1 /* 2 * Copyright (c) 2001, 2014, 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 #if !defined(__clang_major__) && defined(__GNUC__) 26 #define ATTRIBUTE_PRINTF(x,y) // FIXME, formats are a mess. 27 #endif 28 29 #include "precompiled.hpp" 30 #include "code/codeCache.hpp" 31 #include "code/icBuffer.hpp" 32 #include "gc_implementation/g1/bufferingOopClosure.hpp" 33 #include "gc_implementation/g1/concurrentG1Refine.hpp" 34 #include "gc_implementation/g1/concurrentG1RefineThread.hpp" 35 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 36 #include "gc_implementation/g1/g1AllocRegion.inline.hpp" 37 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 38 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 39 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 40 #include "gc_implementation/g1/g1EvacFailure.hpp" 41 #include "gc_implementation/g1/g1GCPhaseTimes.hpp" 42 #include "gc_implementation/g1/g1Log.hpp" 43 #include "gc_implementation/g1/g1MarkSweep.hpp" 44 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 45 #include "gc_implementation/g1/g1RemSet.inline.hpp" 46 #include "gc_implementation/g1/g1StringDedup.hpp" 47 #include "gc_implementation/g1/g1YCTypes.hpp" 48 #include "gc_implementation/g1/heapRegion.inline.hpp" 49 #include "gc_implementation/g1/heapRegionRemSet.hpp" 50 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 51 #include "gc_implementation/g1/vm_operations_g1.hpp" 52 #include "gc_implementation/shared/gcHeapSummary.hpp" 53 #include "gc_implementation/shared/gcTimer.hpp" 54 #include "gc_implementation/shared/gcTrace.hpp" 55 #include "gc_implementation/shared/gcTraceTime.hpp" 56 #include "gc_implementation/shared/isGCActiveMark.hpp" 57 #include "memory/gcLocker.inline.hpp" 58 #include "memory/generationSpec.hpp" 59 #include "memory/iterator.hpp" 60 #include "memory/referenceProcessor.hpp" 61 #include "oops/oop.inline.hpp" 62 #include "oops/oop.pcgc.inline.hpp" 63 #include "runtime/prefetch.inline.hpp" 64 #include "runtime/orderAccess.inline.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "utilities/ticks.hpp" 67 68 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 69 70 // turn it on so that the contents of the young list (scan-only / 71 // to-be-collected) are printed at "strategic" points before / during 72 // / after the collection --- this is useful for debugging 73 #define YOUNG_LIST_VERBOSE 0 74 // CURRENT STATUS 75 // This file is under construction. Search for "FIXME". 76 77 // INVARIANTS/NOTES 78 // 79 // All allocation activity covered by the G1CollectedHeap interface is 80 // serialized by acquiring the HeapLock. This happens in mem_allocate 81 // and allocate_new_tlab, which are the "entry" points to the 82 // allocation code from the rest of the JVM. (Note that this does not 83 // apply to TLAB allocation, which is not part of this interface: it 84 // is done by clients of this interface.) 85 86 // Notes on implementation of parallelism in different tasks. 87 // 88 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 89 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 90 // It does use run_task() which sets _n_workers in the task. 91 // G1ParTask executes g1_process_strong_roots() -> 92 // SharedHeap::process_strong_roots() which calls eventually to 93 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 94 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 95 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 96 // 97 98 // Local to this file. 99 100 class RefineCardTableEntryClosure: public CardTableEntryClosure { 101 bool _concurrent; 102 public: 103 RefineCardTableEntryClosure() : _concurrent(true) { } 104 105 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 106 bool oops_into_cset = G1CollectedHeap::heap()->g1_rem_set()->refine_card(card_ptr, worker_i, false); 107 // This path is executed by the concurrent refine or mutator threads, 108 // concurrently, and so we do not care if card_ptr contains references 109 // that point into the collection set. 110 assert(!oops_into_cset, "should be"); 111 112 if (_concurrent && SuspendibleThreadSet::should_yield()) { 113 // Caller will actually yield. 