1 /* 2 * Copyright (c) 2001, 2016, 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/concurrentMarkThread.inline.hpp" 28 #include "gc/g1/g1CollectedHeap.inline.hpp" 29 #include "gc/g1/g1CollectorPolicy.hpp" 30 #include "gc/g1/g1ConcurrentMark.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_cset_region_length(0), 121 _survivor_cset_region_length(0), 122 _old_cset_region_length(0), 123 124 _collection_set(NULL), 125 _collection_set_bytes_used_before(0), 126 127 // Incremental CSet attributes 128 _inc_cset_build_state(Inactive), 129 _inc_cset_head(NULL), 130 _inc_cset_tail(NULL), 131 _inc_cset_bytes_used_before(0), 132 _inc_cset_max_finger(NULL), 133 _inc_cset_recorded_rs_lengths(0), 134 _inc_cset_recorded_rs_lengths_diffs(0), 135 _inc_cset_predicted_elapsed_time_ms(0.0), 136 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 137 138 // add here any more surv rate groups 139 _recorded_survivor_regions(0), 140 _recorded_survivor_head(NULL), 141 _recorded_survivor_tail(NULL), 142 _survivors_age_table(true), 143 144 _gc_overhead_perc(0.0), 145 146 _bytes_allocated_in_old_since_last_gc(0), 147 _ihop_control(NULL), 148 _initial_mark_to_mixed() { 149 150 // SurvRateGroups below must be initialized after the predictor because they 151 // indirectly use it through this object passed to their constructor. 152 _short_lived_surv_rate_group = 153 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 154 _survivor_surv_rate_group = 155 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 156 157 // Set up the region size and associated fields. Given that the 158 // policy is created before the heap, we have to set this up here, 159 // so it's done as soon as possible. 160 161 // It would have been natural to pass initial_heap_byte_size() and 162 // max_heap_byte_size() to setup_heap_region_size() but those have 163 // not been set up at this point since they should be aligned with 164 // the region size. So, there is a circular dependency here. We base 165 // the region size on the heap size, but the heap size should be 166 // aligned with the region size. To get around this we use the 167 // unaligned values for the heap. 168 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 169 HeapRegionRemSet::setup_remset_size(); 170 171 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 172 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 173 clear_ratio_check_data(); 174 175 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 176 177 int index = MIN2(_parallel_gc_threads - 1, 7); 178 179 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 180 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 181 _cost_scan_hcc_seq->add(0.0); 182 _young_cards_per_entry_ratio_seq->add( 183 young_cards_per_entry_ratio_defaults[index]); 184 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 185 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 186 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 187 _young_other_cost_per_region_ms_seq->add( 188 young_other_cost_per_region_ms_defaults[index]); 189 _non_young_other_cost_per_region_ms_seq->add( 190 non_young_other_cost_per_region_ms_defaults[index]); 191 192 // Below, we might need to calculate the pause time target based on 193 // the pause interval. When we do so we are going to give G1 maximum 194 // flexibility and allow it to do pauses when it needs to. So, we'll 195 // arrange that the pause interval to be pause time target + 1 to 196 // ensure that a) the pause time target is maximized with respect to 197 // the pause interval and b) we maintain the invariant that pause 198 // time target < pause interval. If the user does not want this 199 // maximum flexibility, they will have to set the pause interval 200 // explicitly. 201 202 // First make sure that, if either parameter is set, its value is 203 // reasonable. 204 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 205 if (MaxGCPauseMillis < 1) { 206 vm_exit_during_initialization("MaxGCPauseMillis should be " 207 "greater than 0"); 208 } 209 } 210 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 211 if (GCPauseIntervalMillis < 1) { 212 vm_exit_during_initialization("GCPauseIntervalMillis should be " 213 "greater than 0"); 214 } 215 } 216 217 // Then, if the pause time target parameter was not set, set it to 218 // the default value. 219 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 220 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 221 // The default pause time target in G1 is 200ms 222 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 223 } else { 224 // We do not allow the pause interval to be set without the 225 // pause time target 226 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 227 "without setting MaxGCPauseMillis"); 228 } 229 } 230 231 // Then, if the interval parameter was not set, set it according to 232 // the pause time target (this will also deal with the case when the 233 // pause time target is the default value). 234 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 235 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 236 } 237 238 // Finally, make sure that the two parameters are consistent. 239 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 240 char buffer[256]; 241 jio_snprintf(buffer, 256, 242 "MaxGCPauseMillis (%u) should be less than " 243 "GCPauseIntervalMillis (%u)", 244 MaxGCPauseMillis, GCPauseIntervalMillis); 245 vm_exit_during_initialization(buffer); 246 } 247 248 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 249 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 250 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 251 252 // start conservatively (around 50ms is about right) 253 _concurrent_mark_remark_times_ms->add(0.05); 254 _concurrent_mark_cleanup_times_ms->add(0.20); 255 _tenuring_threshold = MaxTenuringThreshold; 256 257 assert(GCTimeRatio > 0, 258 "we should have set it to a default value set_g1_gc_flags() " 259 "if a user set it to 0"); 260 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 261 262 uintx reserve_perc = G1ReservePercent; 263 // Put an artificial ceiling on this so that it's not set to a silly value. 264 if (reserve_perc > 50) { 265 reserve_perc = 50; 266 warning("G1ReservePercent is set to a value that is too large, " 267 "it's been updated to " UINTX_FORMAT, reserve_perc); 268 } 269 _reserve_factor = (double) reserve_perc / 100.0; 270 // This will be set when the heap is expanded 271 // for the first time during initialization. 272 _reserve_regions = 0; 273 274 _cset_chooser = new CollectionSetChooser(); 275 } 276 277 G1CollectorPolicy::~G1CollectorPolicy() { 278 delete _ihop_control; 279 } 280 281 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 282 return _predictor.get_new_prediction(seq); 283 } 284 285 size_t G1CollectorPolicy::get_new_size_prediction(TruncatedSeq const* seq) const { 286 return (size_t)get_new_prediction(seq); 287 } 288 289 void G1CollectorPolicy::initialize_alignments() { 290 _space_alignment = HeapRegion::GrainBytes; 291 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 292 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 293 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 294 } 295 296 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 297 298 // There are three command line options related to the young gen size: 299 // NewSize, MaxNewSize and NewRatio (There is also -Xmn, but that is 300 // just a short form for NewSize==MaxNewSize). G1 will use its internal 301 // heuristics to calculate the actual young gen size, so these options 302 // basically only limit the range within which G1 can pick a young gen 303 // size. Also, these are general options taking byte sizes. G1 will 304 // internally work with a number of regions instead. So, some rounding 305 // will occur. 306 // 307 // If nothing related to the the young gen size is set on the command 308 // line we should allow the young gen to be between G1NewSizePercent 309 // and G1MaxNewSizePercent of the heap size. This means that every time 310 // the heap size changes, the limits for the young gen size will be 311 // recalculated. 312 // 313 // If only -XX:NewSize is set we should use the specified value as the 314 // minimum size for young gen. Still using G1MaxNewSizePercent of the 315 // heap as maximum. 316 // 317 // If only -XX:MaxNewSize is set we should use the specified value as the 318 // maximum size for young gen. Still using G1NewSizePercent of the heap 319 // as minimum. 320 // 321 // If -XX:NewSize and -XX:MaxNewSize are both specified we use these values. 322 // No updates when the heap size changes. There is a special case when 323 // NewSize==MaxNewSize. This is interpreted as "fixed" and will use a 324 // different heuristic for calculating the collection set when we do mixed 325 // collection. 326 // 327 // If only -XX:NewRatio is set we should use the specified ratio of the heap 328 // as both min and max. This will be interpreted as "fixed" just like the 329 // NewSize==MaxNewSize case above. But we will update the min and max 330 // every time the heap size changes. 331 // 332 // NewSize and MaxNewSize override NewRatio. So, NewRatio is ignored if it is 333 // combined with either NewSize or MaxNewSize. (A warning message is printed.) 334 class G1YoungGenSizer : public CHeapObj<mtGC> { 335 private: 336 enum SizerKind { 337 SizerDefaults, 338 SizerNewSizeOnly, 339 SizerMaxNewSizeOnly, 340 SizerMaxAndNewSize, 341 SizerNewRatio 342 }; 343 SizerKind _sizer_kind; 344 uint _min_desired_young_length; 345 uint _max_desired_young_length; 346 bool _adaptive_size; 347 uint calculate_default_min_length(uint new_number_of_heap_regions); 348 uint calculate_default_max_length(uint new_number_of_heap_regions); 349 350 // Update the given values for minimum and maximum young gen length in regions 351 // given the number of heap regions depending on the kind of sizing algorithm. 352 void recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length); 353 354 public: 355 G1YoungGenSizer(); 356 // Calculate the maximum length of the young gen given the number of regions 357 // depending on the sizing algorithm. 