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 #include "utilities/pair.hpp" 43 44 // Different defaults for different number of GC threads 45 // They were chosen by running GCOld and SPECjbb on debris with different 46 // numbers of GC threads and choosing them based on the results 47 48 // all the same 49 static double rs_length_diff_defaults[] = { 50 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 51 }; 52 53 static double cost_per_card_ms_defaults[] = { 54 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 55 }; 56 57 // all the same 58 static double young_cards_per_entry_ratio_defaults[] = { 59 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 60 }; 61 62 static double cost_per_entry_ms_defaults[] = { 63 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 64 }; 65 66 static double cost_per_byte_ms_defaults[] = { 67 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 68 }; 69 70 // these should be pretty consistent 71 static double constant_other_time_ms_defaults[] = { 72 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 73 }; 74 75 76 static double young_other_cost_per_region_ms_defaults[] = { 77 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 78 }; 79 80 static double non_young_other_cost_per_region_ms_defaults[] = { 81 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 82 }; 83 84 G1CollectorPolicy::G1CollectorPolicy() : 85 _predictor(G1ConfidencePercent / 100.0), 86 _parallel_gc_threads(ParallelGCThreads), 87 88 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 89 _stop_world_start(0.0), 90 91 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 92 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 93 94 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _prev_collection_pause_end_ms(0.0), 96 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 106 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 107 _non_young_other_cost_per_region_ms_seq( 108 new TruncatedSeq(TruncatedSeqLength)), 109 110 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 111 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 112 113 _pause_time_target_ms((double) MaxGCPauseMillis), 114 115 _recent_prev_end_times_for_all_gcs_sec( 116 new TruncatedSeq(NumPrevPausesForHeuristics)), 117 118 _recent_avg_pause_time_ratio(0.0), 119 _rs_lengths_prediction(0), 120 _max_survivor_regions(0), 121 122 _eden_used_bytes_before_gc(0), 123 _survivor_used_bytes_before_gc(0), 124 _heap_used_bytes_before_gc(0), 125 _metaspace_used_bytes_before_gc(0), 126 _eden_capacity_bytes_before_gc(0), 127 _heap_capacity_bytes_before_gc(0), 128 129 _eden_cset_region_length(0), 130 _survivor_cset_region_length(0), 131 _old_cset_region_length(0), 132 133 _collection_set(NULL), 134 _collection_set_bytes_used_before(0), 135 136 // Incremental CSet attributes 137 _inc_cset_build_state(Inactive), 138 _inc_cset_head(NULL), 139 _inc_cset_tail(NULL), 140 _inc_cset_bytes_used_before(0), 141 _inc_cset_max_finger(NULL), 142 _inc_cset_recorded_rs_lengths(0), 143 _inc_cset_recorded_rs_lengths_diffs(0), 144 _inc_cset_predicted_elapsed_time_ms(0.0), 145 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 146 147 // add here any more surv rate groups 148 _recorded_survivor_regions(0), 149 _recorded_survivor_head(NULL), 150 _recorded_survivor_tail(NULL), 151 _survivors_age_table(true), 152 153 _gc_overhead_perc(0.0), 154 155 _last_old_allocated_bytes(0), 156 _ihop_control(NULL), 157 _initial_mark_to_mixed() { 158 159 // SurvRateGroups below must be initialized after the predictor because they 160 // indirectly use it through this object passed to their constructor. 161 _short_lived_surv_rate_group = 162 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 163 _survivor_surv_rate_group = 164 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 165 166 // Set up the region size and associated fields. Given that the 167 // policy is created before the heap, we have to set this up here, 168 // so it's done as soon as possible. 169 170 // It would have been natural to pass initial_heap_byte_size() and 171 // max_heap_byte_size() to setup_heap_region_size() but those have 172 // not been set up at this point since they should be aligned with 173 // the region size. So, there is a circular dependency here. We base 174 // the region size on the heap size, but the heap size should be 175 // aligned with the region size. To get around this we use the 176 // unaligned values for the heap. 177 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 178 HeapRegionRemSet::setup_remset_size(); 179 180 G1ErgoVerbose::initialize(); 181 if (PrintAdaptiveSizePolicy) { 182 // Currently, we only use a single switch for all the heuristics. 183 G1ErgoVerbose::set_enabled(true); 184 // Given that we don't currently have a verboseness level 185 // parameter, we'll hardcode this to high. This can be easily 186 // changed in the future. 187 G1ErgoVerbose::set_level(ErgoHigh); 188 } else { 189 G1ErgoVerbose::set_enabled(false); 190 } 191 192 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 193 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 194 195 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 196 197 int index = MIN2(_parallel_gc_threads - 1, 7); 198 199 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 200 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 201 _cost_scan_hcc_seq->add(0.0); 202 _young_cards_per_entry_ratio_seq->add( 203 young_cards_per_entry_ratio_defaults[index]); 204 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 205 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 206 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 207 _young_other_cost_per_region_ms_seq->add( 208 young_other_cost_per_region_ms_defaults[index]); 209 _non_young_other_cost_per_region_ms_seq->add( 210 non_young_other_cost_per_region_ms_defaults[index]); 211 212 // Below, we might need to calculate the pause time target based on 213 // the pause interval. When we do so we are going to give G1 maximum 214 // flexibility and allow it to do pauses when it needs to. So, we'll 215 // arrange that the pause interval to be pause time target + 1 to 216 // ensure that a) the pause time target is maximized with respect to 217 // the pause interval and b) we maintain the invariant that pause 218 // time target < pause interval. If the user does not want this 219 // maximum flexibility, they will have to set the pause interval 220 // explicitly. 221 222 // First make sure that, if either parameter is set, its value is 223 // reasonable. 224 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 225 if (MaxGCPauseMillis < 1) { 226 vm_exit_during_initialization("MaxGCPauseMillis should be " 227 "greater than 0"); 228 } 229 } 230 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 231 if (GCPauseIntervalMillis < 1) { 232 vm_exit_during_initialization("GCPauseIntervalMillis should be " 233 "greater than 0"); 234 } 235 } 236 237 // Then, if the pause time target parameter was not set, set it to 238 // the default value. 239 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 240 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 241 // The default pause time target in G1 is 200ms 242 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 243 } else { 244 // We do not allow the pause interval to be set without the 245 // pause time target 246 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 247 "without setting MaxGCPauseMillis"); 248 } 249 } 250 251 // Then, if the interval parameter was not set, set it according to 252 // the pause time target (this will also deal with the case when the 253 // pause time target is the default value). 254 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 255 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 256 } 257 258 // Finally, make sure that the two parameters are consistent. 259 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 260 char buffer[256]; 261 jio_snprintf(buffer, 256, 262 "MaxGCPauseMillis (%u) should be less than " 263 "GCPauseIntervalMillis (%u)", 264 MaxGCPauseMillis, GCPauseIntervalMillis); 265 vm_exit_during_initialization(buffer); 266 } 267 268 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 269 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 270 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 271 272 // start conservatively (around 50ms is about right) 273 _concurrent_mark_remark_times_ms->add(0.05); 274 _concurrent_mark_cleanup_times_ms->add(0.20); 275 _tenuring_threshold = MaxTenuringThreshold; 276 277 assert(GCTimeRatio > 0, 278 "we should have set it to a default value set_g1_gc_flags() " 279 "if a user set it to 0"); 280 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 281 282 uintx reserve_perc = G1ReservePercent; 283 // Put an artificial ceiling on this so that it's not set to a silly value. 284 if (reserve_perc > 50) { 285 reserve_perc = 50; 286 warning("G1ReservePercent is set to a value that is too large, " 287 "it's been updated to " UINTX_FORMAT, reserve_perc); 288 } 289 _reserve_factor = (double) reserve_perc / 100.0; 290 // This will be set when the heap is expanded 291 // for the first time during initialization. 292 _reserve_regions = 0; 293 294 _collectionSetChooser = new CollectionSetChooser(); 295 } 296 297 G1CollectorPolicy::~G1CollectorPolicy() { 298 delete _ihop_control; 299 } 300 301 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 302 return _predictor.get_new_prediction(seq); 303 } 304 305 void G1CollectorPolicy::initialize_alignments() { 306 _space_alignment = HeapRegion::GrainBytes; 307 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 308 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 309 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 310 } 311 312 void G1CollectorPolicy::initialize_flags() { 313 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 314 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 315 } 316 317 if (SurvivorRatio < 1) { 318 vm_exit_during_initialization("Invalid survivor ratio specified"); 319 } 320 CollectorPolicy::initialize_flags(); 321 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 322 } 323 324 void G1CollectorPolicy::post_heap_initialize() { 325 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 326 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 327 if (max_young_size != MaxNewSize) { 328 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 329 } 330 331 _ihop_control = create_ihop_control(); 332 } 333 334 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 335 336 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 337 _min_desired_young_length(0), _max_desired_young_length(0) { 338 if (FLAG_IS_CMDLINE(NewRatio)) { 339 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 340 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 341 } else { 342 _sizer_kind = SizerNewRatio; 343 _adaptive_size = false; 344 return; 345 } 346 } 347 348 if (NewSize > MaxNewSize) { 349 if (FLAG_IS_CMDLINE(MaxNewSize)) { 350 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 351 "A new max generation size of " SIZE_FORMAT "k will be used.", 352 NewSize/K, MaxNewSize/K, NewSize/K); 353 } 354 MaxNewSize = NewSize; 355 } 356 357 if (FLAG_IS_CMDLINE(NewSize)) { 358 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 359 1U); 360 if (FLAG_IS_CMDLINE(MaxNewSize)) { 361 _max_desired_young_length = 362 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 363 1U); 364 _sizer_kind = SizerMaxAndNewSize; 365 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 366 } else { 367 _sizer_kind = SizerNewSizeOnly; 368 } 369 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 370 _max_desired_young_length = 371 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 372 1U); 373 _sizer_kind = SizerMaxNewSizeOnly; 374 } 375 } 376 377 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 378 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 379 return MAX2(1U, default_value); 380 } 381 382 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 383 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 384 return MAX2(1U, default_value); 385 } 386 387 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 388 assert(number_of_heap_regions > 0, "Heap must be initialized"); 389 390 switch (_sizer_kind) { 391 case SizerDefaults: 392 *min_young_length = calculate_default_min_length(number_of_heap_regions); 393 *max_young_length = calculate_default_max_length(number_of_heap_regions); 394 break; 395 case SizerNewSizeOnly: 396 *max_young_length = calculate_default_max_length(number_of_heap_regions); 397 *max_young_length = MAX2(*min_young_length, *max_young_length); 398 break; 399 case SizerMaxNewSizeOnly: 400 *min_young_length = calculate_default_min_length(number_of_heap_regions); 401 *min_young_length = MIN2(*min_young_length, *max_young_length); 402 break; 403 case SizerMaxAndNewSize: 404 // Do nothing. Values set on the command line, don't update them at runtime. 