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->print("[%s", GCCauseString("GC pause", _g1->gc_cause()) 889 .append(gcs_are_young() ? " (young)" : " (mixed)")); 890 } 891 892 // We only need to do this here as the policy will only be applied 893 // to the GC we're about to start. so, no point is calculating this 894 // every time we calculate / recalculate the target young length. 895 update_survivors_policy(); 896 897 assert(_g1->used() == _g1->recalculate_used(), 898 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT, 899 _g1->used(), _g1->recalculate_used())); 900 901 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 902 _all_stop_world_times_ms->add(s_w_t_ms); 903 _stop_world_start = 0.0; 904 905 _cur_collection_start_sec = start_time_sec; 906 _cur_collection_pause_used_at_start_bytes = start_used; 907 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); 908 _pending_cards = _g1->pending_card_num(); 909 _max_pending_cards = _g1->max_pending_card_num(); 910 911 _bytes_in_collection_set_before_gc = 0; 912 _bytes_copied_during_gc = 0; 913 914 YoungList* young_list = _g1->young_list(); 915 _eden_bytes_before_gc = young_list->eden_used_bytes(); 916 _survivor_bytes_before_gc = young_list->survivor_used_bytes(); 917 _capacity_before_gc = _g1->capacity(); 918 919 #ifdef DEBUG 920 // initialise these to something well known so that we can spot 921 // if they are not set properly 922 923 for (int i = 0; i < _parallel_gc_threads; ++i) { 924 _par_last_gc_worker_start_times_ms[i] = -1234.0; 925 _par_last_ext_root_scan_times_ms[i] = -1234.0; 926 _par_last_satb_filtering_times_ms[i] = -1234.0; 927 _par_last_update_rs_times_ms[i] = -1234.0; 928 _par_last_update_rs_processed_buffers[i] = -1234.0; 929 _par_last_scan_rs_times_ms[i] = -1234.0; 930 _par_last_obj_copy_times_ms[i] = -1234.0; 931 _par_last_termination_times_ms[i] = -1234.0; 932 _par_last_termination_attempts[i] = -1234.0; 933 _par_last_gc_worker_end_times_ms[i] = -1234.0; 934 _par_last_gc_worker_times_ms[i] = -1234.0; 935 _par_last_gc_worker_other_times_ms[i] = -1234.0; 936 } 937 #endif 938 939 for (int i = 0; i < _aux_num; ++i) { 940 _cur_aux_times_ms[i] = 0.0; 941 _cur_aux_times_set[i] = false; 942 } 943 944 // This is initialized to zero here and is set during the evacuation 945 // pause if we actually waited for the root region scanning to finish. 946 _root_region_scan_wait_time_ms = 0.0; 947 948 _last_gc_was_young = false; 949 950 // do that for any other surv rate groups 951 _short_lived_surv_rate_group->stop_adding_regions(); 952 _survivors_age_table.clear(); 953 954 assert( verify_young_ages(), "region age verification" ); 955 } 956 957 void G1CollectorPolicy::record_concurrent_mark_init_end(double 958 mark_init_elapsed_time_ms) { 959 _during_marking = true; 960 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); 961 clear_during_initial_mark_pause(); 962 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 963 } 964 965 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 966 _mark_remark_start_sec = os::elapsedTime(); 967 _during_marking = false; 968 } 969 970 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 971 double end_time_sec = os::elapsedTime(); 972 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 973 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 974 _cur_mark_stop_world_time_ms += elapsed_time_ms; 975 _prev_collection_pause_end_ms += elapsed_time_ms; 976 977 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); 978 } 979 980 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 981 _mark_cleanup_start_sec = os::elapsedTime(); 982 } 983 984 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 985 _last_young_gc = true; 986 _in_marking_window = false; 987 } 988 989 void G1CollectorPolicy::record_concurrent_pause() { 990 if (_stop_world_start > 0.0) { 991 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 992 _all_yield_times_ms->add(yield_ms); 993 } 994 } 995 996 void G1CollectorPolicy::record_concurrent_pause_end() { 997 } 998 999 template<class T> 1000 T sum_of(T* sum_arr, int start, int n, int N) { 1001 T sum = (T)0; 1002 for (int i = 0; i < n; i++) { 1003 int j = (start + i) % N; 1004 sum += sum_arr[j]; 1005 } 1006 return sum; 1007 } 1008 1009 void G1CollectorPolicy::print_par_stats(int level, 1010 const char* str, 1011 double* data) { 1012 double min = data[0], max = data[0]; 1013 double total = 0.0; 1014 LineBuffer buf(level); 1015 buf.append("[%s (ms):", str); 1016 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1017 double val = data[i]; 1018 if (val < min) 1019 min = val; 1020 if (val > max) 1021 max = val; 1022 total += val; 1023 if (G1Log::finest()) { 1024 buf.append(" %.1lf", val); 1025 } 1026 } 1027 1028 if (G1Log::finest()) { 1029 buf.append_and_print_cr(""); 1030 } 1031 double avg = total / (double) no_of_gc_threads(); 1032 buf.append_and_print_cr(" Avg: %.1lf Min: %.1lf Max: %.1lf Diff: %.1lf]", 1033 avg, min, max, max - min); 1034 } 1035 1036 void G1CollectorPolicy::print_par_sizes(int level, 1037 const char* str, 1038 double* data) { 1039 double min = data[0], max = data[0]; 1040 double total = 0.0; 1041 LineBuffer buf(level); 1042 buf.append("[%s :", str); 1043 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1044 double val = data[i]; 1045 if (val < min) 1046 min = val; 1047 if (val > max) 1048 max = val; 1049 total += val; 1050 buf.append(" %d", (int) val); 1051 } 1052 buf.append_and_print_cr(""); 1053 double avg = total / (double) no_of_gc_threads(); 1054 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]", 1055 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min); 1056 } 1057 1058 void G1CollectorPolicy::print_stats(int level, 1059 const char* str, 1060 double value) { 1061 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value); 1062 } 1063 1064 void G1CollectorPolicy::print_stats(int level, 1065 const char* str, 1066 int value) { 1067 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value); 1068 } 1069 1070 double G1CollectorPolicy::avg_value(double* data) { 1071 if (G1CollectedHeap::use_parallel_gc_threads()) { 1072 double ret = 0.0; 1073 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1074 ret += data[i]; 1075 } 1076 return ret / (double) no_of_gc_threads(); 1077 } else { 1078 return data[0]; 1079 } 1080 } 1081 1082 double G1CollectorPolicy::max_value(double* data) { 1083 if (G1CollectedHeap::use_parallel_gc_threads()) { 1084 double ret = data[0]; 1085 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1086 if (data[i] > ret) { 1087 ret = data[i]; 1088 } 1089 } 1090 return ret; 1091 } else { 1092 return data[0]; 1093 } 1094 } 1095 1096 double G1CollectorPolicy::sum_of_values(double* data) { 1097 if (G1CollectedHeap::use_parallel_gc_threads()) { 1098 double sum = 0.0; 1099 for (uint i = 0; i < no_of_gc_threads(); i++) { 1100 sum += data[i]; 1101 } 1102 return sum; 1103 } else { 1104 return data[0]; 1105 } 1106 } 1107 1108 double G1CollectorPolicy::max_sum(double* data1, double* data2) { 1109 double ret = data1[0] + data2[0]; 1110 1111 if (G1CollectedHeap::use_parallel_gc_threads()) { 1112 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1113 double data = data1[i] + data2[i]; 1114 if (data > ret) { 1115 ret = data; 1116 } 1117 } 1118 } 1119 return ret; 1120 } 1121 1122 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 1123 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 1124 return false; 1125 } 1126 1127 size_t marking_initiating_used_threshold = 1128 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 1129 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 1130 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 1131 1132 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 1133 if (gcs_are_young()) { 1134 ergo_verbose5(ErgoConcCycles, 1135 "request concurrent cycle initiation", 1136 ergo_format_reason("occupancy higher than threshold") 1137 ergo_format_byte("occupancy") 1138 ergo_format_byte("allocation request") 1139 ergo_format_byte_perc("threshold") 1140 ergo_format_str("source"), 1141 cur_used_bytes, 1142 alloc_byte_size, 1143 marking_initiating_used_threshold, 1144 (double) InitiatingHeapOccupancyPercent, 1145 source); 1146 return true; 1147 } else { 1148 ergo_verbose5(ErgoConcCycles, 1149 "do not request concurrent cycle initiation", 1150 ergo_format_reason("still doing mixed collections") 1151 ergo_format_byte("occupancy") 1152 ergo_format_byte("allocation request") 1153 ergo_format_byte_perc("threshold") 1154 ergo_format_str("source"), 1155 cur_used_bytes, 1156 alloc_byte_size, 1157 marking_initiating_used_threshold, 1158 (double) InitiatingHeapOccupancyPercent, 1159 source); 1160 } 1161 } 1162 1163 return false; 1164 } 1165 1166 // Anything below that is considered to be zero 1167 #define MIN_TIMER_GRANULARITY 0.0000001 1168 1169 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) { 1170 double end_time_sec = os::elapsedTime(); 1171 double elapsed_ms = _last_pause_time_ms; 1172 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1173 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1174 "otherwise, the subtraction below does not make sense"); 1175 size_t rs_size = 1176 _cur_collection_pause_used_regions_at_start - cset_region_length(); 1177 size_t cur_used_bytes = _g1->used(); 1178 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1179 bool last_pause_included_initial_mark = false; 1180 bool update_stats = !