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/g1GCPhaseTimes.hpp" 32 #include "gc/g1/heapRegion.inline.hpp" 33 #include "gc/g1/heapRegionRemSet.hpp" 34 #include "gc/shared/gcPolicyCounters.hpp" 35 #include "logging/log.hpp" 36 #include "runtime/arguments.hpp" 37 #include "runtime/java.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "utilities/debug.hpp" 40 41 // Different defaults for different number of GC threads 42 // They were chosen by running GCOld and SPECjbb on debris with different 43 // numbers of GC threads and choosing them based on the results 44 45 // all the same 46 static double rs_length_diff_defaults[] = { 47 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 48 }; 49 50 static double cost_per_card_ms_defaults[] = { 51 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 52 }; 53 54 // all the same 55 static double young_cards_per_entry_ratio_defaults[] = { 56 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 57 }; 58 59 static double cost_per_entry_ms_defaults[] = { 60 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 61 }; 62 63 static double cost_per_byte_ms_defaults[] = { 64 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 65 }; 66 67 // these should be pretty consistent 68 static double constant_other_time_ms_defaults[] = { 69 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 70 }; 71 72 73 static double young_other_cost_per_region_ms_defaults[] = { 74 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 75 }; 76 77 static double non_young_other_cost_per_region_ms_defaults[] = { 78 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 79 }; 80 81 G1CollectorPolicy::G1CollectorPolicy() : 82 _predictor(G1ConfidencePercent / 100.0), 83 _parallel_gc_threads(ParallelGCThreads), 84 85 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 86 _stop_world_start(0.0), 87 88 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 89 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 91 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 92 _prev_collection_pause_end_ms(0.0), 93 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 94 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _non_young_other_cost_per_region_ms_seq( 105 new TruncatedSeq(TruncatedSeqLength)), 106 107 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 108 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 109 110 _pause_time_target_ms((double) MaxGCPauseMillis), 111 112 _recent_prev_end_times_for_all_gcs_sec( 113 new TruncatedSeq(NumPrevPausesForHeuristics)), 114 115 _recent_avg_pause_time_ratio(0.0), 116 _rs_lengths_prediction(0), 117 _max_survivor_regions(0), 118 119 _eden_used_bytes_before_gc(0), 120 _survivor_used_bytes_before_gc(0), 121 _old_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 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 175 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 176 177 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 178 179 int index = MIN2(_parallel_gc_threads - 1, 7); 180 181 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 182 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 183 _cost_scan_hcc_seq->add(0.0); 184 _young_cards_per_entry_ratio_seq->add( 185 young_cards_per_entry_ratio_defaults[index]); 186 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 187 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 188 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 189 _young_other_cost_per_region_ms_seq->add( 190 young_other_cost_per_region_ms_defaults[index]); 191 _non_young_other_cost_per_region_ms_seq->add( 192 non_young_other_cost_per_region_ms_defaults[index]); 193 194 // Below, we might need to calculate the pause time target based on 195 // the pause interval. When we do so we are going to give G1 maximum 196 // flexibility and allow it to do pauses when it needs to. So, we'll 197 // arrange that the pause interval to be pause time target + 1 to 198 // ensure that a) the pause time target is maximized with respect to 199 // the pause interval and b) we maintain the invariant that pause 200 // time target < pause interval. If the user does not want this 201 // maximum flexibility, they will have to set the pause interval 202 // explicitly. 203 204 // First make sure that, if either parameter is set, its value is 205 // reasonable. 206 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 207 if (MaxGCPauseMillis < 1) { 208 vm_exit_during_initialization("MaxGCPauseMillis should be " 209 "greater than 0"); 210 } 211 } 212 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 213 if (GCPauseIntervalMillis < 1) { 214 vm_exit_during_initialization("GCPauseIntervalMillis should be " 215 "greater than 0"); 216 } 217 } 218 219 // Then, if the pause time target parameter was not set, set it to 220 // the default value. 221 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 222 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 223 // The default pause time target in G1 is 200ms 224 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 225 } else { 226 // We do not allow the pause interval to be set without the 227 // pause time target 228 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 229 "without setting MaxGCPauseMillis"); 230 } 231 } 232 233 // Then, if the interval parameter was not set, set it according to 234 // the pause time target (this will also deal with the case when the 235 // pause time target is the default value). 236 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 237 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 238 } 239 240 // Finally, make sure that the two parameters are consistent. 241 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 242 char buffer[256]; 243 jio_snprintf(buffer, 256, 244 "MaxGCPauseMillis (%u) should be less than " 245 "GCPauseIntervalMillis (%u)", 246 MaxGCPauseMillis, GCPauseIntervalMillis); 247 vm_exit_during_initialization(buffer); 248 } 249 250 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 251 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 252 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 253 254 // start conservatively (around 50ms is about right) 255 _concurrent_mark_remark_times_ms->add(0.05); 256 _concurrent_mark_cleanup_times_ms->add(0.20); 257 _tenuring_threshold = MaxTenuringThreshold; 258 259 assert(GCTimeRatio > 0, 260 "we should have set it to a default value set_g1_gc_flags() " 261 "if a user set it to 0"); 262 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 263 264 uintx reserve_perc = G1ReservePercent; 265 // Put an artificial ceiling on this so that it's not set to a silly value. 266 if (reserve_perc > 50) { 267 reserve_perc = 50; 268 warning("G1ReservePercent is set to a value that is too large, " 269 "it's been updated to " UINTX_FORMAT, reserve_perc); 270 } 271 _reserve_factor = (double) reserve_perc / 100.0; 272 // This will be set when the heap is expanded 273 // for the first time during initialization. 274 _reserve_regions = 0; 275 276 _collectionSetChooser = new CollectionSetChooser(); 277 } 278 279 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 280 return _predictor.get_new_prediction(seq); 281 } 282 283 void G1CollectorPolicy::initialize_alignments() { 284 _space_alignment = HeapRegion::GrainBytes; 285 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 286 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 287 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 288 } 289 290 void G1CollectorPolicy::initialize_flags() { 291 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 292 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 293 } 294 295 if (SurvivorRatio < 1) { 296 vm_exit_during_initialization("Invalid survivor ratio specified"); 297 } 298 CollectorPolicy::initialize_flags(); 299 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 300 } 301 302 void G1CollectorPolicy::post_heap_initialize() { 303 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 304 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 305 if (max_young_size != MaxNewSize) { 306 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 307 } 308 } 309 310 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 311 312 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 313 _min_desired_young_length(0), _max_desired_young_length(0) { 314 if (FLAG_IS_CMDLINE(NewRatio)) { 315 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 316 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 317 } else { 318 _sizer_kind = SizerNewRatio; 319 _adaptive_size = false; 320 return; 321 } 322 } 323 324 if (NewSize > MaxNewSize) { 325 if (FLAG_IS_CMDLINE(MaxNewSize)) { 326 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 327 "A new max generation size of " SIZE_FORMAT "k will be used.", 328 NewSize/K, MaxNewSize/K, NewSize/K); 329 } 330 MaxNewSize = NewSize; 331 } 332 333 if (FLAG_IS_CMDLINE(NewSize)) { 334 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 335 1U); 336 if (FLAG_IS_CMDLINE(MaxNewSize)) { 337 _max_desired_young_length = 338 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 339 1U); 340 _sizer_kind = SizerMaxAndNewSize; 341 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 342 } else { 343 _sizer_kind = SizerNewSizeOnly; 344 } 345 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 346 _max_desired_young_length = 347 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 348 1U); 349 _sizer_kind = SizerMaxNewSizeOnly; 350 } 351 } 352 353 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 354 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 355 return MAX2(1U, default_value); 356 } 357 358 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 359 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 360 return MAX2(1U, default_value); 361 } 362 363 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 364 assert(number_of_heap_regions > 0, "Heap must be initialized"); 365 366 switch (_sizer_kind) { 367 case SizerDefaults: 368 *min_young_length = calculate_default_min_length(number_of_heap_regions); 369 *max_young_length = calculate_default_max_length(number_of_heap_regions); 370 break; 371 case SizerNewSizeOnly: 372 *max_young_length = calculate_default_max_length(number_of_heap_regions); 373 *max_young_length = MAX2(*min_young_length, *max_young_length); 374 break; 375 case SizerMaxNewSizeOnly: 376 *min_young_length = calculate_default_min_length(number_of_heap_regions); 377 *min_young_length = MIN2(*min_young_length, *max_young_length); 378 break; 379 case SizerMaxAndNewSize: 380 // Do nothing. Values set on the command line, don't update them at runtime. 