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 _stop_world_start(0.0), 137 138 _cur_clear_ct_time_ms(0.0), 139 _root_region_scan_wait_time_ms(0.0), 140 141 _cur_ref_proc_time_ms(0.0), 142 _cur_ref_enq_time_ms(0.0), 143 144 #ifndef PRODUCT 145 _min_clear_cc_time_ms(-1.0), 146 _max_clear_cc_time_ms(-1.0), 147 _cur_clear_cc_time_ms(0.0), 148 _cum_clear_cc_time_ms(0.0), 149 _num_cc_clears(0L), 150 #endif 151 152 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 153 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), 154 155 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 156 _prev_collection_pause_end_ms(0.0), 157 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 158 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), 159 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 160 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 161 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), 162 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 163 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 164 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 165 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), 166 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 167 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), 168 _non_young_other_cost_per_region_ms_seq( 169 new TruncatedSeq(TruncatedSeqLength)), 170 171 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), 172 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), 173 174 _pause_time_target_ms((double) MaxGCPauseMillis), 175 176 _gcs_are_young(true), 177 178 _during_marking(false), 179 _in_marking_window(false), 180 _in_marking_window_im(false), 181 182 _recent_prev_end_times_for_all_gcs_sec( 183 new TruncatedSeq(NumPrevPausesForHeuristics)), 184 185 _recent_avg_pause_time_ratio(0.0), 186 187 _initiate_conc_mark_if_possible(false), 188 _during_initial_mark_pause(false), 189 _last_young_gc(false), 190 _last_gc_was_young(false), 191 192 _eden_bytes_before_gc(0), 193 _survivor_bytes_before_gc(0), 194 _capacity_before_gc(0), 195 196 _eden_cset_region_length(0), 197 _survivor_cset_region_length(0), 198 _old_cset_region_length(0), 199 200 _collection_set(NULL), 201 _collection_set_bytes_used_before(0), 202 203 // Incremental CSet attributes 204 _inc_cset_build_state(Inactive), 205 _inc_cset_head(NULL), 206 _inc_cset_tail(NULL), 207 _inc_cset_bytes_used_before(0), 208 _inc_cset_max_finger(NULL), 209 _inc_cset_recorded_rs_lengths(0), 210 _inc_cset_recorded_rs_lengths_diffs(0), 211 _inc_cset_predicted_elapsed_time_ms(0.0), 212 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), 213 214 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 215 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 216 #endif // _MSC_VER 217 218 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", 219 G1YoungSurvRateNumRegionsSummary)), 220 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", 221 G1YoungSurvRateNumRegionsSummary)), 222 // add here any more surv rate groups 223 _recorded_survivor_regions(0), 224 _recorded_survivor_head(NULL), 225 _recorded_survivor_tail(NULL), 226 _survivors_age_table(true), 227 228 _gc_overhead_perc(0.0) { 229 230 // Set up the region size and associated fields. Given that the 231 // policy is created before the heap, we have to set this up here, 232 // so it's done as soon as possible. 233 HeapRegion::setup_heap_region_size(Arguments::min_heap_size()); 234 HeapRegionRemSet::setup_remset_size(); 235 236 G1ErgoVerbose::initialize(); 237 if (PrintAdaptiveSizePolicy) { 238 // Currently, we only use a single switch for all the heuristics. 239 G1ErgoVerbose::set_enabled(true); 240 // Given that we don't currently have a verboseness level 241 // parameter, we'll hardcode this to high. This can be easily 242 // changed in the future. 243 G1ErgoVerbose::set_level(ErgoHigh); 244 } else { 245 G1ErgoVerbose::set_enabled(false); 246 } 247 248 // Verify PLAB sizes 249 const size_t region_size = HeapRegion::GrainWords; 250 if (YoungPLABSize > region_size || OldPLABSize > region_size) { 251 char buffer[128]; 252 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT, 253 OldPLABSize > region_size ? "Old" : "Young", region_size); 254 vm_exit_during_initialization(buffer); 255 } 256 257 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); 258 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; 259 260 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads]; 261 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads]; 262 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads]; 263 264 _par_last_update_rs_times_ms = new double[_parallel_gc_threads]; 265 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; 266 267 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; 268 269 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads]; 270 271 _par_last_termination_times_ms = new double[_parallel_gc_threads]; 272 _par_last_termination_attempts = new double[_parallel_gc_threads]; 273 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads]; 274 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads]; 275 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads]; 276 277 int index; 278 if (ParallelGCThreads == 0) 279 index = 0; 280 else if (ParallelGCThreads > 8) 281 index = 7; 282 else 283 index = ParallelGCThreads - 1; 284 285 _pending_card_diff_seq->add(0.0); 286 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); 287 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); 288 _young_cards_per_entry_ratio_seq->add( 289 young_cards_per_entry_ratio_defaults[index]); 290 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); 291 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); 292 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); 293 _young_other_cost_per_region_ms_seq->add( 294 young_other_cost_per_region_ms_defaults[index]); 295 _non_young_other_cost_per_region_ms_seq->add( 296 non_young_other_cost_per_region_ms_defaults[index]); 297 298 // Below, we might need to calculate the pause time target based on 299 // the pause interval. When we do so we are going to give G1 maximum 300 // flexibility and allow it to do pauses when it needs to. So, we'll 301 // arrange that the pause interval to be pause time target + 1 to 302 // ensure that a) the pause time target is maximized with respect to 303 // the pause interval and b) we maintain the invariant that pause 304 // time target < pause interval. If the user does not want this 305 // maximum flexibility, they will have to set the pause interval 306 // explicitly. 307 308 // First make sure that, if either parameter is set, its value is 309 // reasonable. 310 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 311 if (MaxGCPauseMillis < 1) { 312 vm_exit_during_initialization("MaxGCPauseMillis should be " 313 "greater than 0"); 314 } 315 } 316 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 317 if (GCPauseIntervalMillis < 1) { 318 vm_exit_during_initialization("GCPauseIntervalMillis should be " 319 "greater than 0"); 320 } 321 } 322 323 // Then, if the pause time target parameter was not set, set it to 324 // the default value. 325 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { 326 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 327 // The default pause time target in G1 is 200ms 328 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); 329 } else { 330 // We do not allow the pause interval to be set without the 331 // pause time target 332 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " 333 "without setting MaxGCPauseMillis"); 334 } 335 } 336 337 // Then, if the interval parameter was not set, set it according to 338 // the pause time target (this will also deal with the case when the 339 // pause time target is the default value). 340 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { 341 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); 342 } 343 344 // Finally, make sure that the two parameters are consistent. 345 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { 346 char buffer[256]; 347 jio_snprintf(buffer, 256, 348 "MaxGCPauseMillis (%u) should be less than " 349 "GCPauseIntervalMillis (%u)", 350 MaxGCPauseMillis, GCPauseIntervalMillis); 351 vm_exit_during_initialization(buffer); 352 } 353 354 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; 355 double time_slice = (double) GCPauseIntervalMillis / 1000.0; 356 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); 357 _sigma = (double) G1ConfidencePercent / 100.0; 358 359 // start conservatively (around 50ms is about right) 360 _concurrent_mark_remark_times_ms->add(0.05); 361 _concurrent_mark_cleanup_times_ms->add(0.20); 362 _tenuring_threshold = MaxTenuringThreshold; 363 // _max_survivor_regions will be calculated by 364 // update_young_list_target_length() during initialization. 365 _max_survivor_regions = 0; 366 367 assert(GCTimeRatio > 0, 368 "we should have set it to a default value set_g1_gc_flags() " 369 "if a user set it to 0"); 370 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); 371 372 uintx reserve_perc = G1ReservePercent; 373 // Put an artificial ceiling on this so that it's not set to a silly value. 374 if (reserve_perc > 50) { 375 reserve_perc = 50; 376 warning("G1ReservePercent is set to a value that is too large, " 377 "it's been updated to %u", reserve_perc); 378 } 379 _reserve_factor = (double) reserve_perc / 100.0; 380 // This will be set when the heap is expanded 381 // for the first time during initialization. 382 _reserve_regions = 0; 383 384 initialize_all(); 385 _collectionSetChooser = new CollectionSetChooser(); 386 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags 387 } 388 389 void G1CollectorPolicy::initialize_flags() { 390 set_min_alignment(HeapRegion::GrainBytes); 391 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name())); 392 if (SurvivorRatio < 1) { 393 vm_exit_during_initialization("Invalid survivor ratio specified"); 394 } 395 CollectorPolicy::initialize_flags(); 396 } 397 398 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) { 399 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max"); 400 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds"); 401 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds"); 402 403 if (FLAG_IS_CMDLINE(NewRatio)) { 404 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { 405 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); 406 } else { 407 _sizer_kind = SizerNewRatio; 408 _adaptive_size = false; 409 return; 410 } 411 } 412 413 if (FLAG_IS_CMDLINE(NewSize)) { 414 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), 415 1U); 416 if (FLAG_IS_CMDLINE(MaxNewSize)) { 417 _max_desired_young_length = 418 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 419 1U); 420 _sizer_kind = SizerMaxAndNewSize; 421 _adaptive_size = _min_desired_young_length == _max_desired_young_length; 422 } else { 423 _sizer_kind = SizerNewSizeOnly; 424 } 425 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { 426 _max_desired_young_length = 427 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), 428 1U); 429 _sizer_kind = SizerMaxNewSizeOnly; 430 } 431 } 432 433 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { 434 uint default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100; 435 return MAX2(1U, default_value); 436 } 437 438 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { 439 uint default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100; 440 return MAX2(1U, default_value); 441 } 442 443 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { 444 assert(new_number_of_heap_regions > 0, "Heap must be initialized"); 445 446 switch (_sizer_kind) { 447 case SizerDefaults: 448 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions); 449 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions); 450 break; 451 case SizerNewSizeOnly: 452 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions); 453 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length); 454 break; 455 case SizerMaxNewSizeOnly: 456 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions); 457 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length); 458 break; 459 case SizerMaxAndNewSize: 460 // Do nothing. Values set on the command line, don't update them at runtime. 