - min_young_length) / 2; 678 while (diff > 0) { 679 uint young_length = min_young_length + diff; 680 if (predict_will_fit(young_length, base_time_ms, 681 base_free_regions, target_pause_time_ms)) { 682 min_young_length = young_length; 683 } else { 684 max_young_length = young_length; 685 } 686 assert(min_young_length < max_young_length, "invariant"); 687 diff = (max_young_length - min_young_length) / 2; 688 } 689 // The results is min_young_length which, according to the 690 // loop invariants, should fit within the target pause time. 691 692 // These are the post-conditions of the binary search above: 693 assert(min_young_length < max_young_length, 694 "otherwise we should have discovered that max_young_length " 695 "fits into the pause target and not done the binary search"); 696 assert(predict_will_fit(min_young_length, base_time_ms, 697 base_free_regions, target_pause_time_ms), 698 "min_young_length, the result of the binary search, should " 699 "fit into the pause target"); 700 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 701 base_free_regions, target_pause_time_ms), 702 "min_young_length, the result of the binary search, should be " 703 "optimal, so no larger length should fit into the pause target"); 704 } 705 } else { 706 // Even the minimum length doesn't fit into the pause time 707 // target, return it as the result nevertheless. 708 } 709 return base_min_length + min_young_length; 710 } 711 712 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 713 double survivor_regions_evac_time = 0.0; 714 for (HeapRegion * r = _recorded_survivor_head; 715 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 716 r = r->get_next_young_region()) { 717 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 718 } 719 return survivor_regions_evac_time; 720 } 721 722 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 723 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 724 725 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 726 if (rs_lengths > _rs_lengths_prediction) { 727 // add 10% to avoid having to recalculate often 728 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 729 update_rs_lengths_prediction(rs_lengths_prediction); 730 731 update_young_list_max_and_target_length(rs_lengths_prediction); 732 } 733 } 734 735 void G1CollectorPolicy::update_rs_lengths_prediction() { 736 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 737 } 738 739 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 740 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 741 _rs_lengths_prediction = prediction; 742 } 743 } 744 745 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 746 bool is_tlab, 747 bool* gc_overhead_limit_was_exceeded) { 748 guarantee(false, "Not using this policy feature yet."); 749 return NULL; 750 } 751 752 // This method controls how a collector handles one or more 753 // of its generations being fully allocated. 754 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 755 bool is_tlab) { 756 guarantee(false, "Not using this policy feature yet."); 757 return NULL; 758 } 759 760 761 #ifndef PRODUCT 762 bool G1CollectorPolicy::verify_young_ages() { 763 HeapRegion* head = _g1->young_list()->first_region(); 764 return 765 verify_young_ages(head, _short_lived_surv_rate_group); 766 // also call verify_young_ages on any additional surv rate groups 767 } 768 769 bool 770 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 771 SurvRateGroup *surv_rate_group) { 772 guarantee( surv_rate_group != NULL, "pre-condition" ); 773 774 const char* name = surv_rate_group->name(); 775 bool ret = true; 776 int prev_age = -1; 777 778 for (HeapRegion* curr = head; 779 curr != NULL; 780 curr = curr->get_next_young_region()) { 781 SurvRateGroup* group = curr->surv_rate_group(); 782 if (group == NULL && !curr->is_survivor()) { 783 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 784 ret = false; 785 } 786 787 if (surv_rate_group == group) { 788 int age = curr->age_in_surv_rate_group(); 789 790 if (age < 0) { 791 gclog_or_tty->print_cr("## %s: encountered negative age", name); 792 ret = false; 793 } 794 795 if (age <= prev_age) { 796 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 797 "(%d, %d)", name, age, prev_age); 798 ret = false; 799 } 800 prev_age = age; 801 } 802 } 803 804 return ret; 805 } 806 #endif // PRODUCT 807 808 void G1CollectorPolicy::record_full_collection_start() { 809 _full_collection_start_sec = os::elapsedTime(); 810 record_heap_size_info_at_start(true /* full */); 811 // Release the future to-space so that it is available for compaction into. 812 collector_state()->set_full_collection(true); 813 } 814 815 void G1CollectorPolicy::record_full_collection_end() { 816 // Consider this like a collection pause for the purposes of allocation 817 // since last pause. 818 double end_sec = os::elapsedTime(); 819 double full_gc_time_sec = end_sec - _full_collection_start_sec; 820 double full_gc_time_ms = full_gc_time_sec * 1000.0; 821 822 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 823 824 update_recent_gc_times(end_sec, full_gc_time_ms); 825 826 collector_state()->set_full_collection(false); 827 828 // "Nuke" the heuristics that control the young/mixed GC 829 // transitions and make sure we start with young GCs after the Full GC. 830 collector_state()->set_gcs_are_young(true); 831 collector_state()->set_last_young_gc(false); 832 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 833 collector_state()->set_during_initial_mark_pause(false); 834 collector_state()->set_in_marking_window(false); 835 collector_state()->set_in_marking_window_im(false); 836 837 _short_lived_surv_rate_group->start_adding_regions(); 838 // also call this on any additional surv rate groups 839 840 record_survivor_regions(0, NULL, NULL); 841 842 _free_regions_at_end_of_collection = _g1->num_free_regions(); 843 // Reset survivors SurvRateGroup. 844 _survivor_surv_rate_group->reset(); 845 update_young_list_max_and_target_length(); 846 update_rs_lengths_prediction(); 847 _collectionSetChooser->clear(); 848 849 _last_old_allocated_bytes = 0; 850 851 record_pause(FullGC, _full_collection_start_sec, end_sec); 852 } 853 854 void G1CollectorPolicy::record_stop_world_start() { 855 _stop_world_start = os::elapsedTime(); 856 } 857 858 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 859 // We only need to do this here as the policy will only be applied 860 // to the GC we're about to start. so, no point is calculating this 861 // every time we calculate / recalculate the target young length. 862 update_survivors_policy(); 863 864 assert(_g1->used() == _g1->recalculate_used(), 865 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 866 _g1->used(), _g1->recalculate_used()); 867 868 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 869 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 870 _stop_world_start = 0.0; 871 872 record_heap_size_info_at_start(false /* full */); 873 874 phase_times()->record_cur_collection_start_sec(start_time_sec); 875 _pending_cards = _g1->pending_card_num(); 876 877 _collection_set_bytes_used_before = 0; 878 _bytes_copied_during_gc = 0; 879 880 collector_state()->set_last_gc_was_young(false); 881 882 // do that for any other surv rate groups 883 _short_lived_surv_rate_group->stop_adding_regions(); 884 _survivors_age_table.clear(); 885 886 assert( verify_young_ages(), "region age verification" ); 887 } 888 889 void G1CollectorPolicy::record_concurrent_mark_init_end(double 890 mark_init_elapsed_time_ms) { 891 collector_state()->set_during_marking(true); 892 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 893 collector_state()->set_during_initial_mark_pause(false); 894 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 895 } 896 897 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 898 _mark_remark_start_sec = os::elapsedTime(); 899 collector_state()->set_during_marking(false); 900 } 901 902 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 903 double end_time_sec = os::elapsedTime(); 904 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 905 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 906 _cur_mark_stop_world_time_ms += elapsed_time_ms; 907 _prev_collection_pause_end_ms += elapsed_time_ms; 908 909 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 910 } 911 912 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 913 _mark_cleanup_start_sec = os::elapsedTime(); 914 } 915 916 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 917 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 918 "skip last young-only gc"); 919 collector_state()->set_last_young_gc(should_continue_with_reclaim); 920 // We abort the marking phase. 