_g1->evacuation_failed(); 1181 set_no_of_gc_threads(no_of_gc_threads); 1182 1183 #ifndef PRODUCT 1184 if (G1YoungSurvRateVerbose) { 1185 gclog_or_tty->print_cr(""); 1186 _short_lived_surv_rate_group->print(); 1187 // do that for any other surv rate groups too 1188 } 1189 #endif // PRODUCT 1190 1191 last_pause_included_initial_mark = during_initial_mark_pause(); 1192 if (last_pause_included_initial_mark) { 1193 record_concurrent_mark_init_end(0.0); 1194 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) { 1195 // Note: this might have already been set, if during the last 1196 // pause we decided to start a cycle but at the beginning of 1197 // this pause we decided to postpone it. That's OK. 1198 set_initiate_conc_mark_if_possible(); 1199 } 1200 1201 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, 1202 end_time_sec, false); 1203 1204 // This assert is exempted when we're doing parallel collection pauses, 1205 // because the fragmentation caused by the parallel GC allocation buffers 1206 // can lead to more memory being used during collection than was used 1207 // before. Best leave this out until the fragmentation problem is fixed. 1208 // Pauses in which evacuation failed can also lead to negative 1209 // collections, since no space is reclaimed from a region containing an 1210 // object whose evacuation failed. 1211 // Further, we're now always doing parallel collection. But I'm still 1212 // leaving this here as a placeholder for a more precise assertion later. 1213 // (DLD, 10/05.) 1214 assert((true || parallel) // Always using GC LABs now. 1215 || _g1->evacuation_failed() 1216 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes, 1217 "Negative collection"); 1218 1219 size_t freed_bytes = 1220 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; 1221 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; 1222 1223 double survival_fraction = 1224 (double)surviving_bytes/ 1225 (double)_collection_set_bytes_used_before; 1226 1227 // These values are used to update the summary information that is 1228 // displayed when TraceGen0Time is enabled, and are output as part 1229 // of the "finer" output, in the non-parallel case. 1230 1231 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); 1232 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms); 1233 double update_rs_time = avg_value(_par_last_update_rs_times_ms); 1234 double update_rs_processed_buffers = 1235 sum_of_values(_par_last_update_rs_processed_buffers); 1236 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); 1237 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); 1238 double termination_time = avg_value(_par_last_termination_times_ms); 1239 1240 double known_time = ext_root_scan_time + 1241 satb_filtering_time + 1242 update_rs_time + 1243 scan_rs_time + 1244 obj_copy_time; 1245 1246 double other_time_ms = elapsed_ms; 1247 1248 // Subtract the root region scanning wait time. It's initialized to 1249 // zero at the start of the pause. 1250 other_time_ms -= _root_region_scan_wait_time_ms; 1251 1252 if (parallel) { 1253 other_time_ms -= _cur_collection_par_time_ms; 1254 } else { 1255 other_time_ms -= known_time; 1256 } 1257 1258 // Now subtract the time taken to fix up roots in generated code 1259 other_time_ms -= _cur_collection_code_root_fixup_time_ms; 1260 1261 // Subtract the time taken to clean the card table from the 1262 // current value of "other time" 1263 other_time_ms -= _cur_clear_ct_time_ms; 1264 1265 // TraceGen0Time and TraceGen1Time summary info updating. 1266 _all_pause_times_ms->add(elapsed_ms); 1267 1268 if (update_stats) { 1269 _summary->record_total_time_ms(elapsed_ms); 1270 _summary->record_other_time_ms(other_time_ms); 1271 1272 MainBodySummary* body_summary = _summary->main_body_summary(); 1273 assert(body_summary != NULL, "should not be null!"); 1274 1275 body_summary->record_root_region_scan_wait_time_ms( 1276 _root_region_scan_wait_time_ms); 1277 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time); 1278 body_summary->record_satb_filtering_time_ms(satb_filtering_time); 1279 body_summary->record_update_rs_time_ms(update_rs_time); 1280 body_summary->record_scan_rs_time_ms(scan_rs_time); 1281 body_summary->record_obj_copy_time_ms(obj_copy_time); 1282 1283 if (parallel) { 1284 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms); 1285 body_summary->record_termination_time_ms(termination_time); 1286 1287 double parallel_known_time = known_time + termination_time; 1288 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time; 1289 body_summary->record_parallel_other_time_ms(parallel_other_time); 1290 } 1291 1292 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms); 1293 1294 // We exempt parallel collection from this check because Alloc Buffer 1295 // fragmentation can produce negative collections. Same with evac 1296 // failure. 1297 // Further, we're now always doing parallel collection. But I'm still 1298 // leaving this here as a placeholder for a more precise assertion later. 1299 // (DLD, 10/05. 1300 assert((true || parallel) 1301 || _g1->evacuation_failed() 1302 || surviving_bytes <= _collection_set_bytes_used_before, 1303 "Or else negative collection!"); 1304 1305 // this is where we update the allocation rate of the application 1306 double app_time_ms = 1307 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); 1308 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1309 // This usually happens due to the timer not having the required 1310 // granularity. Some Linuxes are the usual culprits. 1311 // We'll just set it to something (arbitrarily) small. 1312 app_time_ms = 1.0; 1313 } 1314 // We maintain the invariant that all objects allocated by mutator 1315 // threads will be allocated out of eden regions. So, we can use 1316 // the eden region number allocated since the previous GC to 1317 // calculate the application's allocate rate. The only exception 1318 // to that is humongous objects that are allocated separately. But 1319 // given that humongous object allocations do not really affect 1320 // either the pause's duration nor when the next pause will take 1321 // place we can safely ignore them here. 1322 uint regions_allocated = eden_cset_region_length(); 1323 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1324 _alloc_rate_ms_seq->add(alloc_rate_ms); 1325 1326 double interval_ms = 1327 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1328 update_recent_gc_times(end_time_sec, elapsed_ms); 1329 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1330 if (recent_avg_pause_time_ratio() < 0.0 || 1331 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1332 #ifndef PRODUCT 1333 // Dump info to allow post-facto debugging 1334 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1335 gclog_or_tty->print_cr("-------------------------------------------"); 1336 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1337 _recent_gc_times_ms->dump(); 1338 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1339 _recent_prev_end_times_for_all_gcs_sec->dump(); 1340 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1341 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1342 // In debug mode, terminate the JVM if the user wants to debug at this point. 