1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/concurrentG1Refine.hpp" 27 #include "gc/g1/concurrentMark.hpp" 28 #include "gc/g1/concurrentMarkThread.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorPolicy.hpp" 31 #include "gc/g1/g1ErgoVerbose.hpp" 32 #include "gc/g1/g1GCPhaseTimes.hpp" 33 #include "gc/g1/g1Log.hpp" 34 #include "gc/g1/heapRegion.inline.hpp" 35 #include "gc/g1/heapRegionRemSet.hpp" 36 #include "gc/shared/gcPolicyCounters.hpp" 37 #include "runtime/arguments.hpp" 38 #include "runtime/java.hpp" 39 #include "runtime/mutexLocker.hpp" 40 #include "utilities/debug.hpp" 41 42 // Different defaults for different number of GC threads 43 // They were chosen by running GCOld and SPECjbb on debris with different 44 // numbers of GC threads and choosing them based on the results 45 46 // all the same 47 static double rs_length_diff_defaults[] = { 48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 49 }; 50 51 static double cost_per_card_ms_defaults[] = { 52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 53 }; 54 55 // all the same 56 static double young_cards_per_entry_ratio_defaults[] = { 57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 58 }; 59 60 static double cost_per_entry_ms_defaults[] = { 61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 62 }; 63 64 static double cost_per_byte_ms_defaults[] = { 65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 66 }; 67 68 // these should be pretty consistent 69 static double constant_other_time_ms_defaults[] = { 70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 71 }; 72 73 74 static double young_other_cost_per_region_ms_defaults[] = { 75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 76 }; 77 78 static double non_young_other_cost_per_region_ms_defaults[] = { 79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 80 }; 81 82 G1CollectorPolicy::G1CollectorPolicy() : 83 _predictor(G1ConfidencePercent / 100.0), 84 _parallel_gc_threads(ParallelGCThreads), 85 86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 87 _stop_world_start(0.0), 88 89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 93 _prev_collection_pause_end_ms(0.0), 94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 105 _non_young_other_cost_per_region_ms_seq( 106 new TruncatedSeq(TruncatedSeqLength)), 107 108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 110 111 _pause_time_target_ms((double) MaxGCPauseMillis), 112 113 _recent_prev_end_times_for_all_gcs_sec( 114 new TruncatedSeq(NumPrevPausesForHeuristics)), 115 116 _recent_avg_pause_time_ratio(0.0), 117 _rs_lengths_prediction(0), 118 _max_survivor_regions(0), 119 120 _eden_used_bytes_before_gc(0), 121 _survivor_used_bytes_before_gc(0), 122 _heap_used_bytes_before_gc(0), 123 _metaspace_used_bytes_before_gc(0), 124 _eden_capacity_bytes_before_gc(0), 125 _heap_capacity_bytes_before_gc(0), 126 127 _eden_cset_region_length(0), 128 _survivor_cset_region_length(0), 129 _old_cset_region_length(0), 130 131 _collection_set(NULL), 132 _collection_set_bytes_used_before(0), 133 134 // Incremental CSet attributes 135 _inc_cset_build_state(Inactive), 136 _inc_cset_head(NULL), 137 _inc_cset_tail(NULL), 138 _inc_cset_bytes_used_before(0), 139 _inc_cset_max_finger(NULL), 140 _inc_cset_recorded_rs_lengths(0), 141 _inc_cset_recorded_rs_lengths_diffs(0), 142 _inc_cset_predicted_elapsed_time_ms(0.0), 143 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 144 145 // add here any more surv rate groups 146 _recorded_survivor_regions(0), 147 _recorded_survivor_head(NULL), 148 _recorded_survivor_tail(NULL), 149 _survivors_age_table(true), 150 151 _gc_overhead_perc(0.0) { 152 153 // SurvRateGroups below must be initialized after the predictor because they 154 // indirectly use it through this object passed to their constructor. 155 _short_lived_surv_rate_group = 156 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 157 _survivor_surv_rate_group = 158 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 159 160 // Set up the region size and associated fields. Given that the 161 // policy is created before the heap, we have to set this up here, 162 // so it's done as soon as possible. 163 164 // It would have been natural to pass initial_heap_byte_size() and 165 // max_heap_byte_size() to setup_heap_region_size() but those have 166 // not been set up at this point since they should be aligned with 167 // the region size. So, there is a circular dependency here. We base 168 // the region size on the heap size, but the heap size should be 169 // aligned with the region size. To get around this we use the 170 // unaligned values for the heap. 171 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 172 HeapRegionRemSet::setup_remset_size(); 173 174 G1ErgoVerbose::initialize(); 175 if (PrintAdaptiveSizePolicy) { 176 // Currently, we only use a single switch for all the heuristics. 177 G1ErgoVerbose::set_enabled(true); 178 // Given that we don't currently have a verboseness level 179 // parameter, we'll hardcode this to high. This can be easily 180 // changed in the future. 181 G1ErgoVerbose::set_level(ErgoHigh); 182 } else { 183 G1ErgoVerbose::set_enabled(false); 184 } 185 186 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 187 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 188 189 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 190 191 int index = MIN2(_parallel_gc_threads - 1, 7); 192 193 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 194 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 195 _cost_scan_hcc_seq->add(0.0); 196 _young_cards_per_entry_ratio_seq->add( 197 young_cards_per_entry_ratio_defaults[index]); 198 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 199 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 200 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 201 _young_other_cost_per_region_ms_seq->add( 202 young_other_cost_per_region_ms_defaults[index]); 203 _non_young_other_cost_per_region_ms_seq->add( 204 non_young_other_cost_per_region_ms_defaults[index]); 205 206 // Below, we might need to calculate the pause time target based on 207 // the pause interval. When we do so we are going to give G1 maximum 208 // flexibility and allow it to do pauses when it needs to. So, we'll 209 // arrange that the pause interval to be pause time target + 1 to 210 // ensure that a) the pause time target is maximized with respect to 211 // the pause interval and b) we maintain the invariant that pause 212 // time target < pause interval. If the user does not want this 213 // maximum flexibility, they will have to set the pause interval 214 // explicitly. 215 216 // First make sure that, if either parameter is set, its value is 217 // reasonable. 218 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 219 if (MaxGCPauseMillis < 1) { 220 vm_exit_during_initialization("MaxGCPauseMillis should be " 221 "greater than 0"); 222 } 223 } 224 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 225 if (GCPauseIntervalMillis < 1) { 226 vm_exit_during_initialization("GCPauseIntervalMillis should be " 227 "greater than 0"); 228 } 229 } 230 231 // Then, if the pause time target parameter was not set, set it to 232 // the default value. 233 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 234 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 235 // The default pause time target in G1 is 200ms 236 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 237 } else { 238 // We do not allow the pause interval to be set without the 239 // pause time target 240 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 241 "without setting MaxGCPauseMillis"); 242 } 243 } 244 245 // Then, if the interval parameter was not set, set it according to 246 // the pause time target (this will also deal with the case when the 247 // pause time target is the default value). 248 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 249 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 250 } 251 252 // Finally, make sure that the two parameters are consistent. 253 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 254 char buffer[256]; 255 jio_snprintf(buffer, 256, 256 "MaxGCPauseMillis (%u) should be less than " 257 "GCPauseIntervalMillis (%u)", 258 MaxGCPauseMillis, GCPauseIntervalMillis); 259 vm_exit_during_initialization(buffer); 260 } 261 262 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 263 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 264 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 265 266 // start conservatively (around 50ms is about right) 267 _concurrent_mark_remark_times_ms->add(0.05); 268 _concurrent_mark_cleanup_times_ms->add(0.20); 269 _tenuring_threshold = MaxTenuringThreshold; 270 271 assert(GCTimeRatio > 0, 272 "we should have set it to a default value set_g1_gc_flags() " 273 "if a user set it to 0"); 274 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 275 276 uintx reserve_perc = G1ReservePercent; 277 // Put an artificial ceiling on this so that it's not set to a silly value. 278 if (reserve_perc > 50) { 279 reserve_perc = 50; 280 warning("G1ReservePercent is set to a value that is too large, " 281 "it's been updated to " UINTX_FORMAT, reserve_perc); 282 } 283 _reserve_factor = (double) reserve_perc / 100.0; 284 // This will be set when the heap is expanded 285 // for the first time during initialization. 