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/g1ErgoVerbose.hpp" 32 #include "gc_implementation/g1/heapRegionRemSet.hpp" 33 #include "gc_implementation/shared/gcPolicyCounters.hpp" 34 #include "runtime/arguments.hpp" 35 #include "runtime/java.hpp" 36 #include "runtime/mutexLocker.hpp" 37 #include "utilities/debug.hpp" 38 39 // Different defaults for different number of GC threads 40 // They were chosen by running GCOld and SPECjbb on debris with different 41 // numbers of GC threads and choosing them based on the results 42 43 // all the same 44 static double rs_length_diff_defaults[] = { 45 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 46 }; 47 48 static double cost_per_card_ms_defaults[] = { 49 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 50 }; 51 52 // all the same 53 static double young_cards_per_entry_ratio_defaults[] = { 54 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 55 }; 56 57 static double cost_per_entry_ms_defaults[] = { 58 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 59 }; 60 61 static double cost_per_byte_ms_defaults[] = { 62 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 63 }; 64 65 // these should be pretty consistent 66 static double constant_other_time_ms_defaults[] = { 67 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 68 }; 69 70 71 static double young_other_cost_per_region_ms_defaults[] = { 72 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 73 }; 74 75 static double non_young_other_cost_per_region_ms_defaults[] = { 76 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 77 }; 78 79 // Help class for avoiding interleaved logging 80 class LineBuffer: public StackObj { 81 82 private: 83 static const int BUFFER_LEN = 1024; 84 static const int INDENT_CHARS = 3; 85 char _buffer[BUFFER_LEN]; 86 int _indent_level; 87 int _cur; 88 89 void vappend(const char* format, va_list ap) { 90 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap); 91 if (res != -1) { 92 _cur += res; 93 } else { 94 DEBUG_ONLY(warning("buffer too small in LineBuffer");) 95 _buffer[BUFFER_LEN -1] = 0; 96 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again 97 } 98 } 99 100 public: 101 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) { 102 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) { 103 _buffer[_cur] = ' '; 104 } 105 } 106 107 #ifndef PRODUCT 108 ~LineBuffer() { 109 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?"); 110 } 111 #endif 112 113 void append(const char* format, ...) { 114 va_list ap; 115 va_start(ap, format); 116 vappend(format, ap); 117 va_end(ap); 118 } 119 120 void append_and_print_cr(const char* format, ...) { 121 va_list ap; 122 va_start(ap, format); 123 vappend(format, ap); 124 va_end(ap); 125 gclog_or_tty->print_cr("%s", _buffer); 126 _cur = _indent_level * INDENT_CHARS; 127 } 128 }; 129 130 G1CollectorPolicy::G1CollectorPolicy() : 131 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads() 132 ? ParallelGCThreads : 1), 133 134 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 135 _all_pause_times_ms(new NumberSeq()), 136 _stop_world_start(0.0), 137 _all_stop_world_times_ms(new NumberSeq()), 138 _all_yield_times_ms(new NumberSeq()), 139 _using_new_ratio_calculations(false), 140 141 _summary(new Summary()), 142 143 _cur_clear_ct_time_ms(0.0), 144 _mark_closure_time_ms(0.0), 145 146 _cur_ref_proc_time_ms(0.0), 147 _cur_ref_enq_time_ms(0.0), 148 149 #ifndef PRODUCT 150 _min_clear_cc_time_ms(-1.0), 151 _max_clear_cc_time_ms(-1.0), 152 _cur_clear_cc_time_ms(0.0), 153 _cum_clear_cc_time_ms(0.0), 154 _num_cc_clears(0L), 155 #endif 156 157 _aux_num(10), 158 _all_aux_times_ms(new NumberSeq[_aux_num]), 159 _cur_aux_start_times_ms(new double[_aux_num]), 160 _cur_aux_times_ms(new double[_aux_num]), 161 _cur_aux_times_set(new bool[_aux_num]), 162 163 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 164 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 165 166 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 167 _prev_collection_pause_end_ms(0.0), 168 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 169 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 170 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 171 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 172 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 173 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 174 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 175 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 176 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 177 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 178 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 179 _non_young_other_cost_per_region_ms_seq( 180 new TruncatedSeq(TruncatedSeqLength)), 181 182 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 183 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 184 185 _pause_time_target_ms((double) MaxGCPauseMillis), 186 187 _gcs_are_young(true), 188 _young_pause_num(0), 189 _mixed_pause_num(0), 190 191 _during_marking(false), 192 _in_marking_window(false), 193 _in_marking_window_im(false), 194 195 _known_garbage_ratio(0.0), 196 _known_garbage_bytes(0), 197 198 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)), 199 200 _recent_prev_end_times_for_all_gcs_sec( 201 new TruncatedSeq(NumPrevPausesForHeuristics)), 202 203 _recent_avg_pause_time_ratio(0.0), 204 205 _all_full_gc_times_ms(new NumberSeq()), 206 207 _initiate_conc_mark_if_possible(false), 208 _during_initial_mark_pause(false), 209 _should_revert_to_young_gcs(false), 210 _last_young_gc(false), 211 _last_gc_was_young(false), 212 213 _eden_bytes_before_gc(0), 214 _survivor_bytes_before_gc(0), 215 _capacity_before_gc(0), 216 217 _prev_collection_pause_used_at_end_bytes(0), 218 219 _eden_cset_region_length(0), 220 _survivor_cset_region_length(0), 221 _old_cset_region_length(0), 222 223 _collection_set(NULL), 224 _collection_set_bytes_used_before(0), 225 226 // Incremental CSet attributes 227 _inc_cset_build_state(Inactive), 228 _inc_cset_head(NULL), 229 _inc_cset_tail(NULL), 230 _inc_cset_bytes_used_before(0), 231 _inc_cset_max_finger(NULL), 232 _inc_cset_recorded_rs_lengths(0), 233 _inc_cset_predicted_elapsed_time_ms(0.0), 234 235 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 236 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 237 #endif // _MSC_VER 238 239 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", 240 G1YoungSurvRateNumRegionsSummary)), 241 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", 242 G1YoungSurvRateNumRegionsSummary)), 243 // add here any more surv rate groups 244 _recorded_survivor_regions(0), 245 _recorded_survivor_head(NULL), 246 _recorded_survivor_tail(NULL), 247 _survivors_age_table(true), 248 249 _gc_overhead_perc(0.0) { 250 251 // Set up the region size and associated fields. Given that the 252 // policy is created before the heap, we have to set this up here, 253 // so it's done as soon as possible. 254 HeapRegion::setup_heap_region_size(Arguments::min_heap_size()); 255 HeapRegionRemSet::setup_remset_size(); 256 257 G1ErgoVerbose::initialize(); 258 if (PrintAdaptiveSizePolicy) { 259 // Currently, we only use a single switch for all the heuristics. 260 G1ErgoVerbose::set_enabled(true); 261 // Given that we don't currently have a verboseness level 262 // parameter, we'll hardcode this to high. This can be easily 263 // changed in the future. 264 G1ErgoVerbose::set_level(ErgoHigh); 265 } else { 266 G1ErgoVerbose::set_enabled(false); 267 } 268 269 // Verify PLAB sizes 270 const size_t region_size = HeapRegion::GrainWords; 271 if (YoungPLABSize > region_size || OldPLABSize > region_size) { 272 char buffer[128]; 273 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT, 274 OldPLABSize > region_size ? "Old" : "Young", region_size); 275 vm_exit_during_initialization(buffer); 276 } 277 278 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 279 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 280 281 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads]; 282 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads]; 283 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads]; 284 285 _par_last_update_rs_times_ms = new double[_parallel_gc_threads]; 286 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; 287 288 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; 289 290 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads]; 291 292 _par_last_termination_times_ms = new double[_parallel_gc_threads]; 293 _par_last_termination_attempts = new double[_parallel_gc_threads]; 294 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads]; 295 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads]; 296 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads]; 297 298 // start conservatively 299 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis; 300 301 int index; 302 if (ParallelGCThreads == 0) 303 index = 0; 304 else if (ParallelGCThreads > 8) 305 index = 7; 306 else 307 index = ParallelGCThreads - 1; 308 309 _pending_card_diff_seq->add(0.0); 310 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 311 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 312 _young_cards_per_entry_ratio_seq->add( 313 young_cards_per_entry_ratio_defaults[index]); 314 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 315 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 316 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 317 _young_other_cost_per_region_ms_seq->add( 318 young_other_cost_per_region_ms_defaults[index]); 319 _non_young_other_cost_per_region_ms_seq->add( 320 non_young_other_cost_per_region_ms_defaults[index]); 321 322 // Below, we might need to calculate the pause time target based on 323 // the pause interval. When we do so we are going to give G1 maximum 324 // flexibility and allow it to do pauses when it needs to. So, we'll 325 // arrange that the pause interval to be pause time target + 1 to 326 // ensure that a) the pause time target is maximized with respect to 327 // the pause interval and b) we maintain the invariant that pause 328 // time target < pause interval. If the user does not want this 329 // maximum flexibility, they will have to set the pause interval 330 // explicitly. 331 332 // First make sure that, if either parameter is set, its value is 333 // reasonable. 334 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 335 if (MaxGCPauseMillis < 1) { 336 vm_exit_during_initialization("MaxGCPauseMillis should be " 337 "greater than 0"); 338 } 339 } 340 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 341 if (GCPauseIntervalMillis < 1) { 342 vm_exit_during_initialization("GCPauseIntervalMillis should be " 343 "greater than 0"); 344 } 345 } 346 347 // Then, if the pause time target parameter was not set, set it to 348 // the default value. 349 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 350 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 351 // The default pause time target in G1 is 200ms 352 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 353 } else { 354 // We do not allow the pause interval to be set without the 355 // pause time target 356 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 357 "without setting MaxGCPauseMillis"); 358 } 359 } 360 361 // Then, if the interval parameter was not set, set it according to 362 // the pause time target (this will also deal with the case when the 363 // pause time target is the default value). 364 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 365 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 366 } 367 368 // Finally, make sure that the two parameters are consistent. 369 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 370 char buffer[256]; 371 jio_snprintf(buffer, 256, 372 "MaxGCPauseMillis (%u) should be less than " 373 "GCPauseIntervalMillis (%u)", 374 MaxGCPauseMillis, GCPauseIntervalMillis); 375 vm_exit_during_initialization(buffer); 376 } 377 378 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 379 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 380 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 381 _sigma = (double) G1ConfidencePercent / 100.0; 382 383 // start conservatively (around 50ms is about right) 384 _concurrent_mark_remark_times_ms->add(0.05); 385 _concurrent_mark_cleanup_times_ms->add(0.20); 386 _tenuring_threshold = MaxTenuringThreshold; 387 // _max_survivor_regions will be calculated by 388 // update_young_list_target_length() during initialization. 