1 /* 2 * Copyright (c) 2001, 2012, 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/g1Log.hpp" 33 #include "gc_implementation/g1/heapRegionRemSet.hpp" 34 #include "gc_implementation/shared/gcPolicyCounters.hpp" 35 #include "runtime/arguments.hpp" 36 #include "runtime/java.hpp" 37 #include "runtime/mutexLocker.hpp" 38 #include "utilities/debug.hpp" 39 40 // Different defaults for different number of GC threads 41 // They were chosen by running GCOld and SPECjbb on debris with different 42 // numbers of GC threads and choosing them based on the results 43 44 // all the same 45 static double rs_length_diff_defaults[] = { 46 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 47 }; 48 49 static double cost_per_card_ms_defaults[] = { 50 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 51 }; 52 53 // all the same 54 static double young_cards_per_entry_ratio_defaults[] = { 55 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 56 }; 57 58 static double cost_per_entry_ms_defaults[] = { 59 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 60 }; 61 62 static double cost_per_byte_ms_defaults[] = { 63 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 64 }; 65 66 // these should be pretty consistent 67 static double constant_other_time_ms_defaults[] = { 68 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 69 }; 70 71 72 static double young_other_cost_per_region_ms_defaults[] = { 73 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 74 }; 75 76 static double non_young_other_cost_per_region_ms_defaults[] = { 77 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 78 }; 79 80 // Help class for avoiding interleaved logging 81 class LineBuffer: public StackObj { 82 83 private: 84 static const int BUFFER_LEN = 1024; 85 static const int INDENT_CHARS = 3; 86 char _buffer[BUFFER_LEN]; 87 int _indent_level; 88 int _cur; 89 90 void vappend(const char* format, va_list ap) { 91 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap); 92 if (res != -1) { 93 _cur += res; 94 } else { 95 DEBUG_ONLY(warning("buffer too small in LineBuffer");) 96 _buffer[BUFFER_LEN -1] = 0; 97 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again 98 } 99 } 100 101 public: 102 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) { 103 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) { 104 _buffer[_cur] = ' '; 105 } 106 } 107 108 #ifndef PRODUCT 109 ~LineBuffer() { 110 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?"); 111 } 112 #endif 113 114 void append(const char* format, ...) { 115 va_list ap; 116 va_start(ap, format); 117 vappend(format, ap); 118 va_end(ap); 119 } 120 121 void append_and_print_cr(const char* format, ...) { 122 va_list ap; 123 va_start(ap, format); 124 vappend(format, ap); 125 va_end(ap); 126 gclog_or_tty->print_cr("%s", _buffer); 127 _cur = _indent_level * INDENT_CHARS; 128 } 129 }; 130 131 G1CollectorPolicy::G1CollectorPolicy() : 132 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads() 133 ? ParallelGCThreads : 1), 134 135 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 136 _all_pause_times_ms(new NumberSeq()), 137 _stop_world_start(0.0), 138 _all_stop_world_times_ms(new NumberSeq()), 139 _all_yield_times_ms(new NumberSeq()), 140 141 _summary(new Summary()), 142 143 _cur_clear_ct_time_ms(0.0), 144 _root_region_scan_wait_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 _recent_prev_end_times_for_all_gcs_sec( 196 new TruncatedSeq(NumPrevPausesForHeuristics)), 197 198 _recent_avg_pause_time_ratio(0.0), 199 200 _all_full_gc_times_ms(new NumberSeq()), 201 202 _initiate_conc_mark_if_possible(false), 203 _during_initial_mark_pause(false), 204 _last_young_gc(false), 205 _last_gc_was_young(false), 206 207 _eden_bytes_before_gc(0), 208 _survivor_bytes_before_gc(0), 209 _capacity_before_gc(0), 210 211 _eden_cset_region_length(0), 212 _survivor_cset_region_length(0), 213 _old_cset_region_length(0), 214 215 _collection_set(NULL), 216 _collection_set_bytes_used_before(0), 217 218 // Incremental CSet attributes 219 _inc_cset_build_state(Inactive), 220 _inc_cset_head(NULL), 221 _inc_cset_tail(NULL), 222 _inc_cset_bytes_used_before(0), 223 _inc_cset_max_finger(NULL), 224 _inc_cset_recorded_rs_lengths(0), 225 _inc_cset_recorded_rs_lengths_diffs(0), 226 _inc_cset_predicted_elapsed_time_ms(0.0), 227 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 228 229 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 230 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 231 #endif // _MSC_VER 232 233 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", 234 G1YoungSurvRateNumRegionsSummary)), 235 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", 236 G1YoungSurvRateNumRegionsSummary)), 237 // add here any more surv rate groups 238 _recorded_survivor_regions(0), 239 _recorded_survivor_head(NULL), 240 _recorded_survivor_tail(NULL), 241 _survivors_age_table(true), 242 243 _gc_overhead_perc(0.0) { 244 245 // Set up the region size and associated fields. Given that the 246 // policy is created before the heap, we have to set this up here, 247 // so it's done as soon as possible. 248 HeapRegion::setup_heap_region_size(Arguments::min_heap_size()); 249 HeapRegionRemSet::setup_remset_size(); 250 251 G1ErgoVerbose::initialize(); 252 if (PrintAdaptiveSizePolicy) { 253 // Currently, we only use a single switch for all the heuristics. 254 G1ErgoVerbose::set_enabled(true); 255 // Given that we don't currently have a verboseness level 256 // parameter, we'll hardcode this to high. This can be easily 257 // changed in the future. 258 G1ErgoVerbose::set_level(ErgoHigh); 259 } else { 260 G1ErgoVerbose::set_enabled(false); 261 } 262 263 // Verify PLAB sizes 264 const size_t region_size = HeapRegion::GrainWords; 265 if (YoungPLABSize > region_size || OldPLABSize > region_size) { 266 char buffer[128]; 267 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT, 268 OldPLABSize > region_size ? "Old" : "Young", region_size); 269 vm_exit_during_initialization(buffer); 270 } 271 272 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 273 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 274 275 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads]; 276 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads]; 277 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads]; 278 279 _par_last_update_rs_times_ms = new double[_parallel_gc_threads]; 280 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; 281 282 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; 283 284 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads]; 285 286 _par_last_termination_times_ms = new double[_parallel_gc_threads]; 287 _par_last_termination_attempts = new double[_parallel_gc_threads]; 288 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads]; 289 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads]; 290 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads]; 291 292 int index; 293 if (ParallelGCThreads == 0) 294 index = 0; 295 else if (ParallelGCThreads > 8) 296 index = 7; 297 else 298 index = ParallelGCThreads - 1; 299 300 _pending_card_diff_seq->add(0.0); 301 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 302 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 303 _young_cards_per_entry_ratio_seq->add( 304 young_cards_per_entry_ratio_defaults[index]); 305 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 306 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 307 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 308 _young_other_cost_per_region_ms_seq->add( 309 young_other_cost_per_region_ms_defaults[index]); 310 _non_young_other_cost_per_region_ms_seq->add( 311 non_young_other_cost_per_region_ms_defaults[index]); 312 313 // Below, we might need to calculate the pause time target based on 314 // the pause interval. When we do so we are going to give G1 maximum 315 // flexibility and allow it to do pauses when it needs to. So, we'll 316 // arrange that the pause interval to be pause time target + 1 to 317 // ensure that a) the pause time target is maximized with respect to 318 // the pause interval and b) we maintain the invariant that pause 319 // time target < pause interval. If the user does not want this 320 // maximum flexibility, they will have to set the pause interval 321 // explicitly. 322 323 // First make sure that, if either parameter is set, its value is 324 // reasonable. 325 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 326 if (MaxGCPauseMillis < 1) { 327 vm_exit_during_initialization("MaxGCPauseMillis should be " 328 "greater than 0"); 329 } 330 } 331 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 332 if (GCPauseIntervalMillis < 1) { 333 vm_exit_during_initialization("GCPauseIntervalMillis should be " 334 "greater than 0"); 335 } 336 } 337 338 // Then, if the pause time target parameter was not set, set it to 339 // the default value. 340 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 341 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 342 // The default pause time target in G1 is 200ms 343 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 344 } else { 345 // We do not allow the pause interval to be set without the 346 // pause time target 347 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 348 "without setting MaxGCPauseMillis"); 349 } 350 } 351 352 // Then, if the interval parameter was not set, set it according to 353 // the pause time target (this will also deal with the case when the 354 // pause time target is the default value). 355 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 356 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 357 } 358 359 // Finally, make sure that the two parameters are consistent. 360 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 361 char buffer[256]; 362 jio_snprintf(buffer, 256, 363 "MaxGCPauseMillis (%u) should be less than " 364 "GCPauseIntervalMillis (%u)", 365 MaxGCPauseMillis, GCPauseIntervalMillis); 366 vm_exit_during_initialization(buffer); 367 } 368 369 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 370 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 371 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 372 _sigma = (double) G1ConfidencePercent / 100.0; 373 374 // start conservatively (around 50ms is about right) 375 _concurrent_mark_remark_times_ms->add(0.05); 376 _concurrent_mark_cleanup_times_ms->add(0.20); 377 _tenuring_threshold = MaxTenuringThreshold; 378 // _max_survivor_regions will be calculated by 379 // update_young_list_target_length() during initialization. 380 _max_survivor_regions = 0; 381 382 assert(GCTimeRatio > 0, 383 "we should have set it to a default value set_g1_gc_flags() " 384 "if a user set it to 0"); 385 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 386 387 uintx reserve_perc = G1ReservePercent; 388 // Put an artificial ceiling on this so that it's not set to a silly value. 389 if (reserve_perc > 50) { 390 reserve_perc = 50; 391 warning("G1ReservePercent is set to a value that is too large, " 392 "it's been updated to %u", reserve_perc); 393 } 394 _reserve_factor = (double) reserve_perc / 100.0; 395 // This will be set when the heap is expanded 396 // for the first time during initialization. 