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