1 /* 2 * Copyright (c) 2001, 2020, 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/g1Analytics.hpp" 27 #include "gc/g1/g1Arguments.hpp" 28 #include "gc/g1/g1CollectedHeap.inline.hpp" 29 #include "gc/g1/g1CollectionSet.hpp" 30 #include "gc/g1/g1CollectionSetCandidates.hpp" 31 #include "gc/g1/g1ConcurrentMark.hpp" 32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp" 33 #include "gc/g1/g1ConcurrentRefine.hpp" 34 #include "gc/g1/g1ConcurrentRefineStats.hpp" 35 #include "gc/g1/g1CollectionSetChooser.hpp" 36 #include "gc/g1/g1HeterogeneousHeapPolicy.hpp" 37 #include "gc/g1/g1HotCardCache.hpp" 38 #include "gc/g1/g1IHOPControl.hpp" 39 #include "gc/g1/g1GCPhaseTimes.hpp" 40 #include "gc/g1/g1Policy.hpp" 41 #include "gc/g1/g1SurvivorRegions.hpp" 42 #include "gc/g1/g1YoungGenSizer.hpp" 43 #include "gc/g1/heapRegion.inline.hpp" 44 #include "gc/g1/heapRegionRemSet.hpp" 45 #include "gc/shared/concurrentGCBreakpoints.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "logging/log.hpp" 48 #include "runtime/arguments.hpp" 49 #include "runtime/globals.hpp" 50 #include "runtime/java.hpp" 51 #include "runtime/mutexLocker.hpp" 52 #include "utilities/debug.hpp" 53 #include "utilities/growableArray.hpp" 54 #include "utilities/pair.hpp" 55 56 G1Policy::G1Policy(STWGCTimer* gc_timer) : 57 _predictor(G1ConfidencePercent / 100.0), 58 _analytics(new G1Analytics(&_predictor)), 59 _remset_tracker(), 60 _mmu_tracker(new G1MMUTrackerQueue(GCPauseIntervalMillis / 1000.0, MaxGCPauseMillis / 1000.0)), 61 _ihop_control(create_ihop_control(&_predictor)), 62 _policy_counters(new GCPolicyCounters("GarbageFirst", 1, 2)), 63 _full_collection_start_sec(0.0), 64 _young_list_desired_length(0), 65 _young_list_target_length(0), 66 _young_list_max_length(0), 67 _eden_surv_rate_group(new G1SurvRateGroup()), 68 _survivor_surv_rate_group(new G1SurvRateGroup()), 69 _reserve_factor((double) G1ReservePercent / 100.0), 70 _reserve_regions(0), 71 _young_gen_sizer(G1YoungGenSizer::create_gen_sizer()), 72 _free_regions_at_end_of_collection(0), 73 _rs_length(0), 74 _rs_length_prediction(0), 75 _pending_cards_at_gc_start(0), 76 _old_gen_alloc_tracker(), 77 _concurrent_start_to_mixed(), 78 _collection_set(NULL), 79 _g1h(NULL), 80 _phase_times_timer(gc_timer), 81 _phase_times(NULL), 82 _mark_remark_start_sec(0), 83 _mark_cleanup_start_sec(0), 84 _tenuring_threshold(MaxTenuringThreshold), 85 _max_survivor_regions(0), 86 _survivors_age_table(true) 87 { 88 } 89 90 G1Policy::~G1Policy() { 91 delete _ihop_control; 92 delete _young_gen_sizer; 93 } 94 95 G1Policy* G1Policy::create_policy(STWGCTimer* gc_timer_stw) { 96 if (G1Arguments::is_heterogeneous_heap()) { 97 return new G1HeterogeneousHeapPolicy(gc_timer_stw); 98 } else { 99 return new G1Policy(gc_timer_stw); 100 } 101 } 102 103 G1CollectorState* G1Policy::collector_state() const { return _g1h->collector_state(); } 104 105 void G1Policy::init(G1CollectedHeap* g1h, G1CollectionSet* collection_set) { 106 _g1h = g1h; 107 _collection_set = collection_set; 108 109 assert(Heap_lock->owned_by_self(), "Locking discipline."); 110 111 _young_gen_sizer->adjust_max_new_size(_g1h->max_expandable_regions()); 112 113 _free_regions_at_end_of_collection = _g1h->num_free_regions(); 114 115 update_young_length_bounds(); 116 // We may immediately start allocating regions and placing them on the 117 // collection set list. Initialize the per-collection set info 118 _collection_set->start_incremental_building(); 119 } 120 121 void G1Policy::note_gc_start() { 122 phase_times()->note_gc_start(); 123 } 124 125 class G1YoungLengthPredictor { 126 const double _base_time_ms; 127 const double _base_free_regions; 128 const double _target_pause_time_ms; 129 const G1Policy* const _policy; 130 131 public: 132 G1YoungLengthPredictor(double base_time_ms, 133 double base_free_regions, 134 double target_pause_time_ms, 135 const G1Policy* policy) : 136 _base_time_ms(base_time_ms), 137 _base_free_regions(base_free_regions), 138 _target_pause_time_ms(target_pause_time_ms), 139 _policy(policy) {} 140 141 bool will_fit(uint young_length) const { 142 if (young_length >= _base_free_regions) { 143 // end condition 1: not enough space for the young regions 144 return false; 145 } 146 147 size_t bytes_to_copy = 0; 148 const double copy_time_ms = _policy->predict_eden_copy_time_ms(young_length, &bytes_to_copy); 149 const double young_other_time_ms = _policy->analytics()->predict_young_other_time_ms(young_length); 150 const double pause_time_ms = _base_time_ms + copy_time_ms + young_other_time_ms; 151 if (pause_time_ms > _target_pause_time_ms) { 152 // end condition 2: prediction is over the target pause time 153 return false; 154 } 155 156 const size_t free_bytes = (_base_free_regions - young_length) * HeapRegion::GrainBytes; 157 158 // When copying, we will likely need more bytes free than is live in the region. 159 // Add some safety margin to factor in the confidence of our guess, and the 160 // natural expected waste. 161 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty 162 // of the calculation: the lower the confidence, the more headroom. 163 // (100 + TargetPLABWastePct) represents the increase in expected bytes during 164 // copying due to anticipated waste in the PLABs. 165 const double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0; 166 const size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy); 167 168 if (expected_bytes_to_copy > free_bytes) { 169 // end condition 3: out-of-space 170 return false; 171 } 172 173 // success! 174 return true; 175 } 176 }; 177 178 void G1Policy::record_new_heap_size(uint new_number_of_regions) { 179 // re-calculate the necessary reserve 180 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 181 // We use ceiling so that if reserve_regions_d is > 0.0 (but 182 // smaller than 1.0) we'll get 1. 183 _reserve_regions = (uint) ceil(reserve_regions_d); 184 185 _young_gen_sizer->heap_size_changed(new_number_of_regions); 186 187 _ihop_control->update_target_occupancy(new_number_of_regions * HeapRegion::GrainBytes); 188 } 189 190 uint G1Policy::calculate_desired_eden_length_by_mmu() const { 191 // One could argue that any useful eden length to keep any MMU would be 1, but 192 // in theory this is possible. Other constraints enforce a minimum eden of 1 193 // anyway. 194 uint desired_min_length = 0; 195 if (use_adaptive_young_list_length()) { 196 double now_sec = os::elapsedTime(); 197 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 198 double alloc_rate_ms = _analytics->predict_alloc_rate_ms(); 199 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 200 } 201 return desired_min_length; 202 } 203 204 void G1Policy::update_young_length_bounds() { 205 update_young_length_bounds(_analytics->predict_rs_length()); 206 } 207 208 void G1Policy::update_young_length_bounds(size_t rs_length) { 209 _young_list_desired_length = calculate_young_desired_length(rs_length); 210 _young_list_target_length = calculate_young_target_length(_young_list_desired_length); 211 _young_list_max_length = calculate_young_max_length(_young_list_target_length); 212 213 log_debug(gc,ergo,heap)("Young list lengths: desired: %u, target: %u, max: %u", 214 _young_list_desired_length, 215 _young_list_target_length, 216 _young_list_max_length); 217 } 218 219 // Calculates desired young gen length. It is calculated from: 220 // 221 // - sizer min/max bounds on young gen 222 // - pause time goal for whole young gen evacuation 223 // - MMU goal influencing eden to make GCs spaced apart. 224 // - a minimum one eden region length. 225 // 226 // We may enter with already allocated eden and survivor regions, that may be 227 // higher than the maximum, or the above goals may result in a desired value 228 // smaller than are already allocated. 229 // The main reason is revising young length, with our without the GCLocker being 230 // active. 231 // 232 uint G1Policy::calculate_young_desired_length(size_t rs_length) const { 233 uint min_young_length_by_sizer = _young_gen_sizer->min_desired_young_length(); 234 uint max_young_length_by_sizer = _young_gen_sizer->max_desired_young_length(); 235 236 assert(min_young_length_by_sizer >= 1, "invariant"); 237 assert(max_young_length_by_sizer >= min_young_length_by_sizer, "invariant"); 238 239 // Absolute minimum eden length. 