1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/concurrentG1Refine.hpp" 27 #include "gc/g1/concurrentMark.hpp" 28 #include "gc/g1/concurrentMarkThread.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorPolicy.hpp" 31 #include "gc/g1/g1ErgoVerbose.hpp" 32 #include "gc/g1/g1GCPhaseTimes.hpp" 33 #include "gc/g1/g1Log.hpp" 34 #include "gc/g1/heapRegion.inline.hpp" 35 #include "gc/g1/heapRegionRemSet.hpp" 36 #include "gc/shared/gcPolicyCounters.hpp" 37 #include "runtime/arguments.hpp" 38 #include "runtime/java.hpp" 39 #include "runtime/mutexLocker.hpp" 40 #include "utilities/debug.hpp" 41 42 // Different defaults for different number of GC threads 43 // They were chosen by running GCOld and SPECjbb on debris with different 44 // numbers of GC threads and choosing them based on the results 45 46 // all the same 47 static double rs_length_diff_defaults[] = { 48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 49 }; 50 51 static double cost_per_card_ms_defaults[] = { 52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 53 }; 54 55 // all the same 56 static double young_cards_per_entry_ratio_defaults[] = { 57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 58 }; 59 60 static double cost_per_entry_ms_defaults[] = { 61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 62 }; 63 64 static double cost_per_byte_ms_defaults[] = { 65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 66 }; 67 68 // these should be pretty consistent 69 static double constant_other_time_ms_defaults[] = { 70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 71 }; 72 73 74 static double young_other_cost_per_region_ms_defaults[] = { 75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 76 }; 77 78 static double non_young_other_cost_per_region_ms_defaults[] = { 79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 80 }; 81 82 G1CollectorPolicy::G1CollectorPolicy() : 83 _predictor(G1ConfidencePercent / 100.0), 84 _parallel_gc_threads(ParallelGCThreads), 85 86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 87 _stop_world_start(0.0), 88 89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 93 _prev_collection_pause_end_ms(0.0), 94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _non_young_other_cost_per_region_ms_seq( 106 new TruncatedSeq(TruncatedSeqLength)), 107 108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 110 111 _pause_time_target_ms((double) MaxGCPauseMillis), 112 113 _recent_prev_end_times_for_all_gcs_sec( 114 new TruncatedSeq(NumPrevPausesForHeuristics)), 115 116 _recent_avg_pause_time_ratio(0.0), 117 _rs_lengths_prediction(0), 118 _max_survivor_regions(0), 119 120 _eden_used_bytes_before_gc(0), 121 _survivor_used_bytes_before_gc(0), 122 _heap_used_bytes_before_gc(0), 123 _metaspace_used_bytes_before_gc(0), 124 _eden_capacity_bytes_before_gc(0), 125 _heap_capacity_bytes_before_gc(0), 126 127 _eden_cset_region_length(0), 128 _survivor_cset_region_length(0), 129 _old_cset_region_length(0), 130 131 _collection_set(NULL), 132 _collection_set_bytes_used_before(0), 133 134 // Incremental CSet attributes 135 _inc_cset_build_state(Inactive), 136 _inc_cset_head(NULL), 137 _inc_cset_tail(NULL), 138 _inc_cset_bytes_used_before(0), 139 _inc_cset_max_finger(NULL), 140 _inc_cset_recorded_rs_lengths(0), 141 _inc_cset_recorded_rs_lengths_diffs(0), 142 _inc_cset_predicted_elapsed_time_ms(0.0), 143 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 144 145 // add here any more surv rate groups 146 _recorded_survivor_regions(0), 147 _recorded_survivor_head(NULL), 148 _recorded_survivor_tail(NULL), 149 _survivors_age_table(true), 150 151 _gc_overhead_perc(0.0) { 152 153 // SurvRateGroups below must be initialized after the predictor because they 154 // indirectly use it through this object passed to their constructor. 155 _short_lived_surv_rate_group = 156 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 157 _survivor_surv_rate_group = 158 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 159 160 // Set up the region size and associated fields. Given that the 161 // policy is created before the heap, we have to set this up here, 162 // so it's done as soon as possible. 163 164 // It would have been natural to pass initial_heap_byte_size() and 165 // max_heap_byte_size() to setup_heap_region_size() but those have 166 // not been set up at this point since they should be aligned with 167 // the region size. So, there is a circular dependency here. We base 168 // the region size on the heap size, but the heap size should be 169 // aligned with the region size. To get around this we use the 170 // unaligned values for the heap. 171 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 172 HeapRegionRemSet::setup_remset_size(); 173 174 G1ErgoVerbose::initialize(); 175 if (PrintAdaptiveSizePolicy) { 176 // Currently, we only use a single switch for all the heuristics. 177 G1ErgoVerbose::set_enabled(true); 178 // Given that we don't currently have a verboseness level 179 // parameter, we'll hardcode this to high. This can be easily 180 // changed in the future. 181 G1ErgoVerbose::set_level(ErgoHigh); 182 } else { 183 G1ErgoVerbose::set_enabled(false); 184 } 185 186 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 187 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 188 189 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 190 191 int index = MIN2(_parallel_gc_threads - 1, 7); 192 193 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 194 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 195 _cost_scan_hcc_seq->add(0.0); 196 _young_cards_per_entry_ratio_seq->add( 197 young_cards_per_entry_ratio_defaults[index]); 198 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 199 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 200 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 201 _young_other_cost_per_region_ms_seq->add( 202 young_other_cost_per_region_ms_defaults[index]); 203 _non_young_other_cost_per_region_ms_seq->add( 204 non_young_other_cost_per_region_ms_defaults[index]); 205 206 // Below, we might need to calculate the pause time target based on 207 // the pause interval. When we do so we are going to give G1 maximum 208 // flexibility and allow it to do pauses when it needs to. So, we'll 209 // arrange that the pause interval to be pause time target + 1 to 210 // ensure that a) the pause time target is maximized with respect to 211 // the pause interval and b) we maintain the invariant that pause 212 // time target < pause interval. If the user does not want this 213 // maximum flexibility, they will have to set the pause interval 214 // explicitly. 215 216 // First make sure that, if either parameter is set, its value is 217 // reasonable. 218 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 219 if (MaxGCPauseMillis < 1) { 220 vm_exit_during_initialization("MaxGCPauseMillis should be " 221 "greater than 0"); 222 } 223 } 224 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 225 if (GCPauseIntervalMillis < 1) { 226 vm_exit_during_initialization("GCPauseIntervalMillis should be " 227 "greater than 0"); 228 } 229 } 230 231 // Then, if the pause time target parameter was not set, set it to 232 // the default value. 233 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 234 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 235 // The default pause time target in G1 is 200ms 236 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 237 } else { 238 // We do not allow the pause interval to be set without the 239 // pause time target 240 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 241 "without setting MaxGCPauseMillis"); 242 } 243 } 244 245 // Then, if the interval parameter was not set, set it according to 246 // the pause time target (this will also deal with the case when the 247 // pause time target is the default value). 248 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 249 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 250 } 251 252 // Finally, make sure that the two parameters are consistent. 253 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 254 char buffer[256]; 255 jio_snprintf(buffer, 256, 256 "MaxGCPauseMillis (%u) should be less than " 257 "GCPauseIntervalMillis (%u)", 258 MaxGCPauseMillis, GCPauseIntervalMillis); 259 vm_exit_during_initialization(buffer); 260 } 261 262 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 263 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 264 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 265 266 // start conservatively (around 50ms is about right) 267 _concurrent_mark_remark_times_ms->add(0.05); 268 _concurrent_mark_cleanup_times_ms->add(0.20); 269 _tenuring_threshold = MaxTenuringThreshold; 270 271 assert(GCTimeRatio > 0, 272 "we should have set it to a default value set_g1_gc_flags() " 273 "if a user set it to 0"); 274 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 275 276 uintx reserve_perc = G1ReservePercent; 277 // Put an artificial ceiling on this so that it's not set to a silly value. 278 if (reserve_perc > 50) { 279 reserve_perc = 50; 280 warning("G1ReservePercent is set to a value that is too large, " 281 "it's been updated to " UINTX_FORMAT, reserve_perc); 282 } 283 _reserve_factor = (double) reserve_perc / 100.0; 284 // This will be set when the heap is expanded 285 // for the first time during initialization. 