114 return false; 115 } 116 // Otherwise, we finished successfully; return true. 117 return true; 118 } 119 120 void set_concurrent(bool b) { _concurrent = b; } 121 }; 122 123 124 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 125 int _calls; 126 G1CollectedHeap* _g1h; 127 CardTableModRefBS* _ctbs; 128 int _histo[256]; 129 public: 130 ClearLoggedCardTableEntryClosure() : 131 _calls(0), _g1h(G1CollectedHeap::heap()), _ctbs(_g1h->g1_barrier_set()) 132 { 133 for (int i = 0; i < 256; i++) _histo[i] = 0; 134 } 135 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 136 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 137 _calls++; 138 unsigned char* ujb = (unsigned char*)card_ptr; 139 int ind = (int)(*ujb); 140 _histo[ind]++; 141 *card_ptr = -1; 142 } 143 return true; 144 } 145 int calls() { return _calls; } 146 void print_histo() { 147 gclog_or_tty->print_cr("Card table value histogram:"); 148 for (int i = 0; i < 256; i++) { 149 if (_histo[i] != 0) { 150 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 151 } 152 } 153 } 154 }; 155 156 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 157 int _calls; 158 G1CollectedHeap* _g1h; 159 CardTableModRefBS* _ctbs; 160 public: 161 RedirtyLoggedCardTableEntryClosure() : 162 _calls(0), _g1h(G1CollectedHeap::heap()), _ctbs(_g1h->g1_barrier_set()) {} 163 164 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 165 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 166 _calls++; 167 *card_ptr = 0; 168 } 169 return true; 170 } 171 int calls() { return _calls; } 172 }; 173 174 YoungList::YoungList(G1CollectedHeap* g1h) : 175 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 176 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 177 guarantee(check_list_empty(false), "just making sure..."); 178 } 179 180 void YoungList::push_region(HeapRegion *hr) { 181 assert(!hr->is_young(), "should not already be young"); 182 assert(hr->get_next_young_region() == NULL, "cause it should!"); 183 184 hr->set_next_young_region(_head); 185 _head = hr; 186 187 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 188 ++_length; 189 } 190 191 void YoungList::add_survivor_region(HeapRegion* hr) { 192 assert(hr->is_survivor(), "should be flagged as survivor region"); 193 assert(hr->get_next_young_region() == NULL, "cause it should!"); 194 195 hr->set_next_young_region(_survivor_head); 196 if (_survivor_head == NULL) { 197 _survivor_tail = hr; 198 } 199 _survivor_head = hr; 200 ++_survivor_length; 201 } 202 203 void YoungList::empty_list(HeapRegion* list) { 204 while (list != NULL) { 205 HeapRegion* next = list->get_next_young_region(); 206 list->set_next_young_region(NULL); 207 list->uninstall_surv_rate_group(); 208 list->set_not_young(); 209 list = next; 210 } 211 } 212 213 void YoungList::empty_list() { 214 assert(check_list_well_formed(), "young list should be well formed"); 215 216 empty_list(_head); 217 _head = NULL; 218 _length = 0; 219 220 empty_list(_survivor_head); 221 _survivor_head = NULL; 222 _survivor_tail = NULL; 223 _survivor_length = 0; 224 225 _last_sampled_rs_lengths = 0; 226 227 assert(check_list_empty(false), "just making sure..."); 228 } 229 230 bool YoungList::check_list_well_formed() { 231 bool ret = true; 232 233 uint length = 0; 234 HeapRegion* curr = _head; 235 HeapRegion* last = NULL; 236 while (curr != NULL) { 237 if (!curr->is_young()) { 238 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 239 "incorrectly tagged (y: %d, surv: %d)", 240 curr->bottom(), curr->end(), 241 curr->is_young(), curr->is_survivor()); 242 ret = false; 243 } 244 ++length; 245 last = curr; 246 curr = curr->get_next_young_region(); 247 } 248 ret = ret && (length == _length); 249 250 if (!ret) { 251 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 252 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 253 length, _length); 254 } 255 256 return ret; 257 } 258 259 bool YoungList::check_list_empty(bool check_sample) { 260 bool ret = true; 261 262 if (_length != 0) { 263 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 264 _length); 265 ret = false; 266 } 267 if (check_sample && _last_sampled_rs_lengths != 0) { 268 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 269 ret = false; 270 } 271 if (_head != NULL) { 272 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 273 ret = false; 274 } 275 if (!ret) { 276 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 277 } 278 279 return ret; 280 } 281 282 void 283 YoungList::rs_length_sampling_init() { 284 _sampled_rs_lengths = 0; 285 _curr = _head; 286 } 287 288 bool 289 YoungList::rs_length_sampling_more() { 290 return _curr != NULL; 291 } 292 293 void 294 YoungList::rs_length_sampling_next() { 295 assert( _curr != NULL, "invariant" ); 296 size_t rs_length = _curr->rem_set()->occupied(); 297 298 _sampled_rs_lengths += rs_length; 299 300 // The current region may not yet have been added to the 301 // incremental collection set (it gets added when it is 302 // retired as the current allocation region). 