461 break; 462 case SizerNewRatio: 463 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1); 464 _max_desired_young_length = _min_desired_young_length; 465 break; 466 default: 467 ShouldNotReachHere(); 468 } 469 470 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values"); 471 } 472 473 void G1CollectorPolicy::init() { 474 // Set aside an initial future to_space. 475 _g1 = G1CollectedHeap::heap(); 476 477 assert(Heap_lock->owned_by_self(), "Locking discipline."); 478 479 initialize_gc_policy_counters(); 480 481 if (adaptive_young_list_length()) { 482 _young_list_fixed_length = 0; 483 } else { 484 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); 485 } 486 _free_regions_at_end_of_collection = _g1->free_regions(); 487 update_young_list_target_length(); 488 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes; 489 490 // We may immediately start allocating regions and placing them on the 491 // collection set list. Initialize the per-collection set info 492 start_incremental_cset_building(); 493 } 494 495 // Create the jstat counters for the policy. 496 void G1CollectorPolicy::initialize_gc_policy_counters() { 497 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); 498 } 499 500 bool G1CollectorPolicy::predict_will_fit(uint young_length, 501 double base_time_ms, 502 uint base_free_regions, 503 double target_pause_time_ms) { 504 if (young_length >= base_free_regions) { 505 // end condition 1: not enough space for the young regions 506 return false; 507 } 508 509 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); 510 size_t bytes_to_copy = 511 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); 512 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); 513 double young_other_time_ms = predict_young_other_time_ms(young_length); 514 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; 515 if (pause_time_ms > target_pause_time_ms) { 516 // end condition 2: prediction is over the target pause time 517 return false; 518 } 519 520 size_t free_bytes = 521 (base_free_regions - young_length) * HeapRegion::GrainBytes; 522 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) { 523 // end condition 3: out-of-space (conservatively!) 524 return false; 525 } 526 527 // success! 528 return true; 529 } 530 531 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { 532 // re-calculate the necessary reserve 533 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; 534 // We use ceiling so that if reserve_regions_d is > 0.0 (but 535 // smaller than 1.0) we'll get 1. 536 _reserve_regions = (uint) ceil(reserve_regions_d); 537 538 _young_gen_sizer->heap_size_changed(new_number_of_regions); 539 } 540 541 uint G1CollectorPolicy::calculate_young_list_desired_min_length( 542 uint base_min_length) { 543 uint desired_min_length = 0; 544 if (adaptive_young_list_length()) { 545 if (_alloc_rate_ms_seq->num() > 3) { 546 double now_sec = os::elapsedTime(); 547 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; 548 double alloc_rate_ms = predict_alloc_rate_ms(); 549 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); 550 } else { 551 // otherwise we don't have enough info to make the prediction 552 } 553 } 554 desired_min_length += base_min_length; 555 // make sure we don't go below any user-defined minimum bound 556 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); 557 } 558 559 uint G1CollectorPolicy::calculate_young_list_desired_max_length() { 560 // Here, we might want to also take into account any additional 561 // constraints (i.e., user-defined minimum bound). Currently, we 562 // effectively don't set this bound. 563 return _young_gen_sizer->max_desired_young_length(); 564 } 565 566 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { 567 if (rs_lengths == (size_t) -1) { 568 // if it's set to the default value (-1), we should predict it; 569 // otherwise, use the given value. 570 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); 571 } 572 573 // Calculate the absolute and desired min bounds. 574 575 // This is how many young regions we already have (currently: the survivors). 576 uint base_min_length = recorded_survivor_regions(); 577 // This is the absolute minimum young length, which ensures that we 578 // can allocate one eden region in the worst-case. 579 uint absolute_min_length = base_min_length + 1; 580 uint desired_min_length = 581 calculate_young_list_desired_min_length(base_min_length); 582 if (desired_min_length < absolute_min_length) { 583 desired_min_length = absolute_min_length; 584 } 585 586 // Calculate the absolute and desired max bounds. 587 588 // We will try our best not to "eat" into the reserve. 589 uint absolute_max_length = 0; 590 if (_free_regions_at_end_of_collection > _reserve_regions) { 591 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; 592 } 593 uint desired_max_length = calculate_young_list_desired_max_length(); 594 if (desired_max_length > absolute_max_length) { 595 desired_max_length = absolute_max_length; 596 } 597 598 uint young_list_target_length = 0; 599 if (adaptive_young_list_length()) { 600 if (gcs_are_young()) { 601 young_list_target_length = 602 calculate_young_list_target_length(rs_lengths, 603 base_min_length, 604 desired_min_length, 605 desired_max_length); 606 _rs_lengths_prediction = rs_lengths; 607 } else { 608 // Don't calculate anything and let the code below bound it to 609 // the desired_min_length, i.e., do the next GC as soon as 610 // possible to maximize how many old regions we can add to it. 611 } 612 } else { 613 // The user asked for a fixed young gen so we'll fix the young gen 614 // whether the next GC is young or mixed. 615 young_list_target_length = _young_list_fixed_length; 616 } 617 618 // Make sure we don't go over the desired max length, nor under the 619 // desired min length. In case they clash, desired_min_length wins 620 // which is why that test is second. 621 if (young_list_target_length > desired_max_length) { 622 young_list_target_length = desired_max_length; 623 } 624 if (young_list_target_length < desired_min_length) { 625 young_list_target_length = desired_min_length; 626 } 627 628 assert(young_list_target_length > recorded_survivor_regions(), 629 "we should be able to allocate at least one eden region"); 630 assert(young_list_target_length >= absolute_min_length, "post-condition"); 631 _young_list_target_length = young_list_target_length; 632 633 update_max_gc_locker_expansion(); 634 } 635 636 uint 637 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, 638 uint base_min_length, 639 uint desired_min_length, 640 uint desired_max_length) { 641 assert(adaptive_young_list_length(), "pre-condition"); 642 assert(gcs_are_young(), "only call this for young GCs"); 643 644 // In case some edge-condition makes the desired max length too small... 645 if (desired_max_length <= desired_min_length) { 646 return desired_min_length; 647 } 648 649 // We'll adjust min_young_length and max_young_length not to include 650 // the already allocated young regions (i.e., so they reflect the 651 // min and max eden regions we'll allocate). The base_min_length 652 // will be reflected in the predictions by the 653 // survivor_regions_evac_time prediction. 654 assert(desired_min_length > base_min_length, "invariant"); 655 uint min_young_length = desired_min_length - base_min_length; 656 assert(desired_max_length > base_min_length, "invariant"); 657 uint max_young_length = desired_max_length - base_min_length; 658 659 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; 660 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); 661 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); 662 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); 663 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); 664 double base_time_ms = 665 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + 666 survivor_regions_evac_time; 667 uint available_free_regions = _free_regions_at_end_of_collection; 668 uint base_free_regions = 0; 669 if (available_free_regions > _reserve_regions) { 670 base_free_regions = available_free_regions - _reserve_regions; 671 } 672 673 // Here, we will make sure that the shortest young length that 674 // makes sense fits within the target pause time. 675 676 if (predict_will_fit(min_young_length, base_time_ms, 677 base_free_regions, target_pause_time_ms)) { 678 // The shortest young length will fit into the target pause time; 679 // we'll now check whether the absolute maximum number of young 680 // regions will fit in the target pause time. If not, we'll do 681 // a binary search between min_young_length and max_young_length. 682 if (predict_will_fit(max_young_length, base_time_ms, 683 base_free_regions, target_pause_time_ms)) { 684 // The maximum young length will fit into the target pause time. 685 // We are done so set min young length to the maximum length (as 686 // the result is assumed to be returned in min_young_length). 687 min_young_length = max_young_length; 688 } else { 689 // The maximum possible number of young regions will not fit within 690 // the target pause time so we'll search for the optimal 691 // length. The loop invariants are: 692 // 693 // min_young_length < max_young_length 694 // min_young_length is known to fit into the target pause time 695 // max_young_length is known not to fit into the target pause time 696 // 697 // Going into the loop we know the above hold as we've just 698 // checked them. Every time around the loop we check whether 699 // the middle value between min_young_length and 700 // max_young_length fits into the target pause time. If it 701 // does, it becomes the new min. If it doesn't, it becomes 702 // the new max. This way we maintain the loop invariants. 