389 _max_survivor_regions = 0; 390 391 assert(GCTimeRatio > 0, 392 "we should have set it to a default value set_g1_gc_flags() " 393 "if a user set it to 0"); 394 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 395 396 uintx reserve_perc = G1ReservePercent; 397 // Put an artificial ceiling on this so that it's not set to a silly value. 398 if (reserve_perc > 50) { 399 reserve_perc = 50; 400 warning("G1ReservePercent is set to a value that is too large, " 401 "it's been updated to %u", reserve_perc); 402 } 403 _reserve_factor = (double) reserve_perc / 100.0; 404 // This will be set when the heap is expanded 405 // for the first time during initialization. 406 _reserve_regions = 0; 407 408 initialize_all(); 409 _collectionSetChooser = new CollectionSetChooser(); 410 } 411 412 // Increment "i", mod "len" 413 static void inc_mod(int& i, int len) { 414 i++; if (i == len) i = 0; 415 } 416 417 void G1CollectorPolicy::initialize_flags() { 418 set_min_alignment(HeapRegion::GrainBytes); 419 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name())); 420 if (SurvivorRatio < 1) { 421 vm_exit_during_initialization("Invalid survivor ratio specified"); 422 } 423 CollectorPolicy::initialize_flags(); 424 } 425 426 // The easiest way to deal with the parsing of the NewSize / 427 // MaxNewSize / etc. parameteres is to re-use the code in the 428 // TwoGenerationCollectorPolicy class. This is similar to what 429 // ParallelScavenge does with its GenerationSizer class (see 430 // ParallelScavengeHeap::initialize()). We might change this in the 431 // future, but it's a good start. 432 class G1YoungGenSizer : public TwoGenerationCollectorPolicy { 433 private: 434 size_t size_to_region_num(size_t byte_size) { 435 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes); 436 } 437 438 public: 439 G1YoungGenSizer() { 440 initialize_flags(); 441 initialize_size_info(); 442 } 443 size_t min_young_region_num() { 444 return size_to_region_num(_min_gen0_size); 445 } 446 size_t initial_young_region_num() { 447 return size_to_region_num(_initial_gen0_size); 448 } 449 size_t max_young_region_num() { 450 return size_to_region_num(_max_gen0_size); 451 } 452 }; 453 454 void G1CollectorPolicy::update_young_list_size_using_newratio(size_t number_of_heap_regions) { 455 assert(number_of_heap_regions > 0, "Heap must be initialized"); 456 size_t young_size = number_of_heap_regions / (NewRatio + 1); 457 _min_desired_young_length = young_size; 458 _max_desired_young_length = young_size; 459 } 460 461 void G1CollectorPolicy::init() { 462 // Set aside an initial future to_space. 463 _g1 = G1CollectedHeap::heap(); 464 465 assert(Heap_lock->owned_by_self(), "Locking discipline."); 466 467 initialize_gc_policy_counters(); 468 469 G1YoungGenSizer sizer; 470 _min_desired_young_length = sizer.min_young_region_num(); 471 _max_desired_young_length = sizer.max_young_region_num(); 472 473 if (FLAG_IS_CMDLINE(NewRatio)) { 474 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 475 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 476 } else { 477 // Treat NewRatio as a fixed size that is only recalculated when the heap size changes 478 update_young_list_size_using_newratio(_g1->n_regions()); 479 _using_new_ratio_calculations = true; 480 } 481 } 482 483 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values"); 484 485 set_adaptive_young_list_length(_min_desired_young_length < _max_desired_young_length); 486 if (adaptive_young_list_length()) { 487 _young_list_fixed_length = 0; 488 } else { 489 assert(_min_desired_young_length == _max_desired_young_length, "Min and max young size differ"); 490 _young_list_fixed_length = _min_desired_young_length; 491 } 492 _free_regions_at_end_of_collection = _g1->free_regions(); 493 update_young_list_target_length(); 494 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes; 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 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 504 } 505 506 bool G1CollectorPolicy::predict_will_fit(size_t young_length, 507 double base_time_ms, 508 size_t base_free_regions, 509 double target_pause_time_ms) { 510 if (young_length >= base_free_regions) { 511 // end condition 1: not enough space for the young regions 512 return false; 513 } 514 515 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1)); 516 size_t bytes_to_copy = 517 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 518 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 519 double young_other_time_ms = predict_young_other_time_ms(young_length); 520 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 521 if (pause_time_ms > target_pause_time_ms) { 522 // end condition 2: prediction is over the target pause time 523 return false; 524 } 525 526 size_t free_bytes = 527 (base_free_regions - young_length) * HeapRegion::GrainBytes; 528 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) { 529 // end condition 3: out-of-space (conservatively!) 530 return false; 531 } 532 533 // success! 534 return true; 535 } 536 537 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) { 538 // re-calculate the necessary reserve 539 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 540 // We use ceiling so that if reserve_regions_d is > 0.0 (but 541 // smaller than 1.0) we'll get 1. 542 _reserve_regions = (size_t) ceil(reserve_regions_d); 543 544 if (_using_new_ratio_calculations) { 545 // -XX:NewRatio was specified so we need to update the 546 // young gen length when the heap size has changed. 547 update_young_list_size_using_newratio(new_number_of_regions); 548 } 549 } 550 551 size_t G1CollectorPolicy::calculate_young_list_desired_min_length( 552 size_t base_min_length) { 553 size_t desired_min_length = 0; 554 if (adaptive_young_list_length()) { 555 if (_alloc_rate_ms_seq->num() > 3) { 556 double now_sec = os::elapsedTime(); 557 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 558 double alloc_rate_ms = predict_alloc_rate_ms(); 559 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms); 560 } else { 561 // otherwise we don't have enough info to make the prediction 562 } 563 } 564 desired_min_length += base_min_length; 565 // make sure we don't go below any user-defined minimum bound 566 return MAX2(_min_desired_young_length, desired_min_length); 567 } 568 569 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() { 570 // Here, we might want to also take into account any additional 571 // constraints (i.e., user-defined minimum bound). Currently, we 572 // effectively don't set this bound. 573 return _max_desired_young_length; 574 } 575 576 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 577 if (rs_lengths == (size_t) -1) { 578 // if it's set to the default value (-1), we should predict it; 579 // otherwise, use the given value. 580 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); 581 } 582 583 // Calculate the absolute and desired min bounds. 584 585 // This is how many young regions we already have (currently: the survivors). 586 size_t base_min_length = recorded_survivor_regions(); 587 // This is the absolute minimum young length, which ensures that we 588 // can allocate one eden region in the worst-case. 589 size_t absolute_min_length = base_min_length + 1; 590 size_t desired_min_length = 591 calculate_young_list_desired_min_length(base_min_length); 592 if (desired_min_length < absolute_min_length) { 593 desired_min_length = absolute_min_length; 594 } 595 596 // Calculate the absolute and desired max bounds. 597 598 // We will try our best not to "eat" into the reserve. 599 size_t absolute_max_length = 0; 600 if (_free_regions_at_end_of_collection > _reserve_regions) { 601 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 602 } 603 size_t desired_max_length = calculate_young_list_desired_max_length(); 604 if (desired_max_length > absolute_max_length) { 605 desired_max_length = absolute_max_length; 606 } 607 608 size_t young_list_target_length = 0; 609 if (adaptive_young_list_length()) { 610 if (gcs_are_young()) { 611 young_list_target_length = 612 calculate_young_list_target_length(rs_lengths, 613 base_min_length, 614 desired_min_length, 615 desired_max_length); 616 _rs_lengths_prediction = rs_lengths; 617 } else { 618 // Don't calculate anything and let the code below bound it to 619 // the desired_min_length, i.e., do the next GC as soon as 620 // possible to maximize how many old regions we can add to it. 621 } 622 } else { 623 if (gcs_are_young()) { 624 young_list_target_length = _young_list_fixed_length; 625 } else { 626 // A bit arbitrary: during mixed GCs we allocate half 627 // the young regions to try to add old regions to the CSet. 628 young_list_target_length = _young_list_fixed_length / 2; 629 // We choose to accept that we might go under the desired min 630 // length given that we intentionally ask for a smaller young gen. 631 desired_min_length = absolute_min_length; 632 } 633 } 634 635 // Make sure we don't go over the desired max length, nor under the 636 // desired min length. In case they clash, desired_min_length wins 637 // which is why that test is second. 638 if (young_list_target_length > desired_max_length) { 639 young_list_target_length = desired_max_length; 640 } 641 if (young_list_target_length < desired_min_length) { 642 young_list_target_length = desired_min_length; 643 } 644 645 assert(young_list_target_length > recorded_survivor_regions(), 646 "we should be able to allocate at least one eden region"); 647 assert(young_list_target_length >= absolute_min_length, "post-condition"); 648 _young_list_target_length = young_list_target_length; 649 650 update_max_gc_locker_expansion(); 651 } 652 653 size_t 654 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 655 size_t base_min_length, 656 size_t desired_min_length, 657 size_t desired_max_length) { 658 assert(adaptive_young_list_length(), "pre-condition"); 659 assert(gcs_are_young(), "only call this for young GCs"); 660 661 // In case some edge-condition makes the desired max length too small... 662 if (desired_max_length <= desired_min_length) { 663 return desired_min_length; 664 } 665 666 // We'll adjust min_young_length and max_young_length not to include 667 // the already allocated young regions (i.e., so they reflect the 668 // min and max eden regions we'll allocate). The base_min_length 669 // will be reflected in the predictions by the 670 // survivor_regions_evac_time prediction. 671 assert(desired_min_length > base_min_length, "invariant"); 672 size_t min_young_length = desired_min_length - base_min_length; 673 assert(desired_max_length > base_min_length, "invariant"); 674 size_t max_young_length = desired_max_length - base_min_length; 675 676 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 677 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 678 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 679 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 680 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 681 double base_time_ms = 682 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 683 survivor_regions_evac_time; 684 size_t available_free_regions = _free_regions_at_end_of_collection; 685 size_t base_free_regions = 0; 686 if (available_free_regions > _reserve_regions) { 687 base_free_regions = available_free_regions - _reserve_regions; 688 } 689 690 // Here, we will make sure that the shortest young length that 691 // makes sense fits within the target pause time. 692 693 if (predict_will_fit(min_young_length, base_time_ms, 694 base_free_regions, target_pause_time_ms)) { 695 // The shortest young length will fit into the target pause time; 696 // we'll now check whether the absolute maximum number of young 697 // regions will fit in the target pause time. If not, we'll do 698 // a binary search between min_young_length and max_young_length. 699 if (predict_will_fit(max_young_length, base_time_ms, 700 base_free_regions, target_pause_time_ms)) { 701 // The maximum young length will fit into the target pause time. 702 // We are done so set min young length to the maximum length (as 703 // the result is assumed to be returned in min_young_length). 