1 /* 2 * Copyright (c) 2001, 2010, 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_implementation/g1/concurrentG1Refine.hpp" 27 #include "gc_implementation/g1/concurrentMark.hpp" 28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 30 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 31 #include "gc_implementation/g1/heapRegionRemSet.hpp" 32 #include "gc_implementation/shared/gcPolicyCounters.hpp" 33 #include "runtime/arguments.hpp" 34 #include "runtime/java.hpp" 35 #include "runtime/mutexLocker.hpp" 36 #include "utilities/debug.hpp" 37 38 #define PREDICTIONS_VERBOSE 0 39 40 // <NEW PREDICTION> 41 42 // Different defaults for different number of GC threads 43 // They were chosen by running GCOld and SPECjbb on debris with different 44 // numbers of GC threads and choosing them based on the results 45 46 // all the same 47 static double rs_length_diff_defaults[] = { 48 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 49 }; 50 51 static double cost_per_card_ms_defaults[] = { 52 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 53 }; 54 55 // all the same 56 static double fully_young_cards_per_entry_ratio_defaults[] = { 57 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 58 }; 59 60 static double cost_per_entry_ms_defaults[] = { 61 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 62 }; 63 64 static double cost_per_byte_ms_defaults[] = { 65 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 66 }; 67 68 // these should be pretty consistent 69 static double constant_other_time_ms_defaults[] = { 70 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 71 }; 72 73 74 static double young_other_cost_per_region_ms_defaults[] = { 75 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 76 }; 77 78 static double non_young_other_cost_per_region_ms_defaults[] = { 79 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 80 }; 81 82 // </NEW PREDICTION> 83 84 G1CollectorPolicy::G1CollectorPolicy() : 85 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads() 86 ? ParallelGCThreads : 1), 87 88 89 _n_pauses(0), 90 _recent_CH_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 91 _recent_G1_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 92 _recent_evac_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 93 _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 94 _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)), 95 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 96 _all_pause_times_ms(new NumberSeq()), 97 _stop_world_start(0.0), 98 _all_stop_world_times_ms(new NumberSeq()), 99 _all_yield_times_ms(new NumberSeq()), 100 101 _all_mod_union_times_ms(new NumberSeq()), 102 103 _summary(new Summary()), 104 105 #ifndef PRODUCT 106 _cur_clear_ct_time_ms(0.0), 107 _min_clear_cc_time_ms(-1.0), 108 _max_clear_cc_time_ms(-1.0), 109 _cur_clear_cc_time_ms(0.0), 110 _cum_clear_cc_time_ms(0.0), 111 _num_cc_clears(0L), 112 #endif 113 114 _region_num_young(0), 115 _region_num_tenured(0), 116 _prev_region_num_young(0), 117 _prev_region_num_tenured(0), 118 119 _aux_num(10), 120 _all_aux_times_ms(new NumberSeq[_aux_num]), 121 _cur_aux_start_times_ms(new double[_aux_num]), 122 _cur_aux_times_ms(new double[_aux_num]), 123 _cur_aux_times_set(new bool[_aux_num]), 124 125 _concurrent_mark_init_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 126 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 127 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 128 129 // <NEW PREDICTION> 130 131 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 132 _prev_collection_pause_end_ms(0.0), 133 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 134 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 135 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 136 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 137 _partially_young_cards_per_entry_ratio_seq( 138 new TruncatedSeq(TruncatedSeqLength)), 139 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 140 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 141 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 142 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 143 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 144 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 145 _non_young_other_cost_per_region_ms_seq( 146 new TruncatedSeq(TruncatedSeqLength)), 147 148 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 149 _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 150 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 151 152 _pause_time_target_ms((double) MaxGCPauseMillis), 153 154 // </NEW PREDICTION> 155 156 _in_young_gc_mode(false), 157 _full_young_gcs(true), 158 _full_young_pause_num(0), 159 _partial_young_pause_num(0), 160 161 _during_marking(false), 162 _in_marking_window(false), 163 _in_marking_window_im(false), 164 165 _known_garbage_ratio(0.0), 166 _known_garbage_bytes(0), 167 168 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)), 169 170 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)), 171 172 _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)), 173 _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)), 174 175 _recent_avg_pause_time_ratio(0.0), 176 _num_markings(0), 177 _n_marks(0), 178 _n_pauses_at_mark_end(0), 179 180 _all_full_gc_times_ms(new NumberSeq()), 181 182 // G1PausesBtwnConcMark defaults to -1 183 // so the hack is to do the cast QQQ FIXME 184 _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark), 185 _n_marks_since_last_pause(0), 186 _initiate_conc_mark_if_possible(false), 187 _during_initial_mark_pause(false), 188 _should_revert_to_full_young_gcs(false), 189 _last_full_young_gc(false), 190 191 _prev_collection_pause_used_at_end_bytes(0), 192 193 _collection_set(NULL), 194 _collection_set_size(0), 195 _collection_set_bytes_used_before(0), 196 197 // Incremental CSet attributes 198 _inc_cset_build_state(Inactive), 199 _inc_cset_head(NULL), 200 _inc_cset_tail(NULL), 201 _inc_cset_size(0), 202 _inc_cset_young_index(0), 203 _inc_cset_bytes_used_before(0), 204 _inc_cset_max_finger(NULL), 205 _inc_cset_recorded_young_bytes(0), 206 _inc_cset_recorded_rs_lengths(0), 207 _inc_cset_predicted_elapsed_time_ms(0.0), 208 _inc_cset_predicted_bytes_to_copy(0), 209 210 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 211 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 212 #endif // _MSC_VER 213 214 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", 215 G1YoungSurvRateNumRegionsSummary)), 216 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", 217 G1YoungSurvRateNumRegionsSummary)), 218 // add here any more surv rate groups 219 _recorded_survivor_regions(0), 220 _recorded_survivor_head(NULL), 221 _recorded_survivor_tail(NULL), 222 _survivors_age_table(true), 223 224 _gc_overhead_perc(0.0) 225 226 { 227 // Set up the region size and associated fields. Given that the 228 // policy is created before the heap, we have to set this up here, 229 // so it's done as soon as possible. 230 HeapRegion::setup_heap_region_size(Arguments::min_heap_size()); 231 HeapRegionRemSet::setup_remset_size(); 232 233 // Verify PLAB sizes 234 const uint region_size = HeapRegion::GrainWords; 235 if (YoungPLABSize > region_size || OldPLABSize > region_size) { 236 char buffer[128]; 237 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most %u", 238 OldPLABSize > region_size ? "Old" : "Young", region_size); 239 vm_exit_during_initialization(buffer); 240 } 241 242 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 243 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 244 245 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads]; 246 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads]; 247 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads]; 248 249 _par_last_update_rs_times_ms = new double[_parallel_gc_threads]; 250 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; 251 252 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; 253 254 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads]; 255 256 _par_last_termination_times_ms = new double[_parallel_gc_threads]; 257 _par_last_termination_attempts = new double[_parallel_gc_threads]; 258 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads]; 259 260 // start conservatively 261 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis; 262 263 // <NEW PREDICTION> 264 265 int index; 266 if (ParallelGCThreads == 0) 267 index = 0; 268 else if (ParallelGCThreads > 8) 269 index = 7; 270 else 271 index = ParallelGCThreads - 1; 272 273 _pending_card_diff_seq->add(0.0); 274 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 275 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 276 _fully_young_cards_per_entry_ratio_seq->add( 277 fully_young_cards_per_entry_ratio_defaults[index]); 278 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 279 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 280 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 281 _young_other_cost_per_region_ms_seq->add( 282 young_other_cost_per_region_ms_defaults[index]); 283 _non_young_other_cost_per_region_ms_seq->add( 284 non_young_other_cost_per_region_ms_defaults[index]); 285 286 // </NEW PREDICTION> 287 288 // Below, we might need to calculate the pause time target based on 289 // the pause interval. When we do so we are going to give G1 maximum 290 // flexibility and allow it to do pauses when it needs to. So, we'll 291 // arrange that the pause interval to be pause time target + 1 to 292 // ensure that a) the pause time target is maximized with respect to 293 // the pause interval and b) we maintain the invariant that pause 294 // time target < pause interval. If the user does not want this 295 // maximum flexibility, they will have to set the pause interval 296 // explicitly. 297 298 // First make sure that, if either parameter is set, its value is 299 // reasonable. 300 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 301 if (MaxGCPauseMillis < 1) { 302 vm_exit_during_initialization("MaxGCPauseMillis should be " 303 "greater than 0"); 304 } 305 } 306 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 307 if (GCPauseIntervalMillis < 1) { 308 vm_exit_during_initialization("GCPauseIntervalMillis should be " 309 "greater than 0"); 310 } 311 } 312 313 // Then, if the pause time target parameter was not set, set it to 314 // the default value. 315 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 316 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 317 // The default pause time target in G1 is 200ms 318 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 319 } else { 320 // We do not allow the pause interval to be set without the 321 // pause time target 322 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 323 "without setting MaxGCPauseMillis"); 324 } 325 } 326 327 // Then, if the interval parameter was not set, set it according to 328 // the pause time target (this will also deal with the case when the 329 // pause time target is the default value). 330 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 331 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 332 } 333 334 // Finally, make sure that the two parameters are consistent. 335 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 336 char buffer[256]; 337 jio_snprintf(buffer, 256, 338 "MaxGCPauseMillis (%u) should be less than " 339 "GCPauseIntervalMillis (%u)", 340 MaxGCPauseMillis, GCPauseIntervalMillis); 341 vm_exit_during_initialization(buffer); 342 } 343 344 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 345 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 346 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 347 _sigma = (double) G1ConfidencePercent / 100.0; 348 349 // start conservatively (around 50ms is about right) 350 _concurrent_mark_init_times_ms->add(0.05); 351 _concurrent_mark_remark_times_ms->add(0.05); 352 _concurrent_mark_cleanup_times_ms->add(0.20); 353 _tenuring_threshold = MaxTenuringThreshold; 354 355 // if G1FixedSurvivorSpaceSize is 0 which means the size is not 356 // fixed, then _max_survivor_regions will be calculated at 357 // calculate_young_list_target_length during initialization 358 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes; 359 360 assert(GCTimeRatio > 0, 361 "we should have set it to a default value set_g1_gc_flags() " 362 "if a user set it to 0"); 363 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 364 365 initialize_all(); 366 } 367 368 // Increment "i", mod "len" 369 static void inc_mod(int& i, int len) { 370 i++; if (i == len) i = 0; 371 } 372 373 void G1CollectorPolicy::initialize_flags() { 374 set_min_alignment(HeapRegion::GrainBytes); 375 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name())); 376 if (SurvivorRatio < 1) { 377 vm_exit_during_initialization("Invalid survivor ratio specified"); 378 } 379 CollectorPolicy::initialize_flags(); 380 } 381 382 // The easiest way to deal with the parsing of the NewSize / 383 // MaxNewSize / etc. parameteres is to re-use the code in the 384 // TwoGenerationCollectorPolicy class. This is similar to what 385 // ParallelScavenge does with its GenerationSizer class (see 386 // ParallelScavengeHeap::initialize()). We might change this in the 387 // future, but it's a good start. 