1343 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1344 #endif // !PRODUCT 1345 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1346 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1347 if (_recent_avg_pause_time_ratio < 0.0) { 1348 _recent_avg_pause_time_ratio = 0.0; 1349 } else { 1350 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1351 _recent_avg_pause_time_ratio = 1.0; 1352 } 1353 } 1354 } 1355 1356 for (int i = 0; i < _aux_num; ++i) { 1357 if (_cur_aux_times_set[i]) { 1358 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]); 1359 } 1360 } 1361 1362 if (G1Log::finer()) { 1363 bool print_marking_info = 1364 _g1->mark_in_progress() && !last_pause_included_initial_mark; 1365 1366 gclog_or_tty->print_cr("%s, %1.8lf secs]", 1367 (last_pause_included_initial_mark) ? " (initial-mark)" : "", 1368 elapsed_ms / 1000.0); 1369 1370 if (_root_region_scan_wait_time_ms > 0.0) { 1371 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms); 1372 } 1373 if (parallel) { 1374 print_stats(1, "Parallel Time", _cur_collection_par_time_ms); 1375 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms); 1376 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms); 1377 if (print_marking_info) { 1378 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms); 1379 } 1380 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); 1381 if (G1Log::finest()) { 1382 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers); 1383 } 1384 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); 1385 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); 1386 print_par_stats(2, "Termination", _par_last_termination_times_ms); 1387 if (G1Log::finest()) { 1388 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts); 1389 } 1390 1391 for (int i = 0; i < _parallel_gc_threads; i++) { 1392 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - 1393 _par_last_gc_worker_start_times_ms[i]; 1394 1395 double worker_known_time = _par_last_ext_root_scan_times_ms[i] + 1396 _par_last_satb_filtering_times_ms[i] + 1397 _par_last_update_rs_times_ms[i] + 1398 _par_last_scan_rs_times_ms[i] + 1399 _par_last_obj_copy_times_ms[i] + 1400 _par_last_termination_times_ms[i]; 1401 1402 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] - 1403 worker_known_time; 1404 } 1405 1406 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms); 1407 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms); 1408 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms); 1409 } else { 1410 print_stats(1, "Ext Root Scanning", ext_root_scan_time); 1411 if (print_marking_info) { 1412 print_stats(1, "SATB Filtering", satb_filtering_time); 1413 } 1414 print_stats(1, "Update RS", update_rs_time); 1415 if (G1Log::finest()) { 1416 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers); 1417 } 1418 print_stats(1, "Scan RS", scan_rs_time); 1419 print_stats(1, "Object Copying", obj_copy_time); 1420 } 1421 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms); 1422 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); 1423 #ifndef PRODUCT 1424 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms); 1425 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms); 1426 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms); 1427 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms); 1428 if (_num_cc_clears > 0) { 1429 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears)); 1430 } 1431 #endif 1432 print_stats(1, "Other", other_time_ms); 1433 print_stats(2, "Choose CSet", 1434 (_recorded_young_cset_choice_time_ms + 1435 _recorded_non_young_cset_choice_time_ms)); 1436 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms); 1437 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms); 1438 print_stats(2, "Free CSet", 1439 (_recorded_young_free_cset_time_ms + 1440 _recorded_non_young_free_cset_time_ms)); 1441 1442 for (int i = 0; i < _aux_num; ++i) { 1443 if (_cur_aux_times_set[i]) { 1444 char buffer[96]; 1445 sprintf(buffer, "Aux%d", i); 1446 print_stats(1, buffer, _cur_aux_times_ms[i]); 1447 } 1448 } 1449 } 1450 1451 bool new_in_marking_window = _in_marking_window; 1452 bool new_in_marking_window_im = false; 1453 if (during_initial_mark_pause()) { 1454 new_in_marking_window = true; 1455 new_in_marking_window_im = true; 1456 } 1457 1458 if (_last_young_gc) { 1459 // This is supposed to to be the "last young GC" before we start 1460 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1461 1462 if (!last_pause_included_initial_mark) { 1463 if (next_gc_should_be_mixed("start mixed GCs", 1464 "do not start mixed GCs")) { 1465 set_gcs_are_young(false); 1466 } 1467 } else { 1468 ergo_verbose0(ErgoMixedGCs, 1469 "do not start mixed GCs", 1470 ergo_format_reason("concurrent cycle is about to start")); 1471 } 1472 _last_young_gc = false; 1473 } 1474 1475 if (!_last_gc_was_young) { 1476 // This is a mixed GC. Here we decide whether to continue doing 1477 // mixed GCs or not. 1478 1479 if (!next_gc_should_be_mixed("continue mixed GCs", 1480 "do not continue mixed GCs")) { 1481 set_gcs_are_young(true); 1482 } 1483 } 1484 1485 _short_lived_surv_rate_group->start_adding_regions(); 1486 // do that for any other surv rate groupsx 1487 1488 if (update_stats) { 1489 double pause_time_ms = elapsed_ms; 1490 1491 size_t diff = 0; 1492 if (_max_pending_cards >= _pending_cards) { 1493 diff = _max_pending_cards - _pending_cards; 1494 } 1495 _pending_card_diff_seq->add((double) diff); 1496 1497 double cost_per_card_ms = 0.0; 1498 if (_pending_cards > 0) { 1499 cost_per_card_ms = update_rs_time / (double) _pending_cards; 1500 _cost_per_card_ms_seq->add(cost_per_card_ms); 1501 } 1502 1503 size_t cards_scanned = _g1->cards_scanned(); 1504 1505 double cost_per_entry_ms = 0.0; 1506 if (cards_scanned > 10) { 1507 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; 1508 if (_last_gc_was_young) { 1509 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1510 } else { 1511 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1512 } 1513 } 1514 1515 if (_max_rs_lengths > 0) { 1516 double cards_per_entry_ratio = 1517 (double) cards_scanned / (double) _max_rs_lengths; 1518 if (_last_gc_was_young) { 1519 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1520 } else { 1521 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1522 } 1523 } 1524 1525 // This is defensive. For a while _max_rs_lengths could get 1526 // smaller than _recorded_rs_lengths which was causing 1527 // rs_length_diff to get very large and mess up the RSet length 1528 // predictions. The reason was unsafe concurrent updates to the 1529 // _inc_cset_recorded_rs_lengths field which the code below guards 1530 // against (see CR 7118202). This bug has now been fixed (see CR 1531 // 7119027). However, I'm still worried that 1532 // _inc_cset_recorded_rs_lengths might still end up somewhat 1533 // inaccurate. The concurrent refinement thread calculates an 1534 // RSet's length concurrently with other CR threads updating it 1535 // which might cause it to calculate the length incorrectly (if, 1536 // say, it's in mid-coarsening). So I'll leave in the defensive 1537 // conditional below just in case. 1538 size_t rs_length_diff = 0; 1539 if (_max_rs_lengths > _recorded_rs_lengths) { 1540 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1541 } 1542 _rs_length_diff_seq->add((double) rs_length_diff); 1543 1544 size_t copied_bytes = surviving_bytes; 1545 double cost_per_byte_ms = 0.0; 1546 if (copied_bytes > 0) { 1547 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; 1548 if (_in_marking_window) { 1549 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1550 } else { 1551 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1552 } 1553 } 1554 1555 double all_other_time_ms = pause_time_ms - 1556 (update_rs_time + scan_rs_time + obj_copy_time + termination_time); 1557 1558 double young_other_time_ms = 0.0; 1559 if (young_cset_region_length() > 0) { 1560 young_other_time_ms = 1561 _recorded_young_cset_choice_time_ms + 1562 _recorded_young_free_cset_time_ms; 1563 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1564 (double) young_cset_region_length()); 1565 } 1566 double non_young_other_time_ms = 0.0; 1567 if (old_cset_region_length() > 0) { 1568 non_young_other_time_ms = 1569 _recorded_non_young_cset_choice_time_ms + 1570 _recorded_non_young_free_cset_time_ms; 1571 1572 