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 #ifndef PRODUCT 1071 // Dump info to allow post-facto debugging 1072 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1073 gclog_or_tty->print_cr("-------------------------------------------"); 1074 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1075 _recent_gc_times_ms->dump(); 1076 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1077 _recent_prev_end_times_for_all_gcs_sec->dump(); 1078 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1079 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1080 // In debug mode, terminate the JVM if the user wants to debug at this point. 1081 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1082 #endif // !PRODUCT 1083 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1084 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1085 if (_recent_avg_pause_time_ratio < 0.0) { 1086 _recent_avg_pause_time_ratio = 0.0; 1087 } else { 1088 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1089 _recent_avg_pause_time_ratio = 1.0; 1090 } 1091 } 1092 } 1093 1094 bool new_in_marking_window = collector_state()->in_marking_window(); 1095 bool new_in_marking_window_im = false; 1096 if (last_pause_included_initial_mark) { 1097 new_in_marking_window = true; 1098 new_in_marking_window_im = true; 1099 } 1100 1101 if (collector_state()->last_young_gc()) { 1102 // This is supposed to to be the "last young GC" before we start 1103 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1104 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1105 1106 if (next_gc_should_be_mixed("start mixed GCs", 1107 "do not start mixed GCs")) { 1108 collector_state()->set_gcs_are_young(false); 1109 } else { 1110 // We aborted the mixed GC phase early. 1111 abort_time_to_mixed_tracking(); 1112 } 1113 1114 collector_state()->set_last_young_gc(false); 1115 } 1116 1117 if (!collector_state()->last_gc_was_young()) { 1118 // This is a mixed GC. Here we decide whether to continue doing 1119 // mixed GCs or not. 1120 if (!next_gc_should_be_mixed("continue mixed GCs", 1121 "do not continue mixed GCs")) { 1122 collector_state()->set_gcs_are_young(true); 1123 1124 maybe_start_marking(); 1125 } 1126 } 1127 1128 _short_lived_surv_rate_group->start_adding_regions(); 1129 // Do that for any other surv rate groups 1130 1131 if (update_stats) { 1132 double cost_per_card_ms = 0.0; 1133 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1134 if (_pending_cards > 0) { 1135 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1136 _cost_per_card_ms_seq->add(cost_per_card_ms); 1137 } 1138 _cost_scan_hcc_seq->add(cost_scan_hcc); 1139 1140 double cost_per_entry_ms = 0.0; 1141 if (cards_scanned > 10) { 1142 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1143 if (collector_state()->last_gc_was_young()) { 1144 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1145 } else { 1146 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1147 } 1148 } 1149 1150 if (_max_rs_lengths > 0) { 1151 double cards_per_entry_ratio = 1152 (double) cards_scanned / (double) _max_rs_lengths; 1153 if (collector_state()->last_gc_was_young()) { 1154 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1155 } else { 1156 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1157 } 1158 } 1159 1160 // This is defensive. For a while _max_rs_lengths could get 1161 // smaller than _recorded_rs_lengths which was causing 1162 // rs_length_diff to get very large and mess up the RSet length 1163 // predictions. The reason was unsafe concurrent updates to the 1164 // _inc_cset_recorded_rs_lengths field which the code below guards 1165 // against (see CR 7118202). This bug has now been fixed (see CR 1166 // 7119027). However, I'm still worried that 1167 // _inc_cset_recorded_rs_lengths might still end up somewhat 1168 // inaccurate. The concurrent refinement thread calculates an 1169 // RSet's length concurrently with other CR threads updating it 1170 // which might cause it to calculate the length incorrectly (if, 1171 // say, it's in mid-coarsening). So I'll leave in the defensive 1172 // conditional below just in case. 1173 size_t rs_length_diff = 0; 1174 if (_max_rs_lengths > _recorded_rs_lengths) { 1175 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1176 } 1177 _rs_length_diff_seq->add((double) rs_length_diff); 1178 1179 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1180 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1181 double cost_per_byte_ms = 0.0; 1182 1183 if (copied_bytes > 0) { 1184 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1185 if (collector_state()->in_marking_window()) { 1186 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1187 } else { 1188 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1189 } 1190 } 1191 1192 if (young_cset_region_length() > 0) { 1193 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1194 young_cset_region_length()); 1195 } 1196 1197 if (old_cset_region_length() > 0) { 1198 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1199 old_cset_region_length()); 1200 } 1201 1202 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1203 1204 _pending_cards_seq->add((double) _pending_cards); 1205 _rs_lengths_seq->add((double) _max_rs_lengths); 1206 } 1207 1208 collector_state()->set_in_marking_window(new_in_marking_window); 1209 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1210 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1211 // IHOP control wants to know the expected young gen length if it were not 1212 // restrained by the heap reserve. Using the actual length would make the 1213 // prediction too small and the limit the young gen every time we get to the 1214 // predicted target occupancy. 1215 size_t last_unrestrained_young_length = 0; 1216 update_young_list_max_and_target_length(&last_unrestrained_young_length); 1217 update_rs_lengths_prediction(); 1218 1219 double marking_to_mixed_time = -1.0; 1220 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1221 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1222 assert(marking_to_mixed_time > 0.0, 1223 "Initial mark to mixed time must be larger than zero but is %.3f", 1224 marking_to_mixed_time); 1225 } 1226 // Only update IHOP information on regular GCs. 1227 if (update_stats) { 1228 update_ihop_statistics(marking_to_mixed_time, 1229 app_time_ms / 1000.0, 1230 _last_old_allocated_bytes, 1231 last_unrestrained_young_length * HeapRegion::GrainBytes); 1232 } 1233 _last_old_allocated_bytes = 0; 1234 1235 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1236 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1237 1238 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1239 1240 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1241 ergo_verbose2(ErgoTiming, 1242 "adjust concurrent refinement thresholds", 1243 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1244 ergo_format_ms("Update RS time goal") 1245 ergo_format_ms("Scan HCC time"), 1246 update_rs_time_goal_ms, 1247 scan_hcc_time_ms); 1248 1249 update_rs_time_goal_ms = 0; 1250 } else { 1251 update_rs_time_goal_ms -= scan_hcc_time_ms; 1252 } 1253 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1254 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1255 update_rs_time_goal_ms); 1256 1257 _collectionSetChooser->verify(); 1258 } 1259 1260 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1261 if (G1UseAdaptiveIHOP) { 1262 // The target occupancy is the total heap occupancy we want to hit. First, we 1263 // want to avoid eating into the reserve intended for young GC (to avoid unnecessary 1264 // throughput loss). Additionally G1 is free to not clean out up to 1265 // G1HeapWastePercent of heap, that space also cannot be used for allocation 1266 // while marking. 