381 break; 382 case SizerNewRatio: 383 *min_young_length = number_of_heap_regions / (NewRatio + 1); 384 *max_young_length = *min_young_length; 385 break; 386 default: 387 ShouldNotReachHere(); 388 } 389 390 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 391 } 392 393 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 394 // We need to pass the desired values because recalculation may not update these 395 // values in some cases. 396 uint temp = _min_desired_young_length; 397 uint result = _max_desired_young_length; 398 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 399 return result; 400 } 401 402 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 403 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 404 &_max_desired_young_length); 405 } 406 407 void G1CollectorPolicy::init() { 408 // Set aside an initial future to_space. 409 _g1 = G1CollectedHeap::heap(); 410 411 assert(Heap_lock->owned_by_self(), "Locking discipline."); 412 413 initialize_gc_policy_counters(); 414 415 if (adaptive_young_list_length()) { 416 _young_list_fixed_length = 0; 417 } else { 418 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 419 } 420 _free_regions_at_end_of_collection = _g1->num_free_regions(); 421 422 update_young_list_max_and_target_length(); 423 // We may immediately start allocating regions and placing them on the 424 // collection set list. Initialize the per-collection set info 425 start_incremental_cset_building(); 426 } 427 428 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 429 phase_times()->note_gc_start(num_active_workers); 430 } 431 432 // Create the jstat counters for the policy. 433 void G1CollectorPolicy::initialize_gc_policy_counters() { 434 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 435 } 436 437 bool G1CollectorPolicy::predict_will_fit(uint young_length, 438 double base_time_ms, 439 uint base_free_regions, 440 double target_pause_time_ms) const { 441 if (young_length >= base_free_regions) { 442 // end condition 1: not enough space for the young regions 443 return false; 444 } 445 446 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 447 size_t bytes_to_copy = 448 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 449 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 450 double young_other_time_ms = predict_young_other_time_ms(young_length); 451 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 452 if (pause_time_ms > target_pause_time_ms) { 453 // end condition 2: prediction is over the target pause time 454 return false; 455 } 456 457 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 458 if ((2.0 /* magic */ * _predictor.sigma()) * bytes_to_copy > free_bytes) { 459 // end condition 3: out-of-space (conservatively!) 460 return false; 461 } 462 463 // success! 464 return true; 465 } 466 467 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 468 // re-calculate the necessary reserve 469 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 470 // We use ceiling so that if reserve_regions_d is > 0.0 (but 471 // smaller than 1.0) we'll get 1. 472 _reserve_regions = (uint) ceil(reserve_regions_d); 473 474 _young_gen_sizer->heap_size_changed(new_number_of_regions); 475 } 476 477 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 478 uint base_min_length) const { 479 uint desired_min_length = 0; 480 if (adaptive_young_list_length()) { 481 if (_alloc_rate_ms_seq->num() > 3) { 482 double now_sec = os::elapsedTime(); 483 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 484 double alloc_rate_ms = predict_alloc_rate_ms(); 485 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 486 } else { 487 // otherwise we don't have enough info to make the prediction 488 } 489 } 490 desired_min_length += base_min_length; 491 // make sure we don't go below any user-defined minimum bound 492 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 493 } 494 495 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 496 // Here, we might want to also take into account any additional 497 // constraints (i.e., user-defined minimum bound). Currently, we 498 // effectively don't set this bound. 499 return _young_gen_sizer->max_desired_young_length(); 500 } 501 502 void G1CollectorPolicy::update_young_list_max_and_target_length() { 503 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 504 } 505 506 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 507 update_young_list_target_length(rs_lengths); 508 update_max_gc_locker_expansion(); 509 } 510 511 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 512 _young_list_target_length = bounded_young_list_target_length(rs_lengths); 513 } 514 515 void G1CollectorPolicy::update_young_list_target_length() { 516 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 517 } 518 519 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths) const { 520 // Calculate the absolute and desired min bounds. 521 522 // This is how many young regions we already have (currently: the survivors). 523 uint base_min_length = recorded_survivor_regions(); 524 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 525 // This is the absolute minimum young length. Ensure that we 526 // will at least have one eden region available for allocation. 527 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 528 // If we shrank the young list target it should not shrink below the current size. 529 desired_min_length = MAX2(desired_min_length, absolute_min_length); 530 // Calculate the absolute and desired max bounds. 531 532 // We will try our best not to "eat" into the reserve. 533 uint absolute_max_length = 0; 534 if (_free_regions_at_end_of_collection > _reserve_regions) { 535 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 536 } 537 uint desired_max_length = calculate_young_list_desired_max_length(); 538 if (desired_max_length > absolute_max_length) { 539 desired_max_length = absolute_max_length; 540 } 541 542 uint young_list_target_length = 0; 543 if (adaptive_young_list_length()) { 544 if (collector_state()->gcs_are_young()) { 545 young_list_target_length = 546 calculate_young_list_target_length(rs_lengths, 547 base_min_length, 548 desired_min_length, 549 desired_max_length); 550 } else { 551 // Don't calculate anything and let the code below bound it to 552 // the desired_min_length, i.e., do the next GC as soon as 553 // possible to maximize how many old regions we can add to it. 554 } 555 } else { 556 // The user asked for a fixed young gen so we'll fix the young gen 557 // whether the next GC is young or mixed. 558 young_list_target_length = _young_list_fixed_length; 559 } 560 561 // Make sure we don't go over the desired max length, nor under the 562 // desired min length. In case they clash, desired_min_length wins 563 // which is why that test is second. 564 if (young_list_target_length > desired_max_length) { 565 young_list_target_length = desired_max_length; 566 } 567 if (young_list_target_length < desired_min_length) { 568 young_list_target_length = desired_min_length; 569 } 570 571 assert(young_list_target_length > recorded_survivor_regions(), 572 "we should be able to allocate at least one eden region"); 573 assert(young_list_target_length >= absolute_min_length, "post-condition"); 574 575 return young_list_target_length; 576 } 577 578 uint 579 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 580 uint base_min_length, 581 uint desired_min_length, 582 uint desired_max_length) const { 583 assert(adaptive_young_list_length(), "pre-condition"); 584 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 585 586 // In case some edge-condition makes the desired max length too small... 587 if (desired_max_length <= desired_min_length) { 588 return desired_min_length; 589 } 590 591 // We'll adjust min_young_length and max_young_length not to include 592 // the already allocated young regions (i.e., so they reflect the 593 // min and max eden regions we'll allocate). The base_min_length 594 // will be reflected in the predictions by the 595 // survivor_regions_evac_time prediction. 596 assert(desired_min_length > base_min_length, "invariant"); 597 uint min_young_length = desired_min_length - base_min_length; 598 assert(desired_max_length > base_min_length, "invariant"); 599 uint max_young_length = desired_max_length - base_min_length; 600 601 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 602 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 603 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 604 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 605 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 606 double base_time_ms = 607 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 608 survivor_regions_evac_time; 609 uint available_free_regions = _free_regions_at_end_of_collection; 610 uint base_free_regions = 0; 611 if (available_free_regions > _reserve_regions) { 612 base_free_regions = available_free_regions - _reserve_regions; 613 } 614 615 // Here, we will make sure that the shortest young length that 616 // makes sense fits within the target pause time. 617 618 if (predict_will_fit(min_young_length, base_time_ms, 619 base_free_regions, target_pause_time_ms)) { 620 // The shortest young length will fit into the target pause time; 621 // we'll now check whether the absolute maximum number of young 622 // regions will fit in the target pause time. If not, we'll do 623 // a binary search between min_young_length and max_young_length. 624 if (predict_will_fit(max_young_length, base_time_ms, 625 base_free_regions, target_pause_time_ms)) { 626 // The maximum young length will fit into the target pause time. 627 // We are done so set min young length to the maximum length (as 628 // the result is assumed to be returned in min_young_length). 