358 uint max_young_length(uint number_of_heap_regions); 359 360 void heap_size_changed(uint new_number_of_heap_regions); 361 uint min_desired_young_length() { 362 return _min_desired_young_length; 363 } 364 uint max_desired_young_length() { 365 return _max_desired_young_length; 366 } 367 368 bool adaptive_young_list_length() const { 369 return _adaptive_size; 370 } 371 }; 372 373 374 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 375 _min_desired_young_length(0), _max_desired_young_length(0) { 376 if (FLAG_IS_CMDLINE(NewRatio)) { 377 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 378 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 379 } else { 380 _sizer_kind = SizerNewRatio; 381 _adaptive_size = false; 382 return; 383 } 384 } 385 386 if (NewSize > MaxNewSize) { 387 if (FLAG_IS_CMDLINE(MaxNewSize)) { 388 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 389 "A new max generation size of " SIZE_FORMAT "k will be used.", 390 NewSize/K, MaxNewSize/K, NewSize/K); 391 } 392 MaxNewSize = NewSize; 393 } 394 395 if (FLAG_IS_CMDLINE(NewSize)) { 396 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 397 1U); 398 if (FLAG_IS_CMDLINE(MaxNewSize)) { 399 _max_desired_young_length = 400 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 401 1U); 402 _sizer_kind = SizerMaxAndNewSize; 403 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 404 } else { 405 _sizer_kind = SizerNewSizeOnly; 406 } 407 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 408 _max_desired_young_length = 409 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 410 1U); 411 _sizer_kind = SizerMaxNewSizeOnly; 412 } 413 } 414 415 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 416 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 417 return MAX2(1U, default_value); 418 } 419 420 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 421 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 422 return MAX2(1U, default_value); 423 } 424 425 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 426 assert(number_of_heap_regions > 0, "Heap must be initialized"); 427 428 switch (_sizer_kind) { 429 case SizerDefaults: 430 *min_young_length = calculate_default_min_length(number_of_heap_regions); 431 *max_young_length = calculate_default_max_length(number_of_heap_regions); 432 break; 433 case SizerNewSizeOnly: 434 *max_young_length = calculate_default_max_length(number_of_heap_regions); 435 *max_young_length = MAX2(*min_young_length, *max_young_length); 436 break; 437 case SizerMaxNewSizeOnly: 438 *min_young_length = calculate_default_min_length(number_of_heap_regions); 439 *min_young_length = MIN2(*min_young_length, *max_young_length); 440 break; 441 case SizerMaxAndNewSize: 442 // Do nothing. Values set on the command line, don't update them at runtime. 443 break; 444 case SizerNewRatio: 445 *min_young_length = number_of_heap_regions / (NewRatio + 1); 446 *max_young_length = *min_young_length; 447 break; 448 default: 449 ShouldNotReachHere(); 450 } 451 452 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 453 } 454 455 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 456 // We need to pass the desired values because recalculation may not update these 457 // values in some cases. 458 uint temp = _min_desired_young_length; 459 uint result = _max_desired_young_length; 460 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 461 return result; 462 } 463 464 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 465 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 466 &_max_desired_young_length); 467 } 468 469 void G1CollectorPolicy::post_heap_initialize() { 470 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 471 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 472 if (max_young_size != MaxNewSize) { 473 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 474 } 475 476 _ihop_control = create_ihop_control(); 477 } 478 479 void G1CollectorPolicy::initialize_flags() { 480 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 481 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 482 } 483 484 if (SurvivorRatio < 1) { 485 vm_exit_during_initialization("Invalid survivor ratio specified"); 486 } 487 CollectorPolicy::initialize_flags(); 488 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 489 } 490 491 492 void G1CollectorPolicy::init() { 493 // Set aside an initial future to_space. 494 _g1 = G1CollectedHeap::heap(); 495 496 assert(Heap_lock->owned_by_self(), "Locking discipline."); 497 498 initialize_gc_policy_counters(); 499 500 if (adaptive_young_list_length()) { 501 _young_list_fixed_length = 0; 502 } else { 503 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 504 } 505 _free_regions_at_end_of_collection = _g1->num_free_regions(); 506 507 update_young_list_max_and_target_length(); 508 // We may immediately start allocating regions and placing them on the 509 // collection set list. Initialize the per-collection set info 510 start_incremental_cset_building(); 511 } 512 513 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 514 phase_times()->note_gc_start(num_active_workers); 515 } 516 517 // Create the jstat counters for the policy. 518 void G1CollectorPolicy::initialize_gc_policy_counters() { 519 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 520 } 521 522 bool G1CollectorPolicy::predict_will_fit(uint young_length, 523 double base_time_ms, 524 uint base_free_regions, 525 double target_pause_time_ms) const { 526 if (young_length >= base_free_regions) { 527 // end condition 1: not enough space for the young regions 528 return false; 529 } 530 531 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 532 size_t bytes_to_copy = 533 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 534 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 535 double young_other_time_ms = predict_young_other_time_ms(young_length); 536 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 537 if (pause_time_ms > target_pause_time_ms) { 538 // end condition 2: prediction is over the target pause time 539 return false; 540 } 541 542 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 543 544 // When copying, we will likely need more bytes free than is live in the region. 545 // Add some safety margin to factor in the confidence of our guess, and the 546 // natural expected waste. 547 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 548 // of the calculation: the lower the confidence, the more headroom. 549 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 550 // copying due to anticipated waste in the PLABs. 551 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 552 size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy); 553 554 if (expected_bytes_to_copy > free_bytes) { 555 // end condition 3: out-of-space 556 return false; 557 } 558 559 // success! 560 return true; 561 } 562 563 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 564 // re-calculate the necessary reserve 565 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 566 // We use ceiling so that if reserve_regions_d is > 0.0 (but 567 // smaller than 1.0) we'll get 1. 568 _reserve_regions = (uint) ceil(reserve_regions_d); 569 570 _young_gen_sizer->heap_size_changed(new_number_of_regions); 571 } 572 573 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 574 uint base_min_length) const { 575 uint desired_min_length = 0; 576 if (adaptive_young_list_length()) { 577 if (_alloc_rate_ms_seq->num() > 3) { 578 double now_sec = os::elapsedTime(); 579 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 580 double alloc_rate_ms = predict_alloc_rate_ms(); 581 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 582 } else { 583 // otherwise we don't have enough info to make the prediction 584 } 585 } 586 desired_min_length += base_min_length; 587 // make sure we don't go below any user-defined minimum bound 588 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 589 } 590 591 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 592 // Here, we might want to also take into account any additional 593 // constraints (i.e., user-defined minimum bound). Currently, we 594 // effectively don't set this bound. 595 return _young_gen_sizer->max_desired_young_length(); 596 } 597 598 uint G1CollectorPolicy::update_young_list_max_and_target_length() { 599 return update_young_list_max_and_target_length(get_new_size_prediction(_rs_lengths_seq)); 600 } 601 602 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 603 uint unbounded_target_length = update_young_list_target_length(rs_lengths); 604 update_max_gc_locker_expansion(); 605 return unbounded_target_length; 606 } 607 608 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 609 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths); 610 _young_list_target_length = young_lengths.first; 611 return young_lengths.second; 612 } 613 614 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const { 615 YoungTargetLengths result; 616 617 // Calculate the absolute and desired min bounds first. 618 619 // This is how many young regions we already have (currently: the survivors). 620 uint base_min_length = recorded_survivor_regions(); 621 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 622 // This is the absolute minimum young length. Ensure that we 623 // will at least have one eden region available for allocation. 624 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 625 // If we shrank the young list target it should not shrink below the current size. 626 desired_min_length = MAX2(desired_min_length, absolute_min_length); 627 // Calculate the absolute and desired max bounds. 628 629 uint desired_max_length = calculate_young_list_desired_max_length(); 630 631 uint young_list_target_length = 0; 632 if (adaptive_young_list_length()) { 633 if (collector_state()->gcs_are_young()) { 634 young_list_target_length = 635 calculate_young_list_target_length(rs_lengths, 636 base_min_length, 637 desired_min_length, 638 desired_max_length); 639 } else { 640 // Don't calculate anything and let the code below bound it to 641 // the desired_min_length, i.e., do the next GC as soon as 642 // possible to maximize how many old regions we can add to it. 643 } 644 } else { 645 // The user asked for a fixed young gen so we'll fix the young gen 646 // whether the next GC is young or mixed. 647 young_list_target_length = _young_list_fixed_length; 648 } 649 650 result.second = young_list_target_length; 651 652 // We will try our best not to "eat" into the reserve. 653 uint absolute_max_length = 0; 654 if (_free_regions_at_end_of_collection > _reserve_regions) { 655 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 656 } 657 if (desired_max_length > absolute_max_length) { 658 desired_max_length = absolute_max_length; 659 } 660 661 // Make sure we don't go over the desired max length, nor under the 662 // desired min length. In case they clash, desired_min_length wins 663 // which is why that test is second. 664 if (young_list_target_length > desired_max_length) { 665 young_list_target_length = desired_max_length; 666 } 667 if (young_list_target_length < desired_min_length) { 668 young_list_target_length = desired_min_length; 669 } 670 671 assert(young_list_target_length > recorded_survivor_regions(), 672 "we should be able to allocate at least one eden region"); 673 assert(young_list_target_length >= absolute_min_length, "post-condition"); 674 675 result.first = young_list_target_length; 676 return result; 677 } 678 679 uint 680 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 681 uint base_min_length, 682 uint desired_min_length, 683 uint desired_max_length) const { 684 assert(adaptive_young_list_length(), "pre-condition"); 685 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 686 687 // In case some edge-condition makes the desired max length too small... 688 if (desired_max_length <= desired_min_length) { 689 return desired_min_length; 690 } 691 692 // We'll adjust min_young_length and max_young_length not to include 693 // the already allocated young regions (i.e., so they reflect the 694 // min and max eden regions we'll allocate). The base_min_length 695 // will be reflected in the predictions by the 696 // survivor_regions_evac_time prediction. 