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1573 (double) old_cset_region_length()); 1574 } 1575 1576 double constant_other_time_ms = all_other_time_ms - 1577 (young_other_time_ms + non_young_other_time_ms); 1578 _constant_other_time_ms_seq->add(constant_other_time_ms); 1579 1580 double survival_ratio = 0.0; 1581 if (_bytes_in_collection_set_before_gc > 0) { 1582 survival_ratio = (double) _bytes_copied_during_gc / 1583 (double) _bytes_in_collection_set_before_gc; 1584 } 1585 1586 _pending_cards_seq->add((double) _pending_cards); 1587 _rs_lengths_seq->add((double) _max_rs_lengths); 1588 } 1589 1590 _in_marking_window = new_in_marking_window; 1591 _in_marking_window_im = new_in_marking_window_im; 1592 _free_regions_at_end_of_collection = _g1->free_regions(); 1593 update_young_list_target_length(); 1594 1595 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1596 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1597 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms); 1598 1599 _collectionSetChooser->verify(); 1600 } 1601 1602 #define EXT_SIZE_FORMAT "%d%s" 1603 #define EXT_SIZE_PARAMS(bytes) \ 1604 byte_size_in_proper_unit((bytes)), \ 1605 proper_unit_for_byte_size((bytes)) 1606 1607 void G1CollectorPolicy::print_heap_transition() { 1608 if (G1Log::finer()) { 1609 YoungList* young_list = _g1->young_list(); 1610 size_t eden_bytes = young_list->eden_used_bytes(); 1611 size_t survivor_bytes = young_list->survivor_used_bytes(); 1612 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes; 1613 size_t used = _g1->used(); 1614 size_t capacity = _g1->capacity(); 1615 size_t eden_capacity = 1616 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes; 1617 1618 gclog_or_tty->print_cr( 1619 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") " 1620 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" " 1621 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->" 1622 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]", 1623 EXT_SIZE_PARAMS(_eden_bytes_before_gc), 1624 EXT_SIZE_PARAMS(_prev_eden_capacity), 1625 EXT_SIZE_PARAMS(eden_bytes), 1626 EXT_SIZE_PARAMS(eden_capacity), 1627 EXT_SIZE_PARAMS(_survivor_bytes_before_gc), 1628 EXT_SIZE_PARAMS(survivor_bytes), 1629 EXT_SIZE_PARAMS(used_before_gc), 1630 EXT_SIZE_PARAMS(_capacity_before_gc), 1631 EXT_SIZE_PARAMS(used), 1632 EXT_SIZE_PARAMS(capacity)); 1633 1634 _prev_eden_capacity = eden_capacity; 1635 } else if (G1Log::fine()) { 1636 _g1->print_size_transition(gclog_or_tty, 1637 _cur_collection_pause_used_at_start_bytes, 1638 _g1->used(), _g1->capacity()); 1639 } 1640 } 1641 1642 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1643 double update_rs_processed_buffers, 1644 double goal_ms) { 1645 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1646 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1647 1648 if (G1UseAdaptiveConcRefinement) { 1649 const int k_gy = 3, k_gr = 6; 1650 const double inc_k = 1.1, dec_k = 0.9; 1651 1652 int g = cg1r->green_zone(); 1653 if (update_rs_time > goal_ms) { 1654 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1655 } else { 1656 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1657 g = (int)MAX2(g * inc_k, g + 1.0); 1658 } 1659 } 1660 // Change the refinement threads params 1661 cg1r->set_green_zone(g); 1662 cg1r->set_yellow_zone(g * k_gy); 1663 cg1r->set_red_zone(g * k_gr); 1664 cg1r->reinitialize_threads(); 1665 1666 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); 1667 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1668 cg1r->yellow_zone()); 1669 // Change the barrier params 1670 dcqs.set_process_completed_threshold(processing_threshold); 1671 dcqs.set_max_completed_queue(cg1r->red_zone()); 1672 } 1673 1674 int curr_queue_size = dcqs.completed_buffers_num(); 1675 if (curr_queue_size >= cg1r->yellow_zone()) { 1676 dcqs.set_completed_queue_padding(curr_queue_size); 1677 } else { 1678 dcqs.set_completed_queue_padding(0); 1679 } 1680 dcqs.notify_if_necessary(); 1681 } 1682 1683 double 1684 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { 1685 size_t rs_length = predict_rs_length_diff(); 1686 size_t card_num; 1687 if (gcs_are_young()) { 1688 card_num = predict_young_card_num(rs_length); 1689 } else { 1690 card_num = predict_non_young_card_num(rs_length); 1691 } 1692 return predict_base_elapsed_time_ms(pending_cards, card_num); 1693 } 1694 1695 double 1696 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1697 size_t scanned_cards) { 1698 return 1699 predict_rs_update_time_ms(pending_cards) + 1700 predict_rs_scan_time_ms(scanned_cards) + 1701 predict_constant_other_time_ms(); 1702 } 1703 1704 double 1705 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1706 bool young) { 1707 size_t rs_length = hr->rem_set()->occupied(); 1708 size_t card_num; 1709 if (gcs_are_young()) { 1710 card_num = predict_young_card_num(rs_length); 1711 } else { 1712 card_num = predict_non_young_card_num(rs_length); 1713 } 1714 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1715 1716 double region_elapsed_time_ms = 1717 predict_rs_scan_time_ms(card_num) + 1718 predict_object_copy_time_ms(bytes_to_copy); 1719 1720 if (young) 1721 region_elapsed_time_ms += predict_young_other_time_ms(1); 1722 else 1723 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1724 1725 return region_elapsed_time_ms; 1726 } 1727 1728 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { 1729 size_t bytes_to_copy; 1730 if (hr->is_marked()) 1731 bytes_to_copy = hr->max_live_bytes(); 1732 else { 1733 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1734 int age = hr->age_in_surv_rate_group(); 1735 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1736 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); 1737 } 1738 return bytes_to_copy; 1739 } 1740 1741 void 1742 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1743 uint survivor_cset_region_length) { 1744 _eden_cset_region_length = eden_cset_region_length; 1745 _survivor_cset_region_length = survivor_cset_region_length; 1746 _old_cset_region_length = 0; 1747 } 1748 1749 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1750 _recorded_rs_lengths = rs_lengths; 1751 } 1752 1753 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1754 double elapsed_ms) { 1755 _recent_gc_times_ms->add(elapsed_ms); 1756 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1757 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1758 } 1759 1760 size_t G1CollectorPolicy::expansion_amount() { 1761 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1762 double threshold = _gc_overhead_perc; 1763 if (recent_gc_overhead > threshold) { 1764 // We will double the existing space, or take 1765 // G1ExpandByPercentOfAvailable % of the available expansion 1766 // space, whichever is smaller, bounded below by a minimum 1767 // expansion (unless that's all that's left.) 1768 const size_t min_expand_bytes = 1*M; 1769 size_t reserved_bytes = _g1->max_capacity(); 1770 size_t committed_bytes = _g1->capacity(); 1771 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1772 size_t expand_bytes; 1773 size_t expand_bytes_via_pct = 1774 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1775 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1776 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1777 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1778 1779 ergo_verbose5(ErgoHeapSizing, 1780 "attempt heap expansion", 1781 ergo_format_reason("recent GC overhead higher than " 1782 "threshold after GC") 1783 ergo_format_perc("recent GC overhead") 1784 ergo_format_perc("threshold") 1785 ergo_format_byte("uncommitted") 1786 ergo_format_byte_perc("calculated expansion amount"), 1787 recent_gc_overhead, threshold, 1788 uncommitted_bytes, 1789 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1790 1791 return expand_bytes; 1792 } else { 1793 return 0; 1794 } 1795 } 1796 1797 class CountCSClosure: public HeapRegionClosure { 1798 G1CollectorPolicy* _g1_policy; 1799 public: 1800 CountCSClosure(G1CollectorPolicy* g1_policy) : 1801 _g1_policy(g1_policy) {} 1802 bool doHeapRegion(HeapRegion* r) { 1803 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); 1804 return false; 1805 } 1806 }; 1807 1808 void G1CollectorPolicy::count_CS_bytes_used() { 1809 CountCSClosure cs_closure(this); 1810 _g1->collection_set_iterate(&cs_closure); 1811 } 