388 class G1YoungGenSizer : public TwoGenerationCollectorPolicy { 389 size_t size_to_region_num(size_t byte_size) { 390 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes); 391 } 392 393 public: 394 G1YoungGenSizer() { 395 initialize_flags(); 396 initialize_size_info(); 397 } 398 399 size_t min_young_region_num() { 400 return size_to_region_num(_min_gen0_size); 401 } 402 size_t initial_young_region_num() { 403 return size_to_region_num(_initial_gen0_size); 404 } 405 size_t max_young_region_num() { 406 return size_to_region_num(_max_gen0_size); 407 } 408 }; 409 410 void G1CollectorPolicy::init() { 411 // Set aside an initial future to_space. 412 _g1 = G1CollectedHeap::heap(); 413 414 assert(Heap_lock->owned_by_self(), "Locking discipline."); 415 416 initialize_gc_policy_counters(); 417 418 if (G1Gen) { 419 _in_young_gc_mode = true; 420 421 G1YoungGenSizer sizer; 422 size_t initial_region_num = sizer.initial_young_region_num(); 423 424 if (UseAdaptiveSizePolicy) { 425 set_adaptive_young_list_length(true); 426 _young_list_fixed_length = 0; 427 } else { 428 set_adaptive_young_list_length(false); 429 _young_list_fixed_length = initial_region_num; 430 } 431 _free_regions_at_end_of_collection = _g1->free_regions(); 432 calculate_young_list_min_length(); 433 guarantee( _young_list_min_length == 0, "invariant, not enough info" ); 434 calculate_young_list_target_length(); 435 } else { 436 _young_list_fixed_length = 0; 437 _in_young_gc_mode = false; 438 } 439 440 // We may immediately start allocating regions and placing them on the 441 // collection set list. Initialize the per-collection set info 442 start_incremental_cset_building(); 443 } 444 445 // Create the jstat counters for the policy. 446 void G1CollectorPolicy::initialize_gc_policy_counters() 447 { 448 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 2 + G1Gen); 449 } 450 451 void G1CollectorPolicy::calculate_young_list_min_length() { 452 _young_list_min_length = 0; 453 454 if (!adaptive_young_list_length()) 455 return; 456 457 if (_alloc_rate_ms_seq->num() > 3) { 458 double now_sec = os::elapsedTime(); 459 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 460 double alloc_rate_ms = predict_alloc_rate_ms(); 461 int min_regions = (int) ceil(alloc_rate_ms * when_ms); 462 int current_region_num = (int) _g1->young_list()->length(); 463 _young_list_min_length = min_regions + current_region_num; 464 } 465 } 466 467 void G1CollectorPolicy::calculate_young_list_target_length() { 468 if (adaptive_young_list_length()) { 469 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); 470 calculate_young_list_target_length(rs_lengths); 471 } else { 472 if (full_young_gcs()) 473 _young_list_target_length = _young_list_fixed_length; 474 else 475 _young_list_target_length = _young_list_fixed_length / 2; 476 477 _young_list_target_length = MAX2(_young_list_target_length, (size_t)1); 478 } 479 calculate_survivors_policy(); 480 } 481 482 void G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths) { 483 guarantee( adaptive_young_list_length(), "pre-condition" ); 484 guarantee( !_in_marking_window || !_last_full_young_gc, "invariant" ); 485 486 double start_time_sec = os::elapsedTime(); 487 size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1ReservePercent); 488 min_reserve_perc = MIN2((size_t) 50, min_reserve_perc); 489 size_t reserve_regions = 490 (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0); 491 492 if (full_young_gcs() && _free_regions_at_end_of_collection > 0) { 493 // we are in fully-young mode and there are free regions in the heap 494 495 double survivor_regions_evac_time = 496 predict_survivor_regions_evac_time(); 497 498 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 499 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 500 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 501 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 502 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, scanned_cards) 503 + survivor_regions_evac_time; 504 505 // the result 506 size_t final_young_length = 0; 507 508 size_t init_free_regions = 509 MAX2((size_t)0, _free_regions_at_end_of_collection - reserve_regions); 510 511 // if we're still under the pause target... 512 if (base_time_ms <= target_pause_time_ms) { 513 // We make sure that the shortest young length that makes sense 514 // fits within the target pause time. 515 size_t min_young_length = 1; 516 517 if (predict_will_fit(min_young_length, base_time_ms, 518 init_free_regions, target_pause_time_ms)) { 519 // The shortest young length will fit within the target pause time; 520 // we'll now check whether the absolute maximum number of young 521 // regions will fit in the target pause time. If not, we'll do 522 // a binary search between min_young_length and max_young_length 523 size_t abs_max_young_length = _free_regions_at_end_of_collection - 1; 524 size_t max_young_length = abs_max_young_length; 525 526 if (max_young_length > min_young_length) { 527 // Let's check if the initial max young length will fit within the 528 // target pause. If so then there is no need to search for a maximal 529 // young length - we'll return the initial maximum 530 531 if (predict_will_fit(max_young_length, base_time_ms, 532 init_free_regions, target_pause_time_ms)) { 533 // The maximum young length will satisfy the target pause time. 534 // We are done so set min young length to this maximum length. 535 // The code after the loop will then set final_young_length using 536 // the value cached in the minimum length. 537 min_young_length = max_young_length; 538 } else { 539 // The maximum possible number of young regions will not fit within 540 // the target pause time so let's search.... 541 542 size_t diff = (max_young_length - min_young_length) / 2; 543 max_young_length = min_young_length + diff; 544 545 while (max_young_length > min_young_length) { 546 if (predict_will_fit(max_young_length, base_time_ms, 547 init_free_regions, target_pause_time_ms)) { 548 549 // The current max young length will fit within the target 550 // pause time. Note we do not exit the loop here. By setting 551 // min = max, and then increasing the max below means that 552 // we will continue searching for an upper bound in the 553 // range [max..max+diff] 554 min_young_length = max_young_length; 555 } 556 diff = (max_young_length - min_young_length) / 2; 557 max_young_length = min_young_length + diff; 558 } 559 // the above loop found a maximal young length that will fit 560 // within the target pause time. 561 } 562 assert(min_young_length <= abs_max_young_length, "just checking"); 563 } 564 final_young_length = min_young_length; 565 } 566 } 567 // and we're done! 568 569 // we should have at least one region in the target young length 570 _young_list_target_length = 571 MAX2((size_t) 1, final_young_length + _recorded_survivor_regions); 572 573 // let's keep an eye of how long we spend on this calculation 574 // right now, I assume that we'll print it when we need it; we 575 // should really adde it to the breakdown of a pause 576 double end_time_sec = os::elapsedTime(); 577 double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0; 578 579 #ifdef TRACE_CALC_YOUNG_LENGTH 580 // leave this in for debugging, just in case 581 gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT ", " 582 "elapsed %1.2lf ms, (%s%s) " SIZE_FORMAT SIZE_FORMAT, 583 target_pause_time_ms, 584 _young_list_target_length 585 elapsed_time_ms, 586 full_young_gcs() ? "full" : "partial", 587 during_initial_mark_pause() ? " i-m" : "", 588 _in_marking_window, 589 _in_marking_window_im); 590 #endif // TRACE_CALC_YOUNG_LENGTH 591 592 if (_young_list_target_length < _young_list_min_length) { 593 // bummer; this means that, if we do a pause when the maximal 594 // length dictates, we'll violate the pause spacing target (the 595 // min length was calculate based on the application's current 596 // alloc rate); 597 598 // so, we have to bite the bullet, and allocate the minimum 599 // number. We'll violate our target, but we just can't meet it. 600 601 #ifdef TRACE_CALC_YOUNG_LENGTH 602 // leave this in for debugging, just in case 603 gclog_or_tty->print_cr("adjusted target length from " 604 SIZE_FORMAT " to " SIZE_FORMAT, 605 _young_list_target_length, _young_list_min_length); 606 #endif // TRACE_CALC_YOUNG_LENGTH 607 608 _young_list_target_length = _young_list_min_length; 609 } 610 } else { 611 // we are in a partially-young mode or we've run out of regions (due 612 // to evacuation failure) 613 614 #ifdef TRACE_CALC_YOUNG_LENGTH 615 // leave this in for debugging, just in case 616 gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT 617 _young_list_min_length); 618 #endif // TRACE_CALC_YOUNG_LENGTH 619 // we'll do the pause as soon as possible by choosing the minimum 620 _young_list_target_length = 621 MAX2(_young_list_min_length, (size_t) 1); 622 } 623 624 _rs_lengths_prediction = rs_lengths; 625 } 626 627 // This is used by: calculate_young_list_target_length(rs_length). It 628 // returns true iff: 629 // the predicted pause time for the given young list will not overflow 630 // the target pause time 631 // and: 632 // the predicted amount of surviving data will not overflow the 633 // the amount of free space available for survivor regions. 634 // 635 bool 636 G1CollectorPolicy::predict_will_fit(size_t young_length, 637 double base_time_ms, 638 size_t init_free_regions, 639 double target_pause_time_ms) { 640 641 if (young_length >= init_free_regions) 642 // end condition 1: not enough space for the young regions 643 return false; 644 645 double accum_surv_rate_adj = 0.0; 646 double accum_surv_rate = 647 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj; 648 649 size_t bytes_to_copy = 650 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 651 652 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 653 654 double young_other_time_ms = 655 predict_young_other_time_ms(young_length); 656 657 double pause_time_ms = 658 base_time_ms + copy_time_ms + young_other_time_ms; 659 660 if (pause_time_ms > target_pause_time_ms) 661 // end condition 2: over the target pause time 662 return false; 663 664 size_t free_bytes = 665 (init_free_regions - young_length) * HeapRegion::GrainBytes; 666 667 if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes) 668 // end condition 3: out of to-space (conservatively) 669 return false; 670 671 // success! 672 return true; 673 } 674 675 double G1CollectorPolicy::predict_survivor_regions_evac_time() { 676 double survivor_regions_evac_time = 0.0; 677 for (HeapRegion * r = _recorded_survivor_head; 678 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 679 r = r->get_next_young_region()) { 680 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true); 681 } 682 return survivor_regions_evac_time; 683 } 684 685 void G1CollectorPolicy::check_prediction_validity() { 686 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 687 688 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 689 if (rs_lengths > _rs_lengths_prediction) { 690 // add 10% to avoid having to recalculate often 691 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 692 calculate_young_list_target_length(rs_lengths_prediction); 693 } 694 } 695 696 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 697 bool is_tlab, 698 bool* gc_overhead_limit_was_exceeded) { 699 guarantee(false, "Not using this policy feature yet."); 700 return NULL; 701 } 702 703 // This method controls how a collector handles one or more 704 // of its generations being fully allocated. 705 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 706 bool is_tlab) { 707 guarantee(false, "Not using this policy feature yet."); 708 return NULL; 709 } 710 711 712 #ifndef PRODUCT 713 bool G1CollectorPolicy::verify_young_ages() { 714 HeapRegion* head = _g1->young_list()->first_region(); 715 return 716 verify_young_ages(head, _short_lived_surv_rate_group); 717 // also call verify_young_ages on any additional surv rate groups 718 } 719 720 bool 721 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 722 SurvRateGroup *surv_rate_group) { 723 guarantee( surv_rate_group != NULL, "pre-condition" ); 724 725 const char* name = surv_rate_group->name(); 726 bool ret = true; 727 int prev_age = -1; 728 729 for (HeapRegion* curr = head; 730 curr != NULL; 731 curr = curr->get_next_young_region()) { 732 SurvRateGroup* group = curr->surv_rate_group(); 733 if (group == NULL && !curr->is_survivor()) { 734 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 735 ret = false; 736 } 737 738 if (surv_rate_group == group) { 739 int age = curr->age_in_surv_rate_group(); 740 741 if (age < 0) { 742 gclog_or_tty->print_cr("## %s: encountered negative age", name); 743 ret = false; 744 } 745 746 if (age <= prev_age) { 747 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 748 "(%d, %d)", name, age, prev_age); 749 ret = false; 750 } 751 prev_age = age; 752 } 753 } 754 755 return ret; 756 } 757 #endif // PRODUCT 758 759 void G1CollectorPolicy::record_full_collection_start() { 760 _cur_collection_start_sec = os::elapsedTime(); 761 // Release the future to-space so that it is available for compaction into. 762 _g1->set_full_collection(); 763 } 764 765 void G1CollectorPolicy::record_full_collection_end() { 766 // Consider this like a collection pause for the purposes of allocation 767 // since last pause. 