1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/concurrentG1Refine.hpp" 27 #include "gc/g1/concurrentMark.hpp" 28 #include "gc/g1/concurrentMarkThread.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorPolicy.hpp" 31 #include "gc/g1/g1GCPhaseTimes.hpp" 32 #include "gc/g1/heapRegion.inline.hpp" 33 #include "gc/g1/heapRegionRemSet.hpp" 34 #include "gc/shared/gcPolicyCounters.hpp" 35 #include "logging/log.hpp" 36 #include "runtime/arguments.hpp" 37 #include "runtime/java.hpp" 38 #include "runtime/mutexLocker.hpp" 39 #include "utilities/debug.hpp" 40 41 // Different defaults for different number of GC threads 42 // They were chosen by running GCOld and SPECjbb on debris with different 43 // numbers of GC threads and choosing them based on the results 44 45 // all the same 46 static double rs_length_diff_defaults[] = { 47 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 48 }; 49 50 static double cost_per_card_ms_defaults[] = { 51 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 52 }; 53 54 // all the same 55 static double young_cards_per_entry_ratio_defaults[] = { 56 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 57 }; 58 59 static double cost_per_entry_ms_defaults[] = { 60 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 61 }; 62 63 static double cost_per_byte_ms_defaults[] = { 64 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 65 }; 66 67 // these should be pretty consistent 68 static double constant_other_time_ms_defaults[] = { 69 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 70 }; 71 72 73 static double young_other_cost_per_region_ms_defaults[] = { 74 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 75 }; 76 77 static double non_young_other_cost_per_region_ms_defaults[] = { 78 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 79 }; 80 81 G1CollectorPolicy::G1CollectorPolicy() : 82 _predictor(G1ConfidencePercent / 100.0), 83 _parallel_gc_threads(ParallelGCThreads), 84 85 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 86 _stop_world_start(0.0), 87 88 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 89 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 90 91 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 92 _prev_collection_pause_end_ms(0.0), 93 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 94 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 95 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)), 96 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 97 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 98 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 99 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 100 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 101 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 102 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 103 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 104 _non_young_other_cost_per_region_ms_seq( 105 new TruncatedSeq(TruncatedSeqLength)), 106 107 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 108 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 109 110 _pause_time_target_ms((double) MaxGCPauseMillis), 111 112 _recent_prev_end_times_for_all_gcs_sec( 113 new TruncatedSeq(NumPrevPausesForHeuristics)), 114 115 _recent_avg_pause_time_ratio(0.0), 116 _rs_lengths_prediction(0), 117 _max_survivor_regions(0), 118 119 _eden_used_bytes_before_gc(0), 120 _survivor_used_bytes_before_gc(0), 121 _old_used_bytes_before_gc(0), 122 _heap_used_bytes_before_gc(0), 123 _metaspace_used_bytes_before_gc(0), 124 _eden_capacity_bytes_before_gc(0), 125 _heap_capacity_bytes_before_gc(0), 126 127 _eden_cset_region_length(0), 128 _survivor_cset_region_length(0), 129 _old_cset_region_length(0), 130 131 _collection_set(NULL), 132 _collection_set_bytes_used_before(0), 133 134 // Incremental CSet attributes 135 _inc_cset_build_state(Inactive), 136 _inc_cset_head(NULL), 137 _inc_cset_tail(NULL), 138 _inc_cset_bytes_used_before(0), 139 _inc_cset_max_finger(NULL), 140 _inc_cset_recorded_rs_lengths(0), 141 _inc_cset_recorded_rs_lengths_diffs(0), 142 _inc_cset_predicted_elapsed_time_ms(0.0), 143 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 144 145 // add here any more surv rate groups 146 _recorded_survivor_regions(0), 147 _recorded_survivor_head(NULL), 148 _recorded_survivor_tail(NULL), 149 _survivors_age_table(true), 150 151 _gc_overhead_perc(0.0) { 152 153 // SurvRateGroups below must be initialized after the predictor because they 154 // indirectly use it through this object passed to their constructor. 155 _short_lived_surv_rate_group = 156 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary); 157 _survivor_surv_rate_group = 158 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary); 159 160 // Set up the region size and associated fields. Given that the 161 // policy is created before the heap, we have to set this up here, 162 // so it's done as soon as possible. 163 164 // It would have been natural to pass initial_heap_byte_size() and 165 // max_heap_byte_size() to setup_heap_region_size() but those have 166 // not been set up at this point since they should be aligned with 167 // the region size. So, there is a circular dependency here. We base 168 // the region size on the heap size, but the heap size should be 169 // aligned with the region size. To get around this we use the 170 // unaligned values for the heap. 171 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); 172 HeapRegionRemSet::setup_remset_size(); 173 174 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 175 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 176 177 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); 178 179 int index = MIN2(_parallel_gc_threads - 1, 7); 180 181 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 182 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 183 _cost_scan_hcc_seq->add(0.0); 184 _young_cards_per_entry_ratio_seq->add( 185 young_cards_per_entry_ratio_defaults[index]); 186 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 187 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 188 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 189 _young_other_cost_per_region_ms_seq->add( 190 young_other_cost_per_region_ms_defaults[index]); 191 _non_young_other_cost_per_region_ms_seq->add( 192 non_young_other_cost_per_region_ms_defaults[index]); 193 194 // Below, we might need to calculate the pause time target based on 195 // the pause interval. When we do so we are going to give G1 maximum 196 // flexibility and allow it to do pauses when it needs to. So, we'll 197 // arrange that the pause interval to be pause time target + 1 to 198 // ensure that a) the pause time target is maximized with respect to 199 // the pause interval and b) we maintain the invariant that pause 200 // time target < pause interval. If the user does not want this 201 // maximum flexibility, they will have to set the pause interval 202 // explicitly. 203 204 // First make sure that, if either parameter is set, its value is 205 // reasonable. 206 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 207 if (MaxGCPauseMillis < 1) { 208 vm_exit_during_initialization("MaxGCPauseMillis should be " 209 "greater than 0"); 210 } 211 } 212 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 213 if (GCPauseIntervalMillis < 1) { 214 vm_exit_during_initialization("GCPauseIntervalMillis should be " 215 "greater than 0"); 216 } 217 } 218 219 // Then, if the pause time target parameter was not set, set it to 220 // the default value. 221 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 222 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 223 // The default pause time target in G1 is 200ms 224 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 225 } else { 226 // We do not allow the pause interval to be set without the 227 // pause time target 228 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 229 "without setting MaxGCPauseMillis"); 230 } 231 } 232 233 // Then, if the interval parameter was not set, set it according to 234 // the pause time target (this will also deal with the case when the 235 // pause time target is the default value). 236 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 237 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 238 } 239 240 // Finally, make sure that the two parameters are consistent. 241 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 242 char buffer[256]; 243 jio_snprintf(buffer, 256, 244 "MaxGCPauseMillis (%u) should be less than " 245 "GCPauseIntervalMillis (%u)", 246 MaxGCPauseMillis, GCPauseIntervalMillis); 247 vm_exit_during_initialization(buffer); 248 } 249 250 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 251 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 252 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 253 254 // start conservatively (around 50ms is about right) 255 _concurrent_mark_remark_times_ms->add(0.05); 256 _concurrent_mark_cleanup_times_ms->add(0.20); 257 _tenuring_threshold = MaxTenuringThreshold; 258 259 assert(GCTimeRatio > 0, 260 "we should have set it to a default value set_g1_gc_flags() " 261 "if a user set it to 0"); 262 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 263 264 uintx reserve_perc = G1ReservePercent; 265 // Put an artificial ceiling on this so that it's not set to a silly value. 266 if (reserve_perc > 50) { 267 reserve_perc = 50; 268 warning("G1ReservePercent is set to a value that is too large, " 269 "it's been updated to " UINTX_FORMAT, reserve_perc); 270 } 271 _reserve_factor = (double) reserve_perc / 100.0; 272 // This will be set when the heap is expanded 273 // for the first time during initialization. 