303 if (_curr->in_collection_set()) { 304 // Update the collection set policy information for this region 305 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 306 } 307 308 _curr = _curr->get_next_young_region(); 309 if (_curr == NULL) { 310 _last_sampled_rs_lengths = _sampled_rs_lengths; 311 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 312 } 313 } 314 315 void 316 YoungList::reset_auxilary_lists() { 317 guarantee( is_empty(), "young list should be empty" ); 318 assert(check_list_well_formed(), "young list should be well formed"); 319 320 // Add survivor regions to SurvRateGroup. 321 _g1h->g1_policy()->note_start_adding_survivor_regions(); 322 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 323 324 int young_index_in_cset = 0; 325 for (HeapRegion* curr = _survivor_head; 326 curr != NULL; 327 curr = curr->get_next_young_region()) { 328 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 329 330 // The region is a non-empty survivor so let's add it to 331 // the incremental collection set for the next evacuation 332 // pause. 333 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 334 young_index_in_cset += 1; 335 } 336 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 337 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 338 339 _head = _survivor_head; 340 _length = _survivor_length; 341 if (_survivor_head != NULL) { 342 assert(_survivor_tail != NULL, "cause it shouldn't be"); 343 assert(_survivor_length > 0, "invariant"); 344 _survivor_tail->set_next_young_region(NULL); 345 } 346 347 // Don't clear the survivor list handles until the start of 348 // the next evacuation pause - we need it in order to re-tag 349 // the survivor regions from this evacuation pause as 'young' 350 // at the start of the next. 351 352 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 353 354 assert(check_list_well_formed(), "young list should be well formed"); 355 } 356 357 void YoungList::print() { 358 HeapRegion* lists[] = {_head, _survivor_head}; 359 const char* names[] = {"YOUNG", "SURVIVOR"}; 360 361 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 362 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 363 HeapRegion *curr = lists[list]; 364 if (curr == NULL) 365 gclog_or_tty->print_cr(" empty"); 366 while (curr != NULL) { 367 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 368 HR_FORMAT_PARAMS(curr), 369 curr->prev_top_at_mark_start(), 370 curr->next_top_at_mark_start(), 371 curr->age_in_surv_rate_group_cond()); 372 curr = curr->get_next_young_region(); 373 } 374 } 375 376 gclog_or_tty->cr(); 377 } 378 379 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 380 { 381 // Claim the right to put the region on the dirty cards region list 382 // by installing a self pointer. 383 HeapRegion* next = hr->get_next_dirty_cards_region(); 384 if (next == NULL) { 385 HeapRegion* res = (HeapRegion*) 386 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 387 NULL); 388 if (res == NULL) { 389 HeapRegion* head; 390 do { 391 // Put the region to the dirty cards region list. 392 head = _dirty_cards_region_list; 393 next = (HeapRegion*) 394 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 395 if (next == head) { 396 assert(hr->get_next_dirty_cards_region() == hr, 397 "hr->get_next_dirty_cards_region() != hr"); 398 if (next == NULL) { 399 // The last region in the list points to itself. 400 hr->set_next_dirty_cards_region(hr); 401 } else { 402 hr->set_next_dirty_cards_region(next); 403 } 404 } 405 } while (next != head); 406 } 407 } 408 } 409 410 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 411 { 412 HeapRegion* head; 413 HeapRegion* hr; 414 do { 415 head = _dirty_cards_region_list; 416 if (head == NULL) { 417 return NULL; 418 } 419 HeapRegion* new_head = head->get_next_dirty_cards_region(); 420 if (head == new_head) { 421 // The last region. 