704 min_young_length = max_young_length; 705 } else { 706 // The maximum possible number of young regions will not fit within 707 // the target pause time so we'll search for the optimal 708 // length. The loop invariants are: 709 // 710 // min_young_length < max_young_length 711 // min_young_length is known to fit into the target pause time 712 // max_young_length is known not to fit into the target pause time 713 // 714 // Going into the loop we know the above hold as we've just 715 // checked them. Every time around the loop we check whether 716 // the middle value between min_young_length and 717 // max_young_length fits into the target pause time. If it 718 // does, it becomes the new min. If it doesn't, it becomes 719 // the new max. This way we maintain the loop invariants. 720 721 assert(min_young_length < max_young_length, "invariant"); 722 size_t diff = (max_young_length - min_young_length) / 2; 723 while (diff > 0) { 724 size_t young_length = min_young_length + diff; 725 if (predict_will_fit(young_length, base_time_ms, 726 base_free_regions, target_pause_time_ms)) { 727 min_young_length = young_length; 728 } else { 729 max_young_length = young_length; 730 } 731 assert(min_young_length < max_young_length, "invariant"); 732 diff = (max_young_length - min_young_length) / 2; 733 } 734 // The results is min_young_length which, according to the 735 // loop invariants, should fit within the target pause time. 736 737 // These are the post-conditions of the binary search above: 738 assert(min_young_length < max_young_length, 739 "otherwise we should have discovered that max_young_length " 740 "fits into the pause target and not done the binary search"); 741 assert(predict_will_fit(min_young_length, base_time_ms, 742 base_free_regions, target_pause_time_ms), 743 "min_young_length, the result of the binary search, should " 744 "fit into the pause target"); 745 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 746 base_free_regions, target_pause_time_ms), 747 "min_young_length, the result of the binary search, should be " 748 "optimal, so no larger length should fit into the pause target"); 749 } 750 } else { 751 // Even the minimum length doesn't fit into the pause time 752 // target, return it as the result nevertheless. 753 } 754 return base_min_length + min_young_length; 755 } 756 757 double G1CollectorPolicy::predict_survivor_regions_evac_time() { 758 double survivor_regions_evac_time = 0.0; 759 for (HeapRegion * r = _recorded_survivor_head; 760 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 761 r = r->get_next_young_region()) { 762 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true); 763 } 764 return survivor_regions_evac_time; 765 } 766 767 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 768 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 769 770 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 771 if (rs_lengths > _rs_lengths_prediction) { 772 // add 10% to avoid having to recalculate often 773 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 774 update_young_list_target_length(rs_lengths_prediction); 775 } 776 } 777 778 779 780 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 781 bool is_tlab, 782 bool* gc_overhead_limit_was_exceeded) { 783 guarantee(false, "Not using this policy feature yet."); 784 return NULL; 785 } 786 787 // This method controls how a collector handles one or more 788 // of its generations being fully allocated. 789 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 790 bool is_tlab) { 791 guarantee(false, "Not using this policy feature yet."); 792 return NULL; 793 } 794 795 796 #ifndef PRODUCT 797 bool G1CollectorPolicy::verify_young_ages() { 798 HeapRegion* head = _g1->young_list()->first_region(); 799 return 800 verify_young_ages(head, _short_lived_surv_rate_group); 801 // also call verify_young_ages on any additional surv rate groups 802 } 803 804 bool 805 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 806 SurvRateGroup *surv_rate_group) { 807 guarantee( surv_rate_group != NULL, "pre-condition" ); 808 809 const char* name = surv_rate_group->name(); 810 bool ret = true; 811 int prev_age = -1; 812 813 for (HeapRegion* curr = head; 814 curr != NULL; 815 curr = curr->get_next_young_region()) { 816 SurvRateGroup* group = curr->surv_rate_group(); 817 if (group == NULL && !curr->is_survivor()) { 818 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 819 ret = false; 820 } 821 822 if (surv_rate_group == group) { 823 int age = curr->age_in_surv_rate_group(); 824 825 if (age < 0) { 826 gclog_or_tty->print_cr("## %s: encountered negative age", name); 827 ret = false; 828 } 829 830 if (age <= prev_age) { 831 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 832 "(%d, %d)", name, age, prev_age); 833 ret = false; 834 } 835 prev_age = age; 836 } 837 } 838 839 return ret; 840 } 841 #endif // PRODUCT 842 843 void G1CollectorPolicy::record_full_collection_start() { 844 _cur_collection_start_sec = os::elapsedTime(); 845 // Release the future to-space so that it is available for compaction into. 846 _g1->set_full_collection(); 847 } 848 849 void G1CollectorPolicy::record_full_collection_end() { 850 // Consider this like a collection pause for the purposes of allocation 851 // since last pause. 852 double end_sec = os::elapsedTime(); 853 double full_gc_time_sec = end_sec - _cur_collection_start_sec; 854 double full_gc_time_ms = full_gc_time_sec * 1000.0; 855 856 _all_full_gc_times_ms->add(full_gc_time_ms); 857 858 update_recent_gc_times(end_sec, full_gc_time_ms); 859 860 _g1->clear_full_collection(); 861 862 // "Nuke" the heuristics that control the young/mixed GC 863 // transitions and make sure we start with young GCs after the Full GC. 864 set_gcs_are_young(true); 865 _last_young_gc = false; 866 _should_revert_to_young_gcs = false; 867 clear_initiate_conc_mark_if_possible(); 868 clear_during_initial_mark_pause(); 869 _known_garbage_bytes = 0; 870 _known_garbage_ratio = 0.0; 871 _in_marking_window = false; 872 _in_marking_window_im = false; 873 874 _short_lived_surv_rate_group->start_adding_regions(); 875 // also call this on any additional surv rate groups 876 877 record_survivor_regions(0, NULL, NULL); 878 879 _free_regions_at_end_of_collection = _g1->free_regions(); 880 // Reset survivors SurvRateGroup. 881 _survivor_surv_rate_group->reset(); 882 update_young_list_target_length(); 883 _collectionSetChooser->updateAfterFullCollection(); 884 } 885 886 void G1CollectorPolicy::record_stop_world_start() { 887 _stop_world_start = os::elapsedTime(); 888 } 889 890 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, 891 size_t start_used) { 892 if (PrintGCDetails) { 893 gclog_or_tty->stamp(PrintGCTimeStamps); 894 gclog_or_tty->print("[GC pause"); 895 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed"); 896 } 897 898 // We only need to do this here as the policy will only be applied 899 // to the GC we're about to start. so, no point is calculating this 900 // every time we calculate / recalculate the target young length. 901 update_survivors_policy(); 902 903 assert(_g1->used() == _g1->recalculate_used(), 904 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT, 905 _g1->used(), _g1->recalculate_used())); 906 907 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 908 _all_stop_world_times_ms->add(s_w_t_ms); 909 _stop_world_start = 0.0; 910 911 _cur_collection_start_sec = start_time_sec; 912 _cur_collection_pause_used_at_start_bytes = start_used; 913 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); 914 _pending_cards = _g1->pending_card_num(); 915 _max_pending_cards = _g1->max_pending_card_num(); 916 917 _bytes_in_collection_set_before_gc = 0; 918 _bytes_copied_during_gc = 0; 919 920 YoungList* young_list = _g1->young_list(); 921 _eden_bytes_before_gc = young_list->eden_used_bytes(); 922 _survivor_bytes_before_gc = young_list->survivor_used_bytes(); 923 _capacity_before_gc = _g1->capacity(); 924 925 #ifdef DEBUG 926 // initialise these to something well known so that we can spot 927 // if they are not set properly 928 929 for (int i = 0; i < _parallel_gc_threads; ++i) { 930 _par_last_gc_worker_start_times_ms[i] = -1234.0; 931 _par_last_ext_root_scan_times_ms[i] = -1234.0; 932 _par_last_mark_stack_scan_times_ms[i] = -1234.0; 933 _par_last_update_rs_times_ms[i] = -1234.0; 934 _par_last_update_rs_processed_buffers[i] = -1234.0; 935 _par_last_scan_rs_times_ms[i] = -1234.0; 936 _par_last_obj_copy_times_ms[i] = -1234.0; 937 _par_last_termination_times_ms[i] = -1234.0; 938 _par_last_termination_attempts[i] = -1234.0; 939 _par_last_gc_worker_end_times_ms[i] = -1234.0; 940 _par_last_gc_worker_times_ms[i] = -1234.0; 941 _par_last_gc_worker_other_times_ms[i] = -1234.0; 942 } 943 #endif 944 945 for (int i = 0; i < _aux_num; ++i) { 946 _cur_aux_times_ms[i] = 0.0; 947 _cur_aux_times_set[i] = false; 948 } 949 950 // This is initialized to zero here and is set during 951 // the evacuation pause if marking is in progress. 952 _cur_satb_drain_time_ms = 0.0; 953 954 _last_gc_was_young = false; 955 956 // do that for any other surv rate groups 957 _short_lived_surv_rate_group->stop_adding_regions(); 958 _survivors_age_table.clear(); 959 960 assert( verify_young_ages(), "region age verification" ); 961 } 962 963 void G1CollectorPolicy::record_concurrent_mark_init_end(double 964 mark_init_elapsed_time_ms) { 965 _during_marking = true; 966 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); 967 clear_during_initial_mark_pause(); 968 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 969 } 970 971 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 972 _mark_remark_start_sec = os::elapsedTime(); 973 _during_marking = false; 974 } 975 976 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 977 double end_time_sec = os::elapsedTime(); 978 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 979 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 980 _cur_mark_stop_world_time_ms += elapsed_time_ms; 981 _prev_collection_pause_end_ms += elapsed_time_ms; 982 983 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); 984 } 985 986 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 987 _mark_cleanup_start_sec = os::elapsedTime(); 988 } 989 990 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 991 _should_revert_to_young_gcs = false; 992 _last_young_gc = true; 993 _in_marking_window = false; 994 } 995 996 void G1CollectorPolicy::record_concurrent_pause() { 997 if (_stop_world_start > 0.0) { 998 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 999 _all_yield_times_ms->add(yield_ms); 1000 } 1001 } 1002 1003 void G1CollectorPolicy::record_concurrent_pause_end() { 1004 } 1005 1006 template<class T> 1007 T sum_of(T* sum_arr, int start, int n, int N) { 1008 T sum = (T)0; 1009 for (int i = 0; i < n; i++) { 1010 int j = (start + i) % N; 1011 sum += sum_arr[j]; 1012 } 1013 return sum; 1014 } 1015 1016 void G1CollectorPolicy::print_par_stats(int level, 1017 const char* str, 1018 double* data) { 1019 double min = data[0], max = data[0]; 1020 double total = 0.0; 1021 LineBuffer buf(level); 1022 buf.append("[%s (ms):", str); 1023 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1024 double val = data[i]; 1025 if (val < min) 1026 min = val; 1027 if (val > max) 1028 max = val; 1029 total += val; 1030 buf.append(" %3.1lf", val); 1031 } 1032 buf.append_and_print_cr(""); 1033 double avg = total / (double) no_of_gc_threads(); 1034 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]", 1035 avg, min, max, max - min); 1036 } 1037 1038 void G1CollectorPolicy::print_par_sizes(int level, 1039 const char* str, 1040 double* data) { 1041 double min = data[0], max = data[0]; 1042 double total = 0.0; 1043 LineBuffer buf(level); 1044 buf.append("[%s :", str); 1045 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1046 double val = data[i]; 1047 if (val < min) 1048 min = val; 1049 if (val > max) 1050 max = val; 1051 total += val; 1052 buf.append(" %d", (int) val); 1053 } 1054 buf.append_and_print_cr(""); 1055 double avg = total / (double) no_of_gc_threads(); 1056 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]", 1057 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min); 1058 } 1059 1060 void G1CollectorPolicy::print_stats(int level, 1061 const char* str, 1062 double value) { 1063 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value); 1064 } 1065 1066 void G1CollectorPolicy::print_stats(int level, 1067 const char* str, 1068 int value) { 1069 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value); 1070 } 1071 1072 double G1CollectorPolicy::avg_value(double* data) { 1073 if (G1CollectedHeap::use_parallel_gc_threads()) { 1074 double ret = 0.0; 1075 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1076 ret += data[i]; 1077 } 1078 return ret / (double) no_of_gc_threads(); 1079 } else { 1080 return data[0]; 1081 } 1082 } 1083 1084 double G1CollectorPolicy::max_value(double* data) { 1085 if (G1CollectedHeap::use_parallel_gc_threads()) { 1086 double ret = data[0]; 1087 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1088 if (data[i] > ret) { 1089 ret = data[i]; 1090 } 1091 } 1092 return ret; 1093 } else { 1094 return data[0]; 1095 } 1096 } 1097 1098 double G1CollectorPolicy::sum_of_values(double* data) { 1099 if (G1CollectedHeap::use_parallel_gc_threads()) { 1100 double sum = 0.0; 1101 for (uint i = 0; i < no_of_gc_threads(); i++) { 1102 sum += data[i]; 1103 } 1104 return sum; 1105 } else { 1106 return data[0]; 1107 } 1108 } 1109 1110 double G1CollectorPolicy::max_sum(double* data1, double* data2) { 1111 double ret = data1[0] + data2[0]; 1112 1113 if (G1CollectedHeap::use_parallel_gc_threads()) { 1114 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1115 double data = data1[i] + data2[i]; 1116 if (data > ret) { 1117 ret = data; 1118 } 1119 } 1120 } 1121 return ret; 1122 } 1123 1124 // Anything below that is considered to be zero 1125 #define MIN_TIMER_GRANULARITY 0.0000001 1126 1127 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) { 1128 double end_time_sec = os::elapsedTime(); 1129 double elapsed_ms = _last_pause_time_ms; 1130 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1131 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1132 "otherwise, the subtraction below does not make sense"); 1133 size_t rs_size = 1134 _cur_collection_pause_used_regions_at_start - cset_region_length(); 1135 size_t cur_used_bytes = _g1->used(); 1136 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1137 bool last_pause_included_initial_mark = false; 1138 bool update_stats = !