274 _reserve_regions = 0; 275 276 _collectionSetChooser = new CollectionSetChooser(); 277 } 278 279 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const { 280 return _predictor.get_new_prediction(seq); 281 } 282 283 void G1CollectorPolicy::initialize_alignments() { 284 _space_alignment = HeapRegion::GrainBytes; 285 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint(); 286 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 287 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); 288 } 289 290 void G1CollectorPolicy::initialize_flags() { 291 if (G1HeapRegionSize != HeapRegion::GrainBytes) { 292 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes); 293 } 294 295 if (SurvivorRatio < 1) { 296 vm_exit_during_initialization("Invalid survivor ratio specified"); 297 } 298 CollectorPolicy::initialize_flags(); 299 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 300 } 301 302 void G1CollectorPolicy::post_heap_initialize() { 303 uintx max_regions = G1CollectedHeap::heap()->max_regions(); 304 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; 305 if (max_young_size != MaxNewSize) { 306 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size); 307 } 308 } 309 310 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); } 311 312 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), 313 _min_desired_young_length(0), _max_desired_young_length(0) { 314 if (FLAG_IS_CMDLINE(NewRatio)) { 315 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 316 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 317 } else { 318 _sizer_kind = SizerNewRatio; 319 _adaptive_size = false; 320 return; 321 } 322 } 323 324 if (NewSize > MaxNewSize) { 325 if (FLAG_IS_CMDLINE(MaxNewSize)) { 326 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " 327 "A new max generation size of " SIZE_FORMAT "k will be used.", 328 NewSize/K, MaxNewSize/K, NewSize/K); 329 } 330 MaxNewSize = NewSize; 331 } 332 333 if (FLAG_IS_CMDLINE(NewSize)) { 334 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 335 1U); 336 if (FLAG_IS_CMDLINE(MaxNewSize)) { 337 _max_desired_young_length = 338 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 339 1U); 340 _sizer_kind = SizerMaxAndNewSize; 341 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 342 } else { 343 _sizer_kind = SizerNewSizeOnly; 344 } 345 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 346 _max_desired_young_length = 347 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 348 1U); 349 _sizer_kind = SizerMaxNewSizeOnly; 350 } 351 } 352 353 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 354 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; 355 return MAX2(1U, default_value); 356 } 357 358 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 359 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; 360 return MAX2(1U, default_value); 361 } 362 363 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { 364 assert(number_of_heap_regions > 0, "Heap must be initialized"); 365 366 switch (_sizer_kind) { 367 case SizerDefaults: 368 *min_young_length = calculate_default_min_length(number_of_heap_regions); 369 *max_young_length = calculate_default_max_length(number_of_heap_regions); 370 break; 371 case SizerNewSizeOnly: 372 *max_young_length = calculate_default_max_length(number_of_heap_regions); 373 *max_young_length = MAX2(*min_young_length, *max_young_length); 374 break; 375 case SizerMaxNewSizeOnly: 376 *min_young_length = calculate_default_min_length(number_of_heap_regions); 377 *min_young_length = MIN2(*min_young_length, *max_young_length); 378 break; 379 case SizerMaxAndNewSize: 380 // Do nothing. Values set on the command line, don't update them at runtime. 381 break; 382 case SizerNewRatio: 383 *min_young_length = number_of_heap_regions / (NewRatio + 1); 384 *max_young_length = *min_young_length; 385 break; 386 default: 387 ShouldNotReachHere(); 388 } 389 390 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); 391 } 392 393 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { 394 // We need to pass the desired values because recalculation may not update these 395 // values in some cases. 396 uint temp = _min_desired_young_length; 397 uint result = _max_desired_young_length; 398 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); 399 return result; 400 } 401 402 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 403 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, 404 &_max_desired_young_length); 405 } 406 407 void G1CollectorPolicy::init() { 408 // Set aside an initial future to_space. 409 _g1 = G1CollectedHeap::heap(); 410 411 assert(Heap_lock->owned_by_self(), "Locking discipline."); 412 413 initialize_gc_policy_counters(); 414 415 if (adaptive_young_list_length()) { 416 _young_list_fixed_length = 0; 417 } else { 418 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 419 } 420 _free_regions_at_end_of_collection = _g1->num_free_regions(); 421 422 update_young_list_max_and_target_length(); 423 // We may immediately start allocating regions and placing them on the 424 // collection set list. Initialize the per-collection set info 425 start_incremental_cset_building(); 426 } 427 428 void G1CollectorPolicy::note_gc_start(uint num_active_workers) { 429 phase_times()->note_gc_start(num_active_workers); 430 } 431 432 // Create the jstat counters for the policy. 433 void G1CollectorPolicy::initialize_gc_policy_counters() { 434 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 435 } 436 437 bool G1CollectorPolicy::predict_will_fit(uint young_length, 438 double base_time_ms, 439 uint base_free_regions, 440 double target_pause_time_ms) const { 441 if (young_length >= base_free_regions) { 442 // end condition 1: not enough space for the young regions 443 return false; 444 } 445 446 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 447 size_t bytes_to_copy = 448 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 449 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 450 double young_other_time_ms = predict_young_other_time_ms(young_length); 451 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 452 if (pause_time_ms > target_pause_time_ms) { 453 // end condition 2: prediction is over the target pause time 454 return false; 455 } 456 457 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes; 458 459 // When copying, we will likely need more bytes free than is live in the region. 460 // Add some safety margin to factor in the confidence of our guess, and the 461 // natural expected waste. 462 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 463 // of the calculation: the lower the confidence, the more headroom. 464 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 465 // copying due to anticipated waste in the PLABs. 466 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 467 size_t expected_bytes_to_copy = safety_factor * bytes_to_copy; 468 469 if (expected_bytes_to_copy > free_bytes) { 470 // end condition 3: out-of-space 471 return false; 472 } 473 474 // success! 475 return true; 476 } 477 478 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 479 // re-calculate the necessary reserve 480 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 481 // We use ceiling so that if reserve_regions_d is > 0.0 (but 482 // smaller than 1.0) we'll get 1. 483 _reserve_regions = (uint) ceil(reserve_regions_d); 484 485 _young_gen_sizer->heap_size_changed(new_number_of_regions); 486 } 487 488 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 489 uint base_min_length) const { 490 uint desired_min_length = 0; 491 if (adaptive_young_list_length()) { 492 if (_alloc_rate_ms_seq->num() > 3) { 493 double now_sec = os::elapsedTime(); 494 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 495 double alloc_rate_ms = predict_alloc_rate_ms(); 496 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 497 } else { 498 // otherwise we don't have enough info to make the prediction 499 } 500 } 501 desired_min_length += base_min_length; 502 // make sure we don't go below any user-defined minimum bound 503 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 504 } 505 506 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const { 507 // Here, we might want to also take into account any additional 508 // constraints (i.e., user-defined minimum bound). Currently, we 509 // effectively don't set this bound. 510 return _young_gen_sizer->max_desired_young_length(); 511 } 512 513 void G1CollectorPolicy::update_young_list_max_and_target_length() { 514 update_young_list_max_and_target_length(get_new_prediction(_rs_lengths_seq)); 515 } 516 517 void G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) { 518 update_young_list_target_length(rs_lengths); 519 update_max_gc_locker_expansion(); 520 } 521 522 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 523 _young_list_target_length = bounded_young_list_target_length(rs_lengths); 524 } 525 526 void G1CollectorPolicy::update_young_list_target_length() { 527 update_young_list_target_length(get_new_prediction(_rs_lengths_seq)); 528 } 529 530 uint G1CollectorPolicy::bounded_young_list_target_length(size_t rs_lengths) const { 531 // Calculate the absolute and desired min bounds. 532 533 // This is how many young regions we already have (currently: the survivors). 534 uint base_min_length = recorded_survivor_regions(); 535 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length); 536 // This is the absolute minimum young length. Ensure that we 537 // will at least have one eden region available for allocation. 538 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1); 539 // If we shrank the young list target it should not shrink below the current size. 540 desired_min_length = MAX2(desired_min_length, absolute_min_length); 541 // Calculate the absolute and desired max bounds. 542 543 // We will try our best not to "eat" into the reserve. 544 uint absolute_max_length = 0; 545 if (_free_regions_at_end_of_collection > _reserve_regions) { 546 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 547 } 548 uint desired_max_length = calculate_young_list_desired_max_length(); 549 if (desired_max_length > absolute_max_length) { 550 desired_max_length = absolute_max_length; 551 } 552 553 uint young_list_target_length = 0; 554 if (adaptive_young_list_length()) { 555 if (collector_state()->gcs_are_young()) { 556 young_list_target_length = 557 calculate_young_list_target_length(rs_lengths, 558 base_min_length, 559 desired_min_length, 560 desired_max_length); 561 } else { 562 // Don't calculate anything and let the code below bound it to 563 // the desired_min_length, i.e., do the next GC as soon as 564 // possible to maximize how many old regions we can add to it. 565 } 566 } else { 567 // The user asked for a fixed young gen so we'll fix the young gen 568 // whether the next GC is young or mixed. 569 young_list_target_length = _young_list_fixed_length; 570 } 571 572 // Make sure we don't go over the desired max length, nor under the 573 // desired min length. In case they clash, desired_min_length wins 574 // which is why that test is second. 575 if (young_list_target_length > desired_max_length) { 576 young_list_target_length = desired_max_length; 577 } 578 if (young_list_target_length < desired_min_length) { 579 young_list_target_length = desired_min_length; 580 } 581 582 assert(young_list_target_length > recorded_survivor_regions(), 583 "we should be able to allocate at least one eden region"); 584 assert(young_list_target_length >= absolute_min_length, "post-condition"); 585 586 return young_list_target_length; 587 } 588 589 uint 590 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 591 uint base_min_length, 592 uint desired_min_length, 593 uint desired_max_length) const { 594 assert(adaptive_young_list_length(), "pre-condition"); 595 assert(collector_state()->gcs_are_young(), "only call this for young GCs"); 596 597 // In case some edge-condition makes the desired max length too small... 