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