1812 1813 void G1CollectorPolicy::print_summary(int level, 1814 const char* str, 1815 NumberSeq* seq) const { 1816 double sum = seq->sum(); 1817 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", 1818 str, sum / 1000.0, seq->avg()); 1819 } 1820 1821 void G1CollectorPolicy::print_summary_sd(int level, 1822 const char* str, 1823 NumberSeq* seq) const { 1824 print_summary(level, str, seq); 1825 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 1826 seq->num(), seq->sd(), seq->maximum()); 1827 } 1828 1829 void G1CollectorPolicy::check_other_times(int level, 1830 NumberSeq* other_times_ms, 1831 NumberSeq* calc_other_times_ms) const { 1832 bool should_print = false; 1833 LineBuffer buf(level + 2); 1834 1835 double max_sum = MAX2(fabs(other_times_ms->sum()), 1836 fabs(calc_other_times_ms->sum())); 1837 double min_sum = MIN2(fabs(other_times_ms->sum()), 1838 fabs(calc_other_times_ms->sum())); 1839 double sum_ratio = max_sum / min_sum; 1840 if (sum_ratio > 1.1) { 1841 should_print = true; 1842 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###"); 1843 } 1844 1845 double max_avg = MAX2(fabs(other_times_ms->avg()), 1846 fabs(calc_other_times_ms->avg())); 1847 double min_avg = MIN2(fabs(other_times_ms->avg()), 1848 fabs(calc_other_times_ms->avg())); 1849 double avg_ratio = max_avg / min_avg; 1850 if (avg_ratio > 1.1) { 1851 should_print = true; 1852 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###"); 1853 } 1854 1855 if (other_times_ms->sum() < -0.01) { 1856 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###"); 1857 } 1858 1859 if (other_times_ms->avg() < -0.01) { 1860 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###"); 1861 } 1862 1863 if (calc_other_times_ms->sum() < -0.01) { 1864 should_print = true; 1865 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###"); 1866 } 1867 1868 if (calc_other_times_ms->avg() < -0.01) { 1869 should_print = true; 1870 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###"); 1871 } 1872 1873 if (should_print) 1874 print_summary(level, "Other(Calc)", calc_other_times_ms); 1875 } 1876 1877 void G1CollectorPolicy::print_summary(PauseSummary* summary) const { 1878 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1879 MainBodySummary* body_summary = summary->main_body_summary(); 1880 if (summary->get_total_seq()->num() > 0) { 1881 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq()); 1882 if (body_summary != NULL) { 1883 print_summary(1, "Root Region Scan Wait", body_summary->get_root_region_scan_wait_seq()); 1884 if (parallel) { 1885 print_summary(1, "Parallel Time", body_summary->get_parallel_seq()); 1886 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1887 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq()); 1888 print_summary(2, "Update RS", body_summary->get_update_rs_seq()); 1889 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq()); 1890 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq()); 1891 print_summary(2, "Termination", body_summary->get_termination_seq()); 1892 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq()); 1893 { 1894 NumberSeq* other_parts[] = { 1895 body_summary->get_ext_root_scan_seq(), 1896 body_summary->get_satb_filtering_seq(), 1897 body_summary->get_update_rs_seq(), 1898 body_summary->get_scan_rs_seq(), 1899 body_summary->get_obj_copy_seq(), 1900 body_summary->get_termination_seq() 1901 }; 1902 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(), 1903 6, other_parts); 1904 check_other_times(2, body_summary->get_parallel_other_seq(), 1905 &calc_other_times_ms); 1906 } 1907 } else { 1908 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq()); 1909 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq()); 1910 print_summary(1, "Update RS", body_summary->get_update_rs_seq()); 1911 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq()); 1912 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq()); 1913 } 1914 } 1915 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq()); 1916 print_summary(1, "Other", summary->get_other_seq()); 1917 { 1918 if (body_summary != NULL) { 1919 NumberSeq calc_other_times_ms; 1920 if (parallel) { 1921 // parallel 1922 NumberSeq* other_parts[] = { 1923 body_summary->get_root_region_scan_wait_seq(), 1924 body_summary->get_parallel_seq(), 1925 body_summary->get_clear_ct_seq() 1926 }; 1927 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 1928 3, other_parts); 1929 } else { 1930 // serial 1931 NumberSeq* other_parts[] = { 1932 body_summary->get_root_region_scan_wait_seq(), 1933 body_summary->get_update_rs_seq(), 1934 body_summary->get_ext_root_scan_seq(), 1935 body_summary->get_satb_filtering_seq(), 1936 body_summary->get_scan_rs_seq(), 1937 body_summary->get_obj_copy_seq() 1938 }; 1939 calc_other_times_ms = NumberSeq(summary->get_total_seq(), 1940 6, other_parts); 1941 } 1942 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms); 1943 } 1944 } 1945 } else { 1946 LineBuffer(1).append_and_print_cr("none"); 1947 } 1948 LineBuffer(0).append_and_print_cr(""); 1949 } 1950 1951 void G1CollectorPolicy::print_tracing_info() const { 1952 if (TraceGen0Time) { 1953 gclog_or_tty->print_cr("ALL PAUSES"); 1954 print_summary_sd(0, "Total", _all_pause_times_ms); 1955 gclog_or_tty->print_cr(""); 1956 gclog_or_tty->print_cr(""); 1957 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 1958 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 1959 gclog_or_tty->print_cr(""); 1960 1961 gclog_or_tty->print_cr("EVACUATION PAUSES"); 1962 print_summary(_summary); 1963 1964 gclog_or_tty->print_cr("MISC"); 1965 print_summary_sd(0, "Stop World", _all_stop_world_times_ms); 1966 print_summary_sd(0, "Yields", _all_yield_times_ms); 1967 for (int i = 0; i < _aux_num; ++i) { 1968 if (_all_aux_times_ms[i].num() > 0) { 1969 char buffer[96]; 1970 sprintf(buffer, "Aux%d", i); 1971 print_summary_sd(0, buffer, &_all_aux_times_ms[i]); 1972 } 1973 } 1974 } 1975 if (TraceGen1Time) { 1976 if (_all_full_gc_times_ms->num() > 0) { 1977 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 1978 _all_full_gc_times_ms->num(), 1979 _all_full_gc_times_ms->sum() / 1000.0); 1980 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg()); 1981 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 1982 _all_full_gc_times_ms->sd(), 1983 _all_full_gc_times_ms->maximum()); 1984 } 1985 } 1986 } 1987 1988 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1989 #ifndef PRODUCT 1990 _short_lived_surv_rate_group->print_surv_rate_summary(); 1991 // add this call for any other surv rate groups 1992 #endif // PRODUCT 1993 } 1994 1995 #ifndef PRODUCT 1996 // for debugging, bit of a hack... 1997 static char* 1998 region_num_to_mbs(int length) { 1999 static char buffer[64]; 2000 double bytes = (double) (length * HeapRegion::GrainBytes); 2001 double mbs = bytes / (double) (1024 * 1024); 2002 sprintf(buffer, "%7.2lfMB", mbs); 2003 return buffer; 2004 } 2005 #endif // PRODUCT 2006 2007 uint G1CollectorPolicy::max_regions(int purpose) { 2008 switch (purpose) { 2009 case GCAllocForSurvived: 2010 return _max_survivor_regions; 2011 case GCAllocForTenured: 2012 return REGIONS_UNLIMITED; 2013 default: 2014 ShouldNotReachHere(); 2015 return REGIONS_UNLIMITED; 2016 }; 2017 } 2018 2019 void G1CollectorPolicy::update_max_gc_locker_expansion() { 2020 uint expansion_region_num = 0; 2021 if (GCLockerEdenExpansionPercent > 0) { 2022 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 2023 double expansion_region_num_d = perc * (double) _young_list_target_length; 2024 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 2025 // less than 1.0) we'll get 1. 2026 expansion_region_num = (uint) ceil(expansion_region_num_d); 2027 } else { 2028 assert(expansion_region_num == 0, "sanity"); 2029 } 2030 _young_list_max_length = _young_list_target_length + expansion_region_num; 2031 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 2032 } 2033 2034 // Calculates survivor space parameters. 2035 void G1CollectorPolicy::update_survivors_policy() { 2036 double max_survivor_regions_d = 2037 (double) _young_list_target_length / (double) SurvivorRatio; 2038 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 2039 // smaller than 1.0) we'll get 1. 