697 assert(desired_min_length > base_min_length, "invariant"); 698 uint min_young_length = desired_min_length - base_min_length; 699 assert(desired_max_length > base_min_length, "invariant"); 700 uint max_young_length = desired_max_length - base_min_length; 701 702 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 703 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 704 size_t pending_cards = get_new_size_prediction(_pending_cards_seq); 705 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 706 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 707 double base_time_ms = 708 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 709 survivor_regions_evac_time; 710 uint available_free_regions = _free_regions_at_end_of_collection; 711 uint base_free_regions = 0; 712 if (available_free_regions > _reserve_regions) { 713 base_free_regions = available_free_regions - _reserve_regions; 714 } 715 716 // Here, we will make sure that the shortest young length that 717 // makes sense fits within the target pause time. 718 719 if (predict_will_fit(min_young_length, base_time_ms, 720 base_free_regions, target_pause_time_ms)) { 721 // The shortest young length will fit into the target pause time; 722 // we'll now check whether the absolute maximum number of young 723 // regions will fit in the target pause time. If not, we'll do 724 // a binary search between min_young_length and max_young_length. 725 if (predict_will_fit(max_young_length, base_time_ms, 726 base_free_regions, target_pause_time_ms)) { 727 // The maximum young length will fit into the target pause time. 728 // We are done so set min young length to the maximum length (as 729 // the result is assumed to be returned in min_young_length). 730 min_young_length = max_young_length; 731 } else { 732 // The maximum possible number of young regions will not fit within 733 // the target pause time so we'll search for the optimal 734 // length. The loop invariants are: 735 // 736 // min_young_length < max_young_length 737 // min_young_length is known to fit into the target pause time 738 // max_young_length is known not to fit into the target pause time 739 // 740 // Going into the loop we know the above hold as we've just 741 // checked them. Every time around the loop we check whether 742 // the middle value between min_young_length and 743 // max_young_length fits into the target pause time. If it 744 // does, it becomes the new min. If it doesn't, it becomes 745 // the new max. This way we maintain the loop invariants. 746 747 assert(min_young_length < max_young_length, "invariant"); 748 uint diff = (max_young_length - min_young_length) / 2; 749 while (diff > 0) { 750 uint young_length = min_young_length + diff; 751 if (predict_will_fit(young_length, base_time_ms, 752 base_free_regions, target_pause_time_ms)) { 753 min_young_length = young_length; 754 } else { 755 max_young_length = young_length; 756 } 757 assert(min_young_length < max_young_length, "invariant"); 758 diff = (max_young_length - min_young_length) / 2; 759 } 760 // The results is min_young_length which, according to the 761 // loop invariants, should fit within the target pause time. 762 763 // These are the post-conditions of the binary search above: 764 assert(min_young_length < max_young_length, 765 "otherwise we should have discovered that max_young_length " 766 "fits into the pause target and not done the binary search"); 767 assert(predict_will_fit(min_young_length, base_time_ms, 768 base_free_regions, target_pause_time_ms), 769 "min_young_length, the result of the binary search, should " 770 "fit into the pause target"); 771 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 772 base_free_regions, target_pause_time_ms), 773 "min_young_length, the result of the binary search, should be " 774 "optimal, so no larger length should fit into the pause target"); 775 } 776 } else { 777 // Even the minimum length doesn't fit into the pause time 778 // target, return it as the result nevertheless. 779 } 780 return base_min_length + min_young_length; 781 } 782 783 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 784 double survivor_regions_evac_time = 0.0; 785 for (HeapRegion * r = _recorded_survivor_head; 786 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 787 r = r->get_next_young_region()) { 788 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 789 } 790 return survivor_regions_evac_time; 791 } 792 793 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 794 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 795 796 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 797 if (rs_lengths > _rs_lengths_prediction) { 798 // add 10% to avoid having to recalculate often 799 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 800 update_rs_lengths_prediction(rs_lengths_prediction); 801 802 update_young_list_max_and_target_length(rs_lengths_prediction); 803 } 804 } 805 806 void G1CollectorPolicy::update_rs_lengths_prediction() { 807 update_rs_lengths_prediction(get_new_size_prediction(_rs_lengths_seq)); 808 } 809 810 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 811 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 812 _rs_lengths_prediction = prediction; 813 } 814 } 815 816 #ifndef PRODUCT 817 bool G1CollectorPolicy::verify_young_ages() { 818 HeapRegion* head = _g1->young_list()->first_region(); 819 return 820 verify_young_ages(head, _short_lived_surv_rate_group); 821 // also call verify_young_ages on any additional surv rate groups 822 } 823 824 bool 825 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 826 SurvRateGroup *surv_rate_group) { 827 guarantee( surv_rate_group != NULL, "pre-condition" ); 828 829 const char* name = surv_rate_group->name(); 830 bool ret = true; 831 int prev_age = -1; 832 833 for (HeapRegion* curr = head; 834 curr != NULL; 835 curr = curr->get_next_young_region()) { 836 SurvRateGroup* group = curr->surv_rate_group(); 837 if (group == NULL && !curr->is_survivor()) { 838 log_error(gc, verify)("## %s: encountered NULL surv_rate_group", name); 839 ret = false; 840 } 841 842 if (surv_rate_group == group) { 843 int age = curr->age_in_surv_rate_group(); 844 845 if (age < 0) { 846 log_error(gc, verify)("## %s: encountered negative age", name); 847 ret = false; 848 } 849 850 if (age <= prev_age) { 851 log_error(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age); 852 ret = false; 853 } 854 prev_age = age; 855 } 856 } 857 858 return ret; 859 } 860 #endif // PRODUCT 861 862 void G1CollectorPolicy::record_full_collection_start() { 863 _full_collection_start_sec = os::elapsedTime(); 864 // Release the future to-space so that it is available for compaction into. 865 collector_state()->set_full_collection(true); 866 } 867 868 void G1CollectorPolicy::record_full_collection_end() { 869 // Consider this like a collection pause for the purposes of allocation 870 // since last pause. 871 double end_sec = os::elapsedTime(); 872 double full_gc_time_sec = end_sec - _full_collection_start_sec; 873 double full_gc_time_ms = full_gc_time_sec * 1000.0; 874 875 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 876 877 update_recent_gc_times(end_sec, full_gc_time_ms); 878 879 collector_state()->set_full_collection(false); 880 881 // "Nuke" the heuristics that control the young/mixed GC 882 // transitions and make sure we start with young GCs after the Full GC. 883 collector_state()->set_gcs_are_young(true); 884 collector_state()->set_last_young_gc(false); 885 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 886 collector_state()->set_during_initial_mark_pause(false); 887 collector_state()->set_in_marking_window(false); 888 collector_state()->set_in_marking_window_im(false); 889 890 _short_lived_surv_rate_group->start_adding_regions(); 891 // also call this on any additional surv rate groups 892 893 record_survivor_regions(0, NULL, NULL); 894 895 _free_regions_at_end_of_collection = _g1->num_free_regions(); 896 // Reset survivors SurvRateGroup. 897 _survivor_surv_rate_group->reset(); 898 update_young_list_max_and_target_length(); 899 update_rs_lengths_prediction(); 900 cset_chooser()->clear(); 901 902 _bytes_allocated_in_old_since_last_gc = 0; 903 904 record_pause(FullGC, _full_collection_start_sec, end_sec); 905 } 906 907 void G1CollectorPolicy::record_stop_world_start() { 908 _stop_world_start = os::elapsedTime(); 909 } 910 911 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 912 // We only need to do this here as the policy will only be applied 913 // to the GC we're about to start. so, no point is calculating this 914 // every time we calculate / recalculate the target young length. 915 update_survivors_policy(); 916 917 assert(_g1->used() == _g1->recalculate_used(), 918 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 919 _g1->used(), _g1->recalculate_used()); 920 921 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 922 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 923 _stop_world_start = 0.0; 924 925 phase_times()->record_cur_collection_start_sec(start_time_sec); 926 _pending_cards = _g1->pending_card_num(); 927 928 _collection_set_bytes_used_before = 0; 929 _bytes_copied_during_gc = 0; 930 931 collector_state()->set_last_gc_was_young(false); 932 933 // do that for any other surv rate groups 934 _short_lived_surv_rate_group->stop_adding_regions(); 935 _survivors_age_table.clear(); 936 937 assert( verify_young_ages(), "region age verification" ); 938 } 939 940 void G1CollectorPolicy::record_concurrent_mark_init_end(double 941 mark_init_elapsed_time_ms) { 942 collector_state()->set_during_marking(true); 943 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 944 collector_state()->set_during_initial_mark_pause(false); 945 } 946 947 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 948 _mark_remark_start_sec = os::elapsedTime(); 949 collector_state()->set_during_marking(false); 950 } 951 952 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 953 double end_time_sec = os::elapsedTime(); 954 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 955 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 956 _prev_collection_pause_end_ms += elapsed_time_ms; 957 958 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 959 } 960 961 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 962 _mark_cleanup_start_sec = os::elapsedTime(); 963 } 964 965 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 966 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 967 "skip last young-only gc"); 968 collector_state()->set_last_young_gc(should_continue_with_reclaim); 969 // We skip the marking phase. 