598 if (desired_max_length <= desired_min_length) { 599 return desired_min_length; 600 } 601 602 // We'll adjust min_young_length and max_young_length not to include 603 // the already allocated young regions (i.e., so they reflect the 604 // min and max eden regions we'll allocate). The base_min_length 605 // will be reflected in the predictions by the 606 // survivor_regions_evac_time prediction. 607 assert(desired_min_length > base_min_length, "invariant"); 608 uint min_young_length = desired_min_length - base_min_length; 609 assert(desired_max_length > base_min_length, "invariant"); 610 uint max_young_length = desired_max_length - base_min_length; 611 612 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 613 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 614 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 615 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 616 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 617 double base_time_ms = 618 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 619 survivor_regions_evac_time; 620 uint available_free_regions = _free_regions_at_end_of_collection; 621 uint base_free_regions = 0; 622 if (available_free_regions > _reserve_regions) { 623 base_free_regions = available_free_regions - _reserve_regions; 624 } 625 626 // Here, we will make sure that the shortest young length that 627 // makes sense fits within the target pause time. 628 629 if (predict_will_fit(min_young_length, base_time_ms, 630 base_free_regions, target_pause_time_ms)) { 631 // The shortest young length will fit into the target pause time; 632 // we'll now check whether the absolute maximum number of young 633 // regions will fit in the target pause time. If not, we'll do 634 // a binary search between min_young_length and max_young_length. 635 if (predict_will_fit(max_young_length, base_time_ms, 636 base_free_regions, target_pause_time_ms)) { 637 // The maximum young length will fit into the target pause time. 638 // We are done so set min young length to the maximum length (as 639 // the result is assumed to be returned in min_young_length). 640 min_young_length = max_young_length; 641 } else { 642 // The maximum possible number of young regions will not fit within 643 // the target pause time so we'll search for the optimal 644 // length. The loop invariants are: 645 // 646 // min_young_length < max_young_length 647 // min_young_length is known to fit into the target pause time 648 // max_young_length is known not to fit into the target pause time 649 // 650 // Going into the loop we know the above hold as we've just 651 // checked them. Every time around the loop we check whether 652 // the middle value between min_young_length and 653 // max_young_length fits into the target pause time. If it 654 // does, it becomes the new min. If it doesn't, it becomes 655 // the new max. This way we maintain the loop invariants. 656 657 assert(min_young_length < max_young_length, "invariant"); 658 uint diff = (max_young_length - min_young_length) / 2; 659 while (diff > 0) { 660 uint young_length = min_young_length + diff; 661 if (predict_will_fit(young_length, base_time_ms, 662 base_free_regions, target_pause_time_ms)) { 663 min_young_length = young_length; 664 } else { 665 max_young_length = young_length; 666 } 667 assert(min_young_length < max_young_length, "invariant"); 668 diff = (max_young_length - min_young_length) / 2; 669 } 670 // The results is min_young_length which, according to the 671 // loop invariants, should fit within the target pause time. 672 673 // These are the post-conditions of the binary search above: 674 assert(min_young_length < max_young_length, 675 "otherwise we should have discovered that max_young_length " 676 "fits into the pause target and not done the binary search"); 677 assert(predict_will_fit(min_young_length, base_time_ms, 678 base_free_regions, target_pause_time_ms), 679 "min_young_length, the result of the binary search, should " 680 "fit into the pause target"); 681 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 682 base_free_regions, target_pause_time_ms), 683 "min_young_length, the result of the binary search, should be " 684 "optimal, so no larger length should fit into the pause target"); 685 } 686 } else { 687 // Even the minimum length doesn't fit into the pause time 688 // target, return it as the result nevertheless. 689 } 690 return base_min_length + min_young_length; 691 } 692 693 double G1CollectorPolicy::predict_survivor_regions_evac_time() const { 694 double survivor_regions_evac_time = 0.0; 695 for (HeapRegion * r = _recorded_survivor_head; 696 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 697 r = r->get_next_young_region()) { 698 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young()); 699 } 700 return survivor_regions_evac_time; 701 } 702 703 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 704 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 705 706 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 707 if (rs_lengths > _rs_lengths_prediction) { 708 // add 10% to avoid having to recalculate often 709 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 710 update_rs_lengths_prediction(rs_lengths_prediction); 711 712 update_young_list_max_and_target_length(rs_lengths_prediction); 713 } 714 } 715 716 void G1CollectorPolicy::update_rs_lengths_prediction() { 717 update_rs_lengths_prediction(get_new_prediction(_rs_lengths_seq)); 718 } 719 720 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) { 721 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) { 722 _rs_lengths_prediction = prediction; 723 } 724 } 725 726 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 727 bool is_tlab, 728 bool* gc_overhead_limit_was_exceeded) { 729 guarantee(false, "Not using this policy feature yet."); 730 return NULL; 731 } 732 733 // This method controls how a collector handles one or more 734 // of its generations being fully allocated. 735 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 736 bool is_tlab) { 737 guarantee(false, "Not using this policy feature yet."); 738 return NULL; 739 } 740 741 742 #ifndef PRODUCT 743 bool G1CollectorPolicy::verify_young_ages() { 744 HeapRegion* head = _g1->young_list()->first_region(); 745 return 746 verify_young_ages(head, _short_lived_surv_rate_group); 747 // also call verify_young_ages on any additional surv rate groups 748 } 749 750 bool 751 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 752 SurvRateGroup *surv_rate_group) { 753 guarantee( surv_rate_group != NULL, "pre-condition" ); 754 755 const char* name = surv_rate_group->name(); 756 bool ret = true; 757 int prev_age = -1; 758 759 for (HeapRegion* curr = head; 760 curr != NULL; 761 curr = curr->get_next_young_region()) { 762 SurvRateGroup* group = curr->surv_rate_group(); 763 if (group == NULL && !curr->is_survivor()) { 764 log_info(gc, verify)("## %s: encountered NULL surv_rate_group", name); 765 ret = false; 766 } 767 768 if (surv_rate_group == group) { 769 int age = curr->age_in_surv_rate_group(); 770 771 if (age < 0) { 772 log_info(gc, verify)("## %s: encountered negative age", name); 773 ret = false; 774 } 775 776 if (age <= prev_age) { 777 log_info(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age); 778 ret = false; 779 } 780 prev_age = age; 781 } 782 } 783 784 return ret; 785 } 786 #endif // PRODUCT 787 788 void G1CollectorPolicy::record_full_collection_start() { 789 _full_collection_start_sec = os::elapsedTime(); 790 record_heap_size_info_at_start(true /* full */); 791 // Release the future to-space so that it is available for compaction into. 792 collector_state()->set_full_collection(true); 793 } 794 795 void G1CollectorPolicy::record_full_collection_end() { 796 // Consider this like a collection pause for the purposes of allocation 797 // since last pause. 798 double end_sec = os::elapsedTime(); 799 double full_gc_time_sec = end_sec - _full_collection_start_sec; 800 double full_gc_time_ms = full_gc_time_sec * 1000.0; 801 802 _trace_old_gen_time_data.record_full_collection(full_gc_time_ms); 803 804 update_recent_gc_times(end_sec, full_gc_time_ms); 805 806 collector_state()->set_full_collection(false); 807 808 // "Nuke" the heuristics that control the young/mixed GC 809 // transitions and make sure we start with young GCs after the Full GC. 810 collector_state()->set_gcs_are_young(true); 811 collector_state()->set_last_young_gc(false); 812 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 813 collector_state()->set_during_initial_mark_pause(false); 814 collector_state()->set_in_marking_window(false); 815 collector_state()->set_in_marking_window_im(false); 816 817 _short_lived_surv_rate_group->start_adding_regions(); 818 // also call this on any additional surv rate groups 819 820 record_survivor_regions(0, NULL, NULL); 821 822 _free_regions_at_end_of_collection = _g1->num_free_regions(); 823 // Reset survivors SurvRateGroup. 824 _survivor_surv_rate_group->reset(); 825 update_young_list_max_and_target_length(); 826 update_rs_lengths_prediction(); 827 _collectionSetChooser->clear(); 828 } 829 830 void G1CollectorPolicy::record_stop_world_start() { 831 _stop_world_start = os::elapsedTime(); 832 } 833 834 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { 835 // We only need to do this here as the policy will only be applied 836 // to the GC we're about to start. so, no point is calculating this 837 // every time we calculate / recalculate the target young length. 