286 _reserve_regions = 0; 287 288 _collectionSetChooser = new CollectionSetChooser(); 289 } 290 291 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 292 return _predictor.get_new_prediction(seq); 293 } 294 295 void G1CollectorPolicy::initialize_alignments() { 296 _space_alignment = HeapRegion::GrainBytes; 297 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 298 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 299 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 300 } 301 302 void G1CollectorPolicy::initialize_flags() { 303 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 304 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 305 } 306 307 if (SurvivorRatio < 1) { 308 vm_exit_during_initialization("Invalid survivor ratio specified"); 309 } 310 CollectorPolicy::initialize_flags(); 311 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 312 } 313 314 void G1CollectorPolicy::post_heap_initialize() { 315 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 316 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 317 if (max_young_size != MaxNewSize) { 318 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 319 } 320 } 321 322 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 323 324 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 325 _min_desired_young_length(0), _max_desired_young_length(0) { 326 if (FLAG_IS_CMDLINE(NewRatio)) { 327 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 328 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 329 } else { 330 _sizer_kind = SizerNewRatio; 331 _adaptive_size = false; 332 return; 333 } 334 } 335 336 if (NewSize > MaxNewSize) { 337 if (FLAG_IS_CMDLINE(MaxNewSize)) { 338 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 339 "A new max generation size of " SIZE_FORMAT "k will be used.", 340 NewSize/K, MaxNewSize/K, NewSize/K); 341 } 342 MaxNewSize = NewSize; 343 } 344 345 if (FLAG_IS_CMDLINE(NewSize)) { 346 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 347 1U); 348 if (FLAG_IS_CMDLINE(MaxNewSize)) { 349 _max_desired_young_length = 350 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 351 1U); 352 _sizer_kind = SizerMaxAndNewSize; 353 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 354 } else { 355 _sizer_kind = SizerNewSizeOnly; 356 } 357 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 358 _max_desired_young_length = 359 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 360 1U); 361 _sizer_kind = SizerMaxNewSizeOnly; 362 } 363 } 364 365 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 366 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 367 return MAX2(1U, default_value); 368 } 369 370 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 371 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 372 return MAX2(1U, default_value); 373 } 374 375 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 376 assert(number_of_heap_regions > 0, "Heap must be initialized"); 377 378 switch (_sizer_kind) { 379 case SizerDefaults: 380 *min_young_length = calculate_default_min_length(number_of_heap_regions); 381 *max_young_length = calculate_default_max_length(number_of_heap_regions); 382 break; 383 case SizerNewSizeOnly: 384 *max_young_length = calculate_default_max_length(number_of_heap_regions); 385 *max_young_length = MAX2(*min_young_length, *max_young_length); 386 break; 387 case SizerMaxNewSizeOnly: 388 *min_young_length = calculate_default_min_length(number_of_heap_regions); 389 *min_young_length = MIN2(*min_young_length, *max_young_length); 390 break; 391 case SizerMaxAndNewSize: 392 // Do nothing. Values set on the command line, don't update them at runtime. 393 break; 394 case SizerNewRatio: 395 *min_young_length = number_of_heap_regions / (NewRatio + 1); 396 *max_young_length = *min_young_length; 397 break; 398 default: 399 ShouldNotReachHere(); 400 } 401 402 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 403 } 404 405 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 406 // We need to pass the desired values because recalculation may not update these 407 // values in some cases. 408 uint temp = _min_desired_young_length; 409 uint result = _max_desired_young_length; 410 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 411 return result; 412 } 413 414 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 415 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 416 &_max_desired_young_length); 417 } 418 419 void G1CollectorPolicy::init() { 420 // Set aside an initial future to_space. 421 _g1 = G1CollectedHeap::heap(); 422 423 assert(Heap_lock->owned_by_self(), "Locking discipline."); 424 425 initialize_gc_policy_counters(); 426 427 if (adaptive_young_list_length()) { 428 _young_list_fixed_length = 0; 429 } else { 430 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 431 } 432 _free_regions_at_end_of_collection = _g1->num_free_regions(); 433 434 update_young_list_max_and_target_length(); 435 // We may immediately start allocating regions and placing them on the 436 // collection set list. Initialize the per-collection set info 437 start_incremental_cset_building(); 438 } 439 440 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 441 phase_times()->note_gc_start(num_active_workers); 442 } 443 444 // Create the jstat counters for the policy. 445 void G1CollectorPolicy::initialize_gc_policy_counters() { 446 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 447 } 448 449 bool G1CollectorPolicy::predict_will_fit(uint young_length, 450 double base_time_ms, 451 uint base_free_regions, 452 double target_pause_time_ms) const { 453 if (young_length >= base_free_regions) { 454 // end condition 1: not enough space for the young regions 455 return false; 456 } 457 458 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 459 size_t bytes_to_copy = 460 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 461 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 462 double young_other_time_ms = predict_young_other_time_ms(young_length); 463 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 464 if (pause_time_ms > target_pause_time_ms) { 465 // end condition 2: prediction is over the target pause time 466 return false; 467 } 468 469 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 470 471 // When copying, we will likely need more bytes free than is live in the region. 472 // Add some safety margin to factor in the confidence of our guess, and the 473 // natural expected waste. 474 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 475 // of the calculation: the lower the confidence, the more headroom. 476 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 477 // copying due to anticipated waste in the PLABs. 478 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 479 size_t expected_bytes_to_copy = safety_factor * bytes_to_copy; 480 481 if (expected_bytes_to_copy > free_bytes) { 482 // end condition 3: out-of-space 483 return false; 484 } 485 486 // success! 487 return true; 488 } 489 490 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 491 // re-calculate the necessary reserve 492 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 493 // We use ceiling so that if reserve_regions_d is > 0.0 (but 494 // smaller than 1.0) we'll get 1. 495 _reserve_regions = (uint) ceil(reserve_regions_d); 496 497 _young_gen_sizer->heap_size_changed(new_number_of_regions); 498 } 499 500 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 501 uint base_min_length) const { 502 uint desired_min_length = 0; 503 if (adaptive_young_list_length()) { 504 if (_alloc_rate_ms_seq->num() > 3) { 505 double now_sec = os::elapsedTime(); 506 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 507 double alloc_rate_ms = predict_alloc_rate_ms(); 508 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 509 } else { 510 // otherwise we don't have enough info to make the prediction 511 } 512 } 513 desired_min_length += base_min_length; 514 // make sure we don't go below any user-defined minimum bound 515 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 516 } 517 518 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 519 // Here, we might want to also take into account any additional 520 // constraints (i.e., user-defined minimum bound). Currently, we 521 // effectively don't set this bound. 522 return _young_gen_sizer->max_desired_young_length(); 523 } 524 525 void G1CollectorPolicy::update_young_list_max_and_target_length() { 526 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 527 } 528 529 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 530 update_young_list_target_length(rs_lengths); 531 update_max_gc_locker_expansion(); 532 } 533 534 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 535 _young_list_target_length = bounded_young_list_target_length(rs_lengths); 536 } 537 538 void G1CollectorPolicy::update_young_list_target_length() { 539 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 540 } 541 542 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths) const { 543 // Calculate the absolute and desired min bounds. 544 545 // This is how many young regions we already have (currently: the survivors). 546 uint base_min_length = recorded_survivor_regions(); 547 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 548 // This is the absolute minimum young length. Ensure that we 549 // will at least have one eden region available for allocation. 550 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 551 // If we shrank the young list target it should not shrink below the current size. 552 desired_min_length = MAX2(desired_min_length, absolute_min_length); 553 // Calculate the absolute and desired max bounds. 554 555 // We will try our best not to "eat" into the reserve. 556 uint absolute_max_length = 0; 557 if (_free_regions_at_end_of_collection > _reserve_regions) { 558 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 559 } 560 uint desired_max_length = calculate_young_list_desired_max_length(); 561 if (desired_max_length > absolute_max_length) { 562 desired_max_length = absolute_max_length; 563 } 564 565 uint young_list_target_length = 0; 566 if (adaptive_young_list_length()) { 567 if (collector_state()->gcs_are_young()) { 568 young_list_target_length = 569 calculate_young_list_target_length(rs_lengths, 570 base_min_length, 571 desired_min_length, 572 desired_max_length); 573 } else { 574 // Don't calculate anything and let the code below bound it to 575 // the desired_min_length, i.e., do the next GC as soon as 576 // possible to maximize how many old regions we can add to it. 577 } 578 } else { 579 // The user asked for a fixed young gen so we'll fix the young gen 580 // whether the next GC is young or mixed. 581 young_list_target_length = _young_list_fixed_length; 582 } 583 584 // Make sure we don't go over the desired max length, nor under the 585 // desired min length. In case they clash, desired_min_length wins 586 // which is why that test is second. 587 if (young_list_target_length > desired_max_length) { 588 young_list_target_length = desired_max_length; 589 } 590 if (young_list_target_length < desired_min_length) { 591 young_list_target_length = desired_min_length; 592 } 593 594 assert(young_list_target_length > recorded_survivor_regions(), 595 "we should be able to allocate at least one eden region"); 596 assert(young_list_target_length >= absolute_min_length, "post-condition"); 597 598 return young_list_target_length; 599 } 600 601 uint 602 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 603 uint base_min_length, 604 uint desired_min_length, 605 uint desired_max_length) const { 606 assert(adaptive_young_list_length(), "pre-condition"); 607 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 608 609 // In case some edge-condition makes the desired max length too small... 610 if (desired_max_length <= desired_min_length) { 611 return desired_min_length; 612 } 613 614 // We'll adjust min_young_length and max_young_length not to include 615 // the already allocated young regions (i.e., so they reflect the 616 // min and max eden regions we'll allocate). The base_min_length 617 // will be reflected in the predictions by the 618 // survivor_regions_evac_time prediction. 