397 _reserve_regions = 0; 398 399 initialize_all(); 400 _collectionSetChooser = new CollectionSetChooser(); 401 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 402 } 403 404 void G1CollectorPolicy::initialize_flags() { 405 set_min_alignment(HeapRegion::GrainBytes); 406 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name())); 407 if (SurvivorRatio < 1) { 408 vm_exit_during_initialization("Invalid survivor ratio specified"); 409 } 410 CollectorPolicy::initialize_flags(); 411 } 412 413 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) { 414 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max"); 415 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds"); 416 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds"); 417 418 if (FLAG_IS_CMDLINE(NewRatio)) { 419 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 420 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 421 } else { 422 _sizer_kind = SizerNewRatio; 423 _adaptive_size = false; 424 return; 425 } 426 } 427 428 if (FLAG_IS_CMDLINE(NewSize)) { 429 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 430 1U); 431 if (FLAG_IS_CMDLINE(MaxNewSize)) { 432 _max_desired_young_length = 433 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 434 1U); 435 _sizer_kind = SizerMaxAndNewSize; 436 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 437 } else { 438 _sizer_kind = SizerNewSizeOnly; 439 } 440 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 441 _max_desired_young_length = 442 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 443 1U); 444 _sizer_kind = SizerMaxNewSizeOnly; 445 } 446 } 447 448 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 449 uint default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100; 450 return MAX2(1U, default_value); 451 } 452 453 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 454 uint default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100; 455 return MAX2(1U, default_value); 456 } 457 458 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 459 assert(new_number_of_heap_regions > 0, "Heap must be initialized"); 460 461 switch (_sizer_kind) { 462 case SizerDefaults: 463 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions); 464 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions); 465 break; 466 case SizerNewSizeOnly: 467 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions); 468 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length); 469 break; 470 case SizerMaxNewSizeOnly: 471 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions); 472 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length); 473 break; 474 case SizerMaxAndNewSize: 475 // Do nothing. Values set on the command line, don't update them at runtime. 476 break; 477 case SizerNewRatio: 478 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1); 479 _max_desired_young_length = _min_desired_young_length; 480 break; 481 default: 482 ShouldNotReachHere(); 483 } 484 485 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values"); 486 } 487 488 void G1CollectorPolicy::init() { 489 // Set aside an initial future to_space. 490 _g1 = G1CollectedHeap::heap(); 491 492 assert(Heap_lock->owned_by_self(), "Locking discipline."); 493 494 initialize_gc_policy_counters(); 495 496 if (adaptive_young_list_length()) { 497 _young_list_fixed_length = 0; 498 } else { 499 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 500 } 501 _free_regions_at_end_of_collection = _g1->free_regions(); 502 update_young_list_target_length(); 503 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes; 504 505 // We may immediately start allocating regions and placing them on the 506 // collection set list. Initialize the per-collection set info 507 start_incremental_cset_building(); 508 } 509 510 // Create the jstat counters for the policy. 511 void G1CollectorPolicy::initialize_gc_policy_counters() { 512 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 513 } 514 515 bool G1CollectorPolicy::predict_will_fit(uint young_length, 516 double base_time_ms, 517 uint base_free_regions, 518 double target_pause_time_ms) { 519 if (young_length >= base_free_regions) { 520 // end condition 1: not enough space for the young regions 521 return false; 522 } 523 524 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 525 size_t bytes_to_copy = 526 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 527 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 528 double young_other_time_ms = predict_young_other_time_ms(young_length); 529 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 530 if (pause_time_ms > target_pause_time_ms) { 531 // end condition 2: prediction is over the target pause time 532 return false; 533 } 534 535 size_t free_bytes = 536 (base_free_regions - young_length) * HeapRegion::GrainBytes; 537 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) { 538 // end condition 3: out-of-space (conservatively!) 539 return false; 540 } 541 542 // success! 543 return true; 544 } 545 546 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 547 // re-calculate the necessary reserve 548 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 549 // We use ceiling so that if reserve_regions_d is > 0.0 (but 550 // smaller than 1.0) we'll get 1. 551 _reserve_regions = (uint) ceil(reserve_regions_d); 552 553 _young_gen_sizer->heap_size_changed(new_number_of_regions); 554 } 555 556 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 557 uint base_min_length) { 558 uint desired_min_length = 0; 559 if (adaptive_young_list_length()) { 560 if (_alloc_rate_ms_seq->num() > 3) { 561 double now_sec = os::elapsedTime(); 562 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 563 double alloc_rate_ms = predict_alloc_rate_ms(); 564 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 565 } else { 566 // otherwise we don't have enough info to make the prediction 567 } 568 } 569 desired_min_length += base_min_length; 570 // make sure we don't go below any user-defined minimum bound 571 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 572 } 573 574 uint G1CollectorPolicy::calculate_young_list_desired_max_length() { 575 // Here, we might want to also take into account any additional 576 // constraints (i.e., user-defined minimum bound). Currently, we 577 // effectively don't set this bound. 578 return _young_gen_sizer->max_desired_young_length(); 579 } 580 581 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 582 if (rs_lengths == (size_t) -1) { 583 // if it's set to the default value (-1), we should predict it; 584 // otherwise, use the given value. 585 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); 586 } 587 588 // Calculate the absolute and desired min bounds. 589 590 // This is how many young regions we already have (currently: the survivors). 591 uint base_min_length = recorded_survivor_regions(); 592 // This is the absolute minimum young length, which ensures that we 593 // can allocate one eden region in the worst-case. 594 uint absolute_min_length = base_min_length + 1; 595 uint desired_min_length = 596 calculate_young_list_desired_min_length(base_min_length); 597 if (desired_min_length < absolute_min_length) { 598 desired_min_length = absolute_min_length; 599 } 600 601 // Calculate the absolute and desired max bounds. 602 603 // We will try our best not to "eat" into the reserve. 604 uint absolute_max_length = 0; 605 if (_free_regions_at_end_of_collection > _reserve_regions) { 606 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 607 } 608 uint desired_max_length = calculate_young_list_desired_max_length(); 609 if (desired_max_length > absolute_max_length) { 610 desired_max_length = absolute_max_length; 611 } 612 613 uint young_list_target_length = 0; 614 if (adaptive_young_list_length()) { 615 if (gcs_are_young()) { 616 young_list_target_length = 617 calculate_young_list_target_length(rs_lengths, 618 base_min_length, 619 desired_min_length, 620 desired_max_length); 621 _rs_lengths_prediction = rs_lengths; 622 } else { 623 // Don't calculate anything and let the code below bound it to 624 // the desired_min_length, i.e., do the next GC as soon as 625 // possible to maximize how many old regions we can add to it. 626 } 627 } else { 628 // The user asked for a fixed young gen so we'll fix the young gen 629 // whether the next GC is young or mixed. 630 young_list_target_length = _young_list_fixed_length; 631 } 632 633 // Make sure we don't go over the desired max length, nor under the 634 // desired min length. In case they clash, desired_min_length wins 635 // which is why that test is second. 636 if (young_list_target_length > desired_max_length) { 637 young_list_target_length = desired_max_length; 638 } 639 if (young_list_target_length < desired_min_length) { 640 young_list_target_length = desired_min_length; 641 } 642 643 assert(young_list_target_length > recorded_survivor_regions(), 644 "we should be able to allocate at least one eden region"); 645 assert(young_list_target_length >= absolute_min_length, "post-condition"); 646 _young_list_target_length = young_list_target_length; 647 648 update_max_gc_locker_expansion(); 649 } 650 651 uint 652 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 653 uint base_min_length, 654 uint desired_min_length, 655 uint desired_max_length) { 656 assert(adaptive_young_list_length(), "pre-condition"); 657 assert(gcs_are_young(), "only call this for young GCs"); 658 659 // In case some edge-condition makes the desired max length too small... 660 if (desired_max_length <= desired_min_length) { 661 return desired_min_length; 662 } 663 664 // We'll adjust min_young_length and max_young_length not to include 665 // the already allocated young regions (i.e., so they reflect the 666 // min and max eden regions we'll allocate). The base_min_length 667 // will be reflected in the predictions by the 668 // survivor_regions_evac_time prediction. 669 assert(desired_min_length > base_min_length, "invariant"); 670 uint min_young_length = desired_min_length - base_min_length; 671 assert(desired_max_length > base_min_length, "invariant"); 672 uint max_young_length = desired_max_length - base_min_length; 673 674 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 675 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 676 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 677 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 678 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 679 double base_time_ms = 680 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 681 survivor_regions_evac_time; 682 uint available_free_regions = _free_regions_at_end_of_collection; 683 uint base_free_regions = 0; 684 if (available_free_regions > _reserve_regions) { 685 base_free_regions = available_free_regions - _reserve_regions; 686 } 687 688 // Here, we will make sure that the shortest young length that 689 // makes sense fits within the target pause time. 690 691 if (predict_will_fit(min_young_length, base_time_ms, 692 base_free_regions, target_pause_time_ms)) { 693 // The shortest young length will fit into the target pause time; 694 // we'll now check whether the absolute maximum number of young 695 // regions will fit in the target pause time. If not, we'll do 696 // a binary search between min_young_length and max_young_length. 697 if (predict_will_fit(max_young_length, base_time_ms, 698 base_free_regions, target_pause_time_ms)) { 699 // The maximum young length will fit into the target pause time. 700 // We are done so set min young length to the maximum length (as 701 // the result is assumed to be returned in min_young_length). 702 min_young_length = max_young_length; 703 } else { 704 // The maximum possible number of young regions will not fit within 705 // the target pause time so we'll search for the optimal 706 // length. The loop invariants are: 707 // 708 // min_young_length < max_young_length 709 // min_young_length is known to fit into the target pause time 710 // max_young_length is known not to fit into the target pause time 711 // 712 // Going into the loop we know the above hold as we've just 713 // checked them. Every time around the loop we check whether 714 // the middle value between min_young_length and 715 // max_young_length fits into the target pause time. If it 716 // does, it becomes the new min. If it doesn't, it becomes 717 // the new max. This way we maintain the loop invariants. 