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/g1GCPhaseTimes.hpp" 33 #include "gc/g1/heapRegion.inline.hpp" 34 #include "gc/g1/heapRegionRemSet.hpp" 35 #include "gc/shared/gcPolicyCounters.hpp" 36 #include "runtime/arguments.hpp" 37 #include "runtime/java.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "utilities/debug.hpp" 40 #include "utilities/pair.hpp" 41 42 // Different defaults for different number of GC threads 43 // They were chosen by running GCOld and SPECjbb on debris with different 44 // numbers of GC threads and choosing them based on the results 45 46 // all the same 47 static double rs_length_diff_defaults[] = { 48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 49 }; 50 51 static double cost_per_card_ms_defaults[] = { 52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 53 }; 54 55 // all the same 56 static double young_cards_per_entry_ratio_defaults[] = { 57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 58 }; 59 60 static double cost_per_entry_ms_defaults[] = { 61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 62 }; 63 64 static double cost_per_byte_ms_defaults[] = { 65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 66 }; 67 68 // these should be pretty consistent 69 static double constant_other_time_ms_defaults[] = { 70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 71 }; 72 73 74 static double young_other_cost_per_region_ms_defaults[] = { 75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 76 }; 77 78 static double non_young_other_cost_per_region_ms_defaults[] = { 79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 80 }; 81 82 G1CollectorPolicy::G1CollectorPolicy() : 83 _predictor(G1ConfidencePercent / 100.0), 84 _parallel_gc_threads(ParallelGCThreads), 85 86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 87 _stop_world_start(0.0), 88 89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 93 _prev_collection_pause_end_ms(0.0), 94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _non_young_other_cost_per_region_ms_seq( 106 new TruncatedSeq(TruncatedSeqLength)), 107 108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 110 111 _pause_time_target_ms((double) MaxGCPauseMillis), 112 113 _recent_prev_end_times_for_all_gcs_sec( 114 new TruncatedSeq(NumPrevPausesForHeuristics)), 115 116 _recent_avg_pause_time_ratio(0.0), 117 _rs_lengths_prediction(0), 118 _max_survivor_regions(0), 119 120 _eden_used_bytes_before_gc(0), 121 _survivor_used_bytes_before_gc(0), 122 _old_used_bytes_before_gc(0), 123 _humongous_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 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 181 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 182 clear_ratio_check_data(); 183 184 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 185 186 int index = MIN2(_parallel_gc_threads - 1, 7); 187 188 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 189 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 190 _cost_scan_hcc_seq->add(0.0); 191 _young_cards_per_entry_ratio_seq->add( 192 young_cards_per_entry_ratio_defaults[index]); 193 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 194 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 195 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 196 _young_other_cost_per_region_ms_seq->add( 197 young_other_cost_per_region_ms_defaults[index]); 198 _non_young_other_cost_per_region_ms_seq->add( 199 non_young_other_cost_per_region_ms_defaults[index]); 200 201 // Below, we might need to calculate the pause time target based on 202 // the pause interval. When we do so we are going to give G1 maximum 203 // flexibility and allow it to do pauses when it needs to. So, we'll 204 // arrange that the pause interval to be pause time target + 1 to 205 // ensure that a) the pause time target is maximized with respect to 206 // the pause interval and b) we maintain the invariant that pause 207 // time target < pause interval. If the user does not want this 208 // maximum flexibility, they will have to set the pause interval 209 // explicitly. 210 211 // First make sure that, if either parameter is set, its value is 212 // reasonable. 213 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 214 if (MaxGCPauseMillis < 1) { 215 vm_exit_during_initialization("MaxGCPauseMillis should be " 216 "greater than 0"); 217 } 218 } 219 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 220 if (GCPauseIntervalMillis < 1) { 221 vm_exit_during_initialization("GCPauseIntervalMillis should be " 222 "greater than 0"); 223 } 224 } 225 226 // Then, if the pause time target parameter was not set, set it to 227 // the default value. 228 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 229 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 230 // The default pause time target in G1 is 200ms 231 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 232 } else { 233 // We do not allow the pause interval to be set without the 234 // pause time target 235 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 236 "without setting MaxGCPauseMillis"); 237 } 238 } 239 240 // Then, if the interval parameter was not set, set it according to 241 // the pause time target (this will also deal with the case when the 242 // pause time target is the default value). 243 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 244 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 245 } 246 247 // Finally, make sure that the two parameters are consistent. 248 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 249 char buffer[256]; 250 jio_snprintf(buffer, 256, 251 "MaxGCPauseMillis (%u) should be less than " 252 "GCPauseIntervalMillis (%u)", 253 MaxGCPauseMillis, GCPauseIntervalMillis); 254 vm_exit_during_initialization(buffer); 255 } 256 257 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 258 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 259 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 260 261 // start conservatively (around 50ms is about right) 262 _concurrent_mark_remark_times_ms->add(0.05); 263 _concurrent_mark_cleanup_times_ms->add(0.20); 264 _tenuring_threshold = MaxTenuringThreshold; 265 266 assert(GCTimeRatio > 0, 267 "we should have set it to a default value set_g1_gc_flags() " 268 "if a user set it to 0"); 269 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 270 271 uintx reserve_perc = G1ReservePercent; 272 // Put an artificial ceiling on this so that it's not set to a silly value. 273 if (reserve_perc > 50) { 274 reserve_perc = 50; 275 warning("G1ReservePercent is set to a value that is too large, " 276 "it's been updated to " UINTX_FORMAT, reserve_perc); 277 } 278 _reserve_factor = (double) reserve_perc / 100.0; 279 // This will be set when the heap is expanded 280 // for the first time during initialization. 281 _reserve_regions = 0; 282 283 _cset_chooser = new CollectionSetChooser(); 284 } 285 286 G1CollectorPolicy::~G1CollectorPolicy() { 287 delete _ihop_control; 288 } 289 290 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 291 return _predictor.get_new_prediction(seq); 292 } 293 294 void G1CollectorPolicy::initialize_alignments() { 295 _space_alignment = HeapRegion::GrainBytes; 296 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 297 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 298 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 299 } 300 301 void G1CollectorPolicy::initialize_flags() { 302 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 303 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 304 } 305 306 if (SurvivorRatio < 1) { 307 vm_exit_during_initialization("Invalid survivor ratio specified"); 308 } 309 CollectorPolicy::initialize_flags(); 310 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 311 } 312 313 void G1CollectorPolicy::post_heap_initialize() { 314 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 315 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 316 if (max_young_size != MaxNewSize) { 317 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 318 } 319 320 _ihop_control = create_ihop_control(); 321 } 322 323 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 324 325 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 326 _min_desired_young_length(0), _max_desired_young_length(0) { 327 if (FLAG_IS_CMDLINE(NewRatio)) { 328 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 329 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 330 } else { 331 _sizer_kind = SizerNewRatio; 332 _adaptive_size = false; 333 return; 334 } 335 } 336 337 if (NewSize > MaxNewSize) { 338 if (FLAG_IS_CMDLINE(MaxNewSize)) { 339 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 340 "A new max generation size of " SIZE_FORMAT "k will be used.", 341 NewSize/K, MaxNewSize/K, NewSize/K); 342 } 343 MaxNewSize = NewSize; 344 } 345 346 if (FLAG_IS_CMDLINE(NewSize)) { 347 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 348 1U); 349 if (FLAG_IS_CMDLINE(MaxNewSize)) { 350 _max_desired_young_length = 351 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 352 1U); 353 _sizer_kind = SizerMaxAndNewSize; 354 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 355 } else { 356 _sizer_kind = SizerNewSizeOnly; 357 } 358 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 359 _max_desired_young_length = 360 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 361 1U); 362 _sizer_kind = SizerMaxNewSizeOnly; 363 } 364 } 365 366 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 367 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 368 return MAX2(1U, default_value); 369 } 370 371 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 372 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 373 return MAX2(1U, default_value); 374 } 375 376 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 377 assert(number_of_heap_regions > 0, "Heap must be initialized"); 378 379 switch (_sizer_kind) { 380 case SizerDefaults: 381 *min_young_length = calculate_default_min_length(number_of_heap_regions); 382 *max_young_length = calculate_default_max_length(number_of_heap_regions); 383 break; 384 case SizerNewSizeOnly: 385 *max_young_length = calculate_default_max_length(number_of_heap_regions); 386 *max_young_length = MAX2(*min_young_length, *max_young_length); 387 break; 388 case SizerMaxNewSizeOnly: 389 *min_young_length = calculate_default_min_length(number_of_heap_regions); 390 *min_young_length = MIN2(*min_young_length, *max_young_length); 391 break; 392 case SizerMaxAndNewSize: 393 // Do nothing. Values set on the command line, don't update them at runtime. 394 break; 395 case SizerNewRatio: 396 *min_young_length = number_of_heap_regions / (NewRatio + 1); 397 *max_young_length = *min_young_length; 398 break; 399 default: 400 ShouldNotReachHere(); 401 } 402 403 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 404 } 405 406 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 407 // We need to pass the desired values because recalculation may not update these 408 // values in some cases. 