619 assert(desired_min_length > base_min_length, "invariant"); 620 uint min_young_length = desired_min_length - base_min_length; 621 assert(desired_max_length > base_min_length, "invariant"); 622 uint max_young_length = desired_max_length - base_min_length; 623 624 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 625 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 626 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 627 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 628 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 629 double base_time_ms = 630 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 631 survivor_regions_evac_time; 632 uint available_free_regions = _free_regions_at_end_of_collection; 633 uint base_free_regions = 0; 634 if (available_free_regions > _reserve_regions) { 635 base_free_regions = available_free_regions - _reserve_regions; 636 } 637 638 // Here, we will make sure that the shortest young length that 639 // makes sense fits within the target pause time. 640 641 if (predict_will_fit(min_young_length, base_time_ms, 642 base_free_regions, target_pause_time_ms)) { 643 // The shortest young length will fit into the target pause time; 644 // we'll now check whether the absolute maximum number of young 645 // regions will fit in the target pause time. If not, we'll do 646 // a binary search between min_young_length and max_young_length. 647 if (predict_will_fit(max_young_length, base_time_ms, 648 base_free_regions, target_pause_time_ms)) { 649 // The maximum young length will fit into the target pause time. 650 // We are done so set min young length to the maximum length (as 651 // the result is assumed to be returned in min_young_length). 652 min_young_length = max_young_length; 653 } else { 654 // The maximum possible number of young regions will not fit within 655 // the target pause time so we'll search for the optimal 656 // length. The loop invariants are: 657 // 658 // min_young_length < max_young_length 659 // min_young_length is known to fit into the target pause time 660 // max_young_length is known not to fit into the target pause time 661 // 662 // Going into the loop we know the above hold as we've just 663 // checked them. Every time around the loop we check whether 664 // the middle value between min_young_length and 665 // max_young_length fits into the target pause time. If it 666 // does, it becomes the new min. If it doesn't, it becomes 667 // the new max. This way we maintain the loop invariants. 668 669 assert(min_young_length < max_young_length, "invariant"); 670 uint diff = (max_young_length - min_young_length) / 2; 671 while (diff > 0) { 672 uint young_length = min_young_length + diff; 673 if (predict_will_fit(young_length, base_time_ms, 674 base_free_regions, target_pause_time_ms)) { 675 min_young_length = young_length; 676 } else { 677 max_young_length = young_length; 678 } 679 assert(min_young_length < max_young_length, "invariant"); 680 diff = (max_young_length - min_young_length) / 2; 681 } 682 // The results is min_young_length which, according to the 683 // loop invariants, should fit within the target pause time. 684 685 // These are the post-conditions of the binary search above: 686 assert(min_young_length < max_young_length, 687 "otherwise we should have discovered that max_young_length " 688 "fits into the pause target and not done the binary search"); 689 assert(predict_will_fit(min_young_length, base_time_ms, 690 base_free_regions, target_pause_time_ms), 691 "min_young_length, the result of the binary search, should " 692 "fit into the pause target"); 693 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 694 base_free_regions, target_pause_time_ms), 695 "min_young_length, the result of the binary search, should be " 696 "optimal, so no larger length should fit into the pause target"); 697 } 698 } else { 699 // Even the minimum length doesn't fit into the pause time 700 // target, return it as the result nevertheless. 701 } 702 return base_min_length + min_young_length; 703 } 704 705 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 706 double survivor_regions_evac_time = 0.0; 707 for (HeapRegion * r = _recorded_survivor_head; 708 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 709 r = r->get_next_young_region()) { 710 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 711 } 712 return survivor_regions_evac_time; 713 } 714 715 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 716 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 717 718 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 719 if (rs_lengths > _rs_lengths_prediction) { 720 // add 10% to avoid having to recalculate often 721 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 722 update_rs_lengths_prediction(rs_lengths_prediction); 723 724 update_young_list_max_and_target_length(rs_lengths_prediction); 725 } 726 } 727 728 void G1CollectorPolicy::update_rs_lengths_prediction() { 729 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 730 } 731 732 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 733 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 734 _rs_lengths_prediction = prediction; 735 } 736 } 737 738 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 739 bool is_tlab, 740 bool* gc_overhead_limit_was_exceeded) { 741 guarantee(false, "Not using this policy feature yet."); 742 return NULL; 743 } 744 745 // This method controls how a collector handles one or more 746 // of its generations being fully allocated. 747 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 748 bool is_tlab) { 749 guarantee(false, "Not using this policy feature yet."); 750 return NULL; 751 } 752 753 754 #ifndef PRODUCT 755 bool G1CollectorPolicy::verify_young_ages() { 756 HeapRegion* head = _g1->young_list()->first_region(); 757 return 758 verify_young_ages(head, _short_lived_surv_rate_group); 759 // also call verify_young_ages on any additional surv rate groups 760 } 761 762 bool 763 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 764 SurvRateGroup *surv_rate_group) { 765 guarantee( surv_rate_group != NULL, "pre-condition" ); 766 767 const char* name = surv_rate_group->name(); 768 bool ret = true; 769 int prev_age = -1; 770 771 for (HeapRegion* curr = head; 772 curr != NULL; 773 curr = curr->get_next_young_region()) { 774 SurvRateGroup* group = curr->surv_rate_group(); 775 if (group == NULL && !curr->is_survivor()) { 776 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 777 ret = false; 778 } 779 780 if (surv_rate_group == group) { 781 int age = curr->age_in_surv_rate_group(); 782 783 if (age < 0) { 784 gclog_or_tty->print_cr("## %s: encountered negative age", name); 785 ret = false; 786 } 787 788 if (age <= prev_age) { 789 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 790 "(%d, %d)", name, age, prev_age); 791 ret = false; 792 } 793 prev_age = age; 794 } 795 } 796 797 return ret; 798 } 799 #endif // PRODUCT 800 801 void G1CollectorPolicy::record_full_collection_start() { 802 _full_collection_start_sec = os::elapsedTime(); 803 record_heap_size_info_at_start(true /* full */); 804 // Release the future to-space so that it is available for compaction into. 805 collector_state()->set_full_collection(true); 806 } 807 808 void G1CollectorPolicy::record_full_collection_end() { 809 // Consider this like a collection pause for the purposes of allocation 810 // since last pause. 811 double end_sec = os::elapsedTime(); 812 double full_gc_time_sec = end_sec - _full_collection_start_sec; 813 double full_gc_time_ms = full_gc_time_sec * 1000.0; 814 815 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 816 817 update_recent_gc_times(end_sec, full_gc_time_ms); 818 819 collector_state()->set_full_collection(false); 820 821 // "Nuke" the heuristics that control the young/mixed GC 822 // transitions and make sure we start with young GCs after the Full GC. 823 collector_state()->set_gcs_are_young(true); 824 collector_state()->set_last_young_gc(false); 825 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 826 collector_state()->set_during_initial_mark_pause(false); 827 collector_state()->set_in_marking_window(false); 828 collector_state()->set_in_marking_window_im(false); 829 830 _short_lived_surv_rate_group->start_adding_regions(); 831 // also call this on any additional surv rate groups 832 833 record_survivor_regions(0, NULL, NULL); 834 835 _free_regions_at_end_of_collection = _g1->num_free_regions(); 836 // Reset survivors SurvRateGroup. 837 _survivor_surv_rate_group->reset(); 838 update_young_list_max_and_target_length(); 839 update_rs_lengths_prediction(); 840 _collectionSetChooser->clear(); 841 } 842 843 void G1CollectorPolicy::record_stop_world_start() { 844 _stop_world_start = os::elapsedTime(); 845 } 846 847 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 848 // We only need to do this here as the policy will only be applied 849 // to the GC we're about to start. so, no point is calculating this 850 // every time we calculate / recalculate the target young length. 851 update_survivors_policy(); 852 853 assert(_g1->used() == _g1->recalculate_used(), 854 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 855 _g1->used(), _g1->recalculate_used()); 856 857 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 858 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 859 _stop_world_start = 0.0; 860 861 record_heap_size_info_at_start(false /* full */); 862 863 phase_times()->record_cur_collection_start_sec(start_time_sec); 864 _pending_cards = _g1->pending_card_num(); 865 866 _collection_set_bytes_used_before = 0; 867 _bytes_copied_during_gc = 0; 868 869 collector_state()->set_last_gc_was_young(false); 870 871 // do that for any other surv rate groups 872 _short_lived_surv_rate_group->stop_adding_regions(); 873 _survivors_age_table.clear(); 874 875 assert( verify_young_ages(), "region age verification" ); 876 } 877 878 void G1CollectorPolicy::record_concurrent_mark_init_end(double 879 mark_init_elapsed_time_ms) { 880 collector_state()->set_during_marking(true); 881 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 882 collector_state()->set_during_initial_mark_pause(false); 883 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 884 } 885 886 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 887 _mark_remark_start_sec = os::elapsedTime(); 888 collector_state()->set_during_marking(false); 889 } 890 891 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 892 double end_time_sec = os::elapsedTime(); 893 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 894 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 895 _cur_mark_stop_world_time_ms += elapsed_time_ms; 896 _prev_collection_pause_end_ms += elapsed_time_ms; 897 898 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec); 899 } 900 901 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 902 _mark_cleanup_start_sec = os::elapsedTime(); 903 } 904 905 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 906 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 907 "skip last young-only gc"); 908 collector_state()->set_last_young_gc(should_continue_with_reclaim); 909 