1267 size_t safe_heap_percentage = (size_t) (G1ReservePercent + G1HeapWastePercent); 1268 size_t target_occupancy = 0; 1269 1270 if (safe_heap_percentage < 100) { 1271 target_occupancy = G1CollectedHeap::heap()->max_capacity() * (100.0 - safe_heap_percentage) / 100.0; 1272 } 1273 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1274 target_occupancy, 1275 &_predictor); 1276 } else { 1277 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1278 G1CollectedHeap::heap()->max_capacity()); 1279 } 1280 } 1281 1282 void G1CollectorPolicy::update_ihop_statistics(double marking_time, 1283 double mutator_time_s, 1284 size_t mutator_alloc_bytes, 1285 size_t young_gen_size) { 1286 bool report = false; 1287 1288 // To avoid using really small times that may be caused by e.g. back-to-back gcs 1289 // we filter them out. 1290 double const min_valid_time = 1e-6; 1291 1292 if (marking_time > min_valid_time) { 1293 _ihop_control->update_time_to_mixed(marking_time); 1294 report = true; 1295 } 1296 1297 // As an approximation for the young gc promotion rates during marking we use 1298 // all of them. In many applications there are only a few if any young gcs during 1299 // marking, which makes any prediction useless. This increases the accuracy of the 1300 // prediction. 1301 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1302 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1303 report = true; 1304 } 1305 1306 if (report) { 1307 report_ihop_statistics(); 1308 } 1309 } 1310 1311 void G1CollectorPolicy::report_ihop_statistics() { 1312 _ihop_control->print(); 1313 } 1314 1315 #define EXT_SIZE_FORMAT "%.1f%s" 1316 #define EXT_SIZE_PARAMS(bytes) \ 1317 byte_size_in_proper_unit((double)(bytes)), \ 1318 proper_unit_for_byte_size((bytes)) 1319 1320 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1321 YoungList* young_list = _g1->young_list(); 1322 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1323 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1324 _heap_capacity_bytes_before_gc = _g1->capacity(); 1325 _heap_used_bytes_before_gc = _g1->used(); 1326 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1327 1328 _eden_capacity_bytes_before_gc = 1329 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1330 1331 if (full) { 1332 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1333 } 1334 } 1335 1336 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1337 size_t bytes_after = _g1->used(); 1338 size_t capacity = _g1->capacity(); 1339 1340 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1341 byte_size_in_proper_unit(bytes_before), 1342 proper_unit_for_byte_size(bytes_before), 1343 byte_size_in_proper_unit(bytes_after), 1344 proper_unit_for_byte_size(bytes_after), 1345 byte_size_in_proper_unit(capacity), 1346 proper_unit_for_byte_size(capacity)); 1347 } 1348 1349 void G1CollectorPolicy::print_heap_transition() const { 1350 print_heap_transition(_heap_used_bytes_before_gc); 1351 } 1352 1353 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1354 YoungList* young_list = _g1->young_list(); 1355 1356 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1357 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1358 size_t heap_used_bytes_after_gc = _g1->used(); 1359 1360 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1361 size_t eden_capacity_bytes_after_gc = 1362 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1363 1364 gclog_or_tty->print( 1365 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1366 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1367 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1368 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1369 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1370 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1371 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1372 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1373 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1374 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1375 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1376 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1377 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1378 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1379 1380 if (full) { 1381 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1382 } 1383 1384 gclog_or_tty->cr(); 1385 } 1386 1387 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1388 phase_times()->print(pause_time_sec); 1389 } 1390 1391 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1392 double update_rs_processed_buffers, 1393 double goal_ms) { 1394 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1395 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1396 1397 if (G1UseAdaptiveConcRefinement) { 1398 const int k_gy = 3, k_gr = 6; 1399 const double inc_k = 1.1, dec_k = 0.9; 1400 1401 int g = cg1r->green_zone(); 1402 if (update_rs_time > goal_ms) { 1403 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1404 } else { 1405 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1406 g = (int)MAX2(g * inc_k, g + 1.0); 1407 } 1408 } 1409 // Change the refinement threads params 1410 cg1r->set_green_zone(g); 1411 cg1r->set_yellow_zone(g * k_gy); 1412 cg1r->set_red_zone(g * k_gr); 1413 cg1r->reinitialize_threads(); 1414 1415 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1416 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1417 cg1r->yellow_zone()); 1418 // Change the barrier params 1419 dcqs.set_process_completed_threshold(processing_threshold); 1420 dcqs.set_max_completed_queue(cg1r->red_zone()); 1421 } 1422 1423 int curr_queue_size = dcqs.completed_buffers_num(); 1424 if (curr_queue_size >= cg1r->yellow_zone()) { 1425 dcqs.set_completed_queue_padding(curr_queue_size); 1426 } else { 1427 dcqs.set_completed_queue_padding(0); 1428 } 1429 dcqs.notify_if_necessary(); 1430 } 1431 1432 