921 if (!should_continue_with_reclaim) { 922 abort_time_to_mixed_tracking(); 923 } 924 collector_state()->set_in_marking_window(false); 925 } 926 927 void G1CollectorPolicy::record_concurrent_pause() { 928 if (_stop_world_start > 0.0) { 929 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 930 _trace_young_gen_time_data.record_yield_time(yield_ms); 931 } 932 } 933 934 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 935 return phase_times()->average_time_ms(phase); 936 } 937 938 double G1CollectorPolicy::young_other_time_ms() const { 939 return phase_times()->young_cset_choice_time_ms() + 940 phase_times()->young_free_cset_time_ms(); 941 } 942 943 double G1CollectorPolicy::non_young_other_time_ms() const { 944 return phase_times()->non_young_cset_choice_time_ms() + 945 phase_times()->non_young_free_cset_time_ms(); 946 947 } 948 949 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 950 return pause_time_ms - 951 average_time_ms(G1GCPhaseTimes::UpdateRS) - 952 average_time_ms(G1GCPhaseTimes::ScanRS) - 953 average_time_ms(G1GCPhaseTimes::ObjCopy) - 954 average_time_ms(G1GCPhaseTimes::Termination); 955 } 956 957 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 958 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 959 } 960 961 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 962 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 963 } 964 965 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 966 if (about_to_start_mixed_phase()) { 967 return false; 968 } 969 970 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 971 972 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 973 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 974 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 975 976 if (marking_request_bytes > marking_initiating_used_threshold) { 977 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 978 ergo_verbose5(ErgoConcCycles, 979 "request concurrent cycle initiation", 980 ergo_format_reason("occupancy higher than threshold") 981 ergo_format_byte("occupancy") 982 ergo_format_byte("allocation request") 983 ergo_format_byte_perc("threshold") 984 ergo_format_str("source"), 985 cur_used_bytes, 986 alloc_byte_size, 987 marking_initiating_used_threshold, 988 (double) marking_initiating_used_threshold / _g1->capacity() * 100, 989 source); 990 return true; 991 } else { 992 ergo_verbose5(ErgoConcCycles, 993 "do not request concurrent cycle initiation", 994 ergo_format_reason("still doing mixed collections") 995 ergo_format_byte("occupancy") 996 ergo_format_byte("allocation request") 997 ergo_format_byte_perc("threshold") 998 ergo_format_str("source"), 999 cur_used_bytes, 1000 alloc_byte_size, 1001 marking_initiating_used_threshold, 1002 (double) InitiatingHeapOccupancyPercent, 1003 source); 1004 } 1005 } 1006 1007 return false; 1008 } 1009 1010 // Anything below that is considered to be zero 1011 #define MIN_TIMER_GRANULARITY 0.0000001 1012 1013 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 1014 double end_time_sec = os::elapsedTime(); 1015 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1016 "otherwise, the subtraction below does not make sense"); 1017 size_t cur_used_bytes = _g1->used(); 1018 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1019 bool last_pause_included_initial_mark = false; 1020 bool update_stats = !_g1->evacuation_failed(); 1021 1022 #ifndef PRODUCT 1023 if (G1YoungSurvRateVerbose) { 1024 gclog_or_tty->cr(); 1025 _short_lived_surv_rate_group->print(); 1026 // do that for any other surv rate groups too 1027 } 1028 #endif // PRODUCT 1029 1030 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec); 1031 1032 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1033 if (last_pause_included_initial_mark) { 1034 record_concurrent_mark_init_end(0.0); 1035 } else { 1036 maybe_start_marking(); 1037 } 1038 1039 double app_time_ms = 1.0; 1040 1041 if (update_stats) { 1042 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1043 // this is where we update the allocation rate of the application 1044 app_time_ms = 1045 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1046 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1047 // This usually happens due to the timer not having the required 1048 // granularity. Some Linuxes are the usual culprits. 1049 // We'll just set it to something (arbitrarily) small. 1050 app_time_ms = 1.0; 1051 } 1052 // We maintain the invariant that all objects allocated by mutator 1053 // threads will be allocated out of eden regions. So, we can use 1054 // the eden region number allocated since the previous GC to 1055 // calculate the application's allocate rate. The only exception 1056 // to that is humongous objects that are allocated separately. But 1057 // given that humongous object allocations do not really affect 1058 // either the pause's duration nor when the next pause will take 1059 // place we can safely ignore them here. 1060 uint regions_allocated = eden_cset_region_length(); 1061 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1062 _alloc_rate_ms_seq->add(alloc_rate_ms); 1063 1064 double interval_ms = 1065 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1066 update_recent_gc_times(end_time_sec, pause_time_ms); 1067 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1068 if (recent_avg_pause_time_ratio() < 0.0 || 1069 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1070 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1071 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1072 if (_recent_avg_pause_time_ratio < 0.0) { 1073 _recent_avg_pause_time_ratio = 0.0; 1074 } else { 1075 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1076 _recent_avg_pause_time_ratio = 1.0; 1077 } 1078 } 1079 } 1080 1081 bool new_in_marking_window = collector_state()->in_marking_window(); 1082 bool new_in_marking_window_im = false; 1083 if (last_pause_included_initial_mark) { 1084 new_in_marking_window = true; 1085 new_in_marking_window_im = true; 1086 } 1087 1088 if (collector_state()->last_young_gc()) { 1089 // This is supposed to to be the "last young GC" before we start 1090 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1091 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1092 1093 if (next_gc_should_be_mixed("start mixed GCs", 1094 "do not start mixed GCs")) { 1095 collector_state()->set_gcs_are_young(false); 1096 } else { 1097 // We aborted the mixed GC phase early. 1098 abort_time_to_mixed_tracking(); 1099 } 1100 1101 collector_state()->set_last_young_gc(false); 1102 } 1103 1104 if (!collector_state()->last_gc_was_young()) { 1105 // This is a mixed GC. Here we decide whether to continue doing 1106 // mixed GCs or not. 1107 if (!next_gc_should_be_mixed("continue mixed GCs", 1108 "do not continue mixed GCs")) { 1109 collector_state()->set_gcs_are_young(true); 1110 1111 maybe_start_marking(); 1112 } 1113 } 1114 1115 _short_lived_surv_rate_group->start_adding_regions(); 1116 // Do that for any other surv rate groups 1117 1118 if (update_stats) { 1119 double cost_per_card_ms = 0.0; 1120 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1121 if (_pending_cards > 0) { 1122 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1123 _cost_per_card_ms_seq->add(cost_per_card_ms); 1124 } 1125 _cost_scan_hcc_seq->add(cost_scan_hcc); 1126 1127 double cost_per_entry_ms = 0.0; 1128 if (cards_scanned > 10) { 1129 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1130 if (collector_state()->last_gc_was_young()) { 1131 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1132 } else { 1133 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1134 } 1135 } 1136 1137 if (_max_rs_lengths > 0) { 1138 double cards_per_entry_ratio = 1139 (double) cards_scanned / (double) _max_rs_lengths; 1140 if (collector_state()->last_gc_was_young()) { 1141 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1142 } else { 1143 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1144 } 1145 } 1146 1147 // This is defensive. For a while _max_rs_lengths could get 1148 // smaller than _recorded_rs_lengths which was causing 1149 // rs_length_diff to get very large and mess up the RSet length 1150 // predictions. The reason was unsafe concurrent updates to the 1151 // _inc_cset_recorded_rs_lengths field which the code below guards 1152 // against (see CR 