1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/concurrentG1Refine.hpp" 27 #include "gc/g1/concurrentMark.hpp" 28 #include "gc/g1/concurrentMarkThread.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorPolicy.hpp" 31 #include "gc/g1/g1ErgoVerbose.hpp" 32 #include "gc/g1/g1GCPhaseTimes.hpp" 33 #include "gc/g1/g1Log.hpp" 34 #include "gc/g1/heapRegion.inline.hpp" 35 #include "gc/g1/heapRegionRemSet.hpp" 36 #include "gc/shared/gcPolicyCounters.hpp" 37 #include "runtime/arguments.hpp" 38 #include "runtime/java.hpp" 39 #include "runtime/mutexLocker.hpp" 40 #include "utilities/debug.hpp" 41 42 // Different defaults for different number of GC threads 43 // They were chosen by running GCOld and SPECjbb on debris with different 44 // numbers of GC threads and choosing them based on the results 45 46 // all the same 47 static double rs_length_diff_defaults[] = { 48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 49 }; 50 51 static double cost_per_card_ms_defaults[] = { 52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 53 }; 54 55 // all the same 56 static double young_cards_per_entry_ratio_defaults[] = { 57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 58 }; 59 60 static double cost_per_entry_ms_defaults[] = { 61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 62 }; 63 64 static double cost_per_byte_ms_defaults[] = { 65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 66 }; 67 68 // these should be pretty consistent 69 static double constant_other_time_ms_defaults[] = { 70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 71 }; 72 73 74 static double young_other_cost_per_region_ms_defaults[] = { 75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 76 }; 77 78 static double non_young_other_cost_per_region_ms_defaults[] = { 79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 80 }; 81 82 G1CollectorPolicy::G1CollectorPolicy() : 83 _predictor(G1ConfidencePercent / 100.0), 84 _parallel_gc_threads(ParallelGCThreads), 85 86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 87 _stop_world_start(0.0), 88 89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 93 _prev_collection_pause_end_ms(0.0), 94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _non_young_other_cost_per_region_ms_seq( 106 new TruncatedSeq(TruncatedSeqLength)), 107 108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 110 111 _pause_time_target_ms((double) MaxGCPauseMillis), 112 113 _recent_prev_end_times_for_all_gcs_sec( 114 new TruncatedSeq(NumPrevPausesForHeuristics)), 115 116 _recent_avg_pause_time_ratio(0.0), 117 _rs_lengths_prediction(0), 118 _max_survivor_regions(0), 119 120 _eden_used_bytes_before_gc(0), 121 _survivor_used_bytes_before_gc(0), 122 _heap_used_bytes_before_gc(0), 123 _metaspace_used_bytes_before_gc(0), 124 _eden_capacity_bytes_before_gc(0), 125 _heap_capacity_bytes_before_gc(0), 126 127 _eden_cset_region_length(0), 128 _survivor_cset_region_length(0), 129 _old_cset_region_length(0), 130 131 _collection_set(NULL), 132 _collection_set_bytes_used_before(0), 133 134 // Incremental CSet attributes 135 _inc_cset_build_state(Inactive), 136 _inc_cset_head(NULL), 137 _inc_cset_tail(NULL), 138 _inc_cset_bytes_used_before(0), 139 _inc_cset_max_finger(NULL), 140 _inc_cset_recorded_rs_lengths(0), 141 _inc_cset_recorded_rs_lengths_diffs(0), 142 _inc_cset_predicted_elapsed_time_ms(0.0), 143 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 144 145 // add here any more surv rate groups 146 _recorded_survivor_regions(0), 147 _recorded_survivor_head(NULL), 148 _recorded_survivor_tail(NULL), 149 _survivors_age_table(true), 150 151 _gc_overhead_perc(0.0) { 152 153 // SurvRateGroups below must be initialized after the predictor because they 154 // indirectly use it through this object passed to their constructor. 155 _short_lived_surv_rate_group = 156 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 157 _survivor_surv_rate_group = 158 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 159 160 // Set up the region size and associated fields. Given that the 161 // policy is created before the heap, we have to set this up here, 162 // so it's done as soon as possible. 163 164 // It would have been natural to pass initial_heap_byte_size() and 165 // max_heap_byte_size() to setup_heap_region_size() but those have 166 // not been set up at this point since they should be aligned with 167 // the region size. So, there is a circular dependency here. We base 168 // the region size on the heap size, but the heap size should be 169 // aligned with the region size. To get around this we use the 170 // unaligned values for the heap. 171 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 172 HeapRegionRemSet::setup_remset_size(); 173 174 G1ErgoVerbose::initialize(); 175 if (PrintAdaptiveSizePolicy) { 176 // Currently, we only use a single switch for all the heuristics. 177 G1ErgoVerbose::set_enabled(true); 178 // Given that we don't currently have a verboseness level 179 // parameter, we'll hardcode this to high. This can be easily 180 // changed in the future. 181 G1ErgoVerbose::set_level(ErgoHigh); 182 } else { 183 G1ErgoVerbose::set_enabled(false); 184 } 185 186 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 187 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 188 189 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 190 191 int index = MIN2(_parallel_gc_threads - 1, 7); 192 193 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 194 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 195 _cost_scan_hcc_seq->add(0.0); 196 _young_cards_per_entry_ratio_seq->add( 197 young_cards_per_entry_ratio_defaults[index]); 198 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 199 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 200 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 201 _young_other_cost_per_region_ms_seq->add( 202 young_other_cost_per_region_ms_defaults[index]); 203 _non_young_other_cost_per_region_ms_seq->add( 204 non_young_other_cost_per_region_ms_defaults[index]); 205 206 // Below, we might need to calculate the pause time target based on 207 // the pause interval. When we do so we are going to give G1 maximum 208 // flexibility and allow it to do pauses when it needs to. So, we'll 209 // arrange that the pause interval to be pause time target + 1 to 210 // ensure that a) the pause time target is maximized with respect to 211 // the pause interval and b) we maintain the invariant that pause 212 // time target < pause interval. If the user does not want this 213 // maximum flexibility, they will have to set the pause interval 214 // explicitly. 215 216 // First make sure that, if either parameter is set, its value is 217 // reasonable. 218 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 219 if (MaxGCPauseMillis < 1) { 220 vm_exit_during_initialization("MaxGCPauseMillis should be " 221 "greater than 0"); 222 } 223 } 224 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 225 if (GCPauseIntervalMillis < 1) { 226 vm_exit_during_initialization("GCPauseIntervalMillis should be " 227 "greater than 0"); 228 } 229 } 230 231 // Then, if the pause time target parameter was not set, set it to 232 // the default value. 233 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 234 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 235 // The default pause time target in G1 is 200ms 236 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 237 } else { 238 // We do not allow the pause interval to be set without the 239 // pause time target 240 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 241 "without setting MaxGCPauseMillis"); 242 } 243 } 244 245 // Then, if the interval parameter was not set, set it according to 246 // the pause time target (this will also deal with the case when the 247 // pause time target is the default value). 248 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 249 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 250 } 251 252 // Finally, make sure that the two parameters are consistent. 253 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 254 char buffer[256]; 255 jio_snprintf(buffer, 256, 256 "MaxGCPauseMillis (%u) should be less than " 257 "GCPauseIntervalMillis (%u)", 258 MaxGCPauseMillis, GCPauseIntervalMillis); 259 vm_exit_during_initialization(buffer); 260 } 261 262 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 263 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 264 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 265 266 // start conservatively (around 50ms is about right) 267 _concurrent_mark_remark_times_ms->add(0.05); 268 _concurrent_mark_cleanup_times_ms->add(0.20); 269 _tenuring_threshold = MaxTenuringThreshold; 270 271 assert(GCTimeRatio > 0, 272 "we should have set it to a default value set_g1_gc_flags() " 273 "if a user set it to 0"); 274 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 275 276 uintx reserve_perc = G1ReservePercent; 277 // Put an artificial ceiling on this so that it's not set to a silly value. 278 if (reserve_perc > 50) { 279 reserve_perc = 50; 280 warning("G1ReservePercent is set to a value that is too large, " 281 "it's been updated to " UINTX_FORMAT, reserve_perc); 282 } 283 _reserve_factor = (double) reserve_perc / 100.0; 284 // This will be set when the heap is expanded 285 // for the first time during initialization. 