409 uint temp = _min_desired_young_length; 410 uint result = _max_desired_young_length; 411 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 412 return result; 413 } 414 415 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 416 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 417 &_max_desired_young_length); 418 } 419 420 void G1CollectorPolicy::init() { 421 // Set aside an initial future to_space. 422 _g1 = G1CollectedHeap::heap(); 423 424 assert(Heap_lock->owned_by_self(), "Locking discipline."); 425 426 initialize_gc_policy_counters(); 427 428 if (adaptive_young_list_length()) { 429 _young_list_fixed_length = 0; 430 } else { 431 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 432 } 433 _free_regions_at_end_of_collection = _g1->num_free_regions(); 434 435 update_young_list_max_and_target_length(); 436 // We may immediately start allocating regions and placing them on the 437 // collection set list. Initialize the per-collection set info 438 start_incremental_cset_building(); 439 } 440 441 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 442 phase_times()->note_gc_start(num_active_workers); 443 } 444 445 // Create the jstat counters for the policy. 446 void G1CollectorPolicy::initialize_gc_policy_counters() { 447 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 448 } 449 450 bool G1CollectorPolicy::predict_will_fit(uint young_length, 451 double base_time_ms, 452 uint base_free_regions, 453 double target_pause_time_ms) const { 454 if (young_length >= base_free_regions) { 455 // end condition 1: not enough space for the young regions 456 return false; 457 } 458 459 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 460 size_t bytes_to_copy = 461 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 462 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 463 double young_other_time_ms = predict_young_other_time_ms(young_length); 464 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 465 if (pause_time_ms > target_pause_time_ms) { 466 // end condition 2: prediction is over the target pause time 467 return false; 468 } 469 470 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 471 472 // When copying, we will likely need more bytes free than is live in the region. 473 // Add some safety margin to factor in the confidence of our guess, and the 474 // natural expected waste. 475 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 476 // of the calculation: the lower the confidence, the more headroom. 477 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 478 // copying due to anticipated waste in the PLABs. 479 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 480 size_t expected_bytes_to_copy = safety_factor * bytes_to_copy; 481 482 if (expected_bytes_to_copy > free_bytes) { 483 // end condition 3: out-of-space 484 return false; 485 } 486 487 // success! 488 return true; 489 } 490 491 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 492 // re-calculate the necessary reserve 493 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 494 // We use ceiling so that if reserve_regions_d is > 0.0 (but 495 // smaller than 1.0) we'll get 1. 496 _reserve_regions = (uint) ceil(reserve_regions_d); 497 498 _young_gen_sizer->heap_size_changed(new_number_of_regions); 499 } 500 501 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 502 uint base_min_length) const { 503 uint desired_min_length = 0; 504 if (adaptive_young_list_length()) { 505 if (_alloc_rate_ms_seq->num() > 3) { 506 double now_sec = os::elapsedTime(); 507 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 508 double alloc_rate_ms = predict_alloc_rate_ms(); 509 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 510 } else { 511 // otherwise we don't have enough info to make the prediction 512 } 513 } 514 desired_min_length += base_min_length; 515 // make sure we don't go below any user-defined minimum bound 516 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 517 } 518 519 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 520 // Here, we might want to also take into account any additional 521 // constraints (i.e., user-defined minimum bound). Currently, we 522 // effectively don't set this bound. 523 return _young_gen_sizer->max_desired_young_length(); 524 } 525 526 uint G1CollectorPolicy::update_young_list_max_and_target_length() { 527 return update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 528 } 529 530 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 531 uint unbounded_target_length = update_young_list_target_length(rs_lengths); 532 update_max_gc_locker_expansion(); 533 return unbounded_target_length; 534 } 535 536 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 537 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths); 538 _young_list_target_length = young_lengths.first; 539 return young_lengths.second; 540 } 541 542 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const { 543 YoungTargetLengths result; 544 545 // Calculate the absolute and desired min bounds first. 546 547 // This is how many young regions we already have (currently: the survivors). 548 uint base_min_length = recorded_survivor_regions(); 549 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 550 // This is the absolute minimum young length. Ensure that we 551 // will at least have one eden region available for allocation. 552 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 553 // If we shrank the young list target it should not shrink below the current size. 554 desired_min_length = MAX2(desired_min_length, absolute_min_length); 555 // Calculate the absolute and desired max bounds. 556 557 uint desired_max_length = calculate_young_list_desired_max_length(); 558 559 uint young_list_target_length = 0; 560 if (adaptive_young_list_length()) { 561 if (collector_state()->gcs_are_young()) { 562 young_list_target_length = 563 calculate_young_list_target_length(rs_lengths, 564 base_min_length, 565 desired_min_length, 566 desired_max_length); 567 } else { 568 // Don't calculate anything and let the code below bound it to 569 // the desired_min_length, i.e., do the next GC as soon as 570 // possible to maximize how many old regions we can add to it. 571 } 572 } else { 573 // The user asked for a fixed young gen so we'll fix the young gen 574 // whether the next GC is young or mixed. 575 young_list_target_length = _young_list_fixed_length; 576 } 577 578 result.second = young_list_target_length; 579 580 // We will try our best not to "eat" into the reserve. 581 uint absolute_max_length = 0; 582 if (_free_regions_at_end_of_collection > _reserve_regions) { 583 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 584 } 585 if (desired_max_length > absolute_max_length) { 586 desired_max_length = absolute_max_length; 587 } 588 589 // Make sure we don't go over the desired max length, nor under the 590 // desired min length. In case they clash, desired_min_length wins 591 // which is why that test is second. 592 if (young_list_target_length > desired_max_length) { 593 young_list_target_length = desired_max_length; 594 } 595 if (young_list_target_length < desired_min_length) { 596 young_list_target_length = desired_min_length; 597 } 598 599 assert(young_list_target_length > recorded_survivor_regions(), 600 "we should be able to allocate at least one eden region"); 601 assert(young_list_target_length >= absolute_min_length, "post-condition"); 602 603 result.first = young_list_target_length; 604 return result; 605 } 606 607 uint 608 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 609 uint base_min_length, 610 uint desired_min_length, 611 uint desired_max_length) const { 612 assert(adaptive_young_list_length(), "pre-condition"); 613 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 614 615 // In case some edge-condition makes the desired max length too small... 616 if (desired_max_length <= desired_min_length) { 617 return desired_min_length; 618 } 619 620 // We'll adjust min_young_length and max_young_length not to include 621 // the already allocated young regions (i.e., so they reflect the 622 // min and max eden regions we'll allocate). The base_min_length 623 // will be reflected in the predictions by the 624 // survivor_regions_evac_time prediction. 625 assert(desired_min_length > base_min_length, "invariant"); 626 uint min_young_length = desired_min_length - base_min_length; 627 assert(desired_max_length > base_min_length, "invariant"); 628 uint max_young_length = desired_max_length - base_min_length; 629 630 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 631 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 632 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 633 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 634 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 635 double base_time_ms = 636 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 637 survivor_regions_evac_time; 638 uint available_free_regions = _free_regions_at_end_of_collection; 639 uint base_free_regions = 0; 640 if (available_free_regions > _reserve_regions) { 641 base_free_regions = available_free_regions - _reserve_regions; 642 } 643 644 // Here, we will make sure that the shortest young length that 645 // makes sense fits within the target pause time. 646 647 if (predict_will_fit(min_young_length, base_time_ms, 648 base_free_regions, target_pause_time_ms)) { 649 // The shortest young length will fit into the target pause time; 650 // we'll now check whether the absolute maximum number of young 651 // regions will fit in the target pause time. If not, we'll do 652 // a binary search between min_young_length and max_young_length. 653 if (predict_will_fit(max_young_length, base_time_ms, 654 base_free_regions, target_pause_time_ms)) { 655 // The maximum young length will fit into the target pause time. 656 // We are done so set min young length to the maximum length (as 657 // the result is assumed to be returned in min_young_length). 658 min_young_length = max_young_length; 659 } else { 660 // The maximum possible number of young regions will not fit within 661 // the target pause time so we'll search for the optimal 662 // length. The loop invariants are: 663 // 664 // min_young_length < max_young_length 665 // min_young_length is known to fit into the target pause time 666 // max_young_length is known not to fit into the target pause time 667 // 668 // Going into the loop we know the above hold as we've just 669 // checked them. Every time around the loop we check whether 670 // the middle value between min_young_length and 671 // max_young_length fits into the target pause time. If it 672 // does, it becomes the new min. If it doesn't, it becomes 673 // the new max. This way we maintain the loop invariants. 