970 if (!should_continue_with_reclaim) { 971 abort_time_to_mixed_tracking(); 972 } 973 collector_state()->set_in_marking_window(false); 974 } 975 976 void G1CollectorPolicy::record_concurrent_pause() { 977 if (_stop_world_start > 0.0) { 978 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 979 _trace_young_gen_time_data.record_yield_time(yield_ms); 980 } 981 } 982 983 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 984 return phase_times()->average_time_ms(phase); 985 } 986 987 double G1CollectorPolicy::young_other_time_ms() const { 988 return phase_times()->young_cset_choice_time_ms() + 989 phase_times()->young_free_cset_time_ms(); 990 } 991 992 double G1CollectorPolicy::non_young_other_time_ms() const { 993 return phase_times()->non_young_cset_choice_time_ms() + 994 phase_times()->non_young_free_cset_time_ms(); 995 996 } 997 998 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 999 return pause_time_ms - 1000 average_time_ms(G1GCPhaseTimes::UpdateRS) - 1001 average_time_ms(G1GCPhaseTimes::ScanRS) - 1002 average_time_ms(G1GCPhaseTimes::ObjCopy) - 1003 average_time_ms(G1GCPhaseTimes::Termination); 1004 } 1005 1006 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 1007 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 1008 } 1009 1010 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 1011 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 1012 } 1013 1014 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 1015 if (about_to_start_mixed_phase()) { 1016 return false; 1017 } 1018 1019 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 1020 1021 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 1022 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 1023 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 1024 1025 bool result = false; 1026 if (marking_request_bytes > marking_initiating_used_threshold) { 1027 result = collector_state()->gcs_are_young() && !collector_state()->last_young_gc(); 1028 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s", 1029 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)", 1030 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1->capacity() * 100, source); 1031 } 1032 1033 return result; 1034 } 1035 1036 // Anything below that is considered to be zero 1037 #define MIN_TIMER_GRANULARITY 0.0000001 1038 1039 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned, size_t heap_used_bytes_before_gc) { 1040 double end_time_sec = os::elapsedTime(); 1041 1042 size_t cur_used_bytes = _g1->used(); 1043 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1044 bool last_pause_included_initial_mark = false; 1045 bool update_stats = !_g1->evacuation_failed(); 1046 1047 NOT_PRODUCT(_short_lived_surv_rate_group->print()); 1048 1049 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec); 1050 1051 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1052 if (last_pause_included_initial_mark) { 1053 record_concurrent_mark_init_end(0.0); 1054 } else { 1055 maybe_start_marking(); 1056 } 1057 1058 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1059 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1060 // This usually happens due to the timer not having the required 1061 // granularity. Some Linuxes are the usual culprits. 1062 // We'll just set it to something (arbitrarily) small. 1063 app_time_ms = 1.0; 1064 } 1065 1066 if (update_stats) { 1067 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1068 // We maintain the invariant that all objects allocated by mutator 1069 // threads will be allocated out of eden regions. So, we can use 1070 // the eden region number allocated since the previous GC to 1071 // calculate the application's allocate rate. The only exception 1072 // to that is humongous objects that are allocated separately. But 1073 // given that humongous object allocations do not really affect 1074 // either the pause's duration nor when the next pause will take 1075 // place we can safely ignore them here. 1076 uint regions_allocated = eden_cset_region_length(); 1077 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1078 _alloc_rate_ms_seq->add(alloc_rate_ms); 1079 1080 double interval_ms = 1081 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1082 update_recent_gc_times(end_time_sec, pause_time_ms); 1083 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1084 if (recent_avg_pause_time_ratio() < 0.0 || 1085 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1086 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1087 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1088 if (_recent_avg_pause_time_ratio < 0.0) { 1089 _recent_avg_pause_time_ratio = 0.0; 1090 } else { 1091 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1092 _recent_avg_pause_time_ratio = 1.0; 1093 } 1094 } 1095 1096 // Compute the ratio of just this last pause time to the entire time range stored 1097 // in the vectors. Comparing this pause to the entire range, rather than only the 1098 // most recent interval, has the effect of smoothing over a possible transient 'burst' 1099 // of more frequent pauses that don't really reflect a change in heap occupancy. 1100 // This reduces the likelihood of a needless heap expansion being triggered. 1101 _last_pause_time_ratio = 1102 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms; 1103 } 1104 1105 bool new_in_marking_window = collector_state()->in_marking_window(); 1106 bool new_in_marking_window_im = false; 1107 if (last_pause_included_initial_mark) { 1108 new_in_marking_window = true; 1109 new_in_marking_window_im = true; 1110 } 1111 1112 if (collector_state()->last_young_gc()) { 1113 // This is supposed to to be the "last young GC" before we start 1114 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1115 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1116 1117 if (next_gc_should_be_mixed("start mixed GCs", 1118 "do not start mixed GCs")) { 1119 collector_state()->set_gcs_are_young(false); 1120 } else { 1121 // We aborted the mixed GC phase early. 1122 abort_time_to_mixed_tracking(); 1123 } 1124 1125 collector_state()->set_last_young_gc(false); 1126 } 1127 1128 if (!collector_state()->last_gc_was_young()) { 1129 // This is a mixed GC. Here we decide whether to continue doing 1130 // mixed GCs or not. 1131 if (!next_gc_should_be_mixed("continue mixed GCs", 1132 "do not continue mixed GCs")) { 1133 collector_state()->set_gcs_are_young(true); 1134 1135 maybe_start_marking(); 1136 } 1137 } 1138 1139 _short_lived_surv_rate_group->start_adding_regions(); 1140 // Do that for any other surv rate groups 1141 1142 if (update_stats) { 1143 double cost_per_card_ms = 0.0; 1144 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1145 if (_pending_cards > 0) { 1146 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1147 _cost_per_card_ms_seq->add(cost_per_card_ms); 1148 } 1149 _cost_scan_hcc_seq->add(cost_scan_hcc); 1150 1151 double cost_per_entry_ms = 0.0; 1152 if (cards_scanned > 10) { 1153 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1154 if (collector_state()->last_gc_was_young()) { 1155 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1156 } else { 1157 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1158 } 1159 } 1160 1161 if (_max_rs_lengths > 0) { 1162 double cards_per_entry_ratio = 1163 (double) cards_scanned / (double) _max_rs_lengths; 1164 if (collector_state()->last_gc_was_young()) { 1165 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1166 } else { 1167 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1168 } 1169 } 1170 1171 // This is defensive. For a while _max_rs_lengths could get 1172 // smaller than _recorded_rs_lengths which was causing 1173 // rs_length_diff to get very large and mess up the RSet length 1174 // predictions. The reason was unsafe concurrent updates to the 1175 // _inc_cset_recorded_rs_lengths field which the code below guards 1176 // against (see CR 7118202). This bug has now been fixed (see CR 1177 // 7119027). However, I'm still worried that 1178 // _inc_cset_recorded_rs_lengths might still end up somewhat 1179 // inaccurate. The concurrent refinement thread calculates an 1180 // RSet's length concurrently with other CR threads updating it 1181 // which might cause it to calculate the length incorrectly (if, 1182 // say, it's in mid-coarsening). So I'll leave in the defensive 1183 // conditional below just in case. 1184 size_t rs_length_diff = 0; 1185 if (_max_rs_lengths > _recorded_rs_lengths) { 1186 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1187 } 1188 _rs_length_diff_seq->add((double) rs_length_diff); 1189 1190 size_t freed_bytes = heap_used_bytes_before_gc - cur_used_bytes; 1191 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1192 double cost_per_byte_ms = 0.0; 1193 1194 if (copied_bytes > 0) { 1195 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1196 if (collector_state()->in_marking_window()) { 1197 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1198 } else { 1199 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1200 } 1201 } 1202 1203 if (young_cset_region_length() > 0) { 1204 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1205 young_cset_region_length()); 1206 } 1207 1208 if (old_cset_region_length() > 0) { 1209 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1210 old_cset_region_length()); 1211 } 1212 1213 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1214 1215 _pending_cards_seq->add((double) _pending_cards); 1216 _rs_lengths_seq->add((double) _max_rs_lengths); 1217 } 1218 1219 collector_state()->set_in_marking_window(new_in_marking_window); 1220 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1221 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1222 // IHOP control wants to know the expected young gen length if it were not 1223 // restrained by the heap reserve. Using the actual length would make the 1224 // prediction too small and the limit the young gen every time we get to the 1225 // predicted target occupancy. 1226 size_t last_unrestrained_young_length = update_young_list_max_and_target_length(); 1227 update_rs_lengths_prediction(); 1228 1229 update_ihop_prediction(app_time_ms / 1000.0, 1230 _bytes_allocated_in_old_since_last_gc, 1231 last_unrestrained_young_length * HeapRegion::GrainBytes); 1232 _bytes_allocated_in_old_since_last_gc = 0; 1233 1234 _ihop_control->send_trace_event(_g1->gc_tracer_stw()); 1235 1236 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1237 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1238 1239 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1240 1241 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1242 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 1243 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms", 1244 update_rs_time_goal_ms, scan_hcc_time_ms); 1245 1246 update_rs_time_goal_ms = 0; 1247 } else { 1248 update_rs_time_goal_ms -= scan_hcc_time_ms; 1249 } 1250 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1251 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1252 update_rs_time_goal_ms); 1253 1254 cset_chooser()->verify(); 1255 } 1256 1257 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const { 1258 if (G1UseAdaptiveIHOP) { 1259 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 1260 G1CollectedHeap::heap()->max_capacity(), 1261 &_predictor, 1262 G1ReservePercent, 1263 G1HeapWastePercent); 1264 } else { 1265 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent, 1266 G1CollectedHeap::heap()->max_capacity()); 1267 } 1268 } 1269 1270 void G1CollectorPolicy::update_ihop_prediction(double mutator_time_s, 1271 size_t mutator_alloc_bytes, 1272 size_t young_gen_size) { 1273 // Always try to update IHOP prediction. Even evacuation failures give information 1274 // about e.g. whether to start IHOP earlier next time. 1275 1276 // Avoid using really small application times that might create samples with 1277 // very high or very low values. They may be caused by e.g. back-to-back gcs. 1278 double const min_valid_time = 1e-6; 1279 1280 bool report = false; 1281 1282 double marking_to_mixed_time = -1.0; 1283 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) { 1284 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time(); 1285 assert(marking_to_mixed_time > 0.0, 1286 "Initial mark to mixed time must be larger than zero but is %.3f", 1287 marking_to_mixed_time); 1288 if (marking_to_mixed_time > min_valid_time) { 1289 _ihop_control->update_marking_length(marking_to_mixed_time); 1290 report = true; 1291 } 1292 } 1293 1294 // As an approximation for the young gc promotion rates during marking we use 1295 // all of them. In many applications there are only a few if any young gcs during 1296 // marking, which makes any prediction useless. This increases the accuracy of the 1297 // prediction. 