240 // Enforcing a minimum eden length helps at startup when the predictors are not 241 // yet trained on the application to avoid unnecessary (but very short) full gcs 242 // on very small (initial) heaps. 243 uint const MinDesiredEdenLength = 1; 244 245 // Calculate the absolute and desired min bounds first. 246 247 // This is how many survivor regions we already have. 248 const uint survivor_length = _g1h->survivor_regions_count(); 249 // Size of the already allocated young gen. 250 const uint allocated_young_length = _g1h->young_regions_count(); 251 // This is the absolute minimum young length that we can return. Ensure that we 252 // don't go below any user-defined minimum bound; but we might have already 253 // allocated more than that for reasons. In this case, use that. 254 uint absolute_min_young_length = MAX2(allocated_young_length, min_young_length_by_sizer); 255 // Calculate the absolute max bounds. After evac failure or when revising the 256 // young length we might have exceeded absolute min length or absolute_max_length, 257 // so adjust the result accordingly. 258 uint absolute_max_young_length = MAX2(max_young_length_by_sizer, absolute_min_young_length); 259 260 uint desired_eden_length_by_mmu = 0; 261 uint desired_eden_length_by_pause = 0; 262 uint desired_eden_length_before_mixed = 0; 263 264 uint desired_young_length = 0; 265 if (use_adaptive_young_list_length()) { 266 desired_eden_length_by_mmu = calculate_desired_eden_length_by_mmu(); 267 268 const size_t pending_cards = _analytics->predict_pending_cards(); 269 double survivor_base_time_ms = predict_base_elapsed_time_ms(pending_cards, rs_length); 270 271 if (!next_gc_should_be_mixed(NULL, NULL)) { 272 desired_eden_length_by_pause = 273 calculate_desired_eden_length_by_pause(survivor_base_time_ms, 274 absolute_min_young_length - survivor_length, 275 absolute_max_young_length - survivor_length); 276 } else { 277 desired_eden_length_before_mixed = 278 calculate_desired_eden_length_before_mixed(survivor_base_time_ms, 279 absolute_min_young_length - survivor_length, 280 absolute_max_young_length - survivor_length); 281 } 282 // Above either sets desired_eden_length_by_pause or desired_eden_length_before_mixed, 283 // the other is zero. Use the one that has been set below. 284 uint desired_eden_length = MAX2(desired_eden_length_by_pause, 285 desired_eden_length_before_mixed); 286 287 // Finally incorporate MMU concerns; assume that it overrides the pause time 288 // goal, as the default value has been chosen to effectively disable it. 289 // Also request at least one eden region, see above for reasons. 290 desired_eden_length = MAX3(desired_eden_length, 291 desired_eden_length_by_mmu, 292 MinDesiredEdenLength); 293 294 desired_young_length = desired_eden_length + survivor_length; 295 } else { 296 // The user asked for a fixed young gen so we'll fix the young gen 297 // whether the next GC is young or mixed. 298 desired_young_length = min_young_length_by_sizer; 299 } 300 // Clamp to absolute min/max after we determined desired lengths. 301 desired_young_length = clamp(desired_young_length, absolute_min_young_length, absolute_max_young_length); 302 303 log_trace(gc, ergo, heap)("Young desired length %u " 304 "survivor length %u " 305 "allocated young length %u " 306 "absolute min young length %u " 307 "absolute max young length %u " 308 "desired eden length by mmu %u " 309 "desired eden length by pause %u " 310 "desired eden length before mixed %u" 311 "desired eden length by default %u", 312 desired_young_length, survivor_length, 313 allocated_young_length, absolute_min_young_length, 314 absolute_max_young_length, desired_eden_length_by_mmu, 315 desired_eden_length_by_pause, 316 desired_eden_length_before_mixed, 317 MinDesiredEdenLength); 318 319 assert(desired_young_length >= allocated_young_length, "must be"); 320 return desired_young_length; 321 } 322 323 // Limit the desired (wished) young length by current free regions. If the request 324 // can be satisfied without using up reserve regions, do so, otherwise eat into 325 // the reserve, giving away at most what the heap sizer allows. 326 uint G1Policy::calculate_young_target_length(uint desired_young_length) const { 327 uint allocated_young_length = _g1h->young_regions_count(); 328 329 uint receiving_additional_eden; 330 if (allocated_young_length >= desired_young_length) { 331 // Already used up all we actually want (may happen as G1 revises the 332 // young list length concurrently, or caused by gclocker). Do not allow more, 333 // potentially resulting in GC. 334 receiving_additional_eden = 0; 335 log_trace(gc, ergo, heap)("Young target length: Already used up desired young %u allocated %u", 336 desired_young_length, 337 allocated_young_length); 338 } else { 339 // Now look at how many free regions are there currently, and the heap reserve. 340 // We will try our best not to "eat" into the reserve as long as we can. If we 341 // do, we at most eat the sizer's minimum regions into the reserve or half the 342 // reserve rounded up (if possible; this is an arbitrary value). 343 344 uint max_to_eat_into_reserve = MIN2(_young_gen_sizer->min_desired_young_length(), 345 (_reserve_regions + 1) / 2); 346 347 log_trace(gc, ergo, heap)("Young target length: Common " 348 "free regions at end of collection %u " 349 "desired young length %u " 350 "reserve region %u " 351 "max to eat into reserve %u", 352 _free_regions_at_end_of_collection, 353 desired_young_length, 354 _reserve_regions, 355 max_to_eat_into_reserve); 356 357 if (_free_regions_at_end_of_collection <= _reserve_regions) { 358 // Fully eat (or already eating) into the reserve, hand back at most absolute_min_length regions. 359 uint receiving_young = MIN3(_free_regions_at_end_of_collection, 360 desired_young_length, 361 max_to_eat_into_reserve); 362 // We could already have allocated more regions than what we could get 363 // above. 364 receiving_additional_eden = allocated_young_length < receiving_young ? 365 receiving_young - allocated_young_length : 0; 366 367 log_trace(gc, ergo, heap)("Young target length: Fully eat into reserve " 368 "receiving young %u receiving additional eden %u", 369 receiving_young, 370 receiving_additional_eden); 371 } else if (_free_regions_at_end_of_collection < (desired_young_length + _reserve_regions)) { 372 // Partially eat into the reserve, at most max_to_eat_into_reserve regions. 373 uint free_outside_reserve = _free_regions_at_end_of_collection - _reserve_regions; 374 assert(free_outside_reserve < desired_young_length, 375 "must be %u %u", 376 free_outside_reserve, desired_young_length); 377 378 uint receiving_within_reserve = MIN2(desired_young_length - free_outside_reserve, 379 max_to_eat_into_reserve); 380 uint receiving_young = free_outside_reserve + receiving_within_reserve; 381 // Again, we could have already allocated more than we could get. 382 receiving_additional_eden = allocated_young_length < receiving_young ? 383 receiving_young - allocated_young_length : 0; 384 385 log_trace(gc, ergo, heap)("Young target length: Partially eat into reserve " 386 "free outside reserve %u " 387 "receiving within reserve %u " 388 "receiving young %u " 389 "receiving additional eden %u", 390 free_outside_reserve, receiving_within_reserve, 391 receiving_young, receiving_additional_eden); 392 } else { 393 // No need to use the reserve. 394 receiving_additional_eden = desired_young_length - allocated_young_length; 395 log_trace(gc, ergo, heap)("Young target length: No need to use reserve " 396 "receiving additional eden %u", 397 receiving_additional_eden); 398 } 399 } 400 401 uint target_young_length = allocated_young_length + receiving_additional_eden; 402 403 assert(target_young_length >= allocated_young_length, "must be"); 404 405 log_trace(gc, ergo, heap)("Young target length: " 406 "young target length %u " 407 "allocated young length %u " 408 "received additional eden %u", 409 target_young_length, allocated_young_length, 410 receiving_additional_eden); 411 return target_young_length; 412 } 413 414 uint G1Policy::calculate_desired_eden_length_by_pause(double base_time_ms, 415 uint min_eden_length, 416 uint max_eden_length) const { 417 assert(use_adaptive_young_list_length(), "pre-condition"); 418 419 assert(min_eden_length <= max_eden_length, "must be %u %u", min_eden_length, max_eden_length); 420 421 // Here, we will make sure that the shortest young length that 422 // makes sense fits within the target pause time. 423 424 G1YoungLengthPredictor p(base_time_ms, 425 _free_regions_at_end_of_collection, 426 _mmu_tracker->max_gc_time() * 1000.0, 427 this); 428 if (p.will_fit(min_eden_length)) { 429 // The shortest young length will fit into the target pause time; 430 // we'll now check whether the absolute maximum number of young 431 // regions will fit in the target pause time. If not, we'll do 432 // a binary search between min_young_length and max_young_length. 