286 _reserve_regions = 0; 287 288 _collectionSetChooser = new CollectionSetChooser(); 289 } 290 291 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 292 return _predictor.get_new_prediction(seq); 293 } 294 295 void G1CollectorPolicy::initialize_alignments() { 296 _space_alignment = HeapRegion::GrainBytes; 297 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 298 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 299 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 300 } 301 302 void G1CollectorPolicy::initialize_flags() { 303 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 304 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 305 } 306 307 if (SurvivorRatio < 1) { 308 vm_exit_during_initialization("Invalid survivor ratio specified"); 309 } 310 CollectorPolicy::initialize_flags(); 311 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 312 } 313 314 void G1CollectorPolicy::post_heap_initialize() { 315 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 316 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 317 if (max_young_size != MaxNewSize) { 318 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 319 } 320 } 321 322 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 323 324 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 325 _min_desired_young_length(0), _max_desired_young_length(0) { 326 if (FLAG_IS_CMDLINE(NewRatio)) { 327 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 328 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 329 } else { 330 _sizer_kind = SizerNewRatio; 331 _adaptive_size = false; 332 return; 333 } 334 } 335 336 if (NewSize > MaxNewSize) { 337 if (FLAG_IS_CMDLINE(MaxNewSize)) { 338 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 339 "A new max generation size of " SIZE_FORMAT "k will be used.", 340 NewSize/K, MaxNewSize/K, NewSize/K); 341 } 342 MaxNewSize = NewSize; 343 } 344 345 if (FLAG_IS_CMDLINE(NewSize)) { 346 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 347 1U); 348 if (FLAG_IS_CMDLINE(MaxNewSize)) { 349 _max_desired_young_length = 350 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 351 1U); 352 _sizer_kind = SizerMaxAndNewSize; 353 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 354 } else { 355 _sizer_kind = SizerNewSizeOnly; 356 } 357 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 358 _max_desired_young_length = 359 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 360 1U); 361 _sizer_kind = SizerMaxNewSizeOnly; 362 } 363 } 364 365 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 366 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 367 return MAX2(1U, default_value); 368 } 369 370 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 371 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 372 return MAX2(1U, default_value); 373 } 374 375 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 376 assert(number_of_heap_regions > 0, "Heap must be initialized"); 377 378 switch (_sizer_kind) { 379 case SizerDefaults: 380 *min_young_length = calculate_default_min_length(number_of_heap_regions); 381 *max_young_length = calculate_default_max_length(number_of_heap_regions); 382 break; 383 case SizerNewSizeOnly: 384 *max_young_length = calculate_default_max_length(number_of_heap_regions); 385 *max_young_length = MAX2(*min_young_length, *max_young_length); 386 break; 387 case SizerMaxNewSizeOnly: 388 *min_young_length = calculate_default_min_length(number_of_heap_regions); 389 *min_young_length = MIN2(*min_young_length, *max_young_length); 390 break; 391 case SizerMaxAndNewSize: 392 // Do nothing. Values set on the command line, don't update them at runtime. 393 break; 394 case SizerNewRatio: 395 *min_young_length = number_of_heap_regions / (NewRatio + 1); 396 *max_young_length = *min_young_length; 397 break; 398 default: 399 ShouldNotReachHere(); 400 } 401 402 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 403 } 404 405 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 406 // We need to pass the desired values because recalculation may not update these 407 // values in some cases. 408 uint temp = _min_desired_young_length; 409 uint result = _max_desired_young_length; 410 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 411 return result; 412 } 413 414 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 415 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 416 &_max_desired_young_length); 417 } 418 419 void G1CollectorPolicy::init() { 420 // Set aside an initial future to_space. 421 _g1 = G1CollectedHeap::heap(); 422 423 assert(Heap_lock->owned_by_self(), "Locking discipline."); 424 425 initialize_gc_policy_counters(); 426 427 if (adaptive_young_list_length()) { 428 _young_list_fixed_length = 0; 429 } else { 430 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 431 } 432 _free_regions_at_end_of_collection = _g1->num_free_regions(); 433 434 update_young_list_max_and_target_length(); 435 // We may immediately start allocating regions and placing them on the 436 // collection set list. Initialize the per-collection set info 437 start_incremental_cset_building(); 438 } 439 440 // Create the jstat counters for the policy. 441 void G1CollectorPolicy::initialize_gc_policy_counters() { 442 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 443 } 444 445 bool G1CollectorPolicy::predict_will_fit(uint young_length, 446 double base_time_ms, 447 uint base_free_regions, 448 double target_pause_time_ms) const { 449 if (young_length >= base_free_regions) { 450 // end condition 1: not enough space for the young regions 451 return false; 452 } 453 454 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 455 size_t bytes_to_copy = 456 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 457 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 458 double young_other_time_ms = predict_young_other_time_ms(young_length); 459 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 460 if (pause_time_ms > target_pause_time_ms) { 461 // end condition 2: prediction is over the target pause time 462 return false; 463 } 464 465 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 466 if ((2.0 /* magic */ * _predictor.sigma()) * bytes_to_copy > free_bytes) { 467 // end condition 3: out-of-space (conservatively!) 468 return false; 469 } 470 471 // success! 472 return true; 473 } 474 475 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 476 // re-calculate the necessary reserve 477 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 478 // We use ceiling so that if reserve_regions_d is > 0.0 (but 479 // smaller than 1.0) we'll get 1. 480 _reserve_regions = (uint) ceil(reserve_regions_d); 481 482 _young_gen_sizer->heap_size_changed(new_number_of_regions); 483 } 484 485 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 486 uint base_min_length) const { 487 uint desired_min_length = 0; 488 if (adaptive_young_list_length()) { 489 if (_alloc_rate_ms_seq->num() > 3) { 490 double now_sec = os::elapsedTime(); 491 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 492 double alloc_rate_ms = predict_alloc_rate_ms(); 493 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 494 } else { 495 // otherwise we don't have enough info to make the prediction 496 } 497 } 498 desired_min_length += base_min_length; 499 // make sure we don't go below any user-defined minimum bound 500 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 501 } 502 503 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 504 // Here, we might want to also take into account any additional 505 // constraints (i.e., user-defined minimum bound). Currently, we 506 // effectively don't set this bound. 507 return _young_gen_sizer->max_desired_young_length(); 508 } 509 510 void G1CollectorPolicy::update_young_list_max_and_target_length() { 511 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 512 } 513 514 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 515 update_young_list_target_length(rs_lengths); 516 update_max_gc_locker_expansion(); 517 } 518 519 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 520 _young_list_target_length = bounded_young_list_target_length(rs_lengths); 521 } 522 523 void G1CollectorPolicy::update_young_list_target_length() { 524 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 525 } 526 527 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths) const { 528 // Calculate the absolute and desired min bounds. 529 530 // This is how many young regions we already have (currently: the survivors). 531 uint base_min_length = recorded_survivor_regions(); 532 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 533 // This is the absolute minimum young length. Ensure that we 534 // will at least have one eden region available for allocation. 535 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 536 // If we shrank the young list target it should not shrink below the current size. 537 desired_min_length = MAX2(desired_min_length, absolute_min_length); 538 // Calculate the absolute and desired max bounds. 539 540 // We will try our best not to "eat" into the reserve. 541 uint absolute_max_length = 0; 542 if (_free_regions_at_end_of_collection > _reserve_regions) { 543 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 544 } 545 uint desired_max_length = calculate_young_list_desired_max_length(); 546 if (desired_max_length > absolute_max_length) { 547 desired_max_length = absolute_max_length; 548 } 549 550 uint young_list_target_length = 0; 551 if (adaptive_young_list_length()) { 552 if (collector_state()->gcs_are_young()) { 553 young_list_target_length = 554 calculate_young_list_target_length(rs_lengths, 555 base_min_length, 556 desired_min_length, 557 desired_max_length); 558 } else { 559 // Don't calculate anything and let the code below bound it to 560 // the desired_min_length, i.e., do the next GC as soon as 561 // possible to maximize how many old regions we can add to it. 562 } 563 } else { 564 // The user asked for a fixed young gen so we'll fix the young gen 565 // whether the next GC is young or mixed. 566 young_list_target_length = _young_list_fixed_length; 567 } 568 569 // Make sure we don't go over the desired max length, nor under the 570 // desired min length. In case they clash, desired_min_length wins 571 // which is why that test is second. 572 if (young_list_target_length > desired_max_length) { 573 young_list_target_length = desired_max_length; 574 } 575 if (young_list_target_length < desired_min_length) { 576 young_list_target_length = desired_min_length; 577 } 578 579 assert(young_list_target_length > recorded_survivor_regions(), 580 "we should be able to allocate at least one eden region"); 581 assert(young_list_target_length >= absolute_min_length, "post-condition"); 582 583 return young_list_target_length; 584 } 585 586 uint 587 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 588 uint base_min_length, 589 uint desired_min_length, 590 uint desired_max_length) const { 591 assert(adaptive_young_list_length(), "pre-condition"); 592 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 593 594 // In case some edge-condition makes the desired max length too small... 595 if (desired_max_length <= desired_min_length) { 596 return desired_min_length; 597 } 598 599 // We'll adjust min_young_length and max_young_length not to include 600 // the already allocated young regions (i.e., so they reflect the 601 // min and max eden regions we'll allocate). The base_min_length 602 // will be reflected in the predictions by the 603 // survivor_regions_evac_time prediction. 604 assert(desired_min_length > base_min_length, "invariant"); 605 uint min_young_length = desired_min_length - base_min_length; 606 assert(desired_max_length > base_min_length, "invariant"); 607 uint max_young_length = desired_max_length - base_min_length; 608 609 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 610 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 611 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 612 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 613 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 614 double base_time_ms = 615 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 616 survivor_regions_evac_time; 617 uint available_free_regions = _free_regions_at_end_of_collection; 618 uint base_free_regions = 0; 619 if (available_free_regions > _reserve_regions) { 620 base_free_regions = available_free_regions - _reserve_regions; 621 } 622 623 // Here, we will make sure that the shortest young length that 624 // makes sense fits within the target pause time. 625 626 if (predict_will_fit(min_young_length, base_time_ms, 627 base_free_regions, target_pause_time_ms)) { 628 // The shortest young length will fit into the target pause time; 629 // we'll now check whether the absolute maximum number of young 630 // regions will fit in the target pause time. If not, we'll do 631 // a binary search between min_young_length and max_young_length. 