674 675 assert(min_young_length < max_young_length, "invariant"); 676 uint diff = (max_young_length - min_young_length) / 2; 677 while (diff > 0) { 678 uint young_length = min_young_length + diff; 679 if (predict_will_fit(young_length, base_time_ms, 680 base_free_regions, target_pause_time_ms)) { 681 min_young_length = young_length; 682 } else { 683 max_young_length = young_length; 684 } 685 assert(min_young_length < max_young_length, "invariant"); 686 diff = (max_young_length - min_young_length) / 2; 687 } 688 // The results is min_young_length which, according to the 689 // loop invariants, should fit within the target pause time. 690 691 // These are the post-conditions of the binary search above: 692 assert(min_young_length < max_young_length, 693 "otherwise we should have discovered that max_young_length " 694 "fits into the pause target and not done the binary search"); 695 assert(predict_will_fit(min_young_length, base_time_ms, 696 base_free_regions, target_pause_time_ms), 697 "min_young_length, the result of the binary search, should " 698 "fit into the pause target"); 699 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 700 base_free_regions, target_pause_time_ms), 701 "min_young_length, the result of the binary search, should be " 702 "optimal, so no larger length should fit into the pause target"); 703 } 704 } else { 705 // Even the minimum length doesn't fit into the pause time 706 // target, return it as the result nevertheless. 707 } 708 return base_min_length + min_young_length; 709 } 710 711 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 712 double survivor_regions_evac_time = 0.0; 713 for (HeapRegion * r = _recorded_survivor_head; 714 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 715 r = r->get_next_young_region()) { 716 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 717 } 718 return survivor_regions_evac_time; 719 } 720 721 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 722 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 723 724 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 725 if (rs_lengths > _rs_lengths_prediction) { 726 // add 10% to avoid having to recalculate often 727 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 728 update_rs_lengths_prediction(rs_lengths_prediction); 729 730 update_young_list_max_and_target_length(rs_lengths_prediction); 731 } 732 } 733 734 void G1CollectorPolicy::update_rs_lengths_prediction() { 735 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 736 } 737 738 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 739 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 740 _rs_lengths_prediction = prediction; 741 } 742 } 743 744 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 745 bool is_tlab, 746 bool* gc_overhead_limit_was_exceeded) { 747 guarantee(false, "Not using this policy feature yet."); 748 return NULL; 749 } 750 751 // This method controls how a collector handles one or more 752 // of its generations being fully allocated. 753 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 754 bool is_tlab) { 755 guarantee(false, "Not using this policy feature yet."); 756 return NULL; 757 } 758 759 760 #ifndef PRODUCT 761 bool G1CollectorPolicy::verify_young_ages() { 762 HeapRegion* head = _g1->young_list()->first_region(); 763 return 764 verify_young_ages(head, _short_lived_surv_rate_group); 765 // also call verify_young_ages on any additional surv rate groups 766 } 767 768 bool 769 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 770 SurvRateGroup *surv_rate_group) { 771 guarantee( surv_rate_group != NULL, "pre-condition" ); 772 773 const char* name = surv_rate_group->name(); 774 bool ret = true; 775 int prev_age = -1; 776 777 for (HeapRegion* curr = head; 778 curr != NULL; 779 curr = curr->get_next_young_region()) { 780 SurvRateGroup* group = curr->surv_rate_group(); 781 if (group == NULL && !curr->is_survivor()) { 782 log_info(gc, verify)("## %s: encountered NULL surv_rate_group", name); 783 ret = false; 784 } 785 786 if (surv_rate_group == group) { 787 int age = curr->age_in_surv_rate_group(); 788 789 if (age < 0) { 790 log_info(gc, verify)("## %s: encountered negative age", name); 791 ret = false; 792 } 793 794 if (age <= prev_age) { 795 log_info(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age); 796 ret = false; 797 } 798 prev_age = age; 799 } 800 } 801 802 return ret; 803 } 804 #endif // PRODUCT 805 806 void G1CollectorPolicy::record_full_collection_start() { 807 _full_collection_start_sec = os::elapsedTime(); 808 record_heap_size_info_at_start(true /* full */); 809 // Release the future to-space so that it is available for compaction into. 810 collector_state()->set_full_collection(true); 811 } 812 813 void G1CollectorPolicy::record_full_collection_end() { 814 // Consider this like a collection pause for the purposes of allocation 815 // since last pause. 816 double end_sec = os::elapsedTime(); 817 double full_gc_time_sec = end_sec - _full_collection_start_sec; 818 double full_gc_time_ms = full_gc_time_sec * 1000.0; 819 820 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 821 822 update_recent_gc_times(end_sec, full_gc_time_ms); 823 824 collector_state()->set_full_collection(false); 825 826 // "Nuke" the heuristics that control the young/mixed GC 827 // transitions and make sure we start with young GCs after the Full GC. 828 collector_state()->set_gcs_are_young(true); 829 collector_state()->set_last_young_gc(false); 830 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 831 collector_state()->set_during_initial_mark_pause(false); 832 collector_state()->set_in_marking_window(false); 833 collector_state()->set_in_marking_window_im(false); 834 835 _short_lived_surv_rate_group->start_adding_regions(); 836 // also call this on any additional surv rate groups 837 838 record_survivor_regions(0, NULL, NULL); 839 840 _free_regions_at_end_of_collection = _g1->num_free_regions(); 841 // Reset survivors SurvRateGroup. 842 _survivor_surv_rate_group->reset(); 843 update_young_list_max_and_target_length(); 844 update_rs_lengths_prediction(); 845 cset_chooser()->clear(); 846 847 _bytes_allocated_in_old_since_last_gc = 0; 848 849 record_pause(FullGC, _full_collection_start_sec, end_sec); 850 } 851 852 void G1CollectorPolicy::record_stop_world_start() { 853 _stop_world_start = os::elapsedTime(); 854 } 855 856 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 857 // We only need to do this here as the policy will only be applied 858 // to the GC we're about to start. so, no point is calculating this 859 // every time we calculate / recalculate the target young length. 860 update_survivors_policy(); 861 862 assert(_g1->used() == _g1->recalculate_used(), 863 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 864 _g1->used(), _g1->recalculate_used()); 865 866 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 867 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 868 _stop_world_start = 0.0; 869 870 record_heap_size_info_at_start(false /* full */); 871 872 phase_times()->record_cur_collection_start_sec(start_time_sec); 873 _pending_cards = _g1->pending_card_num(); 874 875 _collection_set_bytes_used_before = 0; 876 _bytes_copied_during_gc = 0; 877 878 collector_state()->set_last_gc_was_young(false); 879 880 // do that for any other surv rate groups 881 _short_lived_surv_rate_group->stop_adding_regions(); 882 _survivors_age_table.clear(); 883 884 assert( verify_young_ages(), "region age verification" ); 885 } 886 887 void G1CollectorPolicy::record_concurrent_mark_init_end(double 888 mark_init_elapsed_time_ms) { 889 collector_state()->set_during_marking(true); 890 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 891 collector_state()->set_during_initial_mark_pause(false); 892 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 893 } 894 895 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 896 _mark_remark_start_sec = os::elapsedTime(); 897 collector_state()->set_during_marking(false); 898 } 899 900 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 901 double end_time_sec = os::elapsedTime(); 902 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 903 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 904 _cur_mark_stop_world_time_ms += elapsed_time_ms; 905 _prev_collection_pause_end_ms += elapsed_time_ms; 906 907 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 908 } 909 910 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 911 _mark_cleanup_start_sec = os::elapsedTime(); 912 } 913 914 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 915 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 916 "skip last young-only gc"); 917 collector_state()->set_last_young_gc(should_continue_with_reclaim); 918 // We skip the marking phase. 919 if (!should_continue_with_reclaim) { 920 abort_time_to_mixed_tracking(); 921 } 922 collector_state()->set_in_marking_window(false); 923 } 924 925 void G1CollectorPolicy::record_concurrent_pause() { 926 if (_stop_world_start > 0.0) { 927 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 928 _trace_young_gen_time_data.record_yield_time(yield_ms); 929 } 930 } 931 932 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 933 return phase_times()->average_time_ms(phase); 934 } 935 936 double G1CollectorPolicy::young_other_time_ms() const { 937 return phase_times()->young_cset_choice_time_ms() + 938 phase_times()->young_free_cset_time_ms(); 939 } 940 941 double G1CollectorPolicy::non_young_other_time_ms() const { 942 return phase_times()->non_young_cset_choice_time_ms() + 943 phase_times()->non_young_free_cset_time_ms(); 944 945 } 946 947 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 948 return pause_time_ms - 949 average_time_ms(G1GCPhaseTimes::UpdateRS) - 950 average_time_ms(G1GCPhaseTimes::ScanRS) - 951 average_time_ms(G1GCPhaseTimes::ObjCopy) - 952 average_time_ms(G1GCPhaseTimes::Termination); 953 } 954 955 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 956 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 957 } 958 959 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 960 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 961 } 962 963 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 964 if (about_to_start_mixed_phase()) { 965 return false; 966 } 967 968 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 969 970 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 971 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 972 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 973 974 bool result = false; 975 if (marking_request_bytes > marking_initiating_used_threshold) { 976 result = collector_state()->gcs_are_young() && !collector_state()->last_young_gc(); 977 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s", 978 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)", 979 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1->capacity() * 100, source); 980 } 981 982 return result; 983 } 984 985 // Anything below that is considered to be zero 986 #define MIN_TIMER_GRANULARITY 0.0000001 987 988 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 989 double end_time_sec = os::elapsedTime(); 990 991 size_t cur_used_bytes = _g1->used(); 992 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 993 bool last_pause_included_initial_mark = false; 994 bool update_stats = !