838 update_survivors_policy(); 839 840 assert(_g1->used() == _g1->recalculate_used(), 841 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT, 842 _g1->used(), _g1->recalculate_used()); 843 844 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 845 _trace_young_gen_time_data.record_start_collection(s_w_t_ms); 846 _stop_world_start = 0.0; 847 848 record_heap_size_info_at_start(false /* full */); 849 850 phase_times()->record_cur_collection_start_sec(start_time_sec); 851 _pending_cards = _g1->pending_card_num(); 852 853 _collection_set_bytes_used_before = 0; 854 _bytes_copied_during_gc = 0; 855 856 collector_state()->set_last_gc_was_young(false); 857 858 // do that for any other surv rate groups 859 _short_lived_surv_rate_group->stop_adding_regions(); 860 _survivors_age_table.clear(); 861 862 assert( verify_young_ages(), "region age verification" ); 863 } 864 865 void G1CollectorPolicy::record_concurrent_mark_init_end(double 866 mark_init_elapsed_time_ms) { 867 collector_state()->set_during_marking(true); 868 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 869 collector_state()->set_during_initial_mark_pause(false); 870 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 871 } 872 873 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 874 _mark_remark_start_sec = os::elapsedTime(); 875 collector_state()->set_during_marking(false); 876 } 877 878 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 879 double end_time_sec = os::elapsedTime(); 880 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 881 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 882 _cur_mark_stop_world_time_ms += elapsed_time_ms; 883 _prev_collection_pause_end_ms += elapsed_time_ms; 884 885 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec); 886 } 887 888 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 889 _mark_cleanup_start_sec = os::elapsedTime(); 890 } 891 892 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 893 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc", 894 "skip last young-only gc"); 895 collector_state()->set_last_young_gc(should_continue_with_reclaim); 896 collector_state()->set_in_marking_window(false); 897 } 898 899 void G1CollectorPolicy::record_concurrent_pause() { 900 if (_stop_world_start > 0.0) { 901 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 902 _trace_young_gen_time_data.record_yield_time(yield_ms); 903 } 904 } 905 906 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 907 return phase_times()->average_time_ms(phase); 908 } 909 910 double G1CollectorPolicy::young_other_time_ms() const { 911 return phase_times()->young_cset_choice_time_ms() + 912 phase_times()->young_free_cset_time_ms(); 913 } 914 915 double G1CollectorPolicy::non_young_other_time_ms() const { 916 return phase_times()->non_young_cset_choice_time_ms() + 917 phase_times()->non_young_free_cset_time_ms(); 918 919 } 920 921 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const { 922 return pause_time_ms - 923 average_time_ms(G1GCPhaseTimes::UpdateRS) - 924 average_time_ms(G1GCPhaseTimes::ScanRS) - 925 average_time_ms(G1GCPhaseTimes::ObjCopy) - 926 average_time_ms(G1GCPhaseTimes::Termination); 927 } 928 929 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const { 930 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms(); 931 } 932 933 bool G1CollectorPolicy::about_to_start_mixed_phase() const { 934 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc(); 935 } 936 937 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 938 if (about_to_start_mixed_phase()) { 939 return false; 940 } 941 942 size_t marking_initiating_used_threshold = 943 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 944 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 945 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 946 947 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 948 if (collector_state()->gcs_are_young() && !collector_state()->last_young_gc()) { 949 log_debug(gc, ergo)("Request concurrent cycle initiation (occupancy higher than threshold)" 950 "occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (" UINTX_FORMAT "%%) source: %s", 951 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, InitiatingHeapOccupancyPercent, source); 952 return true; 953 } else { 954 log_debug(gc, ergo)("Do not request concurrent cycle initiation (still doing mixed collections)" 955 "occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (" UINTX_FORMAT "%%) source: %s", 956 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, InitiatingHeapOccupancyPercent, source); } 957 } 958 959 return false; 960 } 961 962 // Anything below that is considered to be zero 963 #define MIN_TIMER_GRANULARITY 0.0000001 964 965 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned) { 966 double end_time_sec = os::elapsedTime(); 967 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 968 "otherwise, the subtraction below does not make sense"); 969 size_t rs_size = 970 _cur_collection_pause_used_regions_at_start - cset_region_length(); 971 size_t cur_used_bytes = _g1->used(); 972 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 973 bool last_pause_included_initial_mark = false; 974 bool update_stats = !_g1->evacuation_failed(); 975 976 NOT_PRODUCT(_short_lived_surv_rate_group->print()); 977 978 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause(); 979 if (last_pause_included_initial_mark) { 980 record_concurrent_mark_init_end(0.0); 981 } else { 982 maybe_start_marking(); 983 } 984 985 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, end_time_sec); 986 987 if (update_stats) { 988 _trace_young_gen_time_data.record_end_collection(pause_time_ms, phase_times()); 989 // this is where we update the allocation rate of the application 990 double app_time_ms = 991 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); 992 if (app_time_ms < MIN_TIMER_GRANULARITY) { 993 // This usually happens due to the timer not having the required 994 // granularity. Some Linuxes are the usual culprits. 995 // We'll just set it to something (arbitrarily) small. 996 app_time_ms = 1.0; 997 } 998 // We maintain the invariant that all objects allocated by mutator 999 // threads will be allocated out of eden regions. So, we can use 1000 // the eden region number allocated since the previous GC to 1001 // calculate the application's allocate rate. The only exception 1002 // to that is humongous objects that are allocated separately. But 1003 // given that humongous object allocations do not really affect 1004 // either the pause's duration nor when the next pause will take 1005 // place we can safely ignore them here. 1006 uint regions_allocated = eden_cset_region_length(); 1007 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1008 _alloc_rate_ms_seq->add(alloc_rate_ms); 1009 1010 double interval_ms = 1011 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1012 update_recent_gc_times(end_time_sec, pause_time_ms); 1013 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1014 if (recent_avg_pause_time_ratio() < 0.0 || 1015 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1016 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1017 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1018 if (_recent_avg_pause_time_ratio < 0.0) { 1019 _recent_avg_pause_time_ratio = 0.0; 1020 } else { 1021 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1022 _recent_avg_pause_time_ratio = 1.0; 1023 } 1024 } 1025 } 1026 1027 bool new_in_marking_window = collector_state()->in_marking_window(); 1028 bool new_in_marking_window_im = false; 1029 if (last_pause_included_initial_mark) { 1030 new_in_marking_window = true; 1031 new_in_marking_window_im = true; 1032 } 1033 1034 if (collector_state()->last_young_gc()) { 1035 // This is supposed to to be the "last young GC" before we start 1036 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1037 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC"); 1038 1039 if (next_gc_should_be_mixed("start mixed GCs", 1040 "do not start mixed GCs")) { 1041 collector_state()->set_gcs_are_young(false); 1042 } 1043 1044 collector_state()->set_last_young_gc(false); 1045 } 1046 1047 if (!collector_state()->last_gc_was_young()) { 1048 // This is a mixed GC. Here we decide whether to continue doing 1049 // mixed GCs or not. 1050 1051 if (!next_gc_should_be_mixed("continue mixed GCs", 1052 "do not continue mixed GCs")) { 1053 collector_state()->set_gcs_are_young(true); 1054 1055 maybe_start_marking(); 1056 } 1057 } 1058 1059 _short_lived_surv_rate_group->start_adding_regions(); 1060 // Do that for any other surv rate groups 1061 1062 if (update_stats) { 1063 double cost_per_card_ms = 0.0; 1064 double cost_scan_hcc = average_time_ms(G1GCPhaseTimes::ScanHCC); 1065 if (_pending_cards > 0) { 1066 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - cost_scan_hcc) / (double) _pending_cards; 1067 _cost_per_card_ms_seq->add(cost_per_card_ms); 1068 } 1069 _cost_scan_hcc_seq->add(cost_scan_hcc); 1070 1071 double cost_per_entry_ms = 0.0; 1072 if (cards_scanned > 10) { 1073 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned; 1074 if (collector_state()->last_gc_was_young()) { 1075 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1076 } else { 1077 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1078 } 1079 } 1080 1081 if (_max_rs_lengths > 0) { 1082 double cards_per_entry_ratio = 1083 (double) cards_scanned / (double) _max_rs_lengths; 1084 if (collector_state()->last_gc_was_young()) { 1085 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1086 } else { 1087 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1088 } 1089 } 1090 1091 // This is defensive. For a while _max_rs_lengths could get 1092 // smaller than _recorded_rs_lengths which was causing 1093 // rs_length_diff to get very large and mess up the RSet length 1094 // predictions. The reason was unsafe concurrent updates to the 1095 // _inc_cset_recorded_rs_lengths field which the code below guards 1096 // against (see CR 7118202). This bug has now been fixed (see CR 1097 // 7119027). However, I'm still worried that 1098 // _inc_cset_recorded_rs_lengths might still end up somewhat 1099 // inaccurate. The concurrent refinement thread calculates an 1100 // RSet's length concurrently with other CR threads updating it 1101 // which might cause it to calculate the length incorrectly (if, 1102 // say, it's in mid-coarsening). So I'll leave in the defensive 1103 // conditional below just in case. 1104 size_t rs_length_diff = 0; 1105 if (_max_rs_lengths > _recorded_rs_lengths) { 1106 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1107 } 1108 _rs_length_diff_seq->add((double) rs_length_diff); 1109 1110 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; 1111 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; 1112 double cost_per_byte_ms = 0.0; 1113 1114 if (copied_bytes > 0) { 1115 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes; 1116 if (collector_state()->in_marking_window()) { 1117 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1118 } else { 1119 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1120 } 1121 } 1122 1123 if (young_cset_region_length() > 0) { 1124 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() / 1125 young_cset_region_length()); 1126 } 1127 1128 if (old_cset_region_length() > 0) { 1129 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() / 1130 old_cset_region_length()); 1131 } 1132 1133 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms)); 1134 1135 _pending_cards_seq->add((double) _pending_cards); 1136 _rs_lengths_seq->add((double) _max_rs_lengths); 1137 } 1138 1139 collector_state()->set_in_marking_window(new_in_marking_window); 1140 collector_state()->set_in_marking_window_im(new_in_marking_window_im); 1141 _free_regions_at_end_of_collection = _g1->num_free_regions(); 1142 update_young_list_max_and_target_length(); 1143 update_rs_lengths_prediction(); 1144 1145 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1146 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1147 1148 double scan_hcc_time_ms = average_time_ms(G1GCPhaseTimes::ScanHCC); 1149 1150 if (update_rs_time_goal_ms < scan_hcc_time_ms) { 1151 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 1152 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms", 1153 update_rs_time_goal_ms, scan_hcc_time_ms); 1154 1155 update_rs_time_goal_ms = 0; 1156 } else { 1157 update_rs_time_goal_ms -= scan_hcc_time_ms; 1158 } 1159 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, 1160 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), 1161 update_rs_time_goal_ms); 1162 1163 _collectionSetChooser->verify(); 1164 } 1165 1166 #define EXT_SIZE_FORMAT "%.1f%s" 1167 #define EXT_SIZE_PARAMS(bytes) \ 1168 byte_size_in_proper_unit((double)(bytes)), \ 1169 proper_unit_for_byte_size((bytes)) 1170 1171 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { 1172 YoungList* young_list = _g1->young_list(); 1173 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); 1174 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); 1175 _heap_capacity_bytes_before_gc = _g1->capacity(); 1176 _heap_used_bytes_before_gc = _g1->used(); 1177 _old_used_bytes_before_gc = _heap_used_bytes_before_gc - _survivor_used_bytes_before_gc - _eden_used_bytes_before_gc; 1178 _cur_collection_pause_used_regions_at_start = _g1->num_used_regions(); 1179 1180 _eden_capacity_bytes_before_gc = 1181 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; 1182 1183 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); 1184 } 1185 1186 void G1CollectorPolicy::print_detailed_heap_transition() const { 1187 YoungList* young_list = _g1->young_list(); 1188 1189 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); 1190 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); 1191 size_t old_used_bytes_after_gc = _g1->used() - eden_used_bytes_after_gc - survivor_used_bytes_after_gc; 1192 1193 size_t heap_capacity_bytes_after_gc = _g1->capacity(); 1194 size_t eden_capacity_bytes_after_gc = 1195 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; 1196 size_t survivor_capacity_bytes_after_gc = _max_survivor_regions * HeapRegion::GrainBytes; 1197 1198 log_info(gc, heap)("Eden: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1199 _eden_used_bytes_before_gc / K, eden_used_bytes_after_gc /K, eden_capacity_bytes_after_gc /K); 1200 1201 log_info(gc, heap)("Survivor: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1202 _survivor_used_bytes_before_gc / K, survivor_used_bytes_after_gc /K, survivor_capacity_bytes_after_gc /K); 1203 1204 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1205 _old_used_bytes_before_gc / K, old_used_bytes_after_gc /K, heap_capacity_bytes_after_gc /K); 1206 1207 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); 1208 } 1209 1210 void G1CollectorPolicy::print_phases(double pause_time_sec) { 1211 phase_times()->print(pause_time_sec); 1212 } 1213 1214 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1215 double update_rs_processed_buffers, 1216 double goal_ms) { 1217 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1218 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1219 1220 if (G1UseAdaptiveConcRefinement) { 1221 const int k_gy = 3, k_gr = 6; 1222 const double inc_k = 1.1, dec_k = 0.9; 1223 1224 int g = cg1r->green_zone(); 1225 if (update_rs_time > goal_ms) { 1226 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1227 } else { 1228 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1229 g = (int)MAX2(g * inc_k, g + 1.0); 1230 } 1231 } 1232 // Change the refinement threads params 1233 cg1r->set_green_zone(g); 1234 cg1r->set_yellow_zone(g * k_gy); 1235 cg1r->set_red_zone(g * k_gr); 1236 cg1r->reinitialize_threads(); 1237 1238 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * _predictor.sigma()), 1); 1239 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1240 cg1r->yellow_zone()); 1241 // Change the barrier params 1242 dcqs.set_process_completed_threshold(processing_threshold); 1243 dcqs.set_max_completed_queue(cg1r->red_zone()); 1244 } 1245 1246 int curr_queue_size = dcqs.completed_buffers_num(); 1247 if (curr_queue_size >= cg1r->yellow_zone()) { 1248 dcqs.set_completed_queue_padding(curr_queue_size); 1249 } else { 1250 dcqs.set_completed_queue_padding(0); 1251 } 1252 dcqs.notify_if_necessary(); 1253 } 1254 1255 size_t G1CollectorPolicy::predict_rs_length_diff() const { 1256 return (size_t) get_new_prediction(_rs_length_diff_seq); 1257 } 1258 1259 double G1CollectorPolicy::predict_alloc_rate_ms() const { 1260 return get_new_prediction(_alloc_rate_ms_seq); 1261 } 1262 1263 double G1CollectorPolicy::predict_cost_per_card_ms() const { 1264 return get_new_prediction(_cost_per_card_ms_seq); 1265 } 1266 1267 double G1CollectorPolicy::predict_scan_hcc_ms() const { 1268 return get_new_prediction(_cost_scan_hcc_seq); 1269 } 1270 1271 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const { 1272 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms(); 1273 } 1274 1275 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const { 1276 return get_new_prediction(_young_cards_per_entry_ratio_seq); 1277 } 1278 1279 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const { 1280 if (_mixed_cards_per_entry_ratio_seq->num() < 2) { 1281 return predict_young_cards_per_entry_ratio(); 1282 } else { 1283 return get_new_prediction(_mixed_cards_per_entry_ratio_seq); 1284 } 1285 } 1286 1287 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const { 1288 return (size_t) (rs_length * predict_young_cards_per_entry_ratio()); 1289 } 1290 1291 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const { 1292 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio()); 1293 } 1294 1295 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const { 1296 if (collector_state()->gcs_are_young()) { 1297 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1298 } else { 1299 return predict_mixed_rs_scan_time_ms(card_num); 1300 } 1301 } 1302 1303 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const { 1304 if (_mixed_cost_per_entry_ms_seq->num() < 3) { 1305 return card_num * get_new_prediction(_cost_per_entry_ms_seq); 1306 } else { 1307 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq); 1308 } 1309 } 1310 1311 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const { 1312 if (_cost_per_byte_ms_during_cm_seq->num() < 3) { 1313 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq); 1314 } else { 1315 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq); 1316 } 1317 } 1318 1319 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const { 1320 if (collector_state()->during_concurrent_mark()) { 1321 return predict_object_copy_time_ms_during_cm(bytes_to_copy); 1322 } else { 1323 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq); 1324 } 1325 } 1326 1327 double G1CollectorPolicy::predict_constant_other_time_ms() const { 1328 return get_new_prediction(_constant_other_time_ms_seq); 1329 } 1330 1331 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const { 1332 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq); 1333 } 1334 1335 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const { 1336 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq); 1337 } 1338 1339 double G1CollectorPolicy::predict_remark_time_ms() const { 1340 return get_new_prediction(_concurrent_mark_remark_times_ms); 1341 } 1342 1343 double G1CollectorPolicy::predict_cleanup_time_ms() const { 1344 return get_new_prediction(_concurrent_mark_cleanup_times_ms); 1345 } 1346 1347 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const { 1348 TruncatedSeq* seq = surv_rate_group->get_seq(age); 1349 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age); 1350 double pred = get_new_prediction(seq); 1351 if (pred > 1.0) { 1352 pred = 1.0; 1353 } 1354 return pred; 1355 } 1356 1357 double G1CollectorPolicy::predict_yg_surv_rate(int age) const { 1358 return predict_yg_surv_rate(age, _short_lived_surv_rate_group); 1359 } 1360 1361 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const { 1362 return _short_lived_surv_rate_group->accum_surv_rate_pred(age); 1363 } 1364 1365 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1366 size_t scanned_cards) const { 1367 return 1368 predict_rs_update_time_ms(pending_cards) + 1369 predict_rs_scan_time_ms(scanned_cards) + 1370 predict_constant_other_time_ms(); 1371 } 1372 1373 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1374 size_t rs_length = predict_rs_length_diff(); 1375 size_t card_num; 1376 if (collector_state()->gcs_are_young()) { 1377 card_num = predict_young_card_num(rs_length); 1378 } else { 1379 card_num = predict_non_young_card_num(rs_length); 1380 } 1381 return predict_base_elapsed_time_ms(pending_cards, card_num); 1382 } 1383 1384 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const { 1385 size_t bytes_to_copy; 1386 if (hr->is_marked()) 1387 bytes_to_copy = hr->max_live_bytes(); 1388 else { 1389 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1390 int age = hr->age_in_surv_rate_group(); 1391 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1392 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate); 1393 } 1394 return bytes_to_copy; 1395 } 1396 1397 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1398 bool for_young_gc) const { 1399 size_t rs_length = hr->rem_set()->occupied(); 1400 size_t card_num; 1401 1402 // Predicting the number of cards is based on which type of GC 1403 // we're predicting for. 1404 if (for_young_gc) { 1405 card_num = predict_young_card_num(rs_length); 1406 } else { 1407 card_num = predict_non_young_card_num(rs_length); 1408 } 1409 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1410 1411 double region_elapsed_time_ms = 1412 predict_rs_scan_time_ms(card_num) + 1413 predict_object_copy_time_ms(bytes_to_copy); 1414 1415 // The prediction of the "other" time for this region is based 1416 // upon the region type and NOT the GC type. 1417 if (hr->is_young()) { 1418 region_elapsed_time_ms += predict_young_other_time_ms(1); 1419 } else { 1420 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1421 } 1422 return region_elapsed_time_ms; 1423 } 1424 1425 void G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1426 uint survivor_cset_region_length) { 1427 _eden_cset_region_length = eden_cset_region_length; 1428 _survivor_cset_region_length = survivor_cset_region_length; 1429 _old_cset_region_length = 0; 1430 } 1431 1432 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1433 _recorded_rs_lengths = rs_lengths; 1434 } 1435 1436 