632 if (predict_will_fit(max_young_length, base_time_ms, 633 base_free_regions, target_pause_time_ms)) { 634 // The maximum young length will fit into the target pause time. 635 // We are done so set min young length to the maximum length (as 636 // the result is assumed to be returned in min_young_length). 637 min_young_length = max_young_length; 638 } else { 639 // The maximum possible number of young regions will not fit within 640 // the target pause time so we'll search for the optimal 641 // length. The loop invariants are: 642 // 643 // min_young_length < max_young_length 644 // min_young_length is known to fit into the target pause time 645 // max_young_length is known not to fit into the target pause time 646 // 647 // Going into the loop we know the above hold as we've just 648 // checked them. Every time around the loop we check whether 649 // the middle value between min_young_length and 650 // max_young_length fits into the target pause time. If it 651 // does, it becomes the new min. If it doesn't, it becomes 652 // the new max. This way we maintain the loop invariants. 653 654 assert(min_young_length < max_young_length, "invariant"); 655 uint diff = (max_young_length - min_young_length) / 2; 656 while (diff > 0) { 657 uint young_length = min_young_length + diff; 658 if (predict_will_fit(young_length, base_time_ms, 659 base_free_regions, target_pause_time_ms)) { 660 min_young_length = young_length; 661 } else { 662 max_young_length = young_length; 663 } 664 assert(min_young_length < max_young_length, "invariant"); 665 diff = (max_young_length - min_young_length) / 2; 666 } 667 // The results is min_young_length which, according to the 668 // loop invariants, should fit within the target pause time. 669 670 // These are the post-conditions of the binary search above: 671 assert(min_young_length < max_young_length, 672 "otherwise we should have discovered that max_young_length " 673 "fits into the pause target and not done the binary search"); 674 assert(predict_will_fit(min_young_length, base_time_ms, 675 base_free_regions, target_pause_time_ms), 676 "min_young_length, the result of the binary search, should " 677 "fit into the pause target"); 678 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 679 base_free_regions, target_pause_time_ms), 680 "min_young_length, the result of the binary search, should be " 681 "optimal, so no larger length should fit into the pause target"); 682 } 683 } else { 684 // Even the minimum length doesn't fit into the pause time 685 // target, return it as the result nevertheless. 686 } 687 return base_min_length + min_young_length; 688 } 689 690 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 691 double survivor_regions_evac_time = 0.0; 692 for (HeapRegion * r = _recorded_survivor_head; 693 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 694 r = r->get_next_young_region()) { 695 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 696 } 697 return survivor_regions_evac_time; 698 } 699 700 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 701 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 702 703 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 704 if (rs_lengths > _rs_lengths_prediction) { 705 // add 10% to avoid having to recalculate often 706 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 707 update_rs_lengths_prediction(rs_lengths_prediction); 708 709 update_young_list_max_and_target_length(rs_lengths_prediction); 710 } 711 } 712 713 void G1CollectorPolicy::update_rs_lengths_prediction() { 714 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 715 } 716 717 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 718 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 719 _rs_lengths_prediction = prediction; 720 } 721 } 722 723 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 724 bool is_tlab, 725 bool* gc_overhead_limit_was_exceeded) { 726 guarantee(false, "Not using this policy feature yet."); 727 return NULL; 728 } 729 730 // This method controls how a collector handles one or more 731 // of its generations being fully allocated. 732 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 733 bool is_tlab) { 734 guarantee(false, "Not using this policy feature yet."); 735 return NULL; 736 } 737 738 739 #ifndef PRODUCT 740 bool G1CollectorPolicy::verify_young_ages() { 741 HeapRegion* head = _g1->young_list()->first_region(); 742 return 743 verify_young_ages(head, _short_lived_surv_rate_group); 744 // also call verify_young_ages on any additional surv rate groups 745 } 746 747 bool 748 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 749 SurvRateGroup *surv_rate_group) { 750 guarantee( surv_rate_group != NULL, "pre-condition" ); 751 752 const char* name = surv_rate_group->name(); 753 bool ret = true; 754 int prev_age = -1; 755 756 for (HeapRegion* curr = head; 757 curr != NULL; 758 curr = curr->get_next_young_region()) { 759 SurvRateGroup* group = curr->surv_rate_group(); 760 if (group == NULL && !curr->is_survivor()) { 761 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 762 ret = false; 763 } 764 765 if (surv_rate_group == group) { 766 int age = curr->age_in_surv_rate_group(); 767 768 if (age < 0) { 769 gclog_or_tty->print_cr("## %s: encountered negative age", name); 770 ret = false; 771 } 772 773 if (age <= prev_age) { 774 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 775 "(%d, %d)", name, age, prev_age); 776 ret = false; 777 } 778 prev_age = age; 779 } 780 } 781 782 return ret; 783 } 784 #endif // PRODUCT 785 786 void G1CollectorPolicy::record_full_collection_start() { 787 _full_collection_start_sec = os::elapsedTime(); 788 record_heap_size_info_at_start(true /* full */); 789 // Release the future to-space so that it is available for compaction into. 790 collector_state()->set_full_collection(true); 791 } 792 793 void G1CollectorPolicy::record_full_collection_end() { 794 // Consider this like a collection pause for the purposes of allocation 795 // since last pause. 796 double end_sec = os::elapsedTime(); 797 double full_gc_time_sec = end_sec - _full_collection_start_sec; 798 double full_gc_time_ms = full_gc_time_sec * 1000.0; 799 800 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 801 802 update_recent_gc_times(end_sec, full_gc_time_ms); 803 804 collector_state()->set_full_collection(false); 805 806 // "Nuke" the heuristics that control the young/mixed GC 807 // transitions and make sure we start with young GCs after the Full GC. 808 collector_state()->set_gcs_are_young(true); 809 collector_state()->set_last_young_gc(false); 810 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 811 collector_state()->set_during_initial_mark_pause(false); 812 collector_state()->set_in_marking_window(false); 813 collector_state()->set_in_marking_window_im(false); 814 815 _short_lived_surv_rate_group->start_adding_regions(); 816 // also call this on any additional surv rate groups 817 818 record_survivor_regions(0, NULL, NULL); 819 820 _free_regions_at_end_of_collection = _g1->num_free_regions(); 821 // Reset survivors SurvRateGroup. 822 _survivor_surv_rate_group->reset(); 823 update_young_list_max_and_target_length(); 824 update_rs_lengths_prediction(); 825 _collectionSetChooser->clear(); 826 } 827 828 void G1CollectorPolicy::record_stop_world_start() { 829 _stop_world_start = os::elapsedTime(); 830 } 831 832 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 833 // We only need to do this here as the policy will only be applied 834 // to the GC we're about to start. so, no point is calculating this 835 // every time we calculate / recalculate the target young length. 836 update_survivors_policy(); 837 838 assert(_g1->used() == _g1->recalculate_used(), 839 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 840 _g1->used(), _g1->recalculate_used()); 841 842 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 843 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 844 _stop_world_start = 0.0; 845 846 record_heap_size_info_at_start(false /* full */); 847 848 phase_times()->record_cur_collection_start_sec(start_time_sec); 849 _pending_cards = _g1->pending_card_num(); 850 851 _collection_set_bytes_used_before = 0; 852 _bytes_copied_during_gc = 0; 853 854 collector_state()->set_last_gc_was_young(false); 855 856 // do that for any other surv rate groups 857 _short_lived_surv_rate_group->stop_adding_regions(); 858 _survivors_age_table.clear(); 859 860 assert( verify_young_ages(), "region age verification" ); 861 } 862 863 void G1CollectorPolicy::record_concurrent_mark_init_end(double 864 mark_init_elapsed_time_ms) { 865 collector_state()->set_during_marking(true); 866 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 867 collector_state()->set_during_initial_mark_pause(false); 868 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 869 } 870 871 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 872 _mark_remark_start_sec = os::elapsedTime(); 873 collector_state()->set_during_marking(false); 874 } 875 876 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 877 double end_time_sec = os::elapsedTime(); 878 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 879 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 880 _cur_mark_stop_world_time_ms += elapsed_time_ms; 881 _prev_collection_pause_end_ms += elapsed_time_ms; 882 883 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec); 884 } 885 886 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 887 _mark_cleanup_start_sec = os::elapsedTime(); 888 } 889 890 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 891 collector_state()->set_last_young_gc(true); 892 collector_state()->set_in_marking_window(false); 893 } 894 895 void G1CollectorPolicy::record_concurrent_pause() { 896 if (_stop_world_start > 0.0) { 897 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 898 _trace_young_gen_time_data.record_yield_time(yield_ms); 899 } 900 } 901 902 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 903 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 904 return false; 905 } 906 907 size_t marking_initiating_used_threshold = 908 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 909 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 910 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 911 912 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 913 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 914 ergo_verbose5(ErgoConcCycles, 915 "request concurrent cycle initiation", 916 ergo_format_reason("occupancy higher than threshold") 917 ergo_format_byte("occupancy") 918 ergo_format_byte("allocation request") 919 ergo_format_byte_perc("threshold") 920 ergo_format_str("source"), 921 cur_used_bytes, 922 alloc_byte_size, 923 marking_initiating_used_threshold, 924 (double) InitiatingHeapOccupancyPercent, 925 source); 926 return true; 927 } else { 928 ergo_verbose5(ErgoConcCycles, 929 "do not request concurrent cycle initiation", 930 ergo_format_reason("still doing mixed collections") 931 ergo_format_byte("occupancy") 932 ergo_format_byte("allocation request") 933 ergo_format_byte_perc("threshold") 934 ergo_format_str("source"), 935 cur_used_bytes, 936 alloc_byte_size, 937 marking_initiating_used_threshold, 938 (double) InitiatingHeapOccupancyPercent, 939 source); 940 } 941 } 942 943 return false; 944 } 945 946 // Anything below that is considered to be zero 947 #define MIN_TIMER_GRANULARITY 0.0000001 948 949 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 950 double end_time_sec = os::elapsedTime(); 951 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 952 "otherwise, the subtraction below does not make sense"); 953 size_t rs_size = 954 _cur_collection_pause_used_regions_at_start - cset_region_length(); 955 size_t cur_used_bytes = _g1->used(); 956 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 957 bool last_pause_included_initial_mark = false; 958 bool update_stats = !