405 break; 406 case SizerNewRatio: 407 *min_young_length = number_of_heap_regions / (NewRatio + 1); 408 *max_young_length = *min_young_length; 409 break; 410 default: 411 ShouldNotReachHere(); 412 } 413 414 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 415 } 416 417 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 418 // We need to pass the desired values because recalculation may not update these 419 // values in some cases. 420 uint temp = _min_desired_young_length; 421 uint result = _max_desired_young_length; 422 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 423 return result; 424 } 425 426 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 427 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 428 &_max_desired_young_length); 429 } 430 431 void G1CollectorPolicy::init() { 432 // Set aside an initial future to_space. 433 _g1 = G1CollectedHeap::heap(); 434 435 assert(Heap_lock->owned_by_self(), "Locking discipline."); 436 437 initialize_gc_policy_counters(); 438 439 if (adaptive_young_list_length()) { 440 _young_list_fixed_length = 0; 441 } else { 442 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 443 } 444 _free_regions_at_end_of_collection = _g1->num_free_regions(); 445 446 update_young_list_max_and_target_length(); 447 // We may immediately start allocating regions and placing them on the 448 // collection set list. Initialize the per-collection set info 449 start_incremental_cset_building(); 450 } 451 452 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 453 phase_times()->note_gc_start(num_active_workers); 454 } 455 456 // Create the jstat counters for the policy. 457 void G1CollectorPolicy::initialize_gc_policy_counters() { 458 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 459 } 460 461 bool G1CollectorPolicy::predict_will_fit(uint young_length, 462 double base_time_ms, 463 uint base_free_regions, 464 double target_pause_time_ms) const { 465 if (young_length >= base_free_regions) { 466 // end condition 1: not enough space for the young regions 467 return false; 468 } 469 470 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 471 size_t bytes_to_copy = 472 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 473 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 474 double young_other_time_ms = predict_young_other_time_ms(young_length); 475 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 476 if (pause_time_ms > target_pause_time_ms) { 477 // end condition 2: prediction is over the target pause time 478 return false; 479 } 480 481 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 482 483 // When copying, we will likely need more bytes free than is live in the region. 484 // Add some safety margin to factor in the confidence of our guess, and the 485 // natural expected waste. 486 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 487 // of the calculation: the lower the confidence, the more headroom. 488 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 489 // copying due to anticipated waste in the PLABs. 490 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 491 size_t expected_bytes_to_copy = safety_factor * bytes_to_copy; 492 493 if (expected_bytes_to_copy > free_bytes) { 494 // end condition 3: out-of-space 495 return false; 496 } 497 498 // success! 499 return true; 500 } 501 502 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 503 // re-calculate the necessary reserve 504 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 505 // We use ceiling so that if reserve_regions_d is > 0.0 (but 506 // smaller than 1.0) we'll get 1. 507 _reserve_regions = (uint) ceil(reserve_regions_d); 508 509 _young_gen_sizer->heap_size_changed(new_number_of_regions); 510 } 511 512 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 513 uint base_min_length) const { 514 uint desired_min_length = 0; 515 if (adaptive_young_list_length()) { 516 if (_alloc_rate_ms_seq->num() > 3) { 517 double now_sec = os::elapsedTime(); 518 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 519 double alloc_rate_ms = predict_alloc_rate_ms(); 520 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 521 } else { 522 // otherwise we don't have enough info to make the prediction 523 } 524 } 525 desired_min_length += base_min_length; 526 // make sure we don't go below any user-defined minimum bound 527 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 528 } 529 530 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 531 // Here, we might want to also take into account any additional 532 // constraints (i.e., user-defined minimum bound). Currently, we 533 // effectively don't set this bound. 534 return _young_gen_sizer->max_desired_young_length(); 535 } 536 537 uint G1CollectorPolicy::update_young_list_max_and_target_length() { 538 return update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 539 } 540 541 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 542 uint unbounded_target_length = update_young_list_target_length(rs_lengths); 543 update_max_gc_locker_expansion(); 544 return unbounded_target_length; 545 } 546 547 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 548 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths); 549 _young_list_target_length = young_lengths.first; 550 return young_lengths.second; 551 } 552 553 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const { 554 YoungTargetLengths result; 555 556 // Calculate the absolute and desired min bounds first. 557 558 // This is how many young regions we already have (currently: the survivors). 559 uint base_min_length = recorded_survivor_regions(); 560 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 561 // This is the absolute minimum young length. Ensure that we 562 // will at least have one eden region available for allocation. 563 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 564 // If we shrank the young list target it should not shrink below the current size. 565 desired_min_length = MAX2(desired_min_length, absolute_min_length); 566 // Calculate the absolute and desired max bounds. 567 568 uint desired_max_length = calculate_young_list_desired_max_length(); 569 570 uint young_list_target_length = 0; 571 if (adaptive_young_list_length()) { 572 if (collector_state()->gcs_are_young()) { 573 young_list_target_length = 574 calculate_young_list_target_length(rs_lengths, 575 base_min_length, 576 desired_min_length, 577 desired_max_length); 578 } else { 579 // Don't calculate anything and let the code below bound it to 580 // the desired_min_length, i.e., do the next GC as soon as 581 // possible to maximize how many old regions we can add to it. 582 } 583 } else { 584 // The user asked for a fixed young gen so we'll fix the young gen 585 // whether the next GC is young or mixed. 586 young_list_target_length = _young_list_fixed_length; 587 } 588 589 result.second = young_list_target_length; 590 591 // We will try our best not to "eat" into the reserve. 592 uint absolute_max_length = 0; 593 if (_free_regions_at_end_of_collection > _reserve_regions) { 594 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 595 } 596 if (desired_max_length > absolute_max_length) { 597 desired_max_length = absolute_max_length; 598 } 599 600 // Make sure we don't go over the desired max length, nor under the 601 // desired min length. In case they clash, desired_min_length wins 602 // which is why that test is second. 603 if (young_list_target_length > desired_max_length) { 604 young_list_target_length = desired_max_length; 605 } 606 if (young_list_target_length < desired_min_length) { 607 young_list_target_length = desired_min_length; 608 } 609 610 assert(young_list_target_length > recorded_survivor_regions(), 611 "we should be able to allocate at least one eden region"); 612 assert(young_list_target_length >= absolute_min_length, "post-condition"); 613 614 result.first = young_list_target_length; 615 return result; 616 } 617 618 uint 619 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 620 uint base_min_length, 621 uint desired_min_length, 622 uint desired_max_length) const { 623 assert(adaptive_young_list_length(), "pre-condition"); 624 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 625 626 // In case some edge-condition makes the desired max length too small... 627 if (desired_max_length <= desired_min_length) { 628 return desired_min_length; 629 } 630 631 // We'll adjust min_young_length and max_young_length not to include 632 // the already allocated young regions (i.e., so they reflect the 633 // min and max eden regions we'll allocate). The base_min_length 634 // will be reflected in the predictions by the 635 // survivor_regions_evac_time prediction. 636 assert(desired_min_length > base_min_length, "invariant"); 637 uint min_young_length = desired_min_length - base_min_length; 638 assert(desired_max_length > base_min_length, "invariant"); 639 uint max_young_length = desired_max_length - base_min_length; 640 641 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 642 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 643 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 644 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 645 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 646 double base_time_ms = 647 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 648 survivor_regions_evac_time; 649 uint available_free_regions = _free_regions_at_end_of_collection; 650 uint base_free_regions = 0; 651 if (available_free_regions > _reserve_regions) { 652 base_free_regions = available_free_regions - _reserve_regions; 653 } 654 655 // Here, we will make sure that the shortest young length that 656 // makes sense fits within the target pause time. 657 658 if (predict_will_fit(min_young_length, base_time_ms, 659 base_free_regions, target_pause_time_ms)) { 660 // The shortest young length will fit into the target pause time; 661 // we'll now check whether the absolute maximum number of young 662 // regions will fit in the target pause time. If not, we'll do 663 // a binary search between min_young_length and max_young_length. 