1298 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) { 1299 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1300 report = true; 1301 } 1302 1303 if (report) { 1304 report_ihop_statistics(); 1305 } 1306 } 1307 1308 void G1CollectorPolicy::report_ihop_statistics() { 1309 _ihop_control->print(); 1310 } 1311 1312 void G1CollectorPolicy::print_phases() { 1313 phase_times()->print(); 1314 } 1315 1316 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1317 double update_rs_processed_buffers, 1318 double goal_ms) { 1319 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1320 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1321 1322 if (G1UseAdaptiveConcRefinement) { 1323 const int k_gy = 3, k_gr = 6; 1324 const double inc_k = 1.1, dec_k = 0.9; 1325 1326 size_t g = cg1r->green_zone(); 1327 if (update_rs_time > goal_ms) { 1328 g = (size_t)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1329 } else { 1330 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1331 g = (size_t)MAX2(g * inc_k, g + 1.0); 1332 } 1333 } 1334 // Change the refinement threads params 1335 cg1r->set_green_zone(g); 1336 cg1r->set_yellow_zone(g * k_gy); 1337 cg1r->set_red_zone(g * k_gr); 1338 cg1r->reinitialize_threads(); 1339 1340 size_t processing_threshold_delta = MAX2<size_t>(cg1r->green_zone() * _predictor.sigma(), 1); 1341 size_t processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1342 cg1r->yellow_zone()); 1343 // Change the barrier params 1344 dcqs.set_process_completed_threshold(processing_threshold); 1345 dcqs.set_max_completed_queue(cg1r->red_zone()); 1346 } 1347 1348 size_t curr_queue_size = dcqs.completed_buffers_num(); 1349 if (curr_queue_size >= cg1r->yellow_zone()) { 1350 dcqs.set_completed_queue_padding(curr_queue_size); 1351 } else { 1352 dcqs.set_completed_queue_padding(0); 1353 } 1354 dcqs.notify_if_necessary(); 1355 } 1356 1357 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1358 return get_new_size_prediction(_rs_length_diff_seq); 1359 } 1360 1361 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1362 return get_new_prediction(_alloc_rate_ms_seq); 1363 } 1364 1365 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1366 return get_new_prediction(_cost_per_card_ms_seq); 1367 } 1368 1369 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1370 return get_new_prediction(_cost_scan_hcc_seq); 1371 } 1372 1373 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1374 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1375 } 1376 1377 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1378 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1379 } 1380 1381 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1382 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1383 return predict_young_cards_per_entry_ratio(); 1384 } else { 1385 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1386 } 1387 } 1388 1389 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1390 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1391 } 1392 1393 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1394 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1395 } 1396 1397 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1398 if (collector_state()->gcs_are_young()) { 1399 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1400 } else { 1401 return predict_mixed_rs_scan_time_ms(card_num); 1402 } 1403 } 1404 1405 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1406 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1407 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1408 } else { 1409 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1410 } 1411 } 1412 1413 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1414 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1415 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1416 } else { 1417 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1418 } 1419 } 1420 1421 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1422 if (collector_state()->during_concurrent_mark()) { 1423 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1424 } else { 1425 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1426 } 1427 } 1428 1429 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1430 return get_new_prediction(_constant_other_time_ms_seq); 1431 } 1432 1433 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1434 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1435 } 1436 1437 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1438 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1439 } 1440 1441 double G1CollectorPolicy::predict_remark_time_ms() const { 1442 return get_new_prediction(_concurrent_mark_remark_times_ms); 1443 } 1444 1445 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1446 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1447 } 1448 1449 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1450 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1451 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1452 double pred = get_new_prediction(seq); 1453 if (pred > 1.0) { 1454 pred = 1.0; 1455 } 1456 return pred; 1457 } 1458 1459 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1460 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1461 } 1462 1463 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1464 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1465 } 1466 1467 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1468 size_t scanned_cards) const { 1469 return 1470 predict_rs_update_time_ms(pending_cards) + 1471 predict_rs_scan_time_ms(scanned_cards) + 1472 predict_constant_other_time_ms(); 1473 } 1474 1475 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1476 size_t rs_length = predict_rs_length_diff(); 1477 size_t card_num; 1478 if (collector_state()->gcs_are_young()) { 1479 card_num = predict_young_card_num(rs_length); 1480 } else { 1481 card_num = predict_non_young_card_num(rs_length); 1482 } 1483 return predict_base_elapsed_time_ms(pending_cards, card_num); 1484 } 1485 1486 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1487 size_t bytes_to_copy; 1488 if (hr->is_marked()) 1489 bytes_to_copy = hr->max_live_bytes(); 1490 else { 1491 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1492 int age = hr->age_in_surv_rate_group(); 1493 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1494 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1495 } 1496 return bytes_to_copy; 1497 } 1498 1499 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1500 bool for_young_gc) const { 1501 size_t rs_length = hr->rem_set()->occupied(); 1502 size_t card_num; 1503 1504 // Predicting the number of cards is based on which type of GC 1505 // we're predicting for. 1506 if (for_young_gc) { 1507 card_num = predict_young_card_num(rs_length); 1508 } else { 1509 card_num = predict_non_young_card_num(rs_length); 1510 } 1511 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1512 1513 double region_elapsed_time_ms = 1514 predict_rs_scan_time_ms(card_num) + 1515 predict_object_copy_time_ms(bytes_to_copy); 1516 1517 // The prediction of the "other" time for this region is based 1518 // upon the region type and NOT the GC type. 1519 if (hr->is_young()) { 1520 region_elapsed_time_ms += predict_young_other_time_ms(1); 1521 } else { 1522 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1523 } 1524 return region_elapsed_time_ms; 1525 } 1526 1527 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1528 uint survivor_cset_region_length) { 1529 _eden_cset_region_length = eden_cset_region_length; 1530 _survivor_cset_region_length = survivor_cset_region_length; 1531 _old_cset_region_length = 0; 1532 } 1533 1534 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1535 _recorded_rs_lengths = rs_lengths; 1536 } 1537 1538 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1539 double elapsed_ms) { 1540 _recent_gc_times_ms->add(elapsed_ms); 1541 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1542 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1543 } 1544 1545 void G1CollectorPolicy::clear_ratio_check_data() { 1546 _ratio_over_threshold_count = 0; 1547 _ratio_over_threshold_sum = 0.0; 1548 _pauses_since_start = 0; 1549 } 1550 1551 size_t G1CollectorPolicy::expansion_amount() { 1552 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1553 double last_gc_overhead = _last_pause_time_ratio * 100.0; 1554 double threshold = _gc_overhead_perc; 1555 size_t expand_bytes = 0; 1556 1557 // If the heap is at less than half its maximum size, scale the threshold down, 1558 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand, 1559 // though the scaling code will likely keep the increase small. 1560 if (_g1->capacity() <= _g1->max_capacity() / 2) { 1561 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2); 1562 threshold = MAX2(threshold, 1.0); 1563 } 1564 1565 // If the last GC time ratio is over the threshold, increment the count of 1566 // times it has been exceeded, and add this ratio to the sum of exceeded 1567 // ratios. 1568 if (last_gc_overhead > threshold) { 1569 _ratio_over_threshold_count++; 1570 _ratio_over_threshold_sum += last_gc_overhead; 1571 } 1572 1573 // Check if we've had enough GC time ratio checks that were over the 1574 // threshold to trigger an expansion. We'll also expand if we've 1575 // reached the end of the history buffer and the average of all entries 1576 // is still over the threshold. This indicates a smaller number of GCs were 1577 // long enough to make the average exceed the threshold. 1578 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics; 1579 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) || 1580 (filled_history_buffer && (recent_gc_overhead > threshold))) { 1581 size_t min_expand_bytes = HeapRegion::GrainBytes; 1582 size_t reserved_bytes = _g1->max_capacity(); 1583 size_t committed_bytes = _g1->capacity(); 1584 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1585 size_t expand_bytes_via_pct = 1586 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1587 double scale_factor = 1.0; 1588 1589 // If the current size is less than 1/4 of the Initial heap size, expand 1590 // by half of the delta between the current and Initial sizes. IE, grow 1591 // back quickly. 