629 min_young_length = max_young_length; 630 } else { 631 // The maximum possible number of young regions will not fit within 632 // the target pause time so we'll search for the optimal 633 // length. The loop invariants are: 634 // 635 // min_young_length < max_young_length 636 // min_young_length is known to fit into the target pause time 637 // max_young_length is known not to fit into the target pause time 638 // 639 // Going into the loop we know the above hold as we've just 640 // checked them. Every time around the loop we check whether 641 // the middle value between min_young_length and 642 // max_young_length fits into the target pause time. If it 643 // does, it becomes the new min. If it doesn't, it becomes 644 // the new max. This way we maintain the loop invariants. 645 646 assert(min_young_length < max_young_length, "invariant"); 647 uint diff = (max_young_length - min_young_length) / 2; 648 while (diff > 0) { 649 uint young_length = min_young_length + diff; 650 if (predict_will_fit(young_length, base_time_ms, 651 base_free_regions, target_pause_time_ms)) { 652 min_young_length = young_length; 653 } else { 654 max_young_length = young_length; 655 } 656 assert(min_young_length < max_young_length, "invariant"); 657 diff = (max_young_length - min_young_length) / 2; 658 } 659 // The results is min_young_length which, according to the 660 // loop invariants, should fit within the target pause time. 661 662 // These are the post-conditions of the binary search above: 663 assert(min_young_length < max_young_length, 664 "otherwise we should have discovered that max_young_length " 665 "fits into the pause target and not done the binary search"); 666 assert(predict_will_fit(min_young_length, base_time_ms, 667 base_free_regions, target_pause_time_ms), 668 "min_young_length, the result of the binary search, should " 669 "fit into the pause target"); 670 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 671 base_free_regions, target_pause_time_ms), 672 "min_young_length, the result of the binary search, should be " 673 "optimal, so no larger length should fit into the pause target"); 674 } 675 } else { 676 // Even the minimum length doesn't fit into the pause time 677 // target, return it as the result nevertheless. 678 } 679 return base_min_length + min_young_length; 680 } 681 682 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 683 double survivor_regions_evac_time = 0.0; 684 for (HeapRegion * r = _recorded_survivor_head; 685 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 686 r = r->get_next_young_region()) { 687 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 688 } 689 return survivor_regions_evac_time; 690 } 691 692 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 693 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 694 695 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 696 if (rs_lengths > _rs_lengths_prediction) { 697 // add 10% to avoid having to recalculate often 698 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 699 update_rs_lengths_prediction(rs_lengths_prediction); 700 701 update_young_list_max_and_target_length(rs_lengths_prediction); 702 } 703 } 704 705 void G1CollectorPolicy::update_rs_lengths_prediction() { 706 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 707 } 708 709 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 710 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 711 _rs_lengths_prediction = prediction; 712 } 713 } 714 715 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 716 bool is_tlab, 717 bool* gc_overhead_limit_was_exceeded) { 718 guarantee(false, "Not using this policy feature yet."); 719 return NULL; 720 } 721 722 // This method controls how a collector handles one or more 723 // of its generations being fully allocated. 724 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 725 bool is_tlab) { 726 guarantee(false, "Not using this policy feature yet."); 727 return NULL; 728 } 729 730 731 #ifndef PRODUCT 732 bool G1CollectorPolicy::verify_young_ages() { 733 HeapRegion* head = _g1->young_list()->first_region(); 734 return 735 verify_young_ages(head, _short_lived_surv_rate_group); 736 // also call verify_young_ages on any additional surv rate groups 737 } 738 739 bool 740 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 741 SurvRateGroup *surv_rate_group) { 742 guarantee( surv_rate_group != NULL, "pre-condition" ); 743 744 const char* name = surv_rate_group->name(); 745 bool ret = true; 746 int prev_age = -1; 747 748 for (HeapRegion* curr = head; 749 curr != NULL; 750 curr = curr->get_next_young_region()) { 751 SurvRateGroup* group = curr->surv_rate_group(); 752 if (group == NULL && !curr->is_survivor()) { 753 log_info(gc, verify)("## %s: encountered NULL surv_rate_group", name); 754 ret = false; 755 } 756 757 if (surv_rate_group == group) { 758 int age = curr->age_in_surv_rate_group(); 759 760 if (age < 0) { 761 log_info(gc, verify)("## %s: encountered negative age", name); 762 ret = false; 763 } 764 765 if (age <= prev_age) { 766 log_info(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age); 767 ret = false; 768 } 769 prev_age = age; 770 } 771 } 772 773 return ret; 774 } 775 #endif // PRODUCT 776 777 void G1CollectorPolicy::record_full_collection_start() { 778 _full_collection_start_sec = os::elapsedTime(); 779 record_heap_size_info_at_start(true /* full */); 780 // Release the future to-space so that it is available for compaction into. 781 collector_state()->set_full_collection(true); 782 } 783 784 void G1CollectorPolicy::record_full_collection_end() { 785 // Consider this like a collection pause for the purposes of allocation 786 // since last pause. 787 double end_sec = os::elapsedTime(); 788 double full_gc_time_sec = end_sec - _full_collection_start_sec; 789 double full_gc_time_ms = full_gc_time_sec * 1000.0; 790 791 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 792 793 update_recent_gc_times(end_sec, full_gc_time_ms); 794 795 collector_state()->set_full_collection(false); 796 797 // "Nuke" the heuristics that control the young/mixed GC 798 // transitions and make sure we start with young GCs after the Full GC. 799 collector_state()->set_gcs_are_young(true); 800 collector_state()->set_last_young_gc(false); 801 collector_state()->set_initiate_conc_mark_if_possible(false); 802 collector_state()->set_during_initial_mark_pause(false); 803 collector_state()->set_in_marking_window(false); 804 collector_state()->set_in_marking_window_im(false); 805 806 _short_lived_surv_rate_group->start_adding_regions(); 807 // also call this on any additional surv rate groups 808 809 record_survivor_regions(0, NULL, NULL); 810 811 _free_regions_at_end_of_collection = _g1->num_free_regions(); 812 // Reset survivors SurvRateGroup. 813 _survivor_surv_rate_group->reset(); 814 update_young_list_max_and_target_length(); 815 update_rs_lengths_prediction(); 816 _collectionSetChooser->clear(); 817 } 818 819 void G1CollectorPolicy::record_stop_world_start() { 820 _stop_world_start = os::elapsedTime(); 821 } 822 823 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 824 // We only need to do this here as the policy will only be applied 825 // to the GC we're about to start. so, no point is calculating this 826 // every time we calculate / recalculate the target young length. 827 update_survivors_policy(); 828 829 assert(_g1->used() == _g1->recalculate_used(), 830 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 831 _g1->used(), _g1->recalculate_used()); 832 833 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 834 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 835 _stop_world_start = 0.0; 836 837 record_heap_size_info_at_start(false /* full */); 838 839 phase_times()->record_cur_collection_start_sec(start_time_sec); 840 _pending_cards = _g1->pending_card_num(); 841 842 _collection_set_bytes_used_before = 0; 843 _bytes_copied_during_gc = 0; 844 845 collector_state()->set_last_gc_was_young(false); 846 847 // do that for any other surv rate groups 848 _short_lived_surv_rate_group->stop_adding_regions(); 849 _survivors_age_table.clear(); 850 851 assert( verify_young_ages(), "region age verification" ); 852 } 853 854 void G1CollectorPolicy::record_concurrent_mark_init_end(double 855 mark_init_elapsed_time_ms) { 856 collector_state()->set_during_marking(true); 857 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 858 collector_state()->set_during_initial_mark_pause(false); 859 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 860 } 861 862 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 863 _mark_remark_start_sec = os::elapsedTime(); 864 collector_state()->set_during_marking(false); 865 } 866 867 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 868 double end_time_sec = os::elapsedTime(); 869 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 870 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 871 _cur_mark_stop_world_time_ms += elapsed_time_ms; 872 _prev_collection_pause_end_ms += elapsed_time_ms; 873 874 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec); 875 } 876 877 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 878 _mark_cleanup_start_sec = os::elapsedTime(); 879 } 880 881 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 882 collector_state()->set_last_young_gc(true); 883 collector_state()->set_in_marking_window(false); 884 } 885 886 void G1CollectorPolicy::record_concurrent_pause() { 887 if (_stop_world_start > 0.0) { 888 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 889 _trace_young_gen_time_data.record_yield_time(yield_ms); 890 } 891 } 892 893 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 894 return phase_times()->average_time_ms(phase); 895 } 896 897 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 898 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 899 return false; 900 } 901 902 size_t marking_initiating_used_threshold = 903 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 904 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 905 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 906 907 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 908 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 909 log_debug(gc, ergo, conc)("Request concurrent cycle initiation (occupancy higher than threshold)" 910 "occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (" UINTX_FORMAT "%%) source: %s", 911 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, InitiatingHeapOccupancyPercent, source); 912 return true; 913 } else { 914 log_debug(gc, ergo, conc)("Do not request concurrent cycle initiation (still doing mixed collections)" 915 "occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (" UINTX_FORMAT "%%) source: %s", 916 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, InitiatingHeapOccupancyPercent, source); } 917 } 918 919 return false; 920 } 921 922 // Anything below that is considered to be zero 923 #define MIN_TIMER_GRANULARITY 0.0000001 924 925 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 926 double end_time_sec = os::elapsedTime(); 927 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 928 "otherwise, the subtraction below does not make sense"); 929 size_t rs_size = 930 _cur_collection_pause_used_regions_at_start - cset_region_length(); 931 size_t cur_used_bytes = _g1->used(); 932 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 933 bool last_pause_included_initial_mark = false; 934 bool update_stats = !