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