433 if (p.will_fit(max_eden_length)) { 434 // The maximum young length will fit into the target pause time. 435 // We are done so set min young length to the maximum length (as 436 // the result is assumed to be returned in min_young_length). 437 min_eden_length = max_eden_length; 438 } else { 439 // The maximum possible number of young regions will not fit within 440 // the target pause time so we'll search for the optimal 441 // length. The loop invariants are: 442 // 443 // min_young_length < max_young_length 444 // min_young_length is known to fit into the target pause time 445 // max_young_length is known not to fit into the target pause time 446 // 447 // Going into the loop we know the above hold as we've just 448 // checked them. Every time around the loop we check whether 449 // the middle value between min_young_length and 450 // max_young_length fits into the target pause time. If it 451 // does, it becomes the new min. If it doesn't, it becomes 452 // the new max. This way we maintain the loop invariants. 453 454 assert(min_eden_length < max_eden_length, "invariant"); 455 uint diff = (max_eden_length - min_eden_length) / 2; 456 while (diff > 0) { 457 uint eden_length = min_eden_length + diff; 458 if (p.will_fit(eden_length)) { 459 min_eden_length = eden_length; 460 } else { 461 max_eden_length = eden_length; 462 } 463 assert(min_eden_length < max_eden_length, "invariant"); 464 diff = (max_eden_length - min_eden_length) / 2; 465 } 466 // The results is min_young_length which, according to the 467 // loop invariants, should fit within the target pause time. 468 469 // These are the post-conditions of the binary search above: 470 assert(min_eden_length < max_eden_length, 471 "otherwise we should have discovered that max_eden_length " 472 "fits into the pause target and not done the binary search"); 473 assert(p.will_fit(min_eden_length), 474 "min_eden_length, the result of the binary search, should " 475 "fit into the pause target"); 476 assert(!p.will_fit(min_eden_length + 1), 477 "min_eden_length, the result of the binary search, should be " 478 "optimal, so no larger length should fit into the pause target"); 479 } 480 } else { 481 // Even the minimum length doesn't fit into the pause time 482 // target, return it as the result nevertheless. 483 } 484 return min_eden_length; 485 } 486 487 uint G1Policy::calculate_desired_eden_length_before_mixed(double survivor_base_time_ms, 488 uint min_eden_length, 489 uint max_eden_length) const { 490 G1CollectionSetCandidates* candidates = _collection_set->candidates(); 491 492 uint min_old_regions_end = MIN2(candidates->cur_idx() + calc_min_old_cset_length(), candidates->num_regions()); 493 double predicted_region_evac_time_ms = survivor_base_time_ms; 494 for (uint i = candidates->cur_idx(); i < min_old_regions_end; i++) { 495 HeapRegion* r = candidates->at(i); 496 predicted_region_evac_time_ms += predict_region_total_time_ms(r, false); 497 } 498 uint desired_eden_length_by_min_cset_length = 499 calculate_desired_eden_length_by_pause(predicted_region_evac_time_ms, 500 min_eden_length, 501 max_eden_length); 502 503 return desired_eden_length_by_min_cset_length; 504 } 505 506 double G1Policy::predict_survivor_regions_evac_time() const { 507 double survivor_regions_evac_time = 0.0; 508 const GrowableArray<HeapRegion*>* survivor_regions = _g1h->survivor()->regions(); 509 for (GrowableArrayIterator<HeapRegion*> it = survivor_regions->begin(); 510 it != survivor_regions->end(); 511 ++it) { 512 survivor_regions_evac_time += predict_region_total_time_ms(*it, collector_state()->in_young_only_phase()); 513 } 514 return survivor_regions_evac_time; 515 } 516 517 G1GCPhaseTimes* G1Policy::phase_times() const { 518 // Lazy allocation because it must follow initialization of all the 519 // OopStorage objects by various other subsystems. 520 if (_phase_times == NULL) { 521 _phase_times = new G1GCPhaseTimes(_phase_times_timer, ParallelGCThreads); 522 } 523 return _phase_times; 524 } 525 526 void G1Policy::revise_young_list_target_length_if_necessary(size_t rs_length) { 527 guarantee(use_adaptive_young_list_length(), "should not call this otherwise" ); 528 529 if (rs_length > _rs_length_prediction) { 530 // add 10% to avoid having to recalculate often 531 size_t rs_length_prediction = rs_length * 1100 / 1000; 532 update_rs_length_prediction(rs_length_prediction); 533 update_young_length_bounds(rs_length_prediction); 534 } 535 } 536 537 void G1Policy::update_rs_length_prediction() { 538 update_rs_length_prediction(_analytics->predict_rs_length()); 539 } 540 541 void G1Policy::update_rs_length_prediction(size_t prediction) { 542 if (collector_state()->in_young_only_phase() && use_adaptive_young_list_length()) { 543 _rs_length_prediction = prediction; 544 } 545 } 546 547 void G1Policy::record_full_collection_start() { 548 _full_collection_start_sec = os::elapsedTime(); 549 // Release the future to-space so that it is available for compaction into. 550 collector_state()->set_in_young_only_phase(false); 551 collector_state()->set_in_full_gc(true); 552 _collection_set->clear_candidates(); 553 _pending_cards_at_gc_start = 0; 554 } 555 556 void G1Policy::record_full_collection_end() { 557 // Consider this like a collection pause for the purposes of allocation 558 // since last pause. 559 double end_sec = os::elapsedTime(); 560 double full_gc_time_sec = end_sec - _full_collection_start_sec; 561 double full_gc_time_ms = full_gc_time_sec * 1000.0; 562 563 _analytics->update_recent_gc_times(end_sec, full_gc_time_ms); 564 565 collector_state()->set_in_full_gc(false); 566 567 // "Nuke" the heuristics that control the young/mixed GC 568 // transitions and make sure we start with young GCs after the Full GC. 569 collector_state()->set_in_young_only_phase(true); 570 collector_state()->set_in_young_gc_before_mixed(false); 571 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0)); 572 collector_state()->set_in_concurrent_start_gc(false); 573 collector_state()->set_mark_or_rebuild_in_progress(false); 574 collector_state()->set_clearing_next_bitmap(false); 575 576 _eden_surv_rate_group->start_adding_regions(); 577 // also call this on any additional surv rate groups 578 579 _free_regions_at_end_of_collection = _g1h->num_free_regions(); 580 _survivor_surv_rate_group->reset(); 581 update_young_length_bounds(); 582 update_rs_length_prediction(); 583 584 _old_gen_alloc_tracker.reset_after_full_gc(); 585 586 record_pause(FullGC, _full_collection_start_sec, end_sec); 587 } 588 589 static void log_refinement_stats(const char* kind, const G1ConcurrentRefineStats& stats) { 590 log_debug(gc, refine, stats) 591 ("%s refinement: %.2fms, refined: " SIZE_FORMAT 592 ", precleaned: " SIZE_FORMAT ", dirtied: " SIZE_FORMAT, 593 kind, 594 stats.refinement_time().seconds() * MILLIUNITS, 595 stats.refined_cards(), 596 stats.precleaned_cards(), 597 stats.dirtied_cards()); 598 } 599 600 void G1Policy::record_concurrent_refinement_stats() { 601 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set(); 602 _pending_cards_at_gc_start = dcqs.num_cards(); 603 604 // Collect per-thread stats, mostly from mutator activity. 605 G1ConcurrentRefineStats mut_stats = dcqs.get_and_reset_refinement_stats(); 606 607 // Collect specialized concurrent refinement thread stats. 608 G1ConcurrentRefine* cr = _g1h->concurrent_refine(); 609 G1ConcurrentRefineStats cr_stats = cr->get_and_reset_refinement_stats(); 610 611 G1ConcurrentRefineStats total_stats = mut_stats + cr_stats; 612 613 log_refinement_stats("Mutator", mut_stats); 614 log_refinement_stats("Concurrent", cr_stats); 615 log_refinement_stats("Total", total_stats); 616 617 // Record the rate at which cards were refined. 