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1433 return (size_t) get_new_prediction(_rs_length_diff_seq); 1434 } 1435 1436 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1437 return get_new_prediction(_alloc_rate_ms_seq); 1438 } 1439 1440 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1441 return get_new_prediction(_cost_per_card_ms_seq); 1442 } 1443 1444 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1445 return get_new_prediction(_cost_scan_hcc_seq); 1446 } 1447 1448 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1449 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1450 } 1451 1452 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1453 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1454 } 1455 1456 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1457 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1458 return predict_young_cards_per_entry_ratio(); 1459 } else { 1460 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1461 } 1462 } 1463 1464 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1465 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1466 } 1467 1468 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1469 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1470 } 1471 1472 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1473 if (collector_state()->gcs_are_young()) { 1474 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1475 } else { 1476 return predict_mixed_rs_scan_time_ms(card_num); 1477 } 1478 } 1479 1480 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1481 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1482 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1483 } else { 1484 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1485 } 1486 } 1487 1488 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1489 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1490 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1491 } else { 1492 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1493 } 1494 } 1495 1496 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1497 if (collector_state()->during_concurrent_mark()) { 1498 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1499 } else { 1500 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1501 } 1502 } 1503 1504 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1505 return get_new_prediction(_constant_other_time_ms_seq); 1506 } 1507 1508 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1509 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1510 } 1511 1512 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1513 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1514 } 1515 1516 double G1CollectorPolicy::predict_remark_time_ms() const { 1517 return get_new_prediction(_concurrent_mark_remark_times_ms); 1518 } 1519 1520 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1521 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1522 } 1523 1524 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1525 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1526 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1527 double pred = get_new_prediction(seq); 1528 if (pred > 1.0) { 1529 pred = 1.0; 1530 } 1531 return pred; 1532 } 1533 1534 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1535 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1536 } 1537 1538 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1539 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1540 } 1541 1542 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1543 size_t scanned_cards) const { 1544 return 1545 predict_rs_update_time_ms(pending_cards) + 1546 predict_rs_scan_time_ms(scanned_cards) + 1547 predict_constant_other_time_ms(); 1548 } 1549 1550 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1551 size_t rs_length = predict_rs_length_diff(); 1552 size_t card_num; 1553 if (collector_state()->gcs_are_young()) { 1554 card_num = predict_young_card_num(rs_length); 1555 } else { 1556 card_num = predict_non_young_card_num(rs_length); 1557 } 1558 return predict_base_elapsed_time_ms(pending_cards, card_num); 1559 } 1560 1561 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1562 size_t bytes_to_copy; 1563 if (hr->is_marked()) 1564 bytes_to_copy = hr->max_live_bytes(); 1565 else { 1566 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1567 int age = hr->age_in_surv_rate_group(); 1568 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1569 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1570 } 1571 return bytes_to_copy; 1572 } 1573 1574 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1575 bool for_young_gc) const { 1576 size_t rs_length = hr->rem_set()->occupied(); 1577 size_t card_num; 1578 1579 // Predicting the number of cards is based on which type of GC 1580 // we're predicting for. 1581 if (for_young_gc) { 1582 card_num = predict_young_card_num(rs_length); 1583 } else { 1584 card_num = predict_non_young_card_num(rs_length); 1585 } 1586 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1587 1588 double region_elapsed_time_ms = 1589 predict_rs_scan_time_ms(card_num) + 1590 predict_object_copy_time_ms(bytes_to_copy); 1591 1592 // The prediction of the "other" time for this region is based 1593 // upon the region type and NOT the GC type. 1594 if (hr->is_young()) { 1595 region_elapsed_time_ms += predict_young_other_time_ms(1); 1596 } else { 1597 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1598 } 1599 return region_elapsed_time_ms; 1600 } 1601 1602 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1603 uint survivor_cset_region_length) { 1604 _eden_cset_region_length = eden_cset_region_length; 1605 _survivor_cset_region_length = survivor_cset_region_length; 1606 _old_cset_region_length = 0; 1607 } 1608 1609 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1610 _recorded_rs_lengths = rs_lengths; 1611 } 1612 1613 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1614 double elapsed_ms) { 1615 _recent_gc_times_ms->add(elapsed_ms); 1616 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1617 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1618 } 1619 1620 size_t G1CollectorPolicy::expansion_amount() const { 1621 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1622 double threshold = _gc_overhead_perc; 1623 if (recent_gc_overhead > threshold) { 1624 // We will double the existing space, or take 1625 // G1ExpandByPercentOfAvailable % of the available expansion 1626 // space, whichever is smaller, bounded below by a minimum 1627 // expansion (unless that's all that's left.) 1628 const size_t min_expand_bytes = 1*M; 1629 size_t reserved_bytes = _g1->max_capacity(); 1630 size_t committed_bytes = _g1->capacity(); 1631 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1632 size_t expand_bytes; 1633 size_t expand_bytes_via_pct = 1634 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1635 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1636 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1637 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1638 1639 ergo_verbose5(ErgoHeapSizing, 1640 "attempt heap expansion", 1641 ergo_format_reason("recent GC overhead higher than " 1642 "threshold after GC") 1643 ergo_format_perc("recent GC overhead") 1644 ergo_format_perc("threshold") 1645 ergo_format_byte("uncommitted") 1646 ergo_format_byte_perc("calculated expansion amount"), 1647 recent_gc_overhead, threshold, 1648 uncommitted_bytes, 1649 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1650 1651 return expand_bytes; 1652 } else { 1653 return 0; 1654 } 1655 } 1656 1657 void G1CollectorPolicy::print_tracing_info() const { 1658 _trace_young_gen_time_data.print(); 1659 _trace_old_gen_time_data.print(); 1660 } 1661 1662 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1663 #ifndef PRODUCT 1664 _short_lived_surv_rate_group->print_surv_rate_summary(); 1665 // add this call for any other surv rate groups 1666 #endif // PRODUCT 1667 } 1668 1669 bool G1CollectorPolicy::is_young_list_full() const { 1670 uint young_list_length = _g1->young_list()->length(); 1671 uint young_list_target_length = _young_list_target_length; 1672 return young_list_length >= young_list_target_length; 1673 } 1674 1675 bool G1CollectorPolicy::can_expand_young_list() const { 1676 uint young_list_length = _g1->young_list()->length(); 1677 uint young_list_max_length = _young_list_max_length; 1678 return young_list_length < young_list_max_length; 1679 } 1680 1681 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1682 uint expansion_region_num = 0; 1683 if (GCLockerEdenExpansionPercent > 0) { 1684 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1685 double expansion_region_num_d = perc * (double) _young_list_target_length; 1686 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1687 // less than 1.0) we'll get 1. 1688 expansion_region_num = (uint) ceil(expansion_region_num_d); 1689 } else { 1690 assert(expansion_region_num == 0, "sanity"); 1691 } 1692 _young_list_max_length = _young_list_target_length + expansion_region_num; 1693 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1694 } 1695 1696 // Calculates survivor space parameters. 1697 void G1CollectorPolicy::update_survivors_policy() { 1698 double max_survivor_regions_d = 1699 (double) _young_list_target_length / (double) SurvivorRatio; 1700 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1701 // smaller than 1.0) we'll get 1. 1702 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1703 1704 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1705 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1706 } 1707 1708 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1709 // We actually check whether we are marking here and not if we are in a 1710 // reclamation phase. This means that we will schedule a concurrent mark 1711 // even while we are still in the process of reclaiming memory. 1712 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1713 if (!during_cycle) { 1714 ergo_verbose1(ErgoConcCycles, 1715 "request concurrent cycle initiation", 1716 ergo_format_reason("requested by GC cause") 1717 ergo_format_str("GC cause"), 1718 GCCause::to_string(gc_cause)); 1719 collector_state()->set_initiate_conc_mark_if_possible(true); 1720 return true; 1721 } else { 1722 ergo_verbose1(ErgoConcCycles, 1723 "do not request concurrent cycle initiation", 1724 ergo_format_reason("concurrent cycle already in progress") 1725 ergo_format_str("GC cause"), 1726 GCCause::to_string(gc_cause)); 1727 return false; 1728 } 1729 } 1730 1731 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1732 // We are about to decide on whether this pause will be an 1733 // initial-mark pause. 1734 1735 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1736 // will set it here if we have to. However, it should be cleared by 1737 // the end of the pause (it's only set for the duration of an 1738 // initial-mark pause). 1739 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1740 1741 if (collector_state()->initiate_conc_mark_if_possible()) { 1742 // We had noticed on a previous pause that the heap occupancy has 1743 // gone over the initiating threshold and we should start a 1744 // concurrent marking cycle. So we might initiate one. 1745 1746 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1747 // Initiate a new initial mark only if there is no marking or reclamation going 1748 // on. 1749 1750 collector_state()->set_during_initial_mark_pause(true); 1751 // And we can now clear initiate_conc_mark_if_possible() as 1752 // we've already acted on it. 1753 collector_state()->set_initiate_conc_mark_if_possible(false); 1754 1755 ergo_verbose0(ErgoConcCycles, 1756 "initiate concurrent cycle", 1757 ergo_format_reason("concurrent cycle initiation requested")); 1758 } else { 1759 // The concurrent marking thread is still finishing up the 1760 // previous cycle. If we start one right now the two cycles 1761 // overlap. In particular, the concurrent marking thread might 1762 // be in the process of clearing the next marking bitmap (which 1763 // we will use for the next cycle if we start one). Starting a 1764 // cycle now will be bad given that parts of the marking 1765 // information might get cleared by the marking thread. And we 1766 // cannot wait for the marking thread to finish the cycle as it 1767 // periodically yields while clearing the next marking bitmap 1768 // and, if it's in a yield point, it's waiting for us to 1769 // finish. So, at this point we will not start a cycle and we'll 1770 // let the concurrent marking thread complete the last one. 1771 ergo_verbose0(ErgoConcCycles, 1772 "do not initiate concurrent cycle", 1773 ergo_format_reason("concurrent cycle already in progress")); 1774 } 1775 } 1776 } 1777 1778 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1779 G1CollectedHeap* _g1h; 1780 CSetChooserParUpdater _cset_updater; 1781 1782 public: 1783 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1784 uint chunk_size) : 1785 _g1h(G1CollectedHeap::heap()), 1786 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1787 1788 bool doHeapRegion(HeapRegion* r) { 1789 // Do we have any marking information for this region? 1790 if (r->is_marked()) { 1791 // We will skip any region that's currently used as an old GC 1792 // alloc region (we should not consider those for collection 1793 // before we fill them up). 