7118202). This bug has now been fixed (see CR 1153 // 7119027). However, I'm still worried that 1154 // _inc_cset_recorded_rs_lengths might still end up somewhat 1155 // inaccurate. The concurrent refinement thread calculates an 1156 // RSet's length concurrently with other CR threads updating it 1157 // which might cause it to calculate the length incorrectly (if, 1158 // say, it's in mid-coarsening). So I'll leave in the defensive 1159 // conditional below just in case. 1160 size_t rs_length_diff = 0; 1161 if (_max_rs_lengths > _recorded_rs_lengths) { 1162 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1163 } 1164 _rs_length_diff_seq->add((double) rs_length_diff); 1165 1166 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1167 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1168 double cost_per_byte_ms = 0.0; 1169 1170 if (copied_bytes > 0) { 1171 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1172 if (collector_state()->in_marking_window()) { 1173 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1174 } else { 1175 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1176 } 1177 } 1178 1179 if (young_cset_region_length() > 0) { 1180 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1181 young_cset_region_length()); 1182 } 1183 1184 if (old_cset_region_length() > 0) { 1185 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1186 old_cset_region_length()); 1187 } 1188 1189 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1190 1191 _pending_cards_seq->add((double) _pending_cards); 1192 _rs_lengths_seq->add((double) _max_rs_lengths); 1193 } 1194 1195 collector_state()->set_in_marking_window(new_in_marking_window); 1196 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1197 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1198 // IHOP control wants to know the expected young gen length if it were not 1199 // restrained by the heap reserve. Using the actual length would make the 1200 // prediction too small and the limit the young gen every time we get to the 1201 // predicted target occupancy. 1202 size_t last_unrestrained_young_length = 0; 1203 update_young_list_max_and_target_length(&last_unrestrained_young_length); 1204 update_rs_lengths_prediction(); 1205 1206 double marking_to_mixed_time = -1.0; 1207 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1208 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1209 assert(marking_to_mixed_time > 0.0, 1210 "Initial mark to mixed time must be larger than zero but is %.3f", 1211 marking_to_mixed_time); 1212 } 1213 // Only update IHOP information on regular GCs. 1214 if (update_stats) { 1215 update_ihop_statistics(marking_to_mixed_time, 1216 app_time_ms / 1000.0, 1217 _last_old_allocated_bytes, 1218 last_unrestrained_young_length * HeapRegion::GrainBytes); 1219 } 1220 _last_old_allocated_bytes = 0; 1221 1222 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1223 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1224 1225 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1226 1227 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1228 ergo_verbose2(ErgoTiming, 1229 "adjust concurrent refinement thresholds", 1230 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1231 ergo_format_ms("Update RS time goal") 1232 ergo_format_ms("Scan HCC time"), 1233 update_rs_time_goal_ms, 1234 scan_hcc_time_ms); 1235 1236 update_rs_time_goal_ms = 0; 1237 } else { 1238 update_rs_time_goal_ms -= scan_hcc_time_ms; 1239 } 1240 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1241 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1242 update_rs_time_goal_ms); 1243 1244 _collectionSetChooser->verify(); 1245 } 1246 1247 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1248 if (G1UseAdaptiveIHOP) { 1249 // The target occupancy is the total heap occupancy we want to hit. First, we 1250 // want to avoid eating into the reserve intended for young GC (to avoid unnecessary 1251 // throughput loss). Additionally G1 is free to not clean out up to 1252 // G1HeapWastePercent of heap, that space also cannot be used for allocation 1253 // while marking. 1254 size_t safe_heap_percentage = (size_t) (G1ReservePercent + G1HeapWastePercent); 1255 size_t target_occupancy = 0; 1256 1257 if (safe_heap_percentage < 100) { 1258 target_occupancy = G1CollectedHeap::heap()->max_capacity() * (100.0 - safe_heap_percentage) / 100.0; 1259 } 1260 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1261 target_occupancy, 1262 &_predictor); 1263 } else { 1264 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1265 G1CollectedHeap::heap()->max_capacity()); 1266 } 1267 } 1268 1269 void G1CollectorPolicy::update_ihop_statistics(double marking_time, 1270 double mutator_time_s, 1271 size_t mutator_alloc_bytes, 1272 size_t young_gen_size) { 1273 bool report = false; 1274 1275 // To avoid using really small times that may be caused by e.g. back-to-back gcs 1276 // we filter them out. 1277 double const min_valid_time = 1e-6; 1278 1279 if (marking_time > min_valid_time) { 1280 _ihop_control->update_time_to_mixed(marking_time); 1281 report = true; 1282 } 1283 1284 // As an approximation for the young gc promotion rates during marking we use 1285 // all of them. In many applications there are only a few if any young gcs during 1286 // marking, which makes any prediction useless. This increases the accuracy of the 1287 // prediction. 1288 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1289 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1290 report = true; 1291 } 1292 1293 if (report) { 1294 report_ihop_statistics(); 1295 } 1296 } 1297 1298 void G1CollectorPolicy::report_ihop_statistics() { 1299 _ihop_control->print(); 1300 } 1301 1302 #define EXT_SIZE_FORMAT "%.1f%s" 1303 #define EXT_SIZE_PARAMS(bytes) \ 1304 byte_size_in_proper_unit((double)(bytes)), \ 1305 proper_unit_for_byte_size((bytes)) 1306 1307 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1308 YoungList* young_list = _g1->young_list(); 1309 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1310 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1311 _heap_capacity_bytes_before_gc = _g1->capacity(); 1312 _heap_used_bytes_before_gc = _g1->used(); 1313 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1314 1315 _eden_capacity_bytes_before_gc = 1316 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1317 1318 if (full) { 1319 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1320 } 1321 } 1322 1323 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1324 size_t bytes_after = _g1->used(); 1325 size_t capacity = _g1->capacity(); 1326 1327 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1328 byte_size_in_proper_unit(bytes_before), 1329 proper_unit_for_byte_size(bytes_before), 1330 byte_size_in_proper_unit(bytes_after), 1331 proper_unit_for_byte_size(bytes_after), 1332 byte_size_in_proper_unit(capacity), 1333 proper_unit_for_byte_size(capacity)); 1334 } 1335 1336 void G1CollectorPolicy::print_heap_transition() const { 1337 print_heap_transition(_heap_used_bytes_before_gc); 1338 } 1339 1340 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1341 YoungList* young_list = _g1->young_list(); 1342 1343 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1344 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1345 size_t heap_used_bytes_after_gc = _g1->used(); 1346 1347 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1348 size_t eden_capacity_bytes_after_gc = 1349 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1350 1351 gclog_or_tty->print( 1352 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1353 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1354 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1355 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1356 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1357 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1358 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1359 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1360 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1361 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1362 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1363 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1364 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1365 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1366 1367 if (full) { 1368 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1369 } 1370 1371 gclog_or_tty->cr(); 1372 } 1373 1374 