286 _reserve_regions = 0; 287 288 _collectionSetChooser = new CollectionSetChooser(); 289 } 290 291 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 292 return _predictor.get_new_prediction(seq); 293 } 294 295 void G1CollectorPolicy::initialize_alignments() { 296 _space_alignment = HeapRegion::GrainBytes; 297 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 298 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 299 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 300 } 301 302 void G1CollectorPolicy::initialize_flags() { 303 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 304 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 305 } 306 307 if (SurvivorRatio < 1) { 308 vm_exit_during_initialization("Invalid survivor ratio specified"); 309 } 310 CollectorPolicy::initialize_flags(); 311 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 312 } 313 314 void G1CollectorPolicy::post_heap_initialize() { 315 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 316 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 317 if (max_young_size != MaxNewSize) { 318 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 319 } 320 } 321 322 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 323 324 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 325 _min_desired_young_length(0), _max_desired_young_length(0) { 326 if (FLAG_IS_CMDLINE(NewRatio)) { 327 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 328 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 329 } else { 330 _sizer_kind = SizerNewRatio; 331 _adaptive_size = false; 332 return; 333 } 334 } 335 336 if (NewSize > MaxNewSize) { 337 if (FLAG_IS_CMDLINE(MaxNewSize)) { 338 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 339 "A new max generation size of " SIZE_FORMAT "k will be used.", 340 NewSize/K, MaxNewSize/K, NewSize/K); 341 } 342 MaxNewSize = NewSize; 343 } 344 345 if (FLAG_IS_CMDLINE(NewSize)) { 346 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 347 1U); 348 if (FLAG_IS_CMDLINE(MaxNewSize)) { 349 _max_desired_young_length = 350 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 351 1U); 352 _sizer_kind = SizerMaxAndNewSize; 353 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 354 } else { 355 _sizer_kind = SizerNewSizeOnly; 356 } 357 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 358 _max_desired_young_length = 359 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 360 1U); 361 _sizer_kind = SizerMaxNewSizeOnly; 362 } 363 } 364 365 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 366 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 367 return MAX2(1U, default_value); 368 } 369 370 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 371 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 372 return MAX2(1U, default_value); 373 } 374 375 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 376 assert(number_of_heap_regions > 0, "Heap must be initialized"); 377 378 switch (_sizer_kind) { 379 case SizerDefaults: 380 *min_young_length = calculate_default_min_length(number_of_heap_regions); 381 *max_young_length = calculate_default_max_length(number_of_heap_regions); 382 break; 383 case SizerNewSizeOnly: 384 *max_young_length = calculate_default_max_length(number_of_heap_regions); 385 *max_young_length = MAX2(*min_young_length, *max_young_length); 386 break; 387 case SizerMaxNewSizeOnly: 388 *min_young_length = calculate_default_min_length(number_of_heap_regions); 389 *min_young_length = MIN2(*min_young_length, *max_young_length); 390 break; 391 case SizerMaxAndNewSize: 392 // Do nothing. Values set on the command line, don't update them at runtime. 393 break; 394 case SizerNewRatio: 395 *min_young_length = number_of_heap_regions / (NewRatio + 1); 396 *max_young_length = *min_young_length; 397 break; 398 default: 399 ShouldNotReachHere(); 400 } 401 402 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 403 } 404 405 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 406 // We need to pass the desired values because recalculation may not update these 407 // values in some cases. 408 uint temp = _min_desired_young_length; 409 uint result = _max_desired_young_length; 410 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 411 return result; 412 } 413 414 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 415 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 416 &_max_desired_young_length); 417 } 418 419 void G1CollectorPolicy::init() { 420 // Set aside an initial future to_space. 421 _g1 = G1CollectedHeap::heap(); 422 423 assert(Heap_lock->owned_by_self(), "Locking discipline."); 424 425 initialize_gc_policy_counters(); 426 427 if (adaptive_young_list_length()) { 428 _young_list_fixed_length = 0; 429 } else { 430 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 431 } 432 _free_regions_at_end_of_collection = _g1->num_free_regions(); 433 434 update_young_list_max_and_target_length(); 435 // We may immediately start allocating regions and placing them on the 436 // collection set list. Initialize the per-collection set info 437 start_incremental_cset_building(); 438 } 439 440 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 441 phase_times()->note_gc_start(num_active_workers); 442 } 443 444 // Create the jstat counters for the policy. 445 void G1CollectorPolicy::initialize_gc_policy_counters() { 446 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 447 } 448 449 bool G1CollectorPolicy::predict_will_fit(uint young_length, 450 double base_time_ms, 451 uint base_free_regions, 452 double target_pause_time_ms) const { 453 if (young_length >= base_free_regions) { 454 // end condition 1: not enough space for the young regions 455 return false; 456 } 457 458 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 459 size_t bytes_to_copy = 460 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 461 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 462 double young_other_time_ms = predict_young_other_time_ms(young_length); 463 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 464 if (pause_time_ms > target_pause_time_ms) { 465 // end condition 2: prediction is over the target pause time 466 return false; 467 } 468 469 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 470 if ((2.0 /* magic */ * _predictor.sigma()) * bytes_to_copy > free_bytes) { 471 // end condition 3: out-of-space (conservatively!) 472 return false; 473 } 474 475 // success! 476 return true; 477 } 478 479 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 480 // re-calculate the necessary reserve 481 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 482 // We use ceiling so that if reserve_regions_d is > 0.0 (but 483 // smaller than 1.0) we'll get 1. 484 _reserve_regions = (uint) ceil(reserve_regions_d); 485 486 _young_gen_sizer->heap_size_changed(new_number_of_regions); 487 } 488 489 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 490 uint base_min_length) const { 491 uint desired_min_length = 0; 492 if (adaptive_young_list_length()) { 493 if (_alloc_rate_ms_seq->num() > 3) { 494 double now_sec = os::elapsedTime(); 495 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 496 double alloc_rate_ms = predict_alloc_rate_ms(); 497 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 498 } else { 499 // otherwise we don't have enough info to make the prediction 500 } 501 } 502 desired_min_length += base_min_length; 503 // make sure we don't go below any user-defined minimum bound 504 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 505 } 506 507 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 508 // Here, we might want to also take into account any additional 509 // constraints (i.e., user-defined minimum bound). Currently, we 510 // effectively don't set this bound. 511 return _young_gen_sizer->max_desired_young_length(); 512 } 513 514 void G1CollectorPolicy::update_young_list_max_and_target_length() { 515 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 516 } 517 518 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 519 update_young_list_target_length(rs_lengths); 520 update_max_gc_locker_expansion(); 521 } 522 523 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 524 _young_list_target_length = bounded_young_list_target_length(rs_lengths); 525 } 526 527 void G1CollectorPolicy::update_young_list_target_length() { 528 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 529 } 530 531 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths) const { 532 // Calculate the absolute and desired min bounds. 533 534 // This is how many young regions we already have (currently: the survivors). 535 uint base_min_length = recorded_survivor_regions(); 536 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 537 // This is the absolute minimum young length. Ensure that we 538 // will at least have one eden region available for allocation. 539 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 540 // If we shrank the young list target it should not shrink below the current size. 541 desired_min_length = MAX2(desired_min_length, absolute_min_length); 542 // Calculate the absolute and desired max bounds. 543 544 // We will try our best not to "eat" into the reserve. 545 uint absolute_max_length = 0; 546 if (_free_regions_at_end_of_collection > _reserve_regions) { 547 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 548 } 549 uint desired_max_length = calculate_young_list_desired_max_length(); 550 if (desired_max_length > absolute_max_length) { 551 desired_max_length = absolute_max_length; 552 } 553 554 uint young_list_target_length = 0; 555 if (adaptive_young_list_length()) { 556 if (collector_state()->gcs_are_young()) { 557 young_list_target_length = 558 calculate_young_list_target_length(rs_lengths, 559 base_min_length, 560 desired_min_length, 561 desired_max_length); 562 } else { 563 // Don't calculate anything and let the code below bound it to 564 // the desired_min_length, i.e., do the next GC as soon as 565 // possible to maximize how many old regions we can add to it. 566 } 567 } else { 568 // The user asked for a fixed young gen so we'll fix the young gen 569 // whether the next GC is young or mixed. 570 young_list_target_length = _young_list_fixed_length; 571 } 572 573 // Make sure we don't go over the desired max length, nor under the 574 // desired min length. In case they clash, desired_min_length wins 575 // which is why that test is second. 576 if (young_list_target_length > desired_max_length) { 577 young_list_target_length = desired_max_length; 578 } 579 if (young_list_target_length < desired_min_length) { 580 young_list_target_length = desired_min_length; 581 } 582 583 assert(young_list_target_length > recorded_survivor_regions(), 584 "we should be able to allocate at least one eden region"); 585 assert(young_list_target_length >= absolute_min_length, "post-condition"); 586 587 return young_list_target_length; 588 } 589 590 uint 591 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 592 uint base_min_length, 593 uint desired_min_length, 594 uint desired_max_length) const { 595 assert(adaptive_young_list_length(), "pre-condition"); 596 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 597 598 // In case some edge-condition makes the desired max length too small... 599 if (desired_max_length <= desired_min_length) { 600 return desired_min_length; 601 } 602 603 // We'll adjust min_young_length and max_young_length not to include 604 // the already allocated young regions (i.e., so they reflect the 605 // min and max eden regions we'll allocate). The base_min_length 606 // will be reflected in the predictions by the 607 // survivor_regions_evac_time prediction. 