768 double end_sec = os::elapsedTime(); 769 double full_gc_time_sec = end_sec - _cur_collection_start_sec; 770 double full_gc_time_ms = full_gc_time_sec * 1000.0; 771 772 _all_full_gc_times_ms->add(full_gc_time_ms); 773 774 update_recent_gc_times(end_sec, full_gc_time_ms); 775 776 _g1->clear_full_collection(); 777 778 // "Nuke" the heuristics that control the fully/partially young GC 779 // transitions and make sure we start with fully young GCs after the 780 // Full GC. 781 set_full_young_gcs(true); 782 _last_full_young_gc = false; 783 _should_revert_to_full_young_gcs = false; 784 clear_initiate_conc_mark_if_possible(); 785 clear_during_initial_mark_pause(); 786 _known_garbage_bytes = 0; 787 _known_garbage_ratio = 0.0; 788 _in_marking_window = false; 789 _in_marking_window_im = false; 790 791 _short_lived_surv_rate_group->start_adding_regions(); 792 // also call this on any additional surv rate groups 793 794 record_survivor_regions(0, NULL, NULL); 795 796 _prev_region_num_young = _region_num_young; 797 _prev_region_num_tenured = _region_num_tenured; 798 799 _free_regions_at_end_of_collection = _g1->free_regions(); 800 // Reset survivors SurvRateGroup. 801 _survivor_surv_rate_group->reset(); 802 calculate_young_list_min_length(); 803 calculate_young_list_target_length(); 804 } 805 806 void G1CollectorPolicy::record_before_bytes(size_t bytes) { 807 _bytes_in_to_space_before_gc += bytes; 808 } 809 810 void G1CollectorPolicy::record_after_bytes(size_t bytes) { 811 _bytes_in_to_space_after_gc += bytes; 812 } 813 814 void G1CollectorPolicy::record_stop_world_start() { 815 _stop_world_start = os::elapsedTime(); 816 } 817 818 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, 819 size_t start_used) { 820 if (PrintGCDetails) { 821 gclog_or_tty->stamp(PrintGCTimeStamps); 822 gclog_or_tty->print("[GC pause"); 823 if (in_young_gc_mode()) 824 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial"); 825 } 826 827 assert(_g1->used_regions() == _g1->recalculate_used_regions(), 828 "sanity"); 829 assert(_g1->used() == _g1->recalculate_used(), "sanity"); 830 831 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 832 _all_stop_world_times_ms->add(s_w_t_ms); 833 _stop_world_start = 0.0; 834 835 _cur_collection_start_sec = start_time_sec; 836 _cur_collection_pause_used_at_start_bytes = start_used; 837 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); 838 _pending_cards = _g1->pending_card_num(); 839 _max_pending_cards = _g1->max_pending_card_num(); 840 841 _bytes_in_to_space_before_gc = 0; 842 _bytes_in_to_space_after_gc = 0; 843 _bytes_in_collection_set_before_gc = 0; 844 845 #ifdef DEBUG 846 // initialise these to something well known so that we can spot 847 // if they are not set properly 848 849 for (int i = 0; i < _parallel_gc_threads; ++i) { 850 _par_last_gc_worker_start_times_ms[i] = -1234.0; 851 _par_last_ext_root_scan_times_ms[i] = -1234.0; 852 _par_last_mark_stack_scan_times_ms[i] = -1234.0; 853 _par_last_update_rs_times_ms[i] = -1234.0; 854 _par_last_update_rs_processed_buffers[i] = -1234.0; 855 _par_last_scan_rs_times_ms[i] = -1234.0; 856 _par_last_obj_copy_times_ms[i] = -1234.0; 857 _par_last_termination_times_ms[i] = -1234.0; 858 _par_last_termination_attempts[i] = -1234.0; 859 _par_last_gc_worker_end_times_ms[i] = -1234.0; 860 } 861 #endif 862 863 for (int i = 0; i < _aux_num; ++i) { 864 _cur_aux_times_ms[i] = 0.0; 865 _cur_aux_times_set[i] = false; 866 } 867 868 _satb_drain_time_set = false; 869 _last_satb_drain_processed_buffers = -1; 870 871 if (in_young_gc_mode()) 872 _last_young_gc_full = false; 873 874 // do that for any other surv rate groups 875 _short_lived_surv_rate_group->stop_adding_regions(); 876 _survivors_age_table.clear(); 877 878 assert( verify_young_ages(), "region age verification" ); 879 } 880 881 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) { 882 _mark_closure_time_ms = mark_closure_time_ms; 883 } 884 885 void G1CollectorPolicy::record_concurrent_mark_init_start() { 886 _mark_init_start_sec = os::elapsedTime(); 887 guarantee(!in_young_gc_mode(), "should not do be here in young GC mode"); 888 } 889 890 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double 891 mark_init_elapsed_time_ms) { 892 _during_marking = true; 893 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); 894 clear_during_initial_mark_pause(); 895 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 896 } 897 898 void G1CollectorPolicy::record_concurrent_mark_init_end() { 899 double end_time_sec = os::elapsedTime(); 900 double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0; 901 _concurrent_mark_init_times_ms->add(elapsed_time_ms); 902 record_concurrent_mark_init_end_pre(elapsed_time_ms); 903 904 _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true); 905 } 906 907 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 908 _mark_remark_start_sec = os::elapsedTime(); 909 _during_marking = false; 910 } 911 912 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 913 double end_time_sec = os::elapsedTime(); 914 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 915 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 916 _cur_mark_stop_world_time_ms += elapsed_time_ms; 917 _prev_collection_pause_end_ms += elapsed_time_ms; 918 919 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); 920 } 921 922 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 923 _mark_cleanup_start_sec = os::elapsedTime(); 924 } 925 926 void 927 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes, 928 size_t max_live_bytes) { 929 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes); 930 record_concurrent_mark_cleanup_end_work2(); 931 } 932 933 void 934 G1CollectorPolicy:: 935 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes, 936 size_t max_live_bytes) { 937 if (_n_marks < 2) _n_marks++; 938 if (G1PolicyVerbose > 0) 939 gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB " 940 " (of " SIZE_FORMAT " MB heap).", 941 max_live_bytes/M, _g1->capacity()/M); 942 } 943 944 // The important thing about this is that it includes "os::elapsedTime". 945 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() { 946 double end_time_sec = os::elapsedTime(); 947 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0; 948 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 949 _cur_mark_stop_world_time_ms += elapsed_time_ms; 950 _prev_collection_pause_end_ms += elapsed_time_ms; 951 952 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true); 953 954 _num_markings++; 955 956 // We did a marking, so reset the "since_last_mark" variables. 957 double considerConcMarkCost = 1.0; 958 // If there are available processors, concurrent activity is free... 959 if (Threads::number_of_non_daemon_threads() * 2 < 960 os::active_processor_count()) { 961 considerConcMarkCost = 0.0; 962 } 963 _n_pauses_at_mark_end = _n_pauses; 964 _n_marks_since_last_pause++; 965 } 966 967 void 968 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 969 if (in_young_gc_mode()) { 970 _should_revert_to_full_young_gcs = false; 971 _last_full_young_gc = true; 972 _in_marking_window = false; 973 if (adaptive_young_list_length()) 974 calculate_young_list_target_length(); 975 } 976 } 977 978 void G1CollectorPolicy::record_concurrent_pause() { 979 if (_stop_world_start > 0.0) { 980 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 981 _all_yield_times_ms->add(yield_ms); 982 } 983 } 984 985 void G1CollectorPolicy::record_concurrent_pause_end() { 986 } 987 988 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() { 989 _cur_CH_strong_roots_end_sec = os::elapsedTime(); 990 _cur_CH_strong_roots_dur_ms = 991 (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0; 992 } 993 994 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() { 995 _cur_G1_strong_roots_end_sec = os::elapsedTime(); 996 _cur_G1_strong_roots_dur_ms = 997 (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0; 998 } 999 1000 template<class T> 1001 T sum_of(T* sum_arr, int start, int n, int N) { 1002 T sum = (T)0; 1003 for (int i = 0; i < n; i++) { 1004 int j = (start + i) % N; 1005 sum += sum_arr[j]; 1006 } 1007 return sum; 1008 } 1009 1010 void G1CollectorPolicy::print_par_stats(int level, 1011 const char* str, 1012 double* data, 1013 bool summary) { 1014 double min = data[0], max = data[0]; 1015 double total = 0.0; 1016 int j; 1017 for (j = 0; j < level; ++j) 1018 gclog_or_tty->print(" "); 1019 gclog_or_tty->print("[%s (ms):", str); 1020 for (uint i = 0; i < ParallelGCThreads; ++i) { 1021 double val = data[i]; 1022 if (val < min) 1023 min = val; 1024 if (val > max) 1025 max = val; 1026 total += val; 1027 gclog_or_tty->print(" %3.1lf", val); 1028 } 1029 if (summary) { 1030 gclog_or_tty->print_cr(""); 1031 double avg = total / (double) ParallelGCThreads; 1032 gclog_or_tty->print(" "); 1033 for (j = 0; j < level; ++j) 1034 gclog_or_tty->print(" "); 1035 gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf", 1036 avg, min, max); 1037 } 1038 gclog_or_tty->print_cr("]"); 1039 } 1040 1041 void G1CollectorPolicy::print_par_sizes(int level, 1042 const char* str, 1043 double* data, 1044 bool summary) { 1045 double min = data[0], max = data[0]; 1046 double total = 0.0; 1047 int j; 1048 for (j = 0; j < level; ++j) 1049 gclog_or_tty->print(" "); 1050 gclog_or_tty->print("[%s :", str); 1051 for (uint i = 0; i < ParallelGCThreads; ++i) { 1052 double val = data[i]; 1053 if (val < min) 1054 min = val; 1055 if (val > max) 1056 max = val; 1057 total += val; 1058 gclog_or_tty->print(" %d", (int) val); 1059 } 1060 if (summary) { 1061 gclog_or_tty->print_cr(""); 1062 double avg = total / (double) ParallelGCThreads; 1063 gclog_or_tty->print(" "); 1064 for (j = 0; j < level; ++j) 1065 gclog_or_tty->print(" "); 1066 gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d", 1067 (int)total, (int)avg, (int)min, (int)max); 1068 } 1069 gclog_or_tty->print_cr("]"); 1070 } 1071 1072 void G1CollectorPolicy::print_stats (int level, 1073 const char* str, 1074 double value) { 1075 for (int j = 0; j < level; ++j) 1076 gclog_or_tty->print(" "); 1077 gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value); 1078 } 1079 1080 void G1CollectorPolicy::print_stats (int level, 1081 const char* str, 1082 int value) { 1083 for (int j = 0; j < level; ++j) 1084 gclog_or_tty->print(" "); 1085 gclog_or_tty->print_cr("[%s: %d]", str, value); 1086 } 1087 1088 double G1CollectorPolicy::avg_value (double* data) { 1089 if (G1CollectedHeap::use_parallel_gc_threads()) { 1090 double ret = 0.0; 1091 for (uint i = 0; i < ParallelGCThreads; ++i) 1092 ret += data[i]; 1093 return ret / (double) ParallelGCThreads; 1094 } else { 1095 return data[0]; 1096 } 1097 } 1098 1099 double G1CollectorPolicy::max_value (double* data) { 1100 if (G1CollectedHeap::use_parallel_gc_threads()) { 1101 double ret = data[0]; 1102 for (uint i = 1; i < ParallelGCThreads; ++i) 1103 if (data[i] > ret) 1104 ret = data[i]; 1105 return ret; 1106 } else { 1107 return data[0]; 1108 } 1109 } 1110 1111 double G1CollectorPolicy::sum_of_values (double* data) { 1112 if (G1CollectedHeap::use_parallel_gc_threads()) { 1113 double sum = 0.0; 1114 for (uint i = 0; i < ParallelGCThreads; i++) 1115 sum += data[i]; 1116 return sum; 1117 } else { 1118 return data[0]; 1119 } 1120 } 1121 1122 double G1CollectorPolicy::max_sum (double* data1, 1123 double* data2) { 1124 double ret = data1[0] + data2[0]; 1125 1126 if (G1CollectedHeap::use_parallel_gc_threads()) { 1127 for (uint i = 1; i < ParallelGCThreads; ++i) { 1128 double data = data1[i] + data2[i]; 1129 if (data > ret) 1130 ret = data; 1131 } 1132 } 1133 return ret; 1134 } 1135 1136 // Anything below that is considered to be zero 1137 #define MIN_TIMER_GRANULARITY 0.0000001 1138 1139 void G1CollectorPolicy::record_collection_pause_end() { 1140 double end_time_sec = os::elapsedTime(); 1141 double elapsed_ms = _last_pause_time_ms; 1142 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1143 double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0; 1144 size_t rs_size = 1145 _cur_collection_pause_used_regions_at_start - collection_set_size(); 1146 size_t cur_used_bytes = _g1->used(); 1147 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1148 bool last_pause_included_initial_mark = false; 1149 bool update_stats = !_g1->evacuation_failed(); 1150 1151 #ifndef PRODUCT 1152 if (G1YoungSurvRateVerbose) { 1153 gclog_or_tty->print_cr(""); 1154 _short_lived_surv_rate_group->print(); 1155 // do that for any other surv rate groups too 1156 } 1157 #endif // PRODUCT 1158 1159 if (in_young_gc_mode()) { 1160 last_pause_included_initial_mark = during_initial_mark_pause(); 1161 if (last_pause_included_initial_mark) 1162 record_concurrent_mark_init_end_pre(0.0); 1163 1164 size_t min_used_targ = 1165 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 1166 1167 1168 if (!_g1->mark_in_progress() && !_last_full_young_gc) { 1169 assert(!last_pause_included_initial_mark, "invariant"); 1170 if (cur_used_bytes > min_used_targ && 1171 cur_used_bytes > _prev_collection_pause_used_at_end_bytes) { 1172 assert(!during_initial_mark_pause(), "we should not see this here"); 1173 1174 // Note: this might have already been set, if during the last 1175 // pause we decided to start a cycle but at the beginning of 1176 // this pause we decided to postpone it. That's OK. 1177 set_initiate_conc_mark_if_possible(); 1178 } 1179 } 1180 1181 _prev_collection_pause_used_at_end_bytes = cur_used_bytes; 1182 } 1183 1184 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, 1185 end_time_sec, false); 1186 1187 guarantee(_cur_collection_pause_used_regions_at_start >= 1188 collection_set_size(), 1189 "Negative RS size?"); 1190 1191 // This assert is exempted when we're doing parallel collection pauses, 1192 // because the fragmentation caused by the parallel GC allocation buffers 1193 // can lead to more memory being used during collection than was used 1194 // before. Best leave this out until the fragmentation problem is fixed. 