_g1->evacuation_failed(); 959 960 #ifndef PRODUCT 961 if (G1YoungSurvRateVerbose) { 962 gclog_or_tty->cr(); 963 _short_lived_surv_rate_group->print(); 964 // do that for any other surv rate groups too 965 } 966 #endif // PRODUCT 967 968 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 969 if (last_pause_included_initial_mark) { 970 record_concurrent_mark_init_end(0.0); 971 } else { 972 maybe_start_marking(); 973 } 974 975 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, end_time_sec); 976 977 if (update_stats) { 978 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 979 // this is where we update the allocation rate of the application 980 double app_time_ms = 981 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 982 if (app_time_ms < MIN_TIMER_GRANULARITY) { 983 // This usually happens due to the timer not having the required 984 // granularity. Some Linuxes are the usual culprits. 985 // We'll just set it to something (arbitrarily) small. 986 app_time_ms = 1.0; 987 } 988 // We maintain the invariant that all objects allocated by mutator 989 // threads will be allocated out of eden regions. So, we can use 990 // the eden region number allocated since the previous GC to 991 // calculate the application's allocate rate. The only exception 992 // to that is humongous objects that are allocated separately. But 993 // given that humongous object allocations do not really affect 994 // either the pause's duration nor when the next pause will take 995 // place we can safely ignore them here. 996 uint regions_allocated = eden_cset_region_length(); 997 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 998 _alloc_rate_ms_seq->add(alloc_rate_ms); 999 1000 double interval_ms = 1001 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1002 update_recent_gc_times(end_time_sec, pause_time_ms); 1003 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1004 if (recent_avg_pause_time_ratio() < 0.0 || 1005 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1006 #ifndef PRODUCT 1007 // Dump info to allow post-facto debugging 1008 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1009 gclog_or_tty->print_cr("-------------------------------------------"); 1010 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1011 _recent_gc_times_ms->dump(); 1012 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1013 _recent_prev_end_times_for_all_gcs_sec->dump(); 1014 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1015 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1016 // In debug mode, terminate the JVM if the user wants to debug at this point. 1017 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1018 #endif // !PRODUCT 1019 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1020 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1021 if (_recent_avg_pause_time_ratio < 0.0) { 1022 _recent_avg_pause_time_ratio = 0.0; 1023 } else { 1024 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1025 _recent_avg_pause_time_ratio = 1.0; 1026 } 1027 } 1028 } 1029 1030 bool new_in_marking_window = collector_state()->in_marking_window(); 1031 bool new_in_marking_window_im = false; 1032 if (last_pause_included_initial_mark) { 1033 new_in_marking_window = true; 1034 new_in_marking_window_im = true; 1035 } 1036 1037 if (collector_state()->last_young_gc()) { 1038 // This is supposed to to be the "last young GC" before we start 1039 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1040 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1041 1042 if (next_gc_should_be_mixed("start mixed GCs", 1043 "do not start mixed GCs")) { 1044 collector_state()->set_gcs_are_young(false); 1045 } 1046 1047 collector_state()->set_last_young_gc(false); 1048 } 1049 1050 if (!collector_state()->last_gc_was_young()) { 1051 // This is a mixed GC. Here we decide whether to continue doing 1052 // mixed GCs or not. 1053 1054 if (!next_gc_should_be_mixed("continue mixed GCs", 1055 "do not continue mixed GCs")) { 1056 collector_state()->set_gcs_are_young(true); 1057 1058 maybe_start_marking(); 1059 } 1060 } 1061 1062 _short_lived_surv_rate_group->start_adding_regions(); 1063 // Do that for any other surv rate groups 1064 1065 if (update_stats) { 1066 double cost_per_card_ms = 0.0; 1067 double cost_scan_hcc = phase_times()->average_time_ms(G1GCPhaseTimes::ScanHCC); 1068 if (_pending_cards > 0) { 1069 cost_per_card_ms = (phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1070 _cost_per_card_ms_seq->add(cost_per_card_ms); 1071 } 1072 _cost_scan_hcc_seq->add(cost_scan_hcc); 1073 1074 double cost_per_entry_ms = 0.0; 1075 if (cards_scanned > 10) { 1076 cost_per_entry_ms = phase_times()->average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1077 if (collector_state()->last_gc_was_young()) { 1078 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1079 } else { 1080 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1081 } 1082 } 1083 1084 if (_max_rs_lengths > 0) { 1085 double cards_per_entry_ratio = 1086 (double) cards_scanned / (double) _max_rs_lengths; 1087 if (collector_state()->last_gc_was_young()) { 1088 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1089 } else { 1090 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1091 } 1092 } 1093 1094 // This is defensive. For a while _max_rs_lengths could get 1095 // smaller than _recorded_rs_lengths which was causing 1096 // rs_length_diff to get very large and mess up the RSet length 1097 // predictions. The reason was unsafe concurrent updates to the 1098 // _inc_cset_recorded_rs_lengths field which the code below guards 1099 // against (see CR 7118202). This bug has now been fixed (see CR 1100 // 7119027). However, I'm still worried that 1101 // _inc_cset_recorded_rs_lengths might still end up somewhat 1102 // inaccurate. The concurrent refinement thread calculates an 1103 // RSet's length concurrently with other CR threads updating it 1104 // which might cause it to calculate the length incorrectly (if, 1105 // say, it's in mid-coarsening). So I'll leave in the defensive 1106 // conditional below just in case. 1107 size_t rs_length_diff = 0; 1108 if (_max_rs_lengths > _recorded_rs_lengths) { 1109 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1110 } 1111 _rs_length_diff_seq->add((double) rs_length_diff); 1112 1113 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1114 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1115 double cost_per_byte_ms = 0.0; 1116 1117 if (copied_bytes > 0) { 1118 cost_per_byte_ms = phase_times()->average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1119 if (collector_state()->in_marking_window()) { 1120 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1121 } else { 1122 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1123 } 1124 } 1125 1126 double all_other_time_ms = pause_time_ms - 1127 (phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS) + phase_times()->average_time_ms(G1GCPhaseTimes::ScanRS) + 1128 phase_times()->average_time_ms(G1GCPhaseTimes::ObjCopy) + phase_times()->average_time_ms(G1GCPhaseTimes::Termination)); 1129 1130 double young_other_time_ms = 0.0; 1131 if (young_cset_region_length() > 0) { 1132 young_other_time_ms = 1133 phase_times()->young_cset_choice_time_ms() + 1134 phase_times()->young_free_cset_time_ms(); 1135 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1136 (double) young_cset_region_length()); 1137 } 1138 double non_young_other_time_ms = 0.0; 1139 if (old_cset_region_length() > 0) { 1140 non_young_other_time_ms = 1141 phase_times()->non_young_cset_choice_time_ms() + 1142 phase_times()->non_young_free_cset_time_ms(); 1143 1144 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1145 (double) old_cset_region_length()); 1146 } 1147 1148 double constant_other_time_ms = all_other_time_ms - 1149 (young_other_time_ms + non_young_other_time_ms); 1150 _constant_other_time_ms_seq->add(constant_other_time_ms); 1151 1152 double survival_ratio = 0.0; 1153 if (_collection_set_bytes_used_before > 0) { 1154 survival_ratio = (double) _bytes_copied_during_gc / 1155 (double) _collection_set_bytes_used_before; 1156 } 1157 1158 _pending_cards_seq->add((double) _pending_cards); 1159 _rs_lengths_seq->add((double) _max_rs_lengths); 1160 } 1161 1162 collector_state()->set_in_marking_window(new_in_marking_window); 1163 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1164 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1165 update_young_list_max_and_target_length(); 1166 update_rs_lengths_prediction(); 1167 1168 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1169 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1170 1171 double scan_hcc_time_ms = phase_times()->average_time_ms(G1GCPhaseTimes::ScanHCC); 1172 1173 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1174 ergo_verbose2(ErgoTiming, 1175 "adjust concurrent refinement thresholds", 1176 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1177 ergo_format_ms("Update RS time goal") 1178 ergo_format_ms("Scan HCC time"), 1179 update_rs_time_goal_ms, 1180 scan_hcc_time_ms); 1181 1182 update_rs_time_goal_ms = 0; 1183 } else { 1184 update_rs_time_goal_ms -= scan_hcc_time_ms; 1185 } 1186 adjust_concurrent_refinement(phase_times()->average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1187 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1188 update_rs_time_goal_ms); 1189 1190 _collectionSetChooser->verify(); 1191 } 1192 1193 #define EXT_SIZE_FORMAT "%.1f%s" 1194 #define EXT_SIZE_PARAMS(bytes) \ 1195 byte_size_in_proper_unit((double)(bytes)), \ 1196 proper_unit_for_byte_size((bytes)) 1197 1198 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1199 YoungList* young_list = _g1->young_list(); 1200 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1201 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1202 _heap_capacity_bytes_before_gc = _g1->capacity(); 1203 _heap_used_bytes_before_gc = _g1->used(); 1204 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1205 1206 _eden_capacity_bytes_before_gc = 1207 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1208 1209 if (full) { 1210 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1211 } 1212 } 1213 1214 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1215 size_t bytes_after = _g1->used(); 1216 size_t capacity = _g1->capacity(); 1217 1218 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1219 byte_size_in_proper_unit(bytes_before), 1220 proper_unit_for_byte_size(bytes_before), 1221 byte_size_in_proper_unit(bytes_after), 1222 proper_unit_for_byte_size(bytes_after), 1223 byte_size_in_proper_unit(capacity), 1224 proper_unit_for_byte_size(capacity)); 1225 } 1226 1227 void G1CollectorPolicy::print_heap_transition() const { 1228 print_heap_transition(_heap_used_bytes_before_gc); 1229 } 1230 1231 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1232 YoungList* young_list = _g1->young_list(); 1233 1234 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1235 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1236 size_t heap_used_bytes_after_gc = _g1->used(); 1237 1238 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1239 size_t eden_capacity_bytes_after_gc = 1240 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1241 1242 gclog_or_tty->print( 1243 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1244 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1245 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1246 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1247 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1248 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1249 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1250 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1251 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1252 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1253 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1254 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1255 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1256 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1257 1258 if (full) { 1259 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1260 } 1261 1262 gclog_or_tty->cr(); 1263 } 1264 1265 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1266 double update_rs_processed_buffers, 1267 double goal_ms) { 1268 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1269 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1270 1271 if (G1UseAdaptiveConcRefinement) { 1272 const int k_gy = 3, k_gr = 6; 1273 const double inc_k = 1.1, dec_k = 0.9; 1274 1275 int g = cg1r->green_zone(); 1276 if (update_rs_time > goal_ms) { 1277 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1278 } else { 1279 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1280 g = (int)MAX2(g * inc_k, g + 1.0); 1281 } 1282 } 1283 // Change the refinement threads params 1284 cg1r->set_green_zone(g); 1285 cg1r->set_yellow_zone(g * k_gy); 1286 cg1r->set_red_zone(g * k_gr); 1287 cg1r->reinitialize_threads(); 1288 1289 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1290 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1291 cg1r->yellow_zone()); 1292 // Change the barrier params 1293 dcqs.set_process_completed_threshold(processing_threshold); 1294 dcqs.set_max_completed_queue(cg1r->red_zone()); 1295 } 1296 1297 int curr_queue_size = dcqs.completed_buffers_num(); 1298 if (curr_queue_size >= cg1r->yellow_zone()) { 1299 dcqs.set_completed_queue_padding(curr_queue_size); 1300 } else { 1301 dcqs.set_completed_queue_padding(0); 1302 } 1303 dcqs.notify_if_necessary(); 1304 } 1305 1306 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1307 return (size_t) get_new_prediction(_rs_length_diff_seq); 1308 } 1309 1310 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1311 return get_new_prediction(_alloc_rate_ms_seq); 1312 } 1313 1314 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1315 return get_new_prediction(_cost_per_card_ms_seq); 1316 } 1317 1318 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1319 return get_new_prediction(_cost_scan_hcc_seq); 1320 } 1321 1322 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1323 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1324 } 1325 1326 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1327 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1328 } 1329 1330 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1331 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1332 return predict_young_cards_per_entry_ratio(); 1333 } else { 1334 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1335 } 1336 } 1337 1338 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1339 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1340 } 1341 1342 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1343 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1344 } 1345 1346 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1347 if (collector_state()->gcs_are_young()) { 1348 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1349 } else { 1350 return predict_mixed_rs_scan_time_ms(card_num); 1351 } 1352 } 1353 1354 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1355 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1356 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1357 } else { 1358 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1359 } 1360 } 1361 1362 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1363 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1364 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1365 } else { 1366 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1367 } 1368 } 1369 1370 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1371 if (collector_state()->during_concurrent_mark()) { 1372 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1373 } else { 1374 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1375 } 1376 } 1377 1378 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1379 return get_new_prediction(_constant_other_time_ms_seq); 1380 } 1381 1382 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1383 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1384 } 1385 1386 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1387 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1388 } 1389 1390 double G1CollectorPolicy::predict_remark_time_ms() const { 1391 return get_new_prediction(_concurrent_mark_remark_times_ms); 1392 } 1393 1394 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1395 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1396 } 1397 1398 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1399 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1400 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1401 double pred = get_new_prediction(seq); 1402 if (pred > 1.0) { 1403 pred = 1.0; 1404 } 1405 return pred; 1406 } 1407 1408 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1409 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1410 } 1411 1412 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1413 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1414 } 1415 1416 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1417 size_t scanned_cards) const { 1418 return 1419 predict_rs_update_time_ms(pending_cards) + 1420 predict_rs_scan_time_ms(scanned_cards) + 1421 predict_constant_other_time_ms(); 1422 } 1423 1424 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1425 size_t rs_length = predict_rs_length_diff(); 1426 size_t card_num; 1427 if (collector_state()->gcs_are_young()) { 1428 card_num = predict_young_card_num(rs_length); 1429 } else { 1430 card_num = predict_non_young_card_num(rs_length); 1431 } 1432 return predict_base_elapsed_time_ms(pending_cards, card_num); 1433 } 1434 1435 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1436 size_t bytes_to_copy; 1437 if (hr->is_marked()) 1438 bytes_to_copy = hr->max_live_bytes(); 1439 else { 1440 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1441 int age = hr->age_in_surv_rate_group(); 1442 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1443 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1444 } 1445 return bytes_to_copy; 1446 } 1447 1448 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1449 bool for_young_gc) const { 1450 size_t rs_length = hr->rem_set()->occupied(); 1451 size_t card_num; 1452 1453 // Predicting the number of cards is based on which type of GC 1454 // we're predicting for. 1455 if (for_young_gc) { 1456 card_num = predict_young_card_num(rs_length); 1457 } else { 1458 card_num = predict_non_young_card_num(rs_length); 1459 } 1460 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1461 1462 double region_elapsed_time_ms = 1463 predict_rs_scan_time_ms(card_num) + 1464 predict_object_copy_time_ms(bytes_to_copy); 1465 1466 // The prediction of the "other" time for this region is based 1467 // upon the region type and NOT the GC type. 1468 if (hr->is_young()) { 1469 region_elapsed_time_ms += predict_young_other_time_ms(1); 1470 } else { 1471 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1472 } 1473 return region_elapsed_time_ms; 1474 } 1475 1476 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1477 uint survivor_cset_region_length) { 1478 _eden_cset_region_length = eden_cset_region_length; 1479 _survivor_cset_region_length = survivor_cset_region_length; 1480 _old_cset_region_length = 0; 1481 } 1482 1483 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1484 _recorded_rs_lengths = rs_lengths; 1485 } 1486 1487 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1488 double elapsed_ms) { 1489 _recent_gc_times_ms->add(elapsed_ms); 1490 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1491 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1492 } 1493 1494 size_t G1CollectorPolicy::expansion_amount() const { 1495 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1496 double threshold = _gc_overhead_perc; 1497 if (recent_gc_overhead > threshold) { 1498 // We will double the existing space, or take 1499 // G1ExpandByPercentOfAvailable % of the available expansion 1500 // space, whichever is smaller, bounded below by a minimum 1501 // expansion (unless that's all that's left.) 1502 const size_t min_expand_bytes = 1*M; 1503 size_t reserved_bytes = _g1->max_capacity(); 1504 size_t committed_bytes = _g1->capacity(); 1505 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1506 size_t expand_bytes; 1507 size_t expand_bytes_via_pct = 1508 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1509 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1510 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1511 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1512 1513 ergo_verbose5(ErgoHeapSizing, 1514 "attempt heap expansion", 1515 ergo_format_reason("recent GC overhead higher than " 1516 "threshold after GC") 1517 ergo_format_perc("recent GC overhead") 1518 ergo_format_perc("threshold") 1519 ergo_format_byte("uncommitted") 1520 ergo_format_byte_perc("calculated expansion amount"), 1521 recent_gc_overhead, threshold, 1522 uncommitted_bytes, 1523 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1524 1525 return expand_bytes; 1526 } else { 1527 return 0; 1528 } 1529 } 1530 1531 void G1CollectorPolicy::print_tracing_info() const { 1532 _trace_young_gen_time_data.print(); 1533 _trace_old_gen_time_data.print(); 1534 } 1535 1536 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1537 #ifndef PRODUCT 1538 _short_lived_surv_rate_group->print_surv_rate_summary(); 1539 // add this call for any other surv rate groups 1540 #endif // PRODUCT 1541 } 1542 1543 bool G1CollectorPolicy::is_young_list_full() const { 1544 uint young_list_length = _g1->young_list()->length(); 1545 uint young_list_target_length = _young_list_target_length; 1546 return young_list_length >= young_list_target_length; 1547 } 1548 1549 bool G1CollectorPolicy::can_expand_young_list() const { 1550 uint young_list_length = _g1->young_list()->length(); 1551 uint young_list_max_length = _young_list_max_length; 1552 return young_list_length < young_list_max_length; 1553 } 1554 1555 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1556 uint expansion_region_num = 0; 1557 if (GCLockerEdenExpansionPercent > 0) { 1558 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1559 double expansion_region_num_d = perc * (double) _young_list_target_length; 1560 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1561 // less than 1.0) we'll get 1. 1562 expansion_region_num = (uint) ceil(expansion_region_num_d); 1563 } else { 1564 assert(expansion_region_num == 0, "sanity"); 1565 } 1566 _young_list_max_length = _young_list_target_length + expansion_region_num; 1567 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1568 } 1569 1570 // Calculates survivor space parameters. 1571 void G1CollectorPolicy::update_survivors_policy() { 1572 double max_survivor_regions_d = 1573 (double) _young_list_target_length / (double) SurvivorRatio; 1574 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1575 // smaller than 1.0) we'll get 1. 1576 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1577 1578 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1579 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1580 } 1581 1582 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 1583 GCCause::Cause gc_cause) { 1584 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1585 if (!during_cycle) { 1586 ergo_verbose1(ErgoConcCycles, 1587 "request concurrent cycle initiation", 1588 ergo_format_reason("requested by GC cause") 1589 ergo_format_str("GC cause"), 1590 GCCause::to_string(gc_cause)); 1591 collector_state()->set_initiate_conc_mark_if_possible(true); 1592 return true; 1593 } else { 1594 ergo_verbose1(ErgoConcCycles, 1595 "do not request concurrent cycle initiation", 1596 ergo_format_reason("concurrent cycle already in progress") 1597 ergo_format_str("GC cause"), 1598 GCCause::to_string(gc_cause)); 1599 return false; 1600 } 1601 } 1602 1603 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1604 // We are about to decide on whether this pause will be an 1605 // initial-mark pause. 1606 1607 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1608 // will set it here if we have to. However, it should be cleared by 1609 // the end of the pause (it's only set for the duration of an 1610 // initial-mark pause). 1611 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1612 1613 if (collector_state()->initiate_conc_mark_if_possible()) { 1614 // We had noticed on a previous pause that the heap occupancy has 1615 // gone over the initiating threshold and we should start a 1616 // concurrent marking cycle. So we might initiate one. 1617 1618 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 1619 if (!during_cycle) { 1620 // The concurrent marking thread is not "during a cycle", i.e., 1621 // it has completed the last one. So we can go ahead and 1622 // initiate a new cycle. 1623 1624 collector_state()->set_during_initial_mark_pause(true); 1625 // We do not allow mixed GCs during marking. 1626 if (!collector_state()->gcs_are_young()) { 1627 collector_state()->set_gcs_are_young(true); 1628 ergo_verbose0(ErgoMixedGCs, 1629 "end mixed GCs", 1630 ergo_format_reason("concurrent cycle is about to start")); 1631 } 1632 1633 // And we can now clear initiate_conc_mark_if_possible() as 1634 // we've already acted on it. 1635 collector_state()->set_initiate_conc_mark_if_possible(false); 1636 1637 ergo_verbose0(ErgoConcCycles, 1638 "initiate concurrent cycle", 1639 ergo_format_reason("concurrent cycle initiation requested")); 1640 } else { 1641 // The concurrent marking thread is still finishing up the 1642 // previous cycle. If we start one right now the two cycles 1643 // overlap. In particular, the concurrent marking thread might 1644 // be in the process of clearing the next marking bitmap (which 1645 // we will use for the next cycle if we start one). Starting a 1646 // cycle now will be bad given that parts of the marking 1647 // information might get cleared by the marking thread. And we 1648 // cannot wait for the marking thread to finish the cycle as it 1649 // periodically yields while clearing the next marking bitmap 1650 // and, if it's in a yield point, it's waiting for us to 1651 // finish. So, at this point we will not start a cycle and we'll 1652 // let the concurrent marking thread complete the last one. 1653 ergo_verbose0(ErgoConcCycles, 1654 "do not initiate concurrent cycle", 1655 ergo_format_reason("concurrent cycle already in progress")); 1656 } 1657 } 1658 } 1659 1660 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1661 G1CollectedHeap* _g1h; 1662 CSetChooserParUpdater _cset_updater; 1663 1664 public: 1665 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1666 uint chunk_size) : 1667 _g1h(G1CollectedHeap::heap()), 1668 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1669 1670 bool doHeapRegion(HeapRegion* r) { 1671 // Do we have any marking information for this region? 1672 if (r->is_marked()) { 1673 // We will skip any region that's currently used as an old GC 1674 // alloc region (we should not consider those for collection 1675 // before we fill them up). 1676 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1677 _cset_updater.add_region(r); 1678 } 1679 } 1680 return false; 1681 } 1682 }; 1683 1684 class ParKnownGarbageTask: public AbstractGangTask { 1685 CollectionSetChooser* _hrSorted; 1686 uint _chunk_size; 1687 G1CollectedHeap* _g1; 1688 HeapRegionClaimer _hrclaimer; 1689 1690 public: 1691 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1692 AbstractGangTask("ParKnownGarbageTask"), 1693 _hrSorted(hrSorted), _chunk_size(chunk_size), 1694 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1695 1696 void work(uint worker_id) { 1697 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1698 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1699 } 1700 }; 1701 1702 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1703 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1704 const uint overpartition_factor = 4; 1705 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1706 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1707 } 1708 1709 void 1710 G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1711 _collectionSetChooser->clear(); 1712 1713 WorkGang* workers = _g1->workers(); 1714 uint n_workers = workers->active_workers(); 1715 1716 uint n_regions = _g1->num_regions(); 1717 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1718 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1719 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1720 workers->run_task(&par_known_garbage_task); 1721 1722 _collectionSetChooser->sort_regions(); 1723 1724 double end_sec = os::elapsedTime(); 1725 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1726 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1727 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1728 _prev_collection_pause_end_ms += elapsed_time_ms; 1729 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec); 1730 } 1731 1732 // Add the heap region at the head of the non-incremental collection set 1733 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1734 assert(_inc_cset_build_state == Active, "Precondition"); 1735 assert(hr->is_old(), "the region should be old"); 1736 1737 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1738 _g1->register_old_region_with_cset(hr); 1739 hr->set_next_in_collection_set(_collection_set); 1740 _collection_set = hr; 1741 _collection_set_bytes_used_before += hr->used(); 1742 size_t rs_length = hr->rem_set()->occupied(); 1743 _recorded_rs_lengths += rs_length; 1744 _old_cset_region_length += 1; 1745 } 1746 1747 // Initialize the per-collection-set information 1748 void G1CollectorPolicy::start_incremental_cset_building() { 1749 assert(_inc_cset_build_state == Inactive, "Precondition"); 1750 1751 _inc_cset_head = NULL; 1752 _inc_cset_tail = NULL; 1753 _inc_cset_bytes_used_before = 0; 1754 1755 _inc_cset_max_finger = 0; 1756 _inc_cset_recorded_rs_lengths = 0; 1757 _inc_cset_recorded_rs_lengths_diffs = 0; 1758 _inc_cset_predicted_elapsed_time_ms = 0.0; 1759 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1760 _inc_cset_build_state = Active; 1761 } 1762 1763 void G1CollectorPolicy::finalize_incremental_cset_building() { 1764 assert(_inc_cset_build_state == Active, "Precondition"); 1765 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1766 1767 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1768 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1769 // that adds a new region to the CSet. Further updates by the 1770 // concurrent refinement thread that samples the young RSet lengths 1771 // are accumulated in the *_diffs fields. Here we add the diffs to 1772 // the "main" fields. 1773 1774 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1775 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1776 } else { 1777 // This is defensive. The diff should in theory be always positive 1778 // as RSets can only grow between GCs. However, given that we 1779 // sample their size concurrently with other threads updating them 1780 // it's possible that we might get the wrong size back, which 1781 // could make the calculations somewhat inaccurate. 