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1437 double elapsed_ms) { 1438 _recent_gc_times_ms->add(elapsed_ms); 1439 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1440 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1441 } 1442 1443 size_t G1CollectorPolicy::expansion_amount() const { 1444 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1445 double threshold = _gc_overhead_perc; 1446 if (recent_gc_overhead > threshold) { 1447 // We will double the existing space, or take 1448 // G1ExpandByPercentOfAvailable % of the available expansion 1449 // space, whichever is smaller, bounded below by a minimum 1450 // expansion (unless that's all that's left.) 1451 const size_t min_expand_bytes = 1*M; 1452 size_t reserved_bytes = _g1->max_capacity(); 1453 size_t committed_bytes = _g1->capacity(); 1454 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1455 size_t expand_bytes; 1456 size_t expand_bytes_via_pct = 1457 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1458 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1459 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1460 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1461 1462 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) " 1463 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B calculated expansion amount: " SIZE_FORMAT "B (" INTX_FORMAT "%%)", 1464 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes_via_pct, G1ExpandByPercentOfAvailable); 1465 1466 return expand_bytes; 1467 } else { 1468 return 0; 1469 } 1470 } 1471 1472 void G1CollectorPolicy::print_tracing_info() const { 1473 _trace_young_gen_time_data.print(); 1474 _trace_old_gen_time_data.print(); 1475 } 1476 1477 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1478 #ifndef PRODUCT 1479 _short_lived_surv_rate_group->print_surv_rate_summary(); 1480 // add this call for any other surv rate groups 1481 #endif // PRODUCT 1482 } 1483 1484 bool G1CollectorPolicy::is_young_list_full() const { 1485 uint young_list_length = _g1->young_list()->length(); 1486 uint young_list_target_length = _young_list_target_length; 1487 return young_list_length >= young_list_target_length; 1488 } 1489 1490 bool G1CollectorPolicy::can_expand_young_list() const { 1491 uint young_list_length = _g1->young_list()->length(); 1492 uint young_list_max_length = _young_list_max_length; 1493 return young_list_length < young_list_max_length; 1494 } 1495 1496 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1497 uint expansion_region_num = 0; 1498 if (GCLockerEdenExpansionPercent > 0) { 1499 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1500 double expansion_region_num_d = perc * (double) _young_list_target_length; 1501 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1502 // less than 1.0) we'll get 1. 1503 expansion_region_num = (uint) ceil(expansion_region_num_d); 1504 } else { 1505 assert(expansion_region_num == 0, "sanity"); 1506 } 1507 _young_list_max_length = _young_list_target_length + expansion_region_num; 1508 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1509 } 1510 1511 // Calculates survivor space parameters. 1512 void G1CollectorPolicy::update_survivors_policy() { 1513 double max_survivor_regions_d = 1514 (double) _young_list_target_length / (double) SurvivorRatio; 1515 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1516 // smaller than 1.0) we'll get 1. 1517 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1518 1519 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1520 HeapRegion::GrainWords * _max_survivor_regions, counters()); 1521 } 1522 1523 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) { 1524 // We actually check whether we are marking here and not if we are in a 1525 // reclamation phase. This means that we will schedule a concurrent mark 1526 // even while we are still in the process of reclaiming memory. 1527 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1528 if (!during_cycle) { 1529 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause)); 1530 collector_state()->set_initiate_conc_mark_if_possible(true); 1531 return true; 1532 } else { 1533 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause)); 1534 return false; 1535 } 1536 } 1537 1538 void G1CollectorPolicy::decide_on_conc_mark_initiation() { 1539 // We are about to decide on whether this pause will be an 1540 // initial-mark pause. 1541 1542 // First, collector_state()->during_initial_mark_pause() should not be already set. We 1543 // will set it here if we have to. However, it should be cleared by 1544 // the end of the pause (it's only set for the duration of an 1545 // initial-mark pause). 1546 assert(!collector_state()->during_initial_mark_pause(), "pre-condition"); 1547 1548 if (collector_state()->initiate_conc_mark_if_possible()) { 1549 // We had noticed on a previous pause that the heap occupancy has 1550 // gone over the initiating threshold and we should start a 1551 // concurrent marking cycle. So we might initiate one. 1552 1553 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) { 1554 // Initiate a new initial mark only if there is no marking or reclamation going 1555 // on. 1556 1557 collector_state()->set_during_initial_mark_pause(true); 1558 // And we can now clear initiate_conc_mark_if_possible() as 1559 // we've already acted on it. 1560 collector_state()->set_initiate_conc_mark_if_possible(false); 1561 1562 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1563 } else { 1564 // The concurrent marking thread is still finishing up the 1565 // previous cycle. If we start one right now the two cycles 1566 // overlap. In particular, the concurrent marking thread might 1567 // be in the process of clearing the next marking bitmap (which 1568 // we will use for the next cycle if we start one). Starting a 1569 // cycle now will be bad given that parts of the marking 1570 // information might get cleared by the marking thread. And we 1571 // cannot wait for the marking thread to finish the cycle as it 1572 // periodically yields while clearing the next marking bitmap 1573 // and, if it's in a yield point, it's waiting for us to 1574 // finish. So, at this point we will not start a cycle and we'll 1575 // let the concurrent marking thread complete the last one. 1576 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1577 } 1578 } 1579 } 1580 1581 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1582 G1CollectedHeap* _g1h; 1583 CSetChooserParUpdater _cset_updater; 1584 1585 public: 1586 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1587 uint chunk_size) : 1588 _g1h(G1CollectedHeap::heap()), 1589 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } 1590 1591 bool doHeapRegion(HeapRegion* r) { 1592 // Do we have any marking information for this region? 1593 if (r->is_marked()) { 1594 // We will skip any region that's currently used as an old GC 1595 // alloc region (we should not consider those for collection 1596 // before we fill them up). 1597 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1598 _cset_updater.add_region(r); 1599 } 1600 } 1601 return false; 1602 } 1603 }; 1604 1605 class ParKnownGarbageTask: public AbstractGangTask { 1606 CollectionSetChooser* _hrSorted; 1607 uint _chunk_size; 1608 G1CollectedHeap* _g1; 1609 HeapRegionClaimer _hrclaimer; 1610 1611 public: 1612 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : 1613 AbstractGangTask("ParKnownGarbageTask"), 1614 _hrSorted(hrSorted), _chunk_size(chunk_size), 1615 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} 1616 1617 void work(uint worker_id) { 1618 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1619 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer); 1620 } 1621 }; 1622 1623 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { 1624 assert(n_workers > 0, "Active gc workers should be greater than 0"); 1625 const uint overpartition_factor = 4; 1626 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); 1627 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); 1628 } 1629 1630 void 1631 G1CollectorPolicy::record_concurrent_mark_cleanup_end() { 1632 _collectionSetChooser->clear(); 1633 1634 WorkGang* workers = _g1->workers(); 1635 uint n_workers = workers->active_workers(); 1636 1637 uint n_regions = _g1->num_regions(); 1638 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); 1639 _collectionSetChooser->prepare_for_par_region_addition(n_workers, n_regions, chunk_size); 1640 ParKnownGarbageTask par_known_garbage_task(_collectionSetChooser, chunk_size, n_workers); 1641 workers->run_task(&par_known_garbage_task); 1642 1643 _collectionSetChooser->sort_regions(); 1644 1645 double end_sec = os::elapsedTime(); 1646 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1647 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 1648 _cur_mark_stop_world_time_ms += elapsed_time_ms; 1649 _prev_collection_pause_end_ms += elapsed_time_ms; 1650 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec); 1651 } 1652 1653 // Add the heap region at the head of the non-incremental collection set 1654 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 1655 assert(_inc_cset_build_state == Active, "Precondition"); 1656 assert(hr->is_old(), "the region should be old"); 1657 1658 assert(!hr->in_collection_set(), "should not already be in the CSet"); 1659 _g1->register_old_region_with_cset(hr); 1660 hr->set_next_in_collection_set(_collection_set); 1661 _collection_set = hr; 1662 _collection_set_bytes_used_before += hr->used(); 1663 size_t rs_length = hr->rem_set()->occupied(); 1664 _recorded_rs_lengths += rs_length; 1665 _old_cset_region_length += 1; 1666 } 1667 1668 // Initialize the per-collection-set information 1669 void G1CollectorPolicy::start_incremental_cset_building() { 1670 assert(_inc_cset_build_state == Inactive, "Precondition"); 1671 1672 _inc_cset_head = NULL; 1673 _inc_cset_tail = NULL; 1674 _inc_cset_bytes_used_before = 0; 1675 1676 _inc_cset_max_finger = 0; 1677 _inc_cset_recorded_rs_lengths = 0; 1678 _inc_cset_recorded_rs_lengths_diffs = 0; 1679 _inc_cset_predicted_elapsed_time_ms = 0.0; 1680 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1681 _inc_cset_build_state = Active; 1682 } 1683 1684 void G1CollectorPolicy::finalize_incremental_cset_building() { 1685 assert(_inc_cset_build_state == Active, "Precondition"); 1686 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1687 1688 // The two "main" fields, _inc_cset_recorded_rs_lengths and 1689 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 1690 // that adds a new region to the CSet. Further updates by the 1691 // concurrent refinement thread that samples the young RSet lengths 1692 // are accumulated in the *_diffs fields. Here we add the diffs to 1693 // the "main" fields. 