_g1->evacuation_failed(); 995 996 NOT_PRODUCT(_short_lived_surv_rate_group->print()); 997 998 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec); 999 1000 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1001 if (last_pause_included_initial_mark) { 1002 record_concurrent_mark_init_end(0.0); 1003 } else { 1004 maybe_start_marking(); 1005 } 1006 1007 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1008 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1009 // This usually happens due to the timer not having the required 1010 // granularity. Some Linuxes are the usual culprits. 1011 // We'll just set it to something (arbitrarily) small. 1012 app_time_ms = 1.0; 1013 } 1014 1015 if (update_stats) { 1016 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1017 // We maintain the invariant that all objects allocated by mutator 1018 // threads will be allocated out of eden regions. So, we can use 1019 // the eden region number allocated since the previous GC to 1020 // calculate the application's allocate rate. The only exception 1021 // to that is humongous objects that are allocated separately. But 1022 // given that humongous object allocations do not really affect 1023 // either the pause's duration nor when the next pause will take 1024 // place we can safely ignore them here. 1025 uint regions_allocated = eden_cset_region_length(); 1026 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1027 _alloc_rate_ms_seq->add(alloc_rate_ms); 1028 1029 double interval_ms = 1030 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1031 update_recent_gc_times(end_time_sec, pause_time_ms); 1032 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1033 if (recent_avg_pause_time_ratio() < 0.0 || 1034 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1035 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1036 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1037 if (_recent_avg_pause_time_ratio < 0.0) { 1038 _recent_avg_pause_time_ratio = 0.0; 1039 } else { 1040 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1041 _recent_avg_pause_time_ratio = 1.0; 1042 } 1043 } 1044 1045 // Compute the ratio of just this last pause time to the entire time range stored 1046 // in the vectors. Comparing this pause to the entire range, rather than only the 1047 // most recent interval, has the effect of smoothing over a possible transient 'burst' 1048 // of more frequent pauses that don't really reflect a change in heap occupancy. 1049 // This reduces the likelihood of a needless heap expansion being triggered. 1050 _last_pause_time_ratio = 1051 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms; 1052 } 1053 1054 bool new_in_marking_window = collector_state()->in_marking_window(); 1055 bool new_in_marking_window_im = false; 1056 if (last_pause_included_initial_mark) { 1057 new_in_marking_window = true; 1058 new_in_marking_window_im = true; 1059 } 1060 1061 if (collector_state()->last_young_gc()) { 1062 // This is supposed to to be the "last young GC" before we start 1063 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1064 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1065 1066 if (next_gc_should_be_mixed("start mixed GCs", 1067 "do not start mixed GCs")) { 1068 collector_state()->set_gcs_are_young(false); 1069 } else { 1070 // We aborted the mixed GC phase early. 1071 abort_time_to_mixed_tracking(); 1072 } 1073 1074 collector_state()->set_last_young_gc(false); 1075 } 1076 1077 if (!collector_state()->last_gc_was_young()) { 1078 // This is a mixed GC. Here we decide whether to continue doing 1079 // mixed GCs or not. 1080 if (!next_gc_should_be_mixed("continue mixed GCs", 1081 "do not continue mixed GCs")) { 1082 collector_state()->set_gcs_are_young(true); 1083 1084 maybe_start_marking(); 1085 } 1086 } 1087 1088 _short_lived_surv_rate_group->start_adding_regions(); 1089 // Do that for any other surv rate groups 1090 1091 if (update_stats) { 1092 double cost_per_card_ms = 0.0; 1093 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1094 if (_pending_cards > 0) { 1095 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1096 _cost_per_card_ms_seq->add(cost_per_card_ms); 1097 } 1098 _cost_scan_hcc_seq->add(cost_scan_hcc); 1099 1100 double cost_per_entry_ms = 0.0; 1101 if (cards_scanned > 10) { 1102 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1103 if (collector_state()->last_gc_was_young()) { 1104 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1105 } else { 1106 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1107 } 1108 } 1109 1110 if (_max_rs_lengths > 0) { 1111 double cards_per_entry_ratio = 1112 (double) cards_scanned / (double) _max_rs_lengths; 1113 if (collector_state()->last_gc_was_young()) { 1114 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1115 } else { 1116 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1117 } 1118 } 1119 1120 // This is defensive. For a while _max_rs_lengths could get 1121 // smaller than _recorded_rs_lengths which was causing 1122 // rs_length_diff to get very large and mess up the RSet length 1123 // predictions. The reason was unsafe concurrent updates to the 1124 // _inc_cset_recorded_rs_lengths field which the code below guards 1125 // against (see CR 7118202). This bug has now been fixed (see CR 1126 // 7119027). However, I'm still worried that 1127 // _inc_cset_recorded_rs_lengths might still end up somewhat 1128 // inaccurate. The concurrent refinement thread calculates an 1129 // RSet's length concurrently with other CR threads updating it 1130 // which might cause it to calculate the length incorrectly (if, 1131 // say, it's in mid-coarsening). So I'll leave in the defensive 1132 // conditional below just in case. 1133 size_t rs_length_diff = 0; 1134 if (_max_rs_lengths > _recorded_rs_lengths) { 1135 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1136 } 1137 _rs_length_diff_seq->add((double) rs_length_diff); 1138 1139 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1140 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1141 double cost_per_byte_ms = 0.0; 1142 1143 if (copied_bytes > 0) { 1144 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1145 if (collector_state()->in_marking_window()) { 1146 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1147 } else { 1148 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1149 } 1150 } 1151 1152 if (young_cset_region_length() > 0) { 1153 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1154 young_cset_region_length()); 1155 } 1156 1157 if (old_cset_region_length() > 0) { 1158 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1159 old_cset_region_length()); 1160 } 1161 1162 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1163 1164 _pending_cards_seq->add((double) _pending_cards); 1165 _rs_lengths_seq->add((double) _max_rs_lengths); 1166 } 1167 1168 collector_state()->set_in_marking_window(new_in_marking_window); 1169 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1170 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1171 // IHOP control wants to know the expected young gen length if it were not 1172 // restrained by the heap reserve. Using the actual length would make the 1173 // prediction too small and the limit the young gen every time we get to the 1174 // predicted target occupancy. 1175 size_t last_unrestrained_young_length = update_young_list_max_and_target_length(); 1176 update_rs_lengths_prediction(); 1177 1178 update_ihop_prediction(app_time_ms / 1000.0, 1179 _bytes_allocated_in_old_since_last_gc, 1180 last_unrestrained_young_length * HeapRegion::GrainBytes); 1181 _bytes_allocated_in_old_since_last_gc = 0; 1182 1183 _ihop_control->send_trace_event(_g1->gc_tracer_stw()); 1184 1185 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1186 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1187 1188 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1189 1190 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1191 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 1192 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms", 1193 update_rs_time_goal_ms, scan_hcc_time_ms); 1194 1195 update_rs_time_goal_ms = 0; 1196 } else { 1197 update_rs_time_goal_ms -= scan_hcc_time_ms; 1198 } 1199 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1200 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1201 update_rs_time_goal_ms); 1202 1203 cset_chooser()->verify(); 1204 } 1205 1206 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1207 if (G1UseAdaptiveIHOP) { 1208 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1209 G1CollectedHeap::heap()->max_capacity(), 1210 &_predictor, 1211 G1ReservePercent, 1212 G1HeapWastePercent); 1213 } else { 1214 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1215 G1CollectedHeap::heap()->max_capacity()); 1216 } 1217 } 1218 1219 void G1CollectorPolicy::update_ihop_prediction(double mutator_time_s, 1220 size_t mutator_alloc_bytes, 1221 size_t young_gen_size) { 1222 // Always try to update IHOP prediction. Even evacuation failures give information 1223 // about e.g. whether to start IHOP earlier next time. 1224 1225 // Avoid using really small application times that might create samples with 1226 // very high or very low values. They may be caused by e.g. back-to-back gcs. 1227 double const min_valid_time = 1e-6; 1228 1229 bool report = false; 1230 1231 double marking_to_mixed_time = -1.0; 1232 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1233 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1234 assert(marking_to_mixed_time > 0.0, 1235 "Initial mark to mixed time must be larger than zero but is %.3f", 1236 marking_to_mixed_time); 1237 if (marking_to_mixed_time > min_valid_time) { 1238 _ihop_control->update_marking_length(marking_to_mixed_time); 1239 report = true; 1240 } 1241 } 1242 1243 // As an approximation for the young gc promotion rates during marking we use 1244 // all of them. In many applications there are only a few if any young gcs during 1245 // marking, which makes any prediction useless. This increases the accuracy of the 1246 // prediction. 