618 // Don't update the rate if the current sample is empty or time is zero. 619 Tickspan refinement_time = total_stats.refinement_time(); 620 size_t refined_cards = total_stats.refined_cards(); 621 if ((refined_cards > 0) && (refinement_time > Tickspan())) { 622 double rate = refined_cards / (refinement_time.seconds() * MILLIUNITS); 623 _analytics->report_concurrent_refine_rate_ms(rate); 624 log_debug(gc, refine, stats)("Concurrent refinement rate: %.2f cards/ms", rate); 625 } 626 627 // Record mutator's card logging rate. 628 double mut_start_time = _analytics->prev_collection_pause_end_ms(); 629 double mut_end_time = phase_times()->cur_collection_start_sec() * MILLIUNITS; 630 double mut_time = mut_end_time - mut_start_time; 631 // Unlike above for conc-refine rate, here we should not require a 632 // non-empty sample, since an application could go some time with only 633 // young-gen or filtered out writes. But we'll ignore unusually short 634 // sample periods, as they may just pollute the predictions. 635 if (mut_time > 1.0) { // Require > 1ms sample time. 636 double dirtied_rate = total_stats.dirtied_cards() / mut_time; 637 _analytics->report_dirtied_cards_rate_ms(dirtied_rate); 638 log_debug(gc, refine, stats)("Generate dirty cards rate: %.2f cards/ms", dirtied_rate); 639 } 640 } 641 642 void G1Policy::record_collection_pause_start(double start_time_sec) { 643 // We only need to do this here as the policy will only be applied 644 // to the GC we're about to start. so, no point is calculating this 645 // every time we calculate / recalculate the target young length. 646 update_survivors_policy(); 647 648 assert(max_survivor_regions() + _g1h->num_used_regions() <= _g1h->max_regions(), 649 "Maximum survivor regions %u plus used regions %u exceeds max regions %u", 650 max_survivor_regions(), _g1h->num_used_regions(), _g1h->max_regions()); 651 assert_used_and_recalculate_used_equal(_g1h); 652 653 phase_times()->record_cur_collection_start_sec(start_time_sec); 654 655 record_concurrent_refinement_stats(); 656 657 _collection_set->reset_bytes_used_before(); 658 659 // do that for any other surv rate groups 660 _eden_surv_rate_group->stop_adding_regions(); 661 _survivors_age_table.clear(); 662 663 assert(_g1h->collection_set()->verify_young_ages(), "region age verification failed"); 664 } 665 666 void G1Policy::record_concurrent_mark_init_end(double mark_init_elapsed_time_ms) { 667 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now"); 668 collector_state()->set_in_concurrent_start_gc(false); 669 } 670 671 void G1Policy::record_concurrent_mark_remark_start() { 672 _mark_remark_start_sec = os::elapsedTime(); 673 } 674 675 void G1Policy::record_concurrent_mark_remark_end() { 676 double end_time_sec = os::elapsedTime(); 677 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 678 _analytics->report_concurrent_mark_remark_times_ms(elapsed_time_ms); 679 _analytics->append_prev_collection_pause_end_ms(elapsed_time_ms); 680 681 record_pause(Remark, _mark_remark_start_sec, end_time_sec); 682 } 683 684 void G1Policy::record_concurrent_mark_cleanup_start() { 685 _mark_cleanup_start_sec = os::elapsedTime(); 686 } 687 688 double G1Policy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const { 689 return phase_times()->average_time_ms(phase); 690 } 691 692 double G1Policy::young_other_time_ms() const { 693 return phase_times()->young_cset_choice_time_ms() + 694 phase_times()->average_time_ms(G1GCPhaseTimes::YoungFreeCSet); 695 } 696 697 double G1Policy::non_young_other_time_ms() const { 698 return phase_times()->non_young_cset_choice_time_ms() + 699 phase_times()->average_time_ms(G1GCPhaseTimes::NonYoungFreeCSet); 700 } 701 702 double G1Policy::other_time_ms(double pause_time_ms) const { 703 return pause_time_ms - phase_times()->cur_collection_par_time_ms(); 704 } 705 706 double G1Policy::constant_other_time_ms(double pause_time_ms) const { 707 return other_time_ms(pause_time_ms) - phase_times()->total_free_cset_time_ms() - phase_times()->total_rebuild_freelist_time_ms(); 708 } 709 710 bool G1Policy::about_to_start_mixed_phase() const { 711 return _g1h->concurrent_mark()->cm_thread()->during_cycle() || collector_state()->in_young_gc_before_mixed(); 712 } 713 714 bool G1Policy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 715 if (about_to_start_mixed_phase()) { 716 return false; 717 } 718 719 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold(); 720 721 size_t cur_used_bytes = _g1h->non_young_capacity_bytes(); 722 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 723 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size; 724 725 bool result = false; 726 if (marking_request_bytes > marking_initiating_used_threshold) { 727 result = collector_state()->in_young_only_phase() && !collector_state()->in_young_gc_before_mixed(); 728 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s", 729 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)", 730 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1h->capacity() * 100, source); 731 } 732 733 return result; 734 } 735 736 double G1Policy::logged_cards_processing_time() const { 737 double all_cards_processing_time = average_time_ms(G1GCPhaseTimes::ScanHR) + average_time_ms(G1GCPhaseTimes::OptScanHR); 738 size_t logged_dirty_cards = phase_times()->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards); 739 size_t scan_heap_roots_cards = phase_times()->sum_thread_work_items(G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::ScanHRScannedCards) + 740 phase_times()->sum_thread_work_items(G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::ScanHRScannedCards); 741 // This may happen if there are duplicate cards in different log buffers. 742 if (logged_dirty_cards > scan_heap_roots_cards) { 743 return all_cards_processing_time + average_time_ms(G1GCPhaseTimes::MergeLB); 744 } 745 return (all_cards_processing_time * logged_dirty_cards / scan_heap_roots_cards) + average_time_ms(G1GCPhaseTimes::MergeLB); 746 } 747 748 // Anything below that is considered to be zero 749 #define MIN_TIMER_GRANULARITY 0.0000001 750 751 void G1Policy::record_collection_pause_end(double pause_time_ms) { 752 G1GCPhaseTimes* p = phase_times(); 753 754 double end_time_sec = os::elapsedTime(); 755 756 PauseKind this_pause = young_gc_pause_kind(); 757 758 bool update_stats = !_g1h->evacuation_failed(); 759 760 record_pause(this_pause, end_time_sec - pause_time_ms / 1000.0, end_time_sec); 761 762 if (is_concurrent_start_pause(this_pause)) { 763 record_concurrent_mark_init_end(0.0); 764 } else { 765 maybe_start_marking(); 766 } 767 768 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _analytics->prev_collection_pause_end_ms()); 769 if (app_time_ms < MIN_TIMER_GRANULARITY) { 770 // This usually happens due to the timer not having the required 771 // granularity. Some Linuxes are the usual culprits. 772 // We'll just set it to something (arbitrarily) small. 773 app_time_ms = 1.0; 774 } 775 776 if (update_stats) { 777 // We maintain the invariant that all objects allocated by mutator 778 // threads will be allocated out of eden regions. So, we can use 779 // the eden region number allocated since the previous GC to 780 // calculate the application's allocate rate. The only exception 781 // to that is humongous objects that are allocated separately. But 782 // given that humongous object allocations do not really affect 783 // either the pause's duration nor when the next pause will take 784 // place we can safely ignore them here. 785 uint regions_allocated = _collection_set->eden_region_length(); 786 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 787 _analytics->report_alloc_rate_ms(alloc_rate_ms); 788 789 _analytics->compute_pause_time_ratios(end_time_sec, pause_time_ms); 790 _analytics->update_recent_gc_times(end_time_sec, pause_time_ms); 791 } 792 793 if (is_last_young_pause(this_pause)) { 794 assert(!is_concurrent_start_pause(this_pause), 795 "The young GC before mixed is not allowed to be concurrent start GC"); 796 // This has been the young GC before we start doing mixed GCs. We already 797 // decided to start mixed GCs much earlier, so there is nothing to do except 798 // advancing the state. 