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 size_t G1CollectorPolicy::get_new_size_prediction(TruncatedSeq const* seq) const { 295 return (size_t)get_new_prediction(seq); 296 } 297 298 void G1CollectorPolicy::initialize_alignments() { 299 _space_alignment = HeapRegion::GrainBytes; 300 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 301 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 302 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 303 } 304 305 void G1CollectorPolicy::initialize_flags() { 306 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 307 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 308 } 309 310 if (SurvivorRatio < 1) { 311 vm_exit_during_initialization("Invalid survivor ratio specified"); 312 } 313 CollectorPolicy::initialize_flags(); 314 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 315 } 316 317 void G1CollectorPolicy::post_heap_initialize() { 318 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 319 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 320 if (max_young_size != MaxNewSize) { 321 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 322 } 323 324 _ihop_control = create_ihop_control(); 325 } 326 327 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 328 329 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 330 _min_desired_young_length(0), _max_desired_young_length(0) { 331 if (FLAG_IS_CMDLINE(NewRatio)) { 332 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 333 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 334 } else { 335 _sizer_kind = SizerNewRatio; 336 _adaptive_size = false; 337 return; 338 } 339 } 340 341 if (NewSize > MaxNewSize) { 342 if (FLAG_IS_CMDLINE(MaxNewSize)) { 343 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 344 "A new max generation size of " SIZE_FORMAT "k will be used.", 345 NewSize/K, MaxNewSize/K, NewSize/K); 346 } 347 MaxNewSize = NewSize; 348 } 349 350 if (FLAG_IS_CMDLINE(NewSize)) { 351 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 352 1U); 353 if (FLAG_IS_CMDLINE(MaxNewSize)) { 354 _max_desired_young_length = 355 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 356 1U); 357 _sizer_kind = SizerMaxAndNewSize; 358 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 359 } else { 360 _sizer_kind = SizerNewSizeOnly; 361 } 362 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 363 _max_desired_young_length = 364 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 365 1U); 366 _sizer_kind = SizerMaxNewSizeOnly; 367 } 368 } 369 370 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 371 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 372 return MAX2(1U, default_value); 373 } 374 375 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 376 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 377 return MAX2(1U, default_value); 378 } 379 380 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 381 assert(number_of_heap_regions > 0, "Heap must be initialized"); 382 383 switch (_sizer_kind) { 384 case SizerDefaults: 385 *min_young_length = calculate_default_min_length(number_of_heap_regions); 386 *max_young_length = calculate_default_max_length(number_of_heap_regions); 387 break; 388 case SizerNewSizeOnly: 389 *max_young_length = calculate_default_max_length(number_of_heap_regions); 390 *max_young_length = MAX2(*min_young_length, *max_young_length); 391 break; 392 case SizerMaxNewSizeOnly: 393 *min_young_length = calculate_default_min_length(number_of_heap_regions); 394 *min_young_length = MIN2(*min_young_length, *max_young_length); 395 break; 396 case SizerMaxAndNewSize: 397 // Do nothing. Values set on the command line, don't update them at runtime. 398 break; 399 case SizerNewRatio: 400 *min_young_length = number_of_heap_regions / (NewRatio + 1); 401 *max_young_length = *min_young_length; 402 break; 403 default: 404 ShouldNotReachHere(); 405 } 406 407 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 408 } 409 410 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 411 // We need to pass the desired values because recalculation may not update these 412 // values in some cases. 413 uint temp = _min_desired_young_length; 414 uint result = _max_desired_young_length; 415 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 416 return result; 417 } 418 419 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 420 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 421 &_max_desired_young_length); 422 } 423 424 void G1CollectorPolicy::init() { 425 // Set aside an initial future to_space. 426 _g1 = G1CollectedHeap::heap(); 427 428 assert(Heap_lock->owned_by_self(), "Locking discipline."); 429 430 initialize_gc_policy_counters(); 431 432 if (adaptive_young_list_length()) { 433 _young_list_fixed_length = 0; 434 } else { 435 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 436 } 437 _free_regions_at_end_of_collection = _g1->num_free_regions(); 438 439 update_young_list_max_and_target_length(); 440 // We may immediately start allocating regions and placing them on the 441 // collection set list. Initialize the per-collection set info 442 start_incremental_cset_building(); 443 } 444 445 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 446 phase_times()->note_gc_start(num_active_workers); 447 } 448 449 // Create the jstat counters for the policy. 450 void G1CollectorPolicy::initialize_gc_policy_counters() { 451 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 452 } 453 454 bool G1CollectorPolicy::predict_will_fit(uint young_length, 455 double base_time_ms, 456 uint base_free_regions, 457 double target_pause_time_ms) const { 458 if (young_length >= base_free_regions) { 459 // end condition 1: not enough space for the young regions 460 return false; 461 } 462 463 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 464 size_t bytes_to_copy = 465 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 466 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 467 double young_other_time_ms = predict_young_other_time_ms(young_length); 468 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 469 if (pause_time_ms > target_pause_time_ms) { 470 // end condition 2: prediction is over the target pause time 471 return false; 472 } 473 474 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 475 476 // When copying, we will likely need more bytes free than is live in the region. 477 // Add some safety margin to factor in the confidence of our guess, and the 478 // natural expected waste. 479 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 480 // of the calculation: the lower the confidence, the more headroom. 481 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 482 // copying due to anticipated waste in the PLABs. 483 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 484 size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy); 485 486 if (expected_bytes_to_copy > free_bytes) { 487 // end condition 3: out-of-space 488 return false; 489 } 490 491 // success! 492 return true; 493 } 494 495 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 496 // re-calculate the necessary reserve 497 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 498 // We use ceiling so that if reserve_regions_d is > 0.0 (but 499 // smaller than 1.0) we'll get 1. 500 _reserve_regions = (uint) ceil(reserve_regions_d); 501 502 _young_gen_sizer->heap_size_changed(new_number_of_regions); 503 } 504 505 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 506 uint base_min_length) const { 507 uint desired_min_length = 0; 508 if (adaptive_young_list_length()) { 509 if (_alloc_rate_ms_seq->num() > 3) { 510 double now_sec = os::elapsedTime(); 511 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 512 double alloc_rate_ms = predict_alloc_rate_ms(); 513 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 514 } else { 515 // otherwise we don't have enough info to make the prediction 516 } 517 } 518 desired_min_length += base_min_length; 519 // make sure we don't go below any user-defined minimum bound 520 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 521 } 522 523 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 524 // Here, we might want to also take into account any additional 525 // constraints (i.e., user-defined minimum bound). Currently, we 526 // effectively don't set this bound. 527 return _young_gen_sizer->max_desired_young_length(); 528 } 529 530 uint G1CollectorPolicy::update_young_list_max_and_target_length() { 531 return update_young_list_max_and_target_length(get_new_size_prediction(_rs_lengths_seq)); 532 } 533 534 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 535 uint unbounded_target_length = update_young_list_target_length(rs_lengths); 536 update_max_gc_locker_expansion(); 537 return unbounded_target_length; 538 } 539 540 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 541 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths); 542 _young_list_target_length = young_lengths.first; 543 return young_lengths.second; 544 } 545 546 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const { 547 YoungTargetLengths result; 548 549 // Calculate the absolute and desired min bounds first. 550 551 // This is how many young regions we already have (currently: the survivors). 552 uint base_min_length = recorded_survivor_regions(); 553 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 554 // This is the absolute minimum young length. Ensure that we 555 // will at least have one eden region available for allocation. 556 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 557 // If we shrank the young list target it should not shrink below the current size. 558 desired_min_length = MAX2(desired_min_length, absolute_min_length); 559 // Calculate the absolute and desired max bounds. 560 561 uint desired_max_length = calculate_young_list_desired_max_length(); 562 563 uint young_list_target_length = 0; 564 if (adaptive_young_list_length()) { 565 if (collector_state()->gcs_are_young()) { 566 young_list_target_length = 567 calculate_young_list_target_length(rs_lengths, 568 base_min_length, 569 desired_min_length, 570 desired_max_length); 571 } else { 572 // Don't calculate anything and let the code below bound it to 573 // the desired_min_length, i.e., do the next GC as soon as 574 // possible to maximize how many old regions we can add to it. 575 } 576 } else { 577 // The user asked for a fixed young gen so we'll fix the young gen 578 // whether the next GC is young or mixed. 