1592 // 1593 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of 1594 // the available expansion space, whichever is smaller, as the base 1595 // expansion size. Then possibly scale this size according to how much the 1596 // threshold has (on average) been exceeded by. If the delta is small 1597 // (less than the StartScaleDownAt value), scale the size down linearly, but 1598 // not by less than MinScaleDownFactor. If the delta is large (greater than 1599 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor 1600 // times the base size. The scaling will be linear in the range from 1601 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words, 1602 // ScaleUpRange sets the rate of scaling up. 1603 if (committed_bytes < InitialHeapSize / 4) { 1604 expand_bytes = (InitialHeapSize - committed_bytes) / 2; 1605 } else { 1606 double const MinScaleDownFactor = 0.2; 1607 double const MaxScaleUpFactor = 2; 1608 double const StartScaleDownAt = _gc_overhead_perc; 1609 double const StartScaleUpAt = _gc_overhead_perc * 1.5; 1610 double const ScaleUpRange = _gc_overhead_perc * 2.0; 1611 1612 double ratio_delta; 1613 if (filled_history_buffer) { 1614 ratio_delta = recent_gc_overhead - threshold; 1615 } else { 1616 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold; 1617 } 1618 1619 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1620 if (ratio_delta < StartScaleDownAt) { 1621 scale_factor = ratio_delta / StartScaleDownAt; 1622 scale_factor = MAX2(scale_factor, MinScaleDownFactor); 1623 } else if (ratio_delta > StartScaleUpAt) { 1624 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange); 1625 scale_factor = MIN2(scale_factor, MaxScaleUpFactor); 1626 } 1627 } 1628 1629 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) " 1630 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)", 1631 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100); 1632 1633 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor); 1634 1635 // Ensure the expansion size is at least the minimum growth amount 1636 // and at most the remaining uncommitted byte size. 1637 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1638 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1639 1640 clear_ratio_check_data(); 1641 } else { 1642 // An expansion was not triggered. If we've started counting, increment 1643 // the number of checks we've made in the current window. If we've 1644 // reached the end of the window without resizing, clear the counters to 1645 // start again the next time we see a ratio above the threshold. 1646 if (_ratio_over_threshold_count > 0) { 1647 _pauses_since_start++; 1648 if (_pauses_since_start > NumPrevPausesForHeuristics) { 1649 clear_ratio_check_data(); 1650 } 1651 } 1652 } 1653 1654 return expand_bytes; 1655 } 1656 1657 void G1CollectorPolicy::print_tracing_info() const { 1658 _trace_young_gen_time_data.print(); 1659 _trace_old_gen_time_data.print(); 1660 } 1661 1662 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1663 #ifndef PRODUCT 1664 _short_lived_surv_rate_group->print_surv_rate_summary(); 1665 // add this call for any other surv rate groups 1666 #endif // PRODUCT 1667 } 1668 1669 bool G1CollectorPolicy::is_young_list_full() const { 1670 uint young_list_length = _g1->young_list()->length(); 1671 uint young_list_target_length = _young_list_target_length; 1672 return young_list_length >= young_list_target_length; 1673 } 1674 1675 bool G1CollectorPolicy::can_expand_young_list() const { 1676 uint young_list_length = _g1->young_list()->length(); 1677 uint young_list_max_length = _young_list_max_length; 1678 return young_list_length < young_list_max_length; 1679 } 1680 1681 bool G1CollectorPolicy::adaptive_young_list_length() const { 1682 return _young_gen_sizer->adaptive_young_list_length(); 1683 } 1684 1685 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1686 uint expansion_region_num = 0; 1687 if (GCLockerEdenExpansionPercent > 0) { 1688 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1689 double expansion_region_num_d = perc * (double) _young_list_target_length; 1690 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1691 // less than 1.0) we'll get 1. 1692 expansion_region_num = (uint) ceil(expansion_region_num_d); 1693 } else { 1694 assert(expansion_region_num == 0, "sanity"); 1695 } 1696 _young_list_max_length = _young_list_target_length + expansion_region_num; 1697 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1698 } 1699 1700 // Calculates survivor space parameters. 1701 void G1CollectorPolicy::update_survivors_policy() { 1702 double max_survivor_regions_d = 1703 (double) _young_list_target_length / (double) SurvivorRatio; 1704 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1705 // smaller than 1.0) we'll get 1. 1706 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1707 1708 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1709 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1710 } 1711 1712 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1713 // We actually check whether we are marking here and not if we are in a 1714 // reclamation phase. This means that we will schedule a concurrent mark 1715 // even while we are still in the process of reclaiming memory. 1716 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1717 if (!during_cycle) { 1718 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause)); 1719 collector_state()->set_initiate_conc_mark_if_possible(true); 1720 return true; 1721 } else { 1722 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause)); 1723 return false; 1724 } 1725 } 1726 1727 void G1CollectorPolicy::initiate_conc_mark() { 1728 collector_state()->set_during_initial_mark_pause(true); 1729 collector_state()->set_initiate_conc_mark_if_possible(false); 1730 } 1731 1732 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1733 // We are about to decide on whether this pause will be an 1734 // initial-mark pause. 1735 1736 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1737 // will set it here if we have to. However, it should be cleared by 1738 // the end of the pause (it's only set for the duration of an 1739 // initial-mark pause). 1740 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1741 1742 if (collector_state()->initiate_conc_mark_if_possible()) { 1743 // We had noticed on a previous pause that the heap occupancy has 1744 // gone over the initiating threshold and we should start a 1745 // concurrent marking cycle. So we might initiate one. 1746 1747 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1748 // Initiate a new initial mark if there is no marking or reclamation going on. 1749 initiate_conc_mark(); 1750 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1751 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) { 1752 // Initiate a user requested initial mark. An initial mark must be young only 1753 // GC, so the collector state must be updated to reflect this. 1754 collector_state()->set_gcs_are_young(true); 1755 collector_state()->set_last_young_gc(false); 1756 1757 abort_time_to_mixed_tracking(); 1758 initiate_conc_mark(); 1759 log_debug(gc, ergo)("Initiate concurrent cycle (user requested concurrent cycle)"); 1760 } else { 1761 // The concurrent marking thread is still finishing up the 1762 // previous cycle. If we start one right now the two cycles 1763 // overlap. In particular, the concurrent marking thread might 1764 // be in the process of clearing the next marking bitmap (which 1765 // we will use for the next cycle if we start one). Starting a 1766 // cycle now will be bad given that parts of the marking 1767 // information might get cleared by the marking thread. And we 1768 // cannot wait for the marking thread to finish the cycle as it 1769 // periodically yields while clearing the next marking bitmap 1770 // and, if it's in a yield point, it's waiting for us to 1771 // finish. So, at this point we will not start a cycle and we'll 1772 // let the concurrent marking thread complete the last one. 1773 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1774 } 1775 } 1776 } 1777 1778 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1779 G1CollectedHeap* _g1h; 1780 CSetChooserParUpdater _cset_updater; 1781 1782 public: 1783 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1784 uint chunk_size) : 1785 _g1h(G1CollectedHeap::heap()), 1786 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1787 1788 bool doHeapRegion(HeapRegion* r) { 1789 // Do we have any marking information for this region? 1790 if (r->is_marked()) { 1791 // We will skip any region that's currently used as an old GC 1792 // alloc region (we should not consider those for collection 1793 // before we fill them up). 1794 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1795 _cset_updater.add_region(r); 1796 } 1797 } 1798 return false; 1799 } 1800 }; 1801 1802 class ParKnownGarbageTask: public AbstractGangTask { 1803 CollectionSetChooser* _hrSorted; 1804 uint _chunk_size; 1805 G1CollectedHeap* _g1; 1806 HeapRegionClaimer _hrclaimer; 1807 1808 public: 1809 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1810 AbstractGangTask("ParKnownGarbageTask"), 1811 _hrSorted(hrSorted), _chunk_size(chunk_size), 1812 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1813 1814 void work(uint worker_id) { 1815 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1816 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1817 } 1818 }; 1819 1820 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1821 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1822 const uint overpartition_factor = 4; 1823 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1824 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1825 } 1826 1827 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1828 cset_chooser()->clear(); 1829 1830 WorkGang* workers = _g1->workers(); 1831 uint n_workers = workers->active_workers(); 1832 1833 uint n_regions = _g1->num_regions(); 1834 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1835 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1836 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers); 1837 workers->run_task(&par_known_garbage_task); 1838 1839 cset_chooser()->sort_regions(); 1840 1841 double end_sec = os::elapsedTime(); 1842 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1843 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1844 _prev_collection_pause_end_ms += elapsed_time_ms; 1845 1846 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1847 } 1848 1849 // Add the heap region at the head of the non-incremental collection set 1850 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1851 assert(_inc_cset_build_state == Active, "Precondition"); 1852 assert(hr->is_old(), "the region should be old"); 1853 1854 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1855 _g1->register_old_region_with_cset(hr); 1856 hr->set_next_in_collection_set(_collection_set); 1857 _collection_set = hr; 1858 _collection_set_bytes_used_before += hr->used(); 1859 size_t rs_length = hr->rem_set()->occupied(); 1860 _recorded_rs_lengths += rs_length; 1861 _old_cset_region_length += 1; 1862 } 1863 1864 // Initialize the per-collection-set information 1865 void G1CollectorPolicy::start_incremental_cset_building() { 1866 assert(_inc_cset_build_state == Inactive, "Precondition"); 1867 1868 _inc_cset_head = NULL; 1869 _inc_cset_tail = NULL; 1870 _inc_cset_bytes_used_before = 0; 1871 1872 _inc_cset_max_finger = 0; 1873 _inc_cset_recorded_rs_lengths = 0; 1874 _inc_cset_recorded_rs_lengths_diffs = 0; 1875 _inc_cset_predicted_elapsed_time_ms = 0.0; 1876 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1877 _inc_cset_build_state = Active; 1878 } 1879 1880 void G1CollectorPolicy::finalize_incremental_cset_building() { 1881 assert(_inc_cset_build_state == Active, "Precondition"); 1882 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1883 1884 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1885 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1886 // that adds a new region to the CSet. Further updates by the 1887 // concurrent refinement thread that samples the young RSet lengths 1888 // are accumulated in the *_diffs fields. Here we add the diffs to 1889 // the "main" fields. 