718 719 assert(min_young_length < max_young_length, "invariant"); 720 uint diff = (max_young_length - min_young_length) / 2; 721 while (diff > 0) { 722 uint young_length = min_young_length + diff; 723 if (predict_will_fit(young_length, base_time_ms, 724 base_free_regions, target_pause_time_ms)) { 725 min_young_length = young_length; 726 } else { 727 max_young_length = young_length; 728 } 729 assert(min_young_length < max_young_length, "invariant"); 730 diff = (max_young_length - min_young_length) / 2; 731 } 732 // The results is min_young_length which, according to the 733 // loop invariants, should fit within the target pause time. 734 735 // These are the post-conditions of the binary search above: 736 assert(min_young_length < max_young_length, 737 "otherwise we should have discovered that max_young_length " 738 "fits into the pause target and not done the binary search"); 739 assert(predict_will_fit(min_young_length, base_time_ms, 740 base_free_regions, target_pause_time_ms), 741 "min_young_length, the result of the binary search, should " 742 "fit into the pause target"); 743 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 744 base_free_regions, target_pause_time_ms), 745 "min_young_length, the result of the binary search, should be " 746 "optimal, so no larger length should fit into the pause target"); 747 } 748 } else { 749 // Even the minimum length doesn't fit into the pause time 750 // target, return it as the result nevertheless. 751 } 752 return base_min_length + min_young_length; 753 } 754 755 double G1CollectorPolicy::predict_survivor_regions_evac_time() { 756 double survivor_regions_evac_time = 0.0; 757 for (HeapRegion * r = _recorded_survivor_head; 758 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 759 r = r->get_next_young_region()) { 760 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true); 761 } 762 return survivor_regions_evac_time; 763 } 764 765 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 766 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 767 768 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 769 if (rs_lengths > _rs_lengths_prediction) { 770 // add 10% to avoid having to recalculate often 771 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 772 update_young_list_target_length(rs_lengths_prediction); 773 } 774 } 775 776 777 778 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 779 bool is_tlab, 780 bool* gc_overhead_limit_was_exceeded) { 781 guarantee(false, "Not using this policy feature yet."); 782 return NULL; 783 } 784 785 // This method controls how a collector handles one or more 786 // of its generations being fully allocated. 787 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 788 bool is_tlab) { 789 guarantee(false, "Not using this policy feature yet."); 790 return NULL; 791 } 792 793 794 #ifndef PRODUCT 795 bool G1CollectorPolicy::verify_young_ages() { 796 HeapRegion* head = _g1->young_list()->first_region(); 797 return 798 verify_young_ages(head, _short_lived_surv_rate_group); 799 // also call verify_young_ages on any additional surv rate groups 800 } 801 802 bool 803 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 804 SurvRateGroup *surv_rate_group) { 805 guarantee( surv_rate_group != NULL, "pre-condition" ); 806 807 const char* name = surv_rate_group->name(); 808 bool ret = true; 809 int prev_age = -1; 810 811 for (HeapRegion* curr = head; 812 curr != NULL; 813 curr = curr->get_next_young_region()) { 814 SurvRateGroup* group = curr->surv_rate_group(); 815 if (group == NULL && !curr->is_survivor()) { 816 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 817 ret = false; 818 } 819 820 if (surv_rate_group == group) { 821 int age = curr->age_in_surv_rate_group(); 822 823 if (age < 0) { 824 gclog_or_tty->print_cr("## %s: encountered negative age", name); 825 ret = false; 826 } 827 828 if (age <= prev_age) { 829 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 830 "(%d, %d)", name, age, prev_age); 831 ret = false; 832 } 833 prev_age = age; 834 } 835 } 836 837 return ret; 838 } 839 #endif // PRODUCT 840 841 void G1CollectorPolicy::record_full_collection_start() { 842 _cur_collection_start_sec = os::elapsedTime(); 843 // Release the future to-space so that it is available for compaction into. 844 _g1->set_full_collection(); 845 } 846 847 void G1CollectorPolicy::record_full_collection_end() { 848 // Consider this like a collection pause for the purposes of allocation 849 // since last pause. 850 double end_sec = os::elapsedTime(); 851 double full_gc_time_sec = end_sec - _cur_collection_start_sec; 852 double full_gc_time_ms = full_gc_time_sec * 1000.0; 853 854 _all_full_gc_times_ms->add(full_gc_time_ms); 855 856 update_recent_gc_times(end_sec, full_gc_time_ms); 857 858 _g1->clear_full_collection(); 859 860 // "Nuke" the heuristics that control the young/mixed GC 861 // transitions and make sure we start with young GCs after the Full GC. 862 set_gcs_are_young(true); 863 _last_young_gc = false; 864 clear_initiate_conc_mark_if_possible(); 865 clear_during_initial_mark_pause(); 866 _in_marking_window = false; 867 _in_marking_window_im = false; 868 869 _short_lived_surv_rate_group->start_adding_regions(); 870 // also call this on any additional surv rate groups 871 872 record_survivor_regions(0, NULL, NULL); 873 874 _free_regions_at_end_of_collection = _g1->free_regions(); 875 // Reset survivors SurvRateGroup. 876 _survivor_surv_rate_group->reset(); 877 update_young_list_target_length(); 878 _collectionSetChooser->clear(); 879 } 880 881 void G1CollectorPolicy::record_stop_world_start() { 882 _stop_world_start = os::elapsedTime(); 883 } 884 885 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, 886 size_t start_used) { 887 if (G1Log::finer()) { 888 gclog_or_tty->stamp(PrintGCTimeStamps); 889 gclog_or_tty->print("[%s", (const char*)GCCauseString("GC pause", _g1->gc_cause()) 890 .append(gcs_are_young() ? " (young)" : " (mixed)")); 891 } 892 893 // We only need to do this here as the policy will only be applied 894 // to the GC we're about to start. so, no point is calculating this 895 // every time we calculate / recalculate the target young length. 896 update_survivors_policy(); 897 898 assert(_g1->used() == _g1->recalculate_used(), 899 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT, 900 _g1->used(), _g1->recalculate_used())); 901 902 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 903 _all_stop_world_times_ms->add(s_w_t_ms); 904 _stop_world_start = 0.0; 905 906 _cur_collection_start_sec = start_time_sec; 907 _cur_collection_pause_used_at_start_bytes = start_used; 908 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); 909 _pending_cards = _g1->pending_card_num(); 910 _max_pending_cards = _g1->max_pending_card_num(); 911 912 _bytes_in_collection_set_before_gc = 0; 913 _bytes_copied_during_gc = 0; 914 915 YoungList* young_list = _g1->young_list(); 916 _eden_bytes_before_gc = young_list->eden_used_bytes(); 917 _survivor_bytes_before_gc = young_list->survivor_used_bytes(); 918 _capacity_before_gc = _g1->capacity(); 919 920 #ifdef DEBUG 921 // initialise these to something well known so that we can spot 922 // if they are not set properly 923 924 for (int i = 0; i < _parallel_gc_threads; ++i) { 925 _par_last_gc_worker_start_times_ms[i] = -1234.0; 926 _par_last_ext_root_scan_times_ms[i] = -1234.0; 927 _par_last_satb_filtering_times_ms[i] = -1234.0; 928 _par_last_update_rs_times_ms[i] = -1234.0; 929 _par_last_update_rs_processed_buffers[i] = -1234.0; 930 _par_last_scan_rs_times_ms[i] = -1234.0; 931 _par_last_obj_copy_times_ms[i] = -1234.0; 932 _par_last_termination_times_ms[i] = -1234.0; 933 _par_last_termination_attempts[i] = -1234.0; 934 _par_last_gc_worker_end_times_ms[i] = -1234.0; 935 _par_last_gc_worker_times_ms[i] = -1234.0; 936 _par_last_gc_worker_other_times_ms[i] = -1234.0; 937 } 938 #endif 939 940 for (int i = 0; i < _aux_num; ++i) { 941 _cur_aux_times_ms[i] = 0.0; 942 _cur_aux_times_set[i] = false; 943 } 944 945 // This is initialized to zero here and is set during the evacuation 946 // pause if we actually waited for the root region scanning to finish. 947 _root_region_scan_wait_time_ms = 0.0; 948 949 _last_gc_was_young = false; 950 951 // do that for any other surv rate groups 952 _short_lived_surv_rate_group->stop_adding_regions(); 953 _survivors_age_table.clear(); 954 955 assert( verify_young_ages(), "region age verification" ); 956 } 957 958 void G1CollectorPolicy::record_concurrent_mark_init_end(double 959 mark_init_elapsed_time_ms) { 960 _during_marking = true; 961 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); 962 clear_during_initial_mark_pause(); 963 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 964 } 965 966 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 967 _mark_remark_start_sec = os::elapsedTime(); 968 _during_marking = false; 969 } 970 971 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 972 double end_time_sec = os::elapsedTime(); 973 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 974 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 975 _cur_mark_stop_world_time_ms += elapsed_time_ms; 976 _prev_collection_pause_end_ms += elapsed_time_ms; 977 978 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); 979 } 980 981 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 982 _mark_cleanup_start_sec = os::elapsedTime(); 983 } 984 985 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 986 _last_young_gc = true; 987 _in_marking_window = false; 988 } 989 990 void G1CollectorPolicy::record_concurrent_pause() { 991 if (_stop_world_start > 0.0) { 992 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 993 _all_yield_times_ms->add(yield_ms); 994 } 995 } 996 997 void G1CollectorPolicy::record_concurrent_pause_end() { 998 } 999 1000 template<class T> 1001 T sum_of(T* sum_arr, int start, int n, int N) { 1002 T sum = (T)0; 1003 for (int i = 0; i < n; i++) { 1004 int j = (start + i) % N; 1005 sum += sum_arr[j]; 1006 } 1007 return sum; 1008 } 1009 1010 void G1CollectorPolicy::print_par_stats(int level, 1011 const char* str, 1012 double* data, 1013 bool showDecimals) { 1014 double min = data[0], max = data[0]; 1015 double total = 0.0; 1016 LineBuffer buf(level); 1017 buf.append("[%s (ms):", str); 1018 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1019 double val = data[i]; 1020 if (val < min) 1021 min = val; 1022 if (val > max) 1023 max = val; 1024 total += val; 1025 if (G1Log::finest()) { 1026 if (showDecimals) { 1027 buf.append(" %.1lf", val); 1028 } else { 1029 buf.append(" %d", (int)val); 1030 } 1031 } 1032 } 1033 1034 if (G1Log::finest()) { 1035 buf.append_and_print_cr(""); 1036 } 1037 double avg = total / (double) no_of_gc_threads(); 1038 if (showDecimals) { 1039 buf.append_and_print_cr(" Min: %.1lf, Avg: %.1lf, Max: %.1lf, Diff: %.1lf, Sum: %.1lf]", 1040 min, avg, max, max - min, total); 1041 } else { 1042 buf.append_and_print_cr(" Min: %d, Avg: %d, Max: %d, Diff: %d, Sum: %d]", 1043 (int)min, (int)avg, (int)max, (int)max - (int)min, (int)total); 1044 } 1045 } 1046 1047 void G1CollectorPolicy::print_stats(int level, 1048 const char* str, 1049 double value) { 1050 LineBuffer(level).append_and_print_cr("[%s: %.1lf ms]", str, value); 1051 } 1052 1053 void G1CollectorPolicy::print_stats(int level, 1054 const char* str, 1055 double value, 1056 int workers) { 1057 LineBuffer(level).append_and_print_cr("[%s: %.1lf ms, GC Workers: %d]", str, value, workers); 1058 } 1059 1060 void G1CollectorPolicy::print_stats(int level, 1061 const char* str, 1062 int value) { 1063 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value); 1064 } 1065 1066 double G1CollectorPolicy::avg_value(double* data) { 1067 if (G1CollectedHeap::use_parallel_gc_threads()) { 1068 double ret = 0.0; 1069 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1070 ret += data[i]; 1071 } 1072 return ret / (double) no_of_gc_threads(); 1073 } else { 1074 return data[0]; 1075 } 1076 } 1077 1078 double G1CollectorPolicy::max_value(double* data) { 1079 if (G1CollectedHeap::use_parallel_gc_threads()) { 1080 double ret = data[0]; 1081 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1082 if (data[i] > ret) { 1083 ret = data[i]; 1084 } 1085 } 1086 return ret; 1087 } else { 1088 return data[0]; 1089 } 1090 } 1091 1092 double G1CollectorPolicy::sum_of_values(double* data) { 1093 if (G1CollectedHeap::use_parallel_gc_threads()) { 