579 young_list_target_length = _young_list_fixed_length; 580 } 581 582 result.second = young_list_target_length; 583 584 // We will try our best not to "eat" into the reserve. 585 uint absolute_max_length = 0; 586 if (_free_regions_at_end_of_collection > _reserve_regions) { 587 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 588 } 589 if (desired_max_length > absolute_max_length) { 590 desired_max_length = absolute_max_length; 591 } 592 593 // Make sure we don't go over the desired max length, nor under the 594 // desired min length. In case they clash, desired_min_length wins 595 // which is why that test is second. 596 if (young_list_target_length > desired_max_length) { 597 young_list_target_length = desired_max_length; 598 } 599 if (young_list_target_length < desired_min_length) { 600 young_list_target_length = desired_min_length; 601 } 602 603 assert(young_list_target_length > recorded_survivor_regions(), 604 "we should be able to allocate at least one eden region"); 605 assert(young_list_target_length >= absolute_min_length, "post-condition"); 606 607 result.first = young_list_target_length; 608 return result; 609 } 610 611 uint 612 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 613 uint base_min_length, 614 uint desired_min_length, 615 uint desired_max_length) const { 616 assert(adaptive_young_list_length(), "pre-condition"); 617 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 618 619 // In case some edge-condition makes the desired max length too small... 620 if (desired_max_length <= desired_min_length) { 621 return desired_min_length; 622 } 623 624 // We'll adjust min_young_length and max_young_length not to include 625 // the already allocated young regions (i.e., so they reflect the 626 // min and max eden regions we'll allocate). The base_min_length 627 // will be reflected in the predictions by the 628 // survivor_regions_evac_time prediction. 629 assert(desired_min_length > base_min_length, "invariant"); 630 uint min_young_length = desired_min_length - base_min_length; 631 assert(desired_max_length > base_min_length, "invariant"); 632 uint max_young_length = desired_max_length - base_min_length; 633 634 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 635 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 636 size_t pending_cards = get_new_size_prediction(_pending_cards_seq); 637 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 638 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 639 double base_time_ms = 640 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 641 survivor_regions_evac_time; 642 uint available_free_regions = _free_regions_at_end_of_collection; 643 uint base_free_regions = 0; 644 if (available_free_regions > _reserve_regions) { 645 base_free_regions = available_free_regions - _reserve_regions; 646 } 647 648 // Here, we will make sure that the shortest young length that 649 // makes sense fits within the target pause time. 650 651 if (predict_will_fit(min_young_length, base_time_ms, 652 base_free_regions, target_pause_time_ms)) { 653 // The shortest young length will fit into the target pause time; 654 // we'll now check whether the absolute maximum number of young 655 // regions will fit in the target pause time. If not, we'll do 656 // a binary search between min_young_length and max_young_length. 657 if (predict_will_fit(max_young_length, base_time_ms, 658 base_free_regions, target_pause_time_ms)) { 659 // The maximum young length will fit into the target pause time. 660 // We are done so set min young length to the maximum length (as 661 // the result is assumed to be returned in min_young_length). 662 min_young_length = max_young_length; 663 } else { 664 // The maximum possible number of young regions will not fit within 665 // the target pause time so we'll search for the optimal 666 // length. The loop invariants are: 667 // 668 // min_young_length < max_young_length 669 // min_young_length is known to fit into the target pause time 670 // max_young_length is known not to fit into the target pause time 671 // 672 // Going into the loop we know the above hold as we've just 673 // checked them. Every time around the loop we check whether 674 // the middle value between min_young_length and 675 // max_young_length fits into the target pause time. If it 676 // does, it becomes the new min. If it doesn't, it becomes 677 // the new max. This way we maintain the loop invariants. 678 679 assert(min_young_length < max_young_length, "invariant"); 680 uint diff = (max_young_length - min_young_length) / 2; 681 while (diff > 0) { 682 uint young_length = min_young_length + diff; 683 if (predict_will_fit(young_length, base_time_ms, 684 base_free_regions, target_pause_time_ms)) { 685 min_young_length = young_length; 686 } else { 687 max_young_length = young_length; 688 } 689 assert(min_young_length < max_young_length, "invariant"); 690 diff = (max_young_length - min_young_length) / 2; 691 } 692 // The results is min_young_length which, according to the 693 // loop invariants, should fit within the target pause time. 694 695 // These are the post-conditions of the binary search above: 696 assert(min_young_length < max_young_length, 697 "otherwise we should have discovered that max_young_length " 698 "fits into the pause target and not done the binary search"); 699 assert(predict_will_fit(min_young_length, base_time_ms, 700 base_free_regions, target_pause_time_ms), 701 "min_young_length, the result of the binary search, should " 702 "fit into the pause target"); 703 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 704 base_free_regions, target_pause_time_ms), 705 "min_young_length, the result of the binary search, should be " 706 "optimal, so no larger length should fit into the pause target"); 707 } 708 } else { 709 // Even the minimum length doesn't fit into the pause time 710 // target, return it as the result nevertheless. 711 } 712 return base_min_length + min_young_length; 713 } 714 715 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 716 double survivor_regions_evac_time = 0.0; 717 for (HeapRegion * r = _recorded_survivor_head; 718 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 719 r = r->get_next_young_region()) { 720 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 721 } 722 return survivor_regions_evac_time; 723 } 724 725 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 726 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 727 728 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 729 if (rs_lengths > _rs_lengths_prediction) { 730 // add 10% to avoid having to recalculate often 731 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 732 update_rs_lengths_prediction(rs_lengths_prediction); 733 734 update_young_list_max_and_target_length(rs_lengths_prediction); 735 } 736 } 737 738 void G1CollectorPolicy::update_rs_lengths_prediction() { 739 update_rs_lengths_prediction(get_new_size_prediction(_rs_lengths_seq)); 740 } 741 742 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 743 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 744 _rs_lengths_prediction = prediction; 745 } 746 } 747 748 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 749 bool is_tlab, 750 bool* gc_overhead_limit_was_exceeded) { 751 guarantee(false, "Not using this policy feature yet."); 752 return NULL; 753 } 754 755 // This method controls how a collector handles one or more 756 // of its generations being fully allocated. 757 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 758 bool is_tlab) { 759 guarantee(false, "Not using this policy feature yet."); 760 return NULL; 761 } 762 763 764 #ifndef PRODUCT 765 bool G1CollectorPolicy::verify_young_ages() { 766 HeapRegion* head = _g1->young_list()->first_region(); 767 return 768 verify_young_ages(head, _short_lived_surv_rate_group); 769 // also call verify_young_ages on any additional surv rate groups 770 } 771 772 bool 773 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 774 SurvRateGroup *surv_rate_group) { 775 guarantee( surv_rate_group != NULL, "pre-condition" ); 776 777 const char* name = surv_rate_group->name(); 778 bool ret = true; 779 int prev_age = -1; 780 781 for (HeapRegion* curr = head; 782 curr != NULL; 783 curr = curr->get_next_young_region()) { 784 SurvRateGroup* group = curr->surv_rate_group(); 785 if (group == NULL && !curr->is_survivor()) { 786 log_info(gc, verify)("## %s: encountered NULL surv_rate_group", name); 787 ret = false; 788 } 789 790 if (surv_rate_group == group) { 791 int age = curr->age_in_surv_rate_group(); 792 793 if (age < 0) { 794 log_info(gc, verify)("## %s: encountered negative age", name); 795 ret = false; 796 } 797 798 if (age <= prev_age) { 799 log_info(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age); 800 ret = false; 801 } 802 prev_age = age; 803 } 804 } 805 806 return ret; 807 } 808 #endif // PRODUCT 809 810 void G1CollectorPolicy::record_full_collection_start() { 811 _full_collection_start_sec = os::elapsedTime(); 812 record_heap_size_info_at_start(true /* full */); 813 // Release the future to-space so that it is available for compaction into. 814 collector_state()->set_full_collection(true); 815 } 816 817 void G1CollectorPolicy::record_full_collection_end() { 818 // Consider this like a collection pause for the purposes of allocation 819 // since last pause. 820 double end_sec = os::elapsedTime(); 821 double full_gc_time_sec = end_sec - _full_collection_start_sec; 822 double full_gc_time_ms = full_gc_time_sec * 1000.0; 823 824 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 825 826 update_recent_gc_times(end_sec, full_gc_time_ms); 827 828 collector_state()->set_full_collection(false); 829 830 // "Nuke" the heuristics that control the young/mixed GC 831 // transitions and make sure we start with young GCs after the Full GC. 832 collector_state()->set_gcs_are_young(true); 833 collector_state()->set_last_young_gc(false); 834 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 835 collector_state()->set_during_initial_mark_pause(false); 836 collector_state()->set_in_marking_window(false); 837 collector_state()->set_in_marking_window_im(false); 838 839 _short_lived_surv_rate_group->start_adding_regions(); 840 // also call this on any additional surv rate groups 841 842 record_survivor_regions(0, NULL, NULL); 843 844 _free_regions_at_end_of_collection = _g1->num_free_regions(); 845 // Reset survivors SurvRateGroup. 