1195 // Pauses in which evacuation failed can also lead to negative 1196 // collections, since no space is reclaimed from a region containing an 1197 // object whose evacuation failed. 1198 // Further, we're now always doing parallel collection. But I'm still 1199 // leaving this here as a placeholder for a more precise assertion later. 1200 // (DLD, 10/05.) 1201 assert((true || parallel) // Always using GC LABs now. 1202 || _g1->evacuation_failed() 1203 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes, 1204 "Negative collection"); 1205 1206 size_t freed_bytes = 1207 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; 1208 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; 1209 1210 double survival_fraction = 1211 (double)surviving_bytes/ 1212 (double)_collection_set_bytes_used_before; 1213 1214 _n_pauses++; 1215 1216 if (update_stats) { 1217 _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms); 1218 _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms); 1219 _recent_evac_times_ms->add(evac_ms); 1220 _recent_pause_times_ms->add(elapsed_ms); 1221 1222 _recent_rs_sizes->add(rs_size); 1223 1224 // We exempt parallel collection from this check because Alloc Buffer 1225 // fragmentation can produce negative collections. Same with evac 1226 // failure. 1227 // Further, we're now always doing parallel collection. But I'm still 1228 // leaving this here as a placeholder for a more precise assertion later. 1229 // (DLD, 10/05. 1230 assert((true || parallel) 1231 || _g1->evacuation_failed() 1232 || surviving_bytes <= _collection_set_bytes_used_before, 1233 "Or else negative collection!"); 1234 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before); 1235 _recent_CS_bytes_surviving->add(surviving_bytes); 1236 1237 // this is where we update the allocation rate of the application 1238 double app_time_ms = 1239 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); 1240 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1241 // This usually happens due to the timer not having the required 1242 // granularity. Some Linuxes are the usual culprits. 1243 // We'll just set it to something (arbitrarily) small. 1244 app_time_ms = 1.0; 1245 } 1246 size_t regions_allocated = 1247 (_region_num_young - _prev_region_num_young) + 1248 (_region_num_tenured - _prev_region_num_tenured); 1249 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1250 _alloc_rate_ms_seq->add(alloc_rate_ms); 1251 _prev_region_num_young = _region_num_young; 1252 _prev_region_num_tenured = _region_num_tenured; 1253 1254 double interval_ms = 1255 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1256 update_recent_gc_times(end_time_sec, elapsed_ms); 1257 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1258 if (recent_avg_pause_time_ratio() < 0.0 || 1259 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1260 #ifndef PRODUCT 1261 // Dump info to allow post-facto debugging 1262 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1263 gclog_or_tty->print_cr("-------------------------------------------"); 1264 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1265 _recent_gc_times_ms->dump(); 1266 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1267 _recent_prev_end_times_for_all_gcs_sec->dump(); 1268 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1269 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1270 // In debug mode, terminate the JVM if the user wants to debug at this point. 1271 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1272 #endif // !PRODUCT 1273 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1274 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1275 if (_recent_avg_pause_time_ratio < 0.0) { 1276 _recent_avg_pause_time_ratio = 0.0; 1277 } else { 1278 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1279 _recent_avg_pause_time_ratio = 1.0; 1280 } 1281 } 1282 } 1283 1284 if (G1PolicyVerbose > 1) { 1285 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses); 1286 } 1287 1288 PauseSummary* summary = _summary; 1289 1290 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); 1291 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms); 1292 double update_rs_time = avg_value(_par_last_update_rs_times_ms); 1293 double update_rs_processed_buffers = 1294 sum_of_values(_par_last_update_rs_processed_buffers); 1295 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); 1296 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); 1297 double termination_time = avg_value(_par_last_termination_times_ms); 1298 1299 double parallel_other_time = _cur_collection_par_time_ms - 1300 (update_rs_time + ext_root_scan_time + mark_stack_scan_time + 1301 scan_rs_time + obj_copy_time + termination_time); 1302 if (update_stats) { 1303 MainBodySummary* body_summary = summary->main_body_summary(); 1304 guarantee(body_summary != NULL, "should not be null!"); 1305 1306 if (_satb_drain_time_set) 1307 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms); 1308 else 1309 body_summary->record_satb_drain_time_ms(0.0); 1310 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time); 1311 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time); 1312 body_summary->record_update_rs_time_ms(update_rs_time); 1313 body_summary->record_scan_rs_time_ms(scan_rs_time); 1314 body_summary->record_obj_copy_time_ms(obj_copy_time); 1315 if (parallel) { 1316 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms); 1317 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms); 1318 body_summary->record_termination_time_ms(termination_time); 1319 body_summary->record_parallel_other_time_ms(parallel_other_time); 1320 } 1321 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms); 1322 } 1323 1324 if (G1PolicyVerbose > 1) { 1325 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n" 1326 " CH Strong: %10.6f ms (avg: %10.6f ms)\n" 1327 " G1 Strong: %10.6f ms (avg: %10.6f ms)\n" 1328 " Evac: %10.6f ms (avg: %10.6f ms)\n" 1329 " ET-RS: %10.6f ms (avg: %10.6f ms)\n" 1330 " |RS|: " SIZE_FORMAT, 1331 elapsed_ms, recent_avg_time_for_pauses_ms(), 1332 _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(), 1333 _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(), 1334 evac_ms, recent_avg_time_for_evac_ms(), 1335 scan_rs_time, 1336 recent_avg_time_for_pauses_ms() - 1337 recent_avg_time_for_G1_strong_ms(), 1338 rs_size); 1339 1340 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K" 1341 " At end " SIZE_FORMAT "K\n" 1342 " garbage : " SIZE_FORMAT "K" 1343 " of " SIZE_FORMAT "K\n" 1344 " survival : %6.2f%% (%6.2f%% avg)", 1345 _cur_collection_pause_used_at_start_bytes/K, 1346 _g1->used()/K, freed_bytes/K, 1347 _collection_set_bytes_used_before/K, 1348 survival_fraction*100.0, 1349 recent_avg_survival_fraction()*100.0); 1350 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f", 1351 recent_avg_pause_time_ratio() * 100.0); 1352 } 1353 1354 double other_time_ms = elapsed_ms; 1355 1356 if (_satb_drain_time_set) { 1357 other_time_ms -= _cur_satb_drain_time_ms; 1358 } 1359 1360 if (parallel) { 1361 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms; 1362 } else { 1363 other_time_ms -= 1364 update_rs_time + 1365 ext_root_scan_time + mark_stack_scan_time + 1366 scan_rs_time + obj_copy_time; 1367 } 1368 1369 if (PrintGCDetails) { 1370 gclog_or_tty->print_cr("%s, %1.8lf secs]", 1371 (last_pause_included_initial_mark) ? " (initial-mark)" : "", 1372 elapsed_ms / 1000.0); 1373 1374 if (_satb_drain_time_set) { 1375 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms); 1376 } 1377 if (_last_satb_drain_processed_buffers >= 0) { 1378 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers); 1379 } 1380 if (parallel) { 1381 print_stats(1, "Parallel Time", _cur_collection_par_time_ms); 1382 print_par_stats(2, "GC Worker Start Time", 1383 _par_last_gc_worker_start_times_ms, false); 1384 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); 1385 print_par_sizes(3, "Processed Buffers", 1386 _par_last_update_rs_processed_buffers, true); 1387 print_par_stats(2, "Ext Root Scanning", 1388 _par_last_ext_root_scan_times_ms); 1389 print_par_stats(2, "Mark Stack Scanning", 1390 _par_last_mark_stack_scan_times_ms); 1391 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); 1392 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); 1393 print_par_stats(2, "Termination", _par_last_termination_times_ms); 1394 print_par_sizes(3, "Termination Attempts", 1395 _par_last_termination_attempts, true); 1396 print_par_stats(2, "GC Worker End Time", 1397 _par_last_gc_worker_end_times_ms, false); 1398 print_stats(2, "Other", parallel_other_time); 1399 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); 1400 } else { 1401 print_stats(1, "Update RS", update_rs_time); 1402 print_stats(2, "Processed Buffers", 1403 (int)update_rs_processed_buffers); 1404 print_stats(1, "Ext Root Scanning", ext_root_scan_time); 1405 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time); 1406 print_stats(1, "Scan RS", scan_rs_time); 1407 print_stats(1, "Object Copying", obj_copy_time); 1408 } 1409 #ifndef PRODUCT 1410 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms); 1411 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms); 1412 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms); 1413 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms); 1414 if (_num_cc_clears > 0) { 1415 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears)); 1416 } 1417 #endif 1418 print_stats(1, "Other", other_time_ms); 1419 print_stats(2, "Choose CSet", _recorded_young_cset_choice_time_ms); 1420 1421 for (int i = 0; i < _aux_num; ++i) { 1422 if (_cur_aux_times_set[i]) { 1423 char buffer[96]; 1424 sprintf(buffer, "Aux%d", i); 1425 print_stats(1, buffer, _cur_aux_times_ms[i]); 1426 } 1427 } 1428 } 1429 if (PrintGCDetails) 1430 gclog_or_tty->print(" ["); 1431 if (PrintGC || PrintGCDetails) 1432 _g1->print_size_transition(gclog_or_tty, 1433 _cur_collection_pause_used_at_start_bytes, 1434 _g1->used(), _g1->capacity()); 1435 if (PrintGCDetails) 1436 gclog_or_tty->print_cr("]"); 1437 1438 _all_pause_times_ms->add(elapsed_ms); 1439 if (update_stats) { 1440 summary->record_total_time_ms(elapsed_ms); 1441 summary->record_other_time_ms(other_time_ms); 1442 } 1443 for (int i = 0; i < _aux_num; ++i) 1444 if (_cur_aux_times_set[i]) 1445 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]); 1446 1447 // Reset marks-between-pauses counter. 1448 _n_marks_since_last_pause = 0; 1449 1450 // Update the efficiency-since-mark vars. 1451 double proc_ms = elapsed_ms * (double) _parallel_gc_threads; 1452 if (elapsed_ms < MIN_TIMER_GRANULARITY) { 1453 // This usually happens due to the timer not having the required 1454 // granularity. Some Linuxes are the usual culprits. 1455 // We'll just set it to something (arbitrarily) small. 1456 proc_ms = 1.0; 1457 } 1458 double cur_efficiency = (double) freed_bytes / proc_ms; 1459 1460 bool new_in_marking_window = _in_marking_window; 1461 bool new_in_marking_window_im = false; 1462 if (during_initial_mark_pause()) { 1463 new_in_marking_window = true; 1464 new_in_marking_window_im = true; 1465 } 1466 1467 if (in_young_gc_mode()) { 1468 if (_last_full_young_gc) { 1469 set_full_young_gcs(false); 1470 _last_full_young_gc = false; 1471 } 1472 1473 if ( !_last_young_gc_full ) { 1474 if ( _should_revert_to_full_young_gcs || 1475 _known_garbage_ratio < 0.05 || 1476 (adaptive_young_list_length() && 1477 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) { 1478 set_full_young_gcs(true); 1479 } 1480 } 1481 _should_revert_to_full_young_gcs = false; 1482 1483 if (_last_young_gc_full && !