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1375 phase_times()->print(pause_time_sec); 1376 } 1377 1378 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1379 double update_rs_processed_buffers, 1380 double goal_ms) { 1381 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1382 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1383 1384 if (G1UseAdaptiveConcRefinement) { 1385 const int k_gy = 3, k_gr = 6; 1386 const double inc_k = 1.1, dec_k = 0.9; 1387 1388 int g = cg1r->green_zone(); 1389 if (update_rs_time > goal_ms) { 1390 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1391 } else { 1392 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1393 g = (int)MAX2(g * inc_k, g + 1.0); 1394 } 1395 } 1396 // Change the refinement threads params 1397 cg1r->set_green_zone(g); 1398 cg1r->set_yellow_zone(g * k_gy); 1399 cg1r->set_red_zone(g * k_gr); 1400 cg1r->reinitialize_threads(); 1401 1402 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1403 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1404 cg1r->yellow_zone()); 1405 // Change the barrier params 1406 dcqs.set_process_completed_threshold(processing_threshold); 1407 dcqs.set_max_completed_queue(cg1r->red_zone()); 1408 } 1409 1410 int curr_queue_size = dcqs.completed_buffers_num(); 1411 if (curr_queue_size >= cg1r->yellow_zone()) { 1412 dcqs.set_completed_queue_padding(curr_queue_size); 1413 } else { 1414 dcqs.set_completed_queue_padding(0); 1415 } 1416 dcqs.notify_if_necessary(); 1417 } 1418 1419 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1420 return (size_t) get_new_prediction(_rs_length_diff_seq); 1421 } 1422 1423 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1424 return get_new_prediction(_alloc_rate_ms_seq); 1425 } 1426 1427 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1428 return get_new_prediction(_cost_per_card_ms_seq); 1429 } 1430 1431 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1432 return get_new_prediction(_cost_scan_hcc_seq); 1433 } 1434 1435 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1436 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1437 } 1438 1439 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1440 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1441 } 1442 1443 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1444 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1445 return predict_young_cards_per_entry_ratio(); 1446 } else { 1447 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1448 } 1449 } 1450 1451 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1452 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1453 } 1454 1455 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1456 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1457 } 1458 1459 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1460 if (collector_state()->gcs_are_young()) { 1461 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1462 } else { 1463 return predict_mixed_rs_scan_time_ms(card_num); 1464 } 1465 } 1466 1467 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1468 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1469 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1470 } else { 1471 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1472 } 1473 } 1474 1475 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1476 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1477 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1478 } else { 1479 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1480 } 1481 } 1482 1483 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1484 if (collector_state()->during_concurrent_mark()) { 1485 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1486 } else { 1487 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1488 } 1489 } 1490 1491 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1492 return get_new_prediction(_constant_other_time_ms_seq); 1493 } 1494 1495 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1496 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1497 } 1498 1499 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1500 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1501 } 1502 1503 double G1CollectorPolicy::predict_remark_time_ms() const { 1504 return get_new_prediction(_concurrent_mark_remark_times_ms); 1505 } 1506 1507 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1508 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1509 } 1510 1511 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1512 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1513 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1514 double pred = get_new_prediction(seq); 1515 if (pred > 1.0) { 1516 pred = 1.0; 1517 } 1518 return pred; 1519 } 1520 1521 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1522 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1523 } 1524 1525 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1526 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1527 } 1528 1529 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1530 size_t scanned_cards) const { 1531 return 1532 predict_rs_update_time_ms(pending_cards) + 1533 predict_rs_scan_time_ms(scanned_cards) + 1534 predict_constant_other_time_ms(); 1535 } 1536 1537 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1538 size_t rs_length = predict_rs_length_diff(); 1539 size_t card_num; 1540 if (collector_state()->gcs_are_young()) { 1541 card_num = predict_young_card_num(rs_length); 1542 } else { 1543 card_num = predict_non_young_card_num(rs_length); 1544 } 1545 return predict_base_elapsed_time_ms(pending_cards, card_num); 1546 } 1547 1548 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1549 size_t bytes_to_copy; 1550 if (hr->is_marked()) 1551 bytes_to_copy = hr->max_live_bytes(); 1552 else { 1553 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1554 int age = hr->age_in_surv_rate_group(); 1555 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1556 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1557 } 1558 return bytes_to_copy; 1559 } 1560 1561 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1562 bool for_young_gc) const { 1563 size_t rs_length = hr->rem_set()->occupied(); 1564 size_t card_num; 1565 1566 // Predicting the number of cards is based on which type of GC 1567 // we're predicting for. 1568 if (for_young_gc) { 1569 card_num = predict_young_card_num(rs_length); 1570 } else { 1571 card_num = predict_non_young_card_num(rs_length); 1572 } 1573 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1574 1575 double region_elapsed_time_ms = 1576 predict_rs_scan_time_ms(card_num) + 1577 predict_object_copy_time_ms(bytes_to_copy); 1578 1579 // The prediction of the "other" time for this region is based 1580 // upon the region type and NOT the GC type. 1581 if (hr->is_young()) { 1582 region_elapsed_time_ms += predict_young_other_time_ms(1); 1583 } else { 1584 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1585 } 1586 return region_elapsed_time_ms; 1587 } 1588 1589 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1590 uint survivor_cset_region_length) { 1591 _eden_cset_region_length = eden_cset_region_length; 1592 _survivor_cset_region_length = survivor_cset_region_length; 1593 _old_cset_region_length = 0; 1594 } 1595 1596 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1597 _recorded_rs_lengths = rs_lengths; 1598 } 1599 1600 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1601 double elapsed_ms) { 1602 _recent_gc_times_ms->add(elapsed_ms); 1603 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1604 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1605 } 1606 1607 size_t G1CollectorPolicy::expansion_amount() const { 1608 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1609 double threshold = _gc_overhead_perc; 1610 if (recent_gc_overhead > threshold) { 1611 // We will double the existing space, or take 1612 // G1ExpandByPercentOfAvailable % of the available expansion 1613 // space, whichever is smaller, bounded below by a minimum 1614 // expansion (unless that's all that's left.) 1615 const size_t min_expand_bytes = 1*M; 1616 size_t reserved_bytes = _g1->max_capacity(); 1617 size_t committed_bytes = _g1->capacity(); 1618 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1619 size_t expand_bytes; 1620 size_t expand_bytes_via_pct = 1621 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1622 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1623 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1624 