846 _survivor_surv_rate_group->reset(); 847 update_young_list_max_and_target_length(); 848 update_rs_lengths_prediction(); 849 cset_chooser()->clear(); 850 851 _bytes_allocated_in_old_since_last_gc = 0; 852 853 record_pause(FullGC, _full_collection_start_sec, end_sec); 854 } 855 856 void G1CollectorPolicy::record_stop_world_start() { 857 _stop_world_start = os::elapsedTime(); 858 } 859 860 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 861 // We only need to do this here as the policy will only be applied 862 // to the GC we're about to start. so, no point is calculating this 863 // every time we calculate / recalculate the target young length. 864 update_survivors_policy(); 865 866 assert(_g1->used() == _g1->recalculate_used(), 867 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 868 _g1->used(), _g1->recalculate_used()); 869 870 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 871 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 872 _stop_world_start = 0.0; 873 874 record_heap_size_info_at_start(false /* full */); 875 876 phase_times()->record_cur_collection_start_sec(start_time_sec); 877 _pending_cards = _g1->pending_card_num(); 878 879 _collection_set_bytes_used_before = 0; 880 _bytes_copied_during_gc = 0; 881 882 collector_state()->set_last_gc_was_young(false); 883 884 // do that for any other surv rate groups 885 _short_lived_surv_rate_group->stop_adding_regions(); 886 _survivors_age_table.clear(); 887 888 assert( verify_young_ages(), "region age verification" ); 889 } 890 891 void G1CollectorPolicy::record_concurrent_mark_init_end(double 892 mark_init_elapsed_time_ms) { 893 collector_state()->set_during_marking(true); 894 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 895 collector_state()->set_during_initial_mark_pause(false); 896 } 897 898 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 899 _mark_remark_start_sec = os::elapsedTime(); 900 collector_state()->set_during_marking(false); 901 } 902 903 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 904 double end_time_sec = os::elapsedTime(); 905 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 906 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 907 _prev_collection_pause_end_ms += elapsed_time_ms; 908 909 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 910 } 911 912 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 913 _mark_cleanup_start_sec = os::elapsedTime(); 914 } 915 916 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 917 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 918 "skip last young-only gc"); 919 collector_state()->set_last_young_gc(should_continue_with_reclaim); 920 // We skip the marking phase. 921 if (!should_continue_with_reclaim) { 922 abort_time_to_mixed_tracking(); 923 } 924 collector_state()->set_in_marking_window(false); 925 } 926 927 void G1CollectorPolicy::record_concurrent_pause() { 928 if (_stop_world_start > 0.0) { 929 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 930 _trace_young_gen_time_data.record_yield_time(yield_ms); 931 } 932 } 933 934 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 935 return phase_times()->average_time_ms(phase); 936 } 937 938 double G1CollectorPolicy::young_other_time_ms() const { 939 return phase_times()->young_cset_choice_time_ms() + 940 phase_times()->young_free_cset_time_ms(); 941 } 942 943 double G1CollectorPolicy::non_young_other_time_ms() const { 944 return phase_times()->non_young_cset_choice_time_ms() + 945 phase_times()->non_young_free_cset_time_ms(); 946 947 } 948 949 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 950 return pause_time_ms - 951 average_time_ms(G1GCPhaseTimes::UpdateRS) - 952 average_time_ms(G1GCPhaseTimes::ScanRS) - 953 average_time_ms(G1GCPhaseTimes::ObjCopy) - 954 average_time_ms(G1GCPhaseTimes::Termination); 955 } 956 957 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 958 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 959 } 960 961 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 962 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 963 } 964 965 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 966 if (about_to_start_mixed_phase()) { 967 return false; 968 } 969 970 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 971 972 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 973 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 974 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 975 976 bool result = false; 977 if (marking_request_bytes > marking_initiating_used_threshold) { 978 result = collector_state()->gcs_are_young() && !collector_state()->last_young_gc(); 979 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s", 980 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)", 981 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1->capacity() * 100, source); 982 } 983 984 return result; 985 } 986 987 // Anything below that is considered to be zero 988 #define MIN_TIMER_GRANULARITY 0.0000001 989 990 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 991 double end_time_sec = os::elapsedTime(); 992 993 size_t cur_used_bytes = _g1->used(); 994 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 995 bool last_pause_included_initial_mark = false; 996 bool update_stats = !_g1->evacuation_failed(); 997 998 NOT_PRODUCT(_short_lived_surv_rate_group->print()); 999 1000 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec); 1001 1002 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1003 if (last_pause_included_initial_mark) { 1004 record_concurrent_mark_init_end(0.0); 1005 } else { 1006 maybe_start_marking(); 1007 } 1008 1009 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1010 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1011 // This usually happens due to the timer not having the required 1012 // granularity. Some Linuxes are the usual culprits. 1013 // We'll just set it to something (arbitrarily) small. 1014 app_time_ms = 1.0; 1015 } 1016 1017 if (update_stats) { 1018 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1019 // We maintain the invariant that all objects allocated by mutator 1020 // threads will be allocated out of eden regions. So, we can use 1021 // the eden region number allocated since the previous GC to 1022 // calculate the application's allocate rate. The only exception 1023 // to that is humongous objects that are allocated separately. But 1024 // given that humongous object allocations do not really affect 1025 // either the pause's duration nor when the next pause will take 1026 // place we can safely ignore them here. 1027 uint regions_allocated = eden_cset_region_length(); 1028 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1029 _alloc_rate_ms_seq->add(alloc_rate_ms); 1030 1031 double interval_ms = 1032 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1033 update_recent_gc_times(end_time_sec, pause_time_ms); 1034 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1035 if (recent_avg_pause_time_ratio() < 0.0 || 1036 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1037 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1038 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1039 if (_recent_avg_pause_time_ratio < 0.0) { 1040 _recent_avg_pause_time_ratio = 0.0; 1041 } else { 1042 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1043 _recent_avg_pause_time_ratio = 1.0; 1044 } 1045 } 1046 1047 // Compute the ratio of just this last pause time to the entire time range stored 1048 // in the vectors. Comparing this pause to the entire range, rather than only the 1049 // most recent interval, has the effect of smoothing over a possible transient 'burst' 1050 // of more frequent pauses that don't really reflect a change in heap occupancy. 1051 // This reduces the likelihood of a needless heap expansion being triggered. 1052 _last_pause_time_ratio = 1053 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms; 1054 } 1055 1056 bool new_in_marking_window = collector_state()->in_marking_window(); 1057 bool new_in_marking_window_im = false; 1058 if (last_pause_included_initial_mark) { 1059 new_in_marking_window = true; 1060 new_in_marking_window_im = true; 1061 } 1062 1063 if (collector_state()->last_young_gc()) { 1064 // This is supposed to to be the "last young GC" before we start 1065 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1066 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1067 1068 if (next_gc_should_be_mixed("start mixed GCs", 1069 "do not start mixed GCs")) { 1070 collector_state()->set_gcs_are_young(false); 1071 } else { 1072 // We aborted the mixed GC phase early. 