1094 double sum = 0.0; 1095 for (uint i = 0; i < no_of_gc_threads(); i++) { 1096 sum += data[i]; 1097 } 1098 return sum; 1099 } else { 1100 return data[0]; 1101 } 1102 } 1103 1104 double G1CollectorPolicy::max_sum(double* data1, double* data2) { 1105 double ret = data1[0] + data2[0]; 1106 1107 if (G1CollectedHeap::use_parallel_gc_threads()) { 1108 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1109 double data = data1[i] + data2[i]; 1110 if (data > ret) { 1111 ret = data; 1112 } 1113 } 1114 } 1115 return ret; 1116 } 1117 1118 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 1119 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 1120 return false; 1121 } 1122 1123 size_t marking_initiating_used_threshold = 1124 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 1125 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 1126 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 1127 1128 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 1129 if (gcs_are_young()) { 1130 ergo_verbose5(ErgoConcCycles, 1131 "request concurrent cycle initiation", 1132 ergo_format_reason("occupancy higher than threshold") 1133 ergo_format_byte("occupancy") 1134 ergo_format_byte("allocation request") 1135 ergo_format_byte_perc("threshold") 1136 ergo_format_str("source"), 1137 cur_used_bytes, 1138 alloc_byte_size, 1139 marking_initiating_used_threshold, 1140 (double) InitiatingHeapOccupancyPercent, 1141 source); 1142 return true; 1143 } else { 1144 ergo_verbose5(ErgoConcCycles, 1145 "do not request concurrent cycle initiation", 1146 ergo_format_reason("still doing mixed collections") 1147 ergo_format_byte("occupancy") 1148 ergo_format_byte("allocation request") 1149 ergo_format_byte_perc("threshold") 1150 ergo_format_str("source"), 1151 cur_used_bytes, 1152 alloc_byte_size, 1153 marking_initiating_used_threshold, 1154 (double) InitiatingHeapOccupancyPercent, 1155 source); 1156 } 1157 } 1158 1159 return false; 1160 } 1161 1162 // Anything below that is considered to be zero 1163 #define MIN_TIMER_GRANULARITY 0.0000001 1164 1165 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) { 1166 double end_time_sec = os::elapsedTime(); 1167 double elapsed_ms = _last_pause_time_ms; 1168 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1169 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1170 "otherwise, the subtraction below does not make sense"); 1171 size_t rs_size = 1172 _cur_collection_pause_used_regions_at_start - cset_region_length(); 1173 size_t cur_used_bytes = _g1->used(); 1174 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1175 bool last_pause_included_initial_mark = false; 1176 bool update_stats = !_g1->evacuation_failed(); 1177 set_no_of_gc_threads(no_of_gc_threads); 1178 1179 #ifndef PRODUCT 1180 if (G1YoungSurvRateVerbose) { 1181 gclog_or_tty->print_cr(""); 1182 _short_lived_surv_rate_group->print(); 1183 // do that for any other surv rate groups too 1184 } 1185 #endif // PRODUCT 1186 1187 last_pause_included_initial_mark = during_initial_mark_pause(); 1188 if (last_pause_included_initial_mark) { 1189 record_concurrent_mark_init_end(0.0); 1190 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) { 1191 // Note: this might have already been set, if during the last 1192 // pause we decided to start a cycle but at the beginning of 1193 // this pause we decided to postpone it. That's OK. 1194 set_initiate_conc_mark_if_possible(); 1195 } 1196 1197 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, 1198 end_time_sec, false); 1199 1200 // This assert is exempted when we're doing parallel collection pauses, 1201 // because the fragmentation caused by the parallel GC allocation buffers 1202 // can lead to more memory being used during collection than was used 1203 // before. Best leave this out until the fragmentation problem is fixed. 1204 // Pauses in which evacuation failed can also lead to negative 1205 // collections, since no space is reclaimed from a region containing an 1206 // object whose evacuation failed. 1207 // Further, we're now always doing parallel collection. But I'm still 1208 // leaving this here as a placeholder for a more precise assertion later. 1209 // (DLD, 10/05.) 1210 assert((true || parallel) // Always using GC LABs now. 1211 || _g1->evacuation_failed() 1212 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes, 1213 "Negative collection"); 1214 1215 size_t freed_bytes = 1216 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; 1217 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; 1218 1219 double survival_fraction = 1220 (double)surviving_bytes/ 1221 (double)_collection_set_bytes_used_before; 1222 1223 // These values are used to update the summary information that is 1224 // displayed when TraceGen0Time is enabled, and are output as part 1225 // of the "finer" output, in the non-parallel case. 1226 1227 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); 1228 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms); 1229 double update_rs_time = avg_value(_par_last_update_rs_times_ms); 1230 double update_rs_processed_buffers = 1231 sum_of_values(_par_last_update_rs_processed_buffers); 1232 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); 1233 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); 1234 double termination_time = avg_value(_par_last_termination_times_ms); 1235 1236 double known_time = ext_root_scan_time + 1237 satb_filtering_time + 1238 update_rs_time + 1239 scan_rs_time + 1240 obj_copy_time; 1241 1242 double other_time_ms = elapsed_ms; 1243 1244 // Subtract the root region scanning wait time. It's initialized to 1245 // zero at the start of the pause. 1246 other_time_ms -= _root_region_scan_wait_time_ms; 1247 1248 if (parallel) { 1249 other_time_ms -= _cur_collection_par_time_ms; 1250 } else { 1251 other_time_ms -= known_time; 1252 } 1253 1254 // Now subtract the time taken to fix up roots in generated code 1255 other_time_ms -= _cur_collection_code_root_fixup_time_ms; 1256 1257 // Subtract the time taken to clean the card table from the 1258 // current value of "other time" 1259 other_time_ms -= _cur_clear_ct_time_ms; 1260 1261 // TraceGen0Time and TraceGen1Time summary info updating. 1262 _all_pause_times_ms->add(elapsed_ms); 1263 1264 if (update_stats) { 1265 _summary->record_total_time_ms(elapsed_ms); 1266 _summary->record_other_time_ms(other_time_ms); 1267 1268 MainBodySummary* body_summary = _summary->main_body_summary(); 1269 assert(body_summary != NULL, "should not be null!"); 1270 1271 body_summary->record_root_region_scan_wait_time_ms( 1272 _root_region_scan_wait_time_ms); 1273 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time); 1274 body_summary->record_satb_filtering_time_ms(satb_filtering_time); 1275 body_summary->record_update_rs_time_ms(update_rs_time); 1276 body_summary->record_scan_rs_time_ms(scan_rs_time); 1277 body_summary->record_obj_copy_time_ms(obj_copy_time); 1278 1279 if (parallel) { 1280 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms); 1281 body_summary->record_termination_time_ms(termination_time); 1282 1283 double parallel_known_time = known_time + termination_time; 1284 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time; 1285 body_summary->record_parallel_other_time_ms(parallel_other_time); 1286 } 1287 1288 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms); 1289 1290 // We exempt parallel collection from this check because Alloc Buffer 1291 // fragmentation can produce negative collections. Same with evac 1292 // failure. 1293 // Further, we're now always doing parallel collection. But I'm still 1294 // leaving this here as a placeholder for a more precise assertion later. 1295 // (DLD, 10/05. 1296 assert((true || parallel) 1297 || _g1->evacuation_failed() 1298 || surviving_bytes <= _collection_set_bytes_used_before, 1299 "Or else negative collection!"); 1300 1301 // this is where we update the allocation rate of the application 1302 double app_time_ms = 1303 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); 1304 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1305 // This usually happens due to the timer not having the required 1306 // granularity. Some Linuxes are the usual culprits. 1307 // We'll just set it to something (arbitrarily) small. 1308 app_time_ms = 1.0; 1309 } 1310 // We maintain the invariant that all objects allocated by mutator 1311 // threads will be allocated out of eden regions. So, we can use 1312 // the eden region number allocated since the previous GC to 1313 // calculate the application's allocate rate. The only exception 1314 // to that is humongous objects that are allocated separately. But 1315 // given that humongous object allocations do not really affect 1316 // either the pause's duration nor when the next pause will take 1317 // place we can safely ignore them here. 1318 uint regions_allocated = eden_cset_region_length(); 1319 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1320 _alloc_rate_ms_seq->add(alloc_rate_ms); 1321 1322 double interval_ms = 1323 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1324 update_recent_gc_times(end_time_sec, elapsed_ms); 1325 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1326 if (recent_avg_pause_time_ratio() < 0.0 || 1327 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1328 #ifndef PRODUCT 1329 // Dump info to allow post-facto debugging 1330 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1331 gclog_or_tty->print_cr("-------------------------------------------"); 1332 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1333 _recent_gc_times_ms->dump(); 1334 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1335 _recent_prev_end_times_for_all_gcs_sec->dump(); 1336 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1337 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1338 // In debug mode, terminate the JVM if the user wants to debug at this point. 1339 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1340 #endif // !PRODUCT 1341 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1342 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1343 if (_recent_avg_pause_time_ratio < 0.0) { 1344 _recent_avg_pause_time_ratio = 0.0; 1345 } else { 1346 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1347 _recent_avg_pause_time_ratio = 1.0; 1348 } 1349 } 1350 } 1351 1352 for (int i = 0; i < _aux_num; ++i) { 1353 if (_cur_aux_times_set[i]) { 1354 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]); 1355 } 1356 } 1357 1358 if (G1Log::finer()) { 1359 bool print_marking_info = 1360 _g1->mark_in_progress() && !last_pause_included_initial_mark; 1361 1362 gclog_or_tty->print_cr("%s, %1.8lf secs]", 1363 (last_pause_included_initial_mark) ? " (initial-mark)" : "", 1364 elapsed_ms / 1000.0); 1365 1366 if (_root_region_scan_wait_time_ms > 0.0) { 1367 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms); 1368 } 1369 if (parallel) { 1370 print_stats(1, "Parallel Time", _cur_collection_par_time_ms, no_of_gc_threads); 1371 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms); 1372 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms); 1373 if (print_marking_info) { 1374 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms); 1375 } 1376 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); 1377 if (G1Log::finest()) { 1378 print_par_stats(3, "Processed Buffers", _par_last_update_rs_processed_buffers, 1379 false /* showDecimals */); 1380 } 1381 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); 1382 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); 1383 print_par_stats(2, "Termination", _par_last_termination_times_ms); 1384 if (G1Log::finest()) { 1385 print_par_stats(3, "Termination Attempts", _par_last_termination_attempts, 1386 false /* showDecimals */); 1387 } 1388 1389 for (int i = 0; i < _parallel_gc_threads; i++) { 1390 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - 1391 _par_last_gc_worker_start_times_ms[i]; 1392 1393 double worker_known_time = _par_last_ext_root_scan_times_ms[i] + 