_g1->evacuation_failed(); 935 936 NOT_PRODUCT(_short_lived_surv_rate_group->print()); 937 938 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 939 if (last_pause_included_initial_mark) { 940 record_concurrent_mark_init_end(0.0); 941 } else if (need_to_start_conc_mark("end of GC")) { 942 // Note: this might have already been set, if during the last 943 // pause we decided to start a cycle but at the beginning of 944 // this pause we decided to postpone it. That's OK. 945 collector_state()->set_initiate_conc_mark_if_possible(true); 946 } 947 948 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, end_time_sec); 949 950 if (update_stats) { 951 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 952 // this is where we update the allocation rate of the application 953 double app_time_ms = 954 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 955 if (app_time_ms < MIN_TIMER_GRANULARITY) { 956 // This usually happens due to the timer not having the required 957 // granularity. Some Linuxes are the usual culprits. 958 // We'll just set it to something (arbitrarily) small. 959 app_time_ms = 1.0; 960 } 961 // We maintain the invariant that all objects allocated by mutator 962 // threads will be allocated out of eden regions. So, we can use 963 // the eden region number allocated since the previous GC to 964 // calculate the application's allocate rate. The only exception 965 // to that is humongous objects that are allocated separately. But 966 // given that humongous object allocations do not really affect 967 // either the pause's duration nor when the next pause will take 968 // place we can safely ignore them here. 969 uint regions_allocated = eden_cset_region_length(); 970 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 971 _alloc_rate_ms_seq->add(alloc_rate_ms); 972 973 double interval_ms = 974 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 975 update_recent_gc_times(end_time_sec, pause_time_ms); 976 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 977 if (recent_avg_pause_time_ratio() < 0.0 || 978 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 979 #ifndef PRODUCT 980 // Dump info to allow post-facto debugging 981 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 982 gclog_or_tty->print_cr("-------------------------------------------"); 983 gclog_or_tty->print_cr("Recent GC Times (ms):"); 984 _recent_gc_times_ms->dump(); 985 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 986 _recent_prev_end_times_for_all_gcs_sec->dump(); 987 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 988 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 989 // In debug mode, terminate the JVM if the user wants to debug at this point. 990 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 991 #endif // !PRODUCT 992 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 993 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 994 if (_recent_avg_pause_time_ratio < 0.0) { 995 _recent_avg_pause_time_ratio = 0.0; 996 } else { 997 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 998 _recent_avg_pause_time_ratio = 1.0; 999 } 1000 } 1001 } 1002 1003 bool new_in_marking_window = collector_state()->in_marking_window(); 1004 bool new_in_marking_window_im = false; 1005 if (last_pause_included_initial_mark) { 1006 new_in_marking_window = true; 1007 new_in_marking_window_im = true; 1008 } 1009 1010 if (collector_state()->last_young_gc()) { 1011 // This is supposed to to be the "last young GC" before we start 1012 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1013 1014 if (!last_pause_included_initial_mark) { 1015 if (next_gc_should_be_mixed("start mixed GCs", 1016 "do not start mixed GCs")) { 1017 collector_state()->set_gcs_are_young(false); 1018 } 1019 } else { 1020 log_debug(gc, ergo)("Do not start mixed GCs (concurrent cycle is about to start)"); 1021 } 1022 collector_state()->set_last_young_gc(false); 1023 } 1024 1025 if (!collector_state()->last_gc_was_young()) { 1026 // This is a mixed GC. Here we decide whether to continue doing 1027 // mixed GCs or not. 1028 1029 if (!next_gc_should_be_mixed("continue mixed GCs", 1030 "do not continue mixed GCs")) { 1031 collector_state()->set_gcs_are_young(true); 1032 } 1033 } 1034 1035 _short_lived_surv_rate_group->start_adding_regions(); 1036 // Do that for any other surv rate groups 1037 1038 if (update_stats) { 1039 double cost_per_card_ms = 0.0; 1040 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1041 if (_pending_cards > 0) { 1042 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1043 _cost_per_card_ms_seq->add(cost_per_card_ms); 1044 } 1045 _cost_scan_hcc_seq->add(cost_scan_hcc); 1046 1047 double cost_per_entry_ms = 0.0; 1048 if (cards_scanned > 10) { 1049 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1050 if (collector_state()->last_gc_was_young()) { 1051 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1052 } else { 1053 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1054 } 1055 } 1056 1057 if (_max_rs_lengths > 0) { 1058 double cards_per_entry_ratio = 1059 (double) cards_scanned / (double) _max_rs_lengths; 1060 if (collector_state()->last_gc_was_young()) { 1061 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1062 } else { 1063 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1064 } 1065 } 1066 1067 // This is defensive. For a while _max_rs_lengths could get 1068 // smaller than _recorded_rs_lengths which was causing 1069 // rs_length_diff to get very large and mess up the RSet length 1070 // predictions. The reason was unsafe concurrent updates to the 1071 // _inc_cset_recorded_rs_lengths field which the code below guards 1072 // against (see CR 7118202). This bug has now been fixed (see CR 1073 // 7119027). However, I'm still worried that 1074 // _inc_cset_recorded_rs_lengths might still end up somewhat 1075 // inaccurate. The concurrent refinement thread calculates an 1076 // RSet's length concurrently with other CR threads updating it 1077 // which might cause it to calculate the length incorrectly (if, 1078 // say, it's in mid-coarsening). So I'll leave in the defensive 1079 // conditional below just in case. 1080 size_t rs_length_diff = 0; 1081 if (_max_rs_lengths > _recorded_rs_lengths) { 1082 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1083 } 1084 _rs_length_diff_seq->add((double) rs_length_diff); 1085 1086 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1087 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1088 double cost_per_byte_ms = 0.0; 1089 1090 if (copied_bytes > 0) { 1091 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1092 if (collector_state()->in_marking_window()) { 1093 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1094 } else { 1095 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1096 } 1097 } 1098 1099 double all_other_time_ms = pause_time_ms - 1100 (average_time_ms(G1GCPhaseTimes::UpdateRS) + average_time_ms(G1GCPhaseTimes::ScanRS) + 1101 average_time_ms(G1GCPhaseTimes::ObjCopy) + average_time_ms(G1GCPhaseTimes::Termination)); 1102 1103 double young_other_time_ms = 0.0; 1104 if (young_cset_region_length() > 0) { 1105 young_other_time_ms = 1106 phase_times()->young_cset_choice_time_ms() + 1107 phase_times()->young_free_cset_time_ms(); 1108 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1109 (double) young_cset_region_length()); 1110 } 1111 double non_young_other_time_ms = 0.0; 1112 if (old_cset_region_length() > 0) { 1113 non_young_other_time_ms = 1114 phase_times()->non_young_cset_choice_time_ms() + 1115 phase_times()->non_young_free_cset_time_ms(); 1116 1117 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1118 (double) old_cset_region_length()); 1119 } 1120 1121 double constant_other_time_ms = all_other_time_ms - 1122 (young_other_time_ms + non_young_other_time_ms); 1123 _constant_other_time_ms_seq->add(constant_other_time_ms); 1124 1125 double survival_ratio = 0.0; 1126 if (_collection_set_bytes_used_before > 0) { 1127 survival_ratio = (double) _bytes_copied_during_gc / 1128 (double) _collection_set_bytes_used_before; 1129 } 1130 1131 _pending_cards_seq->add((double) _pending_cards); 1132 _rs_lengths_seq->add((double) _max_rs_lengths); 1133 } 1134 1135 collector_state()->set_in_marking_window(new_in_marking_window); 1136 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1137 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1138 update_young_list_max_and_target_length(); 1139 update_rs_lengths_prediction(); 1140 1141 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1142 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1143 1144 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1145 1146 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1147 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 1148 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms", 