1073 abort_time_to_mixed_tracking(); 1074 } 1075 1076 collector_state()->set_last_young_gc(false); 1077 } 1078 1079 if (!collector_state()->last_gc_was_young()) { 1080 // This is a mixed GC. Here we decide whether to continue doing 1081 // mixed GCs or not. 1082 if (!next_gc_should_be_mixed("continue mixed GCs", 1083 "do not continue mixed GCs")) { 1084 collector_state()->set_gcs_are_young(true); 1085 1086 maybe_start_marking(); 1087 } 1088 } 1089 1090 _short_lived_surv_rate_group->start_adding_regions(); 1091 // Do that for any other surv rate groups 1092 1093 if (update_stats) { 1094 double cost_per_card_ms = 0.0; 1095 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1096 if (_pending_cards > 0) { 1097 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1098 _cost_per_card_ms_seq->add(cost_per_card_ms); 1099 } 1100 _cost_scan_hcc_seq->add(cost_scan_hcc); 1101 1102 double cost_per_entry_ms = 0.0; 1103 if (cards_scanned > 10) { 1104 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1105 if (collector_state()->last_gc_was_young()) { 1106 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1107 } else { 1108 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1109 } 1110 } 1111 1112 if (_max_rs_lengths > 0) { 1113 double cards_per_entry_ratio = 1114 (double) cards_scanned / (double) _max_rs_lengths; 1115 if (collector_state()->last_gc_was_young()) { 1116 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1117 } else { 1118 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1119 } 1120 } 1121 1122 // This is defensive. For a while _max_rs_lengths could get 1123 // smaller than _recorded_rs_lengths which was causing 1124 // rs_length_diff to get very large and mess up the RSet length 1125 // predictions. The reason was unsafe concurrent updates to the 1126 // _inc_cset_recorded_rs_lengths field which the code below guards 1127 // against (see CR 7118202). This bug has now been fixed (see CR 1128 // 7119027). However, I'm still worried that 1129 // _inc_cset_recorded_rs_lengths might still end up somewhat 1130 // inaccurate. The concurrent refinement thread calculates an 1131 // RSet's length concurrently with other CR threads updating it 1132 // which might cause it to calculate the length incorrectly (if, 1133 // say, it's in mid-coarsening). So I'll leave in the defensive 1134 // conditional below just in case. 1135 size_t rs_length_diff = 0; 1136 if (_max_rs_lengths > _recorded_rs_lengths) { 1137 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1138 } 1139 _rs_length_diff_seq->add((double) rs_length_diff); 1140 1141 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1142 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1143 double cost_per_byte_ms = 0.0; 1144 1145 if (copied_bytes > 0) { 1146 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1147 if (collector_state()->in_marking_window()) { 1148 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1149 } else { 1150 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1151 } 1152 } 1153 1154 if (young_cset_region_length() > 0) { 1155 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1156 young_cset_region_length()); 1157 } 1158 1159 if (old_cset_region_length() > 0) { 1160 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1161 old_cset_region_length()); 1162 } 1163 1164 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1165 1166 _pending_cards_seq->add((double) _pending_cards); 1167 _rs_lengths_seq->add((double) _max_rs_lengths); 1168 } 1169 1170 collector_state()->set_in_marking_window(new_in_marking_window); 1171 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1172 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1173 // IHOP control wants to know the expected young gen length if it were not 1174 // restrained by the heap reserve. Using the actual length would make the 1175 // prediction too small and the limit the young gen every time we get to the 1176 // predicted target occupancy. 1177 size_t last_unrestrained_young_length = update_young_list_max_and_target_length(); 1178 update_rs_lengths_prediction(); 1179 1180 update_ihop_prediction(app_time_ms / 1000.0, 1181 _bytes_allocated_in_old_since_last_gc, 1182 last_unrestrained_young_length * HeapRegion::GrainBytes); 1183 _bytes_allocated_in_old_since_last_gc = 0; 1184 1185 _ihop_control->send_trace_event(_g1->gc_tracer_stw()); 1186 1187 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1188 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1189 1190 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1191 1192 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1193 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 1194 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms", 1195 update_rs_time_goal_ms, scan_hcc_time_ms); 1196 1197 update_rs_time_goal_ms = 0; 1198 } else { 1199 update_rs_time_goal_ms -= scan_hcc_time_ms; 1200 } 1201 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1202 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1203 update_rs_time_goal_ms); 1204 1205 cset_chooser()->verify(); 1206 } 1207 1208 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1209 if (G1UseAdaptiveIHOP) { 1210 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1211 G1CollectedHeap::heap()->max_capacity(), 1212 &_predictor, 1213 G1ReservePercent, 1214 G1HeapWastePercent); 1215 } else { 1216 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1217 G1CollectedHeap::heap()->max_capacity()); 1218 } 1219 } 1220 1221 void G1CollectorPolicy::update_ihop_prediction(double mutator_time_s, 1222 size_t mutator_alloc_bytes, 1223 size_t young_gen_size) { 1224 // Always try to update IHOP prediction. Even evacuation failures give information 1225 // about e.g. whether to start IHOP earlier next time. 1226 1227 // Avoid using really small application times that might create samples with 1228 // very high or very low values. They may be caused by e.g. back-to-back gcs. 1229 double const min_valid_time = 1e-6; 1230 1231 bool report = false; 1232 1233 double marking_to_mixed_time = -1.0; 1234 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1235 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1236 assert(marking_to_mixed_time > 0.0, 1237 "Initial mark to mixed time must be larger than zero but is %.3f", 1238 marking_to_mixed_time); 1239 if (marking_to_mixed_time > min_valid_time) { 1240 _ihop_control->update_marking_length(marking_to_mixed_time); 1241 report = true; 1242 } 1243 } 1244 1245 // As an approximation for the young gc promotion rates during marking we use 1246 // all of them. In many applications there are only a few if any young gcs during 1247 // marking, which makes any prediction useless. This increases the accuracy of the 1248 // prediction. 1249 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1250 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1251 report = true; 1252 } 1253 1254 if (report) { 1255 report_ihop_statistics(); 1256 } 1257 } 1258 1259 void G1CollectorPolicy::report_ihop_statistics() { 1260 _ihop_control->print(); 1261 } 1262 1263 #define EXT_SIZE_FORMAT "%.1f%s" 1264 #define EXT_SIZE_PARAMS(bytes) \ 1265 byte_size_in_proper_unit((double)(bytes)), \ 1266 proper_unit_for_byte_size((bytes)) 1267 1268 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1269 YoungList* young_list = _g1->young_list(); 1270 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1271 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1272 _heap_capacity_bytes_before_gc = _g1->capacity(); 1273 _old_used_bytes_before_gc = _g1->old_regions_count() * HeapRegion::GrainBytes; 1274 _humongous_used_bytes_before_gc = _g1->humongous_regions_count() * HeapRegion::GrainBytes; 1275 _heap_used_bytes_before_gc = _g1->used(); 1276 _eden_capacity_bytes_before_gc = (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1277 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1278 } 1279 1280 void G1CollectorPolicy::print_detailed_heap_transition() const { 1281 YoungList* young_list = _g1->young_list(); 1282 1283 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1284 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1285 size_t heap_used_bytes_after_gc = _g1->used(); 1286 size_t old_used_bytes_after_gc = _g1->old_regions_count() * HeapRegion::GrainBytes; 1287 size_t humongous_used_bytes_after_gc = _g1->humongous_regions_count() * HeapRegion::GrainBytes; 1288 1289 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1290 size_t eden_capacity_bytes_after_gc = 1291 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1292 size_t survivor_capacity_bytes_after_gc = _max_survivor_regions * HeapRegion::GrainBytes; 1293 1294 log_info(gc, heap)("Eden: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1295 _eden_used_bytes_before_gc / K, eden_used_bytes_after_gc /K, eden_capacity_bytes_after_gc /K); 1296 log_info(gc, heap)("Survivor: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1297 _survivor_used_bytes_before_gc / K, survivor_used_bytes_after_gc /K, survivor_capacity_bytes_after_gc /K); 1298 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K", 1299 _old_used_bytes_before_gc / K, old_used_bytes_after_gc /K); 1300 log_info(gc, heap)("Humongous: " SIZE_FORMAT "K->" SIZE_FORMAT "K", 1301 _humongous_used_bytes_before_gc / K, humongous_used_bytes_after_gc /K); 1302 1303 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1304 } 1305 1306 void G1CollectorPolicy::print_phases() { 1307 phase_times()->print(); 1308 } 1309 1310 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1311 double update_rs_processed_buffers, 1312 double goal_ms) { 1313 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1314 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1315 1316 if (G1UseAdaptiveConcRefinement) { 1317 const int k_gy = 3, k_gr = 6; 1318 const double inc_k = 1.1, dec_k = 0.9; 1319 1320 int g = cg1r->green_zone(); 1321 if (update_rs_time > goal_ms) { 1322 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1323 } else { 1324 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1325 g = (int)MAX2(g * inc_k, g + 1.0); 1326 } 1327 } 1328 // Change the refinement threads params 1329 cg1r->set_green_zone(g); 1330 cg1r->set_yellow_zone(g * k_gy); 1331 cg1r->set_red_zone(g * k_gr); 1332 cg1r->reinitialize_threads(); 1333 1334 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1335 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1336 cg1r->yellow_zone()); 1337 // Change the barrier params 1338 dcqs.set_process_completed_threshold(processing_threshold); 1339 dcqs.set_max_completed_queue(cg1r->red_zone()); 1340 } 1341 1342 int curr_queue_size = dcqs.completed_buffers_num(); 1343 if (curr_queue_size >= cg1r->yellow_zone()) { 1344 dcqs.set_completed_queue_padding(curr_queue_size); 1345 } else { 1346 dcqs.set_completed_queue_padding(0); 1347 } 1348 dcqs.notify_if_necessary(); 1349 } 1350 1351 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1352 return get_new_size_prediction(_rs_length_diff_seq); 1353 } 1354 1355 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1356 return get_new_prediction(_alloc_rate_ms_seq); 1357 } 1358 1359 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1360 return get_new_prediction(_cost_per_card_ms_seq); 1361 } 1362 1363 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1364 return get_new_prediction(_cost_scan_hcc_seq); 1365 } 1366 1367 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1368 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1369 } 1370 1371 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1372 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1373 } 1374 1375 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1376 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1377 return predict_young_cards_per_entry_ratio(); 1378 } else { 1379 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1380 } 1381 } 1382 1383 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1384 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1385 } 1386 1387 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1388 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1389 } 1390 1391 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1392 if (collector_state()->gcs_are_young()) { 1393 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1394 } else { 1395 return predict_mixed_rs_scan_time_ms(card_num); 1396 } 1397 } 1398 1399 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1400 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1401 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1402 } else { 1403 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1404 } 1405 } 1406 1407 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1408 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1409 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1410 } else { 1411 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1412 } 1413 } 1414 1415 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1416 if (collector_state()->during_concurrent_mark()) { 1417 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1418 } else { 1419 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1420 } 1421 } 1422 1423 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1424 return get_new_prediction(_constant_other_time_ms_seq); 1425 } 1426 1427 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1428 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1429 } 1430 1431 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1432 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1433 } 1434 1435 double G1CollectorPolicy::predict_remark_time_ms() const { 1436 return get_new_prediction(_concurrent_mark_remark_times_ms); 1437 } 1438 1439 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1440 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1441 } 1442 1443 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1444 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1445 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1446 double pred = get_new_prediction(seq); 1447 if (pred > 1.0) { 1448 pred = 1.0; 1449 } 1450 return pred; 1451 } 1452 1453 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1454 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1455 } 1456 1457 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1458 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1459 } 1460 1461 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1462 size_t scanned_cards) const { 1463 return 1464 predict_rs_update_time_ms(pending_cards) + 1465 predict_rs_scan_time_ms(scanned_cards) + 1466 predict_constant_other_time_ms(); 1467 } 1468 1469 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1470 size_t rs_length = predict_rs_length_diff(); 1471 size_t card_num; 1472 if (collector_state()->gcs_are_young()) { 1473 card_num = predict_young_card_num(rs_length); 1474 } else { 1475 card_num = predict_non_young_card_num(rs_length); 1476 } 1477 return predict_base_elapsed_time_ms(pending_cards, card_num); 1478 } 1479 1480 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1481 size_t bytes_to_copy; 1482 if (hr->is_marked()) 1483 bytes_to_copy = hr->max_live_bytes(); 1484 else { 1485 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1486 int age = hr->age_in_surv_rate_group(); 1487 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1488 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1489 } 1490 return bytes_to_copy; 1491 } 1492 1493 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1494 bool for_young_gc) const { 1495 size_t rs_length = hr->rem_set()->occupied(); 1496 size_t card_num; 1497 1498 // Predicting the number of cards is based on which type of GC 1499 // we're predicting for. 1500 if (for_young_gc) { 1501 card_num = predict_young_card_num(rs_length); 1502 } else { 1503 card_num = predict_non_young_card_num(rs_length); 1504 } 1505 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1506 1507 double region_elapsed_time_ms = 1508 predict_rs_scan_time_ms(card_num) + 1509 predict_object_copy_time_ms(bytes_to_copy); 1510 1511 // The prediction of the "other" time for this region is based 1512 // upon the region type and NOT the GC type. 1513 if (hr->is_young()) { 1514 region_elapsed_time_ms += predict_young_other_time_ms(1); 1515 } else { 1516 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1517 } 1518 return region_elapsed_time_ms; 1519 } 1520 1521 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1522 uint survivor_cset_region_length) { 1523 _eden_cset_region_length = eden_cset_region_length; 1524 _survivor_cset_region_length = survivor_cset_region_length; 1525 _old_cset_region_length = 0; 1526 } 1527 1528 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1529 _recorded_rs_lengths = rs_lengths; 1530 } 1531 1532 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1533 double elapsed_ms) { 1534 _recent_gc_times_ms->add(elapsed_ms); 1535 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1536 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1537 } 1538 1539 void G1CollectorPolicy::clear_ratio_check_data() { 1540 _ratio_over_threshold_count = 0; 1541 _ratio_over_threshold_sum = 0.0; 1542 _pauses_since_start = 0; 1543 } 1544 1545 size_t G1CollectorPolicy::expansion_amount() { 1546 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1547 double last_gc_overhead = _last_pause_time_ratio * 100.0; 1548 double threshold = _gc_overhead_perc; 1549 size_t expand_bytes = 0; 1550 1551 // If the heap is at less than half its maximum size, scale the threshold down, 1552 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand, 1553 // though the scaling code will likely keep the increase small. 1554 if (_g1->capacity() <= _g1->max_capacity() / 2) { 1555 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2); 1556 threshold = MAX2(threshold, 1.0); 1557 } 1558 1559 // If the last GC time ratio is over the threshold, increment the count of 1560 // times it has been exceeded, and add this ratio to the sum of exceeded 1561 // ratios. 1562 if (last_gc_overhead > threshold) { 1563 _ratio_over_threshold_count++; 1564 _ratio_over_threshold_sum += last_gc_overhead; 1565 } 1566 1567 // Check if we've had enough GC time ratio checks that were over the 1568 // threshold to trigger an expansion. We'll also expand if we've 1569 // reached the end of the history buffer and the average of all entries 1570 // is still over the threshold. This indicates a smaller number of GCs were 1571 // long enough to make the average exceed the threshold. 1572 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics; 1573 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) || 1574 (filled_history_buffer && (recent_gc_overhead > threshold))) { 1575 size_t min_expand_bytes = HeapRegion::GrainBytes; 1576 size_t reserved_bytes = _g1->max_capacity(); 1577 size_t committed_bytes = _g1->capacity(); 1578 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1579 size_t expand_bytes_via_pct = 1580 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1581 double scale_factor = 1.0; 1582 1583 // If the current size is less than 1/4 of the Initial heap size, expand 1584 // by half of the delta between the current and Initial sizes. IE, grow 1585 // back quickly. 1586 // 1587 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of 1588 // the available expansion space, whichever is smaller, as the base 1589 // expansion size. Then possibly scale this size according to how much the 1590 // threshold has (on average) been exceeded by. If the delta is small 1591 // (less than the StartScaleDownAt value), scale the size down linearly, but 1592 // not by less than MinScaleDownFactor. If the delta is large (greater than 1593 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor 1594 // times the base size. The scaling will be linear in the range from 1595 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words, 1596 // ScaleUpRange sets the rate of scaling up. 1597 if (committed_bytes < InitialHeapSize / 4) { 1598 expand_bytes = (InitialHeapSize - committed_bytes) / 2; 1599 } else { 1600 double const MinScaleDownFactor = 0.2; 1601 double const MaxScaleUpFactor = 2; 1602 double const StartScaleDownAt = _gc_overhead_perc; 1603 double const StartScaleUpAt = _gc_overhead_perc * 1.5; 1604 double const ScaleUpRange = _gc_overhead_perc * 2.0; 1605 1606 double ratio_delta; 1607 if (filled_history_buffer) { 1608 ratio_delta = recent_gc_overhead - threshold; 1609 } else { 1610 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold; 1611 } 1612 1613 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1614 if (ratio_delta < StartScaleDownAt) { 1615 scale_factor = ratio_delta / StartScaleDownAt; 1616 scale_factor = MAX2(scale_factor, MinScaleDownFactor); 1617 } else if (ratio_delta > StartScaleUpAt) { 1618 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange); 1619 scale_factor = MIN2(scale_factor, MaxScaleUpFactor); 1620 } 1621 } 1622 1623 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) " 1624 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)", 1625 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100); 1626 1627 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor); 1628 1629 // Ensure the expansion size is at least the minimum growth amount 1630 // and at most the remaining uncommitted byte size. 1631 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1632 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1633 1634 clear_ratio_check_data(); 1635 } else { 1636 // An expansion was not triggered. If we've started counting, increment 1637 // the number of checks we've made in the current window. If we've 1638 // reached the end of the window without resizing, clear the counters to 1639 // start again the next time we see a ratio above the threshold. 