799 collector_state()->set_in_young_only_phase(false); 800 collector_state()->set_in_young_gc_before_mixed(false); 801 } else if (is_mixed_pause(this_pause)) { 802 // This is a mixed GC. Here we decide whether to continue doing more 803 // mixed GCs or not. 804 if (!next_gc_should_be_mixed("continue mixed GCs", 805 "do not continue mixed GCs")) { 806 collector_state()->set_in_young_only_phase(true); 807 808 clear_collection_set_candidates(); 809 maybe_start_marking(); 810 } 811 } else { 812 assert(is_young_only_pause(this_pause), "must be"); 813 } 814 815 _eden_surv_rate_group->start_adding_regions(); 816 817 double merge_hcc_time_ms = average_time_ms(G1GCPhaseTimes::MergeHCC); 818 if (update_stats) { 819 size_t const total_log_buffer_cards = p->sum_thread_work_items(G1GCPhaseTimes::MergeHCC, G1GCPhaseTimes::MergeHCCDirtyCards) + 820 p->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards); 821 // Update prediction for card merge; MergeRSDirtyCards includes the cards from the Eager Reclaim phase. 822 size_t const total_cards_merged = p->sum_thread_work_items(G1GCPhaseTimes::MergeRS, G1GCPhaseTimes::MergeRSDirtyCards) + 823 p->sum_thread_work_items(G1GCPhaseTimes::OptMergeRS, G1GCPhaseTimes::MergeRSDirtyCards) + 824 total_log_buffer_cards; 825 826 // The threshold for the number of cards in a given sampling which we consider 827 // large enough so that the impact from setup and other costs is negligible. 828 size_t const CardsNumSamplingThreshold = 10; 829 830 if (total_cards_merged > CardsNumSamplingThreshold) { 831 double avg_time_merge_cards = average_time_ms(G1GCPhaseTimes::MergeER) + 832 average_time_ms(G1GCPhaseTimes::MergeRS) + 833 average_time_ms(G1GCPhaseTimes::MergeHCC) + 834 average_time_ms(G1GCPhaseTimes::MergeLB) + 835 average_time_ms(G1GCPhaseTimes::OptMergeRS); 836 _analytics->report_cost_per_card_merge_ms(avg_time_merge_cards / total_cards_merged, 837 is_young_only_pause(this_pause)); 838 } 839 840 // Update prediction for card scan 841 size_t const total_cards_scanned = p->sum_thread_work_items(G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::ScanHRScannedCards) + 842 p->sum_thread_work_items(G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::ScanHRScannedCards); 843 844 if (total_cards_scanned > CardsNumSamplingThreshold) { 845 double avg_time_dirty_card_scan = average_time_ms(G1GCPhaseTimes::ScanHR) + 846 average_time_ms(G1GCPhaseTimes::OptScanHR); 847 848 _analytics->report_cost_per_card_scan_ms(avg_time_dirty_card_scan / total_cards_scanned, 849 is_young_only_pause(this_pause)); 850 } 851 852 // Update prediction for the ratio between cards from the remembered 853 // sets and actually scanned cards from the remembered sets. 854 // Cards from the remembered sets are all cards not duplicated by cards from 855 // the logs. 856 // Due to duplicates in the log buffers, the number of actually scanned cards 857 // can be smaller than the cards in the log buffers. 858 const size_t from_rs_length_cards = (total_cards_scanned > total_log_buffer_cards) ? total_cards_scanned - total_log_buffer_cards : 0; 859 double merge_to_scan_ratio = 0.0; 860 if (total_cards_scanned > 0) { 861 merge_to_scan_ratio = (double) from_rs_length_cards / total_cards_scanned; 862 } 863 _analytics->report_card_merge_to_scan_ratio(merge_to_scan_ratio, 864 is_young_only_pause(this_pause)); 865 866 const size_t recorded_rs_length = _collection_set->recorded_rs_length(); 867 const size_t rs_length_diff = _rs_length > recorded_rs_length ? _rs_length - recorded_rs_length : 0; 868 _analytics->report_rs_length_diff(rs_length_diff); 869 870 // Update prediction for copy cost per byte 871 size_t copied_bytes = p->sum_thread_work_items(G1GCPhaseTimes::MergePSS, G1GCPhaseTimes::MergePSSCopiedBytes); 872 873 if (copied_bytes > 0) { 874 double cost_per_byte_ms = (average_time_ms(G1GCPhaseTimes::ObjCopy) + average_time_ms(G1GCPhaseTimes::OptObjCopy)) / copied_bytes; 875 _analytics->report_cost_per_byte_ms(cost_per_byte_ms, collector_state()->mark_or_rebuild_in_progress()); 876 } 877 878 if (_collection_set->young_region_length() > 0) { 879 _analytics->report_young_other_cost_per_region_ms(young_other_time_ms() / 880 _collection_set->young_region_length()); 881 } 882 883 if (_collection_set->old_region_length() > 0) { 884 _analytics->report_non_young_other_cost_per_region_ms(non_young_other_time_ms() / 885 _collection_set->old_region_length()); 886 } 887 888 _analytics->report_constant_other_time_ms(constant_other_time_ms(pause_time_ms)); 889 890 // Do not update RS lengths and the number of pending cards with information from mixed gc: 891 // these are is wildly different to during young only gc and mess up young gen sizing right 892 // after the mixed gc phase. 893 // During mixed gc we do not use them for young gen sizing. 894 if (is_young_only_pause(this_pause)) { 895 _analytics->report_pending_cards((double) _pending_cards_at_gc_start); 896 _analytics->report_rs_length((double) _rs_length); 897 } 898 } 899 900 assert(!(is_concurrent_start_pause(this_pause) && collector_state()->mark_or_rebuild_in_progress()), 901 "If the last pause has been concurrent start, we should not have been in the marking window"); 902 if (is_concurrent_start_pause(this_pause)) { 903 collector_state()->set_mark_or_rebuild_in_progress(true); 904 } 905 906 _free_regions_at_end_of_collection = _g1h->num_free_regions(); 907 908 update_rs_length_prediction(); 909 910 // Do not update dynamic IHOP due to G1 periodic collection as it is highly likely 911 // that in this case we are not running in a "normal" operating mode. 912 if (_g1h->gc_cause() != GCCause::_g1_periodic_collection) { 913 update_young_length_bounds(); 914 915 _old_gen_alloc_tracker.reset_after_young_gc(app_time_ms / 1000.0); 916 update_ihop_prediction(_old_gen_alloc_tracker.last_cycle_duration(), 917 _old_gen_alloc_tracker.last_cycle_old_bytes(), 918 is_young_only_pause(this_pause)); 919 920 _ihop_control->send_trace_event(_g1h->gc_tracer_stw()); 921 } else { 922 // Any garbage collection triggered as periodic collection resets the time-to-mixed 923 // measurement. Periodic collection typically means that the application is "inactive", i.e. 924 // the marking threads may have received an uncharacterisic amount of cpu time 925 // for completing the marking, i.e. are faster than expected. 926 // This skews the predicted marking length towards smaller values which might cause 927 // the mark start being too late. 928 _concurrent_start_to_mixed.reset(); 929 } 930 931 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 932 double scan_logged_cards_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 933 934 if (scan_logged_cards_time_goal_ms < merge_hcc_time_ms) { 935 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)." 936 "Logged Cards Scan time goal: %1.2fms Scan HCC time: %1.2fms", 937 scan_logged_cards_time_goal_ms, merge_hcc_time_ms); 938 939 scan_logged_cards_time_goal_ms = 0; 940 } else { 941 scan_logged_cards_time_goal_ms -= merge_hcc_time_ms; 942 } 943 944 double const logged_cards_time = logged_cards_processing_time(); 945 946 log_debug(gc, ergo, refine)("Concurrent refinement times: Logged Cards Scan time goal: %1.2fms Logged Cards Scan time: %1.2fms HCC time: %1.2fms", 947 scan_logged_cards_time_goal_ms, logged_cards_time, merge_hcc_time_ms); 948 949 _g1h->concurrent_refine()->adjust(logged_cards_time, 950 phase_times()->sum_thread_work_items(G1GCPhaseTimes::MergeLB, G1GCPhaseTimes::MergeLBDirtyCards), 951 scan_logged_cards_time_goal_ms); 952 } 953 954 G1IHOPControl* G1Policy::create_ihop_control(const G1Predictions* predictor){ 955 if (G1UseAdaptiveIHOP) { 956 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent, 957 predictor, 958 G1ReservePercent, 959 G1HeapWastePercent); 960 } else { 961 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent); 962 } 963 } 964 965 void G1Policy::update_ihop_prediction(double mutator_time_s, 966 size_t mutator_alloc_bytes, 967 bool this_gc_was_young_only) { 968 // Always try to update IHOP prediction. Even evacuation failures give information 969 // about e.g. whether to start IHOP earlier next time. 970 971 // Avoid using really small application times that might create samples with 972 // very high or very low values. They may be caused by e.g. back-to-back gcs. 973 double const min_valid_time = 1e-6; 974 975 bool report = false; 976 977 double marking_to_mixed_time = -1.0; 978 if (!this_gc_was_young_only && _concurrent_start_to_mixed.has_result()) { 979 marking_to_mixed_time = _concurrent_start_to_mixed.last_marking_time(); 980 assert(marking_to_mixed_time > 0.0, 981 "Concurrent start to mixed time must be larger than zero but is %.3f", 982 marking_to_mixed_time); 983 if (marking_to_mixed_time > min_valid_time) { 984 _ihop_control->update_marking_length(marking_to_mixed_time); 985 report = true; 986 } 987 } 988 989 // As an approximation for the young gc promotion rates during marking we use 990 // all of them. In many applications there are only a few if any young gcs during 991 // marking, which makes any prediction useless. This increases the accuracy of the 992 // prediction. 