664 if (predict_will_fit(max_young_length, base_time_ms, 665 base_free_regions, target_pause_time_ms)) { 666 // The maximum young length will fit into the target pause time. 667 // We are done so set min young length to the maximum length (as 668 // the result is assumed to be returned in min_young_length). 669 min_young_length = max_young_length; 670 } else { 671 // The maximum possible number of young regions will not fit within 672 // the target pause time so we'll search for the optimal 673 // length. The loop invariants are: 674 // 675 // min_young_length < max_young_length 676 // min_young_length is known to fit into the target pause time 677 // max_young_length is known not to fit into the target pause time 678 // 679 // Going into the loop we know the above hold as we've just 680 // checked them. Every time around the loop we check whether 681 // the middle value between min_young_length and 682 // max_young_length fits into the target pause time. If it 683 // does, it becomes the new min. If it doesn't, it becomes 684 // the new max. This way we maintain the loop invariants. 685 686 assert(min_young_length < max_young_length, "invariant"); 687 uint diff = (max_young_length - min_young_length) / 2; 688 while (diff > 0) { 689 uint young_length = min_young_length + diff; 690 if (predict_will_fit(young_length, base_time_ms, 691 base_free_regions, target_pause_time_ms)) { 692 min_young_length = young_length; 693 } else { 694 max_young_length = young_length; 695 } 696 assert(min_young_length < max_young_length, "invariant"); 697 diff = (max_young_length - min_young_length) / 2; 698 } 699 // The results is min_young_length which, according to the 700 // loop invariants, should fit within the target pause time. 701 702 // These are the post-conditions of the binary search above: 703 assert(min_young_length < max_young_length, 704 "otherwise we should have discovered that max_young_length " 705 "fits into the pause target and not done the binary search"); 706 assert(predict_will_fit(min_young_length, base_time_ms, 707 base_free_regions, target_pause_time_ms), 708 "min_young_length, the result of the binary search, should " 709 "fit into the pause target"); 710 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 711 base_free_regions, target_pause_time_ms), 712 "min_young_length, the result of the binary search, should be " 713 "optimal, so no larger length should fit into the pause target"); 714 } 715 } else { 716 // Even the minimum length doesn't fit into the pause time 717 // target, return it as the result nevertheless. 718 } 719 return base_min_length + min_young_length; 720 } 721 722 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 723 double survivor_regions_evac_time = 0.0; 724 for (HeapRegion * r = _recorded_survivor_head; 725 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 726 r = r->get_next_young_region()) { 727 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 728 } 729 return survivor_regions_evac_time; 730 } 731 732 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 733 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 734 735 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 736 if (rs_lengths > _rs_lengths_prediction) { 737 // add 10% to avoid having to recalculate often 738 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 739 update_rs_lengths_prediction(rs_lengths_prediction); 740 741 update_young_list_max_and_target_length(rs_lengths_prediction); 742 } 743 } 744 745 void G1CollectorPolicy::update_rs_lengths_prediction() { 746 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 747 } 748 749 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 750 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 751 _rs_lengths_prediction = prediction; 752 } 753 } 754 755 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 756 bool is_tlab, 757 bool* gc_overhead_limit_was_exceeded) { 758 guarantee(false, "Not using this policy feature yet."); 759 return NULL; 760 } 761 762 // This method controls how a collector handles one or more 763 // of its generations being fully allocated. 764 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 765 bool is_tlab) { 766 guarantee(false, "Not using this policy feature yet."); 767 return NULL; 768 } 769 770 771 #ifndef PRODUCT 772 bool G1CollectorPolicy::verify_young_ages() { 773 HeapRegion* head = _g1->young_list()->first_region(); 774 return 775 verify_young_ages(head, _short_lived_surv_rate_group); 776 // also call verify_young_ages on any additional surv rate groups 777 } 778 779 bool 780 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 781 SurvRateGroup *surv_rate_group) { 782 guarantee( surv_rate_group != NULL, "pre-condition" ); 783 784 const char* name = surv_rate_group->name(); 785 bool ret = true; 786 int prev_age = -1; 787 788 for (HeapRegion* curr = head; 789 curr != NULL; 790 curr = curr->get_next_young_region()) { 791 SurvRateGroup* group = curr->surv_rate_group(); 792 if (group == NULL && !curr->is_survivor()) { 793 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 794 ret = false; 795 } 796 797 if (surv_rate_group == group) { 798 int age = curr->age_in_surv_rate_group(); 799 800 if (age < 0) { 801 gclog_or_tty->print_cr("## %s: encountered negative age", name); 802 ret = false; 803 } 804 805 if (age <= prev_age) { 806 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 807 "(%d, %d)", name, age, prev_age); 808 ret = false; 809 } 810 prev_age = age; 811 } 812 } 813 814 return ret; 815 } 816 #endif // PRODUCT 817 818 void G1CollectorPolicy::record_full_collection_start() { 819 _full_collection_start_sec = os::elapsedTime(); 820 record_heap_size_info_at_start(true /* full */); 821 // Release the future to-space so that it is available for compaction into. 822 collector_state()->set_full_collection(true); 823 } 824 825 void G1CollectorPolicy::record_full_collection_end() { 826 // Consider this like a collection pause for the purposes of allocation 827 // since last pause. 828 double end_sec = os::elapsedTime(); 829 double full_gc_time_sec = end_sec - _full_collection_start_sec; 830 double full_gc_time_ms = full_gc_time_sec * 1000.0; 831 832 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 833 834 update_recent_gc_times(end_sec, full_gc_time_ms); 835 836 collector_state()->set_full_collection(false); 837 838 // "Nuke" the heuristics that control the young/mixed GC 839 // transitions and make sure we start with young GCs after the Full GC. 840 collector_state()->set_gcs_are_young(true); 841 collector_state()->set_last_young_gc(false); 842 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 843 collector_state()->set_during_initial_mark_pause(false); 844 collector_state()->set_in_marking_window(false); 845 collector_state()->set_in_marking_window_im(false); 846 847 _short_lived_surv_rate_group->start_adding_regions(); 848 // also call this on any additional surv rate groups 849 850 record_survivor_regions(0, NULL, NULL); 851 852 _free_regions_at_end_of_collection = _g1->num_free_regions(); 853 // Reset survivors SurvRateGroup. 854 _survivor_surv_rate_group->reset(); 855 update_young_list_max_and_target_length(); 856 update_rs_lengths_prediction(); 857 _collectionSetChooser->clear(); 858 859 _last_old_allocated_bytes = 0; 860 861 record_pause(FullGC, _full_collection_start_sec, end_sec); 862 } 863 864 void G1CollectorPolicy::record_stop_world_start() { 865 _stop_world_start = os::elapsedTime(); 866 } 867 868 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 869 // We only need to do this here as the policy will only be applied 870 // to the GC we're about to start. so, no point is calculating this 871 // every time we calculate / recalculate the target young length. 872 update_survivors_policy(); 873 874 assert(_g1->used() == _g1->recalculate_used(), 875 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 876 _g1->used(), _g1->recalculate_used()); 877 878 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 879 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 880 _stop_world_start = 0.0; 881 882 record_heap_size_info_at_start(false /* full */); 883 884 phase_times()->record_cur_collection_start_sec(start_time_sec); 885 _pending_cards = _g1->pending_card_num(); 886 887 _collection_set_bytes_used_before = 0; 888 _bytes_copied_during_gc = 0; 889 890 collector_state()->set_last_gc_was_young(false); 891 892 // do that for any other surv rate groups 893 _short_lived_surv_rate_group->stop_adding_regions(); 894 _survivors_age_table.clear(); 895 896 assert( verify_young_ages(), "region age verification" ); 897 } 898 899 void G1CollectorPolicy::record_concurrent_mark_init_end(double 900 mark_init_elapsed_time_ms) { 901 collector_state()->set_during_marking(true); 902 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 903 collector_state()->set_during_initial_mark_pause(false); 904 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 905 } 906 907 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 908 _mark_remark_start_sec = os::elapsedTime(); 909 collector_state()->set_during_marking(false); 910 } 911 912 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 913 double end_time_sec = os::elapsedTime(); 914 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 915 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 916 _cur_mark_stop_world_time_ms += elapsed_time_ms; 917 _prev_collection_pause_end_ms += elapsed_time_ms; 918 919 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 920 } 921 922 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 923 _mark_cleanup_start_sec = os::elapsedTime(); 924 } 925 926 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 927 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 928 "skip last young-only gc"); 929 collector_state()->set_last_young_gc(should_continue_with_reclaim); 930 // We skip the marking phase. 