1394 _par_last_satb_filtering_times_ms[i] + 1395 _par_last_update_rs_times_ms[i] + 1396 _par_last_scan_rs_times_ms[i] + 1397 _par_last_obj_copy_times_ms[i] + 1398 _par_last_termination_times_ms[i]; 1399 1400 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] - 1401 worker_known_time; 1402 } 1403 1404 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms); 1405 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms); 1406 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms); 1407 } else { 1408 print_stats(1, "Ext Root Scanning", ext_root_scan_time); 1409 if (print_marking_info) { 1410 print_stats(1, "SATB Filtering", satb_filtering_time); 1411 } 1412 print_stats(1, "Update RS", update_rs_time); 1413 if (G1Log::finest()) { 1414 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers); 1415 } 1416 print_stats(1, "Scan RS", scan_rs_time); 1417 print_stats(1, "Object Copying", obj_copy_time); 1418 } 1419 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms); 1420 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); 1421 #ifndef PRODUCT 1422 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms); 1423 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms); 1424 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms); 1425 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms); 1426 if (_num_cc_clears > 0) { 1427 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears)); 1428 } 1429 #endif 1430 print_stats(1, "Other", other_time_ms); 1431 print_stats(2, "Choose CSet", 1432 (_recorded_young_cset_choice_time_ms + 1433 _recorded_non_young_cset_choice_time_ms)); 1434 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms); 1435 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms); 1436 print_stats(2, "Free CSet", 1437 (_recorded_young_free_cset_time_ms + 1438 _recorded_non_young_free_cset_time_ms)); 1439 1440 for (int i = 0; i < _aux_num; ++i) { 1441 if (_cur_aux_times_set[i]) { 1442 char buffer[96]; 1443 sprintf(buffer, "Aux%d", i); 1444 print_stats(1, buffer, _cur_aux_times_ms[i]); 1445 } 1446 } 1447 } 1448 1449 bool new_in_marking_window = _in_marking_window; 1450 bool new_in_marking_window_im = false; 1451 if (during_initial_mark_pause()) { 1452 new_in_marking_window = true; 1453 new_in_marking_window_im = true; 1454 } 1455 1456 if (_last_young_gc) { 1457 // This is supposed to to be the "last young GC" before we start 1458 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1459 1460 if (!last_pause_included_initial_mark) { 1461 if (next_gc_should_be_mixed("start mixed GCs", 1462 "do not start mixed GCs")) { 1463 set_gcs_are_young(false); 1464 } 1465 } else { 1466 ergo_verbose0(ErgoMixedGCs, 1467 "do not start mixed GCs", 1468 ergo_format_reason("concurrent cycle is about to start")); 1469 } 1470 _last_young_gc = false; 1471 } 1472 1473 if (!_last_gc_was_young) { 1474 // This is a mixed GC. Here we decide whether to continue doing 1475 // mixed GCs or not. 1476 1477 if (!next_gc_should_be_mixed("continue mixed GCs", 1478 "do not continue mixed GCs")) { 1479 set_gcs_are_young(true); 1480 } 1481 } 1482 1483 _short_lived_surv_rate_group->start_adding_regions(); 1484 // do that for any other surv rate groupsx 1485 1486 if (update_stats) { 1487 double pause_time_ms = elapsed_ms; 1488 1489 size_t diff = 0; 1490 if (_max_pending_cards >= _pending_cards) { 1491 diff = _max_pending_cards - _pending_cards; 1492 } 1493 _pending_card_diff_seq->add((double) diff); 1494 1495 double cost_per_card_ms = 0.0; 1496 if (_pending_cards > 0) { 1497 cost_per_card_ms = update_rs_time / (double) _pending_cards; 1498 _cost_per_card_ms_seq->add(cost_per_card_ms); 1499 } 1500 1501 size_t cards_scanned = _g1->cards_scanned(); 1502 1503 double cost_per_entry_ms = 0.0; 1504 if (cards_scanned > 10) { 1505 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; 1506 if (_last_gc_was_young) { 1507 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1508 } else { 1509 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1510 } 1511 } 1512 1513 if (_max_rs_lengths > 0) { 1514 double cards_per_entry_ratio = 1515 (double) cards_scanned / (double) _max_rs_lengths; 1516 if (_last_gc_was_young) { 1517 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1518 } else { 1519 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1520 } 1521 } 1522 1523 // This is defensive. For a while _max_rs_lengths could get 1524 // smaller than _recorded_rs_lengths which was causing 1525 // rs_length_diff to get very large and mess up the RSet length 1526 // predictions. The reason was unsafe concurrent updates to the 1527 // _inc_cset_recorded_rs_lengths field which the code below guards 1528 // against (see CR 7118202). This bug has now been fixed (see CR 1529 // 7119027). However, I'm still worried that 1530 // _inc_cset_recorded_rs_lengths might still end up somewhat 1531 // inaccurate. The concurrent refinement thread calculates an 1532 // RSet's length concurrently with other CR threads updating it 1533 // which might cause it to calculate the length incorrectly (if, 1534 // say, it's in mid-coarsening). So I'll leave in the defensive 1535 // conditional below just in case. 1536 size_t rs_length_diff = 0; 1537 if (_max_rs_lengths > _recorded_rs_lengths) { 1538 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1539 } 1540 _rs_length_diff_seq->add((double) rs_length_diff); 1541 1542 size_t copied_bytes = surviving_bytes; 1543 double cost_per_byte_ms = 0.0; 1544 if (copied_bytes > 0) { 1545 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; 1546 if (_in_marking_window) { 1547 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1548 } else { 1549 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1550 } 1551 } 1552 1553 double all_other_time_ms = pause_time_ms - 1554 (update_rs_time + scan_rs_time + obj_copy_time + termination_time); 1555 1556 double young_other_time_ms = 0.0; 1557 if (young_cset_region_length() > 0) { 1558 young_other_time_ms = 1559 _recorded_young_cset_choice_time_ms + 1560 _recorded_young_free_cset_time_ms; 1561 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1562 (double) young_cset_region_length()); 1563 } 1564 double non_young_other_time_ms = 0.0; 1565 if (old_cset_region_length() > 0) { 1566 non_young_other_time_ms = 1567 _recorded_non_young_cset_choice_time_ms + 1568 _recorded_non_young_free_cset_time_ms; 1569 1570 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1571 (double) old_cset_region_length()); 1572 } 1573 1574 double constant_other_time_ms = all_other_time_ms - 1575 (young_other_time_ms + non_young_other_time_ms); 1576 _constant_other_time_ms_seq->add(constant_other_time_ms); 1577 1578 double survival_ratio = 0.0; 1579 if (_bytes_in_collection_set_before_gc > 0) { 1580 survival_ratio = (double) _bytes_copied_during_gc / 1581 (double) _bytes_in_collection_set_before_gc; 1582 } 1583 1584 _pending_cards_seq->add((double) _pending_cards); 1585 _rs_lengths_seq->add((double) _max_rs_lengths); 1586 } 1587 1588 _in_marking_window = new_in_marking_window; 1589 _in_marking_window_im = new_in_marking_window_im; 1590 _free_regions_at_end_of_collection = _g1->free_regions(); 1591 update_young_list_target_length(); 1592 1593 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1594 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1595 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms); 1596 1597 _collectionSetChooser->verify(); 1598 } 1599 1600 #define EXT_SIZE_FORMAT "%.1f%s" 1601 #define EXT_SIZE_PARAMS(bytes) \ 1602 byte_size_in_proper_unit((double)(bytes)), \ 1603 proper_unit_for_byte_size((bytes)) 1604 1605 void G1CollectorPolicy::print_heap_transition() { 1606 if (G1Log::finer()) { 1607 YoungList* young_list = _g1->young_list(); 1608 size_t eden_bytes = young_list->eden_used_bytes(); 1609 size_t survivor_bytes = young_list->survivor_used_bytes(); 1610 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes; 1611 size_t used = _g1->used(); 1612 size_t capacity = _g1->capacity(); 1613 size_t eden_capacity = 1614 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes; 1615 1616 gclog_or_tty->print_cr( 1617 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") " 1618 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" " 1619 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->" 1620 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]", 1621 EXT_SIZE_PARAMS(_eden_bytes_before_gc), 1622 EXT_SIZE_PARAMS(_prev_eden_capacity), 1623 EXT_SIZE_PARAMS(eden_bytes), 1624 EXT_SIZE_PARAMS(eden_capacity), 1625 EXT_SIZE_PARAMS(_survivor_bytes_before_gc), 1626 EXT_SIZE_PARAMS(survivor_bytes), 1627 EXT_SIZE_PARAMS(used_before_gc), 1628 EXT_SIZE_PARAMS(_capacity_before_gc), 1629 EXT_SIZE_PARAMS(used), 1630 EXT_SIZE_PARAMS(capacity)); 1631 1632 _prev_eden_capacity = eden_capacity; 1633 } else if (G1Log::fine()) { 1634 _g1->print_size_transition(gclog_or_tty, 1635 _cur_collection_pause_used_at_start_bytes, 1636 _g1->used(), _g1->capacity()); 1637 } 1638 } 1639 1640 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1641 double update_rs_processed_buffers, 1642 double goal_ms) { 1643 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1644 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1645 1646 if (G1UseAdaptiveConcRefinement) { 1647 const int k_gy = 3, k_gr = 6; 1648 const double inc_k = 1.1, dec_k = 0.9; 1649 1650 int g = cg1r->green_zone(); 1651 if (update_rs_time > goal_ms) { 1652 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1653 } else { 1654 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1655 g = (int)MAX2(g * inc_k, g + 1.0); 1656 } 1657 } 1658 // Change the refinement threads params 1659 cg1r->set_green_zone(g); 1660 cg1r->set_yellow_zone(g * k_gy); 1661 cg1r->set_red_zone(g * k_gr); 1662 cg1r->reinitialize_threads(); 1663 1664 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); 1665 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1666 cg1r->yellow_zone()); 1667 // Change the barrier params 1668 dcqs.set_process_completed_threshold(processing_threshold); 1669 dcqs.set_max_completed_queue(cg1r->red_zone()); 1670 } 1671 1672 int curr_queue_size = dcqs.completed_buffers_num(); 1673 if (curr_queue_size >= cg1r->yellow_zone()) { 1674 dcqs.set_completed_queue_padding(curr_queue_size); 1675 } else { 1676 dcqs.set_completed_queue_padding(0); 1677 } 1678 dcqs.notify_if_necessary(); 1679 } 1680 1681 double 1682 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { 1683 size_t rs_length = predict_rs_length_diff(); 1684 size_t card_num; 1685 if (gcs_are_young()) { 1686 card_num = predict_young_card_num(rs_length); 1687 } else { 1688 card_num = predict_non_young_card_num(rs_length); 1689 } 1690 return predict_base_elapsed_time_ms(pending_cards, card_num); 1691 } 1692 1693 double 1694 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1695 size_t scanned_cards) { 1696 return 1697 predict_rs_update_time_ms(pending_cards) + 1698 predict_rs_scan_time_ms(scanned_cards) + 1699 predict_constant_other_time_ms(); 1700 } 1701 1702 double 1703 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1704 bool young) { 1705 size_t rs_length = hr->rem_set()->occupied(); 1706 size_t card_num; 1707 if (gcs_are_young()) { 1708 card_num = predict_young_card_num(rs_length); 1709 } else { 1710 card_num = predict_non_young_card_num(rs_length); 1711 } 1712 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1713 1714 double region_elapsed_time_ms = 1715 predict_rs_scan_time_ms(card_num) + 1716 predict_object_copy_time_ms(bytes_to_copy); 1717 1718 if (young) 1719 region_elapsed_time_ms += predict_young_other_time_ms(1); 1720 else 1721 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1722 1723 return region_elapsed_time_ms; 1724 } 1725 1726 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { 1727 size_t bytes_to_copy; 1728 if (hr->is_marked()) 1729 bytes_to_copy = hr->max_live_bytes(); 1730 else { 1731 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1732 int age = hr->age_in_surv_rate_group(); 1733 