_during_marking) 1484 _young_gc_eff_seq->add(cur_efficiency); 1485 } 1486 1487 _short_lived_surv_rate_group->start_adding_regions(); 1488 // do that for any other surv rate groupsx 1489 1490 // <NEW PREDICTION> 1491 1492 if (update_stats) { 1493 double pause_time_ms = elapsed_ms; 1494 1495 size_t diff = 0; 1496 if (_max_pending_cards >= _pending_cards) 1497 diff = _max_pending_cards - _pending_cards; 1498 _pending_card_diff_seq->add((double) diff); 1499 1500 double cost_per_card_ms = 0.0; 1501 if (_pending_cards > 0) { 1502 cost_per_card_ms = update_rs_time / (double) _pending_cards; 1503 _cost_per_card_ms_seq->add(cost_per_card_ms); 1504 } 1505 1506 size_t cards_scanned = _g1->cards_scanned(); 1507 1508 double cost_per_entry_ms = 0.0; 1509 if (cards_scanned > 10) { 1510 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; 1511 if (_last_young_gc_full) 1512 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1513 else 1514 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1515 } 1516 1517 if (_max_rs_lengths > 0) { 1518 double cards_per_entry_ratio = 1519 (double) cards_scanned / (double) _max_rs_lengths; 1520 if (_last_young_gc_full) 1521 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1522 else 1523 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1524 } 1525 1526 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1527 if (rs_length_diff >= 0) 1528 _rs_length_diff_seq->add((double) rs_length_diff); 1529 1530 size_t copied_bytes = surviving_bytes; 1531 double cost_per_byte_ms = 0.0; 1532 if (copied_bytes > 0) { 1533 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; 1534 if (_in_marking_window) 1535 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1536 else 1537 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1538 } 1539 1540 double all_other_time_ms = pause_time_ms - 1541 (update_rs_time + scan_rs_time + obj_copy_time + 1542 _mark_closure_time_ms + termination_time); 1543 1544 double young_other_time_ms = 0.0; 1545 if (_recorded_young_regions > 0) { 1546 young_other_time_ms = 1547 _recorded_young_cset_choice_time_ms + 1548 _recorded_young_free_cset_time_ms; 1549 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1550 (double) _recorded_young_regions); 1551 } 1552 double non_young_other_time_ms = 0.0; 1553 if (_recorded_non_young_regions > 0) { 1554 non_young_other_time_ms = 1555 _recorded_non_young_cset_choice_time_ms + 1556 _recorded_non_young_free_cset_time_ms; 1557 1558 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1559 (double) _recorded_non_young_regions); 1560 } 1561 1562 double constant_other_time_ms = all_other_time_ms - 1563 (young_other_time_ms + non_young_other_time_ms); 1564 _constant_other_time_ms_seq->add(constant_other_time_ms); 1565 1566 double survival_ratio = 0.0; 1567 if (_bytes_in_collection_set_before_gc > 0) { 1568 survival_ratio = (double) bytes_in_to_space_during_gc() / 1569 (double) _bytes_in_collection_set_before_gc; 1570 } 1571 1572 _pending_cards_seq->add((double) _pending_cards); 1573 _scanned_cards_seq->add((double) cards_scanned); 1574 _rs_lengths_seq->add((double) _max_rs_lengths); 1575 1576 double expensive_region_limit_ms = 1577 (double) MaxGCPauseMillis - predict_constant_other_time_ms(); 1578 if (expensive_region_limit_ms < 0.0) { 1579 // this means that the other time was predicted to be longer than 1580 // than the max pause time 1581 expensive_region_limit_ms = (double) MaxGCPauseMillis; 1582 } 1583 _expensive_region_limit_ms = expensive_region_limit_ms; 1584 1585 if (PREDICTIONS_VERBOSE) { 1586 gclog_or_tty->print_cr(""); 1587 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d " 1588 "REGIONS %d %d %d " 1589 "PENDING_CARDS %d %d " 1590 "CARDS_SCANNED %d %d " 1591 "RS_LENGTHS %d %d " 1592 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf " 1593 "SURVIVAL_RATIO %1.6lf %1.6lf " 1594 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf " 1595 "OTHER_YOUNG %1.6lf %1.6lf " 1596 "OTHER_NON_YOUNG %1.6lf %1.6lf " 1597 "VTIME_DIFF %1.6lf TERMINATION %1.6lf " 1598 "ELAPSED %1.6lf %1.6lf ", 1599 _cur_collection_start_sec, 1600 (!_last_young_gc_full) ? 2 : 1601 (last_pause_included_initial_mark) ? 1 : 0, 1602 _recorded_region_num, 1603 _recorded_young_regions, 1604 _recorded_non_young_regions, 1605 _predicted_pending_cards, _pending_cards, 1606 _predicted_cards_scanned, cards_scanned, 1607 _predicted_rs_lengths, _max_rs_lengths, 1608 _predicted_rs_update_time_ms, update_rs_time, 1609 _predicted_rs_scan_time_ms, scan_rs_time, 1610 _predicted_survival_ratio, survival_ratio, 1611 _predicted_object_copy_time_ms, obj_copy_time, 1612 _predicted_constant_other_time_ms, constant_other_time_ms, 1613 _predicted_young_other_time_ms, young_other_time_ms, 1614 _predicted_non_young_other_time_ms, 1615 non_young_other_time_ms, 1616 _vtime_diff_ms, termination_time, 1617 _predicted_pause_time_ms, elapsed_ms); 1618 } 1619 1620 if (G1PolicyVerbose > 0) { 1621 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms", 1622 _predicted_pause_time_ms, 1623 (_within_target) ? "within" : "outside", 1624 elapsed_ms); 1625 } 1626 1627 } 1628 1629 _in_marking_window = new_in_marking_window; 1630 _in_marking_window_im = new_in_marking_window_im; 1631 _free_regions_at_end_of_collection = _g1->free_regions(); 1632 calculate_young_list_min_length(); 1633 calculate_young_list_target_length(); 1634 1635 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1636 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1637 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms); 1638 // </NEW PREDICTION> 1639 } 1640 1641 // <NEW PREDICTION> 1642 1643 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1644 double update_rs_processed_buffers, 1645 double goal_ms) { 1646 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1647 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1648 1649 if (G1UseAdaptiveConcRefinement) { 1650 const int k_gy = 3, k_gr = 6; 1651 const double inc_k = 1.1, dec_k = 0.9; 1652 1653 int g = cg1r->green_zone(); 1654 if (update_rs_time > goal_ms) { 1655 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1656 } else { 1657 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1658 g = (int)MAX2(g * inc_k, g + 1.0); 1659 } 1660 } 1661 // Change the refinement threads params 1662 cg1r->set_green_zone(g); 1663 cg1r->set_yellow_zone(g * k_gy); 1664 cg1r->set_red_zone(g * k_gr); 1665 cg1r->reinitialize_threads(); 1666 1667 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); 1668 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1669 cg1r->yellow_zone()); 1670 // Change the barrier params 1671 dcqs.set_process_completed_threshold(processing_threshold); 1672 dcqs.set_max_completed_queue(cg1r->red_zone()); 1673 } 1674 1675 int curr_queue_size = dcqs.completed_buffers_num(); 1676 if (curr_queue_size >= cg1r->yellow_zone()) { 1677 dcqs.set_completed_queue_padding(curr_queue_size); 1678 } else { 1679 dcqs.set_completed_queue_padding(0); 1680 } 1681 dcqs.notify_if_necessary(); 1682 } 1683 1684 double 1685 G1CollectorPolicy:: 1686 predict_young_collection_elapsed_time_ms(size_t adjustment) { 1687 guarantee( adjustment == 0 || adjustment == 1, "invariant" ); 1688 1689 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1690 size_t young_num = g1h->young_list()->length(); 1691 if (young_num == 0) 1692 return 0.0; 1693 1694 young_num += adjustment; 1695 size_t pending_cards = predict_pending_cards(); 1696 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() + 1697 predict_rs_length_diff(); 1698 size_t card_num; 1699 if (full_young_gcs()) 1700 card_num = predict_young_card_num(rs_lengths); 1701 else 1702 card_num = predict_non_young_card_num(rs_lengths); 1703 size_t young_byte_size = young_num * HeapRegion::GrainBytes; 1704 double accum_yg_surv_rate = 1705 _short_lived_surv_rate_group->accum_surv_rate(adjustment); 1706 1707 size_t bytes_to_copy = 1708 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes); 1709 1710 return 1711 predict_rs_update_time_ms(pending_cards) + 1712 predict_rs_scan_time_ms(card_num) + 1713 predict_object_copy_time_ms(bytes_to_copy) + 1714 predict_young_other_time_ms(young_num) + 1715 predict_constant_other_time_ms(); 1716 } 1717 1718 double 1719 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { 1720 size_t rs_length = predict_rs_length_diff(); 1721 size_t card_num; 1722 if (full_young_gcs()) 1723 card_num = predict_young_card_num(rs_length); 1724 else 1725 card_num = predict_non_young_card_num(rs_length); 1726 return predict_base_elapsed_time_ms(pending_cards, card_num); 1727 } 1728 1729 double 1730 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1731 size_t scanned_cards) { 1732 return 1733 predict_rs_update_time_ms(pending_cards) + 1734 predict_rs_scan_time_ms(scanned_cards) + 1735 predict_constant_other_time_ms(); 1736 } 1737 1738 double 1739 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1740 bool young) { 1741 size_t rs_length = hr->rem_set()->occupied(); 1742 size_t card_num; 1743 if (full_young_gcs()) 1744 card_num = predict_young_card_num(rs_length); 1745 else 1746 card_num = predict_non_young_card_num(rs_length); 1747 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1748 1749 double region_elapsed_time_ms = 1750 predict_rs_scan_time_ms(card_num) + 1751 predict_object_copy_time_ms(bytes_to_copy); 1752 1753 if (young) 1754 region_elapsed_time_ms += predict_young_other_time_ms(1); 1755 else 1756 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1757 1758 return region_elapsed_time_ms; 1759 } 1760 1761 size_t 1762 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { 1763 size_t bytes_to_copy; 1764 if (hr->is_marked()) 1765 bytes_to_copy = hr->max_live_bytes(); 1766 else { 1767 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1, 1768 "invariant" ); 1769 int age = hr->age_in_surv_rate_group(); 1770 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1771 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); 1772 } 1773 1774 return bytes_to_copy; 1775 } 1776 1777 void 1778 G1CollectorPolicy::start_recording_regions() { 1779 _recorded_rs_lengths = 0; 1780 _recorded_young_regions = 0; 1781 _recorded_non_young_regions = 0; 1782 1783 #if PREDICTIONS_VERBOSE 1784 _recorded_marked_bytes = 0; 1785 _recorded_young_bytes = 0; 1786 _predicted_bytes_to_copy = 0; 1787 _predicted_rs_lengths = 0; 1788 _predicted_cards_scanned = 0; 1789 #endif // PREDICTIONS_VERBOSE 1790 } 1791 1792 void 1793 G1CollectorPolicy::record_cset_region_info(HeapRegion* hr, bool young) { 1794 #if PREDICTIONS_VERBOSE 1795 if (!young) { 1796 _recorded_marked_bytes += hr->max_live_bytes(); 1797 } 1798 _predicted_bytes_to_copy += predict_bytes_to_copy(hr); 1799 #endif // PREDICTIONS_VERBOSE 1800 1801 size_t rs_length = hr->rem_set()->occupied(); 1802 _recorded_rs_lengths += rs_length; 1803 } 1804 1805 void 1806 G1CollectorPolicy::record_non_young_cset_region(HeapRegion* hr) { 1807 assert(!hr->is_young(), "should not call this"); 1808 ++_recorded_non_young_regions; 1809 record_cset_region_info(hr, false); 1810 } 1811 1812 void 1813 G1CollectorPolicy::set_recorded_young_regions(size_t n_regions) { 1814 _recorded_young_regions = n_regions; 1815 } 1816 1817 void G1CollectorPolicy::set_recorded_young_bytes(size_t bytes) { 1818 #if PREDICTIONS_VERBOSE 1819 _recorded_young_bytes = bytes; 1820 #endif // PREDICTIONS_VERBOSE 1821 } 1822 1823 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1824 _recorded_rs_lengths = rs_lengths; 1825 } 1826 1827 void G1CollectorPolicy::set_predicted_bytes_to_copy(size_t bytes) { 1828 _predicted_bytes_to_copy = bytes; 1829 } 1830 1831 void 1832 G1CollectorPolicy::end_recording_regions() { 1833 // The _predicted_pause_time_ms field is referenced in code 1834 // not under PREDICTIONS_VERBOSE. Let's initialize it. 1835 _predicted_pause_time_ms = -1.0; 1836 1837 #if PREDICTIONS_VERBOSE 1838 _predicted_pending_cards = predict_pending_cards(); 1839 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff(); 1840 if (full_young_gcs()) 1841 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths); 1842 else 1843 _predicted_cards_scanned += 1844 predict_non_young_card_num(_predicted_rs_lengths); 1845 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions; 1846 1847 _predicted_rs_update_time_ms = 1848 predict_rs_update_time_ms(_g1->pending_card_num()); 1849 _predicted_rs_scan_time_ms = 1850 predict_rs_scan_time_ms(_predicted_cards_scanned); 1851 _predicted_object_copy_time_ms = 1852 predict_object_copy_time_ms(_predicted_bytes_to_copy); 1853 _predicted_constant_other_time_ms = 1854 predict_constant_other_time_ms(); 1855 _predicted_young_other_time_ms = 1856 predict_young_other_time_ms(_recorded_young_regions); 1857 _predicted_non_young_other_time_ms = 1858 predict_non_young_other_time_ms(_recorded_non_young_regions); 1859 1860 _predicted_pause_time_ms = 1861 _predicted_rs_update_time_ms + 1862 _predicted_rs_scan_time_ms + 1863 _predicted_object_copy_time_ms + 1864 _predicted_constant_other_time_ms + 1865 _predicted_young_other_time_ms + 1866 _predicted_non_young_other_time_ms; 1867 #endif // PREDICTIONS_VERBOSE 1868 } 1869 1870 void G1CollectorPolicy::check_if_region_is_too_expensive(double 1871 predicted_time_ms) { 1872 // I don't think we need to do this when in young GC mode since 1873 // marking will be initiated next time we hit the soft limit anyway... 