608 assert(desired_min_length > base_min_length, "invariant"); 609 uint min_young_length = desired_min_length - base_min_length; 610 assert(desired_max_length > base_min_length, "invariant"); 611 uint max_young_length = desired_max_length - base_min_length; 612 613 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 614 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 615 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 616 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 617 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 618 double base_time_ms = 619 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 620 survivor_regions_evac_time; 621 uint available_free_regions = _free_regions_at_end_of_collection; 622 uint base_free_regions = 0; 623 if (available_free_regions > _reserve_regions) { 624 base_free_regions = available_free_regions - _reserve_regions; 625 } 626 627 // Here, we will make sure that the shortest young length that 628 // makes sense fits within the target pause time. 629 630 if (predict_will_fit(min_young_length, base_time_ms, 631 base_free_regions, target_pause_time_ms)) { 632 // The shortest young length will fit into the target pause time; 633 // we'll now check whether the absolute maximum number of young 634 // regions will fit in the target pause time. If not, we'll do 635 // a binary search between min_young_length and max_young_length. 636 if (predict_will_fit(max_young_length, base_time_ms, 637 base_free_regions, target_pause_time_ms)) { 638 // The maximum young length will fit into the target pause time. 639 // We are done so set min young length to the maximum length (as 640 // the result is assumed to be returned in min_young_length). 641 min_young_length = max_young_length; 642 } else { 643 // The maximum possible number of young regions will not fit within 644 // the target pause time so we'll search for the optimal 645 // length. The loop invariants are: 646 // 647 // min_young_length < max_young_length 648 // min_young_length is known to fit into the target pause time 649 // max_young_length is known not to fit into the target pause time 650 // 651 // Going into the loop we know the above hold as we've just 652 // checked them. Every time around the loop we check whether 653 // the middle value between min_young_length and 654 // max_young_length fits into the target pause time. If it 655 // does, it becomes the new min. If it doesn't, it becomes 656 // the new max. This way we maintain the loop invariants. 657 658 assert(min_young_length < max_young_length, "invariant"); 659 uint diff = (max_young_length - min_young_length) / 2; 660 while (diff > 0) { 661 uint young_length = min_young_length + diff; 662 if (predict_will_fit(young_length, base_time_ms, 663 base_free_regions, target_pause_time_ms)) { 664 min_young_length = young_length; 665 } else { 666 max_young_length = young_length; 667 } 668 assert(min_young_length < max_young_length, "invariant"); 669 diff = (max_young_length - min_young_length) / 2; 670 } 671 // The results is min_young_length which, according to the 672 // loop invariants, should fit within the target pause time. 673 674 // These are the post-conditions of the binary search above: 675 assert(min_young_length < max_young_length, 676 "otherwise we should have discovered that max_young_length " 677 "fits into the pause target and not done the binary search"); 678 assert(predict_will_fit(min_young_length, base_time_ms, 679 base_free_regions, target_pause_time_ms), 680 "min_young_length, the result of the binary search, should " 681 "fit into the pause target"); 682 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 683 base_free_regions, target_pause_time_ms), 684 "min_young_length, the result of the binary search, should be " 685 "optimal, so no larger length should fit into the pause target"); 686 } 687 } else { 688 // Even the minimum length doesn't fit into the pause time 689 // target, return it as the result nevertheless. 690 } 691 return base_min_length + min_young_length; 692 } 693 694 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 695 double survivor_regions_evac_time = 0.0; 696 for (HeapRegion * r = _recorded_survivor_head; 697 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 698 r = r->get_next_young_region()) { 699 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 700 } 701 return survivor_regions_evac_time; 702 } 703 704 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 705 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 706 707 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 708 if (rs_lengths > _rs_lengths_prediction) { 709 // add 10% to avoid having to recalculate often 710 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 711 update_rs_lengths_prediction(rs_lengths_prediction); 712 713 update_young_list_max_and_target_length(rs_lengths_prediction); 714 } 715 } 716 717 void G1CollectorPolicy::update_rs_lengths_prediction() { 718 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 719 } 720 721 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 722 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 723 _rs_lengths_prediction = prediction; 724 } 725 } 726 727 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 728 bool is_tlab, 729 bool* gc_overhead_limit_was_exceeded) { 730 guarantee(false, "Not using this policy feature yet."); 731 return NULL; 732 } 733 734 // This method controls how a collector handles one or more 735 // of its generations being fully allocated. 736 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 737 bool is_tlab) { 738 guarantee(false, "Not using this policy feature yet."); 739 return NULL; 740 } 741 742 743 #ifndef PRODUCT 744 bool G1CollectorPolicy::verify_young_ages() { 745 HeapRegion* head = _g1->young_list()->first_region(); 746 return 747 verify_young_ages(head, _short_lived_surv_rate_group); 748 // also call verify_young_ages on any additional surv rate groups 749 } 750 751 bool 752 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 753 SurvRateGroup *surv_rate_group) { 754 guarantee( surv_rate_group != NULL, "pre-condition" ); 755 756 const char* name = surv_rate_group->name(); 757 bool ret = true; 758 int prev_age = -1; 759 760 for (HeapRegion* curr = head; 761 curr != NULL; 762 curr = curr->get_next_young_region()) { 763 SurvRateGroup* group = curr->surv_rate_group(); 764 if (group == NULL && !curr->is_survivor()) { 765 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 766 ret = false; 767 } 768 769 if (surv_rate_group == group) { 770 int age = curr->age_in_surv_rate_group(); 771 772 if (age < 0) { 773 gclog_or_tty->print_cr("## %s: encountered negative age", name); 774 ret = false; 775 } 776 777 if (age <= prev_age) { 778 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 779 "(%d, %d)", name, age, prev_age); 780 ret = false; 781 } 782 prev_age = age; 783 } 784 } 785 786 return ret; 787 } 788 #endif // PRODUCT 789 790 void G1CollectorPolicy::record_full_collection_start() { 791 _full_collection_start_sec = os::elapsedTime(); 792 record_heap_size_info_at_start(true /* full */); 793 // Release the future to-space so that it is available for compaction into. 794 collector_state()->set_full_collection(true); 795 } 796 797 void G1CollectorPolicy::record_full_collection_end() { 798 // Consider this like a collection pause for the purposes of allocation 799 // since last pause. 800 double end_sec = os::elapsedTime(); 801 double full_gc_time_sec = end_sec - _full_collection_start_sec; 802 double full_gc_time_ms = full_gc_time_sec * 1000.0; 803 804 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 805 806 update_recent_gc_times(end_sec, full_gc_time_ms); 807 808 collector_state()->set_full_collection(false); 809 810 // "Nuke" the heuristics that control the young/mixed GC 811 // transitions and make sure we start with young GCs after the Full GC. 812 collector_state()->set_gcs_are_young(true); 813 collector_state()->set_last_young_gc(false); 814 collector_state()->set_initiate_conc_mark_if_possible(false); 815 collector_state()->set_during_initial_mark_pause(false); 816 collector_state()->set_in_marking_window(false); 817 collector_state()->set_in_marking_window_im(false); 818 819 _short_lived_surv_rate_group->start_adding_regions(); 820 // also call this on any additional surv rate groups 821 822 record_survivor_regions(0, NULL, NULL); 823 824 _free_regions_at_end_of_collection = _g1->num_free_regions(); 825 // Reset survivors SurvRateGroup. 826 _survivor_surv_rate_group->reset(); 827 update_young_list_max_and_target_length(); 828 update_rs_lengths_prediction(); 829 _collectionSetChooser->clear(); 830 } 831 832 void G1CollectorPolicy::record_stop_world_start() { 833 _stop_world_start = os::elapsedTime(); 834 } 835 836 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 837 // We only need to do this here as the policy will only be applied 838 // to the GC we're about to start. so, no point is calculating this 839 // every time we calculate / recalculate the target young length. 840 update_survivors_policy(); 841 842 assert(_g1->used() == _g1->recalculate_used(), 843 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 844 _g1->used(), _g1->recalculate_used()); 845 846 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 847 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 848 _stop_world_start = 0.0; 849 850 record_heap_size_info_at_start(false /* full */); 851 852 phase_times()->record_cur_collection_start_sec(start_time_sec); 853 _pending_cards = _g1->pending_card_num(); 854 855 _collection_set_bytes_used_before = 0; 856 _bytes_copied_during_gc = 0; 857 858 collector_state()->set_last_gc_was_young(false); 859 860 // do that for any other surv rate groups 861 _short_lived_surv_rate_group->stop_adding_regions(); 862 _survivors_age_table.clear(); 863 864 assert( verify_young_ages(), "region age verification" ); 865 } 866 867 void G1CollectorPolicy::record_concurrent_mark_init_end(double 868 mark_init_elapsed_time_ms) { 869 collector_state()->set_during_marking(true); 870 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 871 collector_state()->set_during_initial_mark_pause(false); 872 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 873 } 874 875 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 876 _mark_remark_start_sec = os::elapsedTime(); 877 collector_state()->set_during_marking(false); 878 } 879 880 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 881 double end_time_sec = os::elapsedTime(); 882 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 883 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 884 _cur_mark_stop_world_time_ms += elapsed_time_ms; 885 _prev_collection_pause_end_ms += elapsed_time_ms; 886 887 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec); 888 } 889 890 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 891 _mark_cleanup_start_sec = os::elapsedTime(); 892 } 893 894 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 895 collector_state()->set_last_young_gc(true); 896 collector_state()->set_in_marking_window(false); 897 } 898 899 void G1CollectorPolicy::record_concurrent_pause() { 900 if (_stop_world_start > 0.0) { 901 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 902 _trace_young_gen_time_data.record_yield_time(yield_ms); 903 } 904 } 905 906 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 907 return