703 704 assert(min_young_length < max_young_length, "invariant"); 705 uint diff = (max_young_length - min_young_length) / 2; 706 while (diff > 0) { 707 uint young_length = min_young_length + diff; 708 if (predict_will_fit(young_length, base_time_ms, 709 base_free_regions, target_pause_time_ms)) { 710 min_young_length = young_length; 711 } else { 712 max_young_length = young_length; 713 } 714 assert(min_young_length < max_young_length, "invariant"); 715 diff = (max_young_length - min_young_length) / 2; 716 } 717 // The results is min_young_length which, according to the 718 // loop invariants, should fit within the target pause time. 719 720 // These are the post-conditions of the binary search above: 721 assert(min_young_length < max_young_length, 722 "otherwise we should have discovered that max_young_length " 723 "fits into the pause target and not done the binary search"); 724 assert(predict_will_fit(min_young_length, base_time_ms, 725 base_free_regions, target_pause_time_ms), 726 "min_young_length, the result of the binary search, should " 727 "fit into the pause target"); 728 assert(!predict_will_fit(min_young_length + 1, base_time_ms, 729 base_free_regions, target_pause_time_ms), 730 "min_young_length, the result of the binary search, should be " 731 "optimal, so no larger length should fit into the pause target"); 732 } 733 } else { 734 // Even the minimum length doesn't fit into the pause time 735 // target, return it as the result nevertheless. 736 } 737 return base_min_length + min_young_length; 738 } 739 740 double G1CollectorPolicy::predict_survivor_regions_evac_time() { 741 double survivor_regions_evac_time = 0.0; 742 for (HeapRegion * r = _recorded_survivor_head; 743 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); 744 r = r->get_next_young_region()) { 745 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true); 746 } 747 return survivor_regions_evac_time; 748 } 749 750 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { 751 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); 752 753 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); 754 if (rs_lengths > _rs_lengths_prediction) { 755 // add 10% to avoid having to recalculate often 756 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; 757 update_young_list_target_length(rs_lengths_prediction); 758 } 759 } 760 761 762 763 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, 764 bool is_tlab, 765 bool* gc_overhead_limit_was_exceeded) { 766 guarantee(false, "Not using this policy feature yet."); 767 return NULL; 768 } 769 770 // This method controls how a collector handles one or more 771 // of its generations being fully allocated. 772 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, 773 bool is_tlab) { 774 guarantee(false, "Not using this policy feature yet."); 775 return NULL; 776 } 777 778 779 #ifndef PRODUCT 780 bool G1CollectorPolicy::verify_young_ages() { 781 HeapRegion* head = _g1->young_list()->first_region(); 782 return 783 verify_young_ages(head, _short_lived_surv_rate_group); 784 // also call verify_young_ages on any additional surv rate groups 785 } 786 787 bool 788 G1CollectorPolicy::verify_young_ages(HeapRegion* head, 789 SurvRateGroup *surv_rate_group) { 790 guarantee( surv_rate_group != NULL, "pre-condition" ); 791 792 const char* name = surv_rate_group->name(); 793 bool ret = true; 794 int prev_age = -1; 795 796 for (HeapRegion* curr = head; 797 curr != NULL; 798 curr = curr->get_next_young_region()) { 799 SurvRateGroup* group = curr->surv_rate_group(); 800 if (group == NULL && !curr->is_survivor()) { 801 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); 802 ret = false; 803 } 804 805 if (surv_rate_group == group) { 806 int age = curr->age_in_surv_rate_group(); 807 808 if (age < 0) { 809 gclog_or_tty->print_cr("## %s: encountered negative age", name); 810 ret = false; 811 } 812 813 if (age <= prev_age) { 814 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " 815 "(%d, %d)", name, age, prev_age); 816 ret = false; 817 } 818 prev_age = age; 819 } 820 } 821 822 return ret; 823 } 824 #endif // PRODUCT 825 826 void G1CollectorPolicy::record_full_collection_start() { 827 _cur_collection_start_sec = os::elapsedTime(); 828 // Release the future to-space so that it is available for compaction into. 829 _g1->set_full_collection(); 830 } 831 832 void G1CollectorPolicy::record_full_collection_end() { 833 // Consider this like a collection pause for the purposes of allocation 834 // since last pause. 835 double end_sec = os::elapsedTime(); 836 double full_gc_time_sec = end_sec - _cur_collection_start_sec; 837 double full_gc_time_ms = full_gc_time_sec * 1000.0; 838 839 _trace_gen1_time_data.record_full_collection(full_gc_time_ms); 840 841 update_recent_gc_times(end_sec, full_gc_time_ms); 842 843 _g1->clear_full_collection(); 844 845 // "Nuke" the heuristics that control the young/mixed GC 846 // transitions and make sure we start with young GCs after the Full GC. 847 set_gcs_are_young(true); 848 _last_young_gc = false; 849 clear_initiate_conc_mark_if_possible(); 850 clear_during_initial_mark_pause(); 851 _in_marking_window = false; 852 _in_marking_window_im = false; 853 854 _short_lived_surv_rate_group->start_adding_regions(); 855 // also call this on any additional surv rate groups 856 857 record_survivor_regions(0, NULL, NULL); 858 859 _free_regions_at_end_of_collection = _g1->free_regions(); 860 // Reset survivors SurvRateGroup. 861 _survivor_surv_rate_group->reset(); 862 update_young_list_target_length(); 863 _collectionSetChooser->clear(); 864 } 865 866 void G1CollectorPolicy::record_stop_world_start() { 867 _stop_world_start = os::elapsedTime(); 868 } 869 870 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, 871 size_t start_used) { 872 if (G1Log::finer()) { 873 gclog_or_tty->stamp(PrintGCTimeStamps); 874 gclog_or_tty->print("[%s", (const char*)GCCauseString("GC pause", _g1->gc_cause()) 875 .append(gcs_are_young() ? " (young)" : " (mixed)")); 876 } 877 878 // We only need to do this here as the policy will only be applied 879 // to the GC we're about to start. so, no point is calculating this 880 // every time we calculate / recalculate the target young length. 881 update_survivors_policy(); 882 883 assert(_g1->used() == _g1->recalculate_used(), 884 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT, 885 _g1->used(), _g1->recalculate_used())); 886 887 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; 888 _trace_gen0_time_data.record_start_collection(s_w_t_ms); 889 _stop_world_start = 0.0; 890 891 _cur_collection_start_sec = start_time_sec; 892 _cur_collection_pause_used_at_start_bytes = start_used; 893 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); 894 _pending_cards = _g1->pending_card_num(); 895 _max_pending_cards = _g1->max_pending_card_num(); 896 897 _bytes_in_collection_set_before_gc = 0; 898 _bytes_copied_during_gc = 0; 899 900 YoungList* young_list = _g1->young_list(); 901 _eden_bytes_before_gc = young_list->eden_used_bytes(); 902 _survivor_bytes_before_gc = young_list->survivor_used_bytes(); 903 _capacity_before_gc = _g1->capacity(); 904 905 #ifdef DEBUG 906 // initialise these to something well known so that we can spot 907 // if they are not set properly 908 909 for (int i = 0; i < _parallel_gc_threads; ++i) { 910 _par_last_gc_worker_start_times_ms[i] = -1234.0; 911 _par_last_ext_root_scan_times_ms[i] = -1234.0; 912 _par_last_satb_filtering_times_ms[i] = -1234.0; 913 _par_last_update_rs_times_ms[i] = -1234.0; 914 _par_last_update_rs_processed_buffers[i] = -1234.0; 915 _par_last_scan_rs_times_ms[i] = -1234.0; 916 _par_last_obj_copy_times_ms[i] = -1234.0; 917 _par_last_termination_times_ms[i] = -1234.0; 918 _par_last_termination_attempts[i] = -1234.0; 919 _par_last_gc_worker_end_times_ms[i] = -1234.0; 920 _par_last_gc_worker_times_ms[i] = -1234.0; 921 _par_last_gc_worker_other_times_ms[i] = -1234.0; 922 } 923 #endif 924 925 // This is initialized to zero here and is set during the evacuation 926 // pause if we actually waited for the root region scanning to finish. 927 _root_region_scan_wait_time_ms = 0.0; 928 929 _last_gc_was_young = false; 930 931 // do that for any other surv rate groups 932 _short_lived_surv_rate_group->stop_adding_regions(); 933 _survivors_age_table.clear(); 934 935 assert( verify_young_ages(), "region age verification" ); 936 } 937 938 void G1CollectorPolicy::record_concurrent_mark_init_end(double 939 mark_init_elapsed_time_ms) { 940 _during_marking = true; 941 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); 942 clear_during_initial_mark_pause(); 943 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; 944 } 945 946 void G1CollectorPolicy::record_concurrent_mark_remark_start() { 947 _mark_remark_start_sec = os::elapsedTime(); 948 _during_marking = false; 949 } 950 951 void G1CollectorPolicy::record_concurrent_mark_remark_end() { 952 double end_time_sec = os::elapsedTime(); 953 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; 954 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); 955 _cur_mark_stop_world_time_ms += elapsed_time_ms; 956 _prev_collection_pause_end_ms += elapsed_time_ms; 957 958 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); 959 } 960 961 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { 962 _mark_cleanup_start_sec = os::elapsedTime(); 963 } 964 965 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { 966 _last_young_gc = true; 967 _in_marking_window = false; 968 } 969 970 void G1CollectorPolicy::record_concurrent_pause() { 971 if (_stop_world_start > 0.0) { 972 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; 973 _trace_gen0_time_data.record_yield_time(yield_ms); 974 } 975 } 976 977 void G1CollectorPolicy::record_concurrent_pause_end() { 978 } 979 980 template<class T> 981 T sum_of(T* sum_arr, int start, int n, int N) { 982 T sum = (T)0; 983 for (int i = 0; i < n; i++) { 984 int j = (start + i) % N; 985 sum += sum_arr[j]; 986 } 987 return sum; 988 } 989 990 void G1CollectorPolicy::print_par_stats(int level, 991 const char* str, 992 double* data, 993 bool showDecimals) { 994 double min = data[0], max = data[0]; 995 double total = 0.0; 996 LineBuffer buf(level); 997 buf.append("[%s (ms):", str); 998 for (uint i = 0; i < no_of_gc_threads(); ++i) { 999 double val = data[i]; 1000 if (val < min) 1001 min = val; 1002 if (val > max) 1003 max = val; 1004 total += val; 1005 if (G1Log::finest()) { 1006 if (showDecimals) { 1007 buf.append(" %.1lf", val); 1008 } else { 1009 buf.append(" %d", (int)val); 1010 } 1011 } 1012 } 1013 1014 if (G1Log::finest()) { 1015 buf.append_and_print_cr(""); 1016 } 1017 double avg = total / (double) no_of_gc_threads(); 1018 if (showDecimals) { 1019 buf.append_and_print_cr(" Min: %.1lf, Avg: %.1lf, Max: %.1lf, Diff: %.1lf, Sum: %.1lf]", 1020 min, avg, max, max - min, total); 1021 } else { 1022 buf.append_and_print_cr(" Min: %d, Avg: %d, Max: %d, Diff: %d, Sum: %d]", 1023 (int)min, (int)avg, (int)max, (int)max - (int)min, (int)total); 1024 } 1025 } 1026 1027 void G1CollectorPolicy::print_stats(int level, 1028 const char* str, 1029 double value) { 1030 LineBuffer(level).append_and_print_cr("[%s: %.1lf ms]", str, value); 1031 } 1032 1033 void G1CollectorPolicy::print_stats(int level, 1034 const char* str, 1035 double value, 1036 int workers) { 1037 LineBuffer(level).append_and_print_cr("[%s: %.1lf ms, GC Workers: %d]", str, value, workers); 1038 } 1039 1040 void G1CollectorPolicy::print_stats(int level, 1041 const char* str, 1042 int value) { 1043 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value); 1044 } 1045 1046 double G1CollectorPolicy::avg_value(double* data) { 1047 if (G1CollectedHeap::use_parallel_gc_threads()) { 1048 double ret = 0.0; 1049 for (uint i = 0; i < no_of_gc_threads(); ++i) { 1050 ret += data[i]; 1051 } 1052 return