1640 if (_ratio_over_threshold_count > 0) { 1641 _pauses_since_start++; 1642 if (_pauses_since_start > NumPrevPausesForHeuristics) { 1643 clear_ratio_check_data(); 1644 } 1645 } 1646 } 1647 1648 return expand_bytes; 1649 } 1650 1651 void G1CollectorPolicy::print_tracing_info() const { 1652 _trace_young_gen_time_data.print(); 1653 _trace_old_gen_time_data.print(); 1654 } 1655 1656 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1657 #ifndef PRODUCT 1658 _short_lived_surv_rate_group->print_surv_rate_summary(); 1659 // add this call for any other surv rate groups 1660 #endif // PRODUCT 1661 } 1662 1663 bool G1CollectorPolicy::is_young_list_full() const { 1664 uint young_list_length = _g1->young_list()->length(); 1665 uint young_list_target_length = _young_list_target_length; 1666 return young_list_length >= young_list_target_length; 1667 } 1668 1669 bool G1CollectorPolicy::can_expand_young_list() const { 1670 uint young_list_length = _g1->young_list()->length(); 1671 uint young_list_max_length = _young_list_max_length; 1672 return young_list_length < young_list_max_length; 1673 } 1674 1675 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1676 uint expansion_region_num = 0; 1677 if (GCLockerEdenExpansionPercent > 0) { 1678 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1679 double expansion_region_num_d = perc * (double) _young_list_target_length; 1680 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1681 // less than 1.0) we'll get 1. 1682 expansion_region_num = (uint) ceil(expansion_region_num_d); 1683 } else { 1684 assert(expansion_region_num == 0, "sanity"); 1685 } 1686 _young_list_max_length = _young_list_target_length + expansion_region_num; 1687 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1688 } 1689 1690 // Calculates survivor space parameters. 1691 void G1CollectorPolicy::update_survivors_policy() { 1692 double max_survivor_regions_d = 1693 (double) _young_list_target_length / (double) SurvivorRatio; 1694 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1695 // smaller than 1.0) we'll get 1. 1696 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1697 1698 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1699 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1700 } 1701 1702 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1703 // We actually check whether we are marking here and not if we are in a 1704 // reclamation phase. This means that we will schedule a concurrent mark 1705 // even while we are still in the process of reclaiming memory. 1706 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1707 if (!during_cycle) { 1708 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause)); 1709 collector_state()->set_initiate_conc_mark_if_possible(true); 1710 return true; 1711 } else { 1712 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause)); 1713 return false; 1714 } 1715 } 1716 1717 void G1CollectorPolicy::initiate_conc_mark() { 1718 collector_state()->set_during_initial_mark_pause(true); 1719 collector_state()->set_initiate_conc_mark_if_possible(false); 1720 } 1721 1722 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1723 // We are about to decide on whether this pause will be an 1724 // initial-mark pause. 1725 1726 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1727 // will set it here if we have to. However, it should be cleared by 1728 // the end of the pause (it's only set for the duration of an 1729 // initial-mark pause). 1730 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1731 1732 if (collector_state()->initiate_conc_mark_if_possible()) { 1733 // We had noticed on a previous pause that the heap occupancy has 1734 // gone over the initiating threshold and we should start a 1735 // concurrent marking cycle. So we might initiate one. 1736 1737 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1738 // Initiate a new initial mark if there is no marking or reclamation going on. 1739 initiate_conc_mark(); 1740 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1741 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) { 1742 // Initiate a user requested initial mark. An initial mark must be young only 1743 // GC, so the collector state must be updated to reflect this. 1744 collector_state()->set_gcs_are_young(true); 1745 collector_state()->set_last_young_gc(false); 1746 1747 abort_time_to_mixed_tracking(); 1748 initiate_conc_mark(); 1749 log_debug(gc, ergo)("Initiate concurrent cycle (user requested concurrent cycle)"); 1750 } else { 1751 // The concurrent marking thread is still finishing up the 1752 // previous cycle. If we start one right now the two cycles 1753 // overlap. In particular, the concurrent marking thread might 1754 // be in the process of clearing the next marking bitmap (which 1755 // we will use for the next cycle if we start one). Starting a 1756 // cycle now will be bad given that parts of the marking 1757 // information might get cleared by the marking thread. And we 1758 // cannot wait for the marking thread to finish the cycle as it 1759 // periodically yields while clearing the next marking bitmap 1760 // and, if it's in a yield point, it's waiting for us to 1761 // finish. So, at this point we will not start a cycle and we'll 1762 // let the concurrent marking thread complete the last one. 1763 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1764 } 1765 } 1766 } 1767 1768 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1769 G1CollectedHeap* _g1h; 1770 CSetChooserParUpdater _cset_updater; 1771 1772 public: 1773 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1774 uint chunk_size) : 1775 _g1h(G1CollectedHeap::heap()), 1776 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1777 1778 bool doHeapRegion(HeapRegion* r) { 1779 // Do we have any marking information for this region? 1780 if (r->is_marked()) { 1781 // We will skip any region that's currently used as an old GC 1782 // alloc region (we should not consider those for collection 1783 // before we fill them up). 1784 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1785 _cset_updater.add_region(r); 1786 } 1787 } 1788 return false; 1789 } 1790 }; 1791 1792 class ParKnownGarbageTask: public AbstractGangTask { 1793 CollectionSetChooser* _hrSorted; 1794 uint _chunk_size; 1795 G1CollectedHeap* _g1; 1796 HeapRegionClaimer _hrclaimer; 1797 1798 public: 1799 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1800 AbstractGangTask("ParKnownGarbageTask"), 1801 _hrSorted(hrSorted), _chunk_size(chunk_size), 1802 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1803 1804 void work(uint worker_id) { 1805 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1806 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1807 } 1808 }; 1809 1810 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1811 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1812 const uint overpartition_factor = 4; 1813 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1814 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1815 } 1816 1817 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1818 cset_chooser()->clear(); 1819 1820 WorkGang* workers = _g1->workers(); 1821 uint n_workers = workers->active_workers(); 1822 1823 uint n_regions = _g1->num_regions(); 1824 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1825 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1826 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers); 1827 workers->run_task(&par_known_garbage_task); 1828 1829 cset_chooser()->sort_regions(); 1830 1831 double end_sec = os::elapsedTime(); 1832 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1833 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1834 _prev_collection_pause_end_ms += elapsed_time_ms; 1835 1836 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1837 } 1838 1839 // Add the heap region at the head of the non-incremental collection set 1840 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1841 assert(_inc_cset_build_state == Active, "Precondition"); 1842 assert(hr->is_old(), "the region should be old"); 1843 1844 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1845 _g1->register_old_region_with_cset(hr); 1846 hr->set_next_in_collection_set(_collection_set); 1847 _collection_set = hr; 1848 _collection_set_bytes_used_before += hr->used(); 1849 size_t rs_length = hr->rem_set()->occupied(); 1850 _recorded_rs_lengths += rs_length; 1851 _old_cset_region_length += 1; 1852 } 1853 1854 // Initialize the per-collection-set information 1855 void G1CollectorPolicy::start_incremental_cset_building() { 1856 assert(_inc_cset_build_state == Inactive, "Precondition"); 1857 1858 _inc_cset_head = NULL; 1859 _inc_cset_tail = NULL; 1860 _inc_cset_bytes_used_before = 0; 1861 1862 _inc_cset_max_finger = 0; 1863 _inc_cset_recorded_rs_lengths = 0; 1864 _inc_cset_recorded_rs_lengths_diffs = 0; 1865 _inc_cset_predicted_elapsed_time_ms = 0.0; 1866 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1867 _inc_cset_build_state = Active; 1868 } 1869 1870 void G1CollectorPolicy::finalize_incremental_cset_building() { 1871 assert(_inc_cset_build_state == Active, "Precondition"); 1872 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1873 1874 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1875 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1876 // that adds a new region to the CSet. Further updates by the 1877 // concurrent refinement thread that samples the young RSet lengths 1878 // are accumulated in the *_diffs fields. Here we add the diffs to 1879 // the "main" fields. 1880 1881 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1882 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1883 } else { 1884 // This is defensive. The diff should in theory be always positive 1885 // as RSets can only grow between GCs. However, given that we 1886 // sample their size concurrently with other threads updating them 1887 // it's possible that we might get the wrong size back, which 1888 // could make the calculations somewhat inaccurate. 1889 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1890 if (_inc_cset_recorded_rs_lengths >= diffs) { 1891 _inc_cset_recorded_rs_lengths -= diffs; 1892 } else { 1893 _inc_cset_recorded_rs_lengths = 0; 1894 } 1895 } 1896 _inc_cset_predicted_elapsed_time_ms += 1897 _inc_cset_predicted_elapsed_time_ms_diffs; 1898 1899 _inc_cset_recorded_rs_lengths_diffs = 0; 1900 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1901 } 1902 1903 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1904 // This routine is used when: 1905 // * adding survivor regions to the incremental cset at the end of an 1906 // evacuation pause, 1907 // * adding the current allocation region to the incremental cset 1908 // when it is retired, and 1909 // * updating existing policy information for a region in the 1910 // incremental cset via young list RSet sampling. 