1890 1891 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1892 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1893 } else { 1894 // This is defensive. The diff should in theory be always positive 1895 // as RSets can only grow between GCs. However, given that we 1896 // sample their size concurrently with other threads updating them 1897 // it's possible that we might get the wrong size back, which 1898 // could make the calculations somewhat inaccurate. 1899 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1900 if (_inc_cset_recorded_rs_lengths >= diffs) { 1901 _inc_cset_recorded_rs_lengths -= diffs; 1902 } else { 1903 _inc_cset_recorded_rs_lengths = 0; 1904 } 1905 } 1906 _inc_cset_predicted_elapsed_time_ms += 1907 _inc_cset_predicted_elapsed_time_ms_diffs; 1908 1909 _inc_cset_recorded_rs_lengths_diffs = 0; 1910 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1911 } 1912 1913 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1914 // This routine is used when: 1915 // * adding survivor regions to the incremental cset at the end of an 1916 // evacuation pause, 1917 // * adding the current allocation region to the incremental cset 1918 // when it is retired, and 1919 // * updating existing policy information for a region in the 1920 // incremental cset via young list RSet sampling. 1921 // Therefore this routine may be called at a safepoint by the 1922 // VM thread, or in-between safepoints by mutator threads (when 1923 // retiring the current allocation region) or a concurrent 1924 // refine thread (RSet sampling). 1925 1926 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1927 size_t used_bytes = hr->used(); 1928 _inc_cset_recorded_rs_lengths += rs_length; 1929 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1930 _inc_cset_bytes_used_before += used_bytes; 1931 1932 // Cache the values we have added to the aggregated information 1933 // in the heap region in case we have to remove this region from 1934 // the incremental collection set, or it is updated by the 1935 // rset sampling code 1936 hr->set_recorded_rs_length(rs_length); 1937 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1938 } 1939 1940 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1941 size_t new_rs_length) { 1942 // Update the CSet information that is dependent on the new RS length 1943 assert(hr->is_young(), "Precondition"); 1944 assert(!SafepointSynchronize::is_at_safepoint(), 1945 "should not be at a safepoint"); 1946 1947 // We could have updated _inc_cset_recorded_rs_lengths and 1948 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1949 // that atomically, as this code is executed by a concurrent 1950 // refinement thread, potentially concurrently with a mutator thread 1951 // allocating a new region and also updating the same fields. To 1952 // avoid the atomic operations we accumulate these updates on two 1953 // separate fields (*_diffs) and we'll just add them to the "main" 1954 // fields at the start of a GC. 1955 1956 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1957 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1958 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1959 1960 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1961 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1962 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1963 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1964 1965 hr->set_recorded_rs_length(new_rs_length); 1966 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1967 } 1968 1969 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1970 assert(hr->is_young(), "invariant"); 1971 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1972 assert(_inc_cset_build_state == Active, "Precondition"); 1973 1974 // We need to clear and set the cached recorded/cached collection set 1975 // information in the heap region here (before the region gets added 1976 // to the collection set). An individual heap region's cached values 1977 // are calculated, aggregated with the policy collection set info, 1978 // and cached in the heap region here (initially) and (subsequently) 1979 // by the Young List sampling code. 1980 1981 size_t rs_length = hr->rem_set()->occupied(); 1982 add_to_incremental_cset_info(hr, rs_length); 1983 1984 HeapWord* hr_end = hr->end(); 1985 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1986 1987 assert(!hr->in_collection_set(), "invariant"); 1988 _g1->register_young_region_with_cset(hr); 1989 assert(hr->next_in_collection_set() == NULL, "invariant"); 1990 } 1991 1992 // Add the region at the RHS of the incremental cset 1993 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1994 // We should only ever be appending survivors at the end of a pause 1995 assert(hr->is_survivor(), "Logic"); 1996 1997 // Do the 'common' stuff 1998 add_region_to_incremental_cset_common(hr); 1999 2000 // Now add the region at the right hand side 2001 if (_inc_cset_tail == NULL) { 2002 assert(_inc_cset_head == NULL, "invariant"); 2003 _inc_cset_head = hr; 2004 } else { 2005 _inc_cset_tail->set_next_in_collection_set(hr); 2006 } 2007 _inc_cset_tail = hr; 2008 } 2009 2010 // Add the region to the LHS of the incremental cset 2011 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2012 // Survivors should be added to the RHS at the end of a pause 2013 assert(hr->is_eden(), "Logic"); 2014 2015 // Do the 'common' stuff 2016 add_region_to_incremental_cset_common(hr); 2017 2018 // Add the region at the left hand side 2019 hr->set_next_in_collection_set(_inc_cset_head); 2020 if (_inc_cset_head == NULL) { 2021 assert(_inc_cset_tail == NULL, "Invariant"); 2022 _inc_cset_tail = hr; 2023 } 2024 _inc_cset_head = hr; 2025 } 2026 2027 #ifndef PRODUCT 2028 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2029 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2030 2031 st->print_cr("\nCollection_set:"); 2032 HeapRegion* csr = list_head; 2033 while (csr != NULL) { 2034 HeapRegion* next = csr->next_in_collection_set(); 2035 assert(csr->in_collection_set(), "bad CS"); 2036 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 2037 HR_FORMAT_PARAMS(csr), 2038 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 2039 csr->age_in_surv_rate_group_cond()); 2040 csr = next; 2041 } 2042 } 2043 #endif // !PRODUCT 2044 2045 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 2046 // Returns the given amount of reclaimable bytes (that represents 2047 // the amount of reclaimable space still to be collected) as a 2048 // percentage of the current heap capacity. 2049 size_t capacity_bytes = _g1->capacity(); 2050 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2051 } 2052 2053 void G1CollectorPolicy::maybe_start_marking() { 2054 if (need_to_start_conc_mark("end of GC")) { 2055 // Note: this might have already been set, if during the last 2056 // pause we decided to start a cycle but at the beginning of 2057 // this pause we decided to postpone it. That's OK. 2058 collector_state()->set_initiate_conc_mark_if_possible(true); 2059 } 2060 } 2061 2062 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const { 2063 assert(!collector_state()->full_collection(), "must be"); 2064 if (collector_state()->during_initial_mark_pause()) { 2065 assert(collector_state()->last_gc_was_young(), "must be"); 2066 assert(!collector_state()->last_young_gc(), "must be"); 2067 return InitialMarkGC; 2068 } else if (collector_state()->last_young_gc()) { 2069 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2070 assert(collector_state()->last_gc_was_young(), "must be"); 2071 return LastYoungGC; 2072 } else if (!collector_state()->last_gc_was_young()) { 2073 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2074 assert(!collector_state()->last_young_gc(), "must be"); 2075 return MixedGC; 2076 } else { 2077 assert(collector_state()->last_gc_was_young(), "must be"); 2078 assert(!collector_state()->during_initial_mark_pause(), "must be"); 2079 assert(!collector_state()->last_young_gc(), "must be"); 2080 return YoungOnlyGC; 2081 } 2082 } 2083 2084 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) { 2085 // Manage the MMU tracker. For some reason it ignores Full GCs. 2086 if (kind != FullGC) { 2087 _mmu_tracker->add_pause(start, end); 2088 } 2089 // Manage the mutator time tracking from initial mark to first mixed gc. 2090 switch (kind) { 2091 case FullGC: 2092 abort_time_to_mixed_tracking(); 2093 break; 2094 case Cleanup: 2095 case Remark: 2096 case YoungOnlyGC: 2097 case LastYoungGC: 2098 _initial_mark_to_mixed.add_pause(end - start); 2099 break; 2100 case InitialMarkGC: 2101 _initial_mark_to_mixed.record_initial_mark_end(end); 2102 break; 2103 case MixedGC: 2104 _initial_mark_to_mixed.record_mixed_gc_start(start); 2105 break; 2106 default: 2107 ShouldNotReachHere(); 2108 } 2109 } 2110 2111 void G1CollectorPolicy::abort_time_to_mixed_tracking() { 2112 _initial_mark_to_mixed.reset(); 2113 } 2114 2115 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2116 const char* false_action_str) const { 2117 if (cset_chooser()->is_empty()) { 2118 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 2119 return false; 2120 } 2121 2122 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 2123 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2124 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2125 double threshold = (double) G1HeapWastePercent; 2126 if (reclaimable_perc <= threshold) { 2127 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2128 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2129 return false; 2130 } 2131 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 2132 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2133 return true; 2134 } 2135 2136 uint G1CollectorPolicy::calc_min_old_cset_length() const { 2137 // The min old CSet region bound is based on the maximum desired 2138 // number of mixed GCs after a cycle. I.e., even if some old regions 2139 // look expensive, we should add them to the CSet anyway to make 2140 // sure we go through the available old regions in no more than the 2141 // maximum desired number of mixed GCs. 2142 // 2143 // The calculation is based on the number of marked regions we added 2144 // to the CSet chooser in the first place, not how many remain, so 2145 // that the result is the same during all mixed GCs that follow a cycle. 