ret / (double) no_of_gc_threads(); 1053 } else { 1054 return data[0]; 1055 } 1056 } 1057 1058 double G1CollectorPolicy::max_value(double* data) { 1059 if (G1CollectedHeap::use_parallel_gc_threads()) { 1060 double ret = data[0]; 1061 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1062 if (data[i] > ret) { 1063 ret = data[i]; 1064 } 1065 } 1066 return ret; 1067 } else { 1068 return data[0]; 1069 } 1070 } 1071 1072 double G1CollectorPolicy::sum_of_values(double* data) { 1073 if (G1CollectedHeap::use_parallel_gc_threads()) { 1074 double sum = 0.0; 1075 for (uint i = 0; i < no_of_gc_threads(); i++) { 1076 sum += data[i]; 1077 } 1078 return sum; 1079 } else { 1080 return data[0]; 1081 } 1082 } 1083 1084 double G1CollectorPolicy::max_sum(double* data1, double* data2) { 1085 double ret = data1[0] + data2[0]; 1086 1087 if (G1CollectedHeap::use_parallel_gc_threads()) { 1088 for (uint i = 1; i < no_of_gc_threads(); ++i) { 1089 double data = data1[i] + data2[i]; 1090 if (data > ret) { 1091 ret = data; 1092 } 1093 } 1094 } 1095 return ret; 1096 } 1097 1098 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { 1099 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { 1100 return false; 1101 } 1102 1103 size_t marking_initiating_used_threshold = 1104 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; 1105 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); 1106 size_t alloc_byte_size = alloc_word_size * HeapWordSize; 1107 1108 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { 1109 if (gcs_are_young()) { 1110 ergo_verbose5(ErgoConcCycles, 1111 "request concurrent cycle initiation", 1112 ergo_format_reason("occupancy higher than threshold") 1113 ergo_format_byte("occupancy") 1114 ergo_format_byte("allocation request") 1115 ergo_format_byte_perc("threshold") 1116 ergo_format_str("source"), 1117 cur_used_bytes, 1118 alloc_byte_size, 1119 marking_initiating_used_threshold, 1120 (double) InitiatingHeapOccupancyPercent, 1121 source); 1122 return true; 1123 } else { 1124 ergo_verbose5(ErgoConcCycles, 1125 "do not request concurrent cycle initiation", 1126 ergo_format_reason("still doing mixed collections") 1127 ergo_format_byte("occupancy") 1128 ergo_format_byte("allocation request") 1129 ergo_format_byte_perc("threshold") 1130 ergo_format_str("source"), 1131 cur_used_bytes, 1132 alloc_byte_size, 1133 marking_initiating_used_threshold, 1134 (double) InitiatingHeapOccupancyPercent, 1135 source); 1136 } 1137 } 1138 1139 return false; 1140 } 1141 1142 // Anything below that is considered to be zero 1143 #define MIN_TIMER_GRANULARITY 0.0000001 1144 1145 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) { 1146 double end_time_sec = os::elapsedTime(); 1147 double elapsed_ms = _last_pause_time_ms; 1148 bool parallel = G1CollectedHeap::use_parallel_gc_threads(); 1149 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), 1150 "otherwise, the subtraction below does not make sense"); 1151 size_t rs_size = 1152 _cur_collection_pause_used_regions_at_start - cset_region_length(); 1153 size_t cur_used_bytes = _g1->used(); 1154 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); 1155 bool last_pause_included_initial_mark = false; 1156 bool update_stats = !_g1->evacuation_failed(); 1157 set_no_of_gc_threads(no_of_gc_threads); 1158 1159 #ifndef PRODUCT 1160 if (G1YoungSurvRateVerbose) { 1161 gclog_or_tty->print_cr(""); 1162 _short_lived_surv_rate_group->print(); 1163 // do that for any other surv rate groups too 1164 } 1165 #endif // PRODUCT 1166 1167 last_pause_included_initial_mark = during_initial_mark_pause(); 1168 if (last_pause_included_initial_mark) { 1169 record_concurrent_mark_init_end(0.0); 1170 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) { 1171 // Note: this might have already been set, if during the last 1172 // pause we decided to start a cycle but at the beginning of 1173 // this pause we decided to postpone it. That's OK. 1174 set_initiate_conc_mark_if_possible(); 1175 } 1176 1177 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, 1178 end_time_sec, false); 1179 1180 size_t freed_bytes = 1181 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; 1182 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; 1183 1184 double survival_fraction = 1185 (double)surviving_bytes/ 1186 (double)_collection_set_bytes_used_before; 1187 1188 // These values are used to update the summary information that is 1189 // displayed when TraceGen0Time is enabled, and are output as part 1190 // of the "finer" output, in the non-parallel case. 1191 1192 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); 1193 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms); 1194 double update_rs_time = avg_value(_par_last_update_rs_times_ms); 1195 double update_rs_processed_buffers = 1196 sum_of_values(_par_last_update_rs_processed_buffers); 1197 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); 1198 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); 1199 double termination_time = avg_value(_par_last_termination_times_ms); 1200 1201 double known_time = ext_root_scan_time + 1202 satb_filtering_time + 1203 update_rs_time + 1204 scan_rs_time + 1205 obj_copy_time; 1206 1207 double other_time_ms = elapsed_ms; 1208 1209 // Subtract the root region scanning wait time. It's initialized to 1210 // zero at the start of the pause. 1211 other_time_ms -= _root_region_scan_wait_time_ms; 1212 1213 if (parallel) { 1214 other_time_ms -= _cur_collection_par_time_ms; 1215 } else { 1216 other_time_ms -= known_time; 1217 } 1218 1219 // Now subtract the time taken to fix up roots in generated code 1220 other_time_ms -= _cur_collection_code_root_fixup_time_ms; 1221 1222 // Subtract the time taken to clean the card table from the 1223 // current value of "other time" 1224 other_time_ms -= _cur_clear_ct_time_ms; 1225 1226 // TraceGen0Time and TraceGen1Time summary info updating. 1227 1228 if (update_stats) { 1229 double parallel_known_time = known_time + termination_time; 1230 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time; 1231 1232 _trace_gen0_time_data.record_end_collection( 1233 elapsed_ms, other_time_ms, _root_region_scan_wait_time_ms, _cur_collection_par_time_ms, 1234 ext_root_scan_time, satb_filtering_time, update_rs_time, scan_rs_time, obj_copy_time, 1235 termination_time, parallel_other_time, _cur_clear_ct_time_ms); 1236 1237 // this is where we update the allocation rate of the application 1238 double app_time_ms = 1239 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); 1240 if (app_time_ms < MIN_TIMER_GRANULARITY) { 1241 // This usually happens due to the timer not having the required 1242 // granularity. Some Linuxes are the usual culprits. 1243 // We'll just set it to something (arbitrarily) small. 1244 app_time_ms = 1.0; 1245 } 1246 // We maintain the invariant that all objects allocated by mutator 1247 // threads will be allocated out of eden regions. So, we can use 1248 // the eden region number allocated since the previous GC to 1249 // calculate the application's allocate rate. The only exception 1250 // to that is humongous objects that are allocated separately. But 1251 // given that humongous object allocations do not really affect 1252 // either the pause's duration nor when the next pause will take 1253 // place we can safely ignore them here. 1254 uint regions_allocated = eden_cset_region_length(); 1255 double alloc_rate_ms = (double) regions_allocated / app_time_ms; 1256 _alloc_rate_ms_seq->add(alloc_rate_ms); 1257 1258 double interval_ms = 1259 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; 1260 update_recent_gc_times(end_time_sec, elapsed_ms); 1261 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; 1262 if (recent_avg_pause_time_ratio() < 0.0 || 1263 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { 1264 #ifndef PRODUCT 1265 // Dump info to allow post-facto debugging 1266 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); 1267 gclog_or_tty->print_cr("-------------------------------------------"); 1268 gclog_or_tty->print_cr("Recent GC Times (ms):"); 1269 _recent_gc_times_ms->dump(); 1270 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); 1271 _recent_prev_end_times_for_all_gcs_sec->dump(); 1272 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", 1273 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); 1274 // In debug mode, terminate the JVM if the user wants to debug at this point. 1275 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); 1276 #endif // !PRODUCT 1277 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in 1278 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. 1279 if (_recent_avg_pause_time_ratio < 0.0) { 1280 _recent_avg_pause_time_ratio = 0.0; 1281 } else { 1282 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); 1283 _recent_avg_pause_time_ratio = 1.0; 1284 } 1285 } 1286 } 1287 1288 if (G1Log::finer()) { 1289 bool print_marking_info = 1290 _g1->mark_in_progress() && !last_pause_included_initial_mark; 1291 1292 gclog_or_tty->print_cr("%s, %1.8lf secs]", 1293 (last_pause_included_initial_mark) ? " (initial-mark)" : "", 1294 elapsed_ms / 1000.0); 1295 1296 if (_root_region_scan_wait_time_ms > 0.0) { 1297 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms); 1298 } 1299 if (parallel) { 1300 print_stats(1, "Parallel Time", _cur_collection_par_time_ms, no_of_gc_threads); 1301 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms); 1302 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms); 1303 if (print_marking_info) { 1304 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms); 1305 } 1306 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); 1307 if (G1Log::finest()) { 1308 print_par_stats(3, "Processed Buffers", _par_last_update_rs_processed_buffers, 1309 false /* showDecimals */); 1310 } 1311 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); 1312 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); 1313 print_par_stats(2, "Termination", _par_last_termination_times_ms); 1314 if (G1Log::finest()) { 1315 print_par_stats(3, "Termination Attempts", _par_last_termination_attempts, 1316 false /* showDecimals */); 1317 } 1318 1319 for (int i = 0; i < _parallel_gc_threads; i++) { 1320 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - 1321 _par_last_gc_worker_start_times_ms[i]; 1322 1323 double worker_known_time = _par_last_ext_root_scan_times_ms[i] + 1324 _par_last_satb_filtering_times_ms[i] + 1325 _par_last_update_rs_times_ms[i] + 1326 _par_last_scan_rs_times_ms[i] + 1327 _par_last_obj_copy_times_ms[i] + 1328 _par_last_termination_times_ms[i]; 1329 1330 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] - 1331 worker_known_time; 1332 } 1333 1334 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms); 1335 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms); 1336 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms); 1337 } else { 1338 print_stats(1, "Ext Root Scanning", ext_root_scan_time); 1339 if (print_marking_info) { 1340 print_stats(1, "SATB Filtering", satb_filtering_time); 1341 } 1342 print_stats(1, "Update RS", update_rs_time); 1343 if (G1Log::finest()) { 1344 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers); 1345 } 1346 print_stats(1, "Scan RS", scan_rs_time); 1347 print_stats(1, "Object Copying", obj_copy_time); 