phase_times()->average_time_ms(phase); 908 } 909 910 double G1CollectorPolicy::young_other_time_ms() const { 911 return phase_times()->young_cset_choice_time_ms() + 912 phase_times()->young_free_cset_time_ms(); 913 } 914 915 double G1CollectorPolicy::non_young_other_time_ms() const { 916 return phase_times()->non_young_cset_choice_time_ms() + 917 phase_times()->non_young_free_cset_time_ms(); 918 919 } 920 921 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 922 return pause_time_ms - 923 average_time_ms(G1GCPhaseTimes::UpdateRS) - 924 average_time_ms(G1GCPhaseTimes::ScanRS) - 925 average_time_ms(G1GCPhaseTimes::ObjCopy) - 926 average_time_ms(G1GCPhaseTimes::Termination); 927 } 928 929 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 930 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 931 } 932 933 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 934 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 935 return false; 936 } 937 938 size_t marking_initiating_used_threshold = 939 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 940 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 941 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 942 943 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 944 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 945 ergo_verbose5(ErgoConcCycles, 946 "request concurrent cycle initiation", 947 ergo_format_reason("occupancy higher than threshold") 948 ergo_format_byte("occupancy") 949 ergo_format_byte("allocation request") 950 ergo_format_byte_perc("threshold") 951 ergo_format_str("source"), 952 cur_used_bytes, 953 alloc_byte_size, 954 marking_initiating_used_threshold, 955 (double) InitiatingHeapOccupancyPercent, 956 source); 957 return true; 958 } else { 959 ergo_verbose5(ErgoConcCycles, 960 "do not request concurrent cycle initiation", 961 ergo_format_reason("still doing mixed collections") 962 ergo_format_byte("occupancy") 963 ergo_format_byte("allocation request") 964 ergo_format_byte_perc("threshold") 965 ergo_format_str("source"), 966 cur_used_bytes, 967 alloc_byte_size, 968 marking_initiating_used_threshold, 969 (double) InitiatingHeapOccupancyPercent, 970 source); 971 } 972 } 973 974 return false; 975 } 976 977 // Anything below that is considered to be zero 978 #define MIN_TIMER_GRANULARITY 0.0000001 979 980 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 981 double end_time_sec = os::elapsedTime(); 982 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 983 "otherwise, the subtraction below does not make sense"); 984 size_t rs_size = 985 _cur_collection_pause_used_regions_at_start - cset_region_length(); 986 size_t cur_used_bytes = _g1->used(); 987 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 988 bool last_pause_included_initial_mark = false; 989 bool update_stats = !_g1->evacuation_failed(); 990 991 #ifndef PRODUCT 992 if (G1YoungSurvRateVerbose) { 993 gclog_or_tty->cr(); 994 _short_lived_surv_rate_group->print(); 995 // do that for any other surv rate groups too 996 } 997 #endif // PRODUCT 998 999 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1000 if (last_pause_included_initial_mark) { 1001 record_concurrent_mark_init_end(0.0); 1002 } else if (need_to_start_conc_mark("end of GC")) { 1003 // Note: this might have already been set, if during the last 1004 // pause we decided to start a cycle but at the beginning of 1005 // this pause we decided to postpone it. That's OK. 1006 collector_state()->set_initiate_conc_mark_if_possible(true); 1007 } 1008 1009 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, end_time_sec); 1010 1011 if (update_stats) { 1012 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1013 // this is where we update the allocation rate of the application 1014 double app_time_ms = 1015 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1016 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1017 // This usually happens due to the timer not having the required 1018 // granularity. Some Linuxes are the usual culprits. 1019 // We'll just set it to something (arbitrarily) small. 1020 app_time_ms = 1.0; 1021 } 1022 // We maintain the invariant that all objects allocated by mutator 1023 // threads will be allocated out of eden regions. So, we can use 1024 // the eden region number allocated since the previous GC to 1025 // calculate the application's allocate rate. The only exception 1026 // to that is humongous objects that are allocated separately. But 1027 // given that humongous object allocations do not really affect 1028 // either the pause's duration nor when the next pause will take 1029 // place we can safely ignore them here. 1030 uint regions_allocated = eden_cset_region_length(); 1031 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1032 _alloc_rate_ms_seq->add(alloc_rate_ms); 1033 1034 double interval_ms = 1035 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1036 update_recent_gc_times(end_time_sec, pause_time_ms); 1037 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1038 if (recent_avg_pause_time_ratio() < 0.0 || 1039 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1040 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1041 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1042 if (_recent_avg_pause_time_ratio < 0.0) { 1043 _recent_avg_pause_time_ratio = 0.0; 1044 } else { 1045 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1046 _recent_avg_pause_time_ratio = 1.0; 1047 } 1048 } 1049 } 1050 1051 bool new_in_marking_window = collector_state()->in_marking_window(); 1052 bool new_in_marking_window_im = false; 1053 if (last_pause_included_initial_mark) { 1054 new_in_marking_window = true; 1055 new_in_marking_window_im = true; 1056 } 1057 1058 if (collector_state()->last_young_gc()) { 1059 // This is supposed to to be the "last young GC" before we start 1060 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1061 1062 if (!last_pause_included_initial_mark) { 1063 if (next_gc_should_be_mixed("start mixed GCs", 1064 "do not start mixed GCs")) { 1065 collector_state()->set_gcs_are_young(false); 1066 } 1067 } else { 1068 ergo_verbose0(ErgoMixedGCs, 1069 "do not start mixed GCs", 1070 ergo_format_reason("concurrent cycle is about to start")); 1071 } 1072 collector_state()->set_last_young_gc(false); 1073 } 1074 1075 if (!collector_state()->last_gc_was_young()) { 1076 // This is a mixed GC. Here we decide whether to continue doing 1077 // mixed GCs or not. 1078 1079 if (!next_gc_should_be_mixed("continue mixed GCs", 1080 "do not continue mixed GCs")) { 1081 collector_state()->set_gcs_are_young(true); 1082 } 1083 } 1084 1085 _short_lived_surv_rate_group->start_adding_regions(); 1086 // Do that for any other surv rate groups 1087 1088 if (update_stats) { 1089 double cost_per_card_ms = 0.0; 1090 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1091 if (_pending_cards > 0) { 1092 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1093 _cost_per_card_ms_seq->add(cost_per_card_ms); 1094 } 1095 _cost_scan_hcc_seq->add(cost_scan_hcc); 1096 1097 double cost_per_entry_ms = 0.0; 1098 if (cards_scanned > 10) { 1099 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1100 if (collector_state()->last_gc_was_young()) { 1101 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1102 } else { 1103 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1104 } 1105 } 1106 1107 if (_max_rs_lengths > 0) { 1108 double cards_per_entry_ratio = 1109 (double) cards_scanned / (double) _max_rs_lengths; 1110 if (collector_state()->last_gc_was_young()) { 1111 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1112 } else { 1113 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1114 } 1115 } 1116 1117 // This is defensive. For a while _max_rs_lengths could get 1118 // smaller than _recorded_rs_lengths which was causing 1119 // rs_length_diff to get very large and mess up the RSet length 1120 // predictions. The reason was unsafe concurrent updates to the 1121 // _inc_cset_recorded_rs_lengths field which the code below guards 1122 // against (see CR 7118202). This bug has now been fixed (see CR 1123 // 7119027). However, I'm still worried that 1124 // _inc_cset_recorded_rs_lengths might still end up somewhat 1125 // inaccurate. The concurrent refinement thread calculates an 1126 // RSet's length concurrently with other CR threads updating it 1127 // which might cause it to calculate the length incorrectly (if, 1128 // say, it's in mid-coarsening). So I'll leave in the defensive 1129 // conditional below just in case. 1130 size_t rs_length_diff = 0; 1131 if (_max_rs_lengths > _recorded_rs_lengths) { 1132 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1133 } 1134 _rs_length_diff_seq->add((double) rs_length_diff); 1135 1136 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1137 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1138 double cost_per_byte_ms = 0.0; 1139 1140 if (copied_bytes > 0) { 1141 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1142 if (collector_state()->in_marking_window()) { 1143 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1144 } else { 1145 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1146 } 1147 } 1148 1149 if (young_cset_region_length() > 0) { 1150 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1151 young_cset_region_length()); 1152 } 1153 1154 if (old_cset_region_length() > 0) { 1155 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1156 old_cset_region_length()); 1157 } 1158 1159 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1160 1161 _pending_cards_seq->add((double) _pending_cards); 1162 _rs_lengths_seq->add((double) _max_rs_lengths); 1163 } 1164 1165 collector_state()->set_in_marking_window(new_in_marking_window); 1166 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1167 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1168 update_young_list_max_and_target_length(); 1169 update_rs_lengths_prediction(); 1170 1171 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1172 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1173 1174 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1175 1176 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1177 ergo_verbose2(ErgoTiming, 1178 "adjust concurrent refinement thresholds", 1179 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1180 ergo_format_ms("Update RS time goal") 1181 ergo_format_ms("Scan HCC time"), 1182 update_rs_time_goal_ms, 1183 scan_hcc_time_ms); 1184 1185 update_rs_time_goal_ms = 0; 1186 } else { 1187 update_rs_time_goal_ms -= scan_hcc_time_ms; 1188 } 1189 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1190 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1191 update_rs_time_goal_ms); 1192 1193 _collectionSetChooser->verify(); 1194 } 1195 1196 #define EXT_SIZE_FORMAT "%.1f%s" 1197 #define