931 if (!should_continue_with_reclaim) { 932 abort_time_to_mixed_tracking(); 933 } 934 collector_state()->set_in_marking_window(false); 935 } 936 937 void G1CollectorPolicy::record_concurrent_pause() { 938 if (_stop_world_start > 0.0) { 939 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 940 _trace_young_gen_time_data.record_yield_time(yield_ms); 941 } 942 } 943 944 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 945 return phase_times()->average_time_ms(phase); 946 } 947 948 double G1CollectorPolicy::young_other_time_ms() const { 949 return phase_times()->young_cset_choice_time_ms() + 950 phase_times()->young_free_cset_time_ms(); 951 } 952 953 double G1CollectorPolicy::non_young_other_time_ms() const { 954 return phase_times()->non_young_cset_choice_time_ms() + 955 phase_times()->non_young_free_cset_time_ms(); 956 957 } 958 959 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 960 return pause_time_ms - 961 average_time_ms(G1GCPhaseTimes::UpdateRS) - 962 average_time_ms(G1GCPhaseTimes::ScanRS) - 963 average_time_ms(G1GCPhaseTimes::ObjCopy) - 964 average_time_ms(G1GCPhaseTimes::Termination); 965 } 966 967 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 968 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 969 } 970 971 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 972 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 973 } 974 975 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 976 if (about_to_start_mixed_phase()) { 977 return false; 978 } 979 980 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 981 982 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 983 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 984 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 985 986 if (marking_request_bytes > marking_initiating_used_threshold) { 987 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 988 ergo_verbose5(ErgoConcCycles, 989 "request concurrent cycle initiation", 990 ergo_format_reason("occupancy higher than threshold") 991 ergo_format_byte("occupancy") 992 ergo_format_byte("allocation request") 993 ergo_format_byte_perc("threshold") 994 ergo_format_str("source"), 995 cur_used_bytes, 996 alloc_byte_size, 997 marking_initiating_used_threshold, 998 (double) marking_initiating_used_threshold / _g1->capacity() * 100, 999 source); 1000 return true; 1001 } else { 1002 ergo_verbose5(ErgoConcCycles, 1003 "do not request concurrent cycle initiation", 1004 ergo_format_reason("still doing mixed collections") 1005 ergo_format_byte("occupancy") 1006 ergo_format_byte("allocation request") 1007 ergo_format_byte_perc("threshold") 1008 ergo_format_str("source"), 1009 cur_used_bytes, 1010 alloc_byte_size, 1011 marking_initiating_used_threshold, 1012 (double) InitiatingHeapOccupancyPercent, 1013 source); 1014 } 1015 } 1016 1017 return false; 1018 } 1019 1020 // Anything below that is considered to be zero 1021 #define MIN_TIMER_GRANULARITY 0.0000001 1022 1023 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 1024 double end_time_sec = os::elapsedTime(); 1025 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1026 "otherwise, the subtraction below does not make sense"); 1027 size_t cur_used_bytes = _g1->used(); 1028 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1029 bool last_pause_included_initial_mark = false; 1030 bool update_stats = !_g1->evacuation_failed(); 1031 1032 #ifndef PRODUCT 1033 if (G1YoungSurvRateVerbose) { 1034 gclog_or_tty->cr(); 1035 _short_lived_surv_rate_group->print(); 1036 // do that for any other surv rate groups too 1037 } 1038 #endif // PRODUCT 1039 1040 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec); 1041 1042 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1043 if (last_pause_included_initial_mark) { 1044 record_concurrent_mark_init_end(0.0); 1045 } else { 1046 maybe_start_marking(); 1047 } 1048 1049 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1050 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1051 // This usually happens due to the timer not having the required 1052 // granularity. Some Linuxes are the usual culprits. 1053 // We'll just set it to something (arbitrarily) small. 1054 app_time_ms = 1.0; 1055 } 1056 1057 if (update_stats) { 1058 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1059 // We maintain the invariant that all objects allocated by mutator 1060 // threads will be allocated out of eden regions. So, we can use 1061 // the eden region number allocated since the previous GC to 1062 // calculate the application's allocate rate. The only exception 1063 // to that is humongous objects that are allocated separately. But 1064 // given that humongous object allocations do not really affect 1065 // either the pause's duration nor when the next pause will take 1066 // place we can safely ignore them here. 1067 uint regions_allocated = eden_cset_region_length(); 1068 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1069 _alloc_rate_ms_seq->add(alloc_rate_ms); 1070 1071 double interval_ms = 1072 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1073 update_recent_gc_times(end_time_sec, pause_time_ms); 1074 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1075 if (recent_avg_pause_time_ratio() < 0.0 || 1076 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1077 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1078 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1079 if (_recent_avg_pause_time_ratio < 0.0) { 1080 _recent_avg_pause_time_ratio = 0.0; 1081 } else { 1082 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1083 _recent_avg_pause_time_ratio = 1.0; 1084 } 1085 } 1086 } 1087 1088 bool new_in_marking_window = collector_state()->in_marking_window(); 1089 bool new_in_marking_window_im = false; 1090 if (last_pause_included_initial_mark) { 1091 new_in_marking_window = true; 1092 new_in_marking_window_im = true; 1093 } 1094 1095 if (collector_state()->last_young_gc()) { 1096 // This is supposed to to be the "last young GC" before we start 1097 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1098 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1099 1100 if (next_gc_should_be_mixed("start mixed GCs", 1101 "do not start mixed GCs")) { 1102 collector_state()->set_gcs_are_young(false); 1103 } else { 1104 // We aborted the mixed GC phase early. 1105 abort_time_to_mixed_tracking(); 1106 } 1107 1108 collector_state()->set_last_young_gc(false); 1109 } 1110 1111 if (!collector_state()->last_gc_was_young()) { 1112 // This is a mixed GC. Here we decide whether to continue doing 1113 // mixed GCs or not. 1114 if (!next_gc_should_be_mixed("continue mixed GCs", 1115 "do not continue mixed GCs")) { 1116 collector_state()->set_gcs_are_young(true); 1117 1118 maybe_start_marking(); 1119 } 1120 } 1121 1122 _short_lived_surv_rate_group->start_adding_regions(); 1123 // Do that for any other surv rate groups 1124 1125 if (update_stats) { 1126 double cost_per_card_ms = 0.0; 1127 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1128 if (_pending_cards > 0) { 1129 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1130 _cost_per_card_ms_seq->add(cost_per_card_ms); 1131 } 1132 _cost_scan_hcc_seq->add(cost_scan_hcc); 1133 1134 double cost_per_entry_ms = 0.0; 1135 if (cards_scanned > 10) { 1136 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1137 if (collector_state()->last_gc_was_young()) { 1138 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1139 } else { 1140 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1141 } 1142 } 1143 1144 if (_max_rs_lengths > 0) { 1145 double cards_per_entry_ratio = 1146 (double) cards_scanned / (double) _max_rs_lengths; 1147 if (collector_state()->last_gc_was_young()) { 1148 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1149 } else { 1150 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1151 } 1152 } 1153 1154 // This is defensive. For a while _max_rs_lengths could get 1155 // smaller than _recorded_rs_lengths which was causing 1156 // rs_length_diff to get very large and mess up the RSet length 1157 // predictions. The reason was unsafe concurrent updates to the 1158 // _inc_cset_recorded_rs_lengths field which the code below guards 1159 // against (see CR 7118202). This bug has now been fixed (see CR 1160 // 7119027). However, I'm still worried that 1161 // _inc_cset_recorded_rs_lengths might still end up somewhat 1162 // inaccurate. The concurrent refinement thread calculates an 1163 // RSet's length concurrently with other CR threads updating it 1164 // which might cause it to calculate the length incorrectly (if, 1165 // say, it's in mid-coarsening). So I'll leave in the defensive 1166 // conditional below just in case. 1167 size_t rs_length_diff = 0; 1168 if (_max_rs_lengths > _recorded_rs_lengths) { 1169 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1170 } 1171 _rs_length_diff_seq->add((double) rs_length_diff); 1172 1173 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1174 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1175 double cost_per_byte_ms = 0.0; 1176 1177 if (copied_bytes > 0) { 1178 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1179 if (collector_state()->in_marking_window()) { 1180 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1181 } else { 1182 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1183 } 1184 } 1185 1186 if (young_cset_region_length() > 0) { 1187 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1188 young_cset_region_length()); 1189 } 1190 1191 if (old_cset_region_length() > 0) { 1192 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1193 old_cset_region_length()); 1194 } 1195 1196 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1197 1198 _pending_cards_seq->add((double) _pending_cards); 1199 _rs_lengths_seq->add((double) _max_rs_lengths); 1200 } 1201 1202 collector_state()->set_in_marking_window(new_in_marking_window); 1203 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1204 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1205 // IHOP control wants to know the expected young gen length if it were not 1206 // restrained by the heap reserve. Using the actual length would make the 1207 // prediction too small and the limit the young gen every time we get to the 1208 // predicted target occupancy. 1209 size_t last_unrestrained_young_length = update_young_list_max_and_target_length(); 1210 update_rs_lengths_prediction(); 1211 1212 // Only update IHOP information on regular GCs. 1213 if (update_stats) { 1214 double marking_to_mixed_time = -1.0; 1215 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1216 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1217 assert(marking_to_mixed_time > 0.0, 1218 "Initial mark to mixed time must be larger than zero but is %.3f", 1219 marking_to_mixed_time); 1220 } 1221 1222 update_ihop_statistics(marking_to_mixed_time, 1223 app_time_ms / 1000.0, 1224 _last_old_allocated_bytes, 1225 last_unrestrained_young_length * HeapRegion::GrainBytes); 1226 } 1227 _last_old_allocated_bytes = 0; 1228 1229 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1230 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1231 1232 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1233 1234 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1235 ergo_verbose2(ErgoTiming, 1236 "adjust concurrent refinement thresholds", 1237 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1238 ergo_format_ms("Update RS time goal") 1239 ergo_format_ms("Scan HCC time"), 1240 update_rs_time_goal_ms, 1241 scan_hcc_time_ms); 1242 1243 update_rs_time_goal_ms = 0; 1244 } else { 1245 update_rs_time_goal_ms -= scan_hcc_time_ms; 1246 } 1247 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1248 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1249 update_rs_time_goal_ms); 1250 1251 _collectionSetChooser->verify(); 1252 } 1253 1254 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1255 if (G1UseAdaptiveIHOP) { 1256 // The target occupancy is the total heap occupancy we want to hit. First, we 1257 // want to avoid eating into the reserve intended for young GC (to avoid unnecessary 1258 // throughput loss). Additionally G1 is free to not clean out up to 1259 // G1HeapWastePercent of heap, that space also cannot be used for allocation 1260 // while marking. 1261 size_t safe_heap_percentage = (size_t) (G1ReservePercent + G1HeapWastePercent); 1262 size_t target_occupancy = 0; 1263 1264 if (safe_heap_percentage < 100) { 1265 target_occupancy = G1CollectedHeap::heap()->max_capacity() * (100.0 - safe_heap_percentage) / 100.0; 1266 } 1267 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1268 target_occupancy, 1269 &_predictor); 1270 } else { 1271 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1272 G1CollectedHeap::heap()->max_capacity()); 1273 } 1274 } 1275 1276 void G1CollectorPolicy::update_ihop_statistics(double marking_time, 1277 double mutator_time_s, 1278 size_t mutator_alloc_bytes, 1279 size_t young_gen_size) { 1280 bool report = false; 1281 1282 // To avoid using really small times that may be caused by e.g. back-to-back gcs 1283 // we filter them out. 1284 double const min_valid_time = 1e-6; 1285 1286 if (marking_time > min_valid_time) { 1287 _ihop_control->update_marking_length(marking_time); 1288 report = true; 1289 } 1290 1291 // As an approximation for the young gc promotion rates during marking we use 1292 // all of them. In many applications there are only a few if any young gcs during 1293 // marking, which makes any prediction useless. This increases the accuracy of the 1294 // prediction. 