1911 // Therefore this routine may be called at a safepoint by the 1912 // VM thread, or in-between safepoints by mutator threads (when 1913 // retiring the current allocation region) or a concurrent 1914 // refine thread (RSet sampling). 1915 1916 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1917 size_t used_bytes = hr->used(); 1918 _inc_cset_recorded_rs_lengths += rs_length; 1919 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1920 _inc_cset_bytes_used_before += used_bytes; 1921 1922 // Cache the values we have added to the aggregated information 1923 // in the heap region in case we have to remove this region from 1924 // the incremental collection set, or it is updated by the 1925 // rset sampling code 1926 hr->set_recorded_rs_length(rs_length); 1927 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1928 } 1929 1930 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1931 size_t new_rs_length) { 1932 // Update the CSet information that is dependent on the new RS length 1933 assert(hr->is_young(), "Precondition"); 1934 assert(!SafepointSynchronize::is_at_safepoint(), 1935 "should not be at a safepoint"); 1936 1937 // We could have updated _inc_cset_recorded_rs_lengths and 1938 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1939 // that atomically, as this code is executed by a concurrent 1940 // refinement thread, potentially concurrently with a mutator thread 1941 // allocating a new region and also updating the same fields. To 1942 // avoid the atomic operations we accumulate these updates on two 1943 // separate fields (*_diffs) and we'll just add them to the "main" 1944 // fields at the start of a GC. 1945 1946 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1947 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1948 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1949 1950 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1951 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1952 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1953 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1954 1955 hr->set_recorded_rs_length(new_rs_length); 1956 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1957 } 1958 1959 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1960 assert(hr->is_young(), "invariant"); 1961 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1962 assert(_inc_cset_build_state == Active, "Precondition"); 1963 1964 // We need to clear and set the cached recorded/cached collection set 1965 // information in the heap region here (before the region gets added 1966 // to the collection set). An individual heap region's cached values 1967 // are calculated, aggregated with the policy collection set info, 1968 // and cached in the heap region here (initially) and (subsequently) 1969 // by the Young List sampling code. 1970 1971 size_t rs_length = hr->rem_set()->occupied(); 1972 add_to_incremental_cset_info(hr, rs_length); 1973 1974 HeapWord* hr_end = hr->end(); 1975 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1976 1977 assert(!hr->in_collection_set(), "invariant"); 1978 _g1->register_young_region_with_cset(hr); 1979 assert(hr->next_in_collection_set() == NULL, "invariant"); 1980 } 1981 1982 // Add the region at the RHS of the incremental cset 1983 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1984 // We should only ever be appending survivors at the end of a pause 1985 assert(hr->is_survivor(), "Logic"); 1986 1987 // Do the 'common' stuff 1988 add_region_to_incremental_cset_common(hr); 1989 1990 // Now add the region at the right hand side 1991 if (_inc_cset_tail == NULL) { 1992 assert(_inc_cset_head == NULL, "invariant"); 1993 _inc_cset_head = hr; 1994 } else { 1995 _inc_cset_tail->set_next_in_collection_set(hr); 1996 } 1997 _inc_cset_tail = hr; 1998 } 1999 2000 // Add the region to the LHS of the incremental cset 2001 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2002 // Survivors should be added to the RHS at the end of a pause 2003 assert(hr->is_eden(), "Logic"); 2004 2005 // Do the 'common' stuff 2006 add_region_to_incremental_cset_common(hr); 2007 2008 // Add the region at the left hand side 2009 hr->set_next_in_collection_set(_inc_cset_head); 2010 if (_inc_cset_head == NULL) { 2011 assert(_inc_cset_tail == NULL, "Invariant"); 2012 _inc_cset_tail = hr; 2013 } 2014 _inc_cset_head = hr; 2015 } 2016 2017 #ifndef PRODUCT 2018 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2019 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2020 2021 st->print_cr("\nCollection_set:"); 2022 HeapRegion* csr = list_head; 2023 while (csr != NULL) { 2024 HeapRegion* next = csr->next_in_collection_set(); 2025 assert(csr->in_collection_set(), "bad CS"); 2026 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2027 HR_FORMAT_PARAMS(csr), 2028 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2029 csr->age_in_surv_rate_group_cond()); 2030 csr = next; 2031 } 2032 } 2033 #endif // !PRODUCT 2034 2035 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2036 // Returns the given amount of reclaimable bytes (that represents 2037 // the amount of reclaimable space still to be collected) as a 2038 // percentage of the current heap capacity. 2039 size_t capacity_bytes = _g1->capacity(); 2040 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2041 } 2042 2043 void G1CollectorPolicy::maybe_start_marking() { 2044 if (need_to_start_conc_mark("end of GC")) { 2045 // Note: this might have already been set, if during the last 2046 // pause we decided to start a cycle but at the beginning of 2047 // this pause we decided to postpone it. That's OK. 2048 collector_state()->set_initiate_conc_mark_if_possible(true); 2049 } 2050 } 2051 2052 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2053 assert(!collector_state()->full_collection(), "must be"); 2054 if (collector_state()->during_initial_mark_pause()) { 2055 assert(collector_state()->last_gc_was_young(), "must be"); 2056 assert(!collector_state()->last_young_gc(), "must be"); 2057 return InitialMarkGC; 2058 } else if (collector_state()->last_young_gc()) { 2059 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2060 assert(collector_state()->last_gc_was_young(), "must be"); 2061 return LastYoungGC; 2062 } else if (!collector_state()->last_gc_was_young()) { 2063 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2064 assert(!collector_state()->last_young_gc(), "must be"); 2065 return MixedGC; 2066 } else { 2067 assert(collector_state()->last_gc_was_young(), "must be"); 2068 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2069 assert(!collector_state()->last_young_gc(), "must be"); 2070 return YoungOnlyGC; 2071 } 2072 } 2073 2074 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2075 // Manage the MMU tracker. For some reason it ignores Full GCs. 2076 if (kind != FullGC) { 2077 _mmu_tracker->add_pause(start, end); 2078 } 2079 // Manage the mutator time tracking from initial mark to first mixed gc. 2080 switch (kind) { 2081 case FullGC: 2082 abort_time_to_mixed_tracking(); 2083 break; 2084 case Cleanup: 2085 case Remark: 2086 case YoungOnlyGC: 2087 case LastYoungGC: 2088 _initial_mark_to_mixed.add_pause(end - start); 2089 break; 2090 case InitialMarkGC: 2091 _initial_mark_to_mixed.record_initial_mark_end(end); 2092 break; 2093 case MixedGC: 2094 _initial_mark_to_mixed.record_mixed_gc_start(start); 2095 break; 2096 default: 2097 ShouldNotReachHere(); 2098 } 2099 } 2100 2101 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2102 _initial_mark_to_mixed.reset(); 2103 } 2104 2105 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2106 const char* false_action_str) const { 2107 if (cset_chooser()->is_empty()) { 2108 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 2109 return false; 2110 } 2111 2112 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2113 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2114 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2115 double threshold = (double) G1HeapWastePercent; 2116 if (reclaimable_perc <= threshold) { 2117 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2118 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2119 return false; 2120 } 2121 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2122 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2123 return true; 2124 } 2125 2126 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2127 // The min old CSet region bound is based on the maximum desired 2128 // number of mixed GCs after a cycle. I.e., even if some old regions 2129 // look expensive, we should add them to the CSet anyway to make 2130 // sure we go through the available old regions in no more than the 2131 // maximum desired number of mixed GCs. 2132 // 2133 // The calculation is based on the number of marked regions we added 2134 // to the CSet chooser in the first place, not how many remain, so 2135 // that the result is the same during all mixed GCs that follow a cycle. 2136 2137 const size_t region_num = (size_t) cset_chooser()->length(); 2138 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2139 size_t result = region_num / gc_num; 2140 // emulate ceiling 2141 if (result * gc_num < region_num) { 2142 result += 1; 2143 } 2144 return (uint) result; 2145 } 2146 2147 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2148 // The max old CSet region bound is based on the threshold expressed 2149 // as a percentage of the heap size. I.e., it should bound the 2150 // number of old regions added to the CSet irrespective of how many 2151 // of them are available. 