EXT_SIZE_PARAMS(bytes) \ 1198 byte_size_in_proper_unit((double)(bytes)), \ 1199 proper_unit_for_byte_size((bytes)) 1200 1201 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1202 YoungList* young_list = _g1->young_list(); 1203 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1204 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1205 _heap_capacity_bytes_before_gc = _g1->capacity(); 1206 _heap_used_bytes_before_gc = _g1->used(); 1207 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1208 1209 _eden_capacity_bytes_before_gc = 1210 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1211 1212 if (full) { 1213 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1214 } 1215 } 1216 1217 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1218 size_t bytes_after = _g1->used(); 1219 size_t capacity = _g1->capacity(); 1220 1221 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1222 byte_size_in_proper_unit(bytes_before), 1223 proper_unit_for_byte_size(bytes_before), 1224 byte_size_in_proper_unit(bytes_after), 1225 proper_unit_for_byte_size(bytes_after), 1226 byte_size_in_proper_unit(capacity), 1227 proper_unit_for_byte_size(capacity)); 1228 } 1229 1230 void G1CollectorPolicy::print_heap_transition() const { 1231 print_heap_transition(_heap_used_bytes_before_gc); 1232 } 1233 1234 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1235 YoungList* young_list = _g1->young_list(); 1236 1237 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1238 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1239 size_t heap_used_bytes_after_gc = _g1->used(); 1240 1241 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1242 size_t eden_capacity_bytes_after_gc = 1243 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1244 1245 gclog_or_tty->print( 1246 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1247 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1248 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1249 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1250 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1251 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1252 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1253 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1254 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1255 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1256 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1257 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1258 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1259 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1260 1261 if (full) { 1262 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1263 } 1264 1265 gclog_or_tty->cr(); 1266 } 1267 1268 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1269 phase_times()->print(pause_time_sec); 1270 } 1271 1272 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1273 double update_rs_processed_buffers, 1274 double goal_ms) { 1275 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1276 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1277 1278 if (G1UseAdaptiveConcRefinement) { 1279 const int k_gy = 3, k_gr = 6; 1280 const double inc_k = 1.1, dec_k = 0.9; 1281 1282 int g = cg1r->green_zone(); 1283 if (update_rs_time > goal_ms) { 1284 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1285 } else { 1286 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1287 g = (int)MAX2(g * inc_k, g + 1.0); 1288 } 1289 } 1290 // Change the refinement threads params 1291 cg1r->set_green_zone(g); 1292 cg1r->set_yellow_zone(g * k_gy); 1293 cg1r->set_red_zone(g * k_gr); 1294 cg1r->reinitialize_threads(); 1295 1296 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1297 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1298 cg1r->yellow_zone()); 1299 // Change the barrier params 1300 dcqs.set_process_completed_threshold(processing_threshold); 1301 dcqs.set_max_completed_queue(cg1r->red_zone()); 1302 } 1303 1304 int curr_queue_size = dcqs.completed_buffers_num(); 1305 if (curr_queue_size >= cg1r->yellow_zone()) { 1306 dcqs.set_completed_queue_padding(curr_queue_size); 1307 } else { 1308 dcqs.set_completed_queue_padding(0); 1309 } 1310 dcqs.notify_if_necessary(); 1311 } 1312 1313 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1314 return (size_t) get_new_prediction(_rs_length_diff_seq); 1315 } 1316 1317 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1318 return get_new_prediction(_alloc_rate_ms_seq); 1319 } 1320 1321 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1322 return get_new_prediction(_cost_per_card_ms_seq); 1323 } 1324 1325 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1326 return get_new_prediction(_cost_scan_hcc_seq); 1327 } 1328 1329 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1330 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1331 } 1332 1333 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1334 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1335 } 1336 1337 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1338 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1339 return predict_young_cards_per_entry_ratio(); 1340 } else { 1341 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1342 } 1343 } 1344 1345 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1346 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1347 } 1348 1349 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1350 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1351 } 1352 1353 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1354 if (collector_state()->gcs_are_young()) { 1355 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1356 } else { 1357 return predict_mixed_rs_scan_time_ms(card_num); 1358 } 1359 } 1360 1361 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1362 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1363 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1364 } else { 1365 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1366 } 1367 } 1368 1369 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1370 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1371 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1372 } else { 1373 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1374 } 1375 } 1376 1377 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1378 if (collector_state()->during_concurrent_mark()) { 1379 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1380 } else { 1381 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1382 } 1383 } 1384 1385 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1386 return get_new_prediction(_constant_other_time_ms_seq); 1387 } 1388 1389 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1390 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1391 } 1392 1393 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1394 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1395 } 1396 1397 double G1CollectorPolicy::predict_remark_time_ms() const { 1398 return get_new_prediction(_concurrent_mark_remark_times_ms); 1399 } 1400 1401 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1402 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1403 } 1404 1405 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1406 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1407 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1408 double pred = get_new_prediction(seq); 1409 if (pred > 1.0) { 1410 pred = 1.0; 1411 } 1412 return pred; 1413 } 1414 1415 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1416 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1417 } 1418 1419 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1420 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1421 } 1422 1423 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1424 size_t scanned_cards) const { 1425 return 1426 predict_rs_update_time_ms(pending_cards) + 1427 predict_rs_scan_time_ms(scanned_cards) + 1428 predict_constant_other_time_ms(); 1429 } 1430 1431 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1432 size_t rs_length = predict_rs_length_diff(); 1433 size_t card_num; 1434 if (collector_state()->gcs_are_young()) { 1435 card_num = predict_young_card_num(rs_length); 1436 } else { 1437 card_num = predict_non_young_card_num(rs_length); 1438 } 1439 return predict_base_elapsed_time_ms(pending_cards, card_num); 1440 } 1441 1442 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1443 size_t bytes_to_copy; 1444 if (hr->is_marked()) 1445 bytes_to_copy = hr->max_live_bytes(); 1446 else { 1447 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1448 int age = hr->age_in_surv_rate_group(); 1449 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1450 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1451 } 1452 return bytes_to_copy; 1453 } 1454 1455 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1456 bool for_young_gc) const { 1457 size_t rs_length = hr->rem_set()->occupied(); 1458 size_t card_num; 1459 1460 // Predicting the number of cards is based on which type of GC 1461 // we're predicting for. 1462 if (for_young_gc) { 1463 card_num = predict_young_card_num(rs_length); 1464 } else { 1465 card_num = predict_non_young_card_num(rs_length); 1466 } 1467 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1468 1469 double region_elapsed_time_ms = 1470 predict_rs_scan_time_ms(card_num) + 1471 predict_object_copy_time_ms(bytes_to_copy); 1472 1473 // The prediction of the "other" time for this region is based 1474 // upon the region type and NOT the GC type. 1475 if (hr->is_young()) { 1476 region_elapsed_time_ms += predict_young_other_time_ms(1); 1477 } else { 1478 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1479 } 1480 return region_elapsed_time_ms; 1481 } 1482 1483 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1484 uint survivor_cset_region_length) { 1485 _eden_cset_region_length = eden_cset_region_length; 1486 _survivor_cset_region_length = survivor_cset_region_length; 1487 _old_cset_region_length = 0; 1488 } 1489 1490 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1491 _recorded_rs_lengths = rs_lengths; 1492 } 1493 1494 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1495 double elapsed_ms) { 1496 _recent_gc_times_ms->add(elapsed_ms); 1497 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1498 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1499 } 1500 1501 size_t G1CollectorPolicy::expansion_amount() const { 1502 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1503 double threshold = _gc_overhead_perc; 1504 if (recent_gc_overhead > threshold) { 1505 // We will double the existing space, or take 1506 // G1ExpandByPercentOfAvailable % of the available expansion 1507 // space, whichever is smaller, bounded below by a minimum 1508 // expansion (unless