1874 if (predicted_time_ms > _expensive_region_limit_ms) { 1875 if (!in_young_gc_mode()) { 1876 set_full_young_gcs(true); 1877 // We might want to do something different here. However, 1878 // right now we don't support the non-generational G1 mode 1879 // (and in fact we are planning to remove the associated code, 1880 // see CR 6814390). So, let's leave it as is and this will be 1881 // removed some time in the future 1882 ShouldNotReachHere(); 1883 set_during_initial_mark_pause(); 1884 } else 1885 // no point in doing another partial one 1886 _should_revert_to_full_young_gcs = true; 1887 } 1888 } 1889 1890 // </NEW PREDICTION> 1891 1892 1893 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1894 double elapsed_ms) { 1895 _recent_gc_times_ms->add(elapsed_ms); 1896 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1897 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1898 } 1899 1900 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() { 1901 if (_recent_pause_times_ms->num() == 0) return (double) MaxGCPauseMillis; 1902 else return _recent_pause_times_ms->avg(); 1903 } 1904 1905 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() { 1906 if (_recent_CH_strong_roots_times_ms->num() == 0) 1907 return (double)MaxGCPauseMillis/3.0; 1908 else return _recent_CH_strong_roots_times_ms->avg(); 1909 } 1910 1911 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() { 1912 if (_recent_G1_strong_roots_times_ms->num() == 0) 1913 return (double)MaxGCPauseMillis/3.0; 1914 else return _recent_G1_strong_roots_times_ms->avg(); 1915 } 1916 1917 double G1CollectorPolicy::recent_avg_time_for_evac_ms() { 1918 if (_recent_evac_times_ms->num() == 0) return (double)MaxGCPauseMillis/3.0; 1919 else return _recent_evac_times_ms->avg(); 1920 } 1921 1922 int G1CollectorPolicy::number_of_recent_gcs() { 1923 assert(_recent_CH_strong_roots_times_ms->num() == 1924 _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync"); 1925 assert(_recent_G1_strong_roots_times_ms->num() == 1926 _recent_evac_times_ms->num(), "Sequence out of sync"); 1927 assert(_recent_evac_times_ms->num() == 1928 _recent_pause_times_ms->num(), "Sequence out of sync"); 1929 assert(_recent_pause_times_ms->num() == 1930 _recent_CS_bytes_used_before->num(), "Sequence out of sync"); 1931 assert(_recent_CS_bytes_used_before->num() == 1932 _recent_CS_bytes_surviving->num(), "Sequence out of sync"); 1933 return _recent_pause_times_ms->num(); 1934 } 1935 1936 double G1CollectorPolicy::recent_avg_survival_fraction() { 1937 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving, 1938 _recent_CS_bytes_used_before); 1939 } 1940 1941 double G1CollectorPolicy::last_survival_fraction() { 1942 return last_survival_fraction_work(_recent_CS_bytes_surviving, 1943 _recent_CS_bytes_used_before); 1944 } 1945 1946 double 1947 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving, 1948 TruncatedSeq* before) { 1949 assert(surviving->num() == before->num(), "Sequence out of sync"); 1950 if (before->sum() > 0.0) { 1951 double recent_survival_rate = surviving->sum() / before->sum(); 1952 // We exempt parallel collection from this check because Alloc Buffer 1953 // fragmentation can produce negative collections. 1954 // Further, we're now always doing parallel collection. But I'm still 1955 // leaving this here as a placeholder for a more precise assertion later. 1956 // (DLD, 10/05.) 1957 assert((true || G1CollectedHeap::use_parallel_gc_threads()) || 1958 _g1->evacuation_failed() || 1959 recent_survival_rate <= 1.0, "Or bad frac"); 1960 return recent_survival_rate; 1961 } else { 1962 return 1.0; // Be conservative. 1963 } 1964 } 1965 1966 double 1967 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving, 1968 TruncatedSeq* before) { 1969 assert(surviving->num() == before->num(), "Sequence out of sync"); 1970 if (surviving->num() > 0 && before->last() > 0.0) { 1971 double last_survival_rate = surviving->last() / before->last(); 1972 // We exempt parallel collection from this check because Alloc Buffer 1973 // fragmentation can produce negative collections. 1974 // Further, we're now always doing parallel collection. But I'm still 1975 // leaving this here as a placeholder for a more precise assertion later. 1976 // (DLD, 10/05.) 1977 assert((true || G1CollectedHeap::use_parallel_gc_threads()) || 1978 last_survival_rate <= 1.0, "Or bad frac"); 1979 return last_survival_rate; 1980 } else { 1981 return 1.0; 1982 } 1983 } 1984 1985 static const int survival_min_obs = 5; 1986 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 }; 1987 static const double min_survival_rate = 0.1; 1988 1989 double 1990 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg, 1991 double latest) { 1992 double res = avg; 1993 if (number_of_recent_gcs() < survival_min_obs) { 1994 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]); 1995 } 1996 res = MAX2(res, latest); 1997 res = MAX2(res, min_survival_rate); 1998 // In the parallel case, LAB fragmentation can produce "negative 1999 // collections"; so can evac failure. Cap at 1.0 2000 res = MIN2(res, 1.0); 2001 return res; 2002 } 2003 2004 size_t G1CollectorPolicy::expansion_amount() { 2005 if ((recent_avg_pause_time_ratio() * 100.0) > _gc_overhead_perc) { 2006 // We will double the existing space, or take 2007 // G1ExpandByPercentOfAvailable % of the available expansion 2008 // space, whichever is smaller, bounded below by a minimum 2009 // expansion (unless that's all that's left.) 2010 const size_t min_expand_bytes = 1*M; 2011 size_t reserved_bytes = _g1->g1_reserved_obj_bytes(); 2012 size_t committed_bytes = _g1->capacity(); 2013 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 2014 size_t expand_bytes; 2015 size_t expand_bytes_via_pct = 2016 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 2017 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 2018 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 2019 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 2020 if (G1PolicyVerbose > 1) { 2021 gclog_or_tty->print("Decided to expand: ratio = %5.2f, " 2022 "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n" 2023 " Answer = %d.\n", 2024 recent_avg_pause_time_ratio(), 2025 byte_size_in_proper_unit(committed_bytes), 2026 proper_unit_for_byte_size(committed_bytes), 2027 byte_size_in_proper_unit(uncommitted_bytes), 2028 proper_unit_for_byte_size(uncommitted_bytes), 2029 byte_size_in_proper_unit(expand_bytes_via_pct), 2030 proper_unit_for_byte_size(expand_bytes_via_pct), 2031 byte_size_in_proper_unit(expand_bytes), 2032 proper_unit_for_byte_size(expand_bytes)); 2033 } 2034 return expand_bytes; 2035 } else { 2036 return 0; 2037 } 2038 } 2039 2040 void G1CollectorPolicy::note_start_of_mark_thread() { 2041 _mark_thread_startup_sec = os::elapsedTime(); 2042 } 2043 2044 class CountCSClosure: public HeapRegionClosure { 2045 G1CollectorPolicy* _g1_policy; 2046 public: 2047 CountCSClosure(G1CollectorPolicy* g1_policy) : 2048 _g1_policy(g1_policy) {} 2049 bool doHeapRegion(HeapRegion* r) { 2050 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); 2051 return false; 2052 } 2053 }; 2054 2055 void G1CollectorPolicy::count_CS_bytes_used() { 2056 CountCSClosure cs_closure(this); 2057 _g1->collection_set_iterate(&cs_closure); 2058 } 2059 2060 static void print_indent(int level) { 2061 for (int j = 0; j < level+1; ++j) 2062 gclog_or_tty->print(" "); 2063 } 2064 2065 void G1CollectorPolicy::print_summary (int level, 2066 const char* str, 2067 NumberSeq* seq) const { 2068 double sum = seq->sum(); 2069 print_indent(level); 2070 gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", 2071 str, sum / 1000.0, seq->avg()); 2072 } 2073 2074 void G1CollectorPolicy::print_summary_sd (int level, 2075 const char* str, 2076 NumberSeq* seq) const { 2077 print_summary(level, str, seq); 2078 print_indent(level + 5); 2079 gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2080 seq->num(), seq->sd(), seq->maximum()); 2081 } 2082 2083 void G1CollectorPolicy::check_other_times(int level, 2084 NumberSeq* other_times_ms, 2085 NumberSeq* calc_other_times_ms) const { 2086 bool should_print = false; 2087 2088 double max_sum = MAX2(fabs(other_times_ms->sum()), 2089 fabs(calc_other_times_ms->sum())); 2090 double min_sum = MIN2(fabs(other_times_ms->sum()), 2091 fabs(calc_other_times_ms->sum())); 2092 double sum_ratio = max_sum / min_sum; 2093 if (sum_ratio > 1.1) { 2094 should_print = true; 2095 print_indent(level + 1); 2096 gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###"); 2097 } 2098 2099 double max_avg = MAX2(fabs(other_times_ms->avg()), 2100 fabs(calc_other_times_ms->avg())); 2101 double min_avg = MIN2(fabs(other_times_ms->avg()), 2102 fabs(calc_other_times_ms->avg())); 2103 double avg_ratio = max_avg / min_avg; 2104 if (avg_ratio > 1.1) { 2105 should_print = true; 2106 print_indent(level + 1); 2107 gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###"); 2108 } 2109 2110 if (other_times_ms->sum() < -0.01) { 2111 print_indent(level + 1); 2112 gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###"); 2113 } 2114 2115 if (other_times_ms->avg() < -0.01) { 2116 print_indent(level + 1); 2117 gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###"); 2118 } 2119 2120 if (calc_other_times_ms->sum() < -0.01) { 2121 should_print = true; 2122 print_indent(level + 1); 2123 gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###"); 2124 } 2125 2126 if (calc_other_times_ms->avg() < -0.01) { 2127 should_print = true; 2128 print_indent(level + 1); 2129 gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###"); 2130 } 2131 2132 if (should_print) 2133 print_summary(level, "Other(Calc)", calc_other_times_ms); 2134 } 2135 2136 void G1CollectorPolicy::print_summary(PauseSummary* summary) const { 2137 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 2138 MainBodySummary* body_summary = summary->main_body_summary(); 2139 if (summary->get_total_seq()->num() > 0) { 2140 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq()); 2141 if (body_summary != NULL) { 2142 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq()); 2143 if (parallel) { 2144 print_summary(1, "Parallel Time", body_summary->get_parallel_seq()); 2145 print_summary(2, "Update RS", body_summary->get_update_rs_seq()); 2146 print_summary(2, "Ext Root Scanning", 2147 body_summary->get_ext_root_scan_seq()); 2148 print_summary(2, "Mark Stack Scanning", 2149 body_summary->get_mark_stack_scan_seq()); 2150 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq()); 2151 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq()); 2152 print_summary(2, "Termination", body_summary->get_termination_seq()); 2153 print_summary(2, "Other", body_summary->get_parallel_other_seq()); 2154 { 2155 NumberSeq* other_parts[] = { 2156 body_summary->get_update_rs_seq(), 2157 body_summary->get_ext_root_scan_seq(), 2158 body_summary->get_mark_stack_scan_seq(), 2159 body_summary->get_scan_rs_seq(), 2160 body_summary->get_obj_copy_seq(), 2161 body_summary->get_termination_seq() 2162 }; 2163 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(), 2164 6, other_parts); 2165 check_other_times(2, body_summary->get_parallel_other_seq(), 2166 &calc_other_times_ms); 2167 } 2168 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq()); 2169 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq()); 2170 } else { 2171 print_summary(1, "Update RS", body_summary->get_update_rs_seq()); 2172 print_summary(1, "Ext Root Scanning", 2173 body_summary->get_ext_root_scan_seq()); 2174 print_summary(1, "Mark Stack Scanning", 2175 body_summary->get_mark_stack_scan_seq()); 2176 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq()); 2177 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq()); 2178 } 2179 } 2180 print_summary(1, "Other", summary->get_other_seq()); 2181 { 2182 if (body_summary != NULL) { 2183 NumberSeq calc_other_times_ms; 2184 if (parallel) { 2185 // parallel 2186 NumberSeq* other_parts[] = { 2187 body_summary->get_satb_drain_seq(), 2188 body_summary->get_parallel_seq(), 2189 body_summary->get_clear_ct_seq() 2190 }; 2191 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 2192 3, other_parts); 2193 } else { 2194 // serial 2195 NumberSeq* other_parts[] = { 2196 body_summary->get_satb_drain_seq(), 2197 body_summary->get_update_rs_seq(), 2198 body_summary->get_ext_root_scan_seq(), 2199 body_summary->get_mark_stack_scan_seq(), 2200 body_summary->get_scan_rs_seq(), 2201 body_summary->get_obj_copy_seq() 2202 }; 2203 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 2204 6, other_parts); 2205 } 2206 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms); 2207 } 2208 } 2209 } else { 2210 print_indent(0); 2211 gclog_or_tty->print_cr("none"); 2212 } 2213 gclog_or_tty->print_cr(""); 2214 } 2215 2216 void G1CollectorPolicy::print_tracing_info() const { 2217 if (TraceGen0Time) { 2218 gclog_or_tty->print_cr("ALL PAUSES"); 2219 print_summary_sd(0, "Total", _all_pause_times_ms); 2220 gclog_or_tty->print_cr(""); 2221 gclog_or_tty->print_cr(""); 2222 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num); 2223 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num); 2224 gclog_or_tty->print_cr(""); 2225 2226 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2227 print_summary(_summary); 2228 2229 gclog_or_tty->print_cr("MISC"); 2230 print_summary_sd(0, "Stop World", _all_stop_world_times_ms); 2231 print_summary_sd(0, "Yields", _all_yield_times_ms); 2232 for (int i = 0; i < _aux_num; ++i) { 2233 if (_all_aux_times_ms[i].num() > 0) { 2234 char buffer[96]; 2235 sprintf(buffer, "Aux%d", i); 2236 print_summary_sd(0, buffer, &_all_aux_times_ms[i]); 2237 } 2238 } 2239 2240 size_t all_region_num = _region_num_young + _region_num_tenured; 2241 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), " 2242 "Tenured %8d (%6.2lf%%)", 2243 all_region_num, 2244 _region_num_young, 2245 (double) _region_num_young / (double) all_region_num * 100.0, 2246 _region_num_tenured, 2247 (double) _region_num_tenured / (double) all_region_num * 100.0); 2248 } 2249 if (TraceGen1Time) { 2250 if (_all_full_gc_times_ms->num() > 0) { 2251 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2252 _all_full_gc_times_ms->num(), 2253 _all_full_gc_times_ms->sum() / 1000.0); 2254 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg()); 2255 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2256 _all_full_gc_times_ms->sd(), 2257 _all_full_gc_times_ms->maximum()); 2258 } 2259 } 2260 } 2261 2262 void G1CollectorPolicy::print_yg_surv_rate_info() const { 2263 #ifndef PRODUCT 2264 _short_lived_surv_rate_group->print_surv_rate_summary(); 2265 // add this call for any other surv rate groups 2266 #endif // PRODUCT 2267 } 2268 2269 bool 2270 G1CollectorPolicy::should_add_next_region_to_young_list() { 2271 assert(in_young_gc_mode(), "should be in young GC mode"); 2272 bool ret; 2273 size_t young_list_length = _g1->young_list()->length(); 2274 size_t young_list_max_length = _young_list_target_length; 2275 if (G1FixedEdenSize) { 2276 young_list_max_length -= _max_survivor_regions; 2277 } 2278 if (young_list_length < young_list_max_length) { 2279 ret = true; 2280 ++_region_num_young; 2281 } else { 2282 ret = false; 2283 ++_region_num_tenured; 2284 } 2285 2286 return ret; 2287 } 2288 2289 #ifndef PRODUCT 2290 // for debugging, bit of a hack... 2291 static char* 2292 region_num_to_mbs(int length) { 2293 static char buffer[64]; 2294 double bytes = (double) (length * HeapRegion::GrainBytes); 2295 double mbs = bytes / (double) (1024 * 1024); 2296 sprintf(buffer, "%7.2lfMB", mbs); 2297 return buffer; 2298 } 2299 #endif // PRODUCT 2300 2301 size_t G1CollectorPolicy::max_regions(int purpose) { 2302 switch (purpose) { 2303 case GCAllocForSurvived: 2304 return _max_survivor_regions; 2305 case GCAllocForTenured: 2306 return REGIONS_UNLIMITED; 2307 default: 2308 ShouldNotReachHere(); 2309 return REGIONS_UNLIMITED; 2310 }; 2311 } 2312 2313 // Calculates survivor space parameters. 