1295 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1296 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1297 report = true; 1298 } 1299 1300 if (report) { 1301 report_ihop_statistics(); 1302 } 1303 } 1304 1305 void G1CollectorPolicy::report_ihop_statistics() { 1306 _ihop_control->print(); 1307 } 1308 1309 #define EXT_SIZE_FORMAT "%.1f%s" 1310 #define EXT_SIZE_PARAMS(bytes) \ 1311 byte_size_in_proper_unit((double)(bytes)), \ 1312 proper_unit_for_byte_size((bytes)) 1313 1314 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1315 YoungList* young_list = _g1->young_list(); 1316 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1317 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1318 _heap_capacity_bytes_before_gc = _g1->capacity(); 1319 _heap_used_bytes_before_gc = _g1->used(); 1320 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1321 1322 _eden_capacity_bytes_before_gc = 1323 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1324 1325 if (full) { 1326 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1327 } 1328 } 1329 1330 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1331 size_t bytes_after = _g1->used(); 1332 size_t capacity = _g1->capacity(); 1333 1334 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1335 byte_size_in_proper_unit(bytes_before), 1336 proper_unit_for_byte_size(bytes_before), 1337 byte_size_in_proper_unit(bytes_after), 1338 proper_unit_for_byte_size(bytes_after), 1339 byte_size_in_proper_unit(capacity), 1340 proper_unit_for_byte_size(capacity)); 1341 } 1342 1343 void G1CollectorPolicy::print_heap_transition() const { 1344 print_heap_transition(_heap_used_bytes_before_gc); 1345 } 1346 1347 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1348 YoungList* young_list = _g1->young_list(); 1349 1350 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1351 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1352 size_t heap_used_bytes_after_gc = _g1->used(); 1353 1354 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1355 size_t eden_capacity_bytes_after_gc = 1356 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1357 1358 gclog_or_tty->print( 1359 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1360 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1361 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1362 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1363 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1364 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1365 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1366 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1367 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1368 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1369 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1370 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1371 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1372 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1373 1374 if (full) { 1375 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1376 } 1377 1378 gclog_or_tty->cr(); 1379 } 1380 1381 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1382 phase_times()->print(pause_time_sec); 1383 } 1384 1385 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1386 double update_rs_processed_buffers, 1387 double goal_ms) { 1388 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1389 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1390 1391 if (G1UseAdaptiveConcRefinement) { 1392 const int k_gy = 3, k_gr = 6; 1393 const double inc_k = 1.1, dec_k = 0.9; 1394 1395 int g = cg1r->green_zone(); 1396 if (update_rs_time > goal_ms) { 1397 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1398 } else { 1399 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1400 g = (int)MAX2(g * inc_k, g + 1.0); 1401 } 1402 } 1403 // Change the refinement threads params 1404 cg1r->set_green_zone(g); 1405 cg1r->set_yellow_zone(g * k_gy); 1406 cg1r->set_red_zone(g * k_gr); 1407 cg1r->reinitialize_threads(); 1408 1409 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1410 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1411 cg1r->yellow_zone()); 1412 // Change the barrier params 1413 dcqs.set_process_completed_threshold(processing_threshold); 1414 dcqs.set_max_completed_queue(cg1r->red_zone()); 1415 } 1416 1417 int curr_queue_size = dcqs.completed_buffers_num(); 1418 if (curr_queue_size >= cg1r->yellow_zone()) { 1419 dcqs.set_completed_queue_padding(curr_queue_size); 1420 } else { 1421 dcqs.set_completed_queue_padding(0); 1422 } 1423 dcqs.notify_if_necessary(); 1424 } 1425 1426 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1427 return (size_t) get_new_prediction(_rs_length_diff_seq); 1428 } 1429 1430 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1431 return get_new_prediction(_alloc_rate_ms_seq); 1432 } 1433 1434 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1435 return get_new_prediction(_cost_per_card_ms_seq); 1436 } 1437 1438 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1439 return get_new_prediction(_cost_scan_hcc_seq); 1440 } 1441 1442 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1443 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1444 } 1445 1446 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1447 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1448 } 1449 1450 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1451 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1452 return predict_young_cards_per_entry_ratio(); 1453 } else { 1454 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1455 } 1456 } 1457 1458 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1459 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1460 } 1461 1462 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1463 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1464 } 1465 1466 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1467 if (collector_state()->gcs_are_young()) { 1468 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1469 } else { 1470 return predict_mixed_rs_scan_time_ms(card_num); 1471 } 1472 } 1473 1474 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1475 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1476 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1477 } else { 1478 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1479 } 1480 } 1481 1482 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1483 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1484 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1485 } else { 1486 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1487 } 1488 } 1489 1490 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1491 if (collector_state()->during_concurrent_mark()) { 1492 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1493 } else { 1494 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1495 } 1496 } 1497 1498 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1499 return get_new_prediction(_constant_other_time_ms_seq); 1500 } 1501 1502 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1503 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1504 } 1505 1506 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1507 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1508 } 1509 1510 double G1CollectorPolicy::predict_remark_time_ms() const { 1511 return get_new_prediction(_concurrent_mark_remark_times_ms); 1512 } 1513 1514 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1515 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1516 } 1517 1518 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1519 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1520 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1521 double pred = get_new_prediction(seq); 1522 if (pred > 1.0) { 1523 pred = 1.0; 1524 } 1525 return pred; 1526 } 1527 1528 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1529 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1530 } 1531 1532 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1533 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1534 } 1535 1536 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1537 size_t scanned_cards) const { 1538 return 1539 predict_rs_update_time_ms(pending_cards) + 1540 predict_rs_scan_time_ms(scanned_cards) + 1541 predict_constant_other_time_ms(); 1542 } 1543 1544 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1545 size_t rs_length = predict_rs_length_diff(); 1546 size_t card_num; 1547 if (collector_state()->gcs_are_young()) { 1548 card_num = predict_young_card_num(rs_length); 1549 } else { 1550 card_num = predict_non_young_card_num(rs_length); 1551 } 1552 return predict_base_elapsed_time_ms(pending_cards, card_num); 1553 } 1554 1555 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1556 size_t bytes_to_copy; 1557 if (hr->is_marked()) 1558 bytes_to_copy = hr->max_live_bytes(); 1559 else { 1560 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1561 int age = hr->age_in_surv_rate_group(); 1562 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1563 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1564 } 1565 return bytes_to_copy; 1566 } 1567 1568 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1569 bool for_young_gc) const { 1570 size_t rs_length = hr->rem_set()->occupied(); 1571 size_t card_num; 1572 1573 // Predicting the number of cards is based on which type of GC 1574 // we're predicting for. 1575 if (for_young_gc) { 1576 card_num = predict_young_card_num(rs_length); 1577 } else { 1578 card_num = predict_non_young_card_num(rs_length); 1579 } 1580 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1581 1582 double region_elapsed_time_ms = 1583 predict_rs_scan_time_ms(card_num) + 1584 predict_object_copy_time_ms(bytes_to_copy); 1585 1586 // The prediction of the "other" time for this region is based 1587 // upon the region type and NOT the GC type. 1588 if (hr->is_young()) { 1589 region_elapsed_time_ms += predict_young_other_time_ms(1); 1590 } else { 1591 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1592 } 1593 return region_elapsed_time_ms; 1594 } 1595 1596 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1597 uint survivor_cset_region_length) { 1598 _eden_cset_region_length = eden_cset_region_length; 1599 _survivor_cset_region_length = survivor_cset_region_length; 1600 _old_cset_region_length = 0; 1601 } 1602 1603 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1604 _recorded_rs_lengths = rs_lengths; 1605 } 1606 1607 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1608 double elapsed_ms) { 1609 _recent_gc_times_ms->add(elapsed_ms); 1610 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1611 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1612 } 1613 1614 size_t G1CollectorPolicy::expansion_amount() const { 1615 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1616 double threshold = _gc_overhead_perc; 1617 if (recent_gc_overhead > threshold) { 1618 // We will double the existing space, or take 1619 // G1ExpandByPercentOfAvailable % of the