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1625 1626 ergo_verbose5(ErgoHeapSizing, 1627 "attempt heap expansion", 1628 ergo_format_reason("recent GC overhead higher than " 1629 "threshold after GC") 1630 ergo_format_perc("recent GC overhead") 1631 ergo_format_perc("threshold") 1632 ergo_format_byte("uncommitted") 1633 ergo_format_byte_perc("calculated expansion amount"), 1634 recent_gc_overhead, threshold, 1635 uncommitted_bytes, 1636 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1637 1638 return expand_bytes; 1639 } else { 1640 return 0; 1641 } 1642 } 1643 1644 void G1CollectorPolicy::print_tracing_info() const { 1645 _trace_young_gen_time_data.print(); 1646 _trace_old_gen_time_data.print(); 1647 } 1648 1649 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1650 #ifndef PRODUCT 1651 _short_lived_surv_rate_group->print_surv_rate_summary(); 1652 // add this call for any other surv rate groups 1653 #endif // PRODUCT 1654 } 1655 1656 bool G1CollectorPolicy::is_young_list_full() const { 1657 uint young_list_length = _g1->young_list()->length(); 1658 uint young_list_target_length = _young_list_target_length; 1659 return young_list_length >= young_list_target_length; 1660 } 1661 1662 bool G1CollectorPolicy::can_expand_young_list() const { 1663 uint young_list_length = _g1->young_list()->length(); 1664 uint young_list_max_length = _young_list_max_length; 1665 return young_list_length < young_list_max_length; 1666 } 1667 1668 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1669 uint expansion_region_num = 0; 1670 if (GCLockerEdenExpansionPercent > 0) { 1671 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1672 double expansion_region_num_d = perc * (double) _young_list_target_length; 1673 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1674 // less than 1.0) we'll get 1. 1675 expansion_region_num = (uint) ceil(expansion_region_num_d); 1676 } else { 1677 assert(expansion_region_num == 0, "sanity"); 1678 } 1679 _young_list_max_length = _young_list_target_length + expansion_region_num; 1680 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1681 } 1682 1683 // Calculates survivor space parameters. 1684 void G1CollectorPolicy::update_survivors_policy() { 1685 double max_survivor_regions_d = 1686 (double) _young_list_target_length / (double) SurvivorRatio; 1687 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1688 // smaller than 1.0) we'll get 1. 1689 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1690 1691 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1692 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1693 } 1694 1695 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1696 // We actually check whether we are marking here and not if we are in a 1697 // reclamation phase. This means that we will schedule a concurrent mark 1698 // even while we are still in the process of reclaiming memory. 1699 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1700 if (!during_cycle) { 1701 ergo_verbose1(ErgoConcCycles, 1702 "request concurrent cycle initiation", 1703 ergo_format_reason("requested by GC cause") 1704 ergo_format_str("GC cause"), 1705 GCCause::to_string(gc_cause)); 1706 collector_state()->set_initiate_conc_mark_if_possible(true); 1707 return true; 1708 } else { 1709 ergo_verbose1(ErgoConcCycles, 1710 "do not request concurrent cycle initiation", 1711 ergo_format_reason("concurrent cycle already in progress") 1712 ergo_format_str("GC cause"), 1713 GCCause::to_string(gc_cause)); 1714 return false; 1715 } 1716 } 1717 1718 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1719 // We are about to decide on whether this pause will be an 1720 // initial-mark pause. 1721 1722 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1723 // will set it here if we have to. However, it should be cleared by 1724 // the end of the pause (it's only set for the duration of an 1725 // initial-mark pause). 1726 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1727 1728 if (collector_state()->initiate_conc_mark_if_possible()) { 1729 // We had noticed on a previous pause that the heap occupancy has 1730 // gone over the initiating threshold and we should start a 1731 // concurrent marking cycle. So we might initiate one. 1732 1733 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1734 // Initiate a new initial mark only if there is no marking or reclamation going 1735 // on. 1736 1737 collector_state()->set_during_initial_mark_pause(true); 1738 // And we can now clear initiate_conc_mark_if_possible() as 1739 // we've already acted on it. 1740 collector_state()->set_initiate_conc_mark_if_possible(false); 1741 1742 ergo_verbose0(ErgoConcCycles, 1743 "initiate concurrent cycle", 1744 ergo_format_reason("concurrent cycle initiation requested")); 1745 } else { 1746 // The concurrent marking thread is still finishing up the 1747 // previous cycle. If we start one right now the two cycles 1748 // overlap. In particular, the concurrent marking thread might 1749 // be in the process of clearing the next marking bitmap (which 1750 // we will use for the next cycle if we start one). Starting a 1751 // cycle now will be bad given that parts of the marking 1752 // information might get cleared by the marking thread. And we 1753 // cannot wait for the marking thread to finish the cycle as it 1754 // periodically yields while clearing the next marking bitmap 1755 // and, if it's in a yield point, it's waiting for us to 1756 // finish. So, at this point we will not start a cycle and we'll 1757 // let the concurrent marking thread complete the last one. 1758 ergo_verbose0(ErgoConcCycles, 1759 "do not initiate concurrent cycle", 1760 ergo_format_reason("concurrent cycle already in progress")); 1761 } 1762 } 1763 } 1764 1765 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1766 G1CollectedHeap* _g1h; 1767 CSetChooserParUpdater _cset_updater; 1768 1769 public: 1770 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1771 uint chunk_size) : 1772 _g1h(G1CollectedHeap::heap()), 1773 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1774 1775 bool doHeapRegion(HeapRegion* r) { 1776 // Do we have any marking information for this region? 1777 if (r->is_marked()) { 1778 // We will skip any region that's currently used as an old GC 1779 // alloc region (we should not consider those for collection 1780 // before we fill them up). 1781 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1782 _cset_updater.add_region(r); 1783 } 1784 } 1785 return false; 1786 } 1787 }; 1788 1789 class ParKnownGarbageTask: public AbstractGangTask { 1790 CollectionSetChooser* _hrSorted; 1791 uint _chunk_size; 1792 G1CollectedHeap* _g1; 1793 HeapRegionClaimer _hrclaimer; 1794 1795 public: 1796 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1797 AbstractGangTask("ParKnownGarbageTask"), 1798 _hrSorted(hrSorted), _chunk_size(chunk_size), 1799 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1800 1801 void work(uint worker_id) { 1802 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1803 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1804 } 1805 }; 1806 1807 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1808 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1809 const uint overpartition_factor = 4; 1810 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1811 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1812 } 1813 1814 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1815 _collectionSetChooser->clear(); 1816 1817 WorkGang* workers = _g1->workers(); 1818 uint n_workers = workers->active_workers(); 1819 1820 uint n_regions = _g1->num_regions(); 1821 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1822 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1823 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1824 workers->run_task(&par_known_garbage_task); 1825 1826 _collectionSetChooser->sort_regions(); 1827 1828 double end_sec = os::elapsedTime(); 1829 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1830 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1831 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1832 _prev_collection_pause_end_ms += elapsed_time_ms; 1833 1834 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1835 } 1836 1837 // Add the heap region at the head of the non-incremental collection set 1838 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1839 assert(_inc_cset_build_state == Active, "Precondition"); 1840 assert(hr->is_old(), "the