1794 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1795 _cset_updater.add_region(r); 1796 } 1797 } 1798 return false; 1799 } 1800 }; 1801 1802 class ParKnownGarbageTask: public AbstractGangTask { 1803 CollectionSetChooser* _hrSorted; 1804 uint _chunk_size; 1805 G1CollectedHeap* _g1; 1806 HeapRegionClaimer _hrclaimer; 1807 1808 public: 1809 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1810 AbstractGangTask("ParKnownGarbageTask"), 1811 _hrSorted(hrSorted), _chunk_size(chunk_size), 1812 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1813 1814 void work(uint worker_id) { 1815 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1816 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1817 } 1818 }; 1819 1820 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1821 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1822 const uint overpartition_factor = 4; 1823 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1824 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1825 } 1826 1827 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1828 _collectionSetChooser->clear(); 1829 1830 WorkGang* workers = _g1->workers(); 1831 uint n_workers = workers->active_workers(); 1832 1833 uint n_regions = _g1->num_regions(); 1834 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1835 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1836 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1837 workers->run_task(&par_known_garbage_task); 1838 1839 _collectionSetChooser->sort_regions(); 1840 1841 double end_sec = os::elapsedTime(); 1842 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1843 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1844 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1845 _prev_collection_pause_end_ms += elapsed_time_ms; 1846 1847 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1848 } 1849 1850 // Add the heap region at the head of the non-incremental collection set 1851 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1852 assert(_inc_cset_build_state == Active, "Precondition"); 1853 assert(hr->is_old(), "the region should be old"); 1854 1855 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1856 _g1->register_old_region_with_cset(hr); 1857 hr->set_next_in_collection_set(_collection_set); 1858 _collection_set = hr; 1859 _collection_set_bytes_used_before += hr->used(); 1860 size_t rs_length = hr->rem_set()->occupied(); 1861 _recorded_rs_lengths += rs_length; 1862 _old_cset_region_length += 1; 1863 } 1864 1865 // Initialize the per-collection-set information 1866 void G1CollectorPolicy::start_incremental_cset_building() { 1867 assert(_inc_cset_build_state == Inactive, "Precondition"); 1868 1869 _inc_cset_head = NULL; 1870 _inc_cset_tail = NULL; 1871 _inc_cset_bytes_used_before = 0; 1872 1873 _inc_cset_max_finger = 0; 1874 _inc_cset_recorded_rs_lengths = 0; 1875 _inc_cset_recorded_rs_lengths_diffs = 0; 1876 _inc_cset_predicted_elapsed_time_ms = 0.0; 1877 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1878 _inc_cset_build_state = Active; 1879 } 1880 1881 void G1CollectorPolicy::finalize_incremental_cset_building() { 1882 assert(_inc_cset_build_state == Active, "Precondition"); 1883 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1884 1885 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1886 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1887 // that adds a new region to the CSet. Further updates by the 1888 // concurrent refinement thread that samples the young RSet lengths 1889 // are accumulated in the *_diffs fields. Here we add the diffs to 1890 // the "main" fields. 1891 1892 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1893 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1894 } else { 1895 // This is defensive. The diff should in theory be always positive 1896 // as RSets can only grow between GCs. However, given that we 1897 // sample their size concurrently with other threads updating them 1898 // it's possible that we might get the wrong size back, which 1899 // could make the calculations somewhat inaccurate. 1900 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1901 if (_inc_cset_recorded_rs_lengths >= diffs) { 1902 _inc_cset_recorded_rs_lengths -= diffs; 1903 } else { 1904 _inc_cset_recorded_rs_lengths = 0; 1905 } 1906 } 1907 _inc_cset_predicted_elapsed_time_ms += 1908 _inc_cset_predicted_elapsed_time_ms_diffs; 1909 1910 _inc_cset_recorded_rs_lengths_diffs = 0; 1911 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1912 } 1913 1914 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1915 // This routine is used when: 1916 // * adding survivor regions to the incremental cset at the end of an 1917 // evacuation pause, 1918 // * adding the current allocation region to the incremental cset 1919 // when it is retired, and 1920 // * updating existing policy information for a region in the 1921 // incremental cset via young list RSet sampling. 1922 // Therefore this routine may be called at a safepoint by the 1923 // VM thread, or in-between safepoints by mutator threads (when 1924 // retiring the current allocation region) or a concurrent 1925 // refine thread (RSet sampling). 1926 1927 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1928 size_t used_bytes = hr->used(); 1929 _inc_cset_recorded_rs_lengths += rs_length; 1930 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1931 _inc_cset_bytes_used_before += used_bytes; 1932 1933 // Cache the values we have added to the aggregated information 1934 // in the heap region in case we have to remove this region from 1935 // the incremental collection set, or it is updated by the 1936 // rset sampling code 1937 hr->set_recorded_rs_length(rs_length); 1938 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1939 } 1940 1941 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1942 size_t new_rs_length) { 1943 // Update the CSet information that is dependent on the new RS length 1944 assert(hr->is_young(), "Precondition"); 1945 assert(!SafepointSynchronize::is_at_safepoint(), 1946 "should not be at a safepoint"); 1947 1948 // We could have updated _inc_cset_recorded_rs_lengths and 1949 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1950 // that atomically, as this code is executed by a concurrent 1951 // refinement thread, potentially concurrently with a mutator thread 1952 // allocating a new region and also updating the same fields. To 1953 // avoid the atomic operations we accumulate these updates on two 1954 // separate fields (*_diffs) and we'll just add them to the "main" 1955 // fields at the start of a GC. 1956 1957 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1958 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1959 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1960 1961 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1962 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1963 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1964 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1965 1966 hr->set_recorded_rs_length(new_rs_length); 1967 