2040 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 2041 2042 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 2043 HeapRegion::GrainWords * _max_survivor_regions); 2044 } 2045 2046 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 2047 GCCause::Cause gc_cause) { 2048 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2049 if (!during_cycle) { 2050 ergo_verbose1(ErgoConcCycles, 2051 "request concurrent cycle initiation", 2052 ergo_format_reason("requested by GC cause") 2053 ergo_format_str("GC cause"), 2054 GCCause::to_string(gc_cause)); 2055 set_initiate_conc_mark_if_possible(); 2056 return true; 2057 } else { 2058 ergo_verbose1(ErgoConcCycles, 2059 "do not request concurrent cycle initiation", 2060 ergo_format_reason("concurrent cycle already in progress") 2061 ergo_format_str("GC cause"), 2062 GCCause::to_string(gc_cause)); 2063 return false; 2064 } 2065 } 2066 2067 void 2068 G1CollectorPolicy::decide_on_conc_mark_initiation() { 2069 // We are about to decide on whether this pause will be an 2070 // initial-mark pause. 2071 2072 // First, during_initial_mark_pause() should not be already set. We 2073 // will set it here if we have to. However, it should be cleared by 2074 // the end of the pause (it's only set for the duration of an 2075 // initial-mark pause). 2076 assert(!during_initial_mark_pause(), "pre-condition"); 2077 2078 if (initiate_conc_mark_if_possible()) { 2079 // We had noticed on a previous pause that the heap occupancy has 2080 // gone over the initiating threshold and we should start a 2081 // concurrent marking cycle. So we might initiate one. 2082 2083 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 2084 if (!during_cycle) { 2085 // The concurrent marking thread is not "during a cycle", i.e., 2086 // it has completed the last one. So we can go ahead and 2087 // initiate a new cycle. 2088 2089 set_during_initial_mark_pause(); 2090 // We do not allow mixed GCs during marking. 2091 if (!gcs_are_young()) { 2092 set_gcs_are_young(true); 2093 ergo_verbose0(ErgoMixedGCs, 2094 "end mixed GCs", 2095 ergo_format_reason("concurrent cycle is about to start")); 2096 } 2097 2098 // And we can now clear initiate_conc_mark_if_possible() as 2099 // we've already acted on it. 2100 clear_initiate_conc_mark_if_possible(); 2101 2102 ergo_verbose0(ErgoConcCycles, 2103 "initiate concurrent cycle", 2104 ergo_format_reason("concurrent cycle initiation requested")); 2105 } else { 2106 // The concurrent marking thread is still finishing up the 2107 // previous cycle. If we start one right now the two cycles 2108 // overlap. In particular, the concurrent marking thread might 2109 // be in the process of clearing the next marking bitmap (which 2110 // we will use for the next cycle if we start one). Starting a 2111 // cycle now will be bad given that parts of the marking 2112 // information might get cleared by the marking thread. And we 2113 // cannot wait for the marking thread to finish the cycle as it 2114 // periodically yields while clearing the next marking bitmap 2115 // and, if it's in a yield point, it's waiting for us to 2116 // finish. So, at this point we will not start a cycle and we'll 2117 // let the concurrent marking thread complete the last one. 2118 ergo_verbose0(ErgoConcCycles, 2119 "do not initiate concurrent cycle", 2120 ergo_format_reason("concurrent cycle already in progress")); 2121 } 2122 } 2123 } 2124 2125 class KnownGarbageClosure: public HeapRegionClosure { 2126 G1CollectedHeap* _g1h; 2127 CollectionSetChooser* _hrSorted; 2128 2129 public: 2130 KnownGarbageClosure(CollectionSetChooser* hrSorted) : 2131 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { } 2132 2133 bool doHeapRegion(HeapRegion* r) { 2134 // We only include humongous regions in collection 2135 // sets when concurrent mark shows that their contained object is 2136 // unreachable. 2137 2138 // Do we have any marking information for this region? 2139 if (r->is_marked()) { 2140 // We will skip any region that's currently used as an old GC 2141 // alloc region (we should not consider those for collection 2142 // before we fill them up). 2143 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 2144 _hrSorted->add_region(r); 2145 } 2146 } 2147 return false; 2148 } 2149 }; 2150 2151 class ParKnownGarbageHRClosure: public HeapRegionClosure { 2152 G1CollectedHeap* _g1h; 2153 CollectionSetChooser* _hrSorted; 2154 uint _marked_regions_added; 2155 size_t _reclaimable_bytes_added; 2156 uint _chunk_size; 2157 uint _cur_chunk_idx; 2158 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) 2159 2160 void get_new_chunk() { 2161 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size); 2162 _cur_chunk_end = _cur_chunk_idx + _chunk_size; 2163 } 2164 void add_region(HeapRegion* r) { 2165 if (_cur_chunk_idx == _cur_chunk_end) { 2166 get_new_chunk(); 2167 } 2168 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); 2169 _hrSorted->set_region(_cur_chunk_idx, r); 2170 _marked_regions_added++; 2171 _reclaimable_bytes_added += r->reclaimable_bytes(); 2172 _cur_chunk_idx++; 2173 } 2174 2175 public: 2176 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 2177 uint chunk_size) : 2178 _g1h(G1CollectedHeap::heap()), 2179 _hrSorted(hrSorted), _chunk_size(chunk_size), 2180 _marked_regions_added(0), _reclaimable_bytes_added(0), 2181 _cur_chunk_idx(0), _cur_chunk_end(0) { } 2182 2183 bool doHeapRegion(HeapRegion* r) { 2184 // Do we have any marking information for this region? 2185 if (r->is_marked()) { 2186 // We will skip any region that's currently used as an old GC 2187 // alloc region (we should not consider those for collection 2188 // before we fill them up). 2189 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 2190 add_region(r); 2191 } 2192 } 2193 return false; 2194 } 2195 uint marked_regions_added() { return _marked_regions_added; } 2196 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; } 2197 }; 2198 2199 class ParKnownGarbageTask: public AbstractGangTask { 2200 CollectionSetChooser* _hrSorted; 2201 uint _chunk_size; 2202 G1CollectedHeap* _g1; 2203 public: 2204 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) : 2205 AbstractGangTask("ParKnownGarbageTask"), 2206 _hrSorted(hrSorted), _chunk_size(chunk_size), 2207 _g1(G1CollectedHeap::heap()) { } 2208 2209 void work(uint worker_id) { 2210 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 2211 2212 // Back to zero for the claim value. 2213 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id, 2214 _g1->workers()->active_workers(), 2215 HeapRegion::InitialClaimValue); 2216 uint regions_added = parKnownGarbageCl.marked_regions_added(); 2217 size_t reclaimable_bytes_added = 2218 parKnownGarbageCl.reclaimable_bytes_added(); 2219 _hrSorted->update_totals(regions_added, reclaimable_bytes_added); 2220 } 2221 }; 2222 2223 void 2224 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) { 2225 _collectionSetChooser->clear(); 2226 2227 uint region_num = _g1->n_regions(); 2228 if (G1CollectedHeap::use_parallel_gc_threads()) { 2229 const uint OverpartitionFactor = 4; 2230 uint WorkUnit; 2231 // The use of MinChunkSize = 8 in the original code 2232 // causes some assertion failures when the total number of 2233 // region is less than 8. The code here tries to fix that. 2234 // Should the original code also be fixed? 2235 if (no_of_gc_threads > 0) { 2236 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U); 2237 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor), 2238 MinWorkUnit); 2239 } else { 2240 assert(no_of_gc_threads > 0, 2241 "The active gc workers should be greater than 0"); 2242 // In a product build do something reasonable to avoid a crash. 