1694 1695 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 1696 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 1697 } else { 1698 // This is defensive. The diff should in theory be always positive 1699 // as RSets can only grow between GCs. However, given that we 1700 // sample their size concurrently with other threads updating them 1701 // it's possible that we might get the wrong size back, which 1702 // could make the calculations somewhat inaccurate. 1703 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 1704 if (_inc_cset_recorded_rs_lengths >= diffs) { 1705 _inc_cset_recorded_rs_lengths -= diffs; 1706 } else { 1707 _inc_cset_recorded_rs_lengths = 0; 1708 } 1709 } 1710 _inc_cset_predicted_elapsed_time_ms += 1711 _inc_cset_predicted_elapsed_time_ms_diffs; 1712 1713 _inc_cset_recorded_rs_lengths_diffs = 0; 1714 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 1715 } 1716 1717 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 1718 // This routine is used when: 1719 // * adding survivor regions to the incremental cset at the end of an 1720 // evacuation pause, 1721 // * adding the current allocation region to the incremental cset 1722 // when it is retired, and 1723 // * updating existing policy information for a region in the 1724 // incremental cset via young list RSet sampling. 1725 // Therefore this routine may be called at a safepoint by the 1726 // VM thread, or in-between safepoints by mutator threads (when 1727 // retiring the current allocation region) or a concurrent 1728 // refine thread (RSet sampling). 1729 1730 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1731 size_t used_bytes = hr->used(); 1732 _inc_cset_recorded_rs_lengths += rs_length; 1733 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 1734 _inc_cset_bytes_used_before += used_bytes; 1735 1736 // Cache the values we have added to the aggregated information 1737 // in the heap region in case we have to remove this region from 1738 // the incremental collection set, or it is updated by the 1739 // rset sampling code 1740 hr->set_recorded_rs_length(rs_length); 1741 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 1742 } 1743 1744 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 1745 size_t new_rs_length) { 1746 // Update the CSet information that is dependent on the new RS length 1747 assert(hr->is_young(), "Precondition"); 1748 assert(!SafepointSynchronize::is_at_safepoint(), 1749 "should not be at a safepoint"); 1750 1751 // We could have updated _inc_cset_recorded_rs_lengths and 1752 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 1753 // that atomically, as this code is executed by a concurrent 1754 // refinement thread, potentially concurrently with a mutator thread 1755 // allocating a new region and also updating the same fields. To 1756 // avoid the atomic operations we accumulate these updates on two 1757 // separate fields (*_diffs) and we'll just add them to the "main" 1758 // fields at the start of a GC. 1759 1760 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 1761 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 1762 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 1763 1764 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 1765 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 1766 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 1767 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 1768 1769 hr->set_recorded_rs_length(new_rs_length); 1770 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 1771 } 1772 1773 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 1774 assert(hr->is_young(), "invariant"); 1775 assert(hr->young_index_in_cset() > -1, "should have already been set"); 1776 assert(_inc_cset_build_state == Active, "Precondition"); 1777 1778 // We need to clear and set the cached recorded/cached collection set 1779 // information in the heap region here (before the region gets added 1780 // to the collection set). An individual heap region's cached values 1781 // are calculated, aggregated with the policy collection set info, 1782 // and cached in the heap region here (initially) and (subsequently) 1783 // by the Young List sampling code. 1784 1785 size_t rs_length = hr->rem_set()->occupied(); 1786 add_to_incremental_cset_info(hr, rs_length); 1787 1788 HeapWord* hr_end = hr->end(); 1789 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 1790 1791 assert(!hr->in_collection_set(), "invariant"); 1792 _g1->register_young_region_with_cset(hr); 1793 assert(hr->next_in_collection_set() == NULL, "invariant"); 1794 } 1795 1796 // Add the region at the RHS of the incremental cset 1797 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 1798 // We should only ever be appending survivors at the end of a pause 1799 assert(hr->is_survivor(), "Logic"); 1800 1801 // Do the 'common' stuff 1802 add_region_to_incremental_cset_common(hr); 1803 1804 // Now add the region at the right hand side 1805 if (_inc_cset_tail == NULL) { 1806 assert(_inc_cset_head == NULL, "invariant"); 1807 _inc_cset_head = hr; 1808 } else { 1809 _inc_cset_tail->set_next_in_collection_set(hr); 1810 } 1811 _inc_cset_tail = hr; 1812 } 1813 1814 // Add the region to the LHS of the incremental cset 1815 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 1816 // Survivors should be added to the RHS at the end of a pause 1817 assert(hr->is_eden(), "Logic"); 1818 1819 // Do the 'common' stuff 1820 add_region_to_incremental_cset_common(hr); 1821 1822 // Add the region at the left hand side 1823 hr->set_next_in_collection_set(_inc_cset_head); 1824 if (_inc_cset_head == NULL) { 1825 assert(_inc_cset_tail == NULL, "Invariant"); 1826 _inc_cset_tail = hr; 1827 } 1828 _inc_cset_head = hr; 1829 } 1830 1831 #ifndef PRODUCT 1832 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 1833 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 1834 1835 st->print_cr("\nCollection_set:"); 1836 HeapRegion* csr = list_head; 1837 while (csr != NULL) { 1838 HeapRegion* next = csr->next_in_collection_set(); 1839 assert(csr->in_collection_set(), "bad CS"); 1840 st->print_cr(" " HR_FORMAT ", P: " PTR_FORMAT "N: " PTR_FORMAT ", age: %4d", 1841 HR_FORMAT_PARAMS(csr), 1842 p2i(csr->prev_top_at_mark_start()), p2i(csr->next_top_at_mark_start()), 1843 csr->age_in_surv_rate_group_cond()); 1844 csr = next; 1845 } 1846 } 1847 #endif // !PRODUCT 1848 1849 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const { 1850 // Returns the given amount of reclaimable bytes (that represents 1851 // the amount of reclaimable space still to be collected) as a 1852 // percentage of the current heap capacity. 1853 size_t capacity_bytes = _g1->capacity(); 1854 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 1855 } 1856 1857 void G1CollectorPolicy::maybe_start_marking() { 1858 if (need_to_start_conc_mark("end of GC")) { 1859 // Note: this might have already been set, if during the last 1860 // pause we decided to start a cycle but at the beginning of 1861 // this pause we decided to postpone it. That's OK. 1862 collector_state()->set_initiate_conc_mark_if_possible(true); 1863 } 1864 } 1865 1866 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 1867 const char* false_action_str) const { 1868 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1869 if (cset_chooser->is_empty()) { 1870 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 1871 return false; 1872 } 1873 1874 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1875 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 1876 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 1877 double threshold = (double) G1HeapWastePercent; 1878 if (reclaimable_perc <= threshold) { 1879 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1880 false_action_str, cset_chooser->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 1881 return false; 1882 } 1883 1884 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1885 true_action_str, cset_chooser->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 1886 return true; 1887 } 1888 1889 uint G1CollectorPolicy::calc_min_old_cset_length() const { 1890 // The min old CSet region bound is based on the maximum desired 1891 // number of mixed GCs after a cycle. I.e., even if some old regions 1892 // look expensive, we should add them to the CSet anyway to make 1893 // sure we go through the available old regions in no more than the 1894 // maximum desired number of mixed GCs. 1895 // 1896 // The calculation is based on the number of marked regions we added 1897 // to the CSet chooser in the first place, not how many remain, so 1898 // that the result is the same during all mixed GCs that follow a cycle. 1899 1900 const size_t region_num = (size_t) _collectionSetChooser->length(); 1901 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 1902 size_t result = region_num / gc_num; 1903 // emulate ceiling 1904 if (result * gc_num < region_num) { 1905 result += 1; 1906 } 1907 return (uint) result; 1908 } 1909 1910 uint G1CollectorPolicy::calc_max_old_cset_length() const { 1911 // The max old CSet region bound is based on the threshold expressed 1912 // as a percentage of the heap size. I.e., it should bound the 1913 // number of old regions added to the CSet irrespective of how many 1914 // of them are available. 1915 1916 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1917 const size_t region_num = g1h->num_regions(); 1918 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 1919 size_t result = region_num * perc / 100; 1920 // emulate ceiling 1921 if (100 * result < region_num * perc) { 1922 result += 1; 1923 } 1924 return (uint) result; 1925 } 1926 1927 1928 double G1CollectorPolicy::finalize_young_cset_part(double target_pause_time_ms) { 1929 double young_start_time_sec = os::elapsedTime(); 1930 1931 YoungList* young_list = _g1->young_list(); 1932 finalize_incremental_cset_building(); 1933 1934 guarantee(target_pause_time_ms > 0.0, 1935 "target_pause_time_ms = %1.6lf should be positive", target_pause_time_ms); 1936 guarantee(_collection_set == NULL, "Precondition"); 1937 1938 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 1939 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); 1940 1941 log_trace(gc, ergo, cset)("Start choosing CSet. pending cards: " SIZE_FORMAT " predicted base time: %1.2fms remaining time: %1.2fms target pause time: %1.2fms", 1942 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); 1943 1944 collector_state()->set_last_gc_was_young(collector_state()->gcs_are_young()); 1945 1946 if (collector_state()->last_gc_was_young()) { 1947 _trace_young_gen_time_data.increment_young_collection_count(); 1948 } else { 1949 _trace_young_gen_time_data.increment_mixed_collection_count(); 1950 } 1951 1952 // The young list is laid with the survivor regions from the previous 1953 // pause are appended to the RHS of the young list, i.e. 1954 // [Newly Young Regions ++ Survivors from last pause]. 