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1968 } 1969 1970 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1971 assert(hr->is_young(), "invariant"); 1972 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1973 assert(_inc_cset_build_state == Active, "Precondition"); 1974 1975 // We need to clear and set the cached recorded/cached collection set 1976 // information in the heap region here (before the region gets added 1977 // to the collection set). An individual heap region's cached values 1978 // are calculated, aggregated with the policy collection set info, 1979 // and cached in the heap region here (initially) and (subsequently) 1980 // by the Young List sampling code. 1981 1982 size_t rs_length = hr->rem_set()->occupied(); 1983 add_to_incremental_cset_info(hr, rs_length); 1984 1985 HeapWord* hr_end = hr->end(); 1986 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1987 1988 assert(!hr->in_collection_set(), "invariant"); 1989 _g1->register_young_region_with_cset(hr); 1990 assert(hr->next_in_collection_set() == NULL, "invariant"); 1991 } 1992 1993 // Add the region at the RHS of the incremental cset 1994 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1995 // We should only ever be appending survivors at the end of a pause 1996 assert(hr->is_survivor(), "Logic"); 1997 1998 // Do the 'common' stuff 1999 add_region_to_incremental_cset_common(hr); 2000 2001 // Now add the region at the right hand side 2002 if (_inc_cset_tail == NULL) { 2003 assert(_inc_cset_head == NULL, "invariant"); 2004 _inc_cset_head = hr; 2005 } else { 2006 _inc_cset_tail->set_next_in_collection_set(hr); 2007 } 2008 _inc_cset_tail = hr; 2009 } 2010 2011 // Add the region to the LHS of the incremental cset 2012 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2013 // Survivors should be added to the RHS at the end of a pause 2014 assert(hr->is_eden(), "Logic"); 2015 2016 // Do the 'common' stuff 2017 add_region_to_incremental_cset_common(hr); 2018 2019 // Add the region at the left hand side 2020 hr->set_next_in_collection_set(_inc_cset_head); 2021 if (_inc_cset_head == NULL) { 2022 assert(_inc_cset_tail == NULL, "Invariant"); 2023 _inc_cset_tail = hr; 2024 } 2025 _inc_cset_head = hr; 2026 } 2027 2028 #ifndef PRODUCT 2029 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2030 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2031 2032 st->print_cr("\nCollection_set:"); 2033 HeapRegion* csr = list_head; 2034 while (csr != NULL) { 2035 HeapRegion* next = csr->next_in_collection_set(); 2036 assert(csr->in_collection_set(), "bad CS"); 2037 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2038 HR_FORMAT_PARAMS(csr), 2039 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2040 csr->age_in_surv_rate_group_cond()); 2041 csr = next; 2042 } 2043 } 2044 #endif // !PRODUCT 2045 2046 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2047 // Returns the given amount of reclaimable bytes (that represents 2048 // the amount of reclaimable space still to be collected) as a 2049 // percentage of the current heap capacity. 2050 size_t capacity_bytes = _g1->capacity(); 2051 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2052 } 2053 2054 void G1CollectorPolicy::maybe_start_marking() { 2055 if (need_to_start_conc_mark("end of GC")) { 2056 // Note: this might have already been set, if during the last 2057 // pause we decided to start a cycle but at the beginning of 2058 // this pause we decided to postpone it. That's OK. 2059 collector_state()->set_initiate_conc_mark_if_possible(true); 2060 } 2061 } 2062 2063 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2064 assert(!collector_state()->full_collection(), "must be"); 2065 if (collector_state()->during_initial_mark_pause()) { 2066 assert(collector_state()->last_gc_was_young(), "must be"); 2067 assert(!collector_state()->last_young_gc(), "must be"); 2068 return InitialMarkGC; 2069 } else if (collector_state()->last_young_gc()) { 2070 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2071 assert(collector_state()->last_gc_was_young(), "must be"); 2072 return LastYoungGC; 2073 } else if (!collector_state()->last_gc_was_young()) { 2074 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2075 assert(!collector_state()->last_young_gc(), "must be"); 2076 return MixedGC; 2077 } else { 2078 assert(collector_state()->last_gc_was_young(), "must be"); 2079 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2080 assert(!collector_state()->last_young_gc(), "must be"); 2081 return YoungOnlyGC; 2082 } 2083 } 2084 2085 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2086 // Manage the MMU tracker. For some reason it ignores Full GCs. 2087 if (kind != FullGC) { 2088 _mmu_tracker->add_pause(start, end); 2089 } 2090 // Manage the mutator time tracking from initial mark to first mixed gc. 2091 switch (kind) { 2092 case FullGC: 2093 abort_time_to_mixed_tracking(); 2094 break; 2095 case Cleanup: 2096 case Remark: 2097 case YoungOnlyGC: 2098 case LastYoungGC: 2099 _initial_mark_to_mixed.add_pause(end - start); 2100 break; 2101 case InitialMarkGC: 2102 _initial_mark_to_mixed.record_initial_mark_end(end); 2103 break; 2104 case MixedGC: 2105 _initial_mark_to_mixed.record_mixed_gc_start(start); 2106 break; 2107 default: 2108 ShouldNotReachHere(); 2109 } 2110 } 2111 2112 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2113 _initial_mark_to_mixed.reset(); 2114 } 2115 2116 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2117 const char* false_action_str) const { 2118 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2119 if (cset_chooser->is_empty()) { 2120 ergo_verbose0(ErgoMixedGCs, 2121 false_action_str, 2122 ergo_format_reason("candidate old regions not available")); 2123 return false; 2124 } 2125 2126 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2127 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2128 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2129 double threshold = (double) G1HeapWastePercent; 2130 if (reclaimable_perc <= threshold) { 2131 ergo_verbose4(ErgoMixedGCs, 2132 false_action_str, 2133 ergo_format_reason("reclaimable percentage not over threshold") 2134 ergo_format_region("candidate old regions") 2135 ergo_format_byte_perc("reclaimable") 2136 ergo_format_perc("threshold"), 2137 cset_chooser->remaining_regions(), 2138 reclaimable_bytes, 2139 reclaimable_perc, threshold); 2140 return false; 2141 } 2142 2143 ergo_verbose4(ErgoMixedGCs, 2144 true_action_str, 2145 ergo_format_reason("candidate old regions available") 2146 ergo_format_region("candidate old regions") 2147 ergo_format_byte_perc("reclaimable") 2148 ergo_format_perc("threshold"), 2149 cset_chooser->remaining_regions(), 2150 reclaimable_bytes, 2151 reclaimable_perc, threshold); 2152 return true; 2153 } 2154 2155 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2156 // The min old CSet region bound is based on the maximum desired 2157 // number of mixed GCs after a cycle. I.e., even if some old regions 2158 // look expensive, we should add them to the CSet anyway to make 2159 // sure we go through the available old regions in no more than the 2160 // maximum desired number of mixed GCs. 2161 // 2162 // The calculation is based on the number of marked regions we added 2163 // to the CSet chooser in the first place, not how many remain, so 2164 // that the result is the same during all mixed GCs that follow a cycle. 