2243 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U); 2244 WorkUnit = 2245 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor), 2246 MinWorkUnit); 2247 } 2248 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(), 2249 WorkUnit); 2250 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, 2251 (int) WorkUnit); 2252 _g1->workers()->run_task(&parKnownGarbageTask); 2253 2254 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 2255 "sanity check"); 2256 } else { 2257 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); 2258 _g1->heap_region_iterate(&knownGarbagecl); 2259 } 2260 2261 _collectionSetChooser->sort_regions(); 2262 2263 double end_sec = os::elapsedTime(); 2264 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 2265 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 2266 _cur_mark_stop_world_time_ms += elapsed_time_ms; 2267 _prev_collection_pause_end_ms += elapsed_time_ms; 2268 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true); 2269 } 2270 2271 // Add the heap region at the head of the non-incremental collection set 2272 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 2273 assert(_inc_cset_build_state == Active, "Precondition"); 2274 assert(!hr->is_young(), "non-incremental add of young region"); 2275 2276 assert(!hr->in_collection_set(), "should not already be in the CSet"); 2277 hr->set_in_collection_set(true); 2278 hr->set_next_in_collection_set(_collection_set); 2279 _collection_set = hr; 2280 _collection_set_bytes_used_before += hr->used(); 2281 _g1->register_region_with_in_cset_fast_test(hr); 2282 size_t rs_length = hr->rem_set()->occupied(); 2283 _recorded_rs_lengths += rs_length; 2284 _old_cset_region_length += 1; 2285 } 2286 2287 // Initialize the per-collection-set information 2288 void G1CollectorPolicy::start_incremental_cset_building() { 2289 assert(_inc_cset_build_state == Inactive, "Precondition"); 2290 2291 _inc_cset_head = NULL; 2292 _inc_cset_tail = NULL; 2293 _inc_cset_bytes_used_before = 0; 2294 2295 _inc_cset_max_finger = 0; 2296 _inc_cset_recorded_rs_lengths = 0; 2297 _inc_cset_recorded_rs_lengths_diffs = 0; 2298 _inc_cset_predicted_elapsed_time_ms = 0.0; 2299 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2300 _inc_cset_build_state = Active; 2301 } 2302 2303 void G1CollectorPolicy::finalize_incremental_cset_building() { 2304 assert(_inc_cset_build_state == Active, "Precondition"); 2305 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2306 2307 // The two "main" fields, _inc_cset_recorded_rs_lengths and 2308 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 2309 // that adds a new region to the CSet. Further updates by the 2310 // concurrent refinement thread that samples the young RSet lengths 2311 // are accumulated in the *_diffs fields. Here we add the diffs to 2312 // the "main" fields. 2313 2314 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 2315 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 2316 } else { 2317 // This is defensive. The diff should in theory be always positive 2318 // as RSets can only grow between GCs. However, given that we 2319 // sample their size concurrently with other threads updating them 2320 // it's possible that we might get the wrong size back, which 2321 // could make the calculations somewhat inaccurate. 2322 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 2323 if (_inc_cset_recorded_rs_lengths >= diffs) { 2324 _inc_cset_recorded_rs_lengths -= diffs; 2325 } else { 2326 _inc_cset_recorded_rs_lengths = 0; 2327 } 2328 } 2329 _inc_cset_predicted_elapsed_time_ms += 2330 _inc_cset_predicted_elapsed_time_ms_diffs; 2331 2332 _inc_cset_recorded_rs_lengths_diffs = 0; 2333 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2334 } 2335 2336 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 2337 // This routine is used when: 2338 // * adding survivor regions to the incremental cset at the end of an 2339 // evacuation pause, 2340 // * adding the current allocation region to the incremental cset 2341 // when it is retired, and 2342 // * updating existing policy information for a region in the 2343 // incremental cset via young list RSet sampling. 2344 // Therefore this routine may be called at a safepoint by the 2345 // VM thread, or in-between safepoints by mutator threads (when 2346 // retiring the current allocation region) or a concurrent 2347 // refine thread (RSet sampling). 2348 2349 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2350 size_t used_bytes = hr->used(); 2351 _inc_cset_recorded_rs_lengths += rs_length; 2352 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 2353 _inc_cset_bytes_used_before += used_bytes; 2354 2355 // Cache the values we have added to the aggregated informtion 2356 // in the heap region in case we have to remove this region from 2357 // the incremental collection set, or it is updated by the 2358 // rset sampling code 2359 hr->set_recorded_rs_length(rs_length); 2360 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 2361 } 2362 2363 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 2364 size_t new_rs_length) { 2365 // Update the CSet information that is dependent on the new RS length 2366 assert(hr->is_young(), "Precondition"); 2367 assert(!SafepointSynchronize::is_at_safepoint(), 2368 "should not be at a safepoint"); 2369 2370 // We could have updated _inc_cset_recorded_rs_lengths and 2371 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 2372 // that atomically, as this code is executed by a concurrent 2373 // refinement thread, potentially concurrently with a mutator thread 2374 // allocating a new region and also updating the same fields. To 2375 // avoid the atomic operations we accumulate these updates on two 2376 // separate fields (*_diffs) and we'll just add them to the "main" 2377 // fields at the start of a GC. 2378 2379 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 2380 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 2381 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 2382 2383 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 2384 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2385 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 2386 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 2387 2388 hr->set_recorded_rs_length(new_rs_length); 2389 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 2390 } 2391 2392 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 2393 assert(hr->is_young(), "invariant"); 2394 assert(hr->young_index_in_cset() > -1, "should have already been set"); 2395 assert(_inc_cset_build_state == Active, "Precondition"); 2396 2397 // We need to clear and set the cached recorded/cached collection set 2398 // information in the heap region here (before the region gets added 2399 // to the collection set). An individual heap region's cached values 2400 // are calculated, aggregated with the policy collection set info, 2401 // and cached in the heap region here (initially) and (subsequently) 2402 // by the Young List sampling code. 2403 2404 size_t rs_length = hr->rem_set()->occupied(); 2405 add_to_incremental_cset_info(hr, rs_length); 2406 2407 HeapWord* hr_end = hr->end(); 2408 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 2409 2410 assert(!hr->in_collection_set(), "invariant"); 2411 hr->set_in_collection_set(true); 2412 assert( hr->next_in_collection_set() == NULL, "invariant"); 2413 2414 _g1->register_region_with_in_cset_fast_test(hr); 2415 } 2416 2417 // Add the region at the RHS of the incremental cset 2418 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 2419 // We should only ever be appending survivors at the end of a pause 2420 assert( hr->is_survivor(), "Logic"); 2421 2422 // Do the 'common' stuff 2423 add_region_to_incremental_cset_common(hr); 2424 2425 // Now add the region at the right hand side 2426 if (_inc_cset_tail == NULL) { 2427 assert(_inc_cset_head == NULL, "invariant"); 2428 _inc_cset_head = hr; 2429 } else { 2430 _inc_cset_tail->set_next_in_collection_set(hr); 2431 } 2432 _inc_cset_tail = hr; 2433 } 2434 2435 // Add the region to the LHS of the incremental cset 2436 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2437 // Survivors should be added to the RHS at the end of a pause 2438 assert(!hr->is_survivor(), "Logic"); 2439 2440 // Do the 'common' stuff 2441 add_region_to_incremental_cset_common(hr); 2442 2443 // Add the region at the left hand side 2444 hr->set_next_in_collection_set(_inc_cset_head); 2445 if (_inc_cset_head == NULL) { 2446 assert(_inc_cset_tail == NULL, "Invariant"); 2447 _inc_cset_tail = hr; 2448 } 2449 _inc_cset_head = hr; 2450 } 2451 2452 #ifndef PRODUCT 2453 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2454 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2455 2456 st->print_cr("\nCollection_set:"); 2457 HeapRegion* csr = list_head; 2458 while (csr != NULL) { 2459 HeapRegion* next = csr->next_in_collection_set(); 2460 assert(csr->in_collection_set(), "bad CS"); 2461 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 2462 HR_FORMAT_PARAMS(csr), 2463 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(), 2464 