1149 update_rs_time_goal_ms, scan_hcc_time_ms); 1150 1151 update_rs_time_goal_ms = 0; 1152 } else { 1153 update_rs_time_goal_ms -= scan_hcc_time_ms; 1154 } 1155 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1156 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1157 update_rs_time_goal_ms); 1158 1159 _collectionSetChooser->verify(); 1160 } 1161 1162 #define EXT_SIZE_FORMAT "%.1f%s" 1163 #define EXT_SIZE_PARAMS(bytes) \ 1164 byte_size_in_proper_unit((double)(bytes)), \ 1165 proper_unit_for_byte_size((bytes)) 1166 1167 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1168 YoungList* young_list = _g1->young_list(); 1169 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1170 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1171 _heap_capacity_bytes_before_gc = _g1->capacity(); 1172 _heap_used_bytes_before_gc = _g1->used(); 1173 _old_used_bytes_before_gc = _heap_used_bytes_before_gc - _survivor_used_bytes_before_gc - _eden_used_bytes_before_gc; 1174 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1175 1176 _eden_capacity_bytes_before_gc = 1177 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1178 1179 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1180 } 1181 1182 void G1CollectorPolicy::print_detailed_heap_transition() const { 1183 YoungList* young_list = _g1->young_list(); 1184 1185 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1186 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1187 size_t old_used_bytes_after_gc = _g1->used() - eden_used_bytes_after_gc - eden_used_bytes_after_gc; 1188 1189 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1190 size_t eden_capacity_bytes_after_gc = 1191 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1192 size_t survivor_capacity_bytes_after_gc = _max_survivor_regions * HeapRegion::GrainBytes; 1193 1194 log_info(gc, heap)("Eden: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1195 _eden_used_bytes_before_gc / K, eden_used_bytes_after_gc /K, eden_capacity_bytes_after_gc /K); 1196 1197 log_info(gc, heap)("Survivor: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1198 _survivor_used_bytes_before_gc / K, survivor_used_bytes_after_gc /K, survivor_capacity_bytes_after_gc /K); 1199 1200 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1201 _old_used_bytes_before_gc / K, old_used_bytes_after_gc /K, heap_capacity_bytes_after_gc /K); 1202 1203 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1204 } 1205 1206 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1207 phase_times()->print(pause_time_sec); 1208 } 1209 1210 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1211 double update_rs_processed_buffers, 1212 double goal_ms) { 1213 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1214 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1215 1216 if (G1UseAdaptiveConcRefinement) { 1217 const int k_gy = 3, k_gr = 6; 1218 const double inc_k = 1.1, dec_k = 0.9; 1219 1220 int g = cg1r->green_zone(); 1221 if (update_rs_time > goal_ms) { 1222 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1223 } else { 1224 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1225 g = (int)MAX2(g * inc_k, g + 1.0); 1226 } 1227 } 1228 // Change the refinement threads params 1229 cg1r->set_green_zone(g); 1230 cg1r->set_yellow_zone(g * k_gy); 1231 cg1r->set_red_zone(g * k_gr); 1232 cg1r->reinitialize_threads(); 1233 1234 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1235 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1236 cg1r->yellow_zone()); 1237 // Change the barrier params 1238 dcqs.set_process_completed_threshold(processing_threshold); 1239 dcqs.set_max_completed_queue(cg1r->red_zone()); 1240 } 1241 1242 int curr_queue_size = dcqs.completed_buffers_num(); 1243 if (curr_queue_size >= cg1r->yellow_zone()) { 1244 dcqs.set_completed_queue_padding(curr_queue_size); 1245 } else { 1246 dcqs.set_completed_queue_padding(0); 1247 } 1248 dcqs.notify_if_necessary(); 1249 } 1250 1251 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1252 return (size_t) get_new_prediction(_rs_length_diff_seq); 1253 } 1254 1255 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1256 return get_new_prediction(_alloc_rate_ms_seq); 1257 } 1258 1259 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1260 return get_new_prediction(_cost_per_card_ms_seq); 1261 } 1262 1263 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1264 return get_new_prediction(_cost_scan_hcc_seq); 1265 } 1266 1267 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1268 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1269 } 1270 1271 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1272 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1273 } 1274 1275 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1276 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1277 return predict_young_cards_per_entry_ratio(); 1278 } else { 1279 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1280 } 1281 } 1282 1283 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1284 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1285 } 1286 1287 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1288 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1289 } 1290 1291 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1292 if (collector_state()->gcs_are_young()) { 1293 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1294 } else { 1295 return predict_mixed_rs_scan_time_ms(card_num); 1296 } 1297 } 1298 1299 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1300 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1301 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1302 } else { 1303 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1304 } 1305 } 1306 1307 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1308 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1309 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1310 } else { 1311 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1312 } 1313 } 1314 1315 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1316 if (collector_state()->during_concurrent_mark()) { 1317 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1318 } else { 1319 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1320 } 1321 } 1322 1323 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1324 return get_new_prediction(_constant_other_time_ms_seq); 1325 } 1326 1327 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1328 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1329 } 1330 1331 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1332 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1333 } 1334 1335 double G1CollectorPolicy::predict_remark_time_ms() const { 1336 return get_new_prediction(_concurrent_mark_remark_times_ms); 1337 } 1338 1339 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1340 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1341 } 1342 1343 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1344 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1345 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1346 double pred = get_new_prediction(seq); 1347 if (pred > 1.0) { 1348 pred = 1.0; 1349 } 1350 return pred; 1351 } 1352 1353 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1354 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1355 } 1356 1357 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1358 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1359 } 1360 1361 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1362 size_t scanned_cards) const { 1363 return 1364 predict_rs_update_time_ms(pending_cards) + 1365 predict_rs_scan_time_ms(scanned_cards) + 1366 predict_constant_other_time_ms(); 1367 } 1368 1369 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1370 size_t rs_length = predict_rs_length_diff(); 1371 size_t card_num; 1372 if (collector_state()->gcs_are_young()) { 1373 card_num = predict_young_card_num(rs_length); 1374 } else { 1375 card_num = predict_non_young_card_num(rs_length); 1376 } 1377 return predict_base_elapsed_time_ms(pending_cards, card_num); 1378 } 1379 1380 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1381 size_t bytes_to_copy; 1382 if (hr->is_marked()) 1383 bytes_to_copy = hr->max_live_bytes(); 1384 else { 1385 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1386 int age = hr->age_in_surv_rate_group(); 1387 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1388 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1389 } 1390 return bytes_to_copy; 1391 } 1392 1393 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1394 bool for_young_gc) const { 1395 size_t rs_length = hr->rem_set()->occupied(); 1396 size_t card_num; 1397 1398 // Predicting the number of cards is based on which type of GC 1399 // we're predicting for. 1400 if (for_young_gc) { 1401 card_num = predict_young_card_num(rs_length); 1402 } else { 1403 card_num = predict_non_young_card_num(rs_length); 1404 } 1405 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1406 1407 double region_elapsed_time_ms = 1408 predict_rs_scan_time_ms(card_num) + 1409 predict_object_copy_time_ms(bytes_to_copy); 1410 1411 // The prediction of the "other" time for this region is based 1412 // upon the region type and NOT the GC type. 1413 if (hr->is_young()) { 1414 region_elapsed_time_ms += predict_young_other_time_ms(1); 1415 } else { 1416 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1417 } 1418 return region_elapsed_time_ms; 1419 } 1420 1421 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1422 uint survivor_cset_region_length) { 1423 _eden_cset_region_length = eden_cset_region_length; 1424 _survivor_cset_region_length = survivor_cset_region_length; 1425 _old_cset_region_length = 0; 1426 } 1427 1428 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1429 _recorded_rs_lengths = rs_lengths; 1430 } 1431 1432 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1433 double elapsed_ms) { 1434 _recent_gc_times_ms->add(elapsed_ms); 1435 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1436 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1437 } 1438 1439 size_t