993 if (this_gc_was_young_only && mutator_time_s > min_valid_time) { 994 // IHOP control wants to know the expected young gen length if it were not 995 // restrained by the heap reserve. Using the actual length would make the 996 // prediction too small and the limit the young gen every time we get to the 997 // predicted target occupancy. 998 size_t young_gen_size = young_list_desired_length() * HeapRegion::GrainBytes; 999 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size); 1000 report = true; 1001 } 1002 1003 if (report) { 1004 report_ihop_statistics(); 1005 } 1006 } 1007 1008 void G1Policy::report_ihop_statistics() { 1009 _ihop_control->print(); 1010 } 1011 1012 void G1Policy::print_phases() { 1013 phase_times()->print(); 1014 } 1015 1016 double G1Policy::predict_base_elapsed_time_ms(size_t pending_cards, 1017 size_t rs_length) const { 1018 size_t effective_scanned_cards = _analytics->predict_scan_card_num(rs_length, collector_state()->in_young_only_phase()); 1019 return 1020 _analytics->predict_card_merge_time_ms(pending_cards + rs_length, collector_state()->in_young_only_phase()) + 1021 _analytics->predict_card_scan_time_ms(effective_scanned_cards, collector_state()->in_young_only_phase()) + 1022 _analytics->predict_constant_other_time_ms() + 1023 predict_survivor_regions_evac_time(); 1024 } 1025 1026 double G1Policy::predict_base_elapsed_time_ms(size_t pending_cards) const { 1027 size_t rs_length = _analytics->predict_rs_length(); 1028 return predict_base_elapsed_time_ms(pending_cards, rs_length); 1029 } 1030 1031 size_t G1Policy::predict_bytes_to_copy(HeapRegion* hr) const { 1032 size_t bytes_to_copy; 1033 if (!hr->is_young()) { 1034 bytes_to_copy = hr->max_live_bytes(); 1035 } else { 1036 bytes_to_copy = (size_t) (hr->used() * hr->surv_rate_prediction(_predictor)); 1037 } 1038 return bytes_to_copy; 1039 } 1040 1041 double G1Policy::predict_eden_copy_time_ms(uint count, size_t* bytes_to_copy) const { 1042 if (count == 0) { 1043 return 0.0; 1044 } 1045 size_t const expected_bytes = _eden_surv_rate_group->accum_surv_rate_pred(count) * HeapRegion::GrainBytes; 1046 if (bytes_to_copy != NULL) { 1047 *bytes_to_copy = expected_bytes; 1048 } 1049 return _analytics->predict_object_copy_time_ms(expected_bytes, collector_state()->mark_or_rebuild_in_progress()); 1050 } 1051 1052 double G1Policy::predict_region_copy_time_ms(HeapRegion* hr) const { 1053 size_t const bytes_to_copy = predict_bytes_to_copy(hr); 1054 return _analytics->predict_object_copy_time_ms(bytes_to_copy, collector_state()->mark_or_rebuild_in_progress()); 1055 } 1056 1057 double G1Policy::predict_region_non_copy_time_ms(HeapRegion* hr, 1058 bool for_young_gc) const { 1059 size_t rs_length = hr->rem_set()->occupied(); 1060 size_t scan_card_num = _analytics->predict_scan_card_num(rs_length, for_young_gc); 1061 1062 double region_elapsed_time_ms = 1063 _analytics->predict_card_merge_time_ms(rs_length, collector_state()->in_young_only_phase()) + 1064 _analytics->predict_card_scan_time_ms(scan_card_num, collector_state()->in_young_only_phase()); 1065 1066 // The prediction of the "other" time for this region is based 1067 // upon the region type and NOT the GC type. 1068 if (hr->is_young()) { 1069 region_elapsed_time_ms += _analytics->predict_young_other_time_ms(1); 1070 } else { 1071 region_elapsed_time_ms += _analytics->predict_non_young_other_time_ms(1); 1072 } 1073 return region_elapsed_time_ms; 1074 } 1075 1076 double G1Policy::predict_region_total_time_ms(HeapRegion* hr, bool for_young_gc) const { 1077 return predict_region_non_copy_time_ms(hr, for_young_gc) + predict_region_copy_time_ms(hr); 1078 } 1079 1080 bool G1Policy::should_allocate_mutator_region() const { 1081 uint young_list_length = _g1h->young_regions_count(); 1082 uint young_list_target_length = _young_list_target_length; 1083 return young_list_length < young_list_target_length; 1084 } 1085 1086 bool G1Policy::can_expand_young_list() const { 1087 uint young_list_length = _g1h->young_regions_count(); 1088 uint young_list_max_length = _young_list_max_length; 1089 return young_list_length < young_list_max_length; 1090 } 1091 1092 bool G1Policy::use_adaptive_young_list_length() const { 1093 return _young_gen_sizer->use_adaptive_young_list_length(); 1094 } 1095 1096 size_t G1Policy::desired_survivor_size(uint max_regions) const { 1097 size_t const survivor_capacity = HeapRegion::GrainWords * max_regions; 1098 return (size_t)((((double)survivor_capacity) * TargetSurvivorRatio) / 100); 1099 } 1100 1101 void G1Policy::print_age_table() { 1102 _survivors_age_table.print_age_table(_tenuring_threshold); 1103 } 1104 1105 uint G1Policy::calculate_young_max_length(uint target_young_length) const { 1106 uint expansion_region_num = 0; 1107 if (GCLockerEdenExpansionPercent > 0) { 1108 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1109 double expansion_region_num_d = perc * (double) _young_list_target_length; 1110 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1111 // less than 1.0) we'll get 1. 1112 expansion_region_num = (uint) ceil(expansion_region_num_d); 1113 } else { 1114 assert(expansion_region_num == 0, "sanity"); 1115 } 1116 uint max_length = target_young_length + expansion_region_num; 1117 assert(target_young_length <= max_length, "post-condition"); 1118 return max_length; 1119 } 1120 1121 // Calculates survivor space parameters. 1122 void G1Policy::update_survivors_policy() { 1123 double max_survivor_regions_d = 1124 (double) _young_list_target_length / (double) SurvivorRatio; 1125 1126 // Calculate desired survivor size based on desired max survivor regions (unconstrained 1127 // by remaining heap). Otherwise we may cause undesired promotions as we are 1128 // already getting close to end of the heap, impacting performance even more. 1129 uint const desired_max_survivor_regions = ceil(max_survivor_regions_d); 1130 size_t const survivor_size = desired_survivor_size(desired_max_survivor_regions); 1131 1132 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(survivor_size); 1133 if (UsePerfData) { 1134 _policy_counters->tenuring_threshold()->set_value(_tenuring_threshold); 1135 _policy_counters->desired_survivor_size()->set_value(survivor_size * oopSize); 1136 } 1137 // The real maximum survivor size is bounded by the number of regions that can 1138 // be allocated into. 1139 _max_survivor_regions = MIN2(desired_max_survivor_regions, 1140 _g1h->num_free_or_available_regions()); 1141 } 1142 1143 bool G1Policy::force_concurrent_start_if_outside_cycle(GCCause::Cause gc_cause) { 1144 // We actually check whether we are marking here and not if we are in a 1145 // reclamation phase. This means that we will schedule a concurrent mark 1146 // even while we are still in the process of reclaiming memory. 1147 bool during_cycle = _g1h->concurrent_mark()->cm_thread()->during_cycle(); 1148 if (!during_cycle) { 1149 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). " 1150 "GC cause: %s", 1151 GCCause::to_string(gc_cause)); 1152 collector_state()->set_initiate_conc_mark_if_possible(true); 1153 return true; 1154 } else { 1155 log_debug(gc, ergo)("Do not request concurrent cycle initiation " 1156 "(concurrent cycle already in progress). GC cause: %s", 1157 GCCause::to_string(gc_cause)); 1158 return false; 1159 } 1160 } 1161 1162 void G1Policy::initiate_conc_mark() { 1163 collector_state()->set_in_concurrent_start_gc(true); 1164 collector_state()->set_initiate_conc_mark_if_possible(false); 1165 } 1166 1167 void G1Policy::decide_on_conc_mark_initiation() { 1168 // We are about to decide on whether this pause will be a 1169 // concurrent start pause. 1170 1171 // First, collector_state()->in_concurrent_start_gc() should not be already set. We 1172 // will set it here if we have to. However, it should be cleared by 1173 // the end of the pause (it's only set for the duration of a 1174 // concurrent start pause). 