1782 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1783 if (_inc_cset_recorded_rs_lengths >= diffs) { 1784 _inc_cset_recorded_rs_lengths -= diffs; 1785 } else { 1786 _inc_cset_recorded_rs_lengths = 0; 1787 } 1788 } 1789 _inc_cset_predicted_elapsed_time_ms += 1790 _inc_cset_predicted_elapsed_time_ms_diffs; 1791 1792 _inc_cset_recorded_rs_lengths_diffs = 0; 1793 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1794 } 1795 1796 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1797 // This routine is used when: 1798 // * adding survivor regions to the incremental cset at the end of an 1799 // evacuation pause, 1800 // * adding the current allocation region to the incremental cset 1801 // when it is retired, and 1802 // * updating existing policy information for a region in the 1803 // incremental cset via young list RSet sampling. 1804 // Therefore this routine may be called at a safepoint by the 1805 // VM thread, or in-between safepoints by mutator threads (when 1806 // retiring the current allocation region) or a concurrent 1807 // refine thread (RSet sampling). 1808 1809 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1810 size_t used_bytes = hr->used(); 1811 _inc_cset_recorded_rs_lengths += rs_length; 1812 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1813 _inc_cset_bytes_used_before += used_bytes; 1814 1815 // Cache the values we have added to the aggregated information 1816 // in the heap region in case we have to remove this region from 1817 // the incremental collection set, or it is updated by the 1818 // rset sampling code 1819 hr->set_recorded_rs_length(rs_length); 1820 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1821 } 1822 1823 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1824 size_t new_rs_length) { 1825 // Update the CSet information that is dependent on the new RS length 1826 assert(hr->is_young(), "Precondition"); 1827 assert(!SafepointSynchronize::is_at_safepoint(), 1828 "should not be at a safepoint"); 1829 1830 // We could have updated _inc_cset_recorded_rs_lengths and 1831 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1832 // that atomically, as this code is executed by a concurrent 1833 // refinement thread, potentially concurrently with a mutator thread 1834 // allocating a new region and also updating the same fields. To 1835 // avoid the atomic operations we accumulate these updates on two 1836 // separate fields (*_diffs) and we'll just add them to the "main" 1837 // fields at the start of a GC. 1838 1839 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1840 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1841 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1842 1843 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1844 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1845 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1846 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1847 1848 hr->set_recorded_rs_length(new_rs_length); 1849 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1850 } 1851 1852 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1853 assert(hr->is_young(), "invariant"); 1854 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1855 assert(_inc_cset_build_state == Active, "Precondition"); 1856 1857 // We need to clear and set the cached recorded/cached collection set 1858 // information in the heap region here (before the region gets added 1859 // to the collection set). An individual heap region's cached values 1860 // are calculated, aggregated with the policy collection set info, 1861 // and cached in the heap region here (initially) and (subsequently) 1862 // by the Young List sampling code. 1863 1864 size_t rs_length = hr->rem_set()->occupied(); 1865 add_to_incremental_cset_info(hr, rs_length); 1866 1867 HeapWord* hr_end = hr->end(); 1868 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1869 1870 assert(!hr->in_collection_set(), "invariant"); 1871 _g1->register_young_region_with_cset(hr); 1872 assert(hr->next_in_collection_set() == NULL, "invariant"); 1873 } 1874 1875 // Add the region at the RHS of the incremental cset 1876 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1877 // We should only ever be appending survivors at the end of a pause 1878 assert(hr->is_survivor(), "Logic"); 1879 1880 // Do the 'common' stuff 1881 add_region_to_incremental_cset_common(hr); 1882 1883 // Now add the region at the right hand side 1884 if (_inc_cset_tail == NULL) { 1885 assert(_inc_cset_head == NULL, "invariant"); 1886 _inc_cset_head = hr; 1887 } else { 1888 _inc_cset_tail->set_next_in_collection_set(hr); 1889 } 1890 _inc_cset_tail = hr; 1891 } 1892 1893 // Add the region to the LHS of the incremental cset 1894 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 1895 // Survivors should be added to the RHS at the end of a pause 1896 assert(hr->is_eden(), "Logic"); 1897 1898 // Do the 'common' stuff 1899 add_region_to_incremental_cset_common(hr); 1900 1901 // Add the region at the left hand side 1902 hr->set_next_in_collection_set(_inc_cset_head); 1903 if (_inc_cset_head == NULL) { 1904 assert(_inc_cset_tail == NULL, "Invariant"); 1905 _inc_cset_tail = hr; 1906 } 1907 _inc_cset_head = hr; 1908 } 1909 1910 #ifndef PRODUCT 1911 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 1912 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 1913 1914 st->print_cr("\nCollection_set:"); 1915 HeapRegion* csr = list_head; 1916 while (csr != NULL) { 1917 HeapRegion* next = csr->next_in_collection_set(); 1918 assert(csr->in_collection_set(), "bad CS"); 1919 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 1920 HR_FORMAT_PARAMS(csr), 1921 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 1922 csr->age_in_surv_rate_group_cond()); 1923 csr = next; 1924 } 1925 } 1926 #endif // !PRODUCT 1927 1928 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 1929 // Returns the given amount of reclaimable bytes (that represents 1930 // the amount of reclaimable space still to be collected) as a 1931 // percentage of the current heap capacity. 1932 size_t capacity_bytes = _g1->capacity(); 1933 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 1934 } 1935 1936 void G1CollectorPolicy::maybe_start_marking() { 1937 if (need_to_start_conc_mark("end of GC")) { 1938 // Note: this might have already been set, if during the last 1939 // pause we decided to start a cycle but at the beginning of 1940 // this pause we decided to postpone it. That's OK. 1941 collector_state()->set_initiate_conc_mark_if_possible(true); 1942 } 1943 } 1944 1945 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 1946 const char* false_action_str) const { 1947 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1948 if (cset_chooser->is_empty()) { 1949 ergo_verbose0(ErgoMixedGCs, 1950 false_action_str, 1951 ergo_format_reason("candidate old regions not available")); 1952 return false; 1953 } 1954 1955 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1956 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 1957 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 1958 double threshold = (double) G1HeapWastePercent; 1959 if (reclaimable_perc <= threshold) { 1960 ergo_verbose4(ErgoMixedGCs, 1961 false_action_str, 1962 ergo_format_reason("reclaimable percentage not over threshold") 1963 ergo_format_region("candidate old regions") 1964 ergo_format_byte_perc("reclaimable") 1965 ergo_format_perc("threshold"), 1966 cset_chooser->remaining_regions(), 1967 reclaimable_bytes, 1968 reclaimable_perc, threshold); 1969 return false; 1970 } 1971 1972 ergo_verbose4(ErgoMixedGCs, 1973 true_action_str, 1974 ergo_format_reason("candidate old regions available") 1975 ergo_format_region("candidate old regions") 1976 ergo_format_byte_perc("reclaimable") 1977 ergo_format_perc("threshold"), 1978 cset_chooser->remaining_regions(), 1979 reclaimable_bytes, 1980 reclaimable_perc, threshold); 1981 return true; 1982 } 1983 1984 uint G1CollectorPolicy::calc_min_old_cset_length() const { 1985 // The min old CSet region bound is based on the maximum desired 1986 // number of mixed GCs after a cycle. I.e., even if some old regions 1987 // look expensive, we should add them to the CSet anyway to make 1988 // sure we go through the available old regions in no more than the 1989 // maximum desired number of mixed GCs. 1990 // 1991 // The calculation is based on the number of marked regions we added 1992 // to the CSet chooser in the first place, not how many remain, so 1993 // that the result is the same during all mixed GCs that follow a cycle. 1994 1995 const size_t region_num = (size_t) _collectionSetChooser->length(); 1996 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 1997 size_t result = region_num / gc_num; 1998 // emulate ceiling 1999 if (result * gc_num < region_num) { 2000 result += 1; 2001 } 2002 return (uint) result; 2003 } 2004 2005 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2006 // The max old CSet region bound is based on the threshold expressed 2007 // as a percentage of the heap size. I.e., it should bound the 2008 // number of old regions added to the CSet irrespective of how many 2009 // of them are available. 2010 2011 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2012 const size_t region_num = g1h->num_regions(); 2013 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2014 size_t result = region_num * perc / 100; 2015 // emulate ceiling 2016 if (100 * result < region_num * perc) { 2017 result += 1; 2018 } 2019 return (uint) result; 2020 } 2021 2022 2023 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2024 double young_start_time_sec = os::elapsedTime(); 2025 2026 YoungList* young_list = _g1->young_list(); 2027 finalize_incremental_cset_building(); 2028 2029 guarantee(target_pause_time_ms > 0.0, 2030 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2031 guarantee(_collection_set == NULL, "Precondition"); 2032 2033 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2034 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2035 2036 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2037 "start choosing CSet", 2038 ergo_format_size("_pending_cards") 2039 ergo_format_ms("predicted base time") 2040 ergo_format_ms("remaining time") 2041 ergo_format_ms("target pause time"), 2042 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2043 2044 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2045 2046 if (collector_state()->last_gc_was_young()) { 2047 _trace_young_gen_time_data.increment_young_collection_count(); 2048 } else { 2049 _trace_young_gen_time_data.increment_mixed_collection_count(); 2050 } 2051 2052 // The young list is laid with the survivor regions from the previous 2053 // pause are appended to the RHS of the young list, i.e. 2054 // [Newly Young Regions ++ Survivors from last pause]. 