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1734 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); 1735 } 1736 return bytes_to_copy; 1737 } 1738 1739 void 1740 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1741 uint survivor_cset_region_length) { 1742 _eden_cset_region_length = eden_cset_region_length; 1743 _survivor_cset_region_length = survivor_cset_region_length; 1744 _old_cset_region_length = 0; 1745 } 1746 1747 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1748 _recorded_rs_lengths = rs_lengths; 1749 } 1750 1751 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1752 double elapsed_ms) { 1753 _recent_gc_times_ms->add(elapsed_ms); 1754 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1755 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1756 } 1757 1758 size_t G1CollectorPolicy::expansion_amount() { 1759 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1760 double threshold = _gc_overhead_perc; 1761 if (recent_gc_overhead > threshold) { 1762 // We will double the existing space, or take 1763 // G1ExpandByPercentOfAvailable % of the available expansion 1764 // space, whichever is smaller, bounded below by a minimum 1765 // expansion (unless that's all that's left.) 1766 const size_t min_expand_bytes = 1*M; 1767 size_t reserved_bytes = _g1->max_capacity(); 1768 size_t committed_bytes = _g1->capacity(); 1769 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1770 size_t expand_bytes; 1771 size_t expand_bytes_via_pct = 1772 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1773 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1774 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1775 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1776 1777 ergo_verbose5(ErgoHeapSizing, 1778 "attempt heap expansion", 1779 ergo_format_reason("recent GC overhead higher than " 1780 "threshold after GC") 1781 ergo_format_perc("recent GC overhead") 1782 ergo_format_perc("threshold") 1783 ergo_format_byte("uncommitted") 1784 ergo_format_byte_perc("calculated expansion amount"), 1785 recent_gc_overhead, threshold, 1786 uncommitted_bytes, 1787 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1788 1789 return expand_bytes; 1790 } else { 1791 return 0; 1792 } 1793 } 1794 1795 class CountCSClosure: public HeapRegionClosure { 1796 G1CollectorPolicy* _g1_policy; 1797 public: 1798 CountCSClosure(G1CollectorPolicy* g1_policy) : 1799 _g1_policy(g1_policy) {} 1800 bool doHeapRegion(HeapRegion* r) { 1801 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); 1802 return false; 1803 } 1804 }; 1805 1806 void G1CollectorPolicy::count_CS_bytes_used() { 1807 CountCSClosure cs_closure(this); 1808 _g1->collection_set_iterate(&cs_closure); 1809 } 1810 1811 void G1CollectorPolicy::print_summary(int level, 1812 const char* str, 1813 NumberSeq* seq) const { 1814 double sum = seq->sum(); 1815 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", 1816 str, sum / 1000.0, seq->avg()); 1817 } 1818 1819 void G1CollectorPolicy::print_summary_sd(int level, 1820 const char* str, 1821 NumberSeq* seq) const { 1822 print_summary(level, str, seq); 1823 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 1824 seq->num(), seq->sd(), seq->maximum()); 1825 } 1826 1827 void G1CollectorPolicy::check_other_times(int level, 1828 NumberSeq* other_times_ms, 1829 NumberSeq* calc_other_times_ms) const { 1830 bool should_print = false; 1831 LineBuffer buf(level + 2); 1832 1833 double max_sum = MAX2(fabs(other_times_ms->sum()), 1834 fabs(calc_other_times_ms->sum())); 1835 double min_sum = MIN2(fabs(other_times_ms->sum()), 1836 fabs(calc_other_times_ms->sum())); 1837 double sum_ratio = max_sum / min_sum; 1838 if (sum_ratio > 1.1) { 1839 should_print = true; 1840 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###"); 1841 } 1842 1843 double max_avg = MAX2(fabs(other_times_ms->avg()), 1844 fabs(calc_other_times_ms->avg())); 1845 double min_avg = MIN2(fabs(other_times_ms->avg()), 1846 fabs(calc_other_times_ms->avg())); 1847 double avg_ratio = max_avg / min_avg; 1848 if (avg_ratio > 1.1) { 1849 should_print = true; 1850 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###"); 1851 } 1852 1853 if (other_times_ms->sum() < -0.01) { 1854 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###"); 1855 } 1856 1857 if (other_times_ms->avg() < -0.01) { 1858 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###"); 1859 } 1860 1861 if (calc_other_times_ms->sum() < -0.01) { 1862 should_print = true; 1863 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###"); 1864 } 1865 1866 if (calc_other_times_ms->avg() < -0.01) { 1867 should_print = true; 1868 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###"); 1869 } 1870 1871 if (should_print) 1872 print_summary(level, "Other(Calc)", calc_other_times_ms); 1873 } 1874 1875 void G1CollectorPolicy::print_summary(PauseSummary* summary) const { 1876 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1877 MainBodySummary* body_summary = summary->main_body_summary(); 1878 if (summary->get_total_seq()->num() > 0) { 1879 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq()); 1880 if (body_summary != NULL) { 1881 print_summary(1, "Root Region Scan Wait", body_summary->get_root_region_scan_wait_seq()); 1882 if (parallel) { 1883 print_summary(1, "Parallel Time", body_summary->get_parallel_seq()); 1884 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1885 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq()); 1886 print_summary(2, "Update RS", body_summary->get_update_rs_seq()); 1887 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq()); 1888 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq()); 1889 print_summary(2, "Termination", body_summary->get_termination_seq()); 1890 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq()); 1891 { 1892 NumberSeq* other_parts[] = { 1893 body_summary->get_ext_root_scan_seq(), 1894 body_summary->get_satb_filtering_seq(), 1895 body_summary->get_update_rs_seq(), 1896 body_summary->get_scan_rs_seq(), 1897 body_summary->get_obj_copy_seq(), 1898 body_summary->get_termination_seq() 1899 }; 1900 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(), 1901 6, other_parts); 1902 check_other_times(2, body_summary->get_parallel_other_seq(), 1903 &calc_other_times_ms); 1904 } 1905 } else { 1906 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1907 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq()); 1908 print_summary(1, "Update RS", body_summary->get_update_rs_seq()); 1909 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq()); 1910 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq()); 1911 } 1912 } 1913 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq()); 1914 print_summary(1, "Other", summary->get_other_seq()); 1915 { 1916 if (body_summary != NULL) { 1917 NumberSeq calc_other_times_ms; 1918 if (parallel) { 1919 // parallel 1920 NumberSeq* other_parts[] = { 1921 body_summary->get_root_region_scan_wait_seq(), 1922 body_summary->get_parallel_seq(), 1923 body_summary->get_clear_ct_seq() 1924 }; 1925 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 1926 3, other_parts); 1927 } else { 1928 // serial 1929 NumberSeq* other_parts[] = { 1930 body_summary->get_root_region_scan_wait_seq(), 1931 body_summary->get_update_rs_seq(), 1932 body_summary->get_ext_root_scan_seq(), 1933 body_summary->get_satb_filtering_seq(), 1934 body_summary->get_scan_rs_seq(), 1935 body_summary->get_obj_copy_seq() 1936 }; 1937 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 1938 6, other_parts); 1939 } 1940 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms); 1941 } 1942 } 1943 } else { 1944 LineBuffer(1).append_and_print_cr("none"); 1945 } 1946 LineBuffer(0).append_and_print_cr(""); 1947 } 1948 1949 void G1CollectorPolicy::print_tracing_info() const { 1950 if (TraceGen0Time) { 1951 gclog_or_tty->print_cr("ALL PAUSES"); 1952 print_summary_sd(0, "Total", _all_pause_times_ms); 1953 gclog_or_tty->print_cr(""); 1954 gclog_or_tty->print_cr(""); 1955 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 1956 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 1957 gclog_or_tty->print_cr(""); 1958 1959 gclog_or_tty->print_cr("EVACUATION PAUSES"); 1960 print_summary(_summary); 1961 1962 gclog_or_tty->print_cr("MISC"); 1963 print_summary_sd(0, "Stop World", _all_stop_world_times_ms); 1964 print_summary_sd(0, "Yields", _all_yield_times_ms); 1965 for (int i = 0; i < _aux_num; ++i) { 1966 if (_all_aux_times_ms[i].num() > 0) { 1967 char buffer[96]; 1968 sprintf(buffer, "Aux%d", i); 1969 print_summary_sd(0, buffer, &_all_aux_times_ms[i]); 1970 } 1971 } 1972 } 1973 if (TraceGen1Time) { 1974 if (_all_full_gc_times_ms->num() > 0) { 1975 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 1976 _all_full_gc_times_ms->num(), 1977 _all_full_gc_times_ms->sum() / 1000.0); 1978 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg()); 1979 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 1980 _all_full_gc_times_ms->sd(), 1981 _all_full_gc_times_ms->maximum()); 1982 } 1983 } 1984 } 1985 1986 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1987 #ifndef PRODUCT 1988 _short_lived_surv_rate_group->print_surv_rate_summary(); 1989 // add this call for any other surv rate groups 1990 #endif // PRODUCT 1991 } 1992 1993 #ifndef PRODUCT 1994 // for debugging, bit of a hack... 1995 static char* 1996 region_num_to_mbs(int length) { 1997 static char buffer[64]; 1998 double bytes = (double) (length * HeapRegion::GrainBytes); 1999 double mbs = bytes / (double) (1024 * 1024); 2000 sprintf(buffer, "%7.2lfMB", mbs); 2001 return buffer; 2002 } 2003 #endif // PRODUCT 2004 2005 uint G1CollectorPolicy::max_regions(int purpose) { 2006 switch (purpose) { 2007 case GCAllocForSurvived: 2008 return _max_survivor_regions; 2009 case GCAllocForTenured: 2010 return REGIONS_UNLIMITED; 2011 default: 2012 ShouldNotReachHere(); 2013 return REGIONS_UNLIMITED; 2014 }; 2015 } 2016 2017 void G1CollectorPolicy::update_max_gc_locker_expansion() { 2018 uint expansion_region_num = 0; 2019 if (GCLockerEdenExpansionPercent > 0) { 2020 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 2021 double expansion_region_num_d = perc * (double) _young_list_target_length; 2022 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 2023 // less than 1.0) we'll get 1. 2024 expansion_region_num = (uint) ceil(expansion_region_num_d); 2025 } else { 2026 assert(expansion_region_num == 0, "sanity"); 2027 } 2028 _young_list_max_length = _young_list_target_length + expansion_region_num; 2029 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 2030 } 2031 2032 // Calculates survivor space parameters. 2033 void G1CollectorPolicy::update_survivors_policy() { 2034 double max_survivor_regions_d = 2035 (double) _young_list_target_length / (double) SurvivorRatio; 2036 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 2037 // smaller than 1.0) we'll get 1. 2038 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 2039 2040 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 2041 HeapRegion::GrainWords * _max_survivor_regions); 2042 } 2043 2044 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 2045 GCCause::Cause gc_cause) { 2046 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2047 if (!during_cycle) { 2048 ergo_verbose1(ErgoConcCycles, 2049 "request concurrent cycle initiation", 2050 ergo_format_reason("requested by GC cause") 2051 ergo_format_str("GC cause"), 2052 GCCause::to_string(gc_cause)); 2053 set_initiate_conc_mark_if_possible(); 2054 return true; 2055 } else { 2056 ergo_verbose1(ErgoConcCycles, 2057 "do not request concurrent cycle initiation", 2058 ergo_format_reason("concurrent cycle already in progress") 2059 ergo_format_str("GC cause"), 2060 GCCause::to_string(gc_cause)); 2061 return false; 2062 } 2063 } 2064 2065 void 2066 G1CollectorPolicy::decide_on_conc_mark_initiation() { 2067 // We are about to decide on whether this pause will be an 2068 // initial-mark pause. 