2152 2153 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2154 const size_t region_num = g1h->num_regions(); 2155 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2156 size_t result = region_num * perc / 100; 2157 // emulate ceiling 2158 if (100 * result < region_num * perc) { 2159 result += 1; 2160 } 2161 return (uint) result; 2162 } 2163 2164 2165 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2166 double young_start_time_sec = os::elapsedTime(); 2167 2168 YoungList* young_list = _g1->young_list(); 2169 finalize_incremental_cset_building(); 2170 2171 guarantee(target_pause_time_ms > 0.0, 2172 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2173 guarantee(_collection_set == NULL, "Precondition"); 2174 2175 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2176 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2177 2178 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", 2179 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2180 2181 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2182 2183 if (collector_state()->last_gc_was_young()) { 2184 _trace_young_gen_time_data.increment_young_collection_count(); 2185 } else { 2186 _trace_young_gen_time_data.increment_mixed_collection_count(); 2187 } 2188 2189 // The young list is laid with the survivor regions from the previous 2190 // pause are appended to the RHS of the young list, i.e. 2191 // [Newly Young Regions ++ Survivors from last pause]. 2192 2193 uint survivor_region_length = young_list->survivor_length(); 2194 uint eden_region_length = young_list->eden_length(); 2195 init_cset_region_lengths(eden_region_length, survivor_region_length); 2196 2197 HeapRegion* hr = young_list->first_survivor_region(); 2198 while (hr != NULL) { 2199 assert(hr->is_survivor(), "badly formed young list"); 2200 // There is a convention that all the young regions in the CSet 2201 // are tagged as "eden", so we do this for the survivors here. We 2202 // use the special set_eden_pre_gc() as it doesn't check that the 2203 // region is free (which is not the case here). 2204 hr->set_eden_pre_gc(); 2205 hr = hr->get_next_young_region(); 2206 } 2207 2208 // Clear the fields that point to the survivor list - they are all young now. 2209 young_list->clear_survivors(); 2210 2211 _collection_set = _inc_cset_head; 2212 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2213 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2214 2215 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", 2216 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms); 2217 2218 // The number of recorded young regions is the incremental 2219 // collection set's current size 2220 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2221 2222 double young_end_time_sec = os::elapsedTime(); 2223 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2224 2225 return time_remaining_ms; 2226 } 2227 2228 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2229 double non_young_start_time_sec = os::elapsedTime(); 2230 double predicted_old_time_ms = 0.0; 2231 2232 2233 if (!collector_state()->gcs_are_young()) { 2234 cset_chooser()->verify(); 2235 const uint min_old_cset_length = calc_min_old_cset_length(); 2236 const uint max_old_cset_length = calc_max_old_cset_length(); 2237 2238 uint expensive_region_num = 0; 2239 bool check_time_remaining = adaptive_young_list_length(); 2240 2241 HeapRegion* hr = cset_chooser()->peek(); 2242 while (hr != NULL) { 2243 if (old_cset_region_length() >= max_old_cset_length) { 2244 // Added maximum number of old regions to the CSet. 2245 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions", 2246 old_cset_region_length(), max_old_cset_length); 2247 break; 2248 } 2249 2250 2251 // Stop adding regions if the remaining reclaimable space is 2252 // not above G1HeapWastePercent. 2253 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2254 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2255 double threshold = (double) G1HeapWastePercent; 2256 if (reclaimable_perc <= threshold) { 2257 // We've added enough old regions that the amount of uncollected 2258 // reclaimable space is at or below the waste threshold. Stop 2259 // adding old regions to the CSet. 2260 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). " 2261 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%", 2262 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2263 break; 2264 } 2265 2266 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2267 if (check_time_remaining) { 2268 if (predicted_time_ms > time_remaining_ms) { 2269 // Too expensive for the current CSet. 2270 2271 if (old_cset_region_length() >= min_old_cset_length) { 2272 // We have added the minimum number of old regions to the CSet, 2273 // we are done with this CSet. 2274 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). " 2275 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions", 2276 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length); 2277 break; 2278 } 2279 2280 // We'll add it anyway given that we haven't reached the 2281 // minimum number of old regions. 2282 expensive_region_num += 1; 2283 } 2284 } else { 2285 if (old_cset_region_length() >= min_old_cset_length) { 2286 // In the non-auto-tuning case, we'll finish adding regions 2287 // to the CSet if we reach the minimum. 2288 2289 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions", 2290 old_cset_region_length(), min_old_cset_length); 2291 break; 2292 } 2293 } 2294 2295 // We will add this region to the CSet. 2296 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2297 predicted_old_time_ms += predicted_time_ms; 2298 cset_chooser()->pop(); // already have region via peek() 2299 _g1->old_set_remove(hr); 2300 add_old_region_to_cset(hr); 2301 2302 hr = cset_chooser()->peek(); 2303 } 2304 if (hr == NULL) { 2305 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)"); 2306 } 2307 2308 if (expensive_region_num > 0) { 2309 // We print the information once here at the end, predicated on 2310 // whether we added any apparently expensive regions or not, to 2311 // avoid generating output per region. 2312 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)." 2313 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms", 2314 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms); 2315 } 2316 2317 cset_chooser()->verify(); 2318 } 2319 2320 stop_incremental_cset_building(); 2321 2322 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f", 2323 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms); 2324 2325 double non_young_end_time_sec = os::elapsedTime(); 2326 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2327 } 2328 2329 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2330 if(TraceYoungGenTime) { 2331 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2332 } 2333 } 2334 2335 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2336 if(TraceYoungGenTime) { 2337 _all_yield_times_ms.add(yield_time_ms); 2338 } 2339 } 2340 2341 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2342 if(TraceYoungGenTime) { 2343 _total.add(pause_time_ms); 2344 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2345 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2346 _parallel.add(phase_times->cur_collection_par_time_ms()); 2347 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2348 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2349 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2350 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2351 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2352 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2353 2354 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2355 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2356 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2357 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2358 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2359 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2360 2361 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2362 _parallel_other.add(parallel_other_time); 2363 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2364 } 2365 } 2366 2367 void TraceYoungGenTimeData::increment_young_collection_count() { 2368 if(TraceYoungGenTime) { 2369 ++_young_pause_num; 2370 } 2371 } 2372 2373 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2374 if(TraceYoungGenTime) { 2375 ++_mixed_pause_num; 2376 } 2377 } 2378 2379 void TraceYoungGenTimeData::print_summary(const char* str, 2380 const NumberSeq* seq) const { 2381 double sum = seq->sum(); 2382 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2383 str, sum / 1000.0, seq->avg()); 2384 } 2385 2386 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2387 const NumberSeq* seq) const { 2388 print_summary(str, seq); 2389 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2390 "(num", seq->num(), seq->sd(), seq->maximum()); 2391 } 2392 2393 void TraceYoungGenTimeData::print() const { 2394 if (!TraceYoungGenTime) { 2395 return; 2396 } 2397 2398 tty->print_cr("ALL PAUSES"); 2399 print_summary_sd(" Total", &_total); 2400 tty->cr(); 2401 tty->cr(); 2402 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2403 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2404 tty->cr(); 2405 2406 tty->print_cr("EVACUATION PAUSES"); 2407 2408 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2409 tty->print_cr("none"); 2410 } else { 2411 print_summary_sd(" Evacuation Pauses", &_total); 2412 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2413 print_summary(" Parallel Time", &_parallel); 2414 print_summary(" Ext Root Scanning", &_ext_root_scan); 2415 print_summary(" SATB Filtering", &_satb_filtering); 2416 print_summary(" Update RS", &_update_rs); 2417 print_summary(" Scan RS", &_scan_rs); 2418 print_summary(" Object Copy", &_obj_copy); 2419 print_summary(" Termination", &_termination); 2420 print_summary(" Parallel Other", &_parallel_other); 2421 print_summary(" Clear CT", &_clear_ct); 2422 print_summary(" Other", &_other); 2423 } 2424 tty->cr(); 2425 2426 tty->print_cr("MISC"); 2427 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2428 print_summary_sd(" Yields", &_all_yield_times_ms); 2429 } 2430 2431 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2432 if (TraceOldGenTime) { 2433 _all_full_gc_times.add(full_gc_time_ms); 2434 } 2435 } 2436 2437 void TraceOldGenTimeData::print() const { 2438 if (!TraceOldGenTime) { 2439 return; 2440 } 2441 2442 if (_all_full_gc_times.num() > 0) { 2443 tty->print("\n%4d full_gcs: total time = %8.2f s", 2444 _all_full_gc_times.num(), 2445 _all_full_gc_times.sum() / 1000.0); 2446 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2447 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2448 _all_full_gc_times.sd(), 2449 _all_full_gc_times.maximum()); 2450 } 2451 }