G1CollectorPolicy::expansion_amount() const { 1440 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1441 double threshold = _gc_overhead_perc; 1442 if (recent_gc_overhead > threshold) { 1443 // We will double the existing space, or take 1444 // G1ExpandByPercentOfAvailable % of the available expansion 1445 // space, whichever is smaller, bounded below by a minimum 1446 // expansion (unless that's all that's left.) 1447 const size_t min_expand_bytes = 1*M; 1448 size_t reserved_bytes = _g1->max_capacity(); 1449 size_t committed_bytes = _g1->capacity(); 1450 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1451 size_t expand_bytes; 1452 size_t expand_bytes_via_pct = 1453 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1454 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1455 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1456 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1457 1458 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) " 1459 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B calculated expansion amount: " SIZE_FORMAT "B (" INTX_FORMAT "%%)", 1460 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes_via_pct, G1ExpandByPercentOfAvailable); 1461 1462 return expand_bytes; 1463 } else { 1464 return 0; 1465 } 1466 } 1467 1468 void G1CollectorPolicy::print_tracing_info() const { 1469 _trace_young_gen_time_data.print(); 1470 _trace_old_gen_time_data.print(); 1471 } 1472 1473 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1474 #ifndef PRODUCT 1475 _short_lived_surv_rate_group->print_surv_rate_summary(); 1476 // add this call for any other surv rate groups 1477 #endif // PRODUCT 1478 } 1479 1480 bool G1CollectorPolicy::is_young_list_full() const { 1481 uint young_list_length = _g1->young_list()->length(); 1482 uint young_list_target_length = _young_list_target_length; 1483 return young_list_length >= young_list_target_length; 1484 } 1485 1486 bool G1CollectorPolicy::can_expand_young_list() const { 1487 uint young_list_length = _g1->young_list()->length(); 1488 uint young_list_max_length = _young_list_max_length; 1489 return young_list_length < young_list_max_length; 1490 } 1491 1492 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1493 uint expansion_region_num = 0; 1494 if (GCLockerEdenExpansionPercent > 0) { 1495 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1496 double expansion_region_num_d = perc * (double) _young_list_target_length; 1497 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1498 // less than 1.0) we'll get 1. 1499 expansion_region_num = (uint) ceil(expansion_region_num_d); 1500 } else { 1501 assert(expansion_region_num == 0, "sanity"); 1502 } 1503 _young_list_max_length = _young_list_target_length + expansion_region_num; 1504 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1505 } 1506 1507 // Calculates survivor space parameters. 1508 void G1CollectorPolicy::update_survivors_policy() { 1509 double max_survivor_regions_d = 1510 (double) _young_list_target_length / (double) SurvivorRatio; 1511 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1512 // smaller than 1.0) we'll get 1. 1513 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1514 1515 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1516 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1517 } 1518 1519 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 1520 GCCause::Cause gc_cause) { 1521 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1522 if (!during_cycle) { 1523 log_debug(gc, ergo, conc)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause)); 1524 collector_state()->set_initiate_conc_mark_if_possible(true); 1525 return true; 1526 } else { 1527 log_debug(gc, ergo, conc)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause)); 1528 return false; 1529 } 1530 } 1531 1532 void 1533 G1CollectorPolicy::decide_on_conc_mark_initiation() { 1534 // We are about to decide on whether this pause will be an 1535 // initial-mark pause. 1536 1537 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1538 // will set it here if we have to. However, it should be cleared by 1539 // the end of the pause (it's only set for the duration of an 1540 // initial-mark pause). 1541 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1542 1543 if (collector_state()->initiate_conc_mark_if_possible()) { 1544 // We had noticed on a previous pause that the heap occupancy has 1545 // gone over the initiating threshold and we should start a 1546 // concurrent marking cycle. So we might initiate one. 1547 1548 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1549 if (!during_cycle) { 1550 // The concurrent marking thread is not "during a cycle", i.e., 1551 // it has completed the last one. So we can go ahead and 1552 // initiate a new cycle. 1553 1554 collector_state()->set_during_initial_mark_pause(true); 1555 // We do not allow mixed GCs during marking. 1556 if (!collector_state()->gcs_are_young()) { 1557 collector_state()->set_gcs_are_young(true); 1558 log_debug(gc, ergo)("End mixed GCs (concurrent cycle is about to start"); 1559 } 1560 1561 // And we can now clear initiate_conc_mark_if_possible() as 1562 // we've already acted on it. 1563 collector_state()->set_initiate_conc_mark_if_possible(false); 1564 log_debug(gc, ergo, conc)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1565 } else { 1566 // The concurrent marking thread is still finishing up the 1567 // previous cycle. If we start one right now the two cycles 1568 // overlap. In particular, the concurrent marking thread might 1569 // be in the process of clearing the next marking bitmap (which 1570 // we will use for the next cycle if we start one). Starting a 1571 // cycle now will be bad given that parts of the marking 1572 // information might get cleared by the marking thread. And we 1573 // cannot wait for the marking thread to finish the cycle as it 1574 // periodically yields while clearing the next marking bitmap 1575 // and, if it's in a yield point, it's waiting for us to 1576 // finish. So, at this point we will not start a cycle and we'll 1577 // let the concurrent marking thread complete the last one. 1578 log_debug(gc, ergo, conc)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1579 } 1580 } 1581 } 1582 1583 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1584 G1CollectedHeap* _g1h; 1585 CSetChooserParUpdater _cset_updater; 1586 1587 public: 1588 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1589 uint chunk_size) : 1590 _g1h(G1CollectedHeap::heap()), 1591 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1592 1593 bool doHeapRegion(HeapRegion* r) { 1594 // Do we have any marking information for this region? 1595 if (r->is_marked()) { 1596 // We will skip any region that's currently used as an old GC 1597 // alloc region (we should not consider those for collection 1598 // before we fill them up). 1599 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1600 _cset_updater.add_region(r); 1601 } 1602 } 1603 return false; 1604 } 1605 }; 1606 1607 class ParKnownGarbageTask: public AbstractGangTask { 1608 CollectionSetChooser* _hrSorted; 1609 uint _chunk_size; 1610 G1CollectedHeap* _g1; 1611 HeapRegionClaimer _hrclaimer; 1612 1613 public: 1614 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1615 AbstractGangTask("ParKnownGarbageTask"), 1616 _hrSorted(hrSorted), _chunk_size(chunk_size), 1617 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1618 1619 void work(uint worker_id) { 1620 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1621 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1622 } 1623 }; 1624 1625 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1626 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1627 const uint overpartition_factor = 4; 1628 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1629 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1630 } 1631 1632 void 1633 G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1634 _collectionSetChooser->clear(); 1635 1636 WorkGang* workers = _g1->workers(); 1637 uint n_workers = workers->active_workers(); 1638 1639 uint n_regions = _g1->num_regions(); 1640 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1641 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1642 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1643 workers->run_task(&par_known_garbage_task); 1644 1645 _collectionSetChooser->sort_regions(); 1646 1647 double end_sec = os::elapsedTime(); 1648 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1649 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1650 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1651 _prev_collection_pause_end_ms += elapsed_time_ms; 1652 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec); 1653 } 1654 1655 // Add the heap region at the head of the non-incremental collection set 1656 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1657 assert(_inc_cset_build_state == Active, "Precondition"); 1658 assert(hr->is_old(), "the region should be old"); 1659 1660 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1661 _g1->register_old_region_with_cset(hr); 1662 hr->set_next_in_collection_set(_collection_set); 1663 _collection_set = hr; 1664 _collection_set_bytes_used_before += hr->used(); 1665 size_t rs_length = hr->rem_set()->occupied(); 1666 _recorded_rs_lengths += rs_length; 1667 _old_cset_region_length += 1; 1668 } 1669 1670 // Initialize the per-collection-set information 1671 void G1CollectorPolicy::start_incremental_cset_building() { 1672 assert(_inc_cset_build_state == Inactive, "Precondition"); 1673 1674 _inc_cset_head = NULL; 1675 _inc_cset_tail = NULL; 1676 _inc_cset_bytes_used_before = 0; 1677 1678 _inc_cset_max_finger = 0; 1679 _inc_cset_recorded_rs_lengths = 0; 1680 _inc_cset_recorded_rs_lengths_diffs = 0; 1681 _inc_cset_predicted_elapsed_time_ms = 0.0; 1682 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1683 _inc_cset_build_state = Active; 1684 } 1685 1686 void G1CollectorPolicy::finalize_incremental_cset_building() { 1687 assert(_inc_cset_build_state == Active, "Precondition"); 1688 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1689 1690 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1691 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1692 // that adds a new region to the CSet. Further updates by the 1693 // concurrent refinement thread that samples the young RSet lengths 1694 // are accumulated in the *_diffs fields. Here we add the diffs to 1695 // the "main" fields. 