available expansion 1620 // space, whichever is smaller, bounded below by a minimum 1621 // expansion (unless that's all that's left.) 1622 const size_t min_expand_bytes = 1*M; 1623 size_t reserved_bytes = _g1->max_capacity(); 1624 size_t committed_bytes = _g1->capacity(); 1625 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1626 size_t expand_bytes; 1627 size_t expand_bytes_via_pct = 1628 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1629 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1630 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1631 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1632 1633 ergo_verbose5(ErgoHeapSizing, 1634 "attempt heap expansion", 1635 ergo_format_reason("recent GC overhead higher than " 1636 "threshold after GC") 1637 ergo_format_perc("recent GC overhead") 1638 ergo_format_perc("threshold") 1639 ergo_format_byte("uncommitted") 1640 ergo_format_byte_perc("calculated expansion amount"), 1641 recent_gc_overhead, threshold, 1642 uncommitted_bytes, 1643 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1644 1645 return expand_bytes; 1646 } else { 1647 return 0; 1648 } 1649 } 1650 1651 void G1CollectorPolicy::print_tracing_info() const { 1652 _trace_young_gen_time_data.print(); 1653 _trace_old_gen_time_data.print(); 1654 } 1655 1656 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1657 #ifndef PRODUCT 1658 _short_lived_surv_rate_group->print_surv_rate_summary(); 1659 // add this call for any other surv rate groups 1660 #endif // PRODUCT 1661 } 1662 1663 bool G1CollectorPolicy::is_young_list_full() const { 1664 uint young_list_length = _g1->young_list()->length(); 1665 uint young_list_target_length = _young_list_target_length; 1666 return young_list_length >= young_list_target_length; 1667 } 1668 1669 bool G1CollectorPolicy::can_expand_young_list() const { 1670 uint young_list_length = _g1->young_list()->length(); 1671 uint young_list_max_length = _young_list_max_length; 1672 return young_list_length < young_list_max_length; 1673 } 1674 1675 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1676 uint expansion_region_num = 0; 1677 if (GCLockerEdenExpansionPercent > 0) { 1678 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1679 double expansion_region_num_d = perc * (double) _young_list_target_length; 1680 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1681 // less than 1.0) we'll get 1. 1682 expansion_region_num = (uint) ceil(expansion_region_num_d); 1683 } else { 1684 assert(expansion_region_num == 0, "sanity"); 1685 } 1686 _young_list_max_length = _young_list_target_length + expansion_region_num; 1687 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1688 } 1689 1690 // Calculates survivor space parameters. 1691 void G1CollectorPolicy::update_survivors_policy() { 1692 double max_survivor_regions_d = 1693 (double) _young_list_target_length / (double) SurvivorRatio; 1694 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1695 // smaller than 1.0) we'll get 1. 1696 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1697 1698 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1699 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1700 } 1701 1702 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1703 // We actually check whether we are marking here and not if we are in a 1704 // reclamation phase. This means that we will schedule a concurrent mark 1705 // even while we are still in the process of reclaiming memory. 1706 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1707 if (!during_cycle) { 1708 ergo_verbose1(ErgoConcCycles, 1709 "request concurrent cycle initiation", 1710 ergo_format_reason("requested by GC cause") 1711 ergo_format_str("GC cause"), 1712 GCCause::to_string(gc_cause)); 1713 collector_state()->set_initiate_conc_mark_if_possible(true); 1714 return true; 1715 } else { 1716 ergo_verbose1(ErgoConcCycles, 1717 "do not request concurrent cycle initiation", 1718 ergo_format_reason("concurrent cycle already in progress") 1719 ergo_format_str("GC cause"), 1720 GCCause::to_string(gc_cause)); 1721 return false; 1722 } 1723 } 1724 1725 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1726 // We are about to decide on whether this pause will be an 1727 // initial-mark pause. 1728 1729 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1730 // will set it here if we have to. However, it should be cleared by 1731 // the end of the pause (it's only set for the duration of an 1732 // initial-mark pause). 1733 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1734 1735 if (collector_state()->initiate_conc_mark_if_possible()) { 1736 // We had noticed on a previous pause that the heap occupancy has 1737 // gone over the initiating threshold and we should start a 1738 // concurrent marking cycle. So we might initiate one. 1739 1740 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1741 // Initiate a new initial mark only if there is no marking or reclamation going 1742 // on. 1743 1744 collector_state()->set_during_initial_mark_pause(true); 1745 // And we can now clear initiate_conc_mark_if_possible() as 1746 // we've already acted on it. 1747 collector_state()->set_initiate_conc_mark_if_possible(false); 1748 1749 ergo_verbose0(ErgoConcCycles, 1750 "initiate concurrent cycle", 1751 ergo_format_reason("concurrent cycle initiation requested")); 1752 } else { 1753 // The concurrent marking thread is still finishing up the 1754 // previous cycle. If we start one right now the two cycles 1755 // overlap. In particular, the concurrent marking thread might 1756 // be in the process of clearing the next marking bitmap (which 1757 // we will use for the next cycle if we start one). Starting a 1758 // cycle now will be bad given that parts of the marking 1759 // information might get cleared by the marking thread. And we 1760 // cannot wait for the marking thread to finish the cycle as it 1761 // periodically yields while clearing the next marking bitmap 1762 // and, if it's in a yield point, it's waiting for us to 1763 // finish. So, at this point we will not start a cycle and we'll 1764 // let the concurrent marking thread complete the last one. 1765 ergo_verbose0(ErgoConcCycles, 1766 "do not initiate concurrent cycle", 1767 ergo_format_reason("concurrent cycle already in progress")); 1768 } 1769 } 1770 } 1771 1772 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1773 G1CollectedHeap* _g1h; 1774 CSetChooserParUpdater _cset_updater; 1775 1776 public: 1777 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1778 uint chunk_size) : 1779 _g1h(G1CollectedHeap::heap()), 1780 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1781 1782 bool doHeapRegion(HeapRegion* r) { 1783 // Do we have any marking information for this region? 1784 if (r->is_marked()) { 1785 // We will skip any region that's currently used as an old GC 1786 // alloc region (we should not consider those for collection 1787 // before we fill them up). 1788 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1789 _cset_updater.add_region(r); 1790 } 1791 } 1792 return false; 1793 } 1794 }; 1795 1796 class ParKnownGarbageTask: public AbstractGangTask { 1797 CollectionSetChooser* _hrSorted; 1798 uint _chunk_size; 1799 G1CollectedHeap* _g1; 1800 HeapRegionClaimer _hrclaimer; 1801 1802 public: 1803 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1804 AbstractGangTask("ParKnownGarbageTask"), 1805 _hrSorted(hrSorted), _chunk_size(chunk_size), 1806 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1807 1808 void work(uint worker_id) { 1809 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1810 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1811 } 1812 }; 1813 1814 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1815 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1816 const uint overpartition_factor = 4; 1817 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1818 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1819 } 1820 1821 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1822 _collectionSetChooser->clear(); 1823 1824 WorkGang* workers = _g1->workers(); 1825 uint n_workers = workers->active_workers(); 1826 1827 uint n_regions = _g1->num_regions(); 1828 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1829 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1830 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1831 workers->run_task(&par_known_garbage_task); 1832 1833 _collectionSetChooser->sort_regions(); 1834 1835 double end_sec = os::elapsedTime(); 1836 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1837 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1838 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1839 _prev_collection_pause_end_ms += elapsed_time_ms; 1840 1841 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1842 } 1843 1844 // Add the heap region at the head of the non-incremental collection set 1845 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1846 assert(_inc_cset_build_state == Active, "Precondition"); 1847 assert(hr->is_old(), "the region should be old"); 1848 1849 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1850 _g1->register_old_region_with_cset(hr); 1851 hr->set_next_in_collection_set(_collection_set); 1852 _collection_set = hr; 1853 _collection_set_bytes_used_before += hr->used(); 1854 size_t rs_length = hr->rem_set()->occupied(); 1855 _recorded_rs_lengths += rs_length; 1856 _old_cset_region_length += 1; 1857 } 1858 1859 // Initialize the per-collection-set information 1860 void G1CollectorPolicy::start_incremental_cset_building() { 1861 assert(_inc_cset_build_state == Inactive, "Precondition"); 1862 1863 _inc_cset_head = NULL; 1864 _inc_cset_tail = NULL; 1865 _inc_cset_bytes_used_before = 0; 1866 1867 _inc_cset_max_finger = 0; 1868 _inc_cset_recorded_rs_lengths = 0; 1869 _inc_cset_recorded_rs_lengths_diffs = 0; 1870 _inc_cset_predicted_elapsed_time_ms = 0.0; 1871 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1872 _inc_cset_build_state = Active; 1873 } 1874 1875 void G1CollectorPolicy::finalize_incremental_cset_building() { 1876 assert(_inc_cset_build_state == Active, "Precondition"); 1877 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1878 1879 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1880 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1881 // that adds a new region to the CSet. Further updates by the 1882 // concurrent refinement thread that samples the young RSet lengths 1883 // are accumulated in the *_diffs fields. Here we add the diffs to 1884 // the "main" fields. 1885 1886 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1887 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1888 } else { 1889 // This is defensive. The diff should in theory be always positive 1890 // as RSets can only grow between GCs. However, given that we 1891 // sample their size concurrently with other threads updating them 1892 // it's possible that we might get the wrong size back, which 1893 // could make the calculations somewhat inaccurate. 1894 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1895 if (_inc_cset_recorded_rs_lengths >= diffs) { 1896 _inc_cset_recorded_rs_lengths -= diffs; 1897 } else { 1898 _inc_cset_recorded_rs_lengths = 0; 1899 } 1900 } 1901 _inc_cset_predicted_elapsed_time_ms += 1902 _inc_cset_predicted_elapsed_time_ms_diffs; 1903 1904 _inc_cset_recorded_rs_lengths_diffs = 0; 1905 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1906 } 1907 1908 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1909 // This routine is used when: 1910 // * adding survivor regions to the incremental cset at the end of an 1911 // evacuation pause, 1912 // * adding the current allocation region to the incremental cset 1913 // when it is retired, and 1914 // * updating existing policy information for a region in the 1915 // incremental cset via young list RSet sampling. 