1348 } 1349 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms); 1350 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); 1351 #ifndef PRODUCT 1352 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms); 1353 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms); 1354 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms); 1355 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms); 1356 if (_num_cc_clears > 0) { 1357 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears)); 1358 } 1359 #endif 1360 print_stats(1, "Other", other_time_ms); 1361 print_stats(2, "Choose CSet", 1362 (_recorded_young_cset_choice_time_ms + 1363 _recorded_non_young_cset_choice_time_ms)); 1364 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms); 1365 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms); 1366 print_stats(2, "Free CSet", 1367 (_recorded_young_free_cset_time_ms + 1368 _recorded_non_young_free_cset_time_ms)); 1369 } 1370 1371 bool new_in_marking_window = _in_marking_window; 1372 bool new_in_marking_window_im = false; 1373 if (during_initial_mark_pause()) { 1374 new_in_marking_window = true; 1375 new_in_marking_window_im = true; 1376 } 1377 1378 if (_last_young_gc) { 1379 // This is supposed to to be the "last young GC" before we start 1380 // doing mixed GCs. Here we decide whether to start mixed GCs or not. 1381 1382 if (!last_pause_included_initial_mark) { 1383 if (next_gc_should_be_mixed("start mixed GCs", 1384 "do not start mixed GCs")) { 1385 set_gcs_are_young(false); 1386 } 1387 } else { 1388 ergo_verbose0(ErgoMixedGCs, 1389 "do not start mixed GCs", 1390 ergo_format_reason("concurrent cycle is about to start")); 1391 } 1392 _last_young_gc = false; 1393 } 1394 1395 if (!_last_gc_was_young) { 1396 // This is a mixed GC. Here we decide whether to continue doing 1397 // mixed GCs or not. 1398 1399 if (!next_gc_should_be_mixed("continue mixed GCs", 1400 "do not continue mixed GCs")) { 1401 set_gcs_are_young(true); 1402 } 1403 } 1404 1405 _short_lived_surv_rate_group->start_adding_regions(); 1406 // do that for any other surv rate groupsx 1407 1408 if (update_stats) { 1409 double pause_time_ms = elapsed_ms; 1410 1411 size_t diff = 0; 1412 if (_max_pending_cards >= _pending_cards) { 1413 diff = _max_pending_cards - _pending_cards; 1414 } 1415 _pending_card_diff_seq->add((double) diff); 1416 1417 double cost_per_card_ms = 0.0; 1418 if (_pending_cards > 0) { 1419 cost_per_card_ms = update_rs_time / (double) _pending_cards; 1420 _cost_per_card_ms_seq->add(cost_per_card_ms); 1421 } 1422 1423 size_t cards_scanned = _g1->cards_scanned(); 1424 1425 double cost_per_entry_ms = 0.0; 1426 if (cards_scanned > 10) { 1427 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; 1428 if (_last_gc_was_young) { 1429 _cost_per_entry_ms_seq->add(cost_per_entry_ms); 1430 } else { 1431 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); 1432 } 1433 } 1434 1435 if (_max_rs_lengths > 0) { 1436 double cards_per_entry_ratio = 1437 (double) cards_scanned / (double) _max_rs_lengths; 1438 if (_last_gc_was_young) { 1439 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1440 } else { 1441 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); 1442 } 1443 } 1444 1445 // This is defensive. For a while _max_rs_lengths could get 1446 // smaller than _recorded_rs_lengths which was causing 1447 // rs_length_diff to get very large and mess up the RSet length 1448 // predictions. The reason was unsafe concurrent updates to the 1449 // _inc_cset_recorded_rs_lengths field which the code below guards 1450 // against (see CR 7118202). This bug has now been fixed (see CR 1451 // 7119027). However, I'm still worried that 1452 // _inc_cset_recorded_rs_lengths might still end up somewhat 1453 // inaccurate. The concurrent refinement thread calculates an 1454 // RSet's length concurrently with other CR threads updating it 1455 // which might cause it to calculate the length incorrectly (if, 1456 // say, it's in mid-coarsening). So I'll leave in the defensive 1457 // conditional below just in case. 1458 size_t rs_length_diff = 0; 1459 if (_max_rs_lengths > _recorded_rs_lengths) { 1460 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; 1461 } 1462 _rs_length_diff_seq->add((double) rs_length_diff); 1463 1464 size_t copied_bytes = surviving_bytes; 1465 double cost_per_byte_ms = 0.0; 1466 if (copied_bytes > 0) { 1467 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; 1468 if (_in_marking_window) { 1469 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); 1470 } else { 1471 _cost_per_byte_ms_seq->add(cost_per_byte_ms); 1472 } 1473 } 1474 1475 double all_other_time_ms = pause_time_ms - 1476 (update_rs_time + scan_rs_time + obj_copy_time + termination_time); 1477 1478 double young_other_time_ms = 0.0; 1479 if (young_cset_region_length() > 0) { 1480 young_other_time_ms = 1481 _recorded_young_cset_choice_time_ms + 1482 _recorded_young_free_cset_time_ms; 1483 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / 1484 (double) young_cset_region_length()); 1485 } 1486 double non_young_other_time_ms = 0.0; 1487 if (old_cset_region_length() > 0) { 1488 non_young_other_time_ms = 1489 _recorded_non_young_cset_choice_time_ms + 1490 _recorded_non_young_free_cset_time_ms; 1491 1492 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / 1493 (double) old_cset_region_length()); 1494 } 1495 1496 double constant_other_time_ms = all_other_time_ms - 1497 (young_other_time_ms + non_young_other_time_ms); 1498 _constant_other_time_ms_seq->add(constant_other_time_ms); 1499 1500 double survival_ratio = 0.0; 1501 if (_bytes_in_collection_set_before_gc > 0) { 1502 survival_ratio = (double) _bytes_copied_during_gc / 1503 (double) _bytes_in_collection_set_before_gc; 1504 } 1505 1506 _pending_cards_seq->add((double) _pending_cards); 1507 _rs_lengths_seq->add((double) _max_rs_lengths); 1508 } 1509 1510 _in_marking_window = new_in_marking_window; 1511 _in_marking_window_im = new_in_marking_window_im; 1512 _free_regions_at_end_of_collection = _g1->free_regions(); 1513 update_young_list_target_length(); 1514 1515 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. 1516 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; 1517 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms); 1518 1519 _collectionSetChooser->verify(); 1520 } 1521 1522 #define EXT_SIZE_FORMAT "%.1f%s" 1523 #define EXT_SIZE_PARAMS(bytes) \ 1524 byte_size_in_proper_unit((double)(bytes)), \ 1525 proper_unit_for_byte_size((bytes)) 1526 1527 void G1CollectorPolicy::print_heap_transition() { 1528 if (G1Log::finer()) { 1529 YoungList* young_list = _g1->young_list(); 1530 size_t eden_bytes = young_list->eden_used_bytes(); 1531 size_t survivor_bytes = young_list->survivor_used_bytes(); 1532 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes; 1533 size_t used = _g1->used(); 1534 size_t capacity = _g1->capacity(); 1535 size_t eden_capacity = 1536 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes; 1537 1538 gclog_or_tty->print_cr( 1539 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") " 1540 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" " 1541 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->" 1542 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]", 1543 EXT_SIZE_PARAMS(_eden_bytes_before_gc), 1544 EXT_SIZE_PARAMS(_prev_eden_capacity), 1545 EXT_SIZE_PARAMS(eden_bytes), 1546 EXT_SIZE_PARAMS(eden_capacity), 1547 EXT_SIZE_PARAMS(_survivor_bytes_before_gc), 1548 EXT_SIZE_PARAMS(survivor_bytes), 1549 EXT_SIZE_PARAMS(used_before_gc), 1550 EXT_SIZE_PARAMS(_capacity_before_gc), 1551 EXT_SIZE_PARAMS(used), 1552 EXT_SIZE_PARAMS(capacity)); 1553 1554 _prev_eden_capacity = eden_capacity; 1555 } else if (G1Log::fine()) { 1556 _g1->print_size_transition(gclog_or_tty, 1557 _cur_collection_pause_used_at_start_bytes, 1558 _g1->used(), _g1->capacity()); 1559 } 1560 } 1561 1562 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, 1563 double update_rs_processed_buffers, 1564 double goal_ms) { 1565 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 1566 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); 1567 1568 if (G1UseAdaptiveConcRefinement) { 1569 const int k_gy = 3, k_gr = 6; 1570 const double inc_k = 1.1, dec_k = 0.9; 1571 1572 int g = cg1r->green_zone(); 1573 if (update_rs_time > goal_ms) { 1574 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. 1575 } else { 1576 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { 1577 g = (int)MAX2(g * inc_k, g + 1.0); 1578 } 1579 } 1580 // Change the refinement threads params 1581 cg1r->set_green_zone(g); 1582 cg1r->set_yellow_zone(g * k_gy); 1583 cg1r->set_red_zone(g * k_gr); 1584 cg1r->reinitialize_threads(); 1585 1586 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); 1587 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, 1588 cg1r->yellow_zone()); 1589 // Change the barrier params 1590 dcqs.set_process_completed_threshold(processing_threshold); 1591 dcqs.set_max_completed_queue(cg1r->red_zone()); 1592 } 1593 1594 int curr_queue_size = dcqs.completed_buffers_num(); 1595 if (curr_queue_size >= cg1r->yellow_zone()) { 1596 dcqs.set_completed_queue_padding(curr_queue_size); 1597 } else { 1598 dcqs.set_completed_queue_padding(0); 1599 } 1600 dcqs.notify_if_necessary(); 1601 } 1602 1603 double 1604 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { 1605 size_t rs_length = predict_rs_length_diff(); 1606 size_t card_num; 1607 if (gcs_are_young()) { 1608 card_num = predict_young_card_num(rs_length); 1609 } else { 1610 card_num = predict_non_young_card_num(rs_length); 1611 } 1612 return predict_base_elapsed_time_ms(pending_cards, card_num); 1613 } 1614 1615 double 1616 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, 1617 size_t scanned_cards) { 1618 return 1619 predict_rs_update_time_ms(pending_cards) + 1620 predict_rs_scan_time_ms(scanned_cards) + 1621 predict_constant_other_time_ms(); 1622 } 1623 1624 double 1625 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, 1626 bool young) { 1627 size_t rs_length = hr->rem_set()->occupied(); 1628 size_t card_num; 1629 if (gcs_are_young()) { 1630 card_num = predict_young_card_num(rs_length); 1631 } else { 1632 card_num = predict_non_young_card_num(rs_length); 1633 } 1634 size_t bytes_to_copy = predict_bytes_to_copy(hr); 1635 1636 double region_elapsed_time_ms = 1637 predict_rs_scan_time_ms(card_num) + 1638 predict_object_copy_time_ms(bytes_to_copy); 1639 1640 if (young) 1641 region_elapsed_time_ms += predict_young_other_time_ms(1); 1642 else 1643 region_elapsed_time_ms += predict_non_young_other_time_ms(1); 1644 1645 return region_elapsed_time_ms; 1646 } 1647 1648 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { 1649 size_t bytes_to_copy; 1650 if (hr->is_marked()) 1651 bytes_to_copy = hr->max_live_bytes(); 1652 else { 1653 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); 1654 int age = hr->age_in_surv_rate_group(); 1655 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); 1656 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); 1657 } 1658 return bytes_to_copy; 1659 } 1660 1661 void 1662 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, 1663 uint survivor_cset_region_length) { 1664 _eden_cset_region_length = eden_cset_region_length; 1665 _survivor_cset_region_length = survivor_cset_region_length; 1666 _old_cset_region_length = 0; 1667 } 1668 1669 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { 1670 _recorded_rs_lengths = rs_lengths; 1671 } 1672 1673 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, 1674 double elapsed_ms) { 1675 _recent_gc_times_ms->add(elapsed_ms); 1676 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); 1677 _prev_collection_pause_end_ms = end_time_sec * 1000.0; 1678 } 1679 1680 size_t G1CollectorPolicy::expansion_amount() { 1681 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; 1682 double threshold = _gc_overhead_perc; 1683 if (recent_gc_overhead > threshold) { 1684 // We will double the existing space, or take 1685 // G1ExpandByPercentOfAvailable % of the available expansion 1686 // space, whichever is smaller, bounded below by a minimum 1687 // expansion (unless that's all that's left.) 1688 const size_t min_expand_bytes = 1*M; 1689 size_t reserved_bytes = _g1->max_capacity(); 1690 size_t committed_bytes = _g1->capacity(); 1691 size_t uncommitted_bytes = reserved_bytes - committed_bytes; 1692 size_t expand_bytes; 1693 size_t expand_bytes_via_pct = 1694 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; 1695 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); 1696 expand_bytes = MAX2(expand_bytes, min_expand_bytes); 1697 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); 1698 1699 ergo_verbose5(ErgoHeapSizing, 1700 "attempt heap expansion", 1701 ergo_format_reason("recent GC overhead higher than " 1702 "threshold after GC") 1703 ergo_format_perc("recent GC overhead") 1704 ergo_format_perc("threshold") 1705 ergo_format_byte("uncommitted") 1706 ergo_format_byte_perc("calculated expansion amount"), 1707 recent_gc_overhead, threshold, 1708 uncommitted_bytes, 1709 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); 1710 1711 return expand_bytes; 1712 } else { 1713 return 0; 1714 } 1715 } 1716 1717 class CountCSClosure: public HeapRegionClosure { 1718 G1CollectorPolicy* _g1_policy; 1719 public: 1720 CountCSClosure(G1CollectorPolicy* g1_policy) : 1721 _g1_policy(g1_policy) {} 1722 bool doHeapRegion(HeapRegion* r) { 1723 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); 1724 return false; 1725 } 1726 }; 1727 1728 void G1CollectorPolicy::count_CS_bytes_used() { 1729 CountCSClosure cs_closure(this); 1730 _g1->collection_set_iterate(&cs_closure); 1731 } 1732 1733 void G1CollectorPolicy::print_tracing_info() const { 1734 _trace_gen0_time_data.print(); 1735 _trace_gen1_time_data.print(); 1736 } 1737 1738 void G1CollectorPolicy::print_yg_surv_rate_info() const { 1739 #ifndef PRODUCT 1740 _short_lived_surv_rate_group->print_surv_rate_summary(); 1741 // add this call for any other surv rate groups 1742 #endif // PRODUCT 1743 } 1744 1745 #ifndef PRODUCT 1746 // for debugging, bit of a hack... 1747 static char* 1748 region_num_to_mbs(int length) { 1749 static char buffer[64]; 1750 double bytes = (double) (length * HeapRegion::GrainBytes); 1751 double mbs = bytes / (double) (1024 * 1024); 1752 sprintf(buffer, "%7.2lfMB", mbs); 1753 return buffer; 1754 } 1755 #endif // PRODUCT 1756 1757 uint G1CollectorPolicy::max_regions(int purpose) { 1758 switch (purpose) { 1759 case GCAllocForSurvived: 1760 return _max_survivor_regions; 1761 case GCAllocForTenured: 1762 return REGIONS_UNLIMITED; 1763 default: 1764 ShouldNotReachHere(); 1765 return REGIONS_UNLIMITED; 1766 }; 1767 } 1768 1769 void G1CollectorPolicy::update_max_gc_locker_expansion() { 1770 uint expansion_region_num = 0; 1771 if (GCLockerEdenExpansionPercent > 0) { 1772 double perc = (double) GCLockerEdenExpansionPercent / 100.0; 1773 double expansion_region_num_d = perc * (double) _young_list_target_length; 1774 // We use ceiling so that if expansion_region_num_d is > 0.0 (but 1775 // less than 1.0) we'll get 1. 1776 expansion_region_num = (uint) ceil(expansion_region_num_d); 1777 } else { 1778 assert(expansion_region_num == 0, "sanity"); 1779 } 1780 _young_list_max_length = _young_list_target_length + expansion_region_num; 1781 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); 1782 } 1783 1784 // Calculates survivor space parameters. 1785 void G1CollectorPolicy::update_survivors_policy() { 1786 double max_survivor_regions_d = 1787 (double) _young_list_target_length / (double) SurvivorRatio; 1788 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but 1789 // smaller than 1.0) we'll get 1. 1790 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); 1791 1792 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( 1793 HeapRegion::GrainWords * _max_survivor_regions); 1794 } 1795 1796 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( 1797 GCCause::Cause gc_cause) { 1798 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1799 if (!during_cycle) { 1800 ergo_verbose1(ErgoConcCycles, 1801 "request concurrent cycle initiation", 1802 ergo_format_reason("requested by GC cause") 1803 ergo_format_str("GC cause"), 1804 GCCause::to_string(gc_cause)); 1805 set_initiate_conc_mark_if_possible(); 1806 return true; 1807 } else { 1808 ergo_verbose1(ErgoConcCycles, 1809 "do not request concurrent cycle initiation", 1810 ergo_format_reason("concurrent cycle already in progress") 1811 ergo_format_str("GC cause"), 1812 GCCause::to_string(gc_cause)); 1813 return false; 1814 } 1815 } 1816 1817 void 1818 G1CollectorPolicy::decide_on_conc_mark_initiation() { 1819 // We are about to decide on whether this pause will be an 1820 // initial-mark pause. 1821 1822 // First, during_initial_mark_pause() should not be already set. We 1823 // will set it here if we have to. However, it should be cleared by 1824 // the end of the pause (it's only set for the duration of an 1825 // initial-mark pause). 1826 assert(!during_initial_mark_pause(), "pre-condition"); 1827 1828 if (initiate_conc_mark_if_possible()) { 1829 // We had noticed on a previous pause that the heap occupancy has 1830 // gone over the initiating threshold and we should start a 1831 // concurrent marking cycle. So we might initiate one. 1832 1833 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); 1834 if (!during_cycle) { 1835 // The concurrent marking thread is not "during a cycle", i.e., 1836 // it has completed the last one. So we can go ahead and 1837 // initiate a new cycle. 1838 1839 set_during_initial_mark_pause(); 1840 // We do not allow mixed GCs during marking. 1841 if (!gcs_are_young()) { 1842 set_gcs_are_young(true); 1843 ergo_verbose0(ErgoMixedGCs, 1844 "end mixed GCs", 1845 ergo_format_reason("concurrent cycle is about to start")); 1846 } 1847 1848 // And we can now clear initiate_conc_mark_if_possible() as 1849 // we've already acted on it. 1850 clear_initiate_conc_mark_if_possible(); 1851 1852 ergo_verbose0(ErgoConcCycles, 1853 "initiate concurrent cycle", 1854 ergo_format_reason("concurrent cycle initiation requested")); 1855 } else { 1856 // The concurrent marking thread is still finishing up the 1857 // previous cycle. If we start one right now the two cycles 1858 // overlap. In particular, the concurrent marking thread might 1859 // be in the process of clearing the next marking bitmap (which 1860 // we will use for the next cycle if we start one). Starting a 1861 // cycle now will be bad given that parts of the marking 1862 // information might get cleared by the marking thread. And we 1863 // cannot wait for the marking thread to finish the cycle as it 1864 // periodically yields while clearing the next marking bitmap 1865 // and, if it's in a yield point, it's waiting for us to 1866 // finish. So, at this point we will not start a cycle and we'll 1867 // let the concurrent marking thread complete the last one. 1868 ergo_verbose0(ErgoConcCycles, 1869 "do not initiate concurrent cycle", 1870 ergo_format_reason("concurrent cycle already in progress")); 1871 } 1872 } 1873 } 1874 1875 class KnownGarbageClosure: public HeapRegionClosure { 1876 G1CollectedHeap* _g1h; 1877 CollectionSetChooser* _hrSorted; 1878 1879 public: 1880 KnownGarbageClosure(CollectionSetChooser* hrSorted) : 1881 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { } 1882 1883 bool doHeapRegion(HeapRegion* r) { 1884 // We only include humongous regions in collection 1885 // sets when concurrent mark shows that their contained object is 1886 // unreachable. 1887 1888 // Do we have any marking information for this region? 1889 if (r->is_marked()) { 1890 // We will skip any region that's currently used as an old GC 1891 // alloc region (we should not consider those for collection 1892 // before we fill them up). 1893 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1894 _hrSorted->add_region(r); 1895 } 1896 } 1897 return false; 1898 } 1899 }; 1900 1901 class ParKnownGarbageHRClosure: public HeapRegionClosure { 1902 G1CollectedHeap* _g1h; 1903 CollectionSetChooser* _hrSorted; 1904 uint _marked_regions_added; 1905 size_t _reclaimable_bytes_added; 1906 uint _chunk_size; 1907 uint _cur_chunk_idx; 1908 uint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) 1909 1910 void get_new_chunk() { 1911 _cur_chunk_idx = _hrSorted->claim_array_chunk(_chunk_size); 1912 _cur_chunk_end = _cur_chunk_idx + _chunk_size; 1913 } 1914 void add_region(HeapRegion* r) { 1915 if (_cur_chunk_idx == _cur_chunk_end) { 1916 get_new_chunk(); 1917 } 1918 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); 1919 _hrSorted->set_region(_cur_chunk_idx, r); 1920 _marked_regions_added++; 1921 _reclaimable_bytes_added += r->reclaimable_bytes(); 1922 _cur_chunk_idx++; 1923 } 1924 1925 public: 1926 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, 1927 uint chunk_size) : 1928 _g1h(G1CollectedHeap::heap()), 1929 _hrSorted(hrSorted), _chunk_size(chunk_size), 1930 _marked_regions_added(0), _reclaimable_bytes_added(0), 1931 _cur_chunk_idx(0), _cur_chunk_end(0) { } 1932 1933 bool doHeapRegion(HeapRegion* r) { 1934 // Do we have any marking information for this region? 1935 if (r->is_marked()) { 1936 // We will skip any region that's currently used as an old GC 1937 // alloc region (we should not consider those for collection 1938 // before we fill them up). 1939 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { 1940 add_region(r); 1941 } 1942 } 1943 return false; 1944 } 1945 uint marked_regions_added() { return _marked_regions_added; } 1946 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; } 1947 }; 1948 1949 class ParKnownGarbageTask: public AbstractGangTask { 1950 CollectionSetChooser* _hrSorted; 1951 uint _chunk_size; 1952 G1CollectedHeap* _g1; 1953 public: 1954 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) : 1955 AbstractGangTask("ParKnownGarbageTask"), 1956 _hrSorted(hrSorted), _chunk_size(chunk_size), 1957 _g1(G1CollectedHeap::heap()) { } 1958 1959 void work(uint worker_id) { 1960 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); 1961 1962 // Back to zero for the claim value. 