csr->age_in_surv_rate_group_cond()); 2465 csr = next; 2466 } 2467 } 2468 #endif // !PRODUCT 2469 2470 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2471 const char* false_action_str) { 2472 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2473 if (cset_chooser->is_empty()) { 2474 ergo_verbose0(ErgoMixedGCs, 2475 false_action_str, 2476 ergo_format_reason("candidate old regions not available")); 2477 return false; 2478 } 2479 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2480 size_t capacity_bytes = _g1->capacity(); 2481 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2482 double threshold = (double) G1HeapWastePercent; 2483 if (perc < threshold) { 2484 ergo_verbose4(ErgoMixedGCs, 2485 false_action_str, 2486 ergo_format_reason("reclaimable percentage lower than threshold") 2487 ergo_format_region("candidate old regions") 2488 ergo_format_byte_perc("reclaimable") 2489 ergo_format_perc("threshold"), 2490 cset_chooser->remaining_regions(), 2491 reclaimable_bytes, perc, threshold); 2492 return false; 2493 } 2494 2495 ergo_verbose4(ErgoMixedGCs, 2496 true_action_str, 2497 ergo_format_reason("candidate old regions available") 2498 ergo_format_region("candidate old regions") 2499 ergo_format_byte_perc("reclaimable") 2500 ergo_format_perc("threshold"), 2501 cset_chooser->remaining_regions(), 2502 reclaimable_bytes, perc, threshold); 2503 return true; 2504 } 2505 2506 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) { 2507 // Set this here - in case we're not doing young collections. 2508 double non_young_start_time_sec = os::elapsedTime(); 2509 2510 YoungList* young_list = _g1->young_list(); 2511 finalize_incremental_cset_building(); 2512 2513 guarantee(target_pause_time_ms > 0.0, 2514 err_msg("target_pause_time_ms = %1.6lf should be positive", 2515 target_pause_time_ms)); 2516 guarantee(_collection_set == NULL, "Precondition"); 2517 2518 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2519 double predicted_pause_time_ms = base_time_ms; 2520 double time_remaining_ms = target_pause_time_ms - base_time_ms; 2521 2522 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2523 "start choosing CSet", 2524 ergo_format_ms("predicted base time") 2525 ergo_format_ms("remaining time") 2526 ergo_format_ms("target pause time"), 2527 base_time_ms, time_remaining_ms, target_pause_time_ms); 2528 2529 HeapRegion* hr; 2530 double young_start_time_sec = os::elapsedTime(); 2531 2532 _collection_set_bytes_used_before = 0; 2533 _last_gc_was_young = gcs_are_young() ? true : false; 2534 2535 if (_last_gc_was_young) { 2536 ++_young_pause_num; 2537 } else { 2538 ++_mixed_pause_num; 2539 } 2540 2541 // The young list is laid with the survivor regions from the previous 2542 // pause are appended to the RHS of the young list, i.e. 2543 // [Newly Young Regions ++ Survivors from last pause]. 2544 2545 uint survivor_region_length = young_list->survivor_length(); 2546 uint eden_region_length = young_list->length() - survivor_region_length; 2547 init_cset_region_lengths(eden_region_length, survivor_region_length); 2548 hr = young_list->first_survivor_region(); 2549 while (hr != NULL) { 2550 assert(hr->is_survivor(), "badly formed young list"); 2551 hr->set_young(); 2552 hr = hr->get_next_young_region(); 2553 } 2554 2555 // Clear the fields that point to the survivor list - they are all young now. 2556 young_list->clear_survivors(); 2557 2558 _collection_set = _inc_cset_head; 2559 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2560 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms; 2561 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; 2562 2563 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2564 "add young regions to CSet", 2565 ergo_format_region("eden") 2566 ergo_format_region("survivors") 2567 ergo_format_ms("predicted young region time"), 2568 eden_region_length, survivor_region_length, 2569 _inc_cset_predicted_elapsed_time_ms); 2570 2571 // The number of recorded young regions is the incremental 2572 // collection set's current size 2573 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2574 2575 double young_end_time_sec = os::elapsedTime(); 2576 _recorded_young_cset_choice_time_ms = 2577 (young_end_time_sec - young_start_time_sec) * 1000.0; 2578 2579 // We are doing young collections so reset this. 2580 non_young_start_time_sec = young_end_time_sec; 2581 2582 if (!gcs_are_young()) { 2583 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2584 cset_chooser->verify(); 2585 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length(); 2586 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length(); 2587 2588 uint expensive_region_num = 0; 2589 bool check_time_remaining = adaptive_young_list_length(); 2590 HeapRegion* hr = cset_chooser->peek(); 2591 while (hr != NULL) { 2592 if (old_cset_region_length() >= max_old_cset_length) { 2593 // Added maximum number of old regions to the CSet. 2594 ergo_verbose2(ErgoCSetConstruction, 2595 "finish adding old regions to CSet", 2596 ergo_format_reason("old CSet region num reached max") 2597 ergo_format_region("old") 2598 ergo_format_region("max"), 2599 old_cset_region_length(), max_old_cset_length); 2600 break; 2601 } 2602 2603 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); 2604 if (check_time_remaining) { 2605 if (predicted_time_ms > time_remaining_ms) { 2606 // Too expensive for the current CSet. 2607 2608 if (old_cset_region_length() >= min_old_cset_length) { 2609 // We have added the minimum number of old regions to the CSet, 2610 // we are done with this CSet. 2611 ergo_verbose4(ErgoCSetConstruction, 2612 "finish adding old regions to CSet", 2613 ergo_format_reason("predicted time is too high") 2614 ergo_format_ms("predicted time") 2615 ergo_format_ms("remaining time") 2616 ergo_format_region("old") 2617 ergo_format_region("min"), 2618 predicted_time_ms, time_remaining_ms, 2619 old_cset_region_length(), min_old_cset_length); 2620 break; 2621 } 2622 2623 // We'll add it anyway given that we haven't reached the 2624 // minimum number of old regions. 2625 expensive_region_num += 1; 2626 } 2627 } else { 2628 if (old_cset_region_length() >= min_old_cset_length) { 2629 // In the non-auto-tuning case, we'll finish adding regions 2630 // to the CSet if we reach the minimum. 2631 ergo_verbose2(ErgoCSetConstruction, 2632 "finish adding old regions to CSet", 2633 ergo_format_reason("old CSet region num reached min") 2634 ergo_format_region("old") 2635 ergo_format_region("min"), 2636 old_cset_region_length(), min_old_cset_length); 2637 break; 2638 } 2639 } 2640 2641 // We will add this region to the CSet. 2642 time_remaining_ms -= predicted_time_ms; 2643 predicted_pause_time_ms += predicted_time_ms; 2644 cset_chooser->remove_and_move_to_next(hr); 2645 _g1->old_set_remove(hr); 2646 add_old_region_to_cset(hr); 2647 2648 hr = cset_chooser->peek(); 2649 } 2650 if (hr == NULL) { 2651 ergo_verbose0(ErgoCSetConstruction, 2652 "finish adding old regions to CSet", 2653 ergo_format_reason("candidate old regions not available")); 2654 } 2655 2656 if (expensive_region_num > 0) { 2657 // We print the information once here at the end, predicated on 2658 // whether we added any apparently expensive regions or not, to 2659 // avoid generating output per region. 2660 ergo_verbose4(ErgoCSetConstruction, 2661 "added expensive regions to CSet", 2662 ergo_format_reason("old CSet region num not reached min") 2663 ergo_format_region("old") 2664 ergo_format_region("expensive") 2665 ergo_format_region("min") 2666 ergo_format_ms("remaining time"), 2667 old_cset_region_length(), 2668 expensive_region_num, 2669 min_old_cset_length, 2670 time_remaining_ms); 2671 } 2672 2673 cset_chooser->verify(); 2674 } 2675 2676 stop_incremental_cset_building(); 2677 2678 count_CS_bytes_used(); 2679 2680 ergo_verbose5(ErgoCSetConstruction, 2681 "finish choosing CSet", 2682 ergo_format_region("eden") 2683 ergo_format_region("survivors") 2684 ergo_format_region("old") 2685 ergo_format_ms("predicted pause time") 2686 ergo_format_ms("target pause time"), 2687 eden_region_length, survivor_region_length, 2688 old_cset_region_length(), 2689 predicted_pause_time_ms, target_pause_time_ms); 2690 2691 double non_young_end_time_sec = os::elapsedTime(); 2692 _recorded_non_young_cset_choice_time_ms = 2693 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; 2694 }