1247 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1248 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1249 report = true; 1250 } 1251 1252 if (report) { 1253 report_ihop_statistics(); 1254 } 1255 } 1256 1257 void G1CollectorPolicy::report_ihop_statistics() { 1258 _ihop_control->print(); 1259 } 1260 1261 #define EXT_SIZE_FORMAT "%.1f%s" 1262 #define EXT_SIZE_PARAMS(bytes) \ 1263 byte_size_in_proper_unit((double)(bytes)), \ 1264 proper_unit_for_byte_size((bytes)) 1265 1266 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1267 YoungList* young_list = _g1->young_list(); 1268 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1269 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1270 _heap_capacity_bytes_before_gc = _g1->capacity(); 1271 _old_used_bytes_before_gc = _g1->old_regions_count() * HeapRegion::GrainBytes; 1272 _humongous_used_bytes_before_gc = _g1->humongous_regions_count() * HeapRegion::GrainBytes; 1273 _heap_used_bytes_before_gc = _g1->used(); 1274 _eden_capacity_bytes_before_gc = (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1275 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1276 } 1277 1278 void G1CollectorPolicy::print_detailed_heap_transition() const { 1279 YoungList* young_list = _g1->young_list(); 1280 1281 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1282 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1283 size_t heap_used_bytes_after_gc = _g1->used(); 1284 size_t old_used_bytes_after_gc = _g1->old_regions_count() * HeapRegion::GrainBytes; 1285 size_t humongous_used_bytes_after_gc = _g1->humongous_regions_count() * HeapRegion::GrainBytes; 1286 1287 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1288 size_t eden_capacity_bytes_after_gc = 1289 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1290 size_t survivor_capacity_bytes_after_gc = _max_survivor_regions * HeapRegion::GrainBytes; 1291 1292 log_info(gc, heap)("Eden: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1293 _eden_used_bytes_before_gc / K, eden_used_bytes_after_gc /K, eden_capacity_bytes_after_gc /K); 1294 log_info(gc, heap)("Survivor: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1295 _survivor_used_bytes_before_gc / K, survivor_used_bytes_after_gc /K, survivor_capacity_bytes_after_gc /K); 1296 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K", 1297 _old_used_bytes_before_gc / K, old_used_bytes_after_gc /K); 1298 log_info(gc, heap)("Humongous: " SIZE_FORMAT "K->" SIZE_FORMAT "K", 1299 _humongous_used_bytes_before_gc / K, humongous_used_bytes_after_gc /K); 1300 1301 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1302 } 1303 1304 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1305 phase_times()->print(pause_time_sec); 1306 } 1307 1308 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1309 double update_rs_processed_buffers, 1310 double goal_ms) { 1311 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1312 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1313 1314 if (G1UseAdaptiveConcRefinement) { 1315 const int k_gy = 3, k_gr = 6; 1316 const double inc_k = 1.1, dec_k = 0.9; 1317 1318 int g = cg1r->green_zone(); 1319 if (update_rs_time > goal_ms) { 1320 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1321 } else { 1322 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1323 g = (int)MAX2(g * inc_k, g + 1.0); 1324 } 1325 } 1326 // Change the refinement threads params 1327 cg1r->set_green_zone(g); 1328 cg1r->set_yellow_zone(g * k_gy); 1329 cg1r->set_red_zone(g * k_gr); 1330 cg1r->reinitialize_threads(); 1331 1332 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1333 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1334 cg1r->yellow_zone()); 1335 // Change the barrier params 1336 dcqs.set_process_completed_threshold(processing_threshold); 1337 dcqs.set_max_completed_queue(cg1r->red_zone()); 1338 } 1339 1340 int curr_queue_size = dcqs.completed_buffers_num(); 1341 if (curr_queue_size >= cg1r->yellow_zone()) { 1342 dcqs.set_completed_queue_padding(curr_queue_size); 1343 } else { 1344 dcqs.set_completed_queue_padding(0); 1345 } 1346 dcqs.notify_if_necessary(); 1347 } 1348 1349 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1350 return (size_t) get_new_prediction(_rs_length_diff_seq); 1351 } 1352 1353 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1354 return get_new_prediction(_alloc_rate_ms_seq); 1355 } 1356 1357 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1358 return get_new_prediction(_cost_per_card_ms_seq); 1359 } 1360 1361 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1362 return get_new_prediction(_cost_scan_hcc_seq); 1363 } 1364 1365 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1366 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1367 } 1368 1369 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1370 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1371 } 1372 1373 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1374 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1375 return predict_young_cards_per_entry_ratio(); 1376 } else { 1377 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1378 } 1379 } 1380 1381 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1382 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1383 } 1384 1385 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1386 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1387 } 1388 1389 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1390 if (collector_state()->gcs_are_young()) { 1391 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1392 } else { 1393 return predict_mixed_rs_scan_time_ms(card_num); 1394 } 1395 } 1396 1397 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1398 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1399 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1400 } else { 1401 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1402 } 1403 } 1404 1405 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1406 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1407 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1408 } else { 1409 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1410 } 1411 } 1412 1413 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1414 if (collector_state()->during_concurrent_mark()) { 1415 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1416 } else { 1417 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1418 } 1419 } 1420 1421 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1422 return get_new_prediction(_constant_other_time_ms_seq); 1423 } 1424 1425 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1426 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1427 } 1428 1429 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1430 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1431 } 1432 1433 double G1CollectorPolicy::predict_remark_time_ms() const { 1434 return get_new_prediction(_concurrent_mark_remark_times_ms); 1435 } 1436 1437 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1438 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1439 } 1440 1441 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1442 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1443 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1444 double pred = get_new_prediction(seq); 1445 if (pred > 1.0) { 1446 pred = 1.0; 1447 } 1448 return pred; 1449 } 1450 1451 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1452 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1453 } 1454 1455 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1456 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1457 } 1458 1459 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1460 size_t scanned_cards) const { 1461 return 1462 predict_rs_update_time_ms(pending_cards) + 1463 predict_rs_scan_time_ms(scanned_cards) + 1464 predict_constant_other_time_ms(); 1465 } 1466 1467 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1468 size_t rs_length = predict_rs_length_diff(); 1469 size_t card_num; 1470 if (collector_state()->gcs_are_young()) { 1471 card_num = predict_young_card_num(rs_length); 1472 } else { 1473 card_num = predict_non_young_card_num(rs_length); 1474 } 1475 return predict_base_elapsed_time_ms(pending_cards, card_num); 1476 } 1477 1478 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1479 size_t bytes_to_copy; 1480 if (hr->is_marked()) 1481 bytes_to_copy = hr->max_live_bytes(); 1482 else { 1483 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1484 int age = hr->age_in_surv_rate_group(); 1485 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1486 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1487 } 1488 return bytes_to_copy; 1489 } 1490 1491 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1492 bool for_young_gc) const { 1493 size_t rs_length = hr->rem_set()->occupied(); 1494 size_t card_num; 1495 1496 // Predicting the number of cards is based on which type of GC 1497 // we're predicting for. 1498 if (for_young_gc) { 1499 card_num = predict_young_card_num(rs_length); 1500 } else { 1501 card_num = predict_non_young_card_num(rs_length); 1502 } 1503 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1504 1505 double region_elapsed_time_ms = 1506 predict_rs_scan_time_ms(card_num) + 1507 predict_object_copy_time_ms(bytes_to_copy); 1508 1509 // The prediction of the "other" time for this region is based 1510 // upon the region type and NOT the GC type. 1511 if (hr->is_young()) { 1512 region_elapsed_time_ms += predict_young_other_time_ms(1); 1513 } else { 1514 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1515 } 1516 return region_elapsed_time_ms; 1517 } 1518 1519 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1520 uint survivor_cset_region_length) { 1521 _eden_cset_region_length = eden_cset_region_length; 1522 _survivor_cset_region_length = survivor_cset_region_length; 1523 _old_cset_region_length = 0; 1524 } 1525 1526 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1527 _recorded_rs_lengths = rs_lengths; 1528 } 1529 1530 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1531 double elapsed_ms) { 1532 _recent_gc_times_ms->add(elapsed_ms); 1533 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1534 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1535 } 1536 1537 void G1CollectorPolicy::clear_ratio_check_data() { 1538 _ratio_over_threshold_count = 0; 1539 _ratio_over_threshold_sum = 0.0; 1540 _pauses_since_start = 0; 1541 } 1542 1543 size_t G1CollectorPolicy::expansion_amount() { 1544 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1545 double last_gc_overhead = _last_pause_time_ratio * 100.0; 1546 double threshold = _gc_overhead_perc; 1547 size_t expand_bytes = 0; 1548 1549 // If the heap is at less than half its maximum size, scale the threshold down, 1550 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand, 1551 // though the scaling code will likely keep the increase small. 1552 if (_g1->capacity() <= _g1->max_capacity() / 2) { 1553 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2); 1554 threshold = MAX2(threshold, 1.0); 1555 } 1556 1557 // If the last GC time ratio is over the threshold, increment the count of 1558 // times it has been exceeded, and add this ratio to the sum of exceeded 1559 // ratios. 1560 if (last_gc_overhead > threshold) { 1561 _ratio_over_threshold_count++; 1562 _ratio_over_threshold_sum += last_gc_overhead; 1563 } 1564 1565 // Check if we've had enough GC time ratio checks that were over the 1566 // threshold to trigger an expansion. We'll also expand if we've 1567 // reached the end of the history buffer and the average of all entries 1568 // is still over the threshold. This indicates a smaller number of GCs were 1569 // long enough to make the average exceed the threshold. 1570 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics; 1571 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) || 1572 (filled_history_buffer && (recent_gc_overhead > threshold))) { 1573 size_t min_expand_bytes = HeapRegion::GrainBytes; 1574 size_t reserved_bytes = _g1->max_capacity(); 1575 size_t committed_bytes = _g1->capacity(); 1576 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1577 size_t expand_bytes_via_pct = 1578 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1579 double scale_factor = 1.0; 1580 1581 // If the current size is less than 1/4 of the Initial heap size, expand 1582 // by half of the delta between the current and Initial sizes. IE, grow 1583 // back quickly. 