collector_state()->set_in_marking_window(false); 910 } 911 912 void G1CollectorPolicy::record_concurrent_pause() { 913 if (_stop_world_start > 0.0) { 914 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 915 _trace_young_gen_time_data.record_yield_time(yield_ms); 916 } 917 } 918 919 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 920 return phase_times()->average_time_ms(phase); 921 } 922 923 double G1CollectorPolicy::young_other_time_ms() const { 924 return phase_times()->young_cset_choice_time_ms() + 925 phase_times()->young_free_cset_time_ms(); 926 } 927 928 double G1CollectorPolicy::non_young_other_time_ms() const { 929 return phase_times()->non_young_cset_choice_time_ms() + 930 phase_times()->non_young_free_cset_time_ms(); 931 932 } 933 934 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 935 return pause_time_ms - 936 average_time_ms(G1GCPhaseTimes::UpdateRS) - 937 average_time_ms(G1GCPhaseTimes::ScanRS) - 938 average_time_ms(G1GCPhaseTimes::ObjCopy) - 939 average_time_ms(G1GCPhaseTimes::Termination); 940 } 941 942 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 943 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 944 } 945 946 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 947 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 948 } 949 950 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 951 if (about_to_start_mixed_phase()) { 952 return false; 953 } 954 955 size_t marking_initiating_used_threshold = 956 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 957 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 958 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 959 960 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 961 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 962 ergo_verbose5(ErgoConcCycles, 963 "request concurrent cycle initiation", 964 ergo_format_reason("occupancy higher than threshold") 965 ergo_format_byte("occupancy") 966 ergo_format_byte("allocation request") 967 ergo_format_byte_perc("threshold") 968 ergo_format_str("source"), 969 cur_used_bytes, 970 alloc_byte_size, 971 marking_initiating_used_threshold, 972 (double) InitiatingHeapOccupancyPercent, 973 source); 974 return true; 975 } else { 976 ergo_verbose5(ErgoConcCycles, 977 "do not request concurrent cycle initiation", 978 ergo_format_reason("still doing mixed collections") 979 ergo_format_byte("occupancy") 980 ergo_format_byte("allocation request") 981 ergo_format_byte_perc("threshold") 982 ergo_format_str("source"), 983 cur_used_bytes, 984 alloc_byte_size, 985 marking_initiating_used_threshold, 986 (double) InitiatingHeapOccupancyPercent, 987 source); 988 } 989 } 990 991 return false; 992 } 993 994 // Anything below that is considered to be zero 995 #define MIN_TIMER_GRANULARITY 0.0000001 996 997 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 998 double end_time_sec = os::elapsedTime(); 999 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1000 "otherwise, the subtraction below does not make sense"); 1001 size_t rs_size = 1002 _cur_collection_pause_used_regions_at_start - cset_region_length(); 1003 size_t cur_used_bytes = _g1->used(); 1004 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1005 bool last_pause_included_initial_mark = false; 1006 bool update_stats = !_g1->evacuation_failed(); 1007 1008 #ifndef PRODUCT 1009 if (G1YoungSurvRateVerbose) { 1010 gclog_or_tty->cr(); 1011 _short_lived_surv_rate_group->print(); 1012 // do that for any other surv rate groups too 1013 } 1014 #endif // PRODUCT 1015 1016 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 1017 if (last_pause_included_initial_mark) { 1018 record_concurrent_mark_init_end(0.0); 1019 } else { 1020 maybe_start_marking(); 1021 } 1022 1023 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, end_time_sec); 1024 1025 if (update_stats) { 1026 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 1027 // this is where we update the allocation rate of the application 1028 double app_time_ms = 1029 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 1030 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1031 // This usually happens due to the timer not having the required 1032 // granularity. Some Linuxes are the usual culprits. 1033 // We'll just set it to something (arbitrarily) small. 1034 app_time_ms = 1.0; 1035 } 1036 // We maintain the invariant that all objects allocated by mutator 1037 // threads will be allocated out of eden regions. So, we can use 1038 // the eden region number allocated since the previous GC to 1039 // calculate the application's allocate rate. The only exception 1040 // to that is humongous objects that are allocated separately. But 1041 // given that humongous object allocations do not really affect 1042 // either the pause's duration nor when the next pause will take 1043 // place we can safely ignore them here. 1044 uint regions_allocated = eden_cset_region_length(); 1045 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1046 _alloc_rate_ms_seq->add(alloc_rate_ms); 1047 1048 double interval_ms = 1049 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1050 update_recent_gc_times(end_time_sec, pause_time_ms); 1051 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1052 if (recent_avg_pause_time_ratio() < 0.0 || 1053 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1054 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1055 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1056 if (_recent_avg_pause_time_ratio < 0.0) { 1057 _recent_avg_pause_time_ratio = 0.0; 1058 } else { 1059 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1060 _recent_avg_pause_time_ratio = 1.0; 1061 } 1062 } 1063 } 1064 1065 bool new_in_marking_window = collector_state()->in_marking_window(); 1066 bool new_in_marking_window_im = false; 1067 if (last_pause_included_initial_mark) { 1068 new_in_marking_window = true; 1069 new_in_marking_window_im = true; 1070 } 1071 1072 if (collector_state()->last_young_gc()) { 1073 // This is supposed to to be the "last young GC" before we start 1074 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1075 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1076 1077 if (next_gc_should_be_mixed("start mixed GCs", 1078 "do not start mixed GCs")) { 1079 collector_state()->set_gcs_are_young(false); 1080 } 1081 1082 collector_state()->set_last_young_gc(false); 1083 } 1084 1085 if (!collector_state()->last_gc_was_young()) { 1086 // This is a mixed GC. Here we decide whether to continue doing 1087 // mixed GCs or not. 1088 1089 if (!next_gc_should_be_mixed("continue mixed GCs", 1090 "do not continue mixed GCs")) { 1091 collector_state()->set_gcs_are_young(true); 1092 1093 maybe_start_marking(); 1094 } 1095 } 1096 1097 _short_lived_surv_rate_group->start_adding_regions(); 1098 // Do that for any other surv rate groups 1099 1100 if (update_stats) { 1101 double cost_per_card_ms = 0.0; 1102 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1103 if (_pending_cards > 0) { 1104 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1105 _cost_per_card_ms_seq->add(cost_per_card_ms); 1106 } 1107 _cost_scan_hcc_seq->add(cost_scan_hcc); 1108 1109 double cost_per_entry_ms = 0.0; 1110 if (cards_scanned > 10) { 1111 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1112 if (collector_state()->last_gc_was_young()) { 1113 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1114 } else { 1115 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1116 } 1117 } 1118 1119 if (_max_rs_lengths > 0) { 1120 double cards_per_entry_ratio = 1121 (double) cards_scanned / (double) _max_rs_lengths; 1122 if (collector_state()->last_gc_was_young()) { 1123 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1124 } else { 1125 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1126 } 1127 } 1128 1129 // This is defensive. For a while _max_rs_lengths could get 1130 // smaller than _recorded_rs_lengths which was causing 1131 // rs_length_diff to get very large and mess up the RSet length 1132 // predictions. The reason was unsafe concurrent updates to the 1133 // _inc_cset_recorded_rs_lengths field which the code below guards 1134 // against (see CR 7118202). This bug has now been fixed (see CR 1135 // 7119027). However, I'm still worried that 1136 // _inc_cset_recorded_rs_lengths might still end up somewhat 1137 // inaccurate. The concurrent refinement thread calculates an 1138 // RSet's length concurrently with other CR threads updating it 1139 // which might cause it to calculate the length incorrectly (if, 1140 // say, it's in mid-coarsening). So I'll leave in the defensive 1141 // conditional below just in case. 1142 size_t rs_length_diff = 0; 1143 if (_max_rs_lengths > _recorded_rs_lengths) { 1144 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1145 } 1146 _rs_length_diff_seq->add((double) rs_length_diff); 1147 1148 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1149 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1150 double cost_per_byte_ms = 0.0; 1151 1152 if (copied_bytes > 0) { 1153 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1154 if (collector_state()->in_marking_window()) { 1155 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1156 } else { 1157 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1158 } 1159 } 1160 1161 if (young_cset_region_length() > 0) { 1162 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1163 young_cset_region_length()); 1164 } 1165 1166 if (old_cset_region_length() > 0) { 1167 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1168 old_cset_region_length()); 1169 } 1170 1171 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1172 1173 _pending_cards_seq->add((double) _pending_cards); 1174 _rs_lengths_seq->add((double) _max_rs_lengths); 1175 } 1176 1177 collector_state()->set_in_marking_window(new_in_marking_window); 1178 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1179 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1180 update_young_list_max_and_target_length(); 1181 update_rs_lengths_prediction(); 1182 1183 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1184 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1185 1186 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1187 1188 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1189 ergo_verbose2(ErgoTiming, 1190 "adjust concurrent refinement thresholds", 1191 ergo_format_reason("Scanning the HCC expected to take longer than Update RS time goal") 1192 ergo_format_ms("Update RS time goal") 1193 ergo_format_ms("Scan HCC time"), 1194 update_rs_time_goal_ms, 1195 scan_hcc_time_ms); 1196 1197 update_rs_time_goal_ms = 0; 1198 } else { 1199 update_rs_time_goal_ms -= scan_hcc_time_ms; 1200 } 1201 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1202 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1203 update_rs_time_goal_ms); 1204 1205 _collectionSetChooser->verify(); 1206 } 1207 1208 #define EXT_SIZE_FORMAT "%.1f%s" 1209 #define EXT_SIZE_PARAMS(bytes) \ 1210 byte_size_in_proper_unit((double)(bytes)), \ 1211 proper_unit_for_byte_size((bytes)) 1212 1213 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1214 YoungList* young_list = _g1->young_list(); 1215 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1216 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1217 _heap_capacity_bytes_before_gc = _g1->capacity(); 1218 _heap_used_bytes_before_gc = _g1->used(); 1219 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1220 1221 _eden_capacity_bytes_before_gc = 1222 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1223 1224 if (full) { 1225 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1226 } 1227 } 1228 1229 void G1CollectorPolicy::print_heap_transition(size_t bytes_before) const { 1230 size_t bytes_after = _g1->used(); 1231 size_t capacity = _g1->capacity(); 1232 1233 gclog_or_tty->print(" " SIZE_FORMAT "%s->" SIZE_FORMAT "%s(" SIZE_FORMAT "%s)", 1234 byte_size_in_proper_unit(bytes_before), 1235 proper_unit_for_byte_size(bytes_before), 1236 byte_size_in_proper_unit(bytes_after), 1237 proper_unit_for_byte_size(bytes_after), 1238 byte_size_in_proper_unit(capacity), 1239 proper_unit_for_byte_size(capacity)); 1240 } 1241 1242 void G1CollectorPolicy::print_heap_transition() const { 1243 print_heap_transition(_heap_used_bytes_before_gc); 1244 } 1245 1246 void G1CollectorPolicy::print_detailed_heap_transition(bool full) const { 1247 YoungList* young_list = _g1->young_list(); 1248 1249 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1250 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1251 size_t heap_used_bytes_after_gc = _g1->used(); 1252 1253 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1254 size_t eden_capacity_bytes_after_gc = 1255 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1256 1257 gclog_or_tty->print( 1258 " [Eden: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ") " 1259 "Survivors: " EXT_SIZE_FORMAT "->" EXT_SIZE_FORMAT " " 1260 "Heap: " EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")->" 1261 EXT_SIZE_FORMAT "(" EXT_SIZE_FORMAT ")]", 1262 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), 1263 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), 1264 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), 1265 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), 1266 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), 1267 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), 1268 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), 1269 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), 1270 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), 1271 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); 1272 1273 if (full) { 1274 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1275 } 1276 1277 gclog_or_tty->cr(); 1278 } 1279 1280 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1281 phase_times()->print(pause_time_sec); 1282 } 1283 1284 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1285 double update_rs_processed_buffers, 1286 double goal_ms) { 1287 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1288 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1289 1290 if (G1UseAdaptiveConcRefinement) { 1291 const int k_gy = 3, k_gr = 6; 1292 const double inc_k = 1.1, dec_k = 0.9; 1293 1294 int g = cg1r->green_zone(); 1295 if (update_rs_time > goal_ms) { 1296 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1297 } else { 1298 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1299 g = (int)MAX2(g * inc_k, g + 1.0); 1300 } 1301 } 1302 // Change the refinement threads params 1303 cg1r->set_green_zone(g); 1304 cg1r->set_yellow_zone(g * k_gy); 1305 cg1r->set_red_zone(g * k_gr); 1306 cg1r->reinitialize_threads(); 1307 1308 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1309 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1310 cg1r->yellow_zone()); 1311 // Change the barrier params 1312 dcqs.set_process_completed_threshold(processing_threshold); 1313 dcqs.set_max_completed_queue(cg1r->red_zone()); 1314 } 1315 1316 int curr_queue_size = dcqs.completed_buffers_num(); 1317 if (curr_queue_size >= cg1r->yellow_zone()) { 1318 dcqs.set_completed_queue_padding(curr_queue_size); 1319 } else { 1320 dcqs.set_completed_queue_padding(0); 1321 } 1322 dcqs.notify_if_necessary(); 1323 } 1324 1325 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1326 return (size_t) get_new_prediction(_rs_length_diff_seq); 1327 } 1328 1329 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1330 return get_new_prediction(_alloc_rate_ms_seq); 1331 } 1332 1333 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1334 return get_new_prediction(_cost_per_card_ms_seq); 1335 } 1336 1337 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1338 return get_new_prediction(_cost_scan_hcc_seq); 1339 } 1340 1341 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1342 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1343 } 1344 1345 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1346 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1347 } 1348 1349 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1350 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1351 return predict_young_cards_per_entry_ratio(); 1352 } else { 1353 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1354 } 1355 } 1356 1357 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1358 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1359 } 1360 1361 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1362 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1363 } 1364 1365 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1366 if (collector_state()->gcs_are_young()) { 1367 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1368 } else { 1369 return predict_mixed_rs_scan_time_ms(card_num); 1370 } 1371 } 1372 1373 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1374 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1375 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1376 } else { 1377 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1378 } 1379 } 1380 1381 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1382 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1383 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1384 } else { 1385 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1386 } 1387 } 1388 1389 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1390 if (collector_state()->during_concurrent_mark()) { 1391 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1392 } else { 1393 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1394 } 1395 } 1396 1397 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1398 return get_new_prediction(_constant_other_time_ms_seq); 1399 } 1400 1401 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1402 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1403 } 1404 1405 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1406 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1407 } 1408 1409 double G1CollectorPolicy::predict_remark_time_ms() const { 1410 return get_new_prediction(_concurrent_mark_remark_times_ms); 1411 } 1412 1413 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1414 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1415 } 1416 1417 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1418 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1419 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1420 double pred = get_new_prediction(seq); 1421 if (pred > 1.0) { 1422 pred = 1.0; 1423 } 1424 return pred; 1425 } 1426 1427 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1428 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1429 } 1430 1431 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1432 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1433 } 1434 1435 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1436 size_t scanned_cards) const { 1437 return 1438 predict_rs_update_time_ms(pending_cards) + 1439 predict_rs_scan_time_ms(scanned_cards) + 1440 predict_constant_other_time_ms(); 1441 } 1442 1443 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1444 size_t rs_length = predict_rs_length_diff(); 1445 size_t card_num; 1446 if (collector_state()->gcs_are_young()) { 1447 card_num = predict_young_card_num(rs_length); 1448 } else { 1449 card_num = predict_non_young_card_num(rs_length); 1450 } 1451 return predict_base_elapsed_time_ms(pending_cards, card_num); 1452 } 1453 1454 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1455 size_t bytes_to_copy; 1456 if (hr->is_marked()) 1457 bytes_to_copy = hr->max_live_bytes(); 1458 else { 1459 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1460 int age = hr->age_in_surv_rate_group(); 1461 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1462 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1463 } 1464 return bytes_to_copy; 1465 } 1466 1467 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1468 bool for_young_gc) const { 1469 size_t rs_length = hr->rem_set()->occupied(); 1470 size_t card_num; 1471 1472 // Predicting the number of cards is based on which type of GC 1473 // we're predicting for. 1474 if (for_young_gc) { 1475 card_num = predict_young_card_num(rs_length); 1476 } else { 1477 card_num = predict_non_young_card_num(rs_length); 1478 } 1479 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1480 1481 double region_elapsed_time_ms = 1482 predict_rs_scan_time_ms(card_num) + 1483 predict_object_copy_time_ms(bytes_to_copy); 1484 1485 // The prediction of the "other" time for this region is based 1486 // upon the region type and NOT the GC type. 1487 if (hr->is_young()) { 1488 region_elapsed_time_ms += predict_young_other_time_ms(1); 1489 } else { 1490 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1491 } 1492 return region_elapsed_time_ms; 1493 } 1494 1495 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1496 uint survivor_cset_region_length) { 1497 _eden_cset_region_length = eden_cset_region_length; 1498 _survivor_cset_region_length = survivor_cset_region_length; 1499 _old_cset_region_length = 0; 1500 } 1501 1502 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1503 _recorded_rs_lengths = rs_lengths; 1504 } 1505 1506 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1507 double elapsed_ms) { 1508 _recent_gc_times_ms->add(elapsed_ms); 1509 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1510 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1511 } 1512 1513 size_t G1CollectorPolicy::expansion_amount() const { 1514 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1515 double threshold = _gc_overhead_perc; 1516 if (recent_gc_overhead > threshold) { 1517 // We will double the existing space, or take 1518 // G1ExpandByPercentOfAvailable % of the available expansion 1519 // space, whichever