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