1955 1956 uint survivor_region_length = young_list->survivor_length(); 1957 uint eden_region_length = young_list->eden_length(); 1958 init_cset_region_lengths(eden_region_length, survivor_region_length); 1959 1960 HeapRegion* hr = young_list->first_survivor_region(); 1961 while (hr != NULL) { 1962 assert(hr->is_survivor(), "badly formed young list"); 1963 // There is a convention that all the young regions in the CSet 1964 // are tagged as "eden", so we do this for the survivors here. We 1965 // use the special set_eden_pre_gc() as it doesn't check that the 1966 // region is free (which is not the case here). 1967 hr->set_eden_pre_gc(); 1968 hr = hr->get_next_young_region(); 1969 } 1970 1971 // Clear the fields that point to the survivor list - they are all young now. 1972 young_list->clear_survivors(); 1973 1974 _collection_set = _inc_cset_head; 1975 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 1976 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); 1977 1978 log_trace(gc, ergo, cset)("Add young regions to CSet. eden: %u regions, survivors: %u regions, predicted young region time: %1.2fms, target pause time: %1.2fms", 1979 eden_region_length, survivor_region_length, _inc_cset_predicted_elapsed_time_ms, target_pause_time_ms); 1980 1981 // The number of recorded young regions is the incremental 1982 // collection set's current size 1983 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 1984 1985 double young_end_time_sec = os::elapsedTime(); 1986 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); 1987 1988 return time_remaining_ms; 1989 } 1990 1991 void G1CollectorPolicy::finalize_old_cset_part(double time_remaining_ms) { 1992 double non_young_start_time_sec = os::elapsedTime(); 1993 double predicted_old_time_ms = 0.0; 1994 1995 1996 if (!collector_state()->gcs_are_young()) { 1997 CollectionSetChooser* cset_chooser = _collectionSetChooser; 1998 cset_chooser->verify(); 1999 const uint min_old_cset_length = calc_min_old_cset_length(); 2000 const uint max_old_cset_length = calc_max_old_cset_length(); 2001 2002 uint expensive_region_num = 0; 2003 bool check_time_remaining = adaptive_young_list_length(); 2004 2005 HeapRegion* hr = cset_chooser->peek(); 2006 while (hr != NULL) { 2007 if (old_cset_region_length() >= max_old_cset_length) { 2008 // Added maximum number of old regions to the CSet. 2009 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached max). old %u regions, max %u regions", 2010 old_cset_region_length(), max_old_cset_length); 2011 break; 2012 } 2013 2014 2015 // Stop adding regions if the remaining reclaimable space is 2016 // not above G1HeapWastePercent. 2017 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2018 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); 2019 double threshold = (double) G1HeapWastePercent; 2020 if (reclaimable_perc <= threshold) { 2021 // We've added enough old regions that the amount of uncollected 2022 // reclaimable space is at or below the waste threshold. Stop 2023 // adding old regions to the CSet. 2024 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (reclaimable percentage not over threshold). " 2025 "old %u regions, max %u regions, reclaimable: " SIZE_FORMAT "B (%1.2f%%) threshold: " UINTX_FORMAT "%%", 2026 old_cset_region_length(), max_old_cset_length, reclaimable_bytes, reclaimable_perc, G1HeapWastePercent); 2027 break; 2028 } 2029 2030 double predicted_time_ms = predict_region_elapsed_time_ms(hr, collector_state()->gcs_are_young()); 2031 if (check_time_remaining) { 2032 if (predicted_time_ms > time_remaining_ms) { 2033 // Too expensive for the current CSet. 2034 2035 if (old_cset_region_length() >= min_old_cset_length) { 2036 // We have added the minimum number of old regions to the CSet, 2037 // we are done with this CSet. 2038 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (predicted time is too high). " 2039 "predicted time: %1.2fms, remaining time: %1.2fms old %u regions, min %u regions", 2040 predicted_time_ms, time_remaining_ms, old_cset_region_length(), min_old_cset_length); 2041 break; 2042 } 2043 2044 // We'll add it anyway given that we haven't reached the 2045 // minimum number of old regions. 2046 expensive_region_num += 1; 2047 } 2048 } else { 2049 if (old_cset_region_length() >= min_old_cset_length) { 2050 // In the non-auto-tuning case, we'll finish adding regions 2051 // to the CSet if we reach the minimum. 2052 2053 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (old CSet region num reached min). old %u regions, min %u regions", 2054 old_cset_region_length(), min_old_cset_length); 2055 break; 2056 } 2057 } 2058 2059 // We will add this region to the CSet. 2060 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 2061 predicted_old_time_ms += predicted_time_ms; 2062 cset_chooser->pop(); // already have region via peek() 2063 _g1->old_set_remove(hr); 2064 add_old_region_to_cset(hr); 2065 2066 hr = cset_chooser->peek(); 2067 } 2068 if (hr == NULL) { 2069 log_debug(gc, ergo, cset)("Finish adding old regions to CSet (candidate old regions not available)"); 2070 } 2071 2072 if (expensive_region_num > 0) { 2073 // We print the information once here at the end, predicated on 2074 // whether we added any apparently expensive regions or not, to 2075 // avoid generating output per region. 2076 log_debug(gc, ergo, cset)("Added expensive regions to CSet (old CSet region num not reached min)." 2077 "old %u regions, expensive: %u regions, min %u regions, remaining time: %1.2fms", 2078 old_cset_region_length(), expensive_region_num, min_old_cset_length, time_remaining_ms); 2079 } 2080 2081 cset_chooser->verify(); 2082 } 2083 2084 stop_incremental_cset_building(); 2085 2086 log_debug(gc, ergo, cset)("Finish choosing CSet. old %u regions, predicted old region time: %1.2fms, time remaining: %1.2f", 2087 old_cset_region_length(), predicted_old_time_ms, time_remaining_ms); 2088 2089 double non_young_end_time_sec = os::elapsedTime(); 2090 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); 2091 } 2092 2093 void TraceYoungGenTimeData::record_start_collection(double time_to_stop_the_world_ms) { 2094 if(TraceYoungGenTime) { 2095 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2096 } 2097 } 2098 2099 void TraceYoungGenTimeData::record_yield_time(double yield_time_ms) { 2100 if(TraceYoungGenTime) { 2101 _all_yield_times_ms.add(yield_time_ms); 2102 } 2103 } 2104 2105 void TraceYoungGenTimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { 2106 if(TraceYoungGenTime) { 2107 _total.add(pause_time_ms); 2108 _other.add(pause_time_ms - phase_times->accounted_time_ms()); 2109 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); 2110 _parallel.add(phase_times->cur_collection_par_time_ms()); 2111 _ext_root_scan.add(phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan)); 2112 _satb_filtering.add(phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering)); 2113 _update_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS)); 2114 _scan_rs.add(phase_times->average_time_ms(G1GCPhaseTimes::ScanRS)); 2115 _obj_copy.add(phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy)); 2116 _termination.add(phase_times->average_time_ms(G1GCPhaseTimes::Termination)); 2117 2118 double parallel_known_time = phase_times->average_time_ms(G1GCPhaseTimes::ExtRootScan) + 2119 phase_times->average_time_ms(G1GCPhaseTimes::SATBFiltering) + 2120 phase_times->average_time_ms(G1GCPhaseTimes::UpdateRS) + 2121 phase_times->average_time_ms(G1GCPhaseTimes::ScanRS) + 2122 phase_times->average_time_ms(G1GCPhaseTimes::ObjCopy) + 2123 phase_times->average_time_ms(G1GCPhaseTimes::Termination); 2124 2125 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; 2126 _parallel_other.add(parallel_other_time); 2127 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); 2128 } 2129 } 2130 2131 void TraceYoungGenTimeData::increment_young_collection_count() { 2132 if(TraceYoungGenTime) { 2133 ++_young_pause_num; 2134 } 2135 } 2136 2137 void TraceYoungGenTimeData::increment_mixed_collection_count() { 2138 if(TraceYoungGenTime) { 2139 ++_mixed_pause_num; 2140 } 2141 } 2142 2143 void TraceYoungGenTimeData::print_summary(const char* str, 2144 const NumberSeq* seq) const { 2145 double sum = seq->sum(); 2146 tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", 2147 str, sum / 1000.0, seq->avg()); 2148 } 2149 2150 void TraceYoungGenTimeData::print_summary_sd(const char* str, 2151 const NumberSeq* seq) const { 2152 print_summary(str, seq); 2153 tty->print_cr("%45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2154 "(num", seq->num(), seq->sd(), seq->maximum()); 2155 } 2156 2157 void TraceYoungGenTimeData::print() const { 2158 if (!TraceYoungGenTime) { 2159 return; 2160 } 2161 2162 tty->print_cr("ALL PAUSES"); 2163 print_summary_sd(" Total", &_total); 2164 tty->cr(); 2165 tty->cr(); 2166 tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2167 tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2168 tty->cr(); 2169 2170 tty->print_cr("EVACUATION PAUSES"); 2171 2172 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2173 tty->print_cr("none"); 2174 } else { 2175 print_summary_sd(" Evacuation Pauses", &_total); 2176 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); 2177 print_summary(" Parallel Time", &_parallel); 2178 print_summary(" Ext Root Scanning", &_ext_root_scan); 2179 print_summary(" SATB Filtering", &_satb_filtering); 2180 print_summary(" Update RS", &_update_rs); 2181 print_summary(" Scan RS", &_scan_rs); 2182 print_summary(" Object Copy", &_obj_copy); 2183 print_summary(" Termination", &_termination); 2184 print_summary(" Parallel Other", &_parallel_other); 2185 print_summary(" Clear CT", &_clear_ct); 2186 print_summary(" Other", &_other); 2187 } 2188 tty->cr(); 2189 2190 tty->print_cr("MISC"); 2191 print_summary_sd(" Stop World", &_all_stop_world_times_ms); 2192 print_summary_sd(" Yields", &_all_yield_times_ms); 2193 } 2194 2195 void TraceOldGenTimeData::record_full_collection(double full_gc_time_ms) { 2196 if (TraceOldGenTime) { 2197 _all_full_gc_times.add(full_gc_time_ms); 2198 } 2199 } 2200 2201 void TraceOldGenTimeData::print() const { 2202 if (!TraceOldGenTime) { 2203 return; 2204 } 2205 2206 if (_all_full_gc_times.num() > 0) { 2207 tty->print("\n%4d full_gcs: total time = %8.2f s", 2208 _all_full_gc_times.num(), 2209 _all_full_gc_times.sum() / 1000.0); 2210 tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2211 tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2212 _all_full_gc_times.sd(), 2213 _all_full_gc_times.maximum()); 2214 } 2215 }