1696 1697 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1698 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1699 } else { 1700 // This is defensive. The diff should in theory be always positive 1701 // as RSets can only grow between GCs. However, given that we 1702 // sample their size concurrently with other threads updating them 1703 // it's possible that we might get the wrong size back, which 1704 // could make the calculations somewhat inaccurate. 1705 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1706 if (_inc_cset_recorded_rs_lengths >= diffs) { 1707 _inc_cset_recorded_rs_lengths -= diffs; 1708 } else { 1709 _inc_cset_recorded_rs_lengths = 0; 1710 } 1711 } 1712 _inc_cset_predicted_elapsed_time_ms += 1713 _inc_cset_predicted_elapsed_time_ms_diffs; 1714 1715 _inc_cset_recorded_rs_lengths_diffs = 0; 1716 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1717 } 1718 1719 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1720 // This routine is used when: 1721 // * adding survivor regions to the incremental cset at the end of an 1722 // evacuation pause, 1723 // * adding the current allocation region to the incremental cset 1724 // when it is retired, and 1725 // * updating existing policy information for a region in the 1726 // incremental cset via young list RSet sampling. 1727 // Therefore this routine may be called at a safepoint by the 1728 // VM thread, or in-between safepoints by mutator threads (when 1729 // retiring the current allocation region) or a concurrent 1730 // refine thread (RSet sampling). 1731 1732 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1733 size_t used_bytes = hr->used(); 1734 _inc_cset_recorded_rs_lengths += rs_length; 1735 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1736 _inc_cset_bytes_used_before += used_bytes; 1737 1738 // Cache the values we have added to the aggregated information 1739 // in the heap region in case we have to remove this region from 1740 // the incremental collection set, or it is updated by the 1741 // rset sampling code 1742 hr->set_recorded_rs_length(rs_length); 1743 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1744 } 1745 1746 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1747 size_t new_rs_length) { 1748 // Update the CSet information that is dependent on the new RS length 1749 assert(hr->is_young(), "Precondition"); 1750 assert(!SafepointSynchronize::is_at_safepoint(), 1751 "should not be at a safepoint"); 1752 1753 // We could have updated _inc_cset_recorded_rs_lengths and 1754 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1755 // that atomically, as this code is executed by a concurrent 1756 // refinement thread, potentially concurrently with a mutator thread 1757 // allocating a new region and also updating the same fields. To 1758 // avoid the atomic operations we accumulate these updates on two 1759 // separate fields (*_diffs) and we'll just add them to the "main" 1760 // fields at the start of a GC. 1761 1762 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1763 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1764 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1765 1766 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1767 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1768 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1769 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1770 1771 hr->set_recorded_rs_length(new_rs_length); 1772 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1773 } 1774 1775 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1776 assert(hr->is_young(), "invariant"); 1777 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1778 assert(_inc_cset_build_state == Active, "Precondition"); 1779 1780 // We need to clear and set the cached recorded/cached collection set 1781 // information in the heap region here (before the region gets added 1782 // to the collection set). An individual heap region's cached values 1783 // are calculated, aggregated with the policy collection set info, 1784 // and cached in the heap region here (initially) and (subsequently) 1785 // by the Young List sampling code. 1786 1787 size_t rs_length = hr->rem_set()->occupied(); 1788 add_to_incremental_cset_info(hr, rs_length); 1789 1790 HeapWord* hr_end = hr->end(); 1791 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1792 1793 assert(!hr->in_collection_set(), "invariant"); 1794 _g1->register_young_region_with_cset(hr); 1795 assert(hr->next_in_collection_set() == NULL, "invariant"); 1796 } 1797 1798 // Add the region at the RHS of the incremental cset 1799 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1800 // We should only ever be appending survivors at the end of a pause 1801 assert(hr->is_survivor(), "Logic"); 1802 1803 // Do the 'common' stuff 1804 add_region_to_incremental_cset_common(hr); 1805 1806 // Now add the region at the right hand side 1807 if (_inc_cset_tail == NULL) { 1808 assert(_inc_cset_head == NULL, "invariant"); 1809 _inc_cset_head = hr; 1810 } else { 1811 _inc_cset_tail->set_next_in_collection_set(hr); 1812 } 1813 _inc_cset_tail = hr; 1814 } 1815 1816 // Add the region to the LHS of the incremental cset 1817 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 1818 // Survivors should be added to the RHS at the end of a pause 1819 assert(hr->is_eden(), "Logic"); 1820 1821 // Do the 'common' stuff 1822 add_region_to_incremental_cset_common(hr); 1823 1824 // Add the region at the left hand side 1825 hr->set_next_in_collection_set(_inc_cset_head); 1826 if (_inc_cset_head == NULL) { 1827 assert(_inc_cset_tail == NULL, "Invariant"); 1828 _inc_cset_tail = hr; 1829 } 1830 _inc_cset_head = hr; 1831 } 1832 1833 #ifndef PRODUCT 1834 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 1835 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 1836 1837 st->print_cr("\nCollection_set:"); 1838 HeapRegion* csr = list_head; 1839 while (csr != NULL) { 1840 HeapRegion* next = csr->next_in_collection_set(); 1841 assert(csr->in_collection_set(), "bad CS"); 1842 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 1843 HR_FORMAT_PARAMS(csr), 1844 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 1845 csr->age_in_surv_rate_group_cond()); 1846 csr = next; 1847 } 1848 } 1849 #endif // !PRODUCT 1850 1851 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 1852 // Returns the given amount of reclaimable bytes (that represents 1853 // the amount of reclaimable space still to be collected) as a 1854 // percentage of the current heap capacity. 1855 size_t capacity_bytes = _g1->capacity(); 1856 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 1857 } 1858 1859 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 1860 const char* false_action_str) const { 1861 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1862 if (cset_chooser->is_empty()) { 1863 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 1864 return false; 1865 } 1866 1867 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1868 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 1869 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 1870 double threshold = (double) G1HeapWastePercent; 1871 if (reclaimable_perc <= threshold) { 1872 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1873 false_action_str, cset_chooser->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 1874 return false; 1875 } 1876 1877 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1878 true_action_str, cset_chooser->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 1879 return true; 1880 } 1881 1882 uint G1CollectorPolicy::calc_min_old_cset_length() const { 1883 // The min old CSet region bound is based on the maximum desired 1884 // number of mixed GCs after a cycle. I.e., even if some old regions 1885 // look expensive, we should add them to the CSet anyway to make 1886 // sure we go through the available old regions in no more than the 1887 // maximum desired number of mixed GCs. 1888 // 1889 // The calculation is based on the number of marked regions we added 1890 // to the CSet chooser in the first place, not how many remain, so 1891 // that the result is the same during all mixed GCs that follow a cycle. 1892 1893 const size_t region_num = (size_t) _collectionSetChooser->length(); 1894 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 1895 size_t result = region_num / gc_num; 1896 // emulate ceiling 1897 if (result * gc_num < region_num) { 1898 result += 1; 1899 } 1900 return (uint) result; 1901 } 1902 1903 uint G1CollectorPolicy::calc_max_old_cset_length() const { 1904 // The max old CSet region bound is based on the threshold expressed 1905 // as a percentage of the heap size. I.e., it should bound the 1906 // number of old regions added to the CSet irrespective of how many 1907 // of them are available. 1908 1909 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1910 const size_t region_num = g1h->num_regions(); 1911 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 1912 size_t result = region_num * perc / 100; 1913 // emulate ceiling 1914 if (100 * result < region_num * perc) { 1915 result += 1; 1916 } 1917 return (uint) result; 1918 } 1919 1920 1921 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 1922 double young_start_time_sec = os::elapsedTime(); 1923 1924 YoungList* young_list = _g1->young_list(); 1925 finalize_incremental_cset_building(); 1926 1927 guarantee(target_pause_time_ms > 0.0, 1928 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 1929 guarantee(_collection_set == NULL, "Precondition"); 1930 1931 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 1932 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 1933 1934 log_trace(gc, ergo, cset)("Start choosing CSet. pending cards: " SIZE_FORMAT " predicted base time: %1.2fms remaining time: %1.2fms target pause time: %1.2fms", 1935 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 1936 1937 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 1938 1939 if (collector_state()->last_gc_was_young()) { 1940 _trace_young_gen_time_data.increment_young_collection_count(); 1941 } else { 1942 _trace_young_gen_time_data.increment_mixed_collection_count(); 1943 } 1944 1945 // The young list is laid with the survivor regions from the previous 1946 // pause are appended to the RHS of the young list, i.e. 1947 // [Newly Young Regions ++ Survivors from last pause]. 