422 new_head = NULL; 423 } 424 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 425 head); 426 } while (hr != head); 427 assert(hr != NULL, "invariant"); 428 hr->set_next_dirty_cards_region(NULL); 429 return hr; 430 } 431 432 #ifdef ASSERT 433 // A region is added to the collection set as it is retired 434 // so an address p can point to a region which will be in the 435 // collection set but has not yet been retired. This method 436 // therefore is only accurate during a GC pause after all 437 // regions have been retired. It is used for debugging 438 // to check if an nmethod has references to objects that can 439 // be move during a partial collection. Though it can be 440 // inaccurate, it is sufficient for G1 because the conservative 441 // implementation of is_scavengable() for G1 will indicate that 442 // all nmethods must be scanned during a partial collection. 443 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 444 HeapRegion* hr = heap_region_containing(p); 445 return hr != NULL && hr->in_collection_set(); 446 } 447 #endif 448 449 // Returns true if the reference points to an object that 450 // can move in an incremental collection. 451 bool G1CollectedHeap::is_scavengable(const void* p) { 452 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 453 G1CollectorPolicy* g1p = g1h->g1_policy(); 454 HeapRegion* hr = heap_region_containing(p); 455 if (hr == NULL) { 456 // null 457 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p)); 458 return false; 459 } else { 460 return !hr->isHumongous(); 461 } 462 } 463 464 void G1CollectedHeap::check_ct_logs_at_safepoint() { 465 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 466 CardTableModRefBS* ct_bs = g1_barrier_set(); 467 468 // Count the dirty cards at the start. 469 CountNonCleanMemRegionClosure count1(this); 470 ct_bs->mod_card_iterate(&count1); 471 int orig_count = count1.n(); 472 473 // First clear the logged cards. 474 ClearLoggedCardTableEntryClosure clear; 475 dcqs.apply_closure_to_all_completed_buffers(&clear); 476 dcqs.iterate_closure_all_threads(&clear, false); 477 clear.print_histo(); 478 479 // Now ensure that there's no dirty cards. 480 CountNonCleanMemRegionClosure count2(this); 481 ct_bs->mod_card_iterate(&count2); 482 if (count2.n() != 0) { 483 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 484 count2.n(), orig_count); 485 } 486 guarantee(count2.n() == 0, "Card table should be clean."); 487 488 RedirtyLoggedCardTableEntryClosure redirty; 489 dcqs.apply_closure_to_all_completed_buffers(&redirty); 490 dcqs.iterate_closure_all_threads(&redirty, false); 491 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 492 clear.calls(), orig_count); 493 guarantee(redirty.calls() == clear.calls(), 494 "Or else mechanism is broken."); 495 496 CountNonCleanMemRegionClosure count3(this); 497 ct_bs->mod_card_iterate(&count3); 498 if (count3.n() != orig_count) { 499 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 500 orig_count, count3.n()); 501 guarantee(count3.n() >= orig_count, "Should have restored them all."); 502 } 503 } 504 505 // Private class members. 506 507 G1CollectedHeap* G1CollectedHeap::_g1h; 508 509 // Private methods. 510 511 HeapRegion* 512 G1CollectedHeap::new_region_try_secondary_free_list(bool is_old) { 513 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 514 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 515 if (!_secondary_free_list.is_empty()) { 516 if (G1ConcRegionFreeingVerbose) { 517 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 518 "secondary_free_list has %u entries", 519 _secondary_free_list.length()); 520 } 521 // It looks as if there are free regions available on the 522 // secondary_free_list. Let's move them to the free_list and try 523 // again to allocate from it. 524 append_secondary_free_list(); 525 526 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 527 "empty we should have moved at least one entry to the free_list"); 528 HeapRegion* res = _free_list.remove_region(is_old); 529 if (G1ConcRegionFreeingVerbose) { 530 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 531 "allocated "HR_FORMAT" from secondary_free_list", 532 HR_FORMAT_PARAMS(res)); 533 } 534 return res; 535 } 536 537 // Wait here until we get notified either when (a) there are no 538 // more free regions coming or (b) some regions have been moved on 539 // the secondary_free_list. 540 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 541 } 542 543 if (G1ConcRegionFreeingVerbose) { 544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 545 "could not allocate from secondary_free_list"); 546 } 547 return NULL; 548 } 549 550 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_expand) { 551 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 552 "the only time we use this to allocate a humongous region is " 553 "when we are allocating a single humongous region"); 554 555 HeapRegion* res; 556 if