region should be old"); 1841 1842 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1843 _g1->register_old_region_with_cset(hr); 1844 hr->set_next_in_collection_set(_collection_set); 1845 _collection_set = hr; 1846 _collection_set_bytes_used_before += hr->used(); 1847 size_t rs_length = hr->rem_set()->occupied(); 1848 _recorded_rs_lengths += rs_length; 1849 _old_cset_region_length += 1; 1850 } 1851 1852 // Initialize the per-collection-set information 1853 void G1CollectorPolicy::start_incremental_cset_building() { 1854 assert(_inc_cset_build_state == Inactive, "Precondition"); 1855 1856 _inc_cset_head = NULL; 1857 _inc_cset_tail = NULL; 1858 _inc_cset_bytes_used_before = 0; 1859 1860 _inc_cset_max_finger = 0; 1861 _inc_cset_recorded_rs_lengths = 0; 1862 _inc_cset_recorded_rs_lengths_diffs = 0; 1863 _inc_cset_predicted_elapsed_time_ms = 0.0; 1864 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1865 _inc_cset_build_state = Active; 1866 } 1867 1868 void G1CollectorPolicy::finalize_incremental_cset_building() { 1869 assert(_inc_cset_build_state == Active, "Precondition"); 1870 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1871 1872 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1873 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1874 // that adds a new region to the CSet. Further updates by the 1875 // concurrent refinement thread that samples the young RSet lengths 1876 // are accumulated in the *_diffs fields. Here we add the diffs to 1877 // the "main" fields. 1878 1879 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1880 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1881 } else { 1882 // This is defensive. The diff should in theory be always positive 1883 // as RSets can only grow between GCs. However, given that we 1884 // sample their size concurrently with other threads updating them 1885 // it's possible that we might get the wrong size back, which 1886 // could make the calculations somewhat inaccurate. 1887 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1888 if (_inc_cset_recorded_rs_lengths >= diffs) { 1889 _inc_cset_recorded_rs_lengths -= diffs; 1890 } else { 1891 _inc_cset_recorded_rs_lengths = 0; 1892 } 1893 } 1894 _inc_cset_predicted_elapsed_time_ms += 1895 _inc_cset_predicted_elapsed_time_ms_diffs; 1896 1897 _inc_cset_recorded_rs_lengths_diffs = 0; 1898 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1899 } 1900 1901 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1902 // This routine is used when: 1903 // * adding survivor regions to the incremental cset at the end of an 1904 // evacuation pause, 1905 // * adding the current allocation region to the incremental cset 1906 // when it is retired, and 1907 // * updating existing policy information for a region in the 1908 // incremental cset via young list RSet sampling. 1909 // Therefore this routine may be called at a safepoint by the 1910 // VM thread, or in-between safepoints by mutator threads (when 1911 // retiring the current allocation region) or a concurrent 1912 // refine thread (RSet sampling). 1913 1914 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1915 size_t used_bytes = hr->used(); 1916 _inc_cset_recorded_rs_lengths += rs_length; 1917 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1918 _inc_cset_bytes_used_before += used_bytes; 1919 1920 // Cache the values we have added to the aggregated information 1921 // in the heap region in case we have to remove this region from 1922 // the incremental collection set, or it is updated by the 1923 // rset sampling code 1924 hr->set_recorded_rs_length(rs_length); 1925 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1926 } 1927 1928 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1929 size_t new_rs_length) { 1930 // Update the CSet information that is dependent on the new RS length 1931 assert(hr->is_young(), "Precondition"); 1932 assert(!SafepointSynchronize::is_at_safepoint(), 1933 "should not be at a safepoint"); 1934 1935 // We could have updated _inc_cset_recorded_rs_lengths and 1936 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1937 // that atomically, as this code is executed by a concurrent 1938 // refinement thread, potentially concurrently with a mutator thread 1939 // allocating a new region and also updating the same fields. To 1940 // avoid the atomic operations we accumulate these updates on two 1941 // separate fields (*_diffs) and we'll just add them to the "main" 1942 // fields at the start of a GC. 1943 1944 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1945 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1946 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1947 1948 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1949 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1950 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1951 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1952 1953 hr->set_recorded_rs_length(new_rs_length); 1954 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1955 } 1956 1957 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1958 assert(hr->is_young(), "invariant"); 1959 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1960 assert(_inc_cset_build_state == Active, "Precondition"); 1961 1962 // We need to clear and set the cached recorded/cached collection set 1963 // information in the heap region here (before the region gets added 1964 // to the collection set). An individual heap region's cached values 1965 // are calculated, aggregated with the policy collection set info, 1966 // and cached in the heap region here (initially) and (subsequently) 1967 // by the Young List sampling code. 1968 1969 size_t rs_length = hr->rem_set()->occupied(); 1970 add_to_incremental_cset_info(hr, rs_length); 1971 1972 HeapWord* hr_end = hr->end(); 1973 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1974 1975 assert(!hr->in_collection_set(), "invariant"); 1976 _g1->register_young_region_with_cset(hr); 1977 assert(hr->next_in_collection_set() == NULL, "invariant"); 1978 } 1979 1980 // Add the region at the RHS of the incremental cset 1981 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1982 // We should only ever be appending survivors at the end of a pause 1983 assert(hr->is_survivor(), "Logic"); 1984 1985 // Do the 'common' stuff 1986 add_region_to_incremental_cset_common(hr); 1987 1988 // Now add the region at the right hand side 1989 if (_inc_cset_tail == NULL) { 1990 assert(_inc_cset_head == NULL, "invariant"); 1991 _inc_cset_head = hr; 1992 } else { 1993 _inc_cset_tail->set_next_in_collection_set(hr); 1994 } 1995 _inc_cset_tail = hr; 1996 } 1997 1998 // Add the region to the LHS of the incremental cset 1999 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2000 // Survivors should be added to the RHS at the end of a pause 2001 assert(hr->is_eden(), "Logic"); 2002 2003 // Do the 'common' stuff 2004 add_region_to_incremental_cset_common(hr); 2005 2006 // Add the region at the left hand side 2007 hr->set_next_in_collection_set(_inc_cset_head); 2008 if (_inc_cset_head == NULL) { 2009 assert(_inc_cset_tail == NULL, "Invariant"); 2010 _inc_cset_tail = hr; 2011 } 2012 _inc_cset_head = hr; 2013 } 2014 2015 #ifndef PRODUCT 2016 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2017 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2018 2019 st->print_cr("\nCollection_set:"); 2020 HeapRegion* csr = list_head; 2021 while (csr != NULL) { 2022 HeapRegion* next = csr->next_in_collection_set(); 2023 assert(csr->in_collection_set(), "bad CS"); 2024 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2025 HR_FORMAT_PARAMS(csr), 2026 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2027 csr->age_in_surv_rate_group_cond()); 2028 csr = next; 2029 } 2030 } 2031 #endif // !PRODUCT 2032 2033 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2034 // Returns the given amount of reclaimable bytes (that represents 2035 // the amount of reclaimable space still to be collected) as a 2036 // percentage of the current heap capacity. 