1584 // 1585 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of 1586 // the available expansion space, whichever is smaller, as the base 1587 // expansion size. Then possibly scale this size according to how much the 1588 // threshold has (on average) been exceeded by. If the delta is small 1589 // (less than the StartScaleDownAt value), scale the size down linearly, but 1590 // not by less than MinScaleDownFactor. If the delta is large (greater than 1591 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor 1592 // times the base size. The scaling will be linear in the range from 1593 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words, 1594 // ScaleUpRange sets the rate of scaling up. 1595 if (committed_bytes < InitialHeapSize / 4) { 1596 expand_bytes = (InitialHeapSize - committed_bytes) / 2; 1597 } else { 1598 double const MinScaleDownFactor = 0.2; 1599 double const MaxScaleUpFactor = 2; 1600 double const StartScaleDownAt = _gc_overhead_perc; 1601 double const StartScaleUpAt = _gc_overhead_perc * 1.5; 1602 double const ScaleUpRange = _gc_overhead_perc * 2.0; 1603 1604 double ratio_delta; 1605 if (filled_history_buffer) { 1606 ratio_delta = recent_gc_overhead - threshold; 1607 } else { 1608 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold; 1609 } 1610 1611 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1612 if (ratio_delta < StartScaleDownAt) { 1613 scale_factor = ratio_delta / StartScaleDownAt; 1614 scale_factor = MAX2(scale_factor, MinScaleDownFactor); 1615 } else if (ratio_delta > StartScaleUpAt) { 1616 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange); 1617 scale_factor = MIN2(scale_factor, MaxScaleUpFactor); 1618 } 1619 } 1620 1621 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) " 1622 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)", 1623 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100); 1624 1625 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor); 1626 1627 // Ensure the expansion size is at least the minimum growth amount 1628 // and at most the remaining uncommitted byte size. 1629 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1630 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1631 1632 clear_ratio_check_data(); 1633 } else { 1634 // An expansion was not triggered. If we've started counting, increment 1635 // the number of checks we've made in the current window. If we've 1636 // reached the end of the window without resizing, clear the counters to 1637 // start again the next time we see a ratio above the threshold. 1638 if (_ratio_over_threshold_count > 0) { 1639 _pauses_since_start++; 1640 if (_pauses_since_start > NumPrevPausesForHeuristics) { 1641 clear_ratio_check_data(); 1642 } 1643 } 1644 } 1645 1646 return expand_bytes; 1647 } 1648 1649 void G1CollectorPolicy::print_tracing_info() const { 1650 _trace_young_gen_time_data.print(); 1651 _trace_old_gen_time_data.print(); 1652 } 1653 1654 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1655 #ifndef PRODUCT 1656 _short_lived_surv_rate_group->print_surv_rate_summary(); 1657 // add this call for any other surv rate groups 1658 #endif // PRODUCT 1659 } 1660 1661 bool G1CollectorPolicy::is_young_list_full() const { 1662 uint young_list_length = _g1->young_list()->length(); 1663 uint young_list_target_length = _young_list_target_length; 1664 return young_list_length >= young_list_target_length; 1665 } 1666 1667 bool G1CollectorPolicy::can_expand_young_list() const { 1668 uint young_list_length = _g1->young_list()->length(); 1669 uint young_list_max_length = _young_list_max_length; 1670 return young_list_length < young_list_max_length; 1671 } 1672 1673 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1674 uint expansion_region_num = 0; 1675 if (GCLockerEdenExpansionPercent > 0) { 1676 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1677 double expansion_region_num_d = perc * (double) _young_list_target_length; 1678 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1679 // less than 1.0) we'll get 1. 1680 expansion_region_num = (uint) ceil(expansion_region_num_d); 1681 } else { 1682 assert(expansion_region_num == 0, "sanity"); 1683 } 1684 _young_list_max_length = _young_list_target_length + expansion_region_num; 1685 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1686 } 1687 1688 // Calculates survivor space parameters. 1689 void G1CollectorPolicy::update_survivors_policy() { 1690 double max_survivor_regions_d = 1691 (double) _young_list_target_length / (double) SurvivorRatio; 1692 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1693 // smaller than 1.0) we'll get 1. 1694 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1695 1696 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1697 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1698 } 1699 1700 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1701 // We actually check whether we are marking here and not if we are in a 1702 // reclamation phase. This means that we will schedule a concurrent mark 1703 // even while we are still in the process of reclaiming memory. 1704 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1705 if (!during_cycle) { 1706 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause)); 1707 collector_state()->set_initiate_conc_mark_if_possible(true); 1708 return true; 1709 } else { 1710 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause)); 1711 return false; 1712 } 1713 } 1714 1715 void G1CollectorPolicy::initiate_conc_mark() { 1716 collector_state()->set_during_initial_mark_pause(true); 1717 collector_state()->set_initiate_conc_mark_if_possible(false); 1718 } 1719 1720 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1721 // We are about to decide on whether this pause will be an 1722 // initial-mark pause. 1723 1724 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1725 // will set it here if we have to. However, it should be cleared by 1726 // the end of the pause (it's only set for the duration of an 1727 // initial-mark pause). 1728 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1729 1730 if (collector_state()->initiate_conc_mark_if_possible()) { 1731 // We had noticed on a previous pause that the heap occupancy has 1732 // gone over the initiating threshold and we should start a 1733 // concurrent marking cycle. So we might initiate one. 1734 1735 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1736 // Initiate a new initial mark if there is no marking or reclamation going on. 1737 initiate_conc_mark(); 1738 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1739 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) { 1740 // Initiate a user requested initial mark. An initial mark must be young only 1741 // GC, so the collector state must be updated to reflect this. 1742 collector_state()->set_gcs_are_young(true); 1743 collector_state()->set_last_young_gc(false); 1744 1745 abort_time_to_mixed_tracking(); 1746 initiate_conc_mark(); 1747 log_debug(gc, ergo)("Initiate concurrent cycle (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 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1762 } 1763 } 1764 } 1765 1766 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1767 G1CollectedHeap* _g1h; 1768 CSetChooserParUpdater _cset_updater; 1769 1770 public: 1771 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1772 uint chunk_size) : 1773 _g1h(G1CollectedHeap::heap()), 1774 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1775 1776 bool doHeapRegion(HeapRegion* r) { 1777 // Do we have any marking information for this region? 1778 if (r->is_marked()) { 1779 // We will skip any region that's currently used as an old GC 1780 // alloc region (we should not consider those for collection 1781 // before we fill them up). 1782 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1783 _cset_updater.add_region(r); 1784 } 1785 } 1786 return false; 1787 } 1788 }; 1789 1790 class ParKnownGarbageTask: public AbstractGangTask { 1791 CollectionSetChooser* _hrSorted; 1792 uint _chunk_size; 1793 G1CollectedHeap* _g1; 1794 HeapRegionClaimer _hrclaimer; 1795 1796 public: 1797 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1798 AbstractGangTask("ParKnownGarbageTask"), 1799 _hrSorted(hrSorted), _chunk_size(chunk_size), 1800 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1801 1802 void work(uint worker_id) { 1803 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1804 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1805 } 1806 }; 1807 1808 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1809 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1810 const uint overpartition_factor = 4; 1811 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1812 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1813 } 1814 1815 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1816 cset_chooser()->clear(); 1817 1818 WorkGang* workers = _g1->workers(); 1819 uint n_workers = workers->active_workers(); 1820 1821 uint n_regions = _g1->num_regions(); 1822 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1823 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1824 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers); 1825 workers->run_task(&par_known_garbage_task); 1826 1827 cset_chooser()->sort_regions(); 1828 1829 double end_sec = os::elapsedTime(); 1830 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1831 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1832 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1833 _prev_collection_pause_end_ms += elapsed_time_ms; 1834 1835 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1836 } 1837 1838 // Add the heap region at the head of the non-incremental collection set 1839 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1840 assert(_inc_cset_build_state == Active, "Precondition"); 1841 assert(hr->is_old(), "the region should be old"); 1842 1843 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1844 _g1->register_old_region_with_cset(hr); 1845 hr->set_next_in_collection_set(_collection_set); 1846 _collection_set = hr; 1847 _collection_set_bytes_used_before += hr->used(); 1848 size_t rs_length = hr->rem_set()->occupied(); 1849 _recorded_rs_lengths += rs_length; 1850 _old_cset_region_length += 1; 1851 } 1852 1853 // Initialize the per-collection-set information 1854 void G1CollectorPolicy::start_incremental_cset_building() { 1855 assert(_inc_cset_build_state == Inactive, "Precondition"); 1856 1857 _inc_cset_head = NULL; 1858 _inc_cset_tail = NULL; 1859 _inc_cset_bytes_used_before = 0; 1860 1861 _inc_cset_max_finger = 0; 1862 _inc_cset_recorded_rs_lengths = 0; 1863 _inc_cset_recorded_rs_lengths_diffs = 0; 1864 _inc_cset_predicted_elapsed_time_ms = 0.0; 1865 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1866 _inc_cset_build_state = Active; 1867 } 1868 1869 void G1CollectorPolicy::finalize_incremental_cset_building() { 1870 assert(_inc_cset_build_state == Active, "Precondition"); 1871 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1872 1873 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1874 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1875 // that adds a new region to the CSet. Further updates by the 1876 // concurrent refinement thread that samples the young RSet lengths 1877 // are accumulated in the *_diffs fields. Here we add the diffs to 1878 // the "main" fields. 