1175 assert(!collector_state()->in_concurrent_start_gc(), "pre-condition"); 1176 1177 if (collector_state()->initiate_conc_mark_if_possible()) { 1178 // We had noticed on a previous pause that the heap occupancy has 1179 // gone over the initiating threshold and we should start a 1180 // concurrent marking cycle. Or we've been explicitly requested 1181 // to start a concurrent marking cycle. Either way, we initiate 1182 // one if not inhibited for some reason. 1183 1184 GCCause::Cause cause = _g1h->gc_cause(); 1185 if ((cause != GCCause::_wb_breakpoint) && 1186 ConcurrentGCBreakpoints::is_controlled()) { 1187 log_debug(gc, ergo)("Do not initiate concurrent cycle (whitebox controlled)"); 1188 } else if (!about_to_start_mixed_phase() && collector_state()->in_young_only_phase()) { 1189 // Initiate a new concurrent start if there is no marking or reclamation going on. 1190 initiate_conc_mark(); 1191 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)"); 1192 } else if (_g1h->is_user_requested_concurrent_full_gc(cause) || 1193 (cause == GCCause::_wb_breakpoint)) { 1194 // Initiate a user requested concurrent start or run to a breakpoint. 1195 // A concurrent start must be young only GC, so the collector state 1196 // must be updated to reflect this. 1197 collector_state()->set_in_young_only_phase(true); 1198 collector_state()->set_in_young_gc_before_mixed(false); 1199 1200 // We might have ended up coming here about to start a mixed phase with a collection set 1201 // active. The following remark might change the change the "evacuation efficiency" of 1202 // the regions in this set, leading to failing asserts later. 1203 // Since the concurrent cycle will recreate the collection set anyway, simply drop it here. 1204 clear_collection_set_candidates(); 1205 abort_time_to_mixed_tracking(); 1206 initiate_conc_mark(); 1207 log_debug(gc, ergo)("Initiate concurrent cycle (%s requested concurrent cycle)", 1208 (cause == GCCause::_wb_breakpoint) ? "run_to breakpoint" : "user"); 1209 } else { 1210 // The concurrent marking thread is still finishing up the 1211 // previous cycle. If we start one right now the two cycles 1212 // overlap. In particular, the concurrent marking thread might 1213 // be in the process of clearing the next marking bitmap (which 1214 // we will use for the next cycle if we start one). Starting a 1215 // cycle now will be bad given that parts of the marking 1216 // information might get cleared by the marking thread. And we 1217 // cannot wait for the marking thread to finish the cycle as it 1218 // periodically yields while clearing the next marking bitmap 1219 // and, if it's in a yield point, it's waiting for us to 1220 // finish. So, at this point we will not start a cycle and we'll 1221 // let the concurrent marking thread complete the last one. 1222 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)"); 1223 } 1224 } 1225 } 1226 1227 void G1Policy::record_concurrent_mark_cleanup_end() { 1228 G1CollectionSetCandidates* candidates = G1CollectionSetChooser::build(_g1h->workers(), _g1h->num_regions()); 1229 _collection_set->set_candidates(candidates); 1230 1231 bool mixed_gc_pending = next_gc_should_be_mixed("request mixed gcs", "request young-only gcs"); 1232 if (!mixed_gc_pending) { 1233 clear_collection_set_candidates(); 1234 abort_time_to_mixed_tracking(); 1235 } 1236 collector_state()->set_in_young_gc_before_mixed(mixed_gc_pending); 1237 collector_state()->set_mark_or_rebuild_in_progress(false); 1238 1239 double end_sec = os::elapsedTime(); 1240 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 1241 _analytics->report_concurrent_mark_cleanup_times_ms(elapsed_time_ms); 1242 _analytics->append_prev_collection_pause_end_ms(elapsed_time_ms); 1243 1244 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec); 1245 } 1246 1247 double G1Policy::reclaimable_bytes_percent(size_t reclaimable_bytes) const { 1248 return percent_of(reclaimable_bytes, _g1h->capacity()); 1249 } 1250 1251 class G1ClearCollectionSetCandidateRemSets : public HeapRegionClosure { 1252 virtual bool do_heap_region(HeapRegion* r) { 1253 r->rem_set()->clear_locked(true /* only_cardset */); 1254 return false; 1255 } 1256 }; 1257 1258 void G1Policy::clear_collection_set_candidates() { 1259 // Clear remembered sets of remaining candidate regions and the actual candidate 1260 // set. 1261 G1ClearCollectionSetCandidateRemSets cl; 1262 _collection_set->candidates()->iterate(&cl); 1263 _collection_set->clear_candidates(); 1264 } 1265 1266 void G1Policy::maybe_start_marking() { 1267 if (need_to_start_conc_mark("end of GC")) { 1268 // Note: this might have already been set, if during the last 1269 // pause we decided to start a cycle but at the beginning of 1270 // this pause we decided to postpone it. That's OK. 1271 collector_state()->set_initiate_conc_mark_if_possible(true); 1272 } 1273 } 1274 1275 bool G1Policy::is_young_only_pause(PauseKind kind) { 1276 assert(kind != FullGC, "must be"); 1277 assert(kind != Remark, "must be"); 1278 assert(kind != Cleanup, "must be"); 1279 return kind == ConcurrentStartGC || kind == LastYoungGC || kind == YoungOnlyGC; 1280 } 1281 1282 bool G1Policy::is_mixed_pause(PauseKind kind) { 1283 assert(kind != FullGC, "must be"); 1284 assert(kind != Remark, "must be"); 1285 assert(kind != Cleanup, "must be"); 1286 return kind == MixedGC; 1287 } 1288 1289 bool G1Policy::is_last_young_pause(PauseKind kind) { 1290 return kind == LastYoungGC; 1291 } 1292 1293 bool G1Policy::is_concurrent_start_pause(PauseKind kind) { 1294 return kind == ConcurrentStartGC; 1295 } 1296 1297 G1Policy::PauseKind G1Policy::young_gc_pause_kind() const { 1298 assert(!collector_state()->in_full_gc(), "must be"); 1299 if (collector_state()->in_concurrent_start_gc()) { 1300 assert(!collector_state()->in_young_gc_before_mixed(), "must be"); 1301 return ConcurrentStartGC; 1302 } else if (collector_state()->in_young_gc_before_mixed()) { 1303 assert(!collector_state()->in_concurrent_start_gc(), "must be"); 1304 return LastYoungGC; 1305 } else if (collector_state()->in_mixed_phase()) { 1306 assert(!collector_state()->in_concurrent_start_gc(), "must be"); 1307 assert(!collector_state()->in_young_gc_before_mixed(), "must be"); 1308 return MixedGC; 1309 } else { 1310 assert(!collector_state()->in_concurrent_start_gc(), "must be"); 1311 assert(!collector_state()->in_young_gc_before_mixed(), "must be"); 1312 return YoungOnlyGC; 1313 } 1314 } 1315 1316 void G1Policy::record_pause(PauseKind kind, double start, double end) { 1317 // Manage the MMU tracker. For some reason it ignores Full GCs. 1318 if (kind != FullGC) { 1319 _mmu_tracker->add_pause(start, end); 1320 } 1321 // Manage the mutator time tracking from concurrent start to first mixed gc. 1322 switch (kind) { 1323 case FullGC: 1324 abort_time_to_mixed_tracking(); 1325 break; 1326 case Cleanup: 1327 case Remark: 1328 case YoungOnlyGC: 1329 case LastYoungGC: 1330 _concurrent_start_to_mixed.add_pause(end - start); 1331 break; 1332 case ConcurrentStartGC: 1333 if (_g1h->gc_cause() != GCCause::_g1_periodic_collection) { 1334 _concurrent_start_to_mixed.record_concurrent_start_end(end); 1335 } 1336 break; 1337 case MixedGC: 1338 _concurrent_start_to_mixed.record_mixed_gc_start(start); 1339 break; 1340 default: 1341 ShouldNotReachHere(); 1342 } 1343 } 1344 1345 void G1Policy::abort_time_to_mixed_tracking() { 1346 _concurrent_start_to_mixed.reset(); 1347 } 1348 1349 bool G1Policy::next_gc_should_be_mixed(const char* true_action_str, 1350 const char* false_action_str) const { 1351 G1CollectionSetCandidates* candidates = _collection_set->candidates(); 1352 1353 if (candidates == NULL || candidates->is_empty()) { 1354 if (false_action_str != NULL) { 1355 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str); 1356 } 1357 return false; 1358 } 1359 1360 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? 