2037 size_t capacity_bytes = _g1->capacity(); 2038 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2039 } 2040 2041 void G1CollectorPolicy::maybe_start_marking() { 2042 if (need_to_start_conc_mark("end of GC")) { 2043 // Note: this might have already been set, if during the last 2044 // pause we decided to start a cycle but at the beginning of 2045 // this pause we decided to postpone it. That's OK. 2046 collector_state()->set_initiate_conc_mark_if_possible(true); 2047 } 2048 } 2049 2050 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2051 assert(!collector_state()->full_collection(), "must be"); 2052 if (collector_state()->during_initial_mark_pause()) { 2053 assert(collector_state()->last_gc_was_young(), "must be"); 2054 assert(!collector_state()->last_young_gc(), "must be"); 2055 return InitialMarkGC; 2056 } else if (collector_state()->last_young_gc()) { 2057 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2058 assert(collector_state()->last_gc_was_young(), "must be"); 2059 return LastYoungGC; 2060 } else if (!collector_state()->last_gc_was_young()) { 2061 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2062 assert(!collector_state()->last_young_gc(), "must be"); 2063 return MixedGC; 2064 } else { 2065 assert(collector_state()->last_gc_was_young(), "must be"); 2066 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2067 assert(!collector_state()->last_young_gc(), "must be"); 2068 return YoungOnlyGC; 2069 } 2070 } 2071 2072 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2073 // Manage the MMU tracker. For some reason it ignores Full GCs. 2074 if (kind != FullGC) { 2075 _mmu_tracker->add_pause(start, end); 2076 } 2077 // Manage the mutator time tracking from initial mark to first mixed gc. 2078 switch (kind) { 2079 case FullGC: 2080 abort_time_to_mixed_tracking(); 2081 break; 2082 case Cleanup: 2083 case Remark: 2084 case YoungOnlyGC: 2085 case LastYoungGC: 2086 _initial_mark_to_mixed.add_pause(end - start); 2087 break; 2088 case InitialMarkGC: 2089 _initial_mark_to_mixed.record_initial_mark_end(end); 2090 break; 2091 case MixedGC: 2092 _initial_mark_to_mixed.record_mixed_gc_start(start); 2093 break; 2094 default: 2095 ShouldNotReachHere(); 2096 } 2097 } 2098 2099 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2100 _initial_mark_to_mixed.reset(); 2101 } 2102 2103 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2104 const char* false_action_str) const { 2105 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2106 if (cset_chooser->is_empty()) { 2107 ergo_verbose0(ErgoMixedGCs, 2108 false_action_str, 2109 ergo_format_reason("candidate old regions not available")); 2110 return false; 2111 } 2112 2113 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2114 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2115 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2116 double threshold = (double) G1HeapWastePercent; 2117 if (reclaimable_perc <= threshold) { 2118 ergo_verbose4(ErgoMixedGCs, 2119 false_action_str, 2120 ergo_format_reason("reclaimable percentage not over threshold") 2121 ergo_format_region("candidate old regions") 2122 ergo_format_byte_perc("reclaimable") 2123 ergo_format_perc("threshold"), 2124 cset_chooser->remaining_regions(), 2125 reclaimable_bytes, 2126 reclaimable_perc, threshold); 2127 return false; 2128 } 2129 2130 ergo_verbose4(ErgoMixedGCs, 2131 true_action_str, 2132 ergo_format_reason("candidate old regions available") 2133 ergo_format_region("candidate old regions") 2134 ergo_format_byte_perc("reclaimable") 2135 ergo_format_perc("threshold"), 2136 cset_chooser->remaining_regions(), 2137 reclaimable_bytes, 2138 reclaimable_perc, threshold); 2139 return true; 2140 } 2141 2142 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2143 // The min old CSet region bound is based on the maximum desired 2144 // number of mixed GCs after a cycle. I.e., even if some old regions 2145 // look expensive, we should add them to the CSet anyway to make 2146 // sure we go through the available old regions in no more than the 2147 // maximum desired number of mixed GCs. 2148 // 2149 // The calculation is based on the number of marked regions we added 2150 // to the CSet chooser in the first place, not how many remain, so 2151 // that the result is the same during all mixed GCs that follow a cycle. 2152 2153 const size_t region_num = (size_t) _collectionSetChooser->length(); 2154 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2155 size_t result = region_num / gc_num; 2156 // emulate ceiling 2157 if (result * gc_num < region_num) { 2158 result += 1; 2159 } 2160 return (uint) result; 2161 } 2162 2163 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2164 // The max old CSet region bound is based on the threshold expressed 2165 // as a percentage of the heap size. I.e., it should bound the 2166 // number of old regions added to the CSet irrespective of how many 2167 // of them are available. 2168 2169 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2170 const size_t region_num = g1h->num_regions(); 2171 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2172 size_t result = region_num * perc / 100; 2173 // emulate ceiling 2174 if (100 * result < region_num * perc) { 2175 result += 1; 2176 } 2177 return (uint) result; 2178 } 2179 2180 2181 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2182 double young_start_time_sec = os::elapsedTime(); 2183 2184 YoungList* young_list = _g1->young_list(); 2185 finalize_incremental_cset_building(); 2186 2187 guarantee(target_pause_time_ms > 0.0, 2188 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2189 guarantee(_collection_set == NULL, "Precondition"); 2190 2191 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2192 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2193 2194 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2195 "start choosing CSet", 2196 ergo_format_size("_pending_cards") 2197 ergo_format_ms("predicted base time") 2198 ergo_format_ms("remaining time") 2199 ergo_format_ms("target pause time"), 2200 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2201 2202 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2203 2204 if (collector_state()->last_gc_was_young()) { 2205 _trace_young_gen_time_data.increment_young_collection_count(); 2206 } else { 2207 _trace_young_gen_time_data.increment_mixed_collection_count(); 2208 } 2209 2210 // The young list is laid with the survivor regions from the previous 2211 // pause are appended to the RHS of the young list, i.e. 2212 // [Newly Young Regions ++ Survivors from last pause]. 2213 2214 uint survivor_region_length = young_list->survivor_length(); 2215 uint eden_region_length = young_list->eden_length(); 2216 init_cset_region_lengths(eden_region_length, survivor_region_length); 2217 2218 HeapRegion* hr = young_list->first_survivor_region(); 2219 while (hr != NULL) { 2220 assert(hr->is_survivor(), "badly formed young list"); 2221 // There is a convention that all the young regions in the CSet 2222 // are tagged as "eden", so we do this for the survivors here. We 2223 // use the special set_eden_pre_gc() as it doesn't check that the 2224 // region is free (which is not the case here). 2225 hr->set_eden_pre_gc(); 2226 hr = hr->get_next_young_region(); 2227 } 2228 2229 // Clear the fields that point to the survivor list - they are all young now. 2230 young_list->clear_survivors(); 2231 2232 _collection_set = _inc_cset_head; 2233 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2234 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2235 2236 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2237 "add young regions to CSet", 2238 ergo_format_region("eden") 2239 ergo_format_region("survivors") 2240 ergo_format_ms("predicted young region time") 2241 ergo_format_ms("target pause time"), 2242 eden_region_length, survivor_region_length, 2243 _inc_cset_predicted_elapsed_time_ms, 2244 target_pause_time_ms); 2245 2246 // The number of recorded young regions is the incremental 2247 // collection set's current size 2248 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2249 2250 double young_end_time_sec = os::elapsedTime(); 2251 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2252 2253 return time_remaining_ms; 2254 } 2255 2256 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2257 double non_young_start_time_sec = os::elapsedTime(); 2258 double predicted_old_time_ms = 0.0; 2259 2260 2261 if (!collector_state()->gcs_are_young()) { 2262 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2263 cset_chooser->verify(); 2264 const uint min_old_cset_length = calc_min_old_cset_length(); 2265 const uint max_old_cset_length = calc_max_old_cset_length(); 2266 2267 uint expensive_region_num = 0; 2268 bool check_time_remaining = adaptive_young_list_length(); 2269 2270 HeapRegion* hr = cset_chooser->peek(); 2271 while (hr != NULL) { 2272 if (old_cset_region_length() >= max_old_cset_length) { 2273 // Added maximum number of old regions to the CSet. 2274 ergo_verbose2(ErgoCSetConstruction, 2275 "finish adding old regions to CSet", 2276 ergo_format_reason("old CSet region num reached max") 2277 ergo_format_region("old") 2278 ergo_format_region("max"), 2279 old_cset_region_length(), max_old_cset_length); 2280 break; 2281 } 2282 2283 2284 // Stop adding regions if the remaining reclaimable space is 2285 // not above G1HeapWastePercent. 