is smaller, bounded below by a minimum 1520 // expansion (unless that's all that's left.) 1521 const size_t min_expand_bytes = 1*M; 1522 size_t reserved_bytes = _g1->max_capacity(); 1523 size_t committed_bytes = _g1->capacity(); 1524 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1525 size_t expand_bytes; 1526 size_t expand_bytes_via_pct = 1527 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1528 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1529 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1530 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1531 1532 ergo_verbose5(ErgoHeapSizing, 1533 "attempt heap expansion", 1534 ergo_format_reason("recent GC overhead higher than " 1535 "threshold after GC") 1536 ergo_format_perc("recent GC overhead") 1537 ergo_format_perc("threshold") 1538 ergo_format_byte("uncommitted") 1539 ergo_format_byte_perc("calculated expansion amount"), 1540 recent_gc_overhead, threshold, 1541 uncommitted_bytes, 1542 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1543 1544 return expand_bytes; 1545 } else { 1546 return 0; 1547 } 1548 } 1549 1550 void G1CollectorPolicy::print_tracing_info() const { 1551 _trace_young_gen_time_data.print(); 1552 _trace_old_gen_time_data.print(); 1553 } 1554 1555 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1556 #ifndef PRODUCT 1557 _short_lived_surv_rate_group->print_surv_rate_summary(); 1558 // add this call for any other surv rate groups 1559 #endif // PRODUCT 1560 } 1561 1562 bool G1CollectorPolicy::is_young_list_full() const { 1563 uint young_list_length = _g1->young_list()->length(); 1564 uint young_list_target_length = _young_list_target_length; 1565 return young_list_length >= young_list_target_length; 1566 } 1567 1568 bool G1CollectorPolicy::can_expand_young_list() const { 1569 uint young_list_length = _g1->young_list()->length(); 1570 uint young_list_max_length = _young_list_max_length; 1571 return young_list_length < young_list_max_length; 1572 } 1573 1574 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1575 uint expansion_region_num = 0; 1576 if (GCLockerEdenExpansionPercent > 0) { 1577 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1578 double expansion_region_num_d = perc * (double) _young_list_target_length; 1579 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1580 // less than 1.0) we'll get 1. 1581 expansion_region_num = (uint) ceil(expansion_region_num_d); 1582 } else { 1583 assert(expansion_region_num == 0, "sanity"); 1584 } 1585 _young_list_max_length = _young_list_target_length + expansion_region_num; 1586 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1587 } 1588 1589 // Calculates survivor space parameters. 1590 void G1CollectorPolicy::update_survivors_policy() { 1591 double max_survivor_regions_d = 1592 (double) _young_list_target_length / (double) SurvivorRatio; 1593 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1594 // smaller than 1.0) we'll get 1. 1595 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1596 1597 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1598 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1599 } 1600 1601 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1602 // We actually check whether we are marking here and not if we are in a 1603 // reclamation phase. This means that we will schedule a concurrent mark 1604 // even while we are still in the process of reclaiming memory. 1605 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1606 if (!during_cycle) { 1607 ergo_verbose1(ErgoConcCycles, 1608 "request concurrent cycle initiation", 1609 ergo_format_reason("requested by GC cause") 1610 ergo_format_str("GC cause"), 1611 GCCause::to_string(gc_cause)); 1612 collector_state()->set_initiate_conc_mark_if_possible(true); 1613 return true; 1614 } else { 1615 ergo_verbose1(ErgoConcCycles, 1616 "do not request concurrent cycle initiation", 1617 ergo_format_reason("concurrent cycle already in progress") 1618 ergo_format_str("GC cause"), 1619 GCCause::to_string(gc_cause)); 1620 return false; 1621 } 1622 } 1623 1624 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1625 // We are about to decide on whether this pause will be an 1626 // initial-mark pause. 1627 1628 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1629 // will set it here if we have to. However, it should be cleared by 1630 // the end of the pause (it's only set for the duration of an 1631 // initial-mark pause). 1632 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1633 1634 if (collector_state()->initiate_conc_mark_if_possible()) { 1635 // We had noticed on a previous pause that the heap occupancy has 1636 // gone over the initiating threshold and we should start a 1637 // concurrent marking cycle. So we might initiate one. 1638 1639 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1640 // Initiate a new initial mark only if there is no marking or reclamation going 1641 // on. 1642 1643 collector_state()->set_during_initial_mark_pause(true); 1644 // And we can now clear initiate_conc_mark_if_possible() as 1645 // we've already acted on it. 1646 collector_state()->set_initiate_conc_mark_if_possible(false); 1647 1648 ergo_verbose0(ErgoConcCycles, 1649 "initiate concurrent cycle", 1650 ergo_format_reason("concurrent cycle initiation requested")); 1651 } else { 1652 // The concurrent marking thread is still finishing up the 1653 // previous cycle. If we start one right now the two cycles 1654 // overlap. In particular, the concurrent marking thread might 1655 // be in the process of clearing the next marking bitmap (which 1656 // we will use for the next cycle if we start one). Starting a 1657 // cycle now will be bad given that parts of the marking 1658 // information might get cleared by the marking thread. And we 1659 // cannot wait for the marking thread to finish the cycle as it 1660 // periodically yields while clearing the next marking bitmap 1661 // and, if it's in a yield point, it's waiting for us to 1662 // finish. So, at this point we will not start a cycle and we'll 1663 // let the concurrent marking thread complete the last one. 1664 ergo_verbose0(ErgoConcCycles, 1665 "do not initiate concurrent cycle", 1666 ergo_format_reason("concurrent cycle already in progress")); 1667 } 1668 } 1669 } 1670 1671 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1672 G1CollectedHeap* _g1h; 1673 CSetChooserParUpdater _cset_updater; 1674 1675 public: 1676 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1677 uint chunk_size) : 1678 _g1h(G1CollectedHeap::heap()), 1679 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1680 1681 bool doHeapRegion(HeapRegion* r) { 1682 // Do we have any marking information for this region? 1683 if (r->is_marked()) { 1684 // We will skip any region that's currently used as an old GC 1685 // alloc region (we should not consider those for collection 1686 // before we fill them up). 1687 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1688 _cset_updater.add_region(r); 1689 } 1690 } 1691 return false; 1692 } 1693 }; 1694 1695 class ParKnownGarbageTask: public AbstractGangTask { 1696 CollectionSetChooser* _hrSorted; 1697 uint _chunk_size; 1698 G1CollectedHeap* _g1; 1699 HeapRegionClaimer _hrclaimer; 1700 1701 public: 1702 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1703 AbstractGangTask("ParKnownGarbageTask"), 1704 _hrSorted(hrSorted), _chunk_size(chunk_size), 1705 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1706 1707 void work(uint worker_id) { 1708 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1709 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1710 } 1711 }; 1712 1713 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1714 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1715 const uint overpartition_factor = 4; 1716 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1717 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1718 } 1719 1720 void 1721 G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1722 _collectionSetChooser->clear(); 1723 1724 WorkGang* workers = _g1->workers(); 1725 uint n_workers = workers->active_workers(); 1726 1727 uint n_regions = _g1->num_regions(); 1728 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1729 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1730 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1731 workers->run_task(&par_known_garbage_task); 1732 1733 _collectionSetChooser->sort_regions(); 1734 1735 double end_sec = os::elapsedTime(); 1736 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1737 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1738 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1739 _prev_collection_pause_end_ms += elapsed_time_ms; 1740 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec); 1741 } 1742 1743 // Add the heap region at the head of the non-incremental collection set 1744 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1745 assert(_inc_cset_build_state == Active, "Precondition"); 1746 assert(hr->is_old(), "the region should be old"); 1747 1748 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1749 _g1->register_old_region_with_cset(hr); 1750 hr->set_next_in_collection_set(_collection_set); 1751 _collection_set = hr; 1752 _collection_set_bytes_used_before += hr->used(); 1753 size_t rs_length = hr->rem_set()->occupied(); 1754 _recorded_rs_lengths += rs_length; 1755 _old_cset_region_length += 1; 1756 } 1757 1758 // Initialize the per-collection-set information 1759 void G1CollectorPolicy::start_incremental_cset_building() { 1760 assert(_inc_cset_build_state == Inactive, "Precondition"); 1761 1762 _inc_cset_head = NULL; 1763 _inc_cset_tail = NULL; 1764 _inc_cset_bytes_used_before = 0; 1765 1766 _inc_cset_max_finger = 0; 1767 _inc_cset_recorded_rs_lengths = 0; 1768 _inc_cset_recorded_rs_lengths_diffs = 0; 1769 _inc_cset_predicted_elapsed_time_ms = 0.0; 1770 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1771 _inc_cset_build_state = Active; 1772 } 1773 1774 void G1CollectorPolicy::finalize_incremental_cset_building() { 1775 assert(_inc_cset_build_state == Active, "Precondition"); 1776 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1777 1778 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1779 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1780 // that adds a new region to the CSet. Further updates by the 1781 // concurrent refinement thread that samples the young RSet lengths 1782 // are accumulated in the *_diffs fields. Here we add the diffs to 1783 // the "main" fields. 1784 1785 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1786 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1787 } else { 1788 // This is defensive. The diff should in theory be always positive 1789 // as RSets can only grow between GCs. However, given that we 1790 // sample their size concurrently with other threads updating them 1791 // it's possible that we might get the wrong size back, which 1792 // could make the calculations somewhat inaccurate. 