1963 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id, 1964 _g1->workers()->active_workers(), 1965 HeapRegion::InitialClaimValue); 1966 uint regions_added = parKnownGarbageCl.marked_regions_added(); 1967 size_t reclaimable_bytes_added = 1968 parKnownGarbageCl.reclaimable_bytes_added(); 1969 _hrSorted->update_totals(regions_added, reclaimable_bytes_added); 1970 } 1971 }; 1972 1973 void 1974 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) { 1975 _collectionSetChooser->clear(); 1976 1977 uint region_num = _g1->n_regions(); 1978 if (G1CollectedHeap::use_parallel_gc_threads()) { 1979 const uint OverpartitionFactor = 4; 1980 uint WorkUnit; 1981 // The use of MinChunkSize = 8 in the original code 1982 // causes some assertion failures when the total number of 1983 // region is less than 8. The code here tries to fix that. 1984 // Should the original code also be fixed? 1985 if (no_of_gc_threads > 0) { 1986 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U); 1987 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor), 1988 MinWorkUnit); 1989 } else { 1990 assert(no_of_gc_threads > 0, 1991 "The active gc workers should be greater than 0"); 1992 // In a product build do something reasonable to avoid a crash. 1993 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U); 1994 WorkUnit = 1995 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor), 1996 MinWorkUnit); 1997 } 1998 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(), 1999 WorkUnit); 2000 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, 2001 (int) WorkUnit); 2002 _g1->workers()->run_task(&parKnownGarbageTask); 2003 2004 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 2005 "sanity check"); 2006 } else { 2007 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); 2008 _g1->heap_region_iterate(&knownGarbagecl); 2009 } 2010 2011 _collectionSetChooser->sort_regions(); 2012 2013 double end_sec = os::elapsedTime(); 2014 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; 2015 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); 2016 _cur_mark_stop_world_time_ms += elapsed_time_ms; 2017 _prev_collection_pause_end_ms += elapsed_time_ms; 2018 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true); 2019 } 2020 2021 // Add the heap region at the head of the non-incremental collection set 2022 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { 2023 assert(_inc_cset_build_state == Active, "Precondition"); 2024 assert(!hr->is_young(), "non-incremental add of young region"); 2025 2026 assert(!hr->in_collection_set(), "should not already be in the CSet"); 2027 hr->set_in_collection_set(true); 2028 hr->set_next_in_collection_set(_collection_set); 2029 _collection_set = hr; 2030 _collection_set_bytes_used_before += hr->used(); 2031 _g1->register_region_with_in_cset_fast_test(hr); 2032 size_t rs_length = hr->rem_set()->occupied(); 2033 _recorded_rs_lengths += rs_length; 2034 _old_cset_region_length += 1; 2035 } 2036 2037 // Initialize the per-collection-set information 2038 void G1CollectorPolicy::start_incremental_cset_building() { 2039 assert(_inc_cset_build_state == Inactive, "Precondition"); 2040 2041 _inc_cset_head = NULL; 2042 _inc_cset_tail = NULL; 2043 _inc_cset_bytes_used_before = 0; 2044 2045 _inc_cset_max_finger = 0; 2046 _inc_cset_recorded_rs_lengths = 0; 2047 _inc_cset_recorded_rs_lengths_diffs = 0; 2048 _inc_cset_predicted_elapsed_time_ms = 0.0; 2049 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2050 _inc_cset_build_state = Active; 2051 } 2052 2053 void G1CollectorPolicy::finalize_incremental_cset_building() { 2054 assert(_inc_cset_build_state == Active, "Precondition"); 2055 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2056 2057 // The two "main" fields, _inc_cset_recorded_rs_lengths and 2058 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread 2059 // that adds a new region to the CSet. Further updates by the 2060 // concurrent refinement thread that samples the young RSet lengths 2061 // are accumulated in the *_diffs fields. Here we add the diffs to 2062 // the "main" fields. 2063 2064 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { 2065 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; 2066 } else { 2067 // This is defensive. The diff should in theory be always positive 2068 // as RSets can only grow between GCs. However, given that we 2069 // sample their size concurrently with other threads updating them 2070 // it's possible that we might get the wrong size back, which 2071 // could make the calculations somewhat inaccurate. 2072 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); 2073 if (_inc_cset_recorded_rs_lengths >= diffs) { 2074 _inc_cset_recorded_rs_lengths -= diffs; 2075 } else { 2076 _inc_cset_recorded_rs_lengths = 0; 2077 } 2078 } 2079 _inc_cset_predicted_elapsed_time_ms += 2080 _inc_cset_predicted_elapsed_time_ms_diffs; 2081 2082 _inc_cset_recorded_rs_lengths_diffs = 0; 2083 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; 2084 } 2085 2086 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { 2087 // This routine is used when: 2088 // * adding survivor regions to the incremental cset at the end of an 2089 // evacuation pause, 2090 // * adding the current allocation region to the incremental cset 2091 // when it is retired, and 2092 // * updating existing policy information for a region in the 2093 // incremental cset via young list RSet sampling. 2094 // Therefore this routine may be called at a safepoint by the 2095 // VM thread, or in-between safepoints by mutator threads (when 2096 // retiring the current allocation region) or a concurrent 2097 // refine thread (RSet sampling). 2098 2099 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2100 size_t used_bytes = hr->used(); 2101 _inc_cset_recorded_rs_lengths += rs_length; 2102 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; 2103 _inc_cset_bytes_used_before += used_bytes; 2104 2105 // Cache the values we have added to the aggregated informtion 2106 // in the heap region in case we have to remove this region from 2107 // the incremental collection set, or it is updated by the 2108 // rset sampling code 2109 hr->set_recorded_rs_length(rs_length); 2110 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); 2111 } 2112 2113 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, 2114 size_t new_rs_length) { 2115 // Update the CSet information that is dependent on the new RS length 2116 assert(hr->is_young(), "Precondition"); 2117 assert(!SafepointSynchronize::is_at_safepoint(), 2118 "should not be at a safepoint"); 2119 2120 // We could have updated _inc_cset_recorded_rs_lengths and 2121 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do 2122 // that atomically, as this code is executed by a concurrent 2123 // refinement thread, potentially concurrently with a mutator thread 2124 // allocating a new region and also updating the same fields. To 2125 // avoid the atomic operations we accumulate these updates on two 2126 // separate fields (*_diffs) and we'll just add them to the "main" 2127 // fields at the start of a GC. 2128 2129 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); 2130 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; 2131 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; 2132 2133 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); 2134 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true); 2135 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; 2136 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; 2137 2138 hr->set_recorded_rs_length(new_rs_length); 2139 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); 2140 } 2141 2142 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { 2143 assert(hr->is_young(), "invariant"); 2144 assert(hr->young_index_in_cset() > -1, "should have already been set"); 2145 assert(_inc_cset_build_state == Active, "Precondition"); 2146 2147 // We need to clear and set the cached recorded/cached collection set 2148 // information in the heap region here (before the region gets added 2149 // to the collection set). An individual heap region's cached values 2150 // are calculated, aggregated with the policy collection set info, 2151 // and cached in the heap region here (initially) and (subsequently) 2152 // by the Young List sampling code. 2153 2154 size_t rs_length = hr->rem_set()->occupied(); 2155 add_to_incremental_cset_info(hr, rs_length); 2156 2157 HeapWord* hr_end = hr->end(); 2158 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); 2159 2160 assert(!hr->in_collection_set(), "invariant"); 2161 hr->set_in_collection_set(true); 2162 assert( hr->next_in_collection_set() == NULL, "invariant"); 2163 2164 _g1->register_region_with_in_cset_fast_test(hr); 2165 } 2166 2167 // Add the region at the RHS of the incremental cset 2168 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { 2169 // We should only ever be appending survivors at the end of a pause 2170 assert( hr->is_survivor(), "Logic"); 2171 2172 // Do the 'common' stuff 2173 add_region_to_incremental_cset_common(hr); 2174 2175 // Now add the region at the right hand side 2176 if (_inc_cset_tail == NULL) { 2177 assert(_inc_cset_head == NULL, "invariant"); 2178 _inc_cset_head = hr; 2179 } else { 2180 _inc_cset_tail->set_next_in_collection_set(hr); 2181 } 2182 _inc_cset_tail = hr; 2183 } 2184 2185 // Add the region to the LHS of the incremental cset 2186 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { 2187 // Survivors should be added to the RHS at the end of a pause 2188 assert(!hr->is_survivor(), "Logic"); 2189 2190 // Do the 'common' stuff 2191 add_region_to_incremental_cset_common(hr); 2192 2193 // Add the region at the left hand side 2194 hr->set_next_in_collection_set(_inc_cset_head); 2195 if (_inc_cset_head == NULL) { 2196 assert(_inc_cset_tail == NULL, "Invariant"); 2197 _inc_cset_tail = hr; 2198 } 2199 _inc_cset_head = hr; 2200 } 2201 2202 #ifndef PRODUCT 2203 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { 2204 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); 2205 2206 st->print_cr("\nCollection_set:"); 2207 HeapRegion* csr = list_head; 2208 while (csr != NULL) { 2209 HeapRegion* next = csr->next_in_collection_set(); 2210 assert(csr->in_collection_set(), "bad CS"); 2211 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 2212 HR_FORMAT_PARAMS(csr), 2213 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(), 2214 csr->age_in_surv_rate_group_cond()); 2215 csr = next; 2216 } 2217 } 2218 #endif // !PRODUCT 2219 2220 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, 2221 const char* false_action_str) { 2222 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2223 if (cset_chooser->is_empty()) { 2224 ergo_verbose0(ErgoMixedGCs, 2225 false_action_str, 2226 ergo_format_reason("candidate old regions not available")); 2227 return false; 2228 } 2229 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); 2230 size_t capacity_bytes = _g1->capacity(); 2231 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; 2232 double threshold = (double) G1HeapWastePercent; 2233 if (perc < threshold) { 2234 ergo_verbose4(ErgoMixedGCs, 2235 false_action_str, 2236 ergo_format_reason("reclaimable percentage lower than threshold") 2237 ergo_format_region("candidate old regions") 2238 ergo_format_byte_perc("reclaimable") 2239 ergo_format_perc("threshold"), 2240 cset_chooser->remaining_regions(), 2241 reclaimable_bytes, perc, threshold); 2242 return false; 2243 } 2244 2245 ergo_verbose4(ErgoMixedGCs, 2246 true_action_str, 2247 ergo_format_reason("candidate old regions available") 2248 ergo_format_region("candidate old regions") 2249 ergo_format_byte_perc("reclaimable") 2250 ergo_format_perc("threshold"), 2251 cset_chooser->remaining_regions(), 2252 reclaimable_bytes, perc, threshold); 2253 return true; 2254 } 2255 2256 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) { 2257 // Set this here - in case we're not doing young collections. 