2314 void G1CollectorPolicy::calculate_survivors_policy() 2315 { 2316 if (G1FixedSurvivorSpaceSize == 0) { 2317 _max_survivor_regions = _young_list_target_length / SurvivorRatio; 2318 } else { 2319 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes; 2320 } 2321 2322 if (G1FixedTenuringThreshold) { 2323 _tenuring_threshold = MaxTenuringThreshold; 2324 } else { 2325 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 2326 HeapRegion::GrainWords * _max_survivor_regions); 2327 } 2328 } 2329 2330 bool 2331 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t 2332 word_size) { 2333 assert(_g1->regions_accounted_for(), "Region leakage!"); 2334 double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 2335 2336 size_t young_list_length = _g1->young_list()->length(); 2337 size_t young_list_max_length = _young_list_target_length; 2338 if (G1FixedEdenSize) { 2339 young_list_max_length -= _max_survivor_regions; 2340 } 2341 bool reached_target_length = young_list_length >= young_list_max_length; 2342 2343 if (in_young_gc_mode()) { 2344 if (reached_target_length) { 2345 assert( young_list_length > 0 && _g1->young_list()->length() > 0, 2346 "invariant" ); 2347 return true; 2348 } 2349 } else { 2350 guarantee( false, "should not reach here" ); 2351 } 2352 2353 return false; 2354 } 2355 2356 #ifndef PRODUCT 2357 class HRSortIndexIsOKClosure: public HeapRegionClosure { 2358 CollectionSetChooser* _chooser; 2359 public: 2360 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) : 2361 _chooser(chooser) {} 2362 2363 bool doHeapRegion(HeapRegion* r) { 2364 if (!r->continuesHumongous()) { 2365 assert(_chooser->regionProperlyOrdered(r), "Ought to be."); 2366 } 2367 return false; 2368 } 2369 }; 2370 2371 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() { 2372 HRSortIndexIsOKClosure cl(_collectionSetChooser); 2373 _g1->heap_region_iterate(&cl); 2374 return true; 2375 } 2376 #endif 2377 2378 bool 2379 G1CollectorPolicy::force_initial_mark_if_outside_cycle() { 2380 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2381 if (!during_cycle) { 2382 set_initiate_conc_mark_if_possible(); 2383 return true; 2384 } else { 2385 return false; 2386 } 2387 } 2388 2389 void 2390 G1CollectorPolicy::decide_on_conc_mark_initiation() { 2391 // We are about to decide on whether this pause will be an 2392 // initial-mark pause. 2393 2394 // First, during_initial_mark_pause() should not be already set. We 2395 // will set it here if we have to. However, it should be cleared by 2396 // the end of the pause (it's only set for the duration of an 2397 // initial-mark pause). 2398 assert(!during_initial_mark_pause(), "pre-condition"); 2399 2400 if (initiate_conc_mark_if_possible()) { 2401 // We had noticed on a previous pause that the heap occupancy has 2402 // gone over the initiating threshold and we should start a 2403 // concurrent marking cycle. So we might initiate one. 2404 2405 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2406 if (!during_cycle) { 2407 // The concurrent marking thread is not "during a cycle", i.e., 2408 // it has completed the last one. So we can go ahead and 2409 // initiate a new cycle. 2410 2411 set_during_initial_mark_pause(); 2412 2413 // And we can now clear initiate_conc_mark_if_possible() as 2414 // we've already acted on it. 2415 clear_initiate_conc_mark_if_possible(); 2416 } else { 2417 // The concurrent marking thread is still finishing up the 2418 // previous cycle. If we start one right now the two cycles 2419 // overlap. In particular, the concurrent marking thread might 2420 // be in the process of clearing the next marking bitmap (which 2421 // we will use for the next cycle if we start one). Starting a 2422 // cycle now will be bad given that parts of the marking 2423 // information might get cleared by the marking thread. And we 2424 // cannot wait for the marking thread to finish the cycle as it 2425 // periodically yields while clearing the next marking bitmap 2426 // and, if it's in a yield point, it's waiting for us to 2427 // finish. So, at this point we will not start a cycle and we'll 2428 // let the concurrent marking thread complete the last one. 2429 } 2430 } 2431 } 2432 2433 void 2434 G1CollectorPolicy_BestRegionsFirst:: 2435 record_collection_pause_start(double start_time_sec, size_t start_used) { 2436 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used); 2437 } 2438 2439 class NextNonCSElemFinder: public HeapRegionClosure { 2440 HeapRegion* _res; 2441 public: 2442 NextNonCSElemFinder(): _res(NULL) {} 2443 bool doHeapRegion(HeapRegion* r) { 2444 if (!r->in_collection_set()) { 2445 _res = r; 2446 return true; 2447 } else { 2448 return false; 2449 } 2450 } 2451 HeapRegion* res() { return _res; } 2452 }; 2453 2454 class KnownGarbageClosure: public HeapRegionClosure { 2455 CollectionSetChooser* _hrSorted; 2456 2457 public: 2458 KnownGarbageClosure(CollectionSetChooser* hrSorted) : 2459 _hrSorted(hrSorted) 2460 {} 2461 2462 bool doHeapRegion(HeapRegion* r) { 2463 // We only include humongous regions in collection 2464 // sets when concurrent mark shows that their contained object is 2465 // unreachable. 2466 2467 // Do we have any marking information for this region? 2468 if (r->is_marked()) { 2469 // We don't include humongous regions in collection 2470 // sets because we collect them immediately at the end of a marking 2471 // cycle. We also don't include young regions because we *must* 2472 // include them in the next collection pause. 2473 if (!r->isHumongous() && !r->is_young()) { 2474 _hrSorted->addMarkedHeapRegion(r); 2475 } 2476 } 2477 return false; 2478 } 2479 }; 2480 2481 class ParKnownGarbageHRClosure: public HeapRegionClosure { 2482 CollectionSetChooser* _hrSorted; 2483 jint _marked_regions_added; 2484 jint _chunk_size; 2485 jint _cur_chunk_idx; 2486 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) 2487 int _worker; 2488 int _invokes; 2489 2490 void get_new_chunk() { 2491 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size); 2492 _cur_chunk_end = _cur_chunk_idx + _chunk_size; 2493 } 2494 void add_region(HeapRegion* r) { 2495 if (_cur_chunk_idx == _cur_chunk_end) { 2496 get_new_chunk(); 2497 } 2498 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); 2499 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r); 2500 _marked_regions_added++; 2501 _cur_chunk_idx++; 2502 } 2503 2504 public: 2505 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 2506 jint chunk_size, 2507 int worker) : 2508 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker), 2509 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0), 2510 _invokes(0) 2511 {} 2512 2513 bool doHeapRegion(HeapRegion* r) { 2514 // We only include humongous regions in collection 2515 // sets when concurrent mark shows that their contained object is 2516 // unreachable. 2517 _invokes++; 2518 2519 // Do we have any marking information for this region? 2520 if (r->is_marked()) { 2521 // We don't include humongous regions in collection 2522 // sets because we collect them immediately at the end of a marking 2523 // cycle. 2524 // We also do not include young regions in collection sets 2525 if (!r->isHumongous() && !r->is_young()) { 2526 add_region(r); 2527 } 2528 } 2529 return false; 2530 } 2531 jint marked_regions_added() { return _marked_regions_added; } 2532 int invokes() { return _invokes; } 2533 }; 2534 2535 class ParKnownGarbageTask: public AbstractGangTask { 2536 CollectionSetChooser* _hrSorted; 2537 jint _chunk_size; 2538 G1CollectedHeap* _g1; 2539 public: 2540 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) : 2541 AbstractGangTask("ParKnownGarbageTask"), 2542 _hrSorted(hrSorted), _chunk_size(chunk_size), 2543 _g1(G1CollectedHeap::heap()) 2544 {} 2545 2546 void work(int i) { 2547 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i); 2548 // Back to zero for the claim value. 2549 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i, 2550 HeapRegion::InitialClaimValue); 2551 jint regions_added = parKnownGarbageCl.marked_regions_added(); 2552 _hrSorted->incNumMarkedHeapRegions(regions_added); 2553 if (G1PrintParCleanupStats) { 2554 gclog_or_tty->print(" Thread %d called %d times, added %d regions to list.\n", 2555 i, parKnownGarbageCl.invokes(), regions_added); 2556 } 2557 } 2558 }; 2559 2560 void 2561 G1CollectorPolicy_BestRegionsFirst:: 2562 record_concurrent_mark_cleanup_end(size_t freed_bytes, 2563 size_t max_live_bytes) { 2564 double start; 2565 if (G1PrintParCleanupStats) start = os::elapsedTime(); 2566 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes); 2567 2568 _collectionSetChooser->clearMarkedHeapRegions(); 2569 double clear_marked_end; 2570 if (G1PrintParCleanupStats) { 2571 clear_marked_end = os::elapsedTime(); 2572 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.", 2573 (clear_marked_end - start)*1000.0); 2574 } 2575 if (G1CollectedHeap::use_parallel_gc_threads()) { 2576 const size_t OverpartitionFactor = 4; 2577 const size_t MinWorkUnit = 8; 2578 const size_t WorkUnit = 2579 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor), 2580 MinWorkUnit); 2581 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(), 2582 WorkUnit); 2583 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, 2584 (int) WorkUnit); 2585 _g1->workers()->run_task(&parKnownGarbageTask); 2586 2587 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 2588 "sanity check"); 2589 } else { 2590 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); 2591 _g1->heap_region_iterate(&knownGarbagecl); 2592 } 2593 double known_garbage_end; 2594 if (G1PrintParCleanupStats) { 2595 known_garbage_end = os::elapsedTime(); 2596 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.", 2597 (known_garbage_end - clear_marked_end)*1000.0); 2598 } 2599 _collectionSetChooser->sortMarkedHeapRegions(); 2600 double sort_end; 2601 if (G1PrintParCleanupStats) { 2602 sort_end = os::elapsedTime(); 2603 gclog_or_tty->print_cr(" sorting: %8.3f ms.", 2604 (sort_end - known_garbage_end)*1000.0); 2605 } 2606 2607 record_concurrent_mark_cleanup_end_work2(); 2608 double work2_end; 2609 if (G1PrintParCleanupStats) { 2610 work2_end = os::elapsedTime(); 2611 gclog_or_tty->print_cr(" work2: %8.3f ms.", 2612 (work2_end - sort_end)*1000.0); 2613 } 2614 } 2615 2616 // Add the heap region at the head of the non-incremental collection set 2617 void G1CollectorPolicy:: 2618 add_to_collection_set(HeapRegion* hr) { 2619 assert(_inc_cset_build_state == Active, "Precondition"); 2620 assert(!hr->is_young(), "non-incremental add of young region"); 2621 2622 if (G1PrintHeapRegions) { 2623 gclog_or_tty->print_cr("added region to cset " 2624 "%d:["PTR_FORMAT", "PTR_FORMAT"], " 2625 "top "PTR_FORMAT", %s", 2626 hr->hrs_index(), hr->bottom(), hr->end(), 2627 hr->top(), hr->is_young() ? "YOUNG" : "NOT_YOUNG"); 2628 } 2629 2630 if (_g1->mark_in_progress()) 2631 _g1->concurrent_mark()->registerCSetRegion(hr); 2632 2633 assert(!hr->in_collection_set(), "should not already be in the CSet"); 2634 hr->set_in_collection_set(true); 2635 hr->set_next_in_collection_set(_collection_set); 2636 _collection_set = hr; 2637 _collection_set_size++; 2638 _collection_set_bytes_used_before += hr->used(); 2639 _g1->register_region_with_in_cset_fast_test(hr); 2640 } 2641 2642 // Initialize the per-collection-set information 2643 void G1CollectorPolicy::start_incremental_cset_building() { 2644 assert(_inc_cset_build_state == Inactive, "Precondition"); 2645 2646 _inc_cset_head = NULL; 2647 _inc_cset_tail = NULL; 2648 _inc_cset_size = 0; 2649 _inc_cset_bytes_used_before = 0; 2650 2651 if (in_young_gc_mode()) { 2652 _inc_cset_young_index = 0; 2653 } 2654 2655 _inc_cset_max_finger = 0; 2656 _inc_cset_recorded_young_bytes = 0; 2657 _inc_cset_recorded_rs_lengths = 0; 2658 _inc_cset_predicted_elapsed_time_ms = 0; 2659 _inc_cset_predicted_bytes_to_copy = 0; 2660 _inc_cset_build_state = Active; 2661 } 2662 2663 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 2664 // This routine is used when: 2665 // * adding survivor regions to the incremental cset at the end of an 2666 // evacuation pause, 2667 // * adding the current allocation region to the incremental cset 2668 // when it is retired, and 2669 // * updating existing policy information for a region in the 2670 // incremental cset via young list RSet sampling. 2671 // Therefore this routine may be called at a safepoint by the 2672 // VM thread, or in-between safepoints by mutator threads (when 2673 // retiring the current allocation region) or a concurrent 2674 // refine thread (RSet sampling). 