that's all that's left.) 1509 const size_t min_expand_bytes = 1*M; 1510 size_t reserved_bytes = _g1->max_capacity(); 1511 size_t committed_bytes = _g1->capacity(); 1512 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1513 size_t expand_bytes; 1514 size_t expand_bytes_via_pct = 1515 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1516 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1517 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1518 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1519 1520 ergo_verbose5(ErgoHeapSizing, 1521 "attempt heap expansion", 1522 ergo_format_reason("recent GC overhead higher than " 1523 "threshold after GC") 1524 ergo_format_perc("recent GC overhead") 1525 ergo_format_perc("threshold") 1526 ergo_format_byte("uncommitted") 1527 ergo_format_byte_perc("calculated expansion amount"), 1528 recent_gc_overhead, threshold, 1529 uncommitted_bytes, 1530 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1531 1532 return expand_bytes; 1533 } else { 1534 return 0; 1535 } 1536 } 1537 1538 void G1CollectorPolicy::print_tracing_info() const { 1539 _trace_young_gen_time_data.print(); 1540 _trace_old_gen_time_data.print(); 1541 } 1542 1543 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1544 #ifndef PRODUCT 1545 _short_lived_surv_rate_group->print_surv_rate_summary(); 1546 // add this call for any other surv rate groups 1547 #endif // PRODUCT 1548 } 1549 1550 bool G1CollectorPolicy::is_young_list_full() const { 1551 uint young_list_length = _g1->young_list()->length(); 1552 uint young_list_target_length = _young_list_target_length; 1553 return young_list_length >= young_list_target_length; 1554 } 1555 1556 bool G1CollectorPolicy::can_expand_young_list() const { 1557 uint young_list_length = _g1->young_list()->length(); 1558 uint young_list_max_length = _young_list_max_length; 1559 return young_list_length < young_list_max_length; 1560 } 1561 1562 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1563 uint expansion_region_num = 0; 1564 if (GCLockerEdenExpansionPercent > 0) { 1565 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1566 double expansion_region_num_d = perc * (double) _young_list_target_length; 1567 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1568 // less than 1.0) we'll get 1. 1569 expansion_region_num = (uint) ceil(expansion_region_num_d); 1570 } else { 1571 assert(expansion_region_num == 0, "sanity"); 1572 } 1573 _young_list_max_length = _young_list_target_length + expansion_region_num; 1574 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1575 } 1576 1577 // Calculates survivor space parameters. 1578 void G1CollectorPolicy::update_survivors_policy() { 1579 double max_survivor_regions_d = 1580 (double) _young_list_target_length / (double) SurvivorRatio; 1581 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1582 // smaller than 1.0) we'll get 1. 1583 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1584 1585 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1586 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1587 } 1588 1589 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 1590 GCCause::Cause gc_cause) { 1591 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1592 if (!during_cycle) { 1593 ergo_verbose1(ErgoConcCycles, 1594 "request concurrent cycle initiation", 1595 ergo_format_reason("requested by GC cause") 1596 ergo_format_str("GC cause"), 1597 GCCause::to_string(gc_cause)); 1598 collector_state()->set_initiate_conc_mark_if_possible(true); 1599 return true; 1600 } else { 1601 ergo_verbose1(ErgoConcCycles, 1602 "do not request concurrent cycle initiation", 1603 ergo_format_reason("concurrent cycle already in progress") 1604 ergo_format_str("GC cause"), 1605 GCCause::to_string(gc_cause)); 1606 return false; 1607 } 1608 } 1609 1610 void 1611 G1CollectorPolicy::decide_on_conc_mark_initiation() { 1612 // We are about to decide on whether this pause will be an 1613 // initial-mark pause. 1614 1615 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1616 // will set it here if we have to. However, it should be cleared by 1617 // the end of the pause (it's only set for the duration of an 1618 // initial-mark pause). 1619 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1620 1621 if (collector_state()->initiate_conc_mark_if_possible()) { 1622 // We had noticed on a previous pause that the heap occupancy has 1623 // gone over the initiating threshold and we should start a 1624 // concurrent marking cycle. So we might initiate one. 1625 1626 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1627 if (!during_cycle) { 1628 // The concurrent marking thread is not "during a cycle", i.e., 1629 // it has completed the last one. So we can go ahead and 1630 // initiate a new cycle. 1631 1632 collector_state()->set_during_initial_mark_pause(true); 1633 // We do not allow mixed GCs during marking. 1634 if (!collector_state()->gcs_are_young()) { 1635 collector_state()->set_gcs_are_young(true); 1636 ergo_verbose0(ErgoMixedGCs, 1637 "end mixed GCs", 1638 ergo_format_reason("concurrent cycle is about to start")); 1639 } 1640 1641 // And we can now clear initiate_conc_mark_if_possible() as 1642 // we've already acted on it. 1643 collector_state()->set_initiate_conc_mark_if_possible(false); 1644 1645 ergo_verbose0(ErgoConcCycles, 1646 "initiate concurrent cycle", 1647 ergo_format_reason("concurrent cycle initiation requested")); 1648 } else { 1649 // The concurrent marking thread is still finishing up the 1650 // previous cycle. If we start one right now the two cycles 1651 // overlap. In particular, the concurrent marking thread might 1652 // be in the process of clearing the next marking bitmap (which 1653 // we will use for the next cycle if we start one). Starting a 1654 // cycle now will be bad given that parts of the marking 1655 // information might get cleared by the marking thread. And we 1656 // cannot wait for the marking thread to finish the cycle as it 1657 // periodically yields while clearing the next marking bitmap 1658 // and, if it's in a yield point, it's waiting for us to 1659 // finish. So, at this point we will not start a cycle and we'll 1660 // let the concurrent marking thread complete the last one. 1661 ergo_verbose0(ErgoConcCycles, 1662 "do not initiate concurrent cycle", 1663 ergo_format_reason("concurrent cycle already in progress")); 1664 } 1665 } 1666 } 1667 1668 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1669 G1CollectedHeap* _g1h; 1670 CSetChooserParUpdater _cset_updater; 1671 1672 public: 1673 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1674 uint chunk_size) : 1675 _g1h(G1CollectedHeap::heap()), 1676 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1677 1678 bool doHeapRegion(HeapRegion* r) { 1679 // Do we have any marking information for this region? 1680 if (r->is_marked()) { 1681 // We will skip any region that's currently used as an old GC 1682 // alloc region (we should not consider those for collection 1683 // before we fill them up). 1684 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1685 _cset_updater.add_region(r); 1686 } 1687 } 1688 return false; 1689 } 1690 }; 1691 1692 class ParKnownGarbageTask: public AbstractGangTask { 1693 CollectionSetChooser* _hrSorted; 1694 uint _chunk_size; 1695 G1CollectedHeap* _g1; 1696 HeapRegionClaimer _hrclaimer; 1697 1698 public: 1699 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1700 AbstractGangTask("ParKnownGarbageTask"), 1701 _hrSorted(hrSorted), _chunk_size(chunk_size), 1702 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1703 1704 void work(uint worker_id) { 1705 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1706 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1707 } 1708 }; 1709 1710 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1711 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1712 const uint overpartition_factor = 4; 1713 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1714 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1715 } 1716 1717 void 1718 G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1719 _collectionSetChooser->clear(); 1720 1721 WorkGang* workers = _g1->workers(); 1722 uint n_workers = workers->active_workers(); 1723 1724 uint n_regions = _g1->num_regions(); 1725 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1726 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1727 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1728 workers->run_task(&par_known_garbage_task); 1729 1730 _collectionSetChooser->sort_regions(); 1731 1732 double end_sec = os::elapsedTime(); 1733 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1734 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1735 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1736 _prev_collection_pause_end_ms += elapsed_time_ms; 1737 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec); 1738 } 1739 1740 // Add the heap region at the head of the non-incremental collection set 1741 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1742 assert(_inc_cset_build_state == Active, "Precondition"); 1743 assert(hr->is_old(), "the region should be old"); 1744 1745 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1746 _g1->register_old_region_with_cset(hr); 1747 hr->set_next_in_collection_set(_collection_set); 1748 _collection_set = hr; 1749 _collection_set_bytes_used_before += hr->used(); 1750 size_t rs_length = hr->rem_set()->occupied(); 1751 _recorded_rs_lengths += rs_length; 1752 _old_cset_region_length += 1; 1753 } 1754 1755 // Initialize the per-collection-set information 1756 void G1CollectorPolicy::start_incremental_cset_building() { 1757 assert(_inc_cset_build_state == Inactive, "Precondition"); 1758 1759 _inc_cset_head = NULL; 1760 _inc_cset_tail = NULL; 1761 _inc_cset_bytes_used_before = 0; 1762 1763 _inc_cset_max_finger = 0; 1764 _inc_cset_recorded_rs_lengths = 0; 1765 _inc_cset_recorded_rs_lengths_diffs = 0; 1766 _inc_cset_predicted_elapsed_time_ms = 0.0; 1767 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1768 _inc_cset_build_state = Active; 1769 } 1770 1771 void G1CollectorPolicy::finalize_incremental_cset_building() { 1772 assert(_inc_cset_build_state == Active, "Precondition"); 1773 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1774 1775 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1776 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1777 // that adds a new region to the CSet. Further updates by the 1778 // concurrent refinement thread that samples the young RSet lengths 1779 // are accumulated in the *_diffs fields. Here we add the diffs to 1780 // the "main" fields. 