1879 1880 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1881 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1882 } else { 1883 // This is defensive. The diff should in theory be always positive 1884 // as RSets can only grow between GCs. However, given that we 1885 // sample their size concurrently with other threads updating them 1886 // it's possible that we might get the wrong size back, which 1887 // could make the calculations somewhat inaccurate. 1888 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1889 if (_inc_cset_recorded_rs_lengths >= diffs) { 1890 _inc_cset_recorded_rs_lengths -= diffs; 1891 } else { 1892 _inc_cset_recorded_rs_lengths = 0; 1893 } 1894 } 1895 _inc_cset_predicted_elapsed_time_ms += 1896 _inc_cset_predicted_elapsed_time_ms_diffs; 1897 1898 _inc_cset_recorded_rs_lengths_diffs = 0; 1899 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1900 } 1901 1902 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1903 // This routine is used when: 1904 // * adding survivor regions to the incremental cset at the end of an 1905 // evacuation pause, 1906 // * adding the current allocation region to the incremental cset 1907 // when it is retired, and 1908 // * updating existing policy information for a region in the 1909 // incremental cset via young list RSet sampling. 1910 // Therefore this routine may be called at a safepoint by the 1911 // VM thread, or in-between safepoints by mutator threads (when 1912 // retiring the current allocation region) or a concurrent 1913 // refine thread (RSet sampling). 1914 1915 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1916 size_t used_bytes = hr->used(); 1917 _inc_cset_recorded_rs_lengths += rs_length; 1918 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1919 _inc_cset_bytes_used_before += used_bytes; 1920 1921 // Cache the values we have added to the aggregated information 1922 // in the heap region in case we have to remove this region from 1923 // the incremental collection set, or it is updated by the 1924 // rset sampling code 1925 hr->set_recorded_rs_length(rs_length); 1926 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1927 } 1928 1929 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1930 size_t new_rs_length) { 1931 // Update the CSet information that is dependent on the new RS length 1932 assert(hr->is_young(), "Precondition"); 1933 assert(!SafepointSynchronize::is_at_safepoint(), 1934 "should not be at a safepoint"); 1935 1936 // We could have updated _inc_cset_recorded_rs_lengths and 1937 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1938 // that atomically, as this code is executed by a concurrent 1939 // refinement thread, potentially concurrently with a mutator thread 1940 // allocating a new region and also updating the same fields. To 1941 // avoid the atomic operations we accumulate these updates on two 1942 // separate fields (*_diffs) and we'll just add them to the "main" 1943 // fields at the start of a GC. 1944 1945 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1946 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1947 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1948 1949 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1950 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1951 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1952 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1953 1954 hr->set_recorded_rs_length(new_rs_length); 1955 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1956 } 1957 1958 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1959 assert(hr->is_young(), "invariant"); 1960 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1961 assert(_inc_cset_build_state == Active, "Precondition"); 1962 1963 // We need to clear and set the cached recorded/cached collection set 1964 // information in the heap region here (before the region gets added 1965 // to the collection set). An individual heap region's cached values 1966 // are calculated, aggregated with the policy collection set info, 1967 // and cached in the heap region here (initially) and (subsequently) 1968 // by the Young List sampling code. 1969 1970 size_t rs_length = hr->rem_set()->occupied(); 1971 add_to_incremental_cset_info(hr, rs_length); 1972 1973 HeapWord* hr_end = hr->end(); 1974 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1975 1976 assert(!hr->in_collection_set(), "invariant"); 1977 _g1->register_young_region_with_cset(hr); 1978 assert(hr->next_in_collection_set() == NULL, "invariant"); 1979 } 1980 1981 // Add the region at the RHS of the incremental cset 1982 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1983 // We should only ever be appending survivors at the end of a pause 1984 assert(hr->is_survivor(), "Logic"); 1985 1986 // Do the 'common' stuff 1987 add_region_to_incremental_cset_common(hr); 1988 1989 // Now add the region at the right hand side 1990 if (_inc_cset_tail == NULL) { 1991 assert(_inc_cset_head == NULL, "invariant"); 1992 _inc_cset_head = hr; 1993 } else { 1994 _inc_cset_tail->set_next_in_collection_set(hr); 1995 } 1996 _inc_cset_tail = hr; 1997 } 1998 1999 // Add the region to the LHS of the incremental cset 2000 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2001 // Survivors should be added to the RHS at the end of a pause 2002 assert(hr->is_eden(), "Logic"); 2003 2004 // Do the 'common' stuff 2005 add_region_to_incremental_cset_common(hr); 2006 2007 // Add the region at the left hand side 2008 hr->set_next_in_collection_set(_inc_cset_head); 2009 if (_inc_cset_head == NULL) { 2010 assert(_inc_cset_tail == NULL, "Invariant"); 2011 _inc_cset_tail = hr; 2012 } 2013 _inc_cset_head = hr; 2014 } 2015 2016 #ifndef PRODUCT 2017 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2018 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2019 2020 st->print_cr("\nCollection_set:"); 2021 HeapRegion* csr = list_head; 2022 while (csr != NULL) { 2023 HeapRegion* next = csr->next_in_collection_set(); 2024 assert(csr->in_collection_set(), "bad CS"); 2025 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2026 HR_FORMAT_PARAMS(csr), 2027 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2028 csr->age_in_surv_rate_group_cond()); 2029 csr = next; 2030 } 2031 } 2032 #endif // !PRODUCT 2033 2034 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2035 // Returns the given amount of reclaimable bytes (that represents 2036 // the amount of reclaimable space still to be collected) as a 2037 // percentage of the current heap capacity. 2038 size_t capacity_bytes = _g1->capacity(); 2039 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2040 } 2041 2042 void G1CollectorPolicy::maybe_start_marking() { 2043 if (need_to_start_conc_mark("end of GC")) { 2044 // Note: this might have already been set, if during the last 2045 // pause we decided to start a cycle but at the beginning of 2046 // this pause we decided to postpone it. That's OK. 2047 collector_state()->set_initiate_conc_mark_if_possible(true); 2048 } 2049 } 2050 2051 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2052 assert(!collector_state()->full_collection(), "must be"); 2053 if (collector_state()->during_initial_mark_pause()) { 2054 assert(collector_state()->last_gc_was_young(), "must be"); 2055 assert(!collector_state()->last_young_gc(), "must be"); 2056 return InitialMarkGC; 2057 } else if (collector_state()->last_young_gc()) { 2058 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2059 assert(collector_state()->last_gc_was_young(), "must be"); 2060 return LastYoungGC; 2061 } else if (!collector_state()->last_gc_was_young()) { 2062 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2063 assert(!collector_state()->last_young_gc(), "must be"); 2064 return MixedGC; 2065 } else { 2066 assert(collector_state()->last_gc_was_young(), "must be"); 2067 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2068 assert(!collector_state()->last_young_gc(), "must be"); 2069 return YoungOnlyGC; 2070 } 2071 } 2072 2073 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2074 // Manage the MMU tracker. For some reason it ignores Full GCs. 2075 if (kind != FullGC) { 2076 _mmu_tracker->add_pause(start, end); 2077 } 2078 // Manage the mutator time tracking from initial mark to first mixed gc. 2079 switch (kind) { 2080 case FullGC: 2081 abort_time_to_mixed_tracking(); 2082 break; 2083 case Cleanup: 2084 case Remark: 2085 case YoungOnlyGC: 2086 case LastYoungGC: 2087 _initial_mark_to_mixed.add_pause(end - start); 2088 break; 2089 case InitialMarkGC: 2090 _initial_mark_to_mixed.record_initial_mark_end(end); 2091 break; 2092 case MixedGC: 2093 _initial_mark_to_mixed.record_mixed_gc_start(start); 2094 break; 2095 default: 2096 ShouldNotReachHere(); 2097 } 2098 } 2099 2100 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2101 _initial_mark_to_mixed.reset(); 2102 } 2103 2104 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2105 const char* false_action_str) const { 2106 if (cset_chooser()->is_empty()) { 2107 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 2108 return false; 2109 } 2110 2111 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2112 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2113 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2114 double threshold = (double) G1HeapWastePercent; 2115 if (reclaimable_perc <= threshold) { 2116 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2117 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2118 return false; 2119 } 2120 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2121 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2122 return true; 2123 } 2124 2125 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2126 // The min old CSet region bound is based on the maximum desired 2127 // number of mixed GCs after a cycle. I.e., even if some old regions 2128 // look expensive, we should add them to the CSet anyway to make 2129 // sure we go through the available old regions in no more than the 2130 // maximum desired number of mixed GCs. 2131 // 2132 // The calculation is based on the number of marked regions we added 2133 // to the CSet chooser in the first place, not how many remain, so 2134 // that the result is the same during all mixed GCs that follow a cycle. 2135 2136 const size_t region_num = (size_t) cset_chooser()->length(); 2137 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2138 size_t result = region_num / gc_num; 2139 // emulate ceiling 2140 if (result * gc_num < region_num) { 2141 result += 1; 2142 } 2143 return (uint) result; 2144 } 2145 2146 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2147 // The max old CSet region bound is based on the threshold expressed 2148 // as a percentage of the heap size. I.e., it should bound the 2149 // number of old regions added to the CSet irrespective of how many 2150 // of them are available. 