1916 // Therefore this routine may be called at a safepoint by the 1917 // VM thread, or in-between safepoints by mutator threads (when 1918 // retiring the current allocation region) or a concurrent 1919 // refine thread (RSet sampling). 1920 1921 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1922 size_t used_bytes = hr->used(); 1923 _inc_cset_recorded_rs_lengths += rs_length; 1924 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1925 _inc_cset_bytes_used_before += used_bytes; 1926 1927 // Cache the values we have added to the aggregated information 1928 // in the heap region in case we have to remove this region from 1929 // the incremental collection set, or it is updated by the 1930 // rset sampling code 1931 hr->set_recorded_rs_length(rs_length); 1932 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1933 } 1934 1935 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1936 size_t new_rs_length) { 1937 // Update the CSet information that is dependent on the new RS length 1938 assert(hr->is_young(), "Precondition"); 1939 assert(!SafepointSynchronize::is_at_safepoint(), 1940 "should not be at a safepoint"); 1941 1942 // We could have updated _inc_cset_recorded_rs_lengths and 1943 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1944 // that atomically, as this code is executed by a concurrent 1945 // refinement thread, potentially concurrently with a mutator thread 1946 // allocating a new region and also updating the same fields. To 1947 // avoid the atomic operations we accumulate these updates on two 1948 // separate fields (*_diffs) and we'll just add them to the "main" 1949 // fields at the start of a GC. 1950 1951 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1952 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1953 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1954 1955 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1956 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1957 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1958 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1959 1960 hr->set_recorded_rs_length(new_rs_length); 1961 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1962 } 1963 1964 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1965 assert(hr->is_young(), "invariant"); 1966 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1967 assert(_inc_cset_build_state == Active, "Precondition"); 1968 1969 // We need to clear and set the cached recorded/cached collection set 1970 // information in the heap region here (before the region gets added 1971 // to the collection set). An individual heap region's cached values 1972 // are calculated, aggregated with the policy collection set info, 1973 // and cached in the heap region here (initially) and (subsequently) 1974 // by the Young List sampling code. 1975 1976 size_t rs_length = hr->rem_set()->occupied(); 1977 add_to_incremental_cset_info(hr, rs_length); 1978 1979 HeapWord* hr_end = hr->end(); 1980 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1981 1982 assert(!hr->in_collection_set(), "invariant"); 1983 _g1->register_young_region_with_cset(hr); 1984 assert(hr->next_in_collection_set() == NULL, "invariant"); 1985 } 1986 1987 // Add the region at the RHS of the incremental cset 1988 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1989 // We should only ever be appending survivors at the end of a pause 1990 assert(hr->is_survivor(), "Logic"); 1991 1992 // Do the 'common' stuff 1993 add_region_to_incremental_cset_common(hr); 1994 1995 // Now add the region at the right hand side 1996 if (_inc_cset_tail == NULL) { 1997 assert(_inc_cset_head == NULL, "invariant"); 1998 _inc_cset_head = hr; 1999 } else { 2000 _inc_cset_tail->set_next_in_collection_set(hr); 2001 } 2002 _inc_cset_tail = hr; 2003 } 2004 2005 // Add the region to the LHS of the incremental cset 2006 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2007 // Survivors should be added to the RHS at the end of a pause 2008 assert(hr->is_eden(), "Logic"); 2009 2010 // Do the 'common' stuff 2011 add_region_to_incremental_cset_common(hr); 2012 2013 // Add the region at the left hand side 2014 hr->set_next_in_collection_set(_inc_cset_head); 2015 if (_inc_cset_head == NULL) { 2016 assert(_inc_cset_tail == NULL, "Invariant"); 2017 _inc_cset_tail = hr; 2018 } 2019 _inc_cset_head = hr; 2020 } 2021 2022 #ifndef PRODUCT 2023 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2024 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2025 2026 st->print_cr("\nCollection_set:"); 2027 HeapRegion* csr = list_head; 2028 while (csr != NULL) { 2029 HeapRegion* next = csr->next_in_collection_set(); 2030 assert(csr->in_collection_set(), "bad CS"); 2031 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2032 HR_FORMAT_PARAMS(csr), 2033 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2034 csr->age_in_surv_rate_group_cond()); 2035 csr = next; 2036 } 2037 } 2038 #endif // !PRODUCT 2039 2040 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2041 // Returns the given amount of reclaimable bytes (that represents 2042 // the amount of reclaimable space still to be collected) as a 2043 // percentage of the current heap capacity. 2044 size_t capacity_bytes = _g1->capacity(); 2045 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2046 } 2047 2048 void G1CollectorPolicy::maybe_start_marking() { 2049 if (need_to_start_conc_mark("end of GC")) { 2050 // Note: this might have already been set, if during the last 2051 // pause we decided to start a cycle but at the beginning of 2052 // this pause we decided to postpone it. That's OK. 2053 collector_state()->set_initiate_conc_mark_if_possible(true); 2054 } 2055 } 2056 2057 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2058 assert(!collector_state()->full_collection(), "must be"); 2059 if (collector_state()->during_initial_mark_pause()) { 2060 assert(collector_state()->last_gc_was_young(), "must be"); 2061 assert(!collector_state()->last_young_gc(), "must be"); 2062 return InitialMarkGC; 2063 } else if (collector_state()->last_young_gc()) { 2064 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2065 assert(collector_state()->last_gc_was_young(), "must be"); 2066 return LastYoungGC; 2067 } else if (!collector_state()->last_gc_was_young()) { 2068 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2069 assert(!collector_state()->last_young_gc(), "must be"); 2070 return MixedGC; 2071 } else { 2072 assert(collector_state()->last_gc_was_young(), "must be"); 2073 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2074 assert(!collector_state()->last_young_gc(), "must be"); 2075 return YoungOnlyGC; 2076 } 2077 } 2078 2079 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2080 // Manage the MMU tracker. For some reason it ignores Full GCs. 2081 if (kind != FullGC) { 2082 _mmu_tracker->add_pause(start, end); 2083 } 2084 // Manage the mutator time tracking from initial mark to first mixed gc. 2085 switch (kind) { 2086 case FullGC: 2087 abort_time_to_mixed_tracking(); 2088 break; 2089 case Cleanup: 2090 case Remark: 2091 case YoungOnlyGC: 2092 case LastYoungGC: 2093 _initial_mark_to_mixed.add_pause(end - start); 2094 break; 2095 case InitialMarkGC: 2096 _initial_mark_to_mixed.record_initial_mark_end(end); 2097 break; 2098 case MixedGC: 2099 _initial_mark_to_mixed.record_mixed_gc_start(start); 2100 break; 2101 default: 2102 ShouldNotReachHere(); 2103 } 2104 } 2105 2106 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2107 _initial_mark_to_mixed.reset(); 2108 } 2109 2110 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2111 const char* false_action_str) const { 2112 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2113 if (cset_chooser->is_empty()) { 2114 ergo_verbose0(ErgoMixedGCs, 2115 false_action_str, 2116 ergo_format_reason("candidate old regions not available")); 2117 return false; 2118 } 2119 2120 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2121 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2122 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2123 double threshold = (double) G1HeapWastePercent; 2124 if (reclaimable_perc <= threshold) { 2125 ergo_verbose4(ErgoMixedGCs, 2126 false_action_str, 2127 ergo_format_reason("reclaimable percentage not over threshold") 2128 ergo_format_region("candidate old regions") 2129 ergo_format_byte_perc("reclaimable") 2130 ergo_format_perc("threshold"), 2131 cset_chooser->remaining_regions(), 2132 reclaimable_bytes, 2133 reclaimable_perc, threshold); 2134 return false; 2135 } 2136 2137 ergo_verbose4(ErgoMixedGCs, 2138 true_action_str, 2139 ergo_format_reason("candidate old regions available") 2140 ergo_format_region("candidate old regions") 2141 ergo_format_byte_perc("reclaimable") 2142 ergo_format_perc("threshold"), 2143 cset_chooser->remaining_regions(), 2144 reclaimable_bytes, 2145 reclaimable_perc, threshold); 2146 return true; 2147 } 2148 2149 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2150 // The min old CSet region bound is based on the maximum desired 2151 // number of mixed GCs after a cycle. I.e., even if some old regions 2152 // look expensive, we should add them to the CSet anyway to make 2153 // sure we go through the available old regions in no more than the 2154 // maximum desired number of mixed GCs. 2155 // 2156 // The calculation is based on the number of marked regions we added 2157 // to the CSet chooser in the first place, not how many remain, so 2158 // that the result is the same during all mixed GCs that follow a cycle. 2159 2160 const size_t region_num = (size_t) _collectionSetChooser->length(); 2161 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2162 size_t result = region_num / gc_num; 2163 // emulate ceiling 2164 if (result * gc_num < region_num) { 2165 result += 1; 2166 } 2167 return (uint) result; 2168 } 2169 2170 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2171 // The max old CSet region bound is based on the threshold expressed 2172 // as a percentage of the heap size. I.e., it should bound the 2173 // number of old regions added to the CSet irrespective of how many 2174 // of them are available. 2175 2176 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2177 const size_t region_num = g1h->num_regions(); 2178 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2179 size_t result = region_num * perc / 100; 2180 // emulate ceiling 2181 if (100 * result < region_num * perc) { 2182 result += 1; 2183 } 2184 return (uint) result; 2185 } 2186 2187 2188 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2189 double young_start_time_sec = os::elapsedTime(); 2190 2191 YoungList* young_list = _g1->young_list(); 2192 finalize_incremental_cset_building(); 2193 2194 guarantee(target_pause_time_ms > 0.0, 2195 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2196 guarantee(_collection_set == NULL, "Precondition"); 2197 2198 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2199 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2200 2201 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2202 "start choosing CSet", 2203 ergo_format_size("_pending_cards") 2204 ergo_format_ms("predicted base time") 2205 ergo_format_ms("remaining time") 2206 ergo_format_ms("target pause time"), 2207 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2208 2209 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2210 2211 if (collector_state()->last_gc_was_young()) { 2212 _trace_young_gen_time_data.increment_young_collection_count(); 2213 } else { 2214 _trace_young_gen_time_data.increment_mixed_collection_count(); 2215 } 2216 2217 // The young list is laid with the survivor regions from the previous 2218 // pause are appended to the RHS of the young list, i.e. 2219 // [Newly Young Regions ++ Survivors from last pause]. 