2069 2070 // First, during_initial_mark_pause() should not be already set. We 2071 // will set it here if we have to. However, it should be cleared by 2072 // the end of the pause (it's only set for the duration of an 2073 // initial-mark pause). 2074 assert(!during_initial_mark_pause(), "pre-condition"); 2075 2076 if (initiate_conc_mark_if_possible()) { 2077 // We had noticed on a previous pause that the heap occupancy has 2078 // gone over the initiating threshold and we should start a 2079 // concurrent marking cycle. So we might initiate one. 2080 2081 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2082 if (!during_cycle) { 2083 // The concurrent marking thread is not "during a cycle", i.e., 2084 // it has completed the last one. So we can go ahead and 2085 // initiate a new cycle. 2086 2087 set_during_initial_mark_pause(); 2088 // We do not allow mixed GCs during marking. 2089 if (!gcs_are_young()) { 2090 set_gcs_are_young(true); 2091 ergo_verbose0(ErgoMixedGCs, 2092 "end mixed GCs", 2093 ergo_format_reason("concurrent cycle is about to start")); 2094 } 2095 2096 // And we can now clear initiate_conc_mark_if_possible() as 2097 // we've already acted on it. 2098 clear_initiate_conc_mark_if_possible(); 2099 2100 ergo_verbose0(ErgoConcCycles, 2101 "initiate concurrent cycle", 2102 ergo_format_reason("concurrent cycle initiation requested")); 2103 } else { 2104 // The concurrent marking thread is still finishing up the 2105 // previous cycle. If we start one right now the two cycles 2106 // overlap. In particular, the concurrent marking thread might 2107 // be in the process of clearing the next marking bitmap (which 2108 // we will use for the next cycle if we start one). Starting a 2109 // cycle now will be bad given that parts of the marking 2110 // information might get cleared by the marking thread. And we 2111 // cannot wait for the marking thread to finish the cycle as it 2112 // periodically yields while clearing the next marking bitmap 2113 // and, if it's in a yield point, it's waiting for us to 2114 // finish. So, at this point we will not start a cycle and we'll 2115 // let the concurrent marking thread complete the last one. 2116 ergo_verbose0(ErgoConcCycles, 2117 "do not initiate concurrent cycle", 2118 ergo_format_reason("concurrent cycle already in progress")); 2119 } 2120 } 2121 } 2122 2123 class KnownGarbageClosure: public HeapRegionClosure { 2124 G1CollectedHeap* _g1h; 2125 CollectionSetChooser* _hrSorted; 2126 2127 public: 2128 KnownGarbageClosure(CollectionSetChooser* hrSorted) : 2129 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { } 2130 2131 bool doHeapRegion(HeapRegion* r) { 2132 // We only include humongous regions in collection 2133 // sets when concurrent mark shows that their contained object is 2134 // unreachable. 2135 2136 // Do we have any marking information for this region? 2137 if (r->is_marked()) { 2138 // We will skip any region that's currently used as an old GC 2139 // alloc region (we should not consider those for collection 2140 // before we fill them up). 2141 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 2142 _hrSorted->add_region(r); 2143 } 2144 } 2145 return false; 2146 } 2147 }; 2148 2149 class ParKnownGarbageHRClosure: public HeapRegionClosure { 2150 G1CollectedHeap* _g1h; 2151 CollectionSetChooser* _hrSorted; 2152 uint _marked_regions_added; 2153 size_t _reclaimable_bytes_added; 2154 uint _chunk_size; 2155 uint _cur_chunk_idx; 2156 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) 2157 2158 void get_new_chunk() { 2159 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size); 2160 _cur_chunk_end = _cur_chunk_idx + _chunk_size; 2161 } 2162 void add_region(HeapRegion* r) { 2163 if (_cur_chunk_idx == _cur_chunk_end) { 2164 get_new_chunk(); 2165 } 2166 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); 2167 _hrSorted->set_region(_cur_chunk_idx, r); 2168 _marked_regions_added++; 2169 _reclaimable_bytes_added += r->reclaimable_bytes(); 2170 _cur_chunk_idx++; 2171 } 2172 2173 public: 2174 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 2175 uint chunk_size) : 2176 _g1h(G1CollectedHeap::heap()), 2177 _hrSorted(hrSorted), _chunk_size(chunk_size), 2178 _marked_regions_added(0), _reclaimable_bytes_added(0), 2179 _cur_chunk_idx(0), _cur_chunk_end(0) { } 2180 2181 bool doHeapRegion(HeapRegion* r) { 2182 // Do we have any marking information for this region? 2183 if (r->is_marked()) { 2184 // We will skip any region that's currently used as an old GC 2185 // alloc region (we should not consider those for collection 2186 // before we fill them up). 2187 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 2188 add_region(r); 2189 } 2190 } 2191 return false; 2192 } 2193 uint marked_regions_added() { return _marked_regions_added; } 2194 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; } 2195 }; 2196 2197 class ParKnownGarbageTask: public AbstractGangTask { 2198 CollectionSetChooser* _hrSorted; 2199 uint _chunk_size; 2200 G1CollectedHeap* _g1; 2201 public: 2202 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) : 2203 AbstractGangTask("ParKnownGarbageTask"), 2204 _hrSorted(hrSorted), _chunk_size(chunk_size), 2205 _g1(G1CollectedHeap::heap()) { } 2206 2207 void work(uint worker_id) { 2208 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 2209 2210 // Back to zero for the claim value. 2211 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id, 2212 _g1->workers()->active_workers(), 2213 HeapRegion::InitialClaimValue); 2214 uint regions_added = parKnownGarbageCl.marked_regions_added(); 2215 size_t reclaimable_bytes_added = 2216 parKnownGarbageCl.reclaimable_bytes_added(); 2217 _hrSorted->update_totals(regions_added, reclaimable_bytes_added); 2218 } 2219 }; 2220 2221 void 2222 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) { 2223 _collectionSetChooser->clear(); 2224 2225 uint region_num = _g1->n_regions(); 2226 if (G1CollectedHeap::use_parallel_gc_threads()) { 2227 const uint OverpartitionFactor = 4; 2228 uint WorkUnit; 2229 // The use of MinChunkSize = 8 in the original code 2230 // causes some assertion failures when the total number of 2231 // region is less than 8. The code here tries to fix that. 2232 // Should the original code also be fixed? 2233 if (no_of_gc_threads > 0) { 2234 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U); 2235 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor), 2236 MinWorkUnit); 2237 } else { 2238 assert(no_of_gc_threads > 0, 2239 "The active gc workers should be greater than 0"); 2240 // In a product build do something reasonable to avoid a crash. 2241 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U); 2242 WorkUnit = 2243 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor), 2244 MinWorkUnit); 2245 } 2246 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(), 2247 WorkUnit); 2248 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, 2249 (int) WorkUnit); 2250 _g1->workers()->run_task(&parKnownGarbageTask); 2251 2252 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 2253 "sanity check"); 2254 } else { 2255 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); 2256 _g1->heap_region_iterate(&knownGarbagecl); 2257 } 2258 2259 _collectionSetChooser->sort_regions(); 2260 2261 double end_sec = os::elapsedTime(); 2262 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 2263 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 2264 _cur_mark_stop_world_time_ms += elapsed_time_ms; 2265 _prev_collection_pause_end_ms += elapsed_time_ms; 2266 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true); 2267 } 2268 2269 // Add the heap region at the head of the non-incremental collection set 2270 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 2271 assert(_inc_cset_build_state == Active, "Precondition"); 2272 assert(!hr->is_young(), "non-incremental add of young region"); 2273 2274 assert(!hr->in_collection_set(), "should not already be in the CSet"); 2275 hr->set_in_collection_set(true); 2276 hr->set_next_in_collection_set(_collection_set); 2277 _collection_set = hr; 2278 _collection_set_bytes_used_before += hr->used(); 2279 _g1->register_region_with_in_cset_fast_test(hr); 2280 size_t rs_length = hr->rem_set()->occupied(); 2281 _recorded_rs_lengths += rs_length; 2282 _old_cset_region_length += 1; 2283 } 2284 2285 // Initialize the per-collection-set information 2286 void G1CollectorPolicy::start_incremental_cset_building() { 2287 assert(_inc_cset_build_state == Inactive, "Precondition"); 2288 2289 _inc_cset_head = NULL; 2290 _inc_cset_tail = NULL; 2291 _inc_cset_bytes_used_before = 0; 2292 2293 _inc_cset_max_finger = 0; 2294 _inc_cset_recorded_rs_lengths = 0; 2295 _inc_cset_recorded_rs_lengths_diffs = 0; 2296 _inc_cset_predicted_elapsed_time_ms = 0.0; 2297 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2298 _inc_cset_build_state = Active; 2299 } 2300 2301 void G1CollectorPolicy::finalize_incremental_cset_building() { 2302 assert(_inc_cset_build_state == Active, "Precondition"); 2303 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2304 2305 // The two "main" fields, _inc_cset_recorded_rs_lengths and 2306 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 2307 // that adds a new region to the CSet. Further updates by the 2308 // concurrent refinement thread that samples the young RSet lengths 2309 // are accumulated in the *_diffs fields. Here we add the diffs to 2310 // the "main" fields. 2311 2312 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 2313 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 2314 } else { 2315 // This is defensive. The diff should in theory be always positive 2316 // as RSets can only grow between GCs. However, given that we 2317 // sample their size concurrently with other threads updating them 2318 // it's possible that we might get the wrong size back, which 2319 // could make the calculations somewhat inaccurate. 2320 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 2321 if (_inc_cset_recorded_rs_lengths >= diffs) { 2322 _inc_cset_recorded_rs_lengths -= diffs; 2323 } else { 2324 _inc_cset_recorded_rs_lengths = 0; 2325 } 2326 } 2327 _inc_cset_predicted_elapsed_time_ms += 2328 _inc_cset_predicted_elapsed_time_ms_diffs; 2329 2330 _inc_cset_recorded_rs_lengths_diffs = 0; 2331 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2332 } 2333 2334 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 2335 // This routine is used when: 2336 // * adding survivor regions to the incremental cset at the end of an 2337 // evacuation pause, 2338 // * adding the current allocation region to the incremental cset 2339 // when it is retired, and 2340 // * updating existing policy information for a region in the 2341 // incremental cset via young list RSet sampling. 2342 // Therefore this routine may be called at a safepoint by the 2343 // VM thread, or in-between safepoints by mutator threads (when 2344 // retiring the current allocation region) or a concurrent 2345 // refine thread (RSet sampling). 