_g1->evacuation_failed(); 1139 set_no_of_gc_threads(no_of_gc_threads); 1140 1141 #ifndef PRODUCT 1142 if (G1YoungSurvRateVerbose) { 1143 gclog_or_tty->print_cr(""); 1144 _short_lived_surv_rate_group->print(); 1145 // do that for any other surv rate groups too 1146 } 1147 #endif // PRODUCT 1148 1149 last_pause_included_initial_mark = during_initial_mark_pause(); 1150 if (last_pause_included_initial_mark) 1151 record_concurrent_mark_init_end(0.0); 1152 1153 size_t marking_initiating_used_threshold = 1154 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 1155 1156 if (!_g1->mark_in_progress() && !_last_young_gc) { 1157 assert(!last_pause_included_initial_mark, "invariant"); 1158 if (cur_used_bytes > marking_initiating_used_threshold) { 1159 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) { 1160 assert(!during_initial_mark_pause(), "we should not see this here"); 1161 1162 ergo_verbose3(ErgoConcCycles, 1163 "request concurrent cycle initiation", 1164 ergo_format_reason("occupancy higher than threshold") 1165 ergo_format_byte("occupancy") 1166 ergo_format_byte_perc("threshold"), 1167 cur_used_bytes, 1168 marking_initiating_used_threshold, 1169 (double) InitiatingHeapOccupancyPercent); 1170 1171 // Note: this might have already been set, if during the last 1172 // pause we decided to start a cycle but at the beginning of 1173 // this pause we decided to postpone it. That's OK. 1174 set_initiate_conc_mark_if_possible(); 1175 } else { 1176 ergo_verbose2(ErgoConcCycles, 1177 "do not request concurrent cycle initiation", 1178 ergo_format_reason("occupancy lower than previous occupancy") 1179 ergo_format_byte("occupancy") 1180 ergo_format_byte("previous occupancy"), 1181 cur_used_bytes, 1182 _prev_collection_pause_used_at_end_bytes); 1183 } 1184 } 1185 } 1186 1187 _prev_collection_pause_used_at_end_bytes = cur_used_bytes; 1188 1189 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, 1190 end_time_sec, false); 1191 1192 // This assert is exempted when we're doing parallel collection pauses, 1193 // because the fragmentation caused by the parallel GC allocation buffers 1194 // can lead to more memory being used during collection than was used 1195 // before. Best leave this out until the fragmentation problem is fixed. 1196 // Pauses in which evacuation failed can also lead to negative 1197 // collections, since no space is reclaimed from a region containing an 1198 // object whose evacuation failed. 1199 // Further, we're now always doing parallel collection. But I'm still 1200 // leaving this here as a placeholder for a more precise assertion later. 1201 // (DLD, 10/05.) 1202 assert((true || parallel) // Always using GC LABs now. 1203 || _g1->evacuation_failed() 1204 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes, 1205 "Negative collection"); 1206 1207 size_t freed_bytes = 1208 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; 1209 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; 1210 1211 double survival_fraction = 1212 (double)surviving_bytes/ 1213 (double)_collection_set_bytes_used_before; 1214 1215 // These values are used to update the summary information that is 1216 // displayed when TraceGen0Time is enabled, and are output as part 1217 // of the PrintGCDetails output, in the non-parallel case. 1218 1219 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); 1220 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms); 1221 double update_rs_time = avg_value(_par_last_update_rs_times_ms); 1222 double update_rs_processed_buffers = 1223 sum_of_values(_par_last_update_rs_processed_buffers); 1224 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); 1225 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); 1226 double termination_time = avg_value(_par_last_termination_times_ms); 1227 1228 double known_time = ext_root_scan_time + 1229 mark_stack_scan_time + 1230 update_rs_time + 1231 scan_rs_time + 1232 obj_copy_time; 1233 1234 double other_time_ms = elapsed_ms; 1235 1236 // Subtract the SATB drain time. It's initialized to zero at the 1237 // start of the pause and is updated during the pause if marking 1238 // is in progress. 1239 other_time_ms -= _cur_satb_drain_time_ms; 1240 1241 if (parallel) { 1242 other_time_ms -= _cur_collection_par_time_ms; 1243 } else { 1244 other_time_ms -= known_time; 1245 } 1246 1247 // Subtract the time taken to clean the card table from the 1248 // current value of "other time" 1249 other_time_ms -= _cur_clear_ct_time_ms; 1250 1251 // Subtract the time spent completing marking in the collection 1252 // set. Note if marking is not in progress during the pause 1253 // the value of _mark_closure_time_ms will be zero. 1254 other_time_ms -= _mark_closure_time_ms; 1255 1256 // TraceGen0Time and TraceGen1Time summary info updating. 1257 _all_pause_times_ms->add(elapsed_ms); 1258 1259 if (update_stats) { 1260 _summary->record_total_time_ms(elapsed_ms); 1261 _summary->record_other_time_ms(other_time_ms); 1262 1263 MainBodySummary* body_summary = _summary->main_body_summary(); 1264 assert(body_summary != NULL, "should not be null!"); 1265 1266 // This will be non-zero iff marking is currently in progress (i.e. 1267 // _g1->mark_in_progress() == true) and the currrent pause was not 1268 // an initial mark pause. Since the body_summary items are NumberSeqs, 1269 // however, they have to be consistent and updated in lock-step with 1270 // each other. Therefore we unconditionally record the SATB drain 1271 // time - even if it's zero. 1272 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms); 1273 1274 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time); 1275 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time); 1276 body_summary->record_update_rs_time_ms(update_rs_time); 1277 body_summary->record_scan_rs_time_ms(scan_rs_time); 1278 body_summary->record_obj_copy_time_ms(obj_copy_time); 1279 1280 if (parallel) { 1281 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms); 1282 body_summary->record_termination_time_ms(termination_time); 1283 1284 double parallel_known_time = known_time + termination_time; 1285 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time; 1286 body_summary->record_parallel_other_time_ms(parallel_other_time); 1287 } 1288 1289 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms); 1290 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms); 1291 1292 // We exempt parallel collection from this check because Alloc Buffer 1293 // fragmentation can produce negative collections. Same with evac 1294 // failure. 1295 // Further, we're now always doing parallel collection. But I'm still 1296 // leaving this here as a placeholder for a more precise assertion later. 1297 // (DLD, 10/05. 1298 assert((true || parallel) 1299 || _g1->evacuation_failed() 1300 || surviving_bytes <= _collection_set_bytes_used_before, 1301 "Or else negative collection!"); 1302 1303 // this is where we update the allocation rate of the application 1304 double app_time_ms = 1305 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); 1306 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1307 // This usually happens due to the timer not having the required 1308 // granularity. Some Linuxes are the usual culprits. 1309 // We'll just set it to something (arbitrarily) small. 1310 app_time_ms = 1.0; 1311 } 1312 // We maintain the invariant that all objects allocated by mutator 1313 // threads will be allocated out of eden regions. So, we can use 1314 // the eden region number allocated since the previous GC to 1315 // calculate the application's allocate rate. The only exception 1316 // to that is humongous objects that are allocated separately. But 1317 // given that humongous object allocations do not really affect 1318 // either the pause's duration nor when the next pause will take 1319 // place we can safely ignore them here. 1320 size_t regions_allocated = eden_cset_region_length(); 1321 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1322 _alloc_rate_ms_seq->add(alloc_rate_ms); 1323 1324 double interval_ms = 1325 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1326 update_recent_gc_times(end_time_sec, elapsed_ms); 1327 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1328 if (recent_avg_pause_time_ratio() < 0.0 || 1329 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1330 #ifndef PRODUCT 1331 // Dump info to allow post-facto debugging 1332 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1333 gclog_or_tty->print_cr("-------------------------------------------"); 1334 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1335 _recent_gc_times_ms->dump(); 1336 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1337 _recent_prev_end_times_for_all_gcs_sec->dump(); 1338 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1339 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1340 // In debug mode, terminate the JVM if the user wants to debug at this point. 