2165 2166 const size_t region_num = (size_t) _collectionSetChooser->length(); 2167 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2168 size_t result = region_num / gc_num; 2169 // emulate ceiling 2170 if (result * gc_num < region_num) { 2171 result += 1; 2172 } 2173 return (uint) result; 2174 } 2175 2176 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2177 // The max old CSet region bound is based on the threshold expressed 2178 // as a percentage of the heap size. I.e., it should bound the 2179 // number of old regions added to the CSet irrespective of how many 2180 // of them are available. 2181 2182 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2183 const size_t region_num = g1h->num_regions(); 2184 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2185 size_t result = region_num * perc / 100; 2186 // emulate ceiling 2187 if (100 * result < region_num * perc) { 2188 result += 1; 2189 } 2190 return (uint) result; 2191 } 2192 2193 2194 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2195 double young_start_time_sec = os::elapsedTime(); 2196 2197 YoungList* young_list = _g1->young_list(); 2198 finalize_incremental_cset_building(); 2199 2200 guarantee(target_pause_time_ms > 0.0, 2201 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2202 guarantee(_collection_set == NULL, "Precondition"); 2203 2204 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2205 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2206 2207 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2208 "start choosing CSet", 2209 ergo_format_size("_pending_cards") 2210 ergo_format_ms("predicted base time") 2211 ergo_format_ms("remaining time") 2212 ergo_format_ms("target pause time"), 2213 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2214 2215 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2216 2217 if (collector_state()->last_gc_was_young()) { 2218 _trace_young_gen_time_data.increment_young_collection_count(); 2219 } else { 2220 _trace_young_gen_time_data.increment_mixed_collection_count(); 2221 } 2222 2223 // The young list is laid with the survivor regions from the previous 2224 // pause are appended to the RHS of the young list, i.e. 2225 // [Newly Young Regions ++ Survivors from last pause]. 2226 2227 uint survivor_region_length = young_list->survivor_length(); 2228 uint eden_region_length = young_list->eden_length(); 2229 init_cset_region_lengths(eden_region_length, survivor_region_length); 2230 2231 HeapRegion* hr = young_list->first_survivor_region(); 2232 while (hr != NULL) { 2233 assert(hr->is_survivor(), "badly formed young list"); 2234 // There is a convention that all the young regions in the CSet 2235 // are tagged as "eden", so we do this for the survivors here. We 2236 // use the special set_eden_pre_gc() as it doesn't check that the 2237 // region is free (which is not the case here). 2238 hr->set_eden_pre_gc(); 2239 hr = hr->get_next_young_region(); 2240 } 2241 2242 // Clear the fields that point to the survivor list - they are all young now. 2243 young_list->clear_survivors(); 2244 2245 _collection_set = _inc_cset_head; 2246 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2247 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2248 2249 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2250 "add young regions to CSet", 2251 ergo_format_region("eden") 2252 ergo_format_region("survivors") 2253 ergo_format_ms("predicted young region time") 2254 ergo_format_ms("target pause time"), 2255 eden_region_length, survivor_region_length, 2256 _inc_cset_predicted_elapsed_time_ms, 2257 target_pause_time_ms); 2258 2259 // The number of recorded young regions is the incremental 2260 // collection set's current size 2261 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2262 2263 double young_end_time_sec = os::elapsedTime(); 2264 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2265 2266 return time_remaining_ms; 2267 } 2268 2269 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2270 double non_young_start_time_sec = os::elapsedTime(); 2271 double predicted_old_time_ms = 0.0; 2272 2273 2274 if (!collector_state()->gcs_are_young()) { 2275 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2276 cset_chooser->verify(); 2277 const uint min_old_cset_length = calc_min_old_cset_length(); 2278 const uint max_old_cset_length = calc_max_old_cset_length(); 2279 2280 uint expensive_region_num = 0; 2281 bool check_time_remaining = adaptive_young_list_length(); 2282 2283 HeapRegion* hr = cset_chooser->peek(); 2284 while (hr != NULL) { 2285 if (old_cset_region_length() >= max_old_cset_length) { 2286 // Added maximum number of old regions to the CSet. 2287 ergo_verbose2(ErgoCSetConstruction, 2288 "finish adding old regions to CSet", 2289 ergo_format_reason("old CSet region num reached max") 2290 ergo_format_region("old") 2291 ergo_format_region("max"), 2292 old_cset_region_length(), max_old_cset_length); 2293 break; 2294 } 2295 2296 2297 // Stop adding regions if the remaining reclaimable space is 2298 // not above G1HeapWastePercent. 2299 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2300 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2301 double threshold = (double) G1HeapWastePercent; 2302 if (reclaimable_perc <= threshold) { 2303 // We've added enough old regions that the amount of uncollected 2304 // reclaimable space is at or below the waste threshold. Stop 2305 // adding old regions to the CSet. 2306 ergo_verbose5(ErgoCSetConstruction, 2307 "finish adding old regions to CSet", 2308 ergo_format_reason("reclaimable percentage not over threshold") 2309 ergo_format_region("old") 2310 ergo_format_region("max") 2311 ergo_format_byte_perc("reclaimable") 2312 ergo_format_perc("threshold"), 2313 old_cset_region_length(), 2314 max_old_cset_length, 2315 reclaimable_bytes, 2316 reclaimable_perc, threshold); 2317 break; 2318 } 2319 2320 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2321 if (check_time_remaining) { 2322 if (predicted_time_ms > time_remaining_ms) { 2323 // Too expensive for the current CSet. 2324 2325 if (old_cset_region_length() >= min_old_cset_length) { 2326 // We have added the minimum number of old regions to the CSet, 2327 // we are done with this CSet. 2328 ergo_verbose4(ErgoCSetConstruction, 2329 "finish adding old regions to CSet", 2330 ergo_format_reason("predicted time is too high") 2331 ergo_format_ms("predicted time") 2332 ergo_format_ms("remaining time") 2333 ergo_format_region("old") 2334 ergo_format_region("min"), 2335 predicted_time_ms, time_remaining_ms, 2336 old_cset_region_length(), min_old_cset_length); 2337 break; 2338 } 2339 2340 // We'll add it anyway given that we haven't reached the 2341 // minimum number of old regions. 