1948 1949 uint survivor_region_length = young_list->survivor_length(); 1950 uint eden_region_length = young_list->eden_length(); 1951 init_cset_region_lengths(eden_region_length, survivor_region_length); 1952 1953 HeapRegion* hr = young_list->first_survivor_region(); 1954 while (hr != NULL) { 1955 assert(hr->is_survivor(), "badly formed young list"); 1956 // There is a convention that all the young regions in the CSet 1957 // are tagged as "eden", so we do this for the survivors here. We 1958 // use the special set_eden_pre_gc() as it doesn't check that the 1959 // region is free (which is not the case here). 1960 hr->set_eden_pre_gc(); 1961 hr = hr->get_next_young_region(); 1962 } 1963 1964 // Clear the fields that point to the survivor list - they are all young now. 1965 young_list->clear_survivors(); 1966 1967 _collection_set = _inc_cset_head; 1968 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 1969 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 1970 1971 log_trace(gc, ergo, cset)("Add young regions to CSet. eden: %u regions, survivors: %u regions, predicted young region time: %1.2fms, target pause time: %1.2fms", 1972 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms); 1973 1974 // The number of recorded young regions is the incremental 1975 // collection set's current size 1976 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 1977 1978 double young_end_time_sec = os::elapsedTime(); 1979 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 1980 1981 return time_remaining_ms; 1982 } 1983 1984 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 1985 double non_young_start_time_sec = os::elapsedTime(); 1986 double predicted_old_time_ms = 0.0; 1987 1988 1989 if (!collector_state()->gcs_are_young()) { 1990 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1991 cset_chooser->verify(); 1992 const uint min_old_cset_length = calc_min_old_cset_length(); 1993 const uint max_old_cset_length = calc_max_old_cset_length(); 1994 1995 uint expensive_region_num = 0; 1996 bool check_time_remaining = adaptive_young_list_length(); 1997 1998 HeapRegion* hr = cset_chooser->peek(); 1999 while (hr != NULL) { 2000 if (old_cset_region_length() >= max_old_cset_length) { 2001 // Added maximum number of old regions to the CSet. 2002 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions", 2003 old_cset_region_length(), max_old_cset_length); 2004 break; 2005 } 2006 2007 2008 // Stop adding regions if the remaining reclaimable space is 2009 // not above G1HeapWastePercent. 2010 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2011 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2012 double threshold = (double) G1HeapWastePercent; 2013 if (reclaimable_perc <= threshold) { 2014 // We've added enough old regions that the amount of uncollected 2015 // reclaimable space is at or below the waste threshold. Stop 2016 // adding old regions to the CSet. 2017 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). " 2018 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%", 2019 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2020 break; 2021 } 2022 2023 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2024 if (check_time_remaining) { 2025 if (predicted_time_ms > time_remaining_ms) { 2026 // Too expensive for the current CSet. 2027 2028 if (old_cset_region_length() >= min_old_cset_length) { 2029 // We have added the minimum number of old regions to the CSet, 2030 // we are done with this CSet. 2031 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). " 2032 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions", 2033 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length); 2034 break; 2035 } 2036 2037 // We'll add it anyway given that we haven't reached the 2038 // minimum number of old regions. 2039 expensive_region_num += 1; 2040 } 2041 } else { 2042 if (old_cset_region_length() >= min_old_cset_length) { 2043 // In the non-auto-tuning case, we'll finish adding regions 2044 // to the CSet if we reach the minimum. 2045 2046 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions", 2047 old_cset_region_length(), min_old_cset_length); 2048 break; 2049 } 2050 } 2051 2052 // We will add this region to the CSet. 2053 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2054 predicted_old_time_ms += predicted_time_ms; 2055 cset_chooser->pop(); // already have region via peek() 2056 _g1->old_set_remove(hr); 2057 add_old_region_to_cset(hr); 2058 2059 hr = cset_chooser->peek(); 2060 } 2061 if (hr == NULL) { 2062 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)"); 2063 } 2064 2065 if (expensive_region_num > 0) { 2066 // We print the information once here at the end, predicated on 2067 // whether we added any apparently expensive regions or not, to 2068 // avoid generating output per region. 2069 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)." 2070 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms", 2071 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms); 2072 } 2073 2074 cset_chooser->verify(); 2075 } 2076 2077 stop_incremental_cset_building(); 2078 2079 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f", 2080 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms); 2081 2082 double non_young_end_time_sec = os::elapsedTime(); 2083 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2084 } 2085 2086 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2087 if(TraceYoungGenTime) { 2088 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2089 } 2090 } 2091 2092 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2093 if(TraceYoungGenTime) { 2094 _all_yield_times_ms.add(yield_time_ms); 2095 } 2096 } 2097 2098 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2099 if(TraceYoungGenTime) { 2100 _total.add(pause_time_ms); 2101 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2102 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2103 _parallel.add(phase_times->cur_collection_par_time_ms()); 2104 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2105 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2106 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2107 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2108 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2109 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2110 2111 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2112 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2113 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2114 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2115 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2116 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2117 2118 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2119 _parallel_other.add(parallel_other_time); 2120 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2121 } 2122 } 2123 2124 void TraceYoungGenTimeData::increment_young_collection_count() { 2125 if(TraceYoungGenTime) { 2126 ++_young_pause_num; 2127 } 2128 } 2129 2130 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2131 if(TraceYoungGenTime) { 2132 ++_mixed_pause_num; 2133 } 2134 } 2135 2136 void TraceYoungGenTimeData::print_summary(const char* str, 2137 const NumberSeq* seq) const { 2138 double sum = seq->sum(); 2139 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2140 str, sum / 1000.0, seq->avg()); 2141 } 2142 2143 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2144 const NumberSeq* seq) const { 2145 print_summary(str, seq); 2146 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2147 "(num", seq->num(), seq->sd(), seq->maximum()); 2148 } 2149 2150 void TraceYoungGenTimeData::print() const { 2151 if (!TraceYoungGenTime) { 2152 return; 2153 } 2154 2155 gclog_or_tty->print_cr("ALL PAUSES"); 2156 print_summary_sd(" Total", &_total); 2157 gclog_or_tty->cr(); 2158 gclog_or_tty->cr(); 2159 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2160 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2161 gclog_or_tty->cr(); 2162 2163 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2164 2165 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2166 gclog_or_tty->print_cr("none"); 2167 } else { 2168 print_summary_sd(" Evacuation Pauses", &_total); 2169 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2170 print_summary(" Parallel Time", &_parallel); 2171 print_summary(" Ext Root Scanning", &_ext_root_scan); 2172 print_summary(" SATB Filtering", &_satb_filtering); 2173 print_summary(" Update RS", &_update_rs); 2174 print_summary(" Scan RS", &_scan_rs); 2175 print_summary(" Object Copy", &_obj_copy); 2176 print_summary(" Termination", &_termination); 2177 print_summary(" Parallel Other", &_parallel_other); 2178 print_summary(" Clear CT", &_clear_ct); 2179 print_summary(" Other", &_other); 2180 } 2181 gclog_or_tty->cr(); 2182 2183 gclog_or_tty->print_cr("MISC"); 2184 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2185 print_summary_sd(" Yields", &_all_yield_times_ms); 2186 } 2187 2188 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2189 if (TraceOldGenTime) { 2190 _all_full_gc_times.add(full_gc_time_ms); 2191 } 2192 } 2193 2194 void TraceOldGenTimeData::print() const { 2195 if (!TraceOldGenTime) { 2196 return; 2197 } 2198 2199 if (_all_full_gc_times.num() > 0) { 2200 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2201 _all_full_gc_times.num(), 2202 _all_full_gc_times.sum() / 1000.0); 2203 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2204 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2205 _all_full_gc_times.sd(), 2206 _all_full_gc_times.maximum()); 2207 } 2208 }