1341 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1342 #endif // !PRODUCT 1343 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1344 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1345 if (_recent_avg_pause_time_ratio < 0.0) { 1346 _recent_avg_pause_time_ratio = 0.0; 1347 } else { 1348 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1349 _recent_avg_pause_time_ratio = 1.0; 1350 } 1351 } 1352 } 1353 1354 for (int i = 0; i < _aux_num; ++i) { 1355 if (_cur_aux_times_set[i]) { 1356 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]); 1357 } 1358 } 1359 1360 // PrintGCDetails output 1361 if (PrintGCDetails) { 1362 bool print_marking_info = 1363 _g1->mark_in_progress() && !last_pause_included_initial_mark; 1364 1365 gclog_or_tty->print_cr("%s, %1.8lf secs]", 1366 (last_pause_included_initial_mark) ? " (initial-mark)" : "", 1367 elapsed_ms / 1000.0); 1368 1369 if (print_marking_info) { 1370 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms); 1371 } 1372 1373 if (parallel) { 1374 print_stats(1, "Parallel Time", _cur_collection_par_time_ms); 1375 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms); 1376 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms); 1377 if (print_marking_info) { 1378 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms); 1379 } 1380 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); 1381 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers); 1382 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); 1383 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); 1384 print_par_stats(2, "Termination", _par_last_termination_times_ms); 1385 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts); 1386 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms); 1387 1388 for (int i = 0; i < _parallel_gc_threads; i++) { 1389 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i]; 1390 1391 double worker_known_time = _par_last_ext_root_scan_times_ms[i] + 1392 _par_last_mark_stack_scan_times_ms[i] + 1393 _par_last_update_rs_times_ms[i] + 1394 _par_last_scan_rs_times_ms[i] + 1395 _par_last_obj_copy_times_ms[i] + 1396 _par_last_termination_times_ms[i]; 1397 1398 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time; 1399 } 1400 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms); 1401 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms); 1402 } else { 1403 print_stats(1, "Ext Root Scanning", ext_root_scan_time); 1404 if (print_marking_info) { 1405 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time); 1406 } 1407 print_stats(1, "Update RS", update_rs_time); 1408 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers); 1409 print_stats(1, "Scan RS", scan_rs_time); 1410 print_stats(1, "Object Copying", obj_copy_time); 1411 } 1412 if (print_marking_info) { 1413 print_stats(1, "Complete CSet Marking", _mark_closure_time_ms); 1414 } 1415 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); 1416 #ifndef PRODUCT 1417 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms); 1418 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms); 1419 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms); 1420 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms); 1421 if (_num_cc_clears > 0) { 1422 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears)); 1423 } 1424 #endif 1425 print_stats(1, "Other", other_time_ms); 1426 print_stats(2, "Choose CSet", 1427 (_recorded_young_cset_choice_time_ms + 1428 _recorded_non_young_cset_choice_time_ms)); 1429 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms); 1430 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms); 1431 print_stats(2, "Free CSet", 1432 (_recorded_young_free_cset_time_ms + 1433 _recorded_non_young_free_cset_time_ms)); 1434 1435 for (int i = 0; i < _aux_num; ++i) { 1436 if (_cur_aux_times_set[i]) { 1437 char buffer[96]; 1438 sprintf(buffer, "Aux%d", i); 1439 print_stats(1, buffer, _cur_aux_times_ms[i]); 1440 } 1441 } 1442 } 1443 1444 // Update the efficiency-since-mark vars. 1445 double proc_ms = elapsed_ms * (double) _parallel_gc_threads; 1446 if (elapsed_ms < MIN_TIMER_GRANULARITY) { 1447 // This usually happens due to the timer not having the required 1448 // granularity. Some Linuxes are the usual culprits. 1449 // We'll just set it to something (arbitrarily) small. 1450 proc_ms = 1.0; 1451 } 1452 double cur_efficiency = (double) freed_bytes / proc_ms; 1453 1454 bool new_in_marking_window = _in_marking_window; 1455 bool new_in_marking_window_im = false; 1456 if (during_initial_mark_pause()) { 1457 new_in_marking_window = true; 1458 new_in_marking_window_im = true; 1459 } 1460 1461 if (_last_young_gc) { 1462 if (!last_pause_included_initial_mark) { 1463 ergo_verbose2(ErgoMixedGCs, 1464 "start mixed GCs", 1465 ergo_format_byte_perc("known garbage"), 1466 _known_garbage_bytes, _known_garbage_ratio * 100.0); 1467 set_gcs_are_young(false); 1468 } else { 1469 ergo_verbose0(ErgoMixedGCs, 1470 "do not start mixed GCs", 1471 ergo_format_reason("concurrent cycle is about to start")); 1472 } 1473 _last_young_gc = false; 1474 } 1475 1476 if (!_last_gc_was_young) { 1477 if (_should_revert_to_young_gcs) { 1478 ergo_verbose2(ErgoMixedGCs, 1479 "end mixed GCs", 1480 ergo_format_reason("mixed GCs end requested") 1481 ergo_format_byte_perc("known garbage"), 1482 _known_garbage_bytes, _known_garbage_ratio * 100.0); 1483 set_gcs_are_young(true); 1484 } else if (_known_garbage_ratio < 0.05) { 1485 ergo_verbose3(ErgoMixedGCs, 1486 "end mixed GCs", 1487 ergo_format_reason("known garbage percent lower than threshold") 1488 ergo_format_byte_perc("known garbage") 1489 ergo_format_perc("threshold"), 1490 _known_garbage_bytes, _known_garbage_ratio * 100.0, 1491 0.05 * 100.0); 1492 set_gcs_are_young(true); 1493 } else if (adaptive_young_list_length() && 1494 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) { 1495 ergo_verbose5(ErgoMixedGCs, 1496 "end mixed GCs", 1497 ergo_format_reason("current GC efficiency lower than " 1498 "predicted young GC efficiency") 1499 ergo_format_double("GC efficiency factor") 1500 ergo_format_double("current GC efficiency") 1501 ergo_format_double("predicted young GC efficiency") 1502 ergo_format_byte_perc("known garbage"), 1503 get_gc_eff_factor(), cur_efficiency, 1504 predict_young_gc_eff(), 1505 _known_garbage_bytes, _known_garbage_ratio * 100.0); 1506 set_gcs_are_young(true); 1507 } 1508 } 1509 _should_revert_to_young_gcs = false; 1510 1511 if (_last_gc_was_young && !_during_marking) { 1512 _young_gc_eff_seq->add(cur_efficiency); 1513 } 1514 1515 _short_lived_surv_rate_group->start_adding_regions(); 1516 // do that for any other surv rate groupsx 1517 1518 if (update_stats) { 1519 double pause_time_ms = elapsed_ms; 1520 1521 size_t diff = 0; 1522 if (_max_pending_cards >= _pending_cards) 1523 diff = _max_pending_cards - _pending_cards; 1524 _pending_card_diff_seq->add((double) diff); 1525 1526 double cost_per_card_ms = 0.0; 1527 if (_pending_cards > 0) { 1528 cost_per_card_ms = update_rs_time / (double) _pending_cards; 1529 _cost_per_card_ms_seq->add(cost_per_card_ms); 1530 } 1531 1532 size_t cards_scanned = _g1->cards_scanned(); 1533 1534 double cost_per_entry_ms = 0.0; 1535 if (cards_scanned > 10) { 1536 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; 1537 if (_last_gc_was_young) { 1538 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1539 } else { 1540 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1541 } 1542 } 1543 1544 if (_max_rs_lengths > 0) { 1545 double cards_per_entry_ratio = 1546 (double) cards_scanned / (double) _max_rs_lengths; 1547 if (_last_gc_was_young) { 1548 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1549 } else { 1550 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1551 } 1552 } 1553 1554 // It turns out that, sometimes, _max_rs_lengths can get smaller 1555 // than _recorded_rs_lengths which causes rs_length_diff to get 1556 // very large and mess up the RSet length predictions. We'll be 1557 // defensive until we work out why this happens. 1558 size_t rs_length_diff = 0; 1559 if (_max_rs_lengths > _recorded_rs_lengths) { 1560 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1561 } 1562 _rs_length_diff_seq->add((double) rs_length_diff); 1563 1564 size_t copied_bytes = surviving_bytes; 1565 double cost_per_byte_ms = 0.0; 1566 if (copied_bytes > 0) { 1567 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; 1568 if (_in_marking_window) { 1569 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1570 } else { 1571 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1572 } 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 (young_cset_region_length() > 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) young_cset_region_length()); 1586 } 1587 double non_young_other_time_ms = 0.0; 1588 if (old_cset_region_length() > 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) old_cset_region_length()); 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_copied_during_gc / 1604 (double) _bytes_in_collection_set_before_gc; 1605 } 1606 1607 _pending_cards_seq->add((double) _pending_cards); 1608 _rs_lengths_seq->add((double) _max_rs_lengths); 1609 1610 double expensive_region_limit_ms = 1611 (double) MaxGCPauseMillis - predict_constant_other_time_ms(); 1612 if (expensive_region_limit_ms < 0.0) { 1613 // this means that the other time was predicted to be longer than 1614 // than the max pause time 1615 expensive_region_limit_ms = (double) MaxGCPauseMillis; 1616 } 1617 _expensive_region_limit_ms = expensive_region_limit_ms; 1618 } 1619 1620 _in_marking_window = new_in_marking_window; 1621 _in_marking_window_im = new_in_marking_window_im; 1622 _free_regions_at_end_of_collection = _g1->free_regions(); 1623 update_young_list_target_length(); 1624 1625 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1626 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1627 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms); 1628 1629 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end."); 1630 } 1631 1632 #define EXT_SIZE_FORMAT "%d%s" 1633 #define EXT_SIZE_PARAMS(bytes) \ 1634 byte_size_in_proper_unit((bytes)), \ 1635 proper_unit_for_byte_size((bytes)) 1636 1637 void G1CollectorPolicy::print_heap_transition() { 1638 if (PrintGCDetails) { 1639 YoungList* young_list = _g1->young_list(); 1640 size_t eden_bytes = young_list->eden_used_bytes(); 1641 size_t survivor_bytes = young_list->survivor_used_bytes(); 1642 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes; 1643 size_t used = _g1->used(); 1644 size_t capacity = _g1->capacity(); 1645 size_t eden_capacity = 1646 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes; 1647 1648 gclog_or_tty->print_cr( 1649 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") " 1650 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" " 1651 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->" 1652 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]", 1653 EXT_SIZE_PARAMS(_eden_bytes_before_gc), 1654 EXT_SIZE_PARAMS(_prev_eden_capacity), 1655 EXT_SIZE_PARAMS(eden_bytes), 1656 EXT_SIZE_PARAMS(eden_capacity), 1657 EXT_SIZE_PARAMS(_survivor_bytes_before_gc), 1658 EXT_SIZE_PARAMS(survivor_bytes), 1659 EXT_SIZE_PARAMS(used_before_gc), 1660 EXT_SIZE_PARAMS(_capacity_before_gc), 1661 EXT_SIZE_PARAMS(used), 1662 EXT_SIZE_PARAMS(capacity)); 1663 1664 _prev_eden_capacity = eden_capacity; 1665 } else if (PrintGC) { 1666 _g1->print_size_transition(gclog_or_tty, 1667 _cur_collection_pause_used_at_start_bytes, 1668 _g1->used(), _g1->capacity()); 1669 } 1670 } 1671 1672 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1673 double update_rs_processed_buffers, 1674 double goal_ms) { 1675 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1676 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1677 1678 if (G1UseAdaptiveConcRefinement) { 1679 const int k_gy = 3, k_gr = 6; 1680 const double inc_k = 1.1, dec_k = 0.9; 1681 1682 int g = cg1r->green_zone(); 1683 if (update_rs_time > goal_ms) { 1684 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1685 } else { 1686 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1687 g = (int)MAX2(g * inc_k, g + 1.0); 1688 } 1689 } 1690 // Change the refinement threads params 1691 cg1r->set_green_zone(g); 1692 cg1r->set_yellow_zone(g * k_gy); 1693 cg1r->set_red_zone(g * k_gr); 1694 cg1r->reinitialize_threads(); 1695 1696 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); 1697 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1698 cg1r->yellow_zone()); 1699 // Change the barrier params 1700 dcqs.set_process_completed_threshold(processing_threshold); 1701 dcqs.set_max_completed_queue(cg1r->red_zone()); 1702 } 1703 1704 int curr_queue_size = dcqs.completed_buffers_num(); 1705 if (curr_queue_size >= cg1r->yellow_zone()) { 1706 dcqs.set_completed_queue_padding(curr_queue_size); 1707 } else { 1708 dcqs.set_completed_queue_padding(0); 1709 } 1710 dcqs.notify_if_necessary(); 1711 } 1712 1713 double 1714 G1CollectorPolicy:: 1715 predict_young_collection_elapsed_time_ms(size_t adjustment) { 1716 guarantee( adjustment == 0 || adjustment == 1, "invariant" ); 1717 1718 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1719 size_t young_num = g1h->young_list()->length(); 1720 if (young_num == 0) 1721 return 0.0; 1722 1723 young_num += adjustment; 1724 size_t pending_cards = predict_pending_cards(); 1725 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() + 1726 predict_rs_length_diff(); 1727 size_t card_num; 1728 if (gcs_are_young()) { 1729 card_num = predict_young_card_num(rs_lengths); 1730 } else { 1731 card_num = predict_non_young_card_num(rs_lengths); 1732 } 1733 size_t young_byte_size = young_num * HeapRegion::GrainBytes; 1734 double accum_yg_surv_rate = 1735 _short_lived_surv_rate_group->accum_surv_rate(adjustment); 1736 1737 size_t bytes_to_copy = 1738 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes); 1739 1740 return 1741 predict_rs_update_time_ms(pending_cards) + 1742 predict_rs_scan_time_ms(card_num) + 1743 predict_object_copy_time_ms(bytes_to_copy) + 1744 predict_young_other_time_ms(young_num) + 1745 predict_constant_other_time_ms(); 1746 } 1747 1748 double 1749 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { 1750 size_t rs_length = predict_rs_length_diff(); 