1781 1782 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1783 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1784 } else { 1785 // This is defensive. The diff should in theory be always positive 1786 // as RSets can only grow between GCs. However, given that we 1787 // sample their size concurrently with other threads updating them 1788 // it's possible that we might get the wrong size back, which 1789 // could make the calculations somewhat inaccurate. 1790 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1791 if (_inc_cset_recorded_rs_lengths >= diffs) { 1792 _inc_cset_recorded_rs_lengths -= diffs; 1793 } else { 1794 _inc_cset_recorded_rs_lengths = 0; 1795 } 1796 } 1797 _inc_cset_predicted_elapsed_time_ms += 1798 _inc_cset_predicted_elapsed_time_ms_diffs; 1799 1800 _inc_cset_recorded_rs_lengths_diffs = 0; 1801 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1802 } 1803 1804 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1805 // This routine is used when: 1806 // * adding survivor regions to the incremental cset at the end of an 1807 // evacuation pause, 1808 // * adding the current allocation region to the incremental cset 1809 // when it is retired, and 1810 // * updating existing policy information for a region in the 1811 // incremental cset via young list RSet sampling. 1812 // Therefore this routine may be called at a safepoint by the 1813 // VM thread, or in-between safepoints by mutator threads (when 1814 // retiring the current allocation region) or a concurrent 1815 // refine thread (RSet sampling). 1816 1817 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1818 size_t used_bytes = hr->used(); 1819 _inc_cset_recorded_rs_lengths += rs_length; 1820 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1821 _inc_cset_bytes_used_before += used_bytes; 1822 1823 // Cache the values we have added to the aggregated information 1824 // in the heap region in case we have to remove this region from 1825 // the incremental collection set, or it is updated by the 1826 // rset sampling code 1827 hr->set_recorded_rs_length(rs_length); 1828 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1829 } 1830 1831 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1832 size_t new_rs_length) { 1833 // Update the CSet information that is dependent on the new RS length 1834 assert(hr->is_young(), "Precondition"); 1835 assert(!SafepointSynchronize::is_at_safepoint(), 1836 "should not be at a safepoint"); 1837 1838 // We could have updated _inc_cset_recorded_rs_lengths and 1839 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1840 // that atomically, as this code is executed by a concurrent 1841 // refinement thread, potentially concurrently with a mutator thread 1842 // allocating a new region and also updating the same fields. To 1843 // avoid the atomic operations we accumulate these updates on two 1844 // separate fields (*_diffs) and we'll just add them to the "main" 1845 // fields at the start of a GC. 1846 1847 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1848 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1849 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1850 1851 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1852 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1853 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1854 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1855 1856 hr->set_recorded_rs_length(new_rs_length); 1857 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1858 } 1859 1860 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1861 assert(hr->is_young(), "invariant"); 1862 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1863 assert(_inc_cset_build_state == Active, "Precondition"); 1864 1865 // We need to clear and set the cached recorded/cached collection set 1866 // information in the heap region here (before the region gets added 1867 // to the collection set). An individual heap region's cached values 1868 // are calculated, aggregated with the policy collection set info, 1869 // and cached in the heap region here (initially) and (subsequently) 1870 // by the Young List sampling code. 1871 1872 size_t rs_length = hr->rem_set()->occupied(); 1873 add_to_incremental_cset_info(hr, rs_length); 1874 1875 HeapWord* hr_end = hr->end(); 1876 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1877 1878 assert(!hr->in_collection_set(), "invariant"); 1879 _g1->register_young_region_with_cset(hr); 1880 assert(hr->next_in_collection_set() == NULL, "invariant"); 1881 } 1882 1883 // Add the region at the RHS of the incremental cset 1884 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1885 // We should only ever be appending survivors at the end of a pause 1886 assert(hr->is_survivor(), "Logic"); 1887 1888 // Do the 'common' stuff 1889 add_region_to_incremental_cset_common(hr); 1890 1891 // Now add the region at the right hand side 1892 if (_inc_cset_tail == NULL) { 1893 assert(_inc_cset_head == NULL, "invariant"); 1894 _inc_cset_head = hr; 1895 } else { 1896 _inc_cset_tail->set_next_in_collection_set(hr); 1897 } 1898 _inc_cset_tail = hr; 1899 } 1900 1901 // Add the region to the LHS of the incremental cset 1902 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 1903 // Survivors should be added to the RHS at the end of a pause 1904 assert(hr->is_eden(), "Logic"); 1905 1906 // Do the 'common' stuff 1907 add_region_to_incremental_cset_common(hr); 1908 1909 // Add the region at the left hand side 1910 hr->set_next_in_collection_set(_inc_cset_head); 1911 if (_inc_cset_head == NULL) { 1912 assert(_inc_cset_tail == NULL, "Invariant"); 1913 _inc_cset_tail = hr; 1914 } 1915 _inc_cset_head = hr; 1916 } 1917 1918 #ifndef PRODUCT 1919 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 1920 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 1921 1922 st->print_cr("\nCollection_set:"); 1923 HeapRegion* csr = list_head; 1924 while (csr != NULL) { 1925 HeapRegion* next = csr->next_in_collection_set(); 1926 assert(csr->in_collection_set(), "bad CS"); 1927 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 1928 HR_FORMAT_PARAMS(csr), 1929 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 1930 csr->age_in_surv_rate_group_cond()); 1931 csr = next; 1932 } 1933 } 1934 #endif // !PRODUCT 1935 1936 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 1937 // Returns the given amount of reclaimable bytes (that represents 1938 // the amount of reclaimable space still to be collected) as a 1939 // percentage of the current heap capacity. 1940 size_t capacity_bytes = _g1->capacity(); 1941 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 1942 } 1943 1944 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 1945 const char* false_action_str) const { 1946 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1947 if (cset_chooser->is_empty()) { 1948 ergo_verbose0(ErgoMixedGCs, 1949 false_action_str, 1950 ergo_format_reason("candidate old regions not available")); 1951 return false; 1952 } 1953 1954 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1955 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 1956 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 1957 double threshold = (double) G1HeapWastePercent; 1958 if (reclaimable_perc <= threshold) { 1959 ergo_verbose4(ErgoMixedGCs, 1960 false_action_str, 1961 ergo_format_reason("reclaimable percentage not over threshold") 1962 ergo_format_region("candidate old regions") 1963 ergo_format_byte_perc("reclaimable") 1964 ergo_format_perc("threshold"), 1965 cset_chooser->remaining_regions(), 1966 reclaimable_bytes, 1967 reclaimable_perc, threshold); 1968 return false; 1969 } 1970 1971 ergo_verbose4(ErgoMixedGCs, 1972 true_action_str, 1973 ergo_format_reason("candidate old regions available") 1974 ergo_format_region("candidate old regions") 1975 ergo_format_byte_perc("reclaimable") 1976 ergo_format_perc("threshold"), 1977 cset_chooser->remaining_regions(), 1978 reclaimable_bytes, 1979 reclaimable_perc, threshold); 1980 return true; 1981 } 1982 1983 uint G1CollectorPolicy::calc_min_old_cset_length() const { 1984 // The min old CSet region bound is based on the maximum desired 1985 // number of mixed GCs after a cycle. I.e., even if some old regions 1986 // look expensive, we should add them to the CSet anyway to make 1987 // sure we go through the available old regions in no more than the 1988 // maximum desired number of mixed GCs. 1989 // 1990 // The calculation is based on the number of marked regions we added 1991 // to the CSet chooser in the first place, not how many remain, so 1992 // that the result is the same during all mixed GCs that follow a cycle. 1993 1994 const size_t region_num = (size_t) _collectionSetChooser->length(); 1995 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 1996 size_t result = region_num / gc_num; 1997 // emulate ceiling 1998 if (result * gc_num < region_num) { 1999 result += 1; 2000 } 2001 return (uint) result; 2002 } 2003 2004 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2005 // The max old CSet region bound is based on the threshold expressed 2006 // as a percentage of the heap size. I.e., it should bound the 2007 // number of old regions added to the CSet irrespective of how many 2008 // of them are available. 2009 2010 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2011 const size_t region_num = g1h->num_regions(); 2012 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2013 size_t result = region_num * perc / 100; 2014 // emulate ceiling 2015 if (100 * result < region_num * perc) { 2016 result += 1; 2017 } 2018 return (uint) result; 2019 } 2020 2021 2022 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2023 double young_start_time_sec = os::elapsedTime(); 2024 2025 YoungList* young_list = _g1->young_list(); 2026 finalize_incremental_cset_building(); 2027 2028 guarantee(target_pause_time_ms > 0.0, 2029 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2030 guarantee(_collection_set == NULL, "Precondition"); 2031 2032 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2033 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2034 2035 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2036 "start choosing CSet", 2037 ergo_format_size("_pending_cards") 2038 ergo_format_ms("predicted base time") 2039 ergo_format_ms("remaining time") 2040 ergo_format_ms("target pause time"), 2041 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2042 2043 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2044 2045 if (collector_state()->last_gc_was_young()) { 2046 _trace_young_gen_time_data.increment_young_collection_count(); 2047 } else { 2048 _trace_young_gen_time_data.increment_mixed_collection_count(); 2049 } 2050 2051 // The young list is laid with the survivor regions from the previous 2052 // pause are appended to the RHS of the young list, i.e. 2053 // [Newly Young Regions ++ Survivors from last pause]. 