1793 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1794 if (_inc_cset_recorded_rs_lengths >= diffs) { 1795 _inc_cset_recorded_rs_lengths -= diffs; 1796 } else { 1797 _inc_cset_recorded_rs_lengths = 0; 1798 } 1799 } 1800 _inc_cset_predicted_elapsed_time_ms += 1801 _inc_cset_predicted_elapsed_time_ms_diffs; 1802 1803 _inc_cset_recorded_rs_lengths_diffs = 0; 1804 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1805 } 1806 1807 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1808 // This routine is used when: 1809 // * adding survivor regions to the incremental cset at the end of an 1810 // evacuation pause, 1811 // * adding the current allocation region to the incremental cset 1812 // when it is retired, and 1813 // * updating existing policy information for a region in the 1814 // incremental cset via young list RSet sampling. 1815 // Therefore this routine may be called at a safepoint by the 1816 // VM thread, or in-between safepoints by mutator threads (when 1817 // retiring the current allocation region) or a concurrent 1818 // refine thread (RSet sampling). 1819 1820 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1821 size_t used_bytes = hr->used(); 1822 _inc_cset_recorded_rs_lengths += rs_length; 1823 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1824 _inc_cset_bytes_used_before += used_bytes; 1825 1826 // Cache the values we have added to the aggregated information 1827 // in the heap region in case we have to remove this region from 1828 // the incremental collection set, or it is updated by the 1829 // rset sampling code 1830 hr->set_recorded_rs_length(rs_length); 1831 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1832 } 1833 1834 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1835 size_t new_rs_length) { 1836 // Update the CSet information that is dependent on the new RS length 1837 assert(hr->is_young(), "Precondition"); 1838 assert(!SafepointSynchronize::is_at_safepoint(), 1839 "should not be at a safepoint"); 1840 1841 // We could have updated _inc_cset_recorded_rs_lengths and 1842 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1843 // that atomically, as this code is executed by a concurrent 1844 // refinement thread, potentially concurrently with a mutator thread 1845 // allocating a new region and also updating the same fields. To 1846 // avoid the atomic operations we accumulate these updates on two 1847 // separate fields (*_diffs) and we'll just add them to the "main" 1848 // fields at the start of a GC. 1849 1850 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1851 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1852 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1853 1854 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1855 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1856 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1857 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1858 1859 hr->set_recorded_rs_length(new_rs_length); 1860 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1861 } 1862 1863 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1864 assert(hr->is_young(), "invariant"); 1865 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1866 assert(_inc_cset_build_state == Active, "Precondition"); 1867 1868 // We need to clear and set the cached recorded/cached collection set 1869 // information in the heap region here (before the region gets added 1870 // to the collection set). An individual heap region's cached values 1871 // are calculated, aggregated with the policy collection set info, 1872 // and cached in the heap region here (initially) and (subsequently) 1873 // by the Young List sampling code. 1874 1875 size_t rs_length = hr->rem_set()->occupied(); 1876 add_to_incremental_cset_info(hr, rs_length); 1877 1878 HeapWord* hr_end = hr->end(); 1879 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1880 1881 assert(!hr->in_collection_set(), "invariant"); 1882 _g1->register_young_region_with_cset(hr); 1883 assert(hr->next_in_collection_set() == NULL, "invariant"); 1884 } 1885 1886 // Add the region at the RHS of the incremental cset 1887 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1888 // We should only ever be appending survivors at the end of a pause 1889 assert(hr->is_survivor(), "Logic"); 1890 1891 // Do the 'common' stuff 1892 add_region_to_incremental_cset_common(hr); 1893 1894 // Now add the region at the right hand side 1895 if (_inc_cset_tail == NULL) { 1896 assert(_inc_cset_head == NULL, "invariant"); 1897 _inc_cset_head = hr; 1898 } else { 1899 _inc_cset_tail->set_next_in_collection_set(hr); 1900 } 1901 _inc_cset_tail = hr; 1902 } 1903 1904 // Add the region to the LHS of the incremental cset 1905 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 1906 // Survivors should be added to the RHS at the end of a pause 1907 assert(hr->is_eden(), "Logic"); 1908 1909 // Do the 'common' stuff 1910 add_region_to_incremental_cset_common(hr); 1911 1912 // Add the region at the left hand side 1913 hr->set_next_in_collection_set(_inc_cset_head); 1914 if (_inc_cset_head == NULL) { 1915 assert(_inc_cset_tail == NULL, "Invariant"); 1916 _inc_cset_tail = hr; 1917 } 1918 _inc_cset_head = hr; 1919 } 1920 1921 #ifndef PRODUCT 1922 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 1923 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 1924 1925 st->print_cr("\nCollection_set:"); 1926 HeapRegion* csr = list_head; 1927 while (csr != NULL) { 1928 HeapRegion* next = csr->next_in_collection_set(); 1929 assert(csr->in_collection_set(), "bad CS"); 1930 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 1931 HR_FORMAT_PARAMS(csr), 1932 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 1933 csr->age_in_surv_rate_group_cond()); 1934 csr = next; 1935 } 1936 } 1937 #endif // !PRODUCT 1938 1939 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 1940 // Returns the given amount of reclaimable bytes (that represents 1941 // the amount of reclaimable space still to be collected) as a 1942 // percentage of the current heap capacity. 1943 size_t capacity_bytes = _g1->capacity(); 1944 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 1945 } 1946 1947 void G1CollectorPolicy::maybe_start_marking() { 1948 if (need_to_start_conc_mark("end of GC")) { 1949 // Note: this might have already been set, if during the last 1950 // pause we decided to start a cycle but at the beginning of 1951 // this pause we decided to postpone it. That's OK. 1952 collector_state()->set_initiate_conc_mark_if_possible(true); 1953 } 1954 } 1955 1956 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 1957 const char* false_action_str) const { 1958 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1959 if (cset_chooser->is_empty()) { 1960 ergo_verbose0(ErgoMixedGCs, 1961 false_action_str, 1962 ergo_format_reason("candidate old regions not available")); 1963 return false; 1964 } 1965 1966 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1967 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 1968 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 1969 double threshold = (double) G1HeapWastePercent; 1970 if (reclaimable_perc <= threshold) { 1971 ergo_verbose4(ErgoMixedGCs, 1972 false_action_str, 1973 ergo_format_reason("reclaimable percentage not over threshold") 1974 ergo_format_region("candidate old regions") 1975 ergo_format_byte_perc("reclaimable") 1976 ergo_format_perc("threshold"), 1977 cset_chooser->remaining_regions(), 1978 reclaimable_bytes, 1979 reclaimable_perc, threshold); 1980 return false; 1981 } 1982 1983 ergo_verbose4(ErgoMixedGCs, 1984 true_action_str, 1985 ergo_format_reason("candidate old regions available") 1986 ergo_format_region("candidate old regions") 1987 ergo_format_byte_perc("reclaimable") 1988 ergo_format_perc("threshold"), 1989 cset_chooser->remaining_regions(), 1990 reclaimable_bytes, 1991 reclaimable_perc, threshold); 1992 return true; 1993 } 1994 1995 uint G1CollectorPolicy::calc_min_old_cset_length() const { 1996 // The min old CSet region bound is based on the maximum desired 1997 // number of mixed GCs after a cycle. I.e., even if some old regions 1998 // look expensive, we should add them to the CSet anyway to make 1999 // sure we go through the available old regions in no more than the 2000 // maximum desired number of mixed GCs. 2001 // 2002 // The calculation is based on the number of marked regions we added 2003 // to the CSet chooser in the first place, not how many remain, so 2004 // that the result is the same during all mixed GCs that follow a cycle. 2005 2006 const size_t region_num = (size_t) _collectionSetChooser->length(); 2007 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 2008 size_t result = region_num / gc_num; 2009 // emulate ceiling 2010 if (result * gc_num < region_num) { 2011 result += 1; 2012 } 2013 return (uint) result; 2014 } 2015 2016 uint G1CollectorPolicy::calc_max_old_cset_length() const { 2017 // The max old CSet region bound is based on the threshold expressed 2018 // as a percentage of the heap size. I.e., it should bound the 2019 // number of old regions added to the CSet irrespective of how many 2020 // of them are available. 2021 2022 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2023 const size_t region_num = g1h->num_regions(); 2024 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 2025 size_t result = region_num * perc / 100; 2026 // emulate ceiling 2027 if (100 * result < region_num * perc) { 2028 result += 1; 2029 } 2030 return (uint) result; 2031 } 2032 2033 2034 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 2035 double young_start_time_sec = os::elapsedTime(); 2036 2037 YoungList* young_list = _g1->young_list(); 2038 finalize_incremental_cset_building(); 2039 2040 guarantee(target_pause_time_ms > 0.0, 2041 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 2042 guarantee(_collection_set == NULL, "Precondition"); 2043 2044 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2045 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 2046 2047 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2048 "start choosing CSet", 2049 ergo_format_size("_pending_cards") 2050 ergo_format_ms("predicted base time") 2051 ergo_format_ms("remaining time") 2052 ergo_format_ms("target pause time"), 2053 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 2054 2055 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 2056 2057 if (collector_state()->last_gc_was_young()) { 2058 _trace_young_gen_time_data.increment_young_collection_count(); 2059 } else { 2060 _trace_young_gen_time_data.increment_mixed_collection_count(); 2061 } 2062 2063 // The young list is laid with the survivor regions from the previous 2064 // pause are appended to the RHS of the young list, i.e. 2065 // [Newly Young Regions ++ Survivors from last pause]. 