2342 expensive_region_num += 1; 2343 } 2344 } else { 2345 if (old_cset_region_length() >= min_old_cset_length) { 2346 // In the non-auto-tuning case, we'll finish adding regions 2347 // to the CSet if we reach the minimum. 2348 ergo_verbose2(ErgoCSetConstruction, 2349 "finish adding old regions to CSet", 2350 ergo_format_reason("old CSet region num reached min") 2351 ergo_format_region("old") 2352 ergo_format_region("min"), 2353 old_cset_region_length(), min_old_cset_length); 2354 break; 2355 } 2356 } 2357 2358 // We will add this region to the CSet. 2359 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2360 predicted_old_time_ms += predicted_time_ms; 2361 cset_chooser->pop(); // already have region via peek() 2362 _g1->old_set_remove(hr); 2363 add_old_region_to_cset(hr); 2364 2365 hr = cset_chooser->peek(); 2366 } 2367 if (hr == NULL) { 2368 ergo_verbose0(ErgoCSetConstruction, 2369 "finish adding old regions to CSet", 2370 ergo_format_reason("candidate old regions not available")); 2371 } 2372 2373 if (expensive_region_num > 0) { 2374 // We print the information once here at the end, predicated on 2375 // whether we added any apparently expensive regions or not, to 2376 // avoid generating output per region. 2377 ergo_verbose4(ErgoCSetConstruction, 2378 "added expensive regions to CSet", 2379 ergo_format_reason("old CSet region num not reached min") 2380 ergo_format_region("old") 2381 ergo_format_region("expensive") 2382 ergo_format_region("min") 2383 ergo_format_ms("remaining time"), 2384 old_cset_region_length(), 2385 expensive_region_num, 2386 min_old_cset_length, 2387 time_remaining_ms); 2388 } 2389 2390 cset_chooser->verify(); 2391 } 2392 2393 stop_incremental_cset_building(); 2394 2395 ergo_verbose3(ErgoCSetConstruction, 2396 "finish choosing CSet", 2397 ergo_format_region("old") 2398 ergo_format_ms("predicted old region time") 2399 ergo_format_ms("time remaining"), 2400 old_cset_region_length(), 2401 predicted_old_time_ms, time_remaining_ms); 2402 2403 double non_young_end_time_sec = os::elapsedTime(); 2404 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2405 } 2406 2407 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2408 if(TraceYoungGenTime) { 2409 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2410 } 2411 } 2412 2413 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2414 if(TraceYoungGenTime) { 2415 _all_yield_times_ms.add(yield_time_ms); 2416 } 2417 } 2418 2419 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2420 if(TraceYoungGenTime) { 2421 _total.add(pause_time_ms); 2422 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2423 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2424 _parallel.add(phase_times->cur_collection_par_time_ms()); 2425 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2426 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2427 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2428 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2429 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2430 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2431 2432 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2433 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2434 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2435 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2436 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2437 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2438 2439 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2440 _parallel_other.add(parallel_other_time); 2441 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2442 } 2443 } 2444 2445 void TraceYoungGenTimeData::increment_young_collection_count() { 2446 if(TraceYoungGenTime) { 2447 ++_young_pause_num; 2448 } 2449 } 2450 2451 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2452 if(TraceYoungGenTime) { 2453 ++_mixed_pause_num; 2454 } 2455 } 2456 2457 void TraceYoungGenTimeData::print_summary(const char* str, 2458 const NumberSeq* seq) const { 2459 double sum = seq->sum(); 2460 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2461 str, sum / 1000.0, seq->avg()); 2462 } 2463 2464 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2465 const NumberSeq* seq) const { 2466 print_summary(str, seq); 2467 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2468 "(num", seq->num(), seq->sd(), seq->maximum()); 2469 } 2470 2471 void TraceYoungGenTimeData::print() const { 2472 if (!TraceYoungGenTime) { 2473 return; 2474 } 2475 2476 gclog_or_tty->print_cr("ALL PAUSES"); 2477 print_summary_sd(" Total", &_total); 2478 gclog_or_tty->cr(); 2479 gclog_or_tty->cr(); 2480 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2481 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2482 gclog_or_tty->cr(); 2483 2484 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2485 2486 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2487 gclog_or_tty->print_cr("none"); 2488 } else { 2489 print_summary_sd(" Evacuation Pauses", &_total); 2490 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2491 print_summary(" Parallel Time", &_parallel); 2492 print_summary(" Ext Root Scanning", &_ext_root_scan); 2493 print_summary(" SATB Filtering", &_satb_filtering); 2494 print_summary(" Update RS", &_update_rs); 2495 print_summary(" Scan RS", &_scan_rs); 2496 print_summary(" Object Copy", &_obj_copy); 2497 print_summary(" Termination", &_termination); 2498 print_summary(" Parallel Other", &_parallel_other); 2499 print_summary(" Clear CT", &_clear_ct); 2500 print_summary(" Other", &_other); 2501 } 2502 gclog_or_tty->cr(); 2503 2504 gclog_or_tty->print_cr("MISC"); 2505 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2506 print_summary_sd(" Yields", &_all_yield_times_ms); 2507 } 2508 2509 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2510 if (TraceOldGenTime) { 2511 _all_full_gc_times.add(full_gc_time_ms); 2512 } 2513 } 2514 2515 void TraceOldGenTimeData::print() const { 2516 if (!TraceOldGenTime) { 2517 return; 2518 } 2519 2520 if (_all_full_gc_times.num() > 0) { 2521 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2522 _all_full_gc_times.num(), 2523 _all_full_gc_times.sum() / 1000.0); 2524 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2525 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2526 _all_full_gc_times.sd(), 2527 _all_full_gc_times.maximum()); 2528 } 2529 }