1751 size_t card_num; 1752 if (gcs_are_young()) { 1753 card_num = predict_young_card_num(rs_length); 1754 } else { 1755 card_num = predict_non_young_card_num(rs_length); 1756 } 1757 return predict_base_elapsed_time_ms(pending_cards, card_num); 1758 } 1759 1760 double 1761 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1762 size_t scanned_cards) { 1763 return 1764 predict_rs_update_time_ms(pending_cards) + 1765 predict_rs_scan_time_ms(scanned_cards) + 1766 predict_constant_other_time_ms(); 1767 } 1768 1769 double 1770 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1771 bool young) { 1772 size_t rs_length = hr->rem_set()->occupied(); 1773 size_t card_num; 1774 if (gcs_are_young()) { 1775 card_num = predict_young_card_num(rs_length); 1776 } else { 1777 card_num = predict_non_young_card_num(rs_length); 1778 } 1779 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1780 1781 double region_elapsed_time_ms = 1782 predict_rs_scan_time_ms(card_num) + 1783 predict_object_copy_time_ms(bytes_to_copy); 1784 1785 if (young) 1786 region_elapsed_time_ms += predict_young_other_time_ms(1); 1787 else 1788 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1789 1790 return region_elapsed_time_ms; 1791 } 1792 1793 size_t 1794 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { 1795 size_t bytes_to_copy; 1796 if (hr->is_marked()) 1797 bytes_to_copy = hr->max_live_bytes(); 1798 else { 1799 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1, 1800 "invariant" ); 1801 int age = hr->age_in_surv_rate_group(); 1802 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1803 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); 1804 } 1805 1806 return bytes_to_copy; 1807 } 1808 1809 void 1810 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length, 1811 size_t survivor_cset_region_length) { 1812 _eden_cset_region_length = eden_cset_region_length; 1813 _survivor_cset_region_length = survivor_cset_region_length; 1814 _old_cset_region_length = 0; 1815 } 1816 1817 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1818 _recorded_rs_lengths = rs_lengths; 1819 } 1820 1821 void G1CollectorPolicy::check_if_region_is_too_expensive(double 1822 predicted_time_ms) { 1823 // I don't think we need to do this when in young GC mode since 1824 // marking will be initiated next time we hit the soft limit anyway... 1825 if (predicted_time_ms > _expensive_region_limit_ms) { 1826 ergo_verbose2(ErgoMixedGCs, 1827 "request mixed GCs end", 1828 ergo_format_reason("predicted region time higher than threshold") 1829 ergo_format_ms("predicted region time") 1830 ergo_format_ms("threshold"), 1831 predicted_time_ms, _expensive_region_limit_ms); 1832 // no point in doing another mixed GC 1833 _should_revert_to_young_gcs = true; 1834 } 1835 } 1836 1837 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1838 double elapsed_ms) { 1839 _recent_gc_times_ms->add(elapsed_ms); 1840 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1841 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1842 } 1843 1844 size_t G1CollectorPolicy::expansion_amount() { 1845 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1846 double threshold = _gc_overhead_perc; 1847 if (recent_gc_overhead > threshold) { 1848 // We will double the existing space, or take 1849 // G1ExpandByPercentOfAvailable % of the available expansion 1850 // space, whichever is smaller, bounded below by a minimum 1851 // expansion (unless that's all that's left.) 1852 const size_t min_expand_bytes = 1*M; 1853 size_t reserved_bytes = _g1->max_capacity(); 1854 size_t committed_bytes = _g1->capacity(); 1855 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1856 size_t expand_bytes; 1857 size_t expand_bytes_via_pct = 1858 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1859 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1860 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1861 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1862 1863 ergo_verbose5(ErgoHeapSizing, 1864 "attempt heap expansion", 1865 ergo_format_reason("recent GC overhead higher than " 1866 "threshold after GC") 1867 ergo_format_perc("recent GC overhead") 1868 ergo_format_perc("threshold") 1869 ergo_format_byte("uncommitted") 1870 ergo_format_byte_perc("calculated expansion amount"), 1871 recent_gc_overhead, threshold, 1872 uncommitted_bytes, 1873 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1874 1875 return expand_bytes; 1876 } else { 1877 return 0; 1878 } 1879 } 1880 1881 class CountCSClosure: public HeapRegionClosure { 1882 G1CollectorPolicy* _g1_policy; 1883 public: 1884 CountCSClosure(G1CollectorPolicy* g1_policy) : 1885 _g1_policy(g1_policy) {} 1886 bool doHeapRegion(HeapRegion* r) { 1887 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); 1888 return false; 1889 } 1890 }; 1891 1892 void G1CollectorPolicy::count_CS_bytes_used() { 1893 CountCSClosure cs_closure(this); 1894 _g1->collection_set_iterate(&cs_closure); 1895 } 1896 1897 void G1CollectorPolicy::print_summary(int level, 1898 const char* str, 1899 NumberSeq* seq) const { 1900 double sum = seq->sum(); 1901 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", 1902 str, sum / 1000.0, seq->avg()); 1903 } 1904 1905 void G1CollectorPolicy::print_summary_sd(int level, 1906 const char* str, 1907 NumberSeq* seq) const { 1908 print_summary(level, str, seq); 1909 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 1910 seq->num(), seq->sd(), seq->maximum()); 1911 } 1912 1913 void G1CollectorPolicy::check_other_times(int level, 1914 NumberSeq* other_times_ms, 1915 NumberSeq* calc_other_times_ms) const { 1916 bool should_print = false; 1917 LineBuffer buf(level + 2); 1918 1919 double max_sum = MAX2(fabs(other_times_ms->sum()), 1920 fabs(calc_other_times_ms->sum())); 1921 double min_sum = MIN2(fabs(other_times_ms->sum()), 1922 fabs(calc_other_times_ms->sum())); 1923 double sum_ratio = max_sum / min_sum; 1924 if (sum_ratio > 1.1) { 1925 should_print = true; 1926 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###"); 1927 } 1928 1929 double max_avg = MAX2(fabs(other_times_ms->avg()), 1930 fabs(calc_other_times_ms->avg())); 1931 double min_avg = MIN2(fabs(other_times_ms->avg()), 1932 fabs(calc_other_times_ms->avg())); 1933 double avg_ratio = max_avg / min_avg; 1934 if (avg_ratio > 1.1) { 1935 should_print = true; 1936 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###"); 1937 } 1938 1939 if (other_times_ms->sum() < -0.01) { 1940 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###"); 1941 } 1942 1943 if (other_times_ms->avg() < -0.01) { 1944 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###"); 1945 } 1946 1947 if (calc_other_times_ms->sum() < -0.01) { 1948 should_print = true; 1949 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###"); 1950 } 1951 1952 if (calc_other_times_ms->avg() < -0.01) { 1953 should_print = true; 1954 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###"); 1955 } 1956 1957 if (should_print) 1958 print_summary(level, "Other(Calc)", calc_other_times_ms); 1959 } 1960 1961 void G1CollectorPolicy::print_summary(PauseSummary* summary) const { 1962 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1963 MainBodySummary* body_summary = summary->main_body_summary(); 1964 if (summary->get_total_seq()->num() > 0) { 1965 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq()); 1966 if (body_summary != NULL) { 1967 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq()); 1968 if (parallel) { 1969 print_summary(1, "Parallel Time", body_summary->get_parallel_seq()); 1970 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1971 print_summary(2, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq()); 1972 print_summary(2, "Update RS", body_summary->get_update_rs_seq()); 1973 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq()); 1974 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq()); 1975 print_summary(2, "Termination", body_summary->get_termination_seq()); 1976 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq()); 1977 { 1978 NumberSeq* other_parts[] = { 1979 body_summary->get_ext_root_scan_seq(), 1980 body_summary->get_mark_stack_scan_seq(), 1981 body_summary->get_update_rs_seq(), 1982 body_summary->get_scan_rs_seq(), 1983 body_summary->get_obj_copy_seq(), 1984 body_summary->get_termination_seq() 1985 }; 1986 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(), 1987 6, other_parts); 1988 check_other_times(2, body_summary->get_parallel_other_seq(), 1989 &calc_other_times_ms); 1990 } 1991 } else { 1992 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1993 print_summary(1, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq()); 1994 print_summary(1, "Update RS", body_summary->get_update_rs_seq()); 1995 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq()); 1996 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq()); 1997 } 1998 } 1999 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq()); 2000 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq()); 2001 print_summary(1, "Other", summary->get_other_seq()); 2002 { 2003 if (body_summary != NULL) { 2004 NumberSeq calc_other_times_ms; 2005 if (parallel) { 2006 // parallel 2007 NumberSeq* other_parts[] = { 2008 body_summary->get_satb_drain_seq(), 2009 body_summary->get_parallel_seq(), 2010 body_summary->get_clear_ct_seq() 2011 }; 2012 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 2013 3, other_parts); 2014 } else { 2015 // serial 2016 NumberSeq* other_parts[] = { 2017 body_summary->get_satb_drain_seq(), 2018 body_summary->get_update_rs_seq(), 2019 body_summary->get_ext_root_scan_seq(), 2020 body_summary->get_mark_stack_scan_seq(), 2021 body_summary->get_scan_rs_seq(), 2022 body_summary->get_obj_copy_seq() 2023 }; 2024 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 2025 6, other_parts); 2026 } 2027 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms); 2028 } 2029 } 2030 } else { 2031 LineBuffer(1).append_and_print_cr("none"); 2032 } 2033 LineBuffer(0).append_and_print_cr(""); 2034 } 2035 2036 void G1CollectorPolicy::print_tracing_info() const { 2037 if (TraceGen0Time) { 2038 gclog_or_tty->print_cr("ALL PAUSES"); 2039 print_summary_sd(0, "Total", _all_pause_times_ms); 2040 gclog_or_tty->print_cr(""); 2041 gclog_or_tty->print_cr(""); 2042 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2043 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2044 gclog_or_tty->print_cr(""); 2045 2046 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2047 print_summary(_summary); 2048 2049 gclog_or_tty->print_cr("MISC"); 2050 print_summary_sd(0, "Stop World", _all_stop_world_times_ms); 2051 print_summary_sd(0, "Yields", _all_yield_times_ms); 2052 for (int i = 0; i < _aux_num; ++i) { 2053 if (_all_aux_times_ms[i].num() > 0) { 2054 char buffer[96]; 2055 sprintf(buffer, "Aux%d", i); 2056 print_summary_sd(0, buffer, &_all_aux_times_ms[i]); 2057 } 2058 } 2059 } 2060 if (TraceGen1Time) { 2061 if (_all_full_gc_times_ms->num() > 0) { 2062 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2063 _all_full_gc_times_ms->num(), 2064 _all_full_gc_times_ms->sum() / 1000.0); 2065 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg()); 2066 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2067 _all_full_gc_times_ms->sd(), 2068 _all_full_gc_times_ms->maximum()); 2069 } 2070 } 2071 } 2072 2073 void G1CollectorPolicy::print_yg_surv_rate_info() const { 2074 #ifndef PRODUCT 2075 _short_lived_surv_rate_group->print_surv_rate_summary(); 2076 // add this call for any other surv rate groups 2077 #endif // PRODUCT 2078 } 2079 2080 #ifndef PRODUCT 2081 // for debugging, bit of a hack... 2082 static char* 2083 region_num_to_mbs(int length) { 2084 static char buffer[64]; 2085 double bytes = (double) (length * HeapRegion::GrainBytes); 2086 double mbs = bytes / (double) (1024 * 1024); 2087 sprintf(buffer, "%7.2lfMB", mbs); 2088 return buffer; 2089 } 2090 #endif // PRODUCT 2091 2092 size_t G1CollectorPolicy::max_regions(int purpose) { 2093 switch (purpose) { 2094 case GCAllocForSurvived: 2095 return _max_survivor_regions; 2096 case GCAllocForTenured: 2097 return REGIONS_UNLIMITED; 2098 default: 2099 ShouldNotReachHere(); 2100 return REGIONS_UNLIMITED; 2101 }; 2102 } 2103 2104 void G1CollectorPolicy::update_max_gc_locker_expansion() { 2105 size_t expansion_region_num = 0; 2106 if (GCLockerEdenExpansionPercent > 0) { 2107 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 2108 double expansion_region_num_d = perc * (double) _young_list_target_length; 2109 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 2110 // less than 1.0) we'll get 1. 