2066 2067 uint survivor_region_length = young_list->survivor_length(); 2068 uint eden_region_length = young_list->eden_length(); 2069 init_cset_region_lengths(eden_region_length, survivor_region_length); 2070 2071 HeapRegion* hr = young_list->first_survivor_region(); 2072 while (hr != NULL) { 2073 assert(hr->is_survivor(), "badly formed young list"); 2074 // There is a convention that all the young regions in the CSet 2075 // are tagged as "eden", so we do this for the survivors here. We 2076 // use the special set_eden_pre_gc() as it doesn't check that the 2077 // region is free (which is not the case here). 2078 hr->set_eden_pre_gc(); 2079 hr = hr->get_next_young_region(); 2080 } 2081 2082 // Clear the fields that point to the survivor list - they are all young now. 2083 young_list->clear_survivors(); 2084 2085 _collection_set = _inc_cset_head; 2086 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2087 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 2088 2089 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, 2090 "add young regions to CSet", 2091 ergo_format_region("eden") 2092 ergo_format_region("survivors") 2093 ergo_format_ms("predicted young region time") 2094 ergo_format_ms("target pause time"), 2095 eden_region_length, survivor_region_length, 2096 _inc_cset_predicted_elapsed_time_ms, 2097 target_pause_time_ms); 2098 2099 // The number of recorded young regions is the incremental 2100 // collection set's current size 2101 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2102 2103 double young_end_time_sec = os::elapsedTime(); 2104 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 2105 2106 return time_remaining_ms; 2107 } 2108 2109 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 2110 double non_young_start_time_sec = os::elapsedTime(); 2111 double predicted_old_time_ms = 0.0; 2112 2113 2114 if (!collector_state()->gcs_are_young()) { 2115 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2116 cset_chooser->verify(); 2117 const uint min_old_cset_length = calc_min_old_cset_length(); 2118 const uint max_old_cset_length = calc_max_old_cset_length(); 2119 2120 uint expensive_region_num = 0; 2121 bool check_time_remaining = adaptive_young_list_length(); 2122 2123 HeapRegion* hr = cset_chooser->peek(); 2124 while (hr != NULL) { 2125 if (old_cset_region_length() >= max_old_cset_length) { 2126 // Added maximum number of old regions to the CSet. 2127 ergo_verbose2(ErgoCSetConstruction, 2128 "finish adding old regions to CSet", 2129 ergo_format_reason("old CSet region num reached max") 2130 ergo_format_region("old") 2131 ergo_format_region("max"), 2132 old_cset_region_length(), max_old_cset_length); 2133 break; 2134 } 2135 2136 2137 // Stop adding regions if the remaining reclaimable space is 2138 // not above G1HeapWastePercent. 2139 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2140 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2141 double threshold = (double) G1HeapWastePercent; 2142 if (reclaimable_perc <= threshold) { 2143 // We've added enough old regions that the amount of uncollected 2144 // reclaimable space is at or below the waste threshold. Stop 2145 // adding old regions to the CSet. 2146 ergo_verbose5(ErgoCSetConstruction, 2147 "finish adding old regions to CSet", 2148 ergo_format_reason("reclaimable percentage not over threshold") 2149 ergo_format_region("old") 2150 ergo_format_region("max") 2151 ergo_format_byte_perc("reclaimable") 2152 ergo_format_perc("threshold"), 2153 old_cset_region_length(), 2154 max_old_cset_length, 2155 reclaimable_bytes, 2156 reclaimable_perc, threshold); 2157 break; 2158 } 2159 2160 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2161 if (check_time_remaining) { 2162 if (predicted_time_ms > time_remaining_ms) { 2163 // Too expensive for the current CSet. 2164 2165 if (old_cset_region_length() >= min_old_cset_length) { 2166 // We have added the minimum number of old regions to the CSet, 2167 // we are done with this CSet. 2168 ergo_verbose4(ErgoCSetConstruction, 2169 "finish adding old regions to CSet", 2170 ergo_format_reason("predicted time is too high") 2171 ergo_format_ms("predicted time") 2172 ergo_format_ms("remaining time") 2173 ergo_format_region("old") 2174 ergo_format_region("min"), 2175 predicted_time_ms, time_remaining_ms, 2176 old_cset_region_length(), min_old_cset_length); 2177 break; 2178 } 2179 2180 // We'll add it anyway given that we haven't reached the 2181 // minimum number of old regions. 2182 expensive_region_num += 1; 2183 } 2184 } else { 2185 if (old_cset_region_length() >= min_old_cset_length) { 2186 // In the non-auto-tuning case, we'll finish adding regions 2187 // to the CSet if we reach the minimum. 2188 ergo_verbose2(ErgoCSetConstruction, 2189 "finish adding old regions to CSet", 2190 ergo_format_reason("old CSet region num reached min") 2191 ergo_format_region("old") 2192 ergo_format_region("min"), 2193 old_cset_region_length(), min_old_cset_length); 2194 break; 2195 } 2196 } 2197 2198 // We will add this region to the CSet. 2199 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2200 predicted_old_time_ms += predicted_time_ms; 2201 cset_chooser->pop(); // already have region via peek() 2202 _g1->old_set_remove(hr); 2203 add_old_region_to_cset(hr); 2204 2205 hr = cset_chooser->peek(); 2206 } 2207 if (hr == NULL) { 2208 ergo_verbose0(ErgoCSetConstruction, 2209 "finish adding old regions to CSet", 2210 ergo_format_reason("candidate old regions not available")); 2211 } 2212 2213 if (expensive_region_num > 0) { 2214 // We print the information once here at the end, predicated on 2215 // whether we added any apparently expensive regions or not, to 2216 // avoid generating output per region. 2217 ergo_verbose4(ErgoCSetConstruction, 2218 "added expensive regions to CSet", 2219 ergo_format_reason("old CSet region num not reached min") 2220 ergo_format_region("old") 2221 ergo_format_region("expensive") 2222 ergo_format_region("min") 2223 ergo_format_ms("remaining time"), 2224 old_cset_region_length(), 2225 expensive_region_num, 2226 min_old_cset_length, 2227 time_remaining_ms); 2228 } 2229 2230 cset_chooser->verify(); 2231 } 2232 2233 stop_incremental_cset_building(); 2234 2235 ergo_verbose3(ErgoCSetConstruction, 2236 "finish choosing CSet", 2237 ergo_format_region("old") 2238 ergo_format_ms("predicted old region time") 2239 ergo_format_ms("time remaining"), 2240 old_cset_region_length(), 2241 predicted_old_time_ms, time_remaining_ms); 2242 2243 double non_young_end_time_sec = os::elapsedTime(); 2244 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2245 } 2246 2247 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2248 if(TraceYoungGenTime) { 2249 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2250 } 2251 } 2252 2253 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2254 if(TraceYoungGenTime) { 2255 _all_yield_times_ms.add(yield_time_ms); 2256 } 2257 } 2258 2259 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2260 if(TraceYoungGenTime) { 2261 _total.add(pause_time_ms); 2262 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2263 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2264 _parallel.add(phase_times->cur_collection_par_time_ms()); 2265 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2266 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2267 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2268 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2269 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2270 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2271 2272 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2273 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2274 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2275 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2276 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2277 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2278 2279 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2280 _parallel_other.add(parallel_other_time); 2281 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2282 } 2283 } 2284 2285 void TraceYoungGenTimeData::increment_young_collection_count() { 2286 if(TraceYoungGenTime) { 2287 ++_young_pause_num; 2288 } 2289 } 2290 2291 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2292 if(TraceYoungGenTime) { 2293 ++_mixed_pause_num; 2294 } 2295 } 2296 2297 void TraceYoungGenTimeData::print_summary(const char* str, 2298 const NumberSeq* seq) const { 2299 double sum = seq->sum(); 2300 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2301 str, sum / 1000.0, seq->avg()); 2302 } 2303 2304 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2305 const NumberSeq* seq) const { 2306 print_summary(str, seq); 2307 gclog_or_tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2308 "(num", seq->num(), seq->sd(), seq->maximum()); 2309 } 2310 2311 void TraceYoungGenTimeData::print() const { 2312 if (!TraceYoungGenTime) { 2313 return; 2314 } 2315 2316 gclog_or_tty->print_cr("ALL PAUSES"); 2317 print_summary_sd(" Total", &_total); 2318 gclog_or_tty->cr(); 2319 gclog_or_tty->cr(); 2320 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2321 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2322 gclog_or_tty->cr(); 2323 2324 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2325 2326 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2327 gclog_or_tty->print_cr("none"); 2328 } else { 2329 print_summary_sd(" Evacuation Pauses", &_total); 2330 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2331 print_summary(" Parallel Time", &_parallel); 2332 print_summary(" Ext Root Scanning", &_ext_root_scan); 2333 print_summary(" SATB Filtering", &_satb_filtering); 2334 print_summary(" Update RS", &_update_rs); 2335 print_summary(" Scan RS", &_scan_rs); 2336 print_summary(" Object Copy", &_obj_copy); 2337 print_summary(" Termination", &_termination); 2338 print_summary(" Parallel Other", &_parallel_other); 2339 print_summary(" Clear CT", &_clear_ct); 2340 print_summary(" Other", &_other); 2341 } 2342 gclog_or_tty->cr(); 2343 2344 gclog_or_tty->print_cr("MISC"); 2345 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2346 print_summary_sd(" Yields", &_all_yield_times_ms); 2347 } 2348 2349 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2350 if (TraceOldGenTime) { 2351 _all_full_gc_times.add(full_gc_time_ms); 2352 } 2353 } 2354 2355 void TraceOldGenTimeData::print() const { 2356 if (!TraceOldGenTime) { 2357 return; 2358 } 2359 2360 if (_all_full_gc_times.num() > 0) { 2361 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2362 _all_full_gc_times.num(), 2363 _all_full_gc_times.sum() / 1000.0); 2364 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2365 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2366 _all_full_gc_times.sd(), 2367 _all_full_gc_times.maximum()); 2368 } 2369 }