2286 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2287 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2288 double threshold = (double) G1HeapWastePercent; 2289 if (reclaimable_perc <= threshold) { 2290 // We've added enough old regions that the amount of uncollected 2291 // reclaimable space is at or below the waste threshold. Stop 2292 // adding old regions to the CSet. 2293 ergo_verbose5(ErgoCSetConstruction, 2294 "finish adding old regions to CSet", 2295 ergo_format_reason("reclaimable percentage not over threshold") 2296 ergo_format_region("old") 2297 ergo_format_region("max") 2298 ergo_format_byte_perc("reclaimable") 2299 ergo_format_perc("threshold"), 2300 old_cset_region_length(), 2301 max_old_cset_length, 2302 reclaimable_bytes, 2303 reclaimable_perc, threshold); 2304 break; 2305 } 2306 2307 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2308 if (check_time_remaining) { 2309 if (predicted_time_ms > time_remaining_ms) { 2310 // Too expensive for the current CSet. 2311 2312 if (old_cset_region_length() >= min_old_cset_length) { 2313 // We have added the minimum number of old regions to the CSet, 2314 // we are done with this CSet. 2315 ergo_verbose4(ErgoCSetConstruction, 2316 "finish adding old regions to CSet", 2317 ergo_format_reason("predicted time is too high") 2318 ergo_format_ms("predicted time") 2319 ergo_format_ms("remaining time") 2320 ergo_format_region("old") 2321 ergo_format_region("min"), 2322 predicted_time_ms, time_remaining_ms, 2323 old_cset_region_length(), min_old_cset_length); 2324 break; 2325 } 2326 2327 // We'll add it anyway given that we haven't reached the 2328 // minimum number of old regions. 2329 expensive_region_num += 1; 2330 } 2331 } else { 2332 if (old_cset_region_length() >= min_old_cset_length) { 2333 // In the non-auto-tuning case, we'll finish adding regions 2334 // to the CSet if we reach the minimum. 2335 ergo_verbose2(ErgoCSetConstruction, 2336 "finish adding old regions to CSet", 2337 ergo_format_reason("old CSet region num reached min") 2338 ergo_format_region("old") 2339 ergo_format_region("min"), 2340 old_cset_region_length(), min_old_cset_length); 2341 break; 2342 } 2343 } 2344 2345 // We will add this region to the CSet. 2346 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2347 predicted_old_time_ms += predicted_time_ms; 2348 cset_chooser->pop(); // already have region via peek() 2349 _g1->old_set_remove(hr); 2350 add_old_region_to_cset(hr); 2351 2352 hr = cset_chooser->peek(); 2353 } 2354 if (hr == NULL) { 2355 ergo_verbose0(ErgoCSetConstruction, 2356 "finish adding old regions to CSet", 2357 ergo_format_reason("candidate old regions not available")); 2358 } 2359 2360 if (expensive_region_num > 0) { 2361 // We print the information once here at the end, predicated on 2362 // whether we added any apparently expensive regions or not, to 2363 // avoid generating output per region. 2364 ergo_verbose4(ErgoCSetConstruction, 2365 "added expensive regions to CSet", 2366 ergo_format_reason("old CSet region num not reached min") 2367 ergo_format_region("old") 2368 ergo_format_region("expensive") 2369 ergo_format_region("min") 2370 ergo_format_ms("remaining time"), 2371 old_cset_region_length(), 2372 expensive_region_num, 2373 min_old_cset_length, 2374 time_remaining_ms); 2375 } 2376 2377 cset_chooser->verify(); 2378 } 2379 2380 stop_incremental_cset_building(); 2381 2382 ergo_verbose3(ErgoCSetConstruction, 2383 "finish choosing CSet", 2384 ergo_format_region("old") 2385 ergo_format_ms("predicted old region time") 2386 ergo_format_ms("time remaining"), 2387 old_cset_region_length(), 2388 predicted_old_time_ms, time_remaining_ms); 2389 2390 double non_young_end_time_sec = os::elapsedTime(); 2391 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2392 } 2393 2394 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2395 if(TraceYoungGenTime) { 2396 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2397 } 2398 } 2399 2400 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2401 if(TraceYoungGenTime) { 2402 _all_yield_times_ms.add(yield_time_ms); 2403 } 2404 } 2405 2406 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2407 if(TraceYoungGenTime) { 2408 _total.add(pause_time_ms); 2409 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2410 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2411 _parallel.add(phase_times->cur_collection_par_time_ms()); 2412 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2413 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2414 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2415 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2416 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2417 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2418 2419 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2420 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2421 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2422 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2423 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2424 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2425 2426 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2427 _parallel_other.add(parallel_other_time); 2428 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2429 } 2430 } 2431 2432 void TraceYoungGenTimeData::increment_young_collection_count() { 2433 if(TraceYoungGenTime) { 2434 ++_young_pause_num; 2435 } 2436 } 2437 2438 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2439 if(TraceYoungGenTime) { 2440 ++_mixed_pause_num; 2441 } 2442 } 2443 2444 void TraceYoungGenTimeData::print_summary(const char* str, 2445 const NumberSeq* seq) const { 2446 double sum = seq->sum(); 2447 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2448 str, sum / 1000.0, seq->avg()); 2449 } 2450 2451 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2452 const NumberSeq* seq) const { 2453 print_summary(str, seq); 2454 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2455 "(num", seq->num(), seq->sd(), seq->maximum()); 2456 } 2457 2458 void TraceYoungGenTimeData::print() const { 2459 if (!TraceYoungGenTime) { 2460 return; 2461 } 2462 2463 gclog_or_tty->print_cr("ALL PAUSES"); 2464 print_summary_sd(" Total", &_total); 2465 gclog_or_tty->cr(); 2466 gclog_or_tty->cr(); 2467 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2468 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2469 gclog_or_tty->cr(); 2470 2471 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2472 2473 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2474 gclog_or_tty->print_cr("none"); 2475 } else { 2476 print_summary_sd(" Evacuation Pauses", &_total); 2477 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2478 print_summary(" Parallel Time", &_parallel); 2479 print_summary(" Ext Root Scanning", &_ext_root_scan); 2480 print_summary(" SATB Filtering", &_satb_filtering); 2481 print_summary(" Update RS", &_update_rs); 2482 print_summary(" Scan RS", &_scan_rs); 2483 print_summary(" Object Copy", &_obj_copy); 2484 print_summary(" Termination", &_termination); 2485 print_summary(" Parallel Other", &_parallel_other); 2486 print_summary(" Clear CT", &_clear_ct); 2487 print_summary(" Other", &_other); 2488 } 2489 gclog_or_tty->cr(); 2490 2491 gclog_or_tty->print_cr("MISC"); 2492 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2493 print_summary_sd(" Yields", &_all_yield_times_ms); 2494 } 2495 2496 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2497 if (TraceOldGenTime) { 2498 _all_full_gc_times.add(full_gc_time_ms); 2499 } 2500 } 2501 2502 void TraceOldGenTimeData::print() const { 2503 if (!TraceOldGenTime) { 2504 return; 2505 } 2506 2507 if (_all_full_gc_times.num() > 0) { 2508 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2509 _all_full_gc_times.num(), 2510 _all_full_gc_times.sum() / 1000.0); 2511 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2512 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2513 _all_full_gc_times.sd(), 2514 _all_full_gc_times.maximum()); 2515 } 2516 }