1361 size_t reclaimable_bytes = candidates->remaining_reclaimable_bytes(); 1362 double reclaimable_percent = reclaimable_bytes_percent(reclaimable_bytes); 1363 double threshold = (double) G1HeapWastePercent; 1364 if (reclaimable_percent <= threshold) { 1365 if (false_action_str != NULL) { 1366 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1367 false_action_str, candidates->num_remaining(), reclaimable_bytes, reclaimable_percent, G1HeapWastePercent); 1368 } 1369 return false; 1370 } 1371 if (true_action_str != NULL) { 1372 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT, 1373 true_action_str, candidates->num_remaining(), reclaimable_bytes, reclaimable_percent, G1HeapWastePercent); 1374 } 1375 return true; 1376 } 1377 1378 uint G1Policy::calc_min_old_cset_length() const { 1379 // The min old CSet region bound is based on the maximum desired 1380 // number of mixed GCs after a cycle. I.e., even if some old regions 1381 // look expensive, we should add them to the CSet anyway to make 1382 // sure we go through the available old regions in no more than the 1383 // maximum desired number of mixed GCs. 1384 // 1385 // The calculation is based on the number of marked regions we added 1386 // to the CSet candidates in the first place, not how many remain, so 1387 // that the result is the same during all mixed GCs that follow a cycle. 1388 1389 const size_t region_num = _collection_set->candidates()->num_regions(); 1390 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); 1391 size_t result = region_num / gc_num; 1392 // emulate ceiling 1393 if (result * gc_num < region_num) { 1394 result += 1; 1395 } 1396 return (uint) result; 1397 } 1398 1399 uint G1Policy::calc_max_old_cset_length() const { 1400 // The max old CSet region bound is based on the threshold expressed 1401 // as a percentage of the heap size. I.e., it should bound the 1402 // number of old regions added to the CSet irrespective of how many 1403 // of them are available. 1404 1405 const G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1406 const size_t region_num = g1h->num_regions(); 1407 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; 1408 size_t result = region_num * perc / 100; 1409 // emulate ceiling 1410 if (100 * result < region_num * perc) { 1411 result += 1; 1412 } 1413 return (uint) result; 1414 } 1415 1416 void G1Policy::calculate_old_collection_set_regions(G1CollectionSetCandidates* candidates, 1417 double time_remaining_ms, 1418 uint& num_initial_regions, 1419 uint& num_optional_regions) { 1420 assert(candidates != NULL, "Must be"); 1421 1422 num_initial_regions = 0; 1423 num_optional_regions = 0; 1424 uint num_expensive_regions = 0; 1425 1426 double predicted_old_time_ms = 0.0; 1427 double predicted_initial_time_ms = 0.0; 1428 double predicted_optional_time_ms = 0.0; 1429 1430 double optional_threshold_ms = time_remaining_ms * optional_prediction_fraction(); 1431 1432 const uint min_old_cset_length = calc_min_old_cset_length(); 1433 const uint max_old_cset_length = MAX2(min_old_cset_length, calc_max_old_cset_length()); 1434 const uint max_optional_regions = max_old_cset_length - min_old_cset_length; 1435 bool check_time_remaining = use_adaptive_young_list_length(); 1436 1437 uint candidate_idx = candidates->cur_idx(); 1438 1439 log_debug(gc, ergo, cset)("Start adding old regions to collection set. Min %u regions, max %u regions, " 1440 "time remaining %1.2fms, optional threshold %1.2fms", 1441 min_old_cset_length, max_old_cset_length, time_remaining_ms, optional_threshold_ms); 1442 1443 HeapRegion* hr = candidates->at(candidate_idx); 1444 while (hr != NULL) { 1445 if (num_initial_regions + num_optional_regions >= max_old_cset_length) { 1446 // Added maximum number of old regions to the CSet. 1447 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Maximum number of regions). " 1448 "Initial %u regions, optional %u regions", 1449 num_initial_regions, num_optional_regions); 1450 break; 1451 } 1452 1453 // Stop adding regions if the remaining reclaimable space is 1454 // not above G1HeapWastePercent. 1455 size_t reclaimable_bytes = candidates->remaining_reclaimable_bytes(); 1456 double reclaimable_percent = reclaimable_bytes_percent(reclaimable_bytes); 1457 double threshold = (double) G1HeapWastePercent; 1458 if (reclaimable_percent <= threshold) { 1459 // We've added enough old regions that the amount of uncollected 1460 // reclaimable space is at or below the waste threshold. Stop 1461 // adding old regions to the CSet. 1462 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Reclaimable percentage below threshold). " 1463 "Reclaimable: " SIZE_FORMAT "%s (%1.2f%%) threshold: " UINTX_FORMAT "%%", 1464 byte_size_in_proper_unit(reclaimable_bytes), proper_unit_for_byte_size(reclaimable_bytes), 1465 reclaimable_percent, G1HeapWastePercent); 1466 break; 1467 } 1468 1469 double predicted_time_ms = predict_region_total_time_ms(hr, false); 1470 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); 1471 // Add regions to old set until we reach the minimum amount 1472 if (num_initial_regions < min_old_cset_length) { 1473 predicted_old_time_ms += predicted_time_ms; 1474 num_initial_regions++; 1475 // Record the number of regions added with no time remaining 1476 if (time_remaining_ms == 0.0) { 1477 num_expensive_regions++; 1478 } 1479 } else if (!check_time_remaining) { 1480 // In the non-auto-tuning case, we'll finish adding regions 1481 // to the CSet if we reach the minimum. 1482 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Region amount reached min)."); 1483 break; 1484 } else { 1485 // Keep adding regions to old set until we reach the optional threshold 1486 if (time_remaining_ms > optional_threshold_ms) { 1487 predicted_old_time_ms += predicted_time_ms; 1488 num_initial_regions++; 1489 } else if (time_remaining_ms > 0) { 1490 // Keep adding optional regions until time is up. 1491 assert(num_optional_regions < max_optional_regions, "Should not be possible."); 1492 predicted_optional_time_ms += predicted_time_ms; 1493 num_optional_regions++; 1494 } else { 1495 log_debug(gc, ergo, cset)("Finish adding old regions to collection set (Predicted time too high)."); 1496 break; 1497 } 1498 } 1499 hr = candidates->at(++candidate_idx); 1500 } 1501 if (hr == NULL) { 1502 log_debug(gc, ergo, cset)("Old candidate collection set empty."); 1503 } 1504 1505 if (num_expensive_regions > 0) { 1506 log_debug(gc, ergo, cset)("Added %u initial old regions to collection set although the predicted time was too high.", 1507 num_expensive_regions); 1508 } 1509 1510 log_debug(gc, ergo, cset)("Finish choosing collection set old regions. Initial: %u, optional: %u, " 1511 "predicted old time: %1.2fms, predicted optional time: %1.2fms, time remaining: %1.2f", 1512 num_initial_regions, num_optional_regions, 1513 predicted_initial_time_ms, predicted_optional_time_ms, time_remaining_ms); 1514 } 1515 1516 void G1Policy::calculate_optional_collection_set_regions(G1CollectionSetCandidates* candidates, 1517 uint const max_optional_regions, 1518 double time_remaining_ms, 1519 uint& num_optional_regions) { 1520 assert(_g1h->collector_state()->in_mixed_phase(), "Should only be called in mixed phase"); 1521 1522 num_optional_regions = 0; 1523 double prediction_ms = 0; 1524 uint candidate_idx = candidates->cur_idx(); 1525 1526 HeapRegion* r = candidates->at(candidate_idx); 1527 while (num_optional_regions < max_optional_regions) { 1528 assert(r != NULL, "Region must exist"); 1529 prediction_ms += predict_region_total_time_ms(r, false); 1530 1531 if (prediction_ms > time_remaining_ms) { 1532 log_debug(gc, ergo, cset)("Prediction %.3fms for region %u does not fit remaining time: %.3fms.", 1533 prediction_ms, r->hrm_index(), time_remaining_ms); 1534 break; 1535 } 1536 // This region will be included in the next optional evacuation. 1537 1538 time_remaining_ms -= prediction_ms; 1539 num_optional_regions++; 1540 r = candidates->at(++candidate_idx); 1541 } 1542 1543 log_debug(gc, ergo, cset)("Prepared %u regions out of %u for optional evacuation. Predicted time: %.3fms", 1544 num_optional_regions, max_optional_regions, prediction_ms); 1545 } 1546 1547 void G1Policy::transfer_survivors_to_cset(const G1SurvivorRegions* survivors) { 1548 note_start_adding_survivor_regions(); 1549 1550 HeapRegion* last = NULL; 1551 for (GrowableArrayIterator<HeapRegion*> it = survivors->regions()->begin(); 1552 it != survivors->regions()->end(); 1553 ++it) { 1554 HeapRegion* curr = *it; 1555 set_region_survivor(curr); 1556 1557 // The region is a non-empty survivor so let's add it to 1558 // the incremental collection set for the next evacuation 1559 // pause. 1560 _collection_set->add_survivor_regions(curr); 1561 1562 last = curr; 1563 } 1564 note_stop_adding_survivor_regions(); 1565 1566 // Don't clear the survivor list handles until the start of 1567 // the next evacuation pause - we need it in order to re-tag 1568 // the survivor regions from this evacuation pause as 'young' 1569 // at the start of the next. 1570 }