2258 double non_young_start_time_sec = os::elapsedTime(); 2259 2260 YoungList* young_list = _g1->young_list(); 2261 finalize_incremental_cset_building(); 2262 2263 guarantee(target_pause_time_ms > 0.0, 2264 err_msg("target_pause_time_ms = %1.6lf should be positive", 2265 target_pause_time_ms)); 2266 guarantee(_collection_set == NULL, "Precondition"); 2267 2268 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); 2269 double predicted_pause_time_ms = base_time_ms; 2270 double time_remaining_ms = target_pause_time_ms - base_time_ms; 2271 2272 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2273 "start choosing CSet", 2274 ergo_format_ms("predicted base time") 2275 ergo_format_ms("remaining time") 2276 ergo_format_ms("target pause time"), 2277 base_time_ms, time_remaining_ms, target_pause_time_ms); 2278 2279 HeapRegion* hr; 2280 double young_start_time_sec = os::elapsedTime(); 2281 2282 _collection_set_bytes_used_before = 0; 2283 _last_gc_was_young = gcs_are_young() ? true : false; 2284 2285 if (_last_gc_was_young) { 2286 _trace_gen0_time_data.increment_young_collection_count(); 2287 } else { 2288 _trace_gen0_time_data.increment_mixed_collection_count(); 2289 } 2290 2291 // The young list is laid with the survivor regions from the previous 2292 // pause are appended to the RHS of the young list, i.e. 2293 // [Newly Young Regions ++ Survivors from last pause]. 2294 2295 uint survivor_region_length = young_list->survivor_length(); 2296 uint eden_region_length = young_list->length() - survivor_region_length; 2297 init_cset_region_lengths(eden_region_length, survivor_region_length); 2298 hr = young_list->first_survivor_region(); 2299 while (hr != NULL) { 2300 assert(hr->is_survivor(), "badly formed young list"); 2301 hr->set_young(); 2302 hr = hr->get_next_young_region(); 2303 } 2304 2305 // Clear the fields that point to the survivor list - they are all young now. 2306 young_list->clear_survivors(); 2307 2308 _collection_set = _inc_cset_head; 2309 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; 2310 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms; 2311 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; 2312 2313 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, 2314 "add young regions to CSet", 2315 ergo_format_region("eden") 2316 ergo_format_region("survivors") 2317 ergo_format_ms("predicted young region time"), 2318 eden_region_length, survivor_region_length, 2319 _inc_cset_predicted_elapsed_time_ms); 2320 2321 // The number of recorded young regions is the incremental 2322 // collection set's current size 2323 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); 2324 2325 double young_end_time_sec = os::elapsedTime(); 2326 _recorded_young_cset_choice_time_ms = 2327 (young_end_time_sec - young_start_time_sec) * 1000.0; 2328 2329 // We are doing young collections so reset this. 2330 non_young_start_time_sec = young_end_time_sec; 2331 2332 if (!gcs_are_young()) { 2333 CollectionSetChooser* cset_chooser = _collectionSetChooser; 2334 cset_chooser->verify(); 2335 const uint min_old_cset_length = cset_chooser->calc_min_old_cset_length(); 2336 const uint max_old_cset_length = cset_chooser->calc_max_old_cset_length(); 2337 2338 uint expensive_region_num = 0; 2339 bool check_time_remaining = adaptive_young_list_length(); 2340 HeapRegion* hr = cset_chooser->peek(); 2341 while (hr != NULL) { 2342 if (old_cset_region_length() >= max_old_cset_length) { 2343 // Added maximum number of old regions to the CSet. 2344 ergo_verbose2(ErgoCSetConstruction, 2345 "finish adding old regions to CSet", 2346 ergo_format_reason("old CSet region num reached max") 2347 ergo_format_region("old") 2348 ergo_format_region("max"), 2349 old_cset_region_length(), max_old_cset_length); 2350 break; 2351 } 2352 2353 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); 2354 if (check_time_remaining) { 2355 if (predicted_time_ms > time_remaining_ms) { 2356 // Too expensive for the current CSet. 2357 2358 if (old_cset_region_length() >= min_old_cset_length) { 2359 // We have added the minimum number of old regions to the CSet, 2360 // we are done with this CSet. 2361 ergo_verbose4(ErgoCSetConstruction, 2362 "finish adding old regions to CSet", 2363 ergo_format_reason("predicted time is too high") 2364 ergo_format_ms("predicted time") 2365 ergo_format_ms("remaining time") 2366 ergo_format_region("old") 2367 ergo_format_region("min"), 2368 predicted_time_ms, time_remaining_ms, 2369 old_cset_region_length(), min_old_cset_length); 2370 break; 2371 } 2372 2373 // We'll add it anyway given that we haven't reached the 2374 // minimum number of old regions. 2375 expensive_region_num += 1; 2376 } 2377 } else { 2378 if (old_cset_region_length() >= min_old_cset_length) { 2379 // In the non-auto-tuning case, we'll finish adding regions 2380 // to the CSet if we reach the minimum. 2381 ergo_verbose2(ErgoCSetConstruction, 2382 "finish adding old regions to CSet", 2383 ergo_format_reason("old CSet region num reached min") 2384 ergo_format_region("old") 2385 ergo_format_region("min"), 2386 old_cset_region_length(), min_old_cset_length); 2387 break; 2388 } 2389 } 2390 2391 // We will add this region to the CSet. 2392 time_remaining_ms -= predicted_time_ms; 2393 predicted_pause_time_ms += predicted_time_ms; 2394 cset_chooser->remove_and_move_to_next(hr); 2395 _g1->old_set_remove(hr); 2396 add_old_region_to_cset(hr); 2397 2398 hr = cset_chooser->peek(); 2399 } 2400 if (hr == NULL) { 2401 ergo_verbose0(ErgoCSetConstruction, 2402 "finish adding old regions to CSet", 2403 ergo_format_reason("candidate old regions not available")); 2404 } 2405 2406 if (expensive_region_num > 0) { 2407 // We print the information once here at the end, predicated on 2408 // whether we added any apparently expensive regions or not, to 2409 // avoid generating output per region. 2410 ergo_verbose4(ErgoCSetConstruction, 2411 "added expensive regions to CSet", 2412 ergo_format_reason("old CSet region num not reached min") 2413 ergo_format_region("old") 2414 ergo_format_region("expensive") 2415 ergo_format_region("min") 2416 ergo_format_ms("remaining time"), 2417 old_cset_region_length(), 2418 expensive_region_num, 2419 min_old_cset_length, 2420 time_remaining_ms); 2421 } 2422 2423 cset_chooser->verify(); 2424 } 2425 2426 stop_incremental_cset_building(); 2427 2428 count_CS_bytes_used(); 2429 2430 ergo_verbose5(ErgoCSetConstruction, 2431 "finish choosing CSet", 2432 ergo_format_region("eden") 2433 ergo_format_region("survivors") 2434 ergo_format_region("old") 2435 ergo_format_ms("predicted pause time") 2436 ergo_format_ms("target pause time"), 2437 eden_region_length, survivor_region_length, 2438 old_cset_region_length(), 2439 predicted_pause_time_ms, target_pause_time_ms); 2440 2441 double non_young_end_time_sec = os::elapsedTime(); 2442 _recorded_non_young_cset_choice_time_ms = 2443 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; 2444 } 2445 2446 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) { 2447 if(TraceGen0Time) { 2448 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); 2449 } 2450 } 2451 2452 void TraceGen0TimeData::record_yield_time(double yield_time_ms) { 2453 if(TraceGen0Time) { 2454 _all_yield_times_ms.add(yield_time_ms); 2455 } 2456 } 2457 2458 void TraceGen0TimeData::record_end_collection( 2459 double total_ms, 2460 double other_ms, 2461 double root_region_scan_wait_ms, 2462 double parallel_ms, 2463 double ext_root_scan_ms, 2464 double satb_filtering_ms, 2465 double update_rs_ms, 2466 double scan_rs_ms, 2467 double obj_copy_ms, 2468 double termination_ms, 2469 double parallel_other_ms, 2470 double clear_ct_ms) 2471 { 2472 if(TraceGen0Time) { 2473 _total.add(total_ms); 2474 _other.add(other_ms); 2475 _root_region_scan_wait.add(root_region_scan_wait_ms); 2476 _parallel.add(parallel_ms); 2477 _ext_root_scan.add(ext_root_scan_ms); 2478 _satb_filtering.add(satb_filtering_ms); 2479 _update_rs.add(update_rs_ms); 2480 _scan_rs.add(scan_rs_ms); 2481 _obj_copy.add(obj_copy_ms); 2482 _termination.add(termination_ms); 2483 _parallel_other.add(parallel_other_ms); 2484 _clear_ct.add(clear_ct_ms); 2485 } 2486 } 2487 2488 void TraceGen0TimeData::increment_young_collection_count() { 2489 if(TraceGen0Time) { 2490 ++_young_pause_num; 2491 } 2492 } 2493 2494 void TraceGen0TimeData::increment_mixed_collection_count() { 2495 if(TraceGen0Time) { 2496 ++_mixed_pause_num; 2497 } 2498 } 2499 2500 void TraceGen0TimeData::print_summary(int level, 2501 const char* str, 2502 const NumberSeq* seq) const { 2503 double sum = seq->sum(); 2504 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", 2505 str, sum / 1000.0, seq->avg()); 2506 } 2507 2508 void TraceGen0TimeData::print_summary_sd(int level, 2509 const char* str, 2510 const NumberSeq* seq) const { 2511 print_summary(level, str, seq); 2512 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", 2513 seq->num(), seq->sd(), seq->maximum()); 2514 } 2515 2516 void TraceGen0TimeData::print() const { 2517 if (!TraceGen0Time) { 2518 return; 2519 } 2520 2521 gclog_or_tty->print_cr("ALL PAUSES"); 2522 print_summary_sd(0, "Total", &_total); 2523 gclog_or_tty->print_cr(""); 2524 gclog_or_tty->print_cr(""); 2525 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); 2526 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); 2527 gclog_or_tty->print_cr(""); 2528 2529 gclog_or_tty->print_cr("EVACUATION PAUSES"); 2530 2531 if (_young_pause_num == 0 && _mixed_pause_num == 0) { 2532 gclog_or_tty->print_cr("none"); 2533 } else { 2534 print_summary_sd(0, "Evacuation Pauses", &_total); 2535 print_summary(1, "Root Region Scan Wait", &_root_region_scan_wait); 2536 print_summary(1, "Parallel Time", &_parallel); 2537 print_summary(2, "Ext Root Scanning", &_ext_root_scan); 2538 print_summary(2, "SATB Filtering", &_satb_filtering); 2539 print_summary(2, "Update RS", &_update_rs); 2540 print_summary(2, "Scan RS", &_scan_rs); 2541 print_summary(2, "Object Copy", &_obj_copy); 2542 print_summary(2, "Termination", &_termination); 2543 print_summary(2, "Parallel Other", &_parallel_other); 2544 print_summary(1, "Clear CT", &_clear_ct); 2545 print_summary(1, "Other", &_other); 2546 } 2547 gclog_or_tty->print_cr(""); 2548 2549 gclog_or_tty->print_cr("MISC"); 2550 print_summary_sd(0, "Stop World", &_all_stop_world_times_ms); 2551 print_summary_sd(0, "Yields", &_all_yield_times_ms); 2552 } 2553 2554 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) { 2555 if (TraceGen1Time) { 2556 _all_full_gc_times.add(full_gc_time_ms); 2557 } 2558 } 2559 2560 void TraceGen1TimeData::print() const { 2561 if (!TraceGen1Time) { 2562 return; 2563 } 2564 2565 if (_all_full_gc_times.num() > 0) { 2566 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", 2567 _all_full_gc_times.num(), 2568 _all_full_gc_times.sum() / 1000.0); 2569 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); 2570 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", 2571 _all_full_gc_times.sd(), 2572 _all_full_gc_times.maximum()); 2573 } 2574 }