2675 2676 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2677 size_t used_bytes = hr->used(); 2678 2679 _inc_cset_recorded_rs_lengths += rs_length; 2680 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 2681 2682 _inc_cset_bytes_used_before += used_bytes; 2683 2684 // Cache the values we have added to the aggregated informtion 2685 // in the heap region in case we have to remove this region from 2686 // the incremental collection set, or it is updated by the 2687 // rset sampling code 2688 hr->set_recorded_rs_length(rs_length); 2689 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 2690 2691 #if PREDICTIONS_VERBOSE 2692 size_t bytes_to_copy = predict_bytes_to_copy(hr); 2693 _inc_cset_predicted_bytes_to_copy += bytes_to_copy; 2694 2695 // Record the number of bytes used in this region 2696 _inc_cset_recorded_young_bytes += used_bytes; 2697 2698 // Cache the values we have added to the aggregated informtion 2699 // in the heap region in case we have to remove this region from 2700 // the incremental collection set, or it is updated by the 2701 // rset sampling code 2702 hr->set_predicted_bytes_to_copy(bytes_to_copy); 2703 #endif // PREDICTIONS_VERBOSE 2704 } 2705 2706 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) { 2707 // This routine is currently only called as part of the updating of 2708 // existing policy information for regions in the incremental cset that 2709 // is performed by the concurrent refine thread(s) as part of young list 2710 // RSet sampling. Therefore we should not be at a safepoint. 2711 2712 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint"); 2713 assert(hr->is_young(), "it should be"); 2714 2715 size_t used_bytes = hr->used(); 2716 size_t old_rs_length = hr->recorded_rs_length(); 2717 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 2718 2719 // Subtract the old recorded/predicted policy information for 2720 // the given heap region from the collection set info. 2721 _inc_cset_recorded_rs_lengths -= old_rs_length; 2722 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms; 2723 2724 _inc_cset_bytes_used_before -= used_bytes; 2725 2726 // Clear the values cached in the heap region 2727 hr->set_recorded_rs_length(0); 2728 hr->set_predicted_elapsed_time_ms(0); 2729 2730 #if PREDICTIONS_VERBOSE 2731 size_t old_predicted_bytes_to_copy = hr->predicted_bytes_to_copy(); 2732 _inc_cset_predicted_bytes_to_copy -= old_predicted_bytes_to_copy; 2733 2734 // Subtract the number of bytes used in this region 2735 _inc_cset_recorded_young_bytes -= used_bytes; 2736 2737 // Clear the values cached in the heap region 2738 hr->set_predicted_bytes_to_copy(0); 2739 #endif // PREDICTIONS_VERBOSE 2740 } 2741 2742 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) { 2743 // Update the collection set information that is dependent on the new RS length 2744 assert(hr->is_young(), "Precondition"); 2745 2746 remove_from_incremental_cset_info(hr); 2747 add_to_incremental_cset_info(hr, new_rs_length); 2748 } 2749 2750 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 2751 assert( hr->is_young(), "invariant"); 2752 assert( hr->young_index_in_cset() == -1, "invariant" ); 2753 assert(_inc_cset_build_state == Active, "Precondition"); 2754 2755 // We need to clear and set the cached recorded/cached collection set 2756 // information in the heap region here (before the region gets added 2757 // to the collection set). An individual heap region's cached values 2758 // are calculated, aggregated with the policy collection set info, 2759 // and cached in the heap region here (initially) and (subsequently) 2760 // by the Young List sampling code. 2761 2762 size_t rs_length = hr->rem_set()->occupied(); 2763 add_to_incremental_cset_info(hr, rs_length); 2764 2765 HeapWord* hr_end = hr->end(); 2766 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 2767 2768 assert(!hr->in_collection_set(), "invariant"); 2769 hr->set_in_collection_set(true); 2770 assert( hr->next_in_collection_set() == NULL, "invariant"); 2771 2772 _inc_cset_size++; 2773 _g1->register_region_with_in_cset_fast_test(hr); 2774 2775 hr->set_young_index_in_cset((int) _inc_cset_young_index); 2776 ++_inc_cset_young_index; 2777 } 2778 2779 // Add the region at the RHS of the incremental cset 2780 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 2781 // We should only ever be appending survivors at the end of a pause 2782 assert( hr->is_survivor(), "Logic"); 2783 2784 // Do the 'common' stuff 2785 add_region_to_incremental_cset_common(hr); 2786 2787 // Now add the region at the right hand side 2788 if (_inc_cset_tail == NULL) { 2789 assert(_inc_cset_head == NULL, "invariant"); 2790 _inc_cset_head = hr; 2791 } else { 2792 _inc_cset_tail->set_next_in_collection_set(hr); 2793 } 2794 _inc_cset_tail = hr; 2795 2796 if (G1PrintHeapRegions) { 2797 gclog_or_tty->print_cr(" added region to incremental cset (RHS) " 2798 "%d:["PTR_FORMAT", "PTR_FORMAT"], " 2799 "top "PTR_FORMAT", young %s", 2800 hr->hrs_index(), hr->bottom(), hr->end(), 2801 hr->top(), (hr->is_young()) ? "YES" : "NO"); 2802 } 2803 } 2804 2805 // Add the region to the LHS of the incremental cset 2806 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2807 // Survivors should be added to the RHS at the end of a pause 2808 assert(!hr->is_survivor(), "Logic"); 2809 2810 // Do the 'common' stuff 2811 add_region_to_incremental_cset_common(hr); 2812 2813 // Add the region at the left hand side 2814 hr->set_next_in_collection_set(_inc_cset_head); 2815 if (_inc_cset_head == NULL) { 2816 assert(_inc_cset_tail == NULL, "Invariant"); 2817 _inc_cset_tail = hr; 2818 } 2819 _inc_cset_head = hr; 2820 2821 if (G1PrintHeapRegions) { 2822 gclog_or_tty->print_cr(" added region to incremental cset (LHS) " 2823 "%d:["PTR_FORMAT", "PTR_FORMAT"], " 2824 "top "PTR_FORMAT", young %s", 2825 hr->hrs_index(), hr->bottom(), hr->end(), 2826 hr->top(), (hr->is_young()) ? "YES" : "NO"); 2827 } 2828 } 2829 2830 #ifndef PRODUCT 2831 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2832 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2833 2834 st->print_cr("\nCollection_set:"); 2835 HeapRegion* csr = list_head; 2836 while (csr != NULL) { 2837 HeapRegion* next = csr->next_in_collection_set(); 2838 assert(csr->in_collection_set(), "bad CS"); 2839 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, " 2840 "age: %4d, y: %d, surv: %d", 2841 csr->bottom(), csr->end(), 2842 csr->top(), 2843 csr->prev_top_at_mark_start(), 2844 csr->next_top_at_mark_start(), 2845 csr->top_at_conc_mark_count(), 2846 csr->age_in_surv_rate_group_cond(), 2847 csr->is_young(), 2848 csr->is_survivor()); 2849 csr = next; 2850 } 2851 } 2852 #endif // !PRODUCT 2853 2854 void 2855 G1CollectorPolicy_BestRegionsFirst::choose_collection_set( 2856 double target_pause_time_ms) { 2857 // Set this here - in case we're not doing young collections. 2858 double non_young_start_time_sec = os::elapsedTime(); 2859 2860 start_recording_regions(); 2861 2862 guarantee(target_pause_time_ms > 0.0, 2863 err_msg("target_pause_time_ms = %1.6lf should be positive", 2864 target_pause_time_ms)); 2865 guarantee(_collection_set == NULL, "Precondition"); 2866 2867 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2868 double predicted_pause_time_ms = base_time_ms; 2869 2870 double time_remaining_ms = target_pause_time_ms - base_time_ms; 2871 2872 // the 10% and 50% values are arbitrary... 2873 if (time_remaining_ms < 0.10 * target_pause_time_ms) { 2874 time_remaining_ms = 0.50 * target_pause_time_ms; 2875 _within_target = false; 2876 } else { 2877 _within_target = true; 2878 } 2879 2880 // We figure out the number of bytes available for future to-space. 2881 // For new regions without marking information, we must assume the 2882 // worst-case of complete survival. If we have marking information for a 2883 // region, we can bound the amount of live data. We can add a number of 2884 // such regions, as long as the sum of the live data bounds does not 2885 // exceed the available evacuation space. 2886 size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes; 2887 2888 size_t expansion_bytes = 2889 _g1->expansion_regions() * HeapRegion::GrainBytes; 2890 2891 _collection_set_bytes_used_before = 0; 2892 _collection_set_size = 0; 2893 2894 // Adjust for expansion and slop. 2895 max_live_bytes = max_live_bytes + expansion_bytes; 2896 2897 assert(_g1->regions_accounted_for(), "Region leakage!"); 2898 2899 HeapRegion* hr; 2900 if (in_young_gc_mode()) { 2901 double young_start_time_sec = os::elapsedTime(); 2902 2903 if (G1PolicyVerbose > 0) { 2904 gclog_or_tty->print_cr("Adding %d young regions to the CSet", 2905 _g1->young_list()->length()); 2906 } 2907 2908 _young_cset_length = 0; 2909 _last_young_gc_full = full_young_gcs() ? true : false; 2910 2911 if (_last_young_gc_full) 2912 ++_full_young_pause_num; 2913 else 2914 ++_partial_young_pause_num; 2915 2916 // The young list is laid with the survivor regions from the previous 2917 // pause are appended to the RHS of the young list, i.e. 2918 // [Newly Young Regions ++ Survivors from last pause]. 2919 2920 hr = _g1->young_list()->first_survivor_region(); 2921 while (hr != NULL) { 2922 assert(hr->is_survivor(), "badly formed young list"); 2923 hr->set_young(); 2924 hr = hr->get_next_young_region(); 2925 } 2926 2927 // Clear the fields that point to the survivor list - they are 2928 // all young now. 2929 _g1->young_list()->clear_survivors(); 2930 2931 if (_g1->mark_in_progress()) 2932 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger); 2933 2934 _young_cset_length = _inc_cset_young_index; 2935 _collection_set = _inc_cset_head; 2936 _collection_set_size = _inc_cset_size; 2937 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2938 2939 // For young regions in the collection set, we assume the worst 2940 // case of complete survival 2941 max_live_bytes -= _inc_cset_size * HeapRegion::GrainBytes; 2942 2943 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms; 2944 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; 2945 2946 // The number of recorded young regions is the incremental 2947 // collection set's current size 2948 set_recorded_young_regions(_inc_cset_size); 2949 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2950 set_recorded_young_bytes(_inc_cset_recorded_young_bytes); 2951 #if PREDICTIONS_VERBOSE 2952 set_predicted_bytes_to_copy(_inc_cset_predicted_bytes_to_copy); 2953 #endif // PREDICTIONS_VERBOSE 2954 2955 if (G1PolicyVerbose > 0) { 2956 gclog_or_tty->print_cr(" Added " PTR_FORMAT " Young Regions to CS.", 2957 _inc_cset_size); 2958 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)", 2959 max_live_bytes/K); 2960 } 2961 2962 assert(_inc_cset_size == _g1->young_list()->length(), "Invariant"); 2963 2964 double young_end_time_sec = os::elapsedTime(); 2965 _recorded_young_cset_choice_time_ms = 2966 (young_end_time_sec - young_start_time_sec) * 1000.0; 2967 2968 // We are doing young collections so reset this. 2969 non_young_start_time_sec = young_end_time_sec; 2970 2971 // Note we can use either _collection_set_size or 2972 // _young_cset_length here 2973 if (_collection_set_size > 0 && _last_young_gc_full) { 2974 // don't bother adding more regions... 2975 goto choose_collection_set_end; 2976 } 2977 } 2978 2979 if (!in_young_gc_mode() || !full_young_gcs()) { 2980 bool should_continue = true; 2981 NumberSeq seq; 2982 double avg_prediction = 100000000000000000.0; // something very large 2983 2984 do { 2985 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms, 2986 avg_prediction); 2987 if (hr != NULL) { 2988 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); 2989 time_remaining_ms -= predicted_time_ms; 2990 predicted_pause_time_ms += predicted_time_ms; 2991 add_to_collection_set(hr); 2992 record_non_young_cset_region(hr); 2993 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes); 2994 if (G1PolicyVerbose > 0) { 2995 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)", 2996 max_live_bytes/K); 2997 } 2998 seq.add(predicted_time_ms); 2999 avg_prediction = seq.avg() + seq.sd(); 3000 } 3001 should_continue = 3002 ( hr != NULL) && 3003 ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0 3004 : _collection_set_size < _young_list_fixed_length ); 3005 } while (should_continue); 3006 3007 if (!adaptive_young_list_length() && 3008 _collection_set_size < _young_list_fixed_length) 3009 _should_revert_to_full_young_gcs = true; 3010 } 3011 3012 choose_collection_set_end: 3013 stop_incremental_cset_building(); 3014 3015 count_CS_bytes_used(); 3016 3017 end_recording_regions(); 3018 3019 double non_young_end_time_sec = os::elapsedTime(); 3020 _recorded_non_young_cset_choice_time_ms = 3021 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; 3022 } 3023 3024 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() { 3025 G1CollectorPolicy::record_full_collection_end(); 3026 _collectionSetChooser->updateAfterFullCollection(); 3027 } 3028 3029 void G1CollectorPolicy_BestRegionsFirst:: 3030 expand_if_possible(size_t numRegions) { 3031 size_t expansion_bytes = numRegions * HeapRegion::GrainBytes; 3032 _g1->expand(expansion_bytes); 3033 } 3034 3035 void G1CollectorPolicy_BestRegionsFirst:: 3036 record_collection_pause_end() { 3037 G1CollectorPolicy::record_collection_pause_end(); 3038 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end."); 3039 }