2346 2347 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2348 size_t used_bytes = hr->used(); 2349 _inc_cset_recorded_rs_lengths += rs_length; 2350 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 2351 _inc_cset_bytes_used_before += used_bytes; 2352 2353 // Cache the values we have added to the aggregated informtion 2354 // in the heap region in case we have to remove this region from 2355 // the incremental collection set, or it is updated by the 2356 // rset sampling code 2357 hr->set_recorded_rs_length(rs_length); 2358 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 2359 } 2360 2361 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 2362 size_t new_rs_length) { 2363 // Update the CSet information that is dependent on the new RS length 2364 assert(hr->is_young(), "Precondition"); 2365 assert(!SafepointSynchronize::is_at_safepoint(), 2366 "should not be at a safepoint"); 2367 2368 // We could have updated _inc_cset_recorded_rs_lengths and 2369 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 2370 // that atomically, as this code is executed by a concurrent 2371 // refinement thread, potentially concurrently with a mutator thread 2372 // allocating a new region and also updating the same fields. To 2373 // avoid the atomic operations we accumulate these updates on two 2374 // separate fields (*_diffs) and we'll just add them to the "main" 2375 // fields at the start of a GC. 2376 2377 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 2378 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 2379 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 2380 2381 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 2382 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2383 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 2384 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 2385 2386 hr->set_recorded_rs_length(new_rs_length); 2387 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 2388 } 2389 2390 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 2391 assert(hr->is_young(), "invariant"); 2392 assert(hr->young_index_in_cset() > -1, "should have already been set"); 2393 assert(_inc_cset_build_state == Active, "Precondition"); 2394 2395 // We need to clear and set the cached recorded/cached collection set 2396 // information in the heap region here (before the region gets added 2397 // to the collection set). An individual heap region's cached values 2398 // are calculated, aggregated with the policy collection set info, 2399 // and cached in the heap region here (initially) and (subsequently) 2400 // by the Young List sampling code. 2401 2402 size_t rs_length = hr->rem_set()->occupied(); 2403 add_to_incremental_cset_info(hr, rs_length); 2404 2405 HeapWord* hr_end = hr->end(); 2406 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 2407 2408 assert(!hr->in_collection_set(), "invariant"); 2409 hr->set_in_collection_set(true); 2410 assert( hr->next_in_collection_set() == NULL, "invariant"); 2411 2412 _g1->register_region_with_in_cset_fast_test(hr); 2413 } 2414 2415 // Add the region at the RHS of the incremental cset 2416 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 2417 // We should only ever be appending survivors at the end of a pause 2418 assert( hr->is_survivor(), "Logic"); 2419 2420 // Do the 'common' stuff 2421 add_region_to_incremental_cset_common(hr); 2422 2423 // Now add the region at the right hand side 2424 if (_inc_cset_tail == NULL) { 2425 assert(_inc_cset_head == NULL, "invariant"); 2426 _inc_cset_head = hr; 2427 } else { 2428 _inc_cset_tail->set_next_in_collection_set(hr); 2429 } 2430 _inc_cset_tail = hr; 2431 } 2432 2433 // Add the region to the LHS of the incremental cset 2434 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2435 // Survivors should be added to the RHS at the end of a pause 2436 assert(!hr->is_survivor(), "Logic"); 2437 2438 // Do the 'common' stuff 2439 add_region_to_incremental_cset_common(hr); 2440 2441 // Add the region at the left hand side 2442 hr->set_next_in_collection_set(_inc_cset_head); 2443 if (_inc_cset_head == NULL) { 2444 assert(_inc_cset_tail == NULL, "Invariant"); 2445 _inc_cset_tail = hr; 2446 } 2447 _inc_cset_head = hr; 2448 } 2449 2450 #ifndef PRODUCT 2451 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2452 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2453 2454 st->print_cr("\nCollection_set:"); 2455 HeapRegion* csr = list_head; 2456 while (csr != NULL) { 2457 HeapRegion* next = csr->next_in_collection_set(); 2458 assert(csr->in_collection_set(), "bad CS"); 2459 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 2460 HR_FORMAT_PARAMS(csr), 2461 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(), 2462 csr->age_in_surv_rate_group_cond()); 2463 csr = next; 2464 } 2465 } 2466 #endif // !PRODUCT 2467 2468 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2469 const char* false_action_str) { 2470 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2471 if (cset_chooser->is_empty()) { 2472 ergo_verbose0(ErgoMixedGCs, 2473 false_action_str, 2474 ergo_format_reason("candidate old regions not available")); 2475 return false; 2476 } 2477 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2478 size_t capacity_bytes = _g1->capacity(); 2479 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2480 double threshold = (double) G1HeapWastePercent; 2481 if (perc < threshold) { 2482 ergo_verbose4(ErgoMixedGCs, 2483 false_action_str, 2484 ergo_format_reason("reclaimable percentage lower than threshold") 2485 ergo_format_region("candidate old regions") 2486 ergo_format_byte_perc("reclaimable") 2487 ergo_format_perc("threshold"), 2488 cset_chooser->remaining_regions(), 2489 reclaimable_bytes, perc, threshold); 2490 return false; 2491 } 2492 2493 ergo_verbose4(ErgoMixedGCs, 2494 true_action_str, 2495 ergo_format_reason("candidate old regions available") 2496 ergo_format_region("candidate old regions") 2497 ergo_format_byte_perc("reclaimable") 2498 ergo_format_perc("threshold"), 2499 cset_chooser->remaining_regions(), 2500 reclaimable_bytes, perc, threshold); 2501 return true; 2502 } 2503 2504 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) { 2505 // Set this here - in case we're not doing young collections. 2506 double non_young_start_time_sec = os::elapsedTime(); 2507 2508 YoungList* young_list = _g1->young_list(); 2509 finalize_incremental_cset_building(); 2510 2511 guarantee(target_pause_time_ms > 0.0, 2512 err_msg("target_pause_time_ms = %1.6lf should be positive", 2513 target_pause_time_ms)); 2514 guarantee(_collection_set == NULL, "Precondition"); 2515 2516 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2517 double predicted_pause_time_ms = base_time_ms; 2518 double time_remaining_ms = target_pause_time_ms - base_time_ms; 2519 2520 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2521 "start choosing CSet", 2522 ergo_format_ms("predicted base time") 2523 ergo_format_ms("remaining time") 2524 ergo_format_ms("target pause time"), 2525 base_time_ms, time_remaining_ms, target_pause_time_ms); 2526 2527 HeapRegion* hr; 2528 double young_start_time_sec = os::elapsedTime(); 2529 2530 _collection_set_bytes_used_before = 0; 2531 _last_gc_was_young = gcs_are_young() ? true : false; 2532 2533 if (_last_gc_was_young) { 2534 ++_young_pause_num; 2535 } else { 2536 ++_mixed_pause_num; 2537 } 2538 2539 // The young list is laid with the survivor regions from the previous 2540 // pause are appended to the RHS of the young list, i.e. 2541 // [Newly Young Regions ++ Survivors from last pause]. 2542 2543 uint survivor_region_length = young_list->survivor_length(); 2544 uint eden_region_length = young_list->length() - survivor_region_length; 2545 init_cset_region_lengths(eden_region_length, survivor_region_length); 2546 hr = young_list->first_survivor_region(); 2547 while (hr != NULL) { 2548 assert(hr->is_survivor(), "badly formed young list"); 2549 hr->set_young(); 2550 hr = hr->get_next_young_region(); 2551 } 2552 2553 // Clear the fields that point to the survivor list - they are all young now. 2554 young_list->clear_survivors(); 2555 2556 _collection_set = _inc_cset_head; 2557 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2558 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms; 2559 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; 2560 2561 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2562 "add young regions to CSet", 2563 ergo_format_region("eden") 2564 ergo_format_region("survivors") 2565 ergo_format_ms("predicted young region time"), 2566 eden_region_length, survivor_region_length, 2567 _inc_cset_predicted_elapsed_time_ms); 2568 2569 // The number of recorded young regions is the incremental 2570 // collection set's current size 2571 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2572 2573 double young_end_time_sec = os::elapsedTime(); 2574 _recorded_young_cset_choice_time_ms = 2575 (young_end_time_sec - young_start_time_sec) * 1000.0; 2576 2577 // We are doing young collections so reset this. 2578 non_young_start_time_sec = young_end_time_sec; 2579 2580 if (!gcs_are_young()) { 2581 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2582 cset_chooser->verify(); 2583 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length(); 2584 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length(); 2585 2586 uint expensive_region_num = 0; 2587 bool check_time_remaining = adaptive_young_list_length(); 2588 HeapRegion* hr = cset_chooser->peek(); 2589 while (hr != NULL) { 2590 if (old_cset_region_length() >= max_old_cset_length) { 2591 // Added maximum number of old regions to the CSet. 2592 ergo_verbose2(ErgoCSetConstruction, 2593 "finish adding old regions to CSet", 2594 ergo_format_reason("old CSet region num reached max") 2595 ergo_format_region("old") 2596 ergo_format_region("max"), 2597 old_cset_region_length(), max_old_cset_length); 2598 break; 2599 } 2600 2601 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); 2602 if (check_time_remaining) { 2603 if (predicted_time_ms > time_remaining_ms) { 2604 // Too expensive for the current CSet. 2605 2606 if (old_cset_region_length() >= min_old_cset_length) { 2607 // We have added the minimum number of old regions to the CSet, 2608 // we are done with this CSet. 2609 ergo_verbose4(ErgoCSetConstruction, 2610 "finish adding old regions to CSet", 2611 ergo_format_reason("predicted time is too high") 2612 ergo_format_ms("predicted time") 2613 ergo_format_ms("remaining time") 2614 ergo_format_region("old") 2615 ergo_format_region("min"), 2616 predicted_time_ms, time_remaining_ms, 2617 old_cset_region_length(), min_old_cset_length); 2618 break; 2619 } 2620 2621 // We'll add it anyway given that we haven't reached the 2622 // minimum number of old regions. 2623 expensive_region_num += 1; 2624 } 2625 } else { 2626 if (old_cset_region_length() >= min_old_cset_length) { 2627 // In the non-auto-tuning case, we'll finish adding regions 2628 // to the CSet if we reach the minimum. 2629 ergo_verbose2(ErgoCSetConstruction, 2630 "finish adding old regions to CSet", 2631 ergo_format_reason("old CSet region num reached min") 2632 ergo_format_region("old") 2633 ergo_format_region("min"), 2634 old_cset_region_length(), min_old_cset_length); 2635 break; 2636 } 2637 } 2638 2639 // We will add this region to the CSet. 2640 time_remaining_ms -= predicted_time_ms; 2641 predicted_pause_time_ms += predicted_time_ms; 2642 cset_chooser->remove_and_move_to_next(hr); 2643 _g1->old_set_remove(hr); 2644 add_old_region_to_cset(hr); 2645 2646 hr = cset_chooser->peek(); 2647 } 2648 if (hr == NULL) { 2649 ergo_verbose0(ErgoCSetConstruction, 2650 "finish adding old regions to CSet", 2651 ergo_format_reason("candidate old regions not available")); 2652 } 2653 2654 if (expensive_region_num > 0) { 2655 // We print the information once here at the end, predicated on 2656 // whether we added any apparently expensive regions or not, to 2657 // avoid generating output per region. 2658 ergo_verbose4(ErgoCSetConstruction, 2659 "added expensive regions to CSet", 2660 ergo_format_reason("old CSet region num not reached min") 2661 ergo_format_region("old") 2662 ergo_format_region("expensive") 2663 ergo_format_region("min") 2664 ergo_format_ms("remaining time"), 2665 old_cset_region_length(), 2666 expensive_region_num, 2667 min_old_cset_length, 2668 time_remaining_ms); 2669 } 2670 2671 cset_chooser->verify(); 2672 } 2673 2674 stop_incremental_cset_building(); 2675 2676 count_CS_bytes_used(); 2677 2678 ergo_verbose5(ErgoCSetConstruction, 2679 "finish choosing CSet", 2680 ergo_format_region("eden") 2681 ergo_format_region("survivors") 2682 ergo_format_region("old") 2683 ergo_format_ms("predicted pause time") 2684 ergo_format_ms("target pause time"), 2685 eden_region_length, survivor_region_length, 2686 old_cset_region_length(), 2687 predicted_pause_time_ms, target_pause_time_ms); 2688 2689 double non_young_end_time_sec = os::elapsedTime(); 2690 _recorded_non_young_cset_choice_time_ms = 2691 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; 2692 }