2111 expansion_region_num = (size_t) ceil(expansion_region_num_d); 2112 } else { 2113 assert(expansion_region_num == 0, "sanity"); 2114 } 2115 _young_list_max_length = _young_list_target_length + expansion_region_num; 2116 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 2117 } 2118 2119 // Calculates survivor space parameters. 2120 void G1CollectorPolicy::update_survivors_policy() { 2121 double max_survivor_regions_d = 2122 (double) _young_list_target_length / (double) SurvivorRatio; 2123 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 2124 // smaller than 1.0) we'll get 1. 2125 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d); 2126 2127 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 2128 HeapRegion::GrainWords * _max_survivor_regions); 2129 } 2130 2131 #ifndef PRODUCT 2132 class HRSortIndexIsOKClosure: public HeapRegionClosure { 2133 CollectionSetChooser* _chooser; 2134 public: 2135 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) : 2136 _chooser(chooser) {} 2137 2138 bool doHeapRegion(HeapRegion* r) { 2139 if (!r->continuesHumongous()) { 2140 assert(_chooser->regionProperlyOrdered(r), "Ought to be."); 2141 } 2142 return false; 2143 } 2144 }; 2145 2146 bool G1CollectorPolicy::assertMarkedBytesDataOK() { 2147 HRSortIndexIsOKClosure cl(_collectionSetChooser); 2148 _g1->heap_region_iterate(&cl); 2149 return true; 2150 } 2151 #endif 2152 2153 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 2154 GCCause::Cause gc_cause) { 2155 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2156 if (!during_cycle) { 2157 ergo_verbose1(ErgoConcCycles, 2158 "request concurrent cycle initiation", 2159 ergo_format_reason("requested by GC cause") 2160 ergo_format_str("GC cause"), 2161 GCCause::to_string(gc_cause)); 2162 set_initiate_conc_mark_if_possible(); 2163 return true; 2164 } else { 2165 ergo_verbose1(ErgoConcCycles, 2166 "do not request concurrent cycle initiation", 2167 ergo_format_reason("concurrent cycle already in progress") 2168 ergo_format_str("GC cause"), 2169 GCCause::to_string(gc_cause)); 2170 return false; 2171 } 2172 } 2173 2174 void 2175 G1CollectorPolicy::decide_on_conc_mark_initiation() { 2176 // We are about to decide on whether this pause will be an 2177 // initial-mark pause. 2178 2179 // First, during_initial_mark_pause() should not be already set. We 2180 // will set it here if we have to. However, it should be cleared by 2181 // the end of the pause (it's only set for the duration of an 2182 // initial-mark pause). 2183 assert(!during_initial_mark_pause(), "pre-condition"); 2184 2185 if (initiate_conc_mark_if_possible()) { 2186 // We had noticed on a previous pause that the heap occupancy has 2187 // gone over the initiating threshold and we should start a 2188 // concurrent marking cycle. So we might initiate one. 2189 2190 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2191 if (!during_cycle) { 2192 // The concurrent marking thread is not "during a cycle", i.e., 2193 // it has completed the last one. So we can go ahead and 2194 // initiate a new cycle. 2195 2196 set_during_initial_mark_pause(); 2197 // We do not allow mixed GCs during marking. 2198 if (!gcs_are_young()) { 2199 set_gcs_are_young(true); 2200 ergo_verbose0(ErgoMixedGCs, 2201 "end mixed GCs", 2202 ergo_format_reason("concurrent cycle is about to start")); 2203 } 2204 2205 // And we can now clear initiate_conc_mark_if_possible() as 2206 // we've already acted on it. 2207 clear_initiate_conc_mark_if_possible(); 2208 2209 ergo_verbose0(ErgoConcCycles, 2210 "initiate concurrent cycle", 2211 ergo_format_reason("concurrent cycle initiation requested")); 2212 } else { 2213 // The concurrent marking thread is still finishing up the 2214 // previous cycle. If we start one right now the two cycles 2215 // overlap. In particular, the concurrent marking thread might 2216 // be in the process of clearing the next marking bitmap (which 2217 // we will use for the next cycle if we start one). Starting a 2218 // cycle now will be bad given that parts of the marking 2219 // information might get cleared by the marking thread. And we 2220 // cannot wait for the marking thread to finish the cycle as it 2221 // periodically yields while clearing the next marking bitmap 2222 // and, if it's in a yield point, it's waiting for us to 2223 // finish. So, at this point we will not start a cycle and we'll 2224 // let the concurrent marking thread complete the last one. 2225 ergo_verbose0(ErgoConcCycles, 2226 "do not initiate concurrent cycle", 2227 ergo_format_reason("concurrent cycle already in progress")); 2228 } 2229 } 2230 } 2231 2232 class KnownGarbageClosure: public HeapRegionClosure { 2233 CollectionSetChooser* _hrSorted; 2234 2235 public: 2236 KnownGarbageClosure(CollectionSetChooser* hrSorted) : 2237 _hrSorted(hrSorted) 2238 {} 2239 2240 bool doHeapRegion(HeapRegion* r) { 2241 // We only include humongous regions in collection 2242 // sets when concurrent mark shows that their contained object is 2243 // unreachable. 2244 2245 // Do we have any marking information for this region? 2246 if (r->is_marked()) { 2247 // We don't include humongous regions in collection 2248 // sets because we collect them immediately at the end of a marking 2249 // cycle. We also don't include young regions because we *must* 2250 // include them in the next collection pause. 2251 if (!r->isHumongous() && !r->is_young()) { 2252 _hrSorted->addMarkedHeapRegion(r); 2253 } 2254 } 2255 return false; 2256 } 2257 }; 2258 2259 class ParKnownGarbageHRClosure: public HeapRegionClosure { 2260 CollectionSetChooser* _hrSorted; 2261 jint _marked_regions_added; 2262 jint _chunk_size; 2263 jint _cur_chunk_idx; 2264 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) 2265 int _worker; 2266 int _invokes; 2267 2268 void get_new_chunk() { 2269 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size); 2270 _cur_chunk_end = _cur_chunk_idx + _chunk_size; 2271 } 2272 void add_region(HeapRegion* r) { 2273 if (_cur_chunk_idx == _cur_chunk_end) { 2274 get_new_chunk(); 2275 } 2276 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); 2277 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r); 2278 _marked_regions_added++; 2279 _cur_chunk_idx++; 2280 } 2281 2282 public: 2283 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 2284 jint chunk_size, 2285 int worker) : 2286 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker), 2287 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0), 2288 _invokes(0) 2289 {} 2290 2291 bool doHeapRegion(HeapRegion* r) { 2292 // We only include humongous regions in collection 2293 // sets when concurrent mark shows that their contained object is 2294 // unreachable. 2295 _invokes++; 2296 2297 // Do we have any marking information for this region? 2298 if (r->is_marked()) { 2299 // We don't include humongous regions in collection 2300 // sets because we collect them immediately at the end of a marking 2301 // cycle. 2302 // We also do not include young regions in collection sets 2303 if (!r->isHumongous() && !r->is_young()) { 2304 add_region(r); 2305 } 2306 } 2307 return false; 2308 } 2309 jint marked_regions_added() { return _marked_regions_added; } 2310 int invokes() { return _invokes; } 2311 }; 2312 2313 class ParKnownGarbageTask: public AbstractGangTask { 2314 CollectionSetChooser* _hrSorted; 2315 jint _chunk_size; 2316 G1CollectedHeap* _g1; 2317 public: 2318 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) : 2319 AbstractGangTask("ParKnownGarbageTask"), 2320 _hrSorted(hrSorted), _chunk_size(chunk_size), 2321 _g1(G1CollectedHeap::heap()) 2322 {} 2323 2324 void work(int i) { 2325 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i); 2326 // Back to zero for the claim value. 2327 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i, 2328 _g1->workers()->active_workers(), 2329 HeapRegion::InitialClaimValue); 2330 jint regions_added = parKnownGarbageCl.marked_regions_added(); 2331 _hrSorted->incNumMarkedHeapRegions(regions_added); 2332 if (G1PrintParCleanupStats) { 2333 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.", 2334 i, parKnownGarbageCl.invokes(), regions_added); 2335 } 2336 } 2337 }; 2338 2339 void 2340 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) { 2341 double start_sec; 2342 if (G1PrintParCleanupStats) { 2343 start_sec = os::elapsedTime(); 2344 } 2345 2346 _collectionSetChooser->clearMarkedHeapRegions(); 2347 double clear_marked_end_sec; 2348 if (G1PrintParCleanupStats) { 2349 clear_marked_end_sec = os::elapsedTime(); 2350 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.", 2351 (clear_marked_end_sec - start_sec) * 1000.0); 2352 } 2353 2354 if (G1CollectedHeap::use_parallel_gc_threads()) { 2355 const size_t OverpartitionFactor = 4; 2356 size_t WorkUnit; 2357 // The use of MinChunkSize = 8 in the original code 2358 // causes some assertion failures when the total number of 2359 // region is less than 8. The code here tries to fix that. 2360 // Should the original code also be fixed? 2361 if (no_of_gc_threads > 0) { 2362 const size_t MinWorkUnit = 2363 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U); 2364 WorkUnit = 2365 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor), 2366 MinWorkUnit); 2367 } else { 2368 assert(no_of_gc_threads > 0, 2369 "The active gc workers should be greater than 0"); 2370 // In a product build do something reasonable to avoid a crash. 2371 const size_t MinWorkUnit = 2372 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U); 2373 WorkUnit = 2374 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor), 2375 MinWorkUnit); 2376 } 2377 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(), 2378 WorkUnit); 2379 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, 2380 (int) WorkUnit); 2381 _g1->workers()->run_task(&parKnownGarbageTask); 2382 2383 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 2384 "sanity check"); 2385 } else { 2386 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); 2387 _g1->heap_region_iterate(&knownGarbagecl); 2388 } 2389 double known_garbage_end_sec; 2390 if (G1PrintParCleanupStats) { 2391 known_garbage_end_sec = os::elapsedTime(); 2392 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.", 2393 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0); 2394 } 2395 2396 _collectionSetChooser->sortMarkedHeapRegions(); 2397 double end_sec = os::elapsedTime(); 2398 if (G1PrintParCleanupStats) { 2399 gclog_or_tty->print_cr(" sorting: %8.3f ms.", 2400 (end_sec - known_garbage_end_sec) * 1000.0); 2401 } 2402 2403 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 2404 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 2405 _cur_mark_stop_world_time_ms += elapsed_time_ms; 2406 _prev_collection_pause_end_ms += elapsed_time_ms; 2407 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true); 2408 } 2409 2410 // Add the heap region at the head of the non-incremental collection set 2411 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 2412 assert(_inc_cset_build_state == Active, "Precondition"); 2413 assert(!hr->is_young(), "non-incremental add of young region"); 2414 2415 if (_g1->mark_in_progress()) 2416 _g1->concurrent_mark()->registerCSetRegion(hr); 2417 2418 assert(!hr->in_collection_set(), "should not already be in the CSet"); 2419 hr->set_in_collection_set(true); 2420 hr->set_next_in_collection_set(_collection_set); 2421 _collection_set = hr; 2422 _collection_set_bytes_used_before += hr->used(); 2423 _g1->register_region_with_in_cset_fast_test(hr); 2424 size_t rs_length = hr->rem_set()->occupied(); 2425 _recorded_rs_lengths += rs_length; 2426 _old_cset_region_length += 1; 2427 } 2428 2429 // Initialize the per-collection-set information 2430 void G1CollectorPolicy::start_incremental_cset_building() { 2431 assert(_inc_cset_build_state == Inactive, "Precondition"); 2432 2433 _inc_cset_head = NULL; 2434 _inc_cset_tail = NULL; 2435 _inc_cset_bytes_used_before = 0; 2436 2437 _inc_cset_max_finger = 0; 2438 _inc_cset_recorded_rs_lengths = 0; 2439 _inc_cset_predicted_elapsed_time_ms = 0; 2440 _inc_cset_build_state = Active; 2441 } 2442 2443 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 2444 // This routine is used when: 2445 // * adding survivor regions to the incremental cset at the end of an 2446 // evacuation pause, 2447 // * adding the current allocation region to the incremental cset 2448 // when it is retired, and 2449 // * updating existing policy information for a region in the 2450 // incremental cset via young list RSet sampling. 