2054 2055 uint survivor_region_length = young_list->survivor_length(); 2056 uint eden_region_length = young_list->eden_length(); 2057 init_cset_region_lengths(eden_region_length, survivor_region_length); 2058 2059 HeapRegion* hr = young_list->first_survivor_region(); 2060 while (hr != NULL) { 2061 assert(hr->is_survivor(), "badly formed young list"); 2062 // There is a convention that all the young regions in the CSet 2063 // are tagged as "eden", so we do this for the survivors here. We 2064 // use the special set_eden_pre_gc() as it doesn't check that the 2065 // region is free (which is not the case here). 2066 hr->set_eden_pre_gc(); 2067 hr = hr->get_next_young_region(); 2068 } 2069 2070 // Clear the fields that point to the survivor list - they are all young now. 2071 young_list->clear_survivors(); 2072 2073 _collection_set = _inc_cset_head; 2074 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2075 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2076 2077 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2078 "add young regions to CSet", 2079 ergo_format_region("eden") 2080 ergo_format_region("survivors") 2081 ergo_format_ms("predicted young region time") 2082 ergo_format_ms("target pause time"), 2083 eden_region_length, survivor_region_length, 2084 _inc_cset_predicted_elapsed_time_ms, 2085 target_pause_time_ms); 2086 2087 // The number of recorded young regions is the incremental 2088 // collection set's current size 2089 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2090 2091 double young_end_time_sec = os::elapsedTime(); 2092 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2093 2094 return time_remaining_ms; 2095 } 2096 2097 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2098 double non_young_start_time_sec = os::elapsedTime(); 2099 double predicted_old_time_ms = 0.0; 2100 2101 2102 if (!collector_state()->gcs_are_young()) { 2103 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2104 cset_chooser->verify(); 2105 const uint min_old_cset_length = calc_min_old_cset_length(); 2106 const uint max_old_cset_length = calc_max_old_cset_length(); 2107 2108 uint expensive_region_num = 0; 2109 bool check_time_remaining = adaptive_young_list_length(); 2110 2111 HeapRegion* hr = cset_chooser->peek(); 2112 while (hr != NULL) { 2113 if (old_cset_region_length() >= max_old_cset_length) { 2114 // Added maximum number of old regions to the CSet. 2115 ergo_verbose2(ErgoCSetConstruction, 2116 "finish adding old regions to CSet", 2117 ergo_format_reason("old CSet region num reached max") 2118 ergo_format_region("old") 2119 ergo_format_region("max"), 2120 old_cset_region_length(), max_old_cset_length); 2121 break; 2122 } 2123 2124 2125 // Stop adding regions if the remaining reclaimable space is 2126 // not above G1HeapWastePercent. 2127 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2128 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2129 double threshold = (double) G1HeapWastePercent; 2130 if (reclaimable_perc <= threshold) { 2131 // We've added enough old regions that the amount of uncollected 2132 // reclaimable space is at or below the waste threshold. Stop 2133 // adding old regions to the CSet. 2134 ergo_verbose5(ErgoCSetConstruction, 2135 "finish adding old regions to CSet", 2136 ergo_format_reason("reclaimable percentage not over threshold") 2137 ergo_format_region("old") 2138 ergo_format_region("max") 2139 ergo_format_byte_perc("reclaimable") 2140 ergo_format_perc("threshold"), 2141 old_cset_region_length(), 2142 max_old_cset_length, 2143 reclaimable_bytes, 2144 reclaimable_perc, threshold); 2145 break; 2146 } 2147 2148 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2149 if (check_time_remaining) { 2150 if (predicted_time_ms > time_remaining_ms) { 2151 // Too expensive for the current CSet. 2152 2153 if (old_cset_region_length() >= min_old_cset_length) { 2154 // We have added the minimum number of old regions to the CSet, 2155 // we are done with this CSet. 2156 ergo_verbose4(ErgoCSetConstruction, 2157 "finish adding old regions to CSet", 2158 ergo_format_reason("predicted time is too high") 2159 ergo_format_ms("predicted time") 2160 ergo_format_ms("remaining time") 2161 ergo_format_region("old") 2162 ergo_format_region("min"), 2163 predicted_time_ms, time_remaining_ms, 2164 old_cset_region_length(), min_old_cset_length); 2165 break; 2166 } 2167 2168 // We'll add it anyway given that we haven't reached the 2169 // minimum number of old regions. 2170 expensive_region_num += 1; 2171 } 2172 } else { 2173 if (old_cset_region_length() >= min_old_cset_length) { 2174 // In the non-auto-tuning case, we'll finish adding regions 2175 // to the CSet if we reach the minimum. 2176 ergo_verbose2(ErgoCSetConstruction, 2177 "finish adding old regions to CSet", 2178 ergo_format_reason("old CSet region num reached min") 2179 ergo_format_region("old") 2180 ergo_format_region("min"), 2181 old_cset_region_length(), min_old_cset_length); 2182 break; 2183 } 2184 } 2185 2186 // We will add this region to the CSet. 2187 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2188 predicted_old_time_ms += predicted_time_ms; 2189 cset_chooser->pop(); // already have region via peek() 2190 _g1->old_set_remove(hr); 2191 add_old_region_to_cset(hr); 2192 2193 hr = cset_chooser->peek(); 2194 } 2195 if (hr == NULL) { 2196 ergo_verbose0(ErgoCSetConstruction, 2197 "finish adding old regions to CSet", 2198 ergo_format_reason("candidate old regions not available")); 2199 } 2200 2201 if (expensive_region_num > 0) { 2202 // We print the information once here at the end, predicated on 2203 // whether we added any apparently expensive regions or not, to 2204 // avoid generating output per region. 2205 ergo_verbose4(ErgoCSetConstruction, 2206 "added expensive regions to CSet", 2207 ergo_format_reason("old CSet region num not reached min") 2208 ergo_format_region("old") 2209 ergo_format_region("expensive") 2210 ergo_format_region("min") 2211 ergo_format_ms("remaining time"), 2212 old_cset_region_length(), 2213 expensive_region_num, 2214 min_old_cset_length, 2215 time_remaining_ms); 2216 } 2217 2218 cset_chooser->verify(); 2219 } 2220 2221 stop_incremental_cset_building(); 2222 2223 ergo_verbose3(ErgoCSetConstruction, 2224 "finish choosing CSet", 2225 ergo_format_region("old") 2226 ergo_format_ms("predicted old region time") 2227 ergo_format_ms("time remaining"), 2228 old_cset_region_length(), 2229 predicted_old_time_ms, time_remaining_ms); 2230 2231 double non_young_end_time_sec = os::elapsedTime(); 2232 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2233 } 2234 2235 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2236 if(TraceYoungGenTime) { 2237 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2238 } 2239 } 2240 2241 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2242 if(TraceYoungGenTime) { 2243 _all_yield_times_ms.add(yield_time_ms); 2244 } 2245 } 2246 2247 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2248 if(TraceYoungGenTime) { 2249 _total.add(pause_time_ms); 2250 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2251 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2252 _parallel.add(phase_times->cur_collection_par_time_ms()); 2253 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2254 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2255 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2256 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2257 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2258 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2259 2260 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2261 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2262 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2263 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2264 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2265 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2266 2267 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2268 _parallel_other.add(parallel_other_time); 2269 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2270 } 2271 } 2272 2273 void TraceYoungGenTimeData::increment_young_collection_count() { 2274 if(TraceYoungGenTime) { 2275 ++_young_pause_num; 2276 } 2277 } 2278 2279 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2280 if(TraceYoungGenTime) { 2281 ++_mixed_pause_num; 2282 } 2283 } 2284 2285 void TraceYoungGenTimeData::print_summary(const char* str, 2286 const NumberSeq* seq) const { 2287 double sum = seq->sum(); 2288 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2289 str, sum / 1000.0, seq->avg()); 2290 } 2291 2292 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2293 const NumberSeq* seq) const { 2294 print_summary(str, seq); 2295 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2296 "(num", seq->num(), seq->sd(), seq->maximum()); 2297 } 2298 2299 void TraceYoungGenTimeData::print() const { 2300 if (!TraceYoungGenTime) { 2301 return; 2302 } 2303 2304 gclog_or_tty->print_cr("ALL PAUSES"); 2305 print_summary_sd(" Total", &_total); 2306 gclog_or_tty->cr(); 2307 gclog_or_tty->cr(); 2308 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2309 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2310 gclog_or_tty->cr(); 2311 2312 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2313 2314 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2315 gclog_or_tty->print_cr("none"); 2316 } else { 2317 print_summary_sd(" Evacuation Pauses", &_total); 2318 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2319 print_summary(" Parallel Time", &_parallel); 2320 print_summary(" Ext Root Scanning", &_ext_root_scan); 2321 print_summary(" SATB Filtering", &_satb_filtering); 2322 print_summary(" Update RS", &_update_rs); 2323 print_summary(" Scan RS", &_scan_rs); 2324 print_summary(" Object Copy", &_obj_copy); 2325 print_summary(" Termination", &_termination); 2326 print_summary(" Parallel Other", &_parallel_other); 2327 print_summary(" Clear CT", &_clear_ct); 2328 print_summary(" Other", &_other); 2329 } 2330 gclog_or_tty->cr(); 2331 2332 gclog_or_tty->print_cr("MISC"); 2333 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2334 print_summary_sd(" Yields", &_all_yield_times_ms); 2335 } 2336 2337 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2338 if (TraceOldGenTime) { 2339 _all_full_gc_times.add(full_gc_time_ms); 2340 } 2341 } 2342 2343 void TraceOldGenTimeData::print() const { 2344 if (!TraceOldGenTime) { 2345 return; 2346 } 2347 2348 if (_all_full_gc_times.num() > 0) { 2349 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2350 _all_full_gc_times.num(), 2351 _all_full_gc_times.sum() / 1000.0); 2352 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2353 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2354 _all_full_gc_times.sd(), 2355 _all_full_gc_times.maximum()); 2356 } 2357 }