2220 2221 uint survivor_region_length = young_list->survivor_length(); 2222 uint eden_region_length = young_list->eden_length(); 2223 init_cset_region_lengths(eden_region_length, survivor_region_length); 2224 2225 HeapRegion* hr = young_list->first_survivor_region(); 2226 while (hr != NULL) { 2227 assert(hr->is_survivor(), "badly formed young list"); 2228 // There is a convention that all the young regions in the CSet 2229 // are tagged as "eden", so we do this for the survivors here. We 2230 // use the special set_eden_pre_gc() as it doesn't check that the 2231 // region is free (which is not the case here). 2232 hr->set_eden_pre_gc(); 2233 hr = hr->get_next_young_region(); 2234 } 2235 2236 // Clear the fields that point to the survivor list - they are all young now. 2237 young_list->clear_survivors(); 2238 2239 _collection_set = _inc_cset_head; 2240 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2241 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2242 2243 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2244 "add young regions to CSet", 2245 ergo_format_region("eden") 2246 ergo_format_region("survivors") 2247 ergo_format_ms("predicted young region time") 2248 ergo_format_ms("target pause time"), 2249 eden_region_length, survivor_region_length, 2250 _inc_cset_predicted_elapsed_time_ms, 2251 target_pause_time_ms); 2252 2253 // The number of recorded young regions is the incremental 2254 // collection set's current size 2255 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2256 2257 double young_end_time_sec = os::elapsedTime(); 2258 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2259 2260 return time_remaining_ms; 2261 } 2262 2263 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2264 double non_young_start_time_sec = os::elapsedTime(); 2265 double predicted_old_time_ms = 0.0; 2266 2267 2268 if (!collector_state()->gcs_are_young()) { 2269 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2270 cset_chooser->verify(); 2271 const uint min_old_cset_length = calc_min_old_cset_length(); 2272 const uint max_old_cset_length = calc_max_old_cset_length(); 2273 2274 uint expensive_region_num = 0; 2275 bool check_time_remaining = adaptive_young_list_length(); 2276 2277 HeapRegion* hr = cset_chooser->peek(); 2278 while (hr != NULL) { 2279 if (old_cset_region_length() >= max_old_cset_length) { 2280 // Added maximum number of old regions to the CSet. 2281 ergo_verbose2(ErgoCSetConstruction, 2282 "finish adding old regions to CSet", 2283 ergo_format_reason("old CSet region num reached max") 2284 ergo_format_region("old") 2285 ergo_format_region("max"), 2286 old_cset_region_length(), max_old_cset_length); 2287 break; 2288 } 2289 2290 2291 // Stop adding regions if the remaining reclaimable space is 2292 // not above G1HeapWastePercent. 2293 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2294 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2295 double threshold = (double) G1HeapWastePercent; 2296 if (reclaimable_perc <= threshold) { 2297 // We've added enough old regions that the amount of uncollected 2298 // reclaimable space is at or below the waste threshold. Stop 2299 // adding old regions to the CSet. 2300 ergo_verbose5(ErgoCSetConstruction, 2301 "finish adding old regions to CSet", 2302 ergo_format_reason("reclaimable percentage not over threshold") 2303 ergo_format_region("old") 2304 ergo_format_region("max") 2305 ergo_format_byte_perc("reclaimable") 2306 ergo_format_perc("threshold"), 2307 old_cset_region_length(), 2308 max_old_cset_length, 2309 reclaimable_bytes, 2310 reclaimable_perc, threshold); 2311 break; 2312 } 2313 2314 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2315 if (check_time_remaining) { 2316 if (predicted_time_ms > time_remaining_ms) { 2317 // Too expensive for the current CSet. 2318 2319 if (old_cset_region_length() >= min_old_cset_length) { 2320 // We have added the minimum number of old regions to the CSet, 2321 // we are done with this CSet. 2322 ergo_verbose4(ErgoCSetConstruction, 2323 "finish adding old regions to CSet", 2324 ergo_format_reason("predicted time is too high") 2325 ergo_format_ms("predicted time") 2326 ergo_format_ms("remaining time") 2327 ergo_format_region("old") 2328 ergo_format_region("min"), 2329 predicted_time_ms, time_remaining_ms, 2330 old_cset_region_length(), min_old_cset_length); 2331 break; 2332 } 2333 2334 // We'll add it anyway given that we haven't reached the 2335 // minimum number of old regions. 2336 expensive_region_num += 1; 2337 } 2338 } else { 2339 if (old_cset_region_length() >= min_old_cset_length) { 2340 // In the non-auto-tuning case, we'll finish adding regions 2341 // to the CSet if we reach the minimum. 2342 ergo_verbose2(ErgoCSetConstruction, 2343 "finish adding old regions to CSet", 2344 ergo_format_reason("old CSet region num reached min") 2345 ergo_format_region("old") 2346 ergo_format_region("min"), 2347 old_cset_region_length(), min_old_cset_length); 2348 break; 2349 } 2350 } 2351 2352 // We will add this region to the CSet. 2353 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2354 predicted_old_time_ms += predicted_time_ms; 2355 cset_chooser->pop(); // already have region via peek() 2356 _g1->old_set_remove(hr); 2357 add_old_region_to_cset(hr); 2358 2359 hr = cset_chooser->peek(); 2360 } 2361 if (hr == NULL) { 2362 ergo_verbose0(ErgoCSetConstruction, 2363 "finish adding old regions to CSet", 2364 ergo_format_reason("candidate old regions not available")); 2365 } 2366 2367 if (expensive_region_num > 0) { 2368 // We print the information once here at the end, predicated on 2369 // whether we added any apparently expensive regions or not, to 2370 // avoid generating output per region. 2371 ergo_verbose4(ErgoCSetConstruction, 2372 "added expensive regions to CSet", 2373 ergo_format_reason("old CSet region num not reached min") 2374 ergo_format_region("old") 2375 ergo_format_region("expensive") 2376 ergo_format_region("min") 2377 ergo_format_ms("remaining time"), 2378 old_cset_region_length(), 2379 expensive_region_num, 2380 min_old_cset_length, 2381 time_remaining_ms); 2382 } 2383 2384 cset_chooser->verify(); 2385 } 2386 2387 stop_incremental_cset_building(); 2388 2389 ergo_verbose3(ErgoCSetConstruction, 2390 "finish choosing CSet", 2391 ergo_format_region("old") 2392 ergo_format_ms("predicted old region time") 2393 ergo_format_ms("time remaining"), 2394 old_cset_region_length(), 2395 predicted_old_time_ms, time_remaining_ms); 2396 2397 double non_young_end_time_sec = os::elapsedTime(); 2398 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2399 } 2400 2401 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2402 if(TraceYoungGenTime) { 2403 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2404 } 2405 } 2406 2407 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2408 if(TraceYoungGenTime) { 2409 _all_yield_times_ms.add(yield_time_ms); 2410 } 2411 } 2412 2413 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2414 if(TraceYoungGenTime) { 2415 _total.add(pause_time_ms); 2416 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2417 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2418 _parallel.add(phase_times->cur_collection_par_time_ms()); 2419 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2420 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2421 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2422 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2423 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2424 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2425 2426 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2427 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2428 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2429 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2430 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2431 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2432 2433 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2434 _parallel_other.add(parallel_other_time); 2435 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2436 } 2437 } 2438 2439 void TraceYoungGenTimeData::increment_young_collection_count() { 2440 if(TraceYoungGenTime) { 2441 ++_young_pause_num; 2442 } 2443 } 2444 2445 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2446 if(TraceYoungGenTime) { 2447 ++_mixed_pause_num; 2448 } 2449 } 2450 2451 void TraceYoungGenTimeData::print_summary(const char* str, 2452 const NumberSeq* seq) const { 2453 double sum = seq->sum(); 2454 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2455 str, sum / 1000.0, seq->avg()); 2456 } 2457 2458 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2459 const NumberSeq* seq) const { 2460 print_summary(str, seq); 2461 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2462 "(num", seq->num(), seq->sd(), seq->maximum()); 2463 } 2464 2465 void TraceYoungGenTimeData::print() const { 2466 if (!TraceYoungGenTime) { 2467 return; 2468 } 2469 2470 gclog_or_tty->print_cr("ALL PAUSES"); 2471 print_summary_sd(" Total", &_total); 2472 gclog_or_tty->cr(); 2473 gclog_or_tty->cr(); 2474 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2475 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2476 gclog_or_tty->cr(); 2477 2478 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2479 2480 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2481 gclog_or_tty->print_cr("none"); 2482 } else { 2483 print_summary_sd(" Evacuation Pauses", &_total); 2484 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2485 print_summary(" Parallel Time", &_parallel); 2486 print_summary(" Ext Root Scanning", &_ext_root_scan); 2487 print_summary(" SATB Filtering", &_satb_filtering); 2488 print_summary(" Update RS", &_update_rs); 2489 print_summary(" Scan RS", &_scan_rs); 2490 print_summary(" Object Copy", &_obj_copy); 2491 print_summary(" Termination", &_termination); 2492 print_summary(" Parallel Other", &_parallel_other); 2493 print_summary(" Clear CT", &_clear_ct); 2494 print_summary(" Other", &_other); 2495 } 2496 gclog_or_tty->cr(); 2497 2498 gclog_or_tty->print_cr("MISC"); 2499 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2500 print_summary_sd(" Yields", &_all_yield_times_ms); 2501 } 2502 2503 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2504 if (TraceOldGenTime) { 2505 _all_full_gc_times.add(full_gc_time_ms); 2506 } 2507 } 2508 2509 void TraceOldGenTimeData::print() const { 2510 if (!TraceOldGenTime) { 2511 return; 2512 } 2513 2514 if (_all_full_gc_times.num() > 0) { 2515 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2516 _all_full_gc_times.num(), 2517 _all_full_gc_times.sum() / 1000.0); 2518 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2519 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2520 _all_full_gc_times.sd(), 2521 _all_full_gc_times.maximum()); 2522 } 2523 }