2451 // Therefore this routine may be called at a safepoint by the 2452 // VM thread, or in-between safepoints by mutator threads (when 2453 // retiring the current allocation region) or a concurrent 2454 // refine thread (RSet sampling). 2455 2456 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2457 size_t used_bytes = hr->used(); 2458 2459 _inc_cset_recorded_rs_lengths += rs_length; 2460 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 2461 2462 _inc_cset_bytes_used_before += used_bytes; 2463 2464 // Cache the values we have added to the aggregated informtion 2465 // in the heap region in case we have to remove this region from 2466 // the incremental collection set, or it is updated by the 2467 // rset sampling code 2468 hr->set_recorded_rs_length(rs_length); 2469 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 2470 } 2471 2472 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) { 2473 // This routine is currently only called as part of the updating of 2474 // existing policy information for regions in the incremental cset that 2475 // is performed by the concurrent refine thread(s) as part of young list 2476 // RSet sampling. Therefore we should not be at a safepoint. 2477 2478 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint"); 2479 assert(hr->is_young(), "it should be"); 2480 2481 size_t used_bytes = hr->used(); 2482 size_t old_rs_length = hr->recorded_rs_length(); 2483 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 2484 2485 // Subtract the old recorded/predicted policy information for 2486 // the given heap region from the collection set info. 2487 _inc_cset_recorded_rs_lengths -= old_rs_length; 2488 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms; 2489 2490 _inc_cset_bytes_used_before -= used_bytes; 2491 2492 // Clear the values cached in the heap region 2493 hr->set_recorded_rs_length(0); 2494 hr->set_predicted_elapsed_time_ms(0); 2495 } 2496 2497 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) { 2498 // Update the collection set information that is dependent on the new RS length 2499 assert(hr->is_young(), "Precondition"); 2500 2501 remove_from_incremental_cset_info(hr); 2502 add_to_incremental_cset_info(hr, new_rs_length); 2503 } 2504 2505 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 2506 assert(hr->is_young(), "invariant"); 2507 assert(hr->young_index_in_cset() > -1, "should have already been set"); 2508 assert(_inc_cset_build_state == Active, "Precondition"); 2509 2510 // We need to clear and set the cached recorded/cached collection set 2511 // information in the heap region here (before the region gets added 2512 // to the collection set). An individual heap region's cached values 2513 // are calculated, aggregated with the policy collection set info, 2514 // and cached in the heap region here (initially) and (subsequently) 2515 // by the Young List sampling code. 2516 2517 size_t rs_length = hr->rem_set()->occupied(); 2518 add_to_incremental_cset_info(hr, rs_length); 2519 2520 HeapWord* hr_end = hr->end(); 2521 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 2522 2523 assert(!hr->in_collection_set(), "invariant"); 2524 hr->set_in_collection_set(true); 2525 assert( hr->next_in_collection_set() == NULL, "invariant"); 2526 2527 _g1->register_region_with_in_cset_fast_test(hr); 2528 } 2529 2530 // Add the region at the RHS of the incremental cset 2531 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 2532 // We should only ever be appending survivors at the end of a pause 2533 assert( hr->is_survivor(), "Logic"); 2534 2535 // Do the 'common' stuff 2536 add_region_to_incremental_cset_common(hr); 2537 2538 // Now add the region at the right hand side 2539 if (_inc_cset_tail == NULL) { 2540 assert(_inc_cset_head == NULL, "invariant"); 2541 _inc_cset_head = hr; 2542 } else { 2543 _inc_cset_tail->set_next_in_collection_set(hr); 2544 } 2545 _inc_cset_tail = hr; 2546 } 2547 2548 // Add the region to the LHS of the incremental cset 2549 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2550 // Survivors should be added to the RHS at the end of a pause 2551 assert(!hr->is_survivor(), "Logic"); 2552 2553 // Do the 'common' stuff 2554 add_region_to_incremental_cset_common(hr); 2555 2556 // Add the region at the left hand side 2557 hr->set_next_in_collection_set(_inc_cset_head); 2558 if (_inc_cset_head == NULL) { 2559 assert(_inc_cset_tail == NULL, "Invariant"); 2560 _inc_cset_tail = hr; 2561 } 2562 _inc_cset_head = hr; 2563 } 2564 2565 #ifndef PRODUCT 2566 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2567 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2568 2569 st->print_cr("\nCollection_set:"); 2570 HeapRegion* csr = list_head; 2571 while (csr != NULL) { 2572 HeapRegion* next = csr->next_in_collection_set(); 2573 assert(csr->in_collection_set(), "bad CS"); 2574 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, " 2575 "age: %4d, y: %d, surv: %d", 2576 csr->bottom(), csr->end(), 2577 csr->top(), 2578 csr->prev_top_at_mark_start(), 2579 csr->next_top_at_mark_start(), 2580 csr->top_at_conc_mark_count(), 2581 csr->age_in_surv_rate_group_cond(), 2582 csr->is_young(), 2583 csr->is_survivor()); 2584 csr = next; 2585 } 2586 } 2587 #endif // !PRODUCT 2588 2589 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) { 2590 // Set this here - in case we're not doing young collections. 2591 double non_young_start_time_sec = os::elapsedTime(); 2592 2593 YoungList* young_list = _g1->young_list(); 2594 2595 guarantee(target_pause_time_ms > 0.0, 2596 err_msg("target_pause_time_ms = %1.6lf should be positive", 2597 target_pause_time_ms)); 2598 guarantee(_collection_set == NULL, "Precondition"); 2599 2600 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2601 double predicted_pause_time_ms = base_time_ms; 2602 2603 double time_remaining_ms = target_pause_time_ms - base_time_ms; 2604 2605 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2606 "start choosing CSet", 2607 ergo_format_ms("predicted base time") 2608 ergo_format_ms("remaining time") 2609 ergo_format_ms("target pause time"), 2610 base_time_ms, time_remaining_ms, target_pause_time_ms); 2611 2612 // the 10% and 50% values are arbitrary... 2613 double threshold = 0.10 * target_pause_time_ms; 2614 if (time_remaining_ms < threshold) { 2615 double prev_time_remaining_ms = time_remaining_ms; 2616 time_remaining_ms = 0.50 * target_pause_time_ms; 2617 ergo_verbose3(ErgoCSetConstruction, 2618 "adjust remaining time", 2619 ergo_format_reason("remaining time lower than threshold") 2620 ergo_format_ms("remaining time") 2621 ergo_format_ms("threshold") 2622 ergo_format_ms("adjusted remaining time"), 2623 prev_time_remaining_ms, threshold, time_remaining_ms); 2624 } 2625 2626 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes; 2627 2628 HeapRegion* hr; 2629 double young_start_time_sec = os::elapsedTime(); 2630 2631 _collection_set_bytes_used_before = 0; 2632 _last_gc_was_young = gcs_are_young() ? true : false; 2633 2634 if (_last_gc_was_young) { 2635 ++_young_pause_num; 2636 } else { 2637 ++_mixed_pause_num; 2638 } 2639 2640 // The young list is laid with the survivor regions from the previous 2641 // pause are appended to the RHS of the young list, i.e. 2642 // [Newly Young Regions ++ Survivors from last pause]. 2643 2644 size_t survivor_region_length = young_list->survivor_length(); 2645 size_t eden_region_length = young_list->length() - survivor_region_length; 2646 init_cset_region_lengths(eden_region_length, survivor_region_length); 2647 hr = young_list->first_survivor_region(); 2648 while (hr != NULL) { 2649 assert(hr->is_survivor(), "badly formed young list"); 2650 hr->set_young(); 2651 hr = hr->get_next_young_region(); 2652 } 2653 2654 // Clear the fields that point to the survivor list - they are all young now. 2655 young_list->clear_survivors(); 2656 2657 if (_g1->mark_in_progress()) 2658 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger); 2659 2660 _collection_set = _inc_cset_head; 2661 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2662 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms; 2663 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; 2664 2665 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2666 "add young regions to CSet", 2667 ergo_format_region("eden") 2668 ergo_format_region("survivors") 2669 ergo_format_ms("predicted young region time"), 2670 eden_region_length, survivor_region_length, 2671 _inc_cset_predicted_elapsed_time_ms); 2672 2673 // The number of recorded young regions is the incremental 2674 // collection set's current size 2675 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2676 2677 double young_end_time_sec = os::elapsedTime(); 2678 _recorded_young_cset_choice_time_ms = 2679 (young_end_time_sec - young_start_time_sec) * 1000.0; 2680 2681 // We are doing young collections so reset this. 2682 non_young_start_time_sec = young_end_time_sec; 2683 2684 if (!gcs_are_young()) { 2685 bool should_continue = true; 2686 NumberSeq seq; 2687 double avg_prediction = 100000000000000000.0; // something very large 2688 2689 double prev_predicted_pause_time_ms = predicted_pause_time_ms; 2690 do { 2691 // Note that add_old_region_to_cset() increments the 2692 // _old_cset_region_length field and cset_region_length() returns the 2693 // sum of _eden_cset_region_length, _survivor_cset_region_length, and 2694 // _old_cset_region_length. So, as old regions are added to the 2695 // CSet, _old_cset_region_length will be incremented and 2696 // cset_region_length(), which is used below, will always reflect 2697 // the the total number of regions added up to this point to the CSet. 2698 2699 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms, 2700 avg_prediction); 2701 if (hr != NULL) { 2702 _g1->old_set_remove(hr); 2703 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); 2704 time_remaining_ms -= predicted_time_ms; 2705 predicted_pause_time_ms += predicted_time_ms; 2706 add_old_region_to_cset(hr); 2707 seq.add(predicted_time_ms); 2708 avg_prediction = seq.avg() + seq.sd(); 2709 } 2710 2711 should_continue = true; 2712 if (hr == NULL) { 2713 // No need for an ergo verbose message here, 2714 // getNextMarkRegion() does this when it returns NULL. 2715 should_continue = false; 2716 } else { 2717 if (adaptive_young_list_length()) { 2718 if (time_remaining_ms < 0.0) { 2719 ergo_verbose1(ErgoCSetConstruction, 2720 "stop adding old regions to CSet", 2721 ergo_format_reason("remaining time is lower than 0") 2722 ergo_format_ms("remaining time"), 2723 time_remaining_ms); 2724 should_continue = false; 2725 } 2726 } else { 2727 if (cset_region_length() >= _young_list_fixed_length) { 2728 ergo_verbose2(ErgoCSetConstruction, 2729 "stop adding old regions to CSet", 2730 ergo_format_reason("CSet length reached target") 2731 ergo_format_region("CSet") 2732 ergo_format_region("young target"), 2733 cset_region_length(), _young_list_fixed_length); 2734 should_continue = false; 2735 } 2736 } 2737 } 2738 } while (should_continue); 2739 2740 if (!adaptive_young_list_length() && 2741 cset_region_length() < _young_list_fixed_length) { 2742 ergo_verbose2(ErgoCSetConstruction, 2743 "request mixed GCs end", 2744 ergo_format_reason("CSet length lower than target") 2745 ergo_format_region("CSet") 2746 ergo_format_region("young target"), 2747 cset_region_length(), _young_list_fixed_length); 2748 _should_revert_to_young_gcs = true; 2749 } 2750 2751 ergo_verbose2(ErgoCSetConstruction | ErgoHigh, 2752 "add old regions to CSet", 2753 ergo_format_region("old") 2754 ergo_format_ms("predicted old region time"), 2755 old_cset_region_length(), 2756 predicted_pause_time_ms - prev_predicted_pause_time_ms); 2757 } 2758 2759 stop_incremental_cset_building(); 2760 2761 count_CS_bytes_used(); 2762 2763 ergo_verbose5(ErgoCSetConstruction, 2764 "finish choosing CSet", 2765 ergo_format_region("eden") 2766 ergo_format_region("survivors") 2767 ergo_format_region("old") 2768 ergo_format_ms("predicted pause time") 2769 ergo_format_ms("target pause time"), 2770 eden_region_length, survivor_region_length, 2771 old_cset_region_length(), 2772 predicted_pause_time_ms, target_pause_time_ms); 2773 2774 double non_young_end_time_sec = os::elapsedTime(); 2775 _recorded_non_young_cset_choice_time_ms = 2776 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; 2777 }