2151 2152 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2153 const size_t region_num = g1h->num_regions(); 2154 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2155 size_t result = region_num * perc / 100; 2156 // emulate ceiling 2157 if (100 * result < region_num * perc) { 2158 result += 1; 2159 } 2160 return (uint) result; 2161 } 2162 2163 2164 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2165 double young_start_time_sec = os::elapsedTime(); 2166 2167 YoungList* young_list = _g1->young_list(); 2168 finalize_incremental_cset_building(); 2169 2170 guarantee(target_pause_time_ms > 0.0, 2171 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2172 guarantee(_collection_set == NULL, "Precondition"); 2173 2174 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2175 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2176 2177 log_trace(gc, ergo, cset)("Start choosing CSet. pending cards: " SIZE_FORMAT " predicted base time: %1.2fms remaining time: %1.2fms target pause time: %1.2fms", 2178 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2179 2180 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2181 2182 if (collector_state()->last_gc_was_young()) { 2183 _trace_young_gen_time_data.increment_young_collection_count(); 2184 } else { 2185 _trace_young_gen_time_data.increment_mixed_collection_count(); 2186 } 2187 2188 // The young list is laid with the survivor regions from the previous 2189 // pause are appended to the RHS of the young list, i.e. 2190 // [Newly Young Regions ++ Survivors from last pause]. 2191 2192 uint survivor_region_length = young_list->survivor_length(); 2193 uint eden_region_length = young_list->eden_length(); 2194 init_cset_region_lengths(eden_region_length, survivor_region_length); 2195 2196 HeapRegion* hr = young_list->first_survivor_region(); 2197 while (hr != NULL) { 2198 assert(hr->is_survivor(), "badly formed young list"); 2199 // There is a convention that all the young regions in the CSet 2200 // are tagged as "eden", so we do this for the survivors here. We 2201 // use the special set_eden_pre_gc() as it doesn't check that the 2202 // region is free (which is not the case here). 2203 hr->set_eden_pre_gc(); 2204 hr = hr->get_next_young_region(); 2205 } 2206 2207 // Clear the fields that point to the survivor list - they are all young now. 2208 young_list->clear_survivors(); 2209 2210 _collection_set = _inc_cset_head; 2211 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2212 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2213 2214 log_trace(gc, ergo, cset)("Add young regions to CSet. eden: %u regions, survivors: %u regions, predicted young region time: %1.2fms, target pause time: %1.2fms", 2215 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms); 2216 2217 // The number of recorded young regions is the incremental 2218 // collection set's current size 2219 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2220 2221 double young_end_time_sec = os::elapsedTime(); 2222 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2223 2224 return time_remaining_ms; 2225 } 2226 2227 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2228 double non_young_start_time_sec = os::elapsedTime(); 2229 double predicted_old_time_ms = 0.0; 2230 2231 2232 if (!collector_state()->gcs_are_young()) { 2233 cset_chooser()->verify(); 2234 const uint min_old_cset_length = calc_min_old_cset_length(); 2235 const uint max_old_cset_length = calc_max_old_cset_length(); 2236 2237 uint expensive_region_num = 0; 2238 bool check_time_remaining = adaptive_young_list_length(); 2239 2240 HeapRegion* hr = cset_chooser()->peek(); 2241 while (hr != NULL) { 2242 if (old_cset_region_length() >= max_old_cset_length) { 2243 // Added maximum number of old regions to the CSet. 2244 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions", 2245 old_cset_region_length(), max_old_cset_length); 2246 break; 2247 } 2248 2249 2250 // Stop adding regions if the remaining reclaimable space is 2251 // not above G1HeapWastePercent. 2252 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2253 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2254 double threshold = (double) G1HeapWastePercent; 2255 if (reclaimable_perc <= threshold) { 2256 // We've added enough old regions that the amount of uncollected 2257 // reclaimable space is at or below the waste threshold. Stop 2258 // adding old regions to the CSet. 2259 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). " 2260 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%", 2261 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2262 break; 2263 } 2264 2265 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2266 if (check_time_remaining) { 2267 if (predicted_time_ms > time_remaining_ms) { 2268 // Too expensive for the current CSet. 2269 2270 if (old_cset_region_length() >= min_old_cset_length) { 2271 // We have added the minimum number of old regions to the CSet, 2272 // we are done with this CSet. 2273 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). " 2274 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions", 2275 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length); 2276 break; 2277 } 2278 2279 // We'll add it anyway given that we haven't reached the 2280 // minimum number of old regions. 2281 expensive_region_num += 1; 2282 } 2283 } else { 2284 if (old_cset_region_length() >= min_old_cset_length) { 2285 // In the non-auto-tuning case, we'll finish adding regions 2286 // to the CSet if we reach the minimum. 2287 2288 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions", 2289 old_cset_region_length(), min_old_cset_length); 2290 break; 2291 } 2292 } 2293 2294 // We will add this region to the CSet. 2295 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2296 predicted_old_time_ms += predicted_time_ms; 2297 cset_chooser()->pop(); // already have region via peek() 2298 _g1->old_set_remove(hr); 2299 add_old_region_to_cset(hr); 2300 2301 hr = cset_chooser()->peek(); 2302 } 2303 if (hr == NULL) { 2304 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)"); 2305 } 2306 2307 if (expensive_region_num > 0) { 2308 // We print the information once here at the end, predicated on 2309 // whether we added any apparently expensive regions or not, to 2310 // avoid generating output per region. 2311 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)." 2312 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms", 2313 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms); 2314 } 2315 2316 cset_chooser()->verify(); 2317 } 2318 2319 stop_incremental_cset_building(); 2320 2321 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f", 2322 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms); 2323 2324 double non_young_end_time_sec = os::elapsedTime(); 2325 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2326 } 2327 2328 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2329 if(TraceYoungGenTime) { 2330 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2331 } 2332 } 2333 2334 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2335 if(TraceYoungGenTime) { 2336 _all_yield_times_ms.add(yield_time_ms); 2337 } 2338 } 2339 2340 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2341 if(TraceYoungGenTime) { 2342 _total.add(pause_time_ms); 2343 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2344 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2345 _parallel.add(phase_times->cur_collection_par_time_ms()); 2346 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2347 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2348 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2349 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2350 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2351 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2352 2353 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2354 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2355 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2356 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2357 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2358 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2359 2360 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2361 _parallel_other.add(parallel_other_time); 2362 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2363 } 2364 } 2365 2366 void TraceYoungGenTimeData::increment_young_collection_count() { 2367 if(TraceYoungGenTime) { 2368 ++_young_pause_num; 2369 } 2370 } 2371 2372 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2373 if(TraceYoungGenTime) { 2374 ++_mixed_pause_num; 2375 } 2376 } 2377 2378 void TraceYoungGenTimeData::print_summary(const char* str, 2379 const NumberSeq* seq) const { 2380 double sum = seq->sum(); 2381 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2382 str, sum / 1000.0, seq->avg()); 2383 } 2384 2385 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2386 const NumberSeq* seq) const { 2387 print_summary(str, seq); 2388 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2389 "(num", seq->num(), seq->sd(), seq->maximum()); 2390 } 2391 2392 void TraceYoungGenTimeData::print() const { 2393 if (!TraceYoungGenTime) { 2394 return; 2395 } 2396 2397 tty->print_cr("ALL PAUSES"); 2398 print_summary_sd(" Total", &_total); 2399 tty->cr(); 2400 tty->cr(); 2401 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2402 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2403 tty->cr(); 2404 2405 tty->print_cr("EVACUATION PAUSES"); 2406 2407 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2408 tty->print_cr("none"); 2409 } else { 2410 print_summary_sd(" Evacuation Pauses", &_total); 2411 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2412 print_summary(" Parallel Time", &_parallel); 2413 print_summary(" Ext Root Scanning", &_ext_root_scan); 2414 print_summary(" SATB Filtering", &_satb_filtering); 2415 print_summary(" Update RS", &_update_rs); 2416 print_summary(" Scan RS", &_scan_rs); 2417 print_summary(" Object Copy", &_obj_copy); 2418 print_summary(" Termination", &_termination); 2419 print_summary(" Parallel Other", &_parallel_other); 2420 print_summary(" Clear CT", &_clear_ct); 2421 print_summary(" Other", &_other); 2422 } 2423 tty->cr(); 2424 2425 tty->print_cr("MISC"); 2426 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2427 print_summary_sd(" Yields", &_all_yield_times_ms); 2428 } 2429 2430 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2431 if (TraceOldGenTime) { 2432 _all_full_gc_times.add(full_gc_time_ms); 2433 } 2434 } 2435 2436 void TraceOldGenTimeData::print() const { 2437 if (!TraceOldGenTime) { 2438 return; 2439 } 2440 2441 if (_all_full_gc_times.num() > 0) { 2442 tty->print("\n%4d full_gcs: total time = %8.2f s", 2443 _all_full_gc_times.num(), 2444 _all_full_gc_times.sum() / 1000.0); 2445 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2446 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2447 _all_full_gc_times.sd(), 2448 _all_full_gc_times.maximum()); 2449 } 2450 }