2146 2147 const size_t region_num = (size_t) cset_chooser()->length(); 2148 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2149 size_t result = region_num / gc_num; 2150 // emulate ceiling 2151 if (result * gc_num < region_num) { 2152 result += 1; 2153 } 2154 return (uint) result; 2155 } 2156 2157 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2158 // The max old CSet region bound is based on the threshold expressed 2159 // as a percentage of the heap size. I.e., it should bound the 2160 // number of old regions added to the CSet irrespective of how many 2161 // of them are available. 2162 2163 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2164 const size_t region_num = g1h->num_regions(); 2165 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2166 size_t result = region_num * perc / 100; 2167 // emulate ceiling 2168 if (100 * result < region_num * perc) { 2169 result += 1; 2170 } 2171 return (uint) result; 2172 } 2173 2174 2175 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2176 double young_start_time_sec = os::elapsedTime(); 2177 2178 YoungList* young_list = _g1->young_list(); 2179 finalize_incremental_cset_building(); 2180 2181 guarantee(target_pause_time_ms > 0.0, 2182 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2183 guarantee(_collection_set == NULL, "Precondition"); 2184 2185 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2186 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2187 2188 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", 2189 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2190 2191 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2192 2193 if (collector_state()->last_gc_was_young()) { 2194 _trace_young_gen_time_data.increment_young_collection_count(); 2195 } else { 2196 _trace_young_gen_time_data.increment_mixed_collection_count(); 2197 } 2198 2199 // The young list is laid with the survivor regions from the previous 2200 // pause are appended to the RHS of the young list, i.e. 2201 // [Newly Young Regions ++ Survivors from last pause]. 2202 2203 uint survivor_region_length = young_list->survivor_length(); 2204 uint eden_region_length = young_list->eden_length(); 2205 init_cset_region_lengths(eden_region_length, survivor_region_length); 2206 2207 HeapRegion* hr = young_list->first_survivor_region(); 2208 while (hr != NULL) { 2209 assert(hr->is_survivor(), "badly formed young list"); 2210 // There is a convention that all the young regions in the CSet 2211 // are tagged as "eden", so we do this for the survivors here. We 2212 // use the special set_eden_pre_gc() as it doesn't check that the 2213 // region is free (which is not the case here). 2214 hr->set_eden_pre_gc(); 2215 hr = hr->get_next_young_region(); 2216 } 2217 2218 // Clear the fields that point to the survivor list - they are all young now. 2219 young_list->clear_survivors(); 2220 2221 _collection_set = _inc_cset_head; 2222 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2223 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2224 2225 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", 2226 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms); 2227 2228 // The number of recorded young regions is the incremental 2229 // collection set's current size 2230 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2231 2232 double young_end_time_sec = os::elapsedTime(); 2233 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2234 2235 return time_remaining_ms; 2236 } 2237 2238 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2239 double non_young_start_time_sec = os::elapsedTime(); 2240 double predicted_old_time_ms = 0.0; 2241 2242 2243 if (!collector_state()->gcs_are_young()) { 2244 cset_chooser()->verify(); 2245 const uint min_old_cset_length = calc_min_old_cset_length(); 2246 const uint max_old_cset_length = calc_max_old_cset_length(); 2247 2248 uint expensive_region_num = 0; 2249 bool check_time_remaining = adaptive_young_list_length(); 2250 2251 HeapRegion* hr = cset_chooser()->peek(); 2252 while (hr != NULL) { 2253 if (old_cset_region_length() >= max_old_cset_length) { 2254 // Added maximum number of old regions to the CSet. 2255 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions", 2256 old_cset_region_length(), max_old_cset_length); 2257 break; 2258 } 2259 2260 2261 // Stop adding regions if the remaining reclaimable space is 2262 // not above G1HeapWastePercent. 2263 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes(); 2264 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2265 double threshold = (double) G1HeapWastePercent; 2266 if (reclaimable_perc <= threshold) { 2267 // We've added enough old regions that the amount of uncollected 2268 // reclaimable space is at or below the waste threshold. Stop 2269 // adding old regions to the CSet. 2270 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). " 2271 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%", 2272 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2273 break; 2274 } 2275 2276 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2277 if (check_time_remaining) { 2278 if (predicted_time_ms > time_remaining_ms) { 2279 // Too expensive for the current CSet. 2280 2281 if (old_cset_region_length() >= min_old_cset_length) { 2282 // We have added the minimum number of old regions to the CSet, 2283 // we are done with this CSet. 2284 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). " 2285 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions", 2286 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length); 2287 break; 2288 } 2289 2290 // We'll add it anyway given that we haven't reached the 2291 // minimum number of old regions. 2292 expensive_region_num += 1; 2293 } 2294 } else { 2295 if (old_cset_region_length() >= min_old_cset_length) { 2296 // In the non-auto-tuning case, we'll finish adding regions 2297 // to the CSet if we reach the minimum. 2298 2299 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions", 2300 old_cset_region_length(), min_old_cset_length); 2301 break; 2302 } 2303 } 2304 2305 // We will add this region to the CSet. 2306 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2307 predicted_old_time_ms += predicted_time_ms; 2308 cset_chooser()->pop(); // already have region via peek() 2309 _g1->old_set_remove(hr); 2310 add_old_region_to_cset(hr); 2311 2312 hr = cset_chooser()->peek(); 2313 } 2314 if (hr == NULL) { 2315 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)"); 2316 } 2317 2318 if (expensive_region_num > 0) { 2319 // We print the information once here at the end, predicated on 2320 // whether we added any apparently expensive regions or not, to 2321 // avoid generating output per region. 2322 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)." 2323 "old: %u regions, expensive: %u regions, min: %u regions, remaining time: %1.2fms", 2324 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms); 2325 } 2326 2327 cset_chooser()->verify(); 2328 } 2329 2330 stop_incremental_cset_building(); 2331 2332 log_debug(gc, ergo, cset)("Finish choosing CSet. old: %u regions, predicted old region time: %1.2fms, time remaining: %1.2f", 2333 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms); 2334 2335 double non_young_end_time_sec = os::elapsedTime(); 2336 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2337 } 2338 2339 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2340 if(TraceYoungGenTime) { 2341 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2342 } 2343 } 2344 2345 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2346 if(TraceYoungGenTime) { 2347 _all_yield_times_ms.add(yield_time_ms); 2348 } 2349 } 2350 2351 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2352 if(TraceYoungGenTime) { 2353 _total.add(pause_time_ms); 2354 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2355 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2356 _parallel.add(phase_times->cur_collection_par_time_ms()); 2357 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2358 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2359 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2360 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2361 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2362 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2363 2364 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2365 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2366 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2367 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2368 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2369 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2370 2371 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2372 _parallel_other.add(parallel_other_time); 2373 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2374 } 2375 } 2376 2377 void TraceYoungGenTimeData::increment_young_collection_count() { 2378 if(TraceYoungGenTime) { 2379 ++_young_pause_num; 2380 } 2381 } 2382 2383 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2384 if(TraceYoungGenTime) { 2385 ++_mixed_pause_num; 2386 } 2387 } 2388 2389 void TraceYoungGenTimeData::print_summary(const char* str, 2390 const NumberSeq* seq) const { 2391 double sum = seq->sum(); 2392 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2393 str, sum / 1000.0, seq->avg()); 2394 } 2395 2396 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2397 const NumberSeq* seq) const { 2398 print_summary(str, seq); 2399 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2400 "(num", seq->num(), seq->sd(), seq->maximum()); 2401 } 2402 2403 void TraceYoungGenTimeData::print() const { 2404 if (!TraceYoungGenTime) { 2405 return; 2406 } 2407 2408 tty->print_cr("ALL PAUSES"); 2409 print_summary_sd(" Total", &_total); 2410 tty->cr(); 2411 tty->cr(); 2412 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2413 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2414 tty->cr(); 2415 2416 tty->print_cr("EVACUATION PAUSES"); 2417 2418 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2419 tty->print_cr("none"); 2420 } else { 2421 print_summary_sd(" Evacuation Pauses", &_total); 2422 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2423 print_summary(" Parallel Time", &_parallel); 2424 print_summary(" Ext Root Scanning", &_ext_root_scan); 2425 print_summary(" SATB Filtering", &_satb_filtering); 2426 print_summary(" Update RS", &_update_rs); 2427 print_summary(" Scan RS", &_scan_rs); 2428 print_summary(" Object Copy", &_obj_copy); 2429 print_summary(" Termination", &_termination); 2430 print_summary(" Parallel Other", &_parallel_other); 2431 print_summary(" Clear CT", &_clear_ct); 2432 print_summary(" Other", &_other); 2433 } 2434 tty->cr(); 2435 2436 tty->print_cr("MISC"); 2437 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2438 print_summary_sd(" Yields", &_all_yield_times_ms); 2439 } 2440 2441 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2442 if (TraceOldGenTime) { 2443 _all_full_gc_times.add(full_gc_time_ms); 2444 } 2445 } 2446 2447 void TraceOldGenTimeData::print() const { 2448 if (!TraceOldGenTime) { 2449 return; 2450 } 2451 2452 if (_all_full_gc_times.num() > 0) { 2453 tty->print("\n%4d full_gcs: total time = %8.2f s", 2454 _all_full_gc_times.num(), 2455 _all_full_gc_times.sum() / 1000.0); 2456 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2457 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2458 _all_full_gc_times.sd(), 2459 _all_full_gc_times.maximum()); 2460 } 2461 }