2055 2056 uint survivor_region_length = young_list->survivor_length(); 2057 uint eden_region_length = young_list->eden_length(); 2058 init_cset_region_lengths(eden_region_length, survivor_region_length); 2059 2060 HeapRegion* hr = young_list->first_survivor_region(); 2061 while (hr != NULL) { 2062 assert(hr->is_survivor(), "badly formed young list"); 2063 // There is a convention that all the young regions in the CSet 2064 // are tagged as "eden", so we do this for the survivors here. We 2065 // use the special set_eden_pre_gc() as it doesn't check that the 2066 // region is free (which is not the case here). 2067 hr->set_eden_pre_gc(); 2068 hr = hr->get_next_young_region(); 2069 } 2070 2071 // Clear the fields that point to the survivor list - they are all young now. 2072 young_list->clear_survivors(); 2073 2074 _collection_set = _inc_cset_head; 2075 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2076 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2077 2078 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2079 "add young regions to CSet", 2080 ergo_format_region("eden") 2081 ergo_format_region("survivors") 2082 ergo_format_ms("predicted young region time") 2083 ergo_format_ms("target pause time"), 2084 eden_region_length, survivor_region_length, 2085 _inc_cset_predicted_elapsed_time_ms, 2086 target_pause_time_ms); 2087 2088 // The number of recorded young regions is the incremental 2089 // collection set's current size 2090 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2091 2092 double young_end_time_sec = os::elapsedTime(); 2093 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2094 2095 return time_remaining_ms; 2096 } 2097 2098 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2099 double non_young_start_time_sec = os::elapsedTime(); 2100 double predicted_old_time_ms = 0.0; 2101 2102 2103 if (!collector_state()->gcs_are_young()) { 2104 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2105 cset_chooser->verify(); 2106 const uint min_old_cset_length = calc_min_old_cset_length(); 2107 const uint max_old_cset_length = calc_max_old_cset_length(); 2108 2109 uint expensive_region_num = 0; 2110 bool check_time_remaining = adaptive_young_list_length(); 2111 2112 HeapRegion* hr = cset_chooser->peek(); 2113 while (hr != NULL) { 2114 if (old_cset_region_length() >= max_old_cset_length) { 2115 // Added maximum number of old regions to the CSet. 2116 ergo_verbose2(ErgoCSetConstruction, 2117 "finish adding old regions to CSet", 2118 ergo_format_reason("old CSet region num reached max") 2119 ergo_format_region("old") 2120 ergo_format_region("max"), 2121 old_cset_region_length(), max_old_cset_length); 2122 break; 2123 } 2124 2125 2126 // Stop adding regions if the remaining reclaimable space is 2127 // not above G1HeapWastePercent. 2128 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2129 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2130 double threshold = (double) G1HeapWastePercent; 2131 if (reclaimable_perc <= threshold) { 2132 // We've added enough old regions that the amount of uncollected 2133 // reclaimable space is at or below the waste threshold. Stop 2134 // adding old regions to the CSet. 2135 ergo_verbose5(ErgoCSetConstruction, 2136 "finish adding old regions to CSet", 2137 ergo_format_reason("reclaimable percentage not over threshold") 2138 ergo_format_region("old") 2139 ergo_format_region("max") 2140 ergo_format_byte_perc("reclaimable") 2141 ergo_format_perc("threshold"), 2142 old_cset_region_length(), 2143 max_old_cset_length, 2144 reclaimable_bytes, 2145 reclaimable_perc, threshold); 2146 break; 2147 } 2148 2149 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2150 if (check_time_remaining) { 2151 if (predicted_time_ms > time_remaining_ms) { 2152 // Too expensive for the current CSet. 2153 2154 if (old_cset_region_length() >= min_old_cset_length) { 2155 // We have added the minimum number of old regions to the CSet, 2156 // we are done with this CSet. 2157 ergo_verbose4(ErgoCSetConstruction, 2158 "finish adding old regions to CSet", 2159 ergo_format_reason("predicted time is too high") 2160 ergo_format_ms("predicted time") 2161 ergo_format_ms("remaining time") 2162 ergo_format_region("old") 2163 ergo_format_region("min"), 2164 predicted_time_ms, time_remaining_ms, 2165 old_cset_region_length(), min_old_cset_length); 2166 break; 2167 } 2168 2169 // We'll add it anyway given that we haven't reached the 2170 // minimum number of old regions. 2171 expensive_region_num += 1; 2172 } 2173 } else { 2174 if (old_cset_region_length() >= min_old_cset_length) { 2175 // In the non-auto-tuning case, we'll finish adding regions 2176 // to the CSet if we reach the minimum. 2177 ergo_verbose2(ErgoCSetConstruction, 2178 "finish adding old regions to CSet", 2179 ergo_format_reason("old CSet region num reached min") 2180 ergo_format_region("old") 2181 ergo_format_region("min"), 2182 old_cset_region_length(), min_old_cset_length); 2183 break; 2184 } 2185 } 2186 2187 // We will add this region to the CSet. 2188 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2189 predicted_old_time_ms += predicted_time_ms; 2190 cset_chooser->pop(); // already have region via peek() 2191 _g1->old_set_remove(hr); 2192 add_old_region_to_cset(hr); 2193 2194 hr = cset_chooser->peek(); 2195 } 2196 if (hr == NULL) { 2197 ergo_verbose0(ErgoCSetConstruction, 2198 "finish adding old regions to CSet", 2199 ergo_format_reason("candidate old regions not available")); 2200 } 2201 2202 if (expensive_region_num > 0) { 2203 // We print the information once here at the end, predicated on 2204 // whether we added any apparently expensive regions or not, to 2205 // avoid generating output per region. 2206 ergo_verbose4(ErgoCSetConstruction, 2207 "added expensive regions to CSet", 2208 ergo_format_reason("old CSet region num not reached min") 2209 ergo_format_region("old") 2210 ergo_format_region("expensive") 2211 ergo_format_region("min") 2212 ergo_format_ms("remaining time"), 2213 old_cset_region_length(), 2214 expensive_region_num, 2215 min_old_cset_length, 2216 time_remaining_ms); 2217 } 2218 2219 cset_chooser->verify(); 2220 } 2221 2222 stop_incremental_cset_building(); 2223 2224 ergo_verbose3(ErgoCSetConstruction, 2225 "finish choosing CSet", 2226 ergo_format_region("old") 2227 ergo_format_ms("predicted old region time") 2228 ergo_format_ms("time remaining"), 2229 old_cset_region_length(), 2230 predicted_old_time_ms, time_remaining_ms); 2231 2232 double non_young_end_time_sec = os::elapsedTime(); 2233 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2234 } 2235 2236 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2237 if(TraceYoungGenTime) { 2238 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2239 } 2240 } 2241 2242 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2243 if(TraceYoungGenTime) { 2244 _all_yield_times_ms.add(yield_time_ms); 2245 } 2246 } 2247 2248 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2249 if(TraceYoungGenTime) { 2250 _total.add(pause_time_ms); 2251 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2252 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2253 _parallel.add(phase_times->cur_collection_par_time_ms()); 2254 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2255 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2256 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2257 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2258 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2259 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2260 2261 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2262 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2263 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2264 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2265 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2266 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2267 2268 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2269 _parallel_other.add(parallel_other_time); 2270 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2271 } 2272 } 2273 2274 void TraceYoungGenTimeData::increment_young_collection_count() { 2275 if(TraceYoungGenTime) { 2276 ++_young_pause_num; 2277 } 2278 } 2279 2280 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2281 if(TraceYoungGenTime) { 2282 ++_mixed_pause_num; 2283 } 2284 } 2285 2286 void TraceYoungGenTimeData::print_summary(const char* str, 2287 const NumberSeq* seq) const { 2288 double sum = seq->sum(); 2289 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2290 str, sum / 1000.0, seq->avg()); 2291 } 2292 2293 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2294 const NumberSeq* seq) const { 2295 print_summary(str, seq); 2296 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2297 "(num", seq->num(), seq->sd(), seq->maximum()); 2298 } 2299 2300 void TraceYoungGenTimeData::print() const { 2301 if (!TraceYoungGenTime) { 2302 return; 2303 } 2304 2305 gclog_or_tty->print_cr("ALL PAUSES"); 2306 print_summary_sd(" Total", &_total); 2307 gclog_or_tty->cr(); 2308 gclog_or_tty->cr(); 2309 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2310 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2311 gclog_or_tty->cr(); 2312 2313 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2314 2315 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2316 gclog_or_tty->print_cr("none"); 2317 } else { 2318 print_summary_sd(" Evacuation Pauses", &_total); 2319 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2320 print_summary(" Parallel Time", &_parallel); 2321 print_summary(" Ext Root Scanning", &_ext_root_scan); 2322 print_summary(" SATB Filtering", &_satb_filtering); 2323 print_summary(" Update RS", &_update_rs); 2324 print_summary(" Scan RS", &_scan_rs); 2325 print_summary(" Object Copy", &_obj_copy); 2326 print_summary(" Termination", &_termination); 2327 print_summary(" Parallel Other", &_parallel_other); 2328 print_summary(" Clear CT", &_clear_ct); 2329 print_summary(" Other", &_other); 2330 } 2331 gclog_or_tty->cr(); 2332 2333 gclog_or_tty->print_cr("MISC"); 2334 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2335 print_summary_sd(" Yields", &_all_yield_times_ms); 2336 } 2337 2338 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2339 if (TraceOldGenTime) { 2340 _all_full_gc_times.add(full_gc_time_ms); 2341 } 2342 } 2343 2344 void TraceOldGenTimeData::print() const { 2345 if (!TraceOldGenTime) { 2346 return; 2347 } 2348 2349 if (_all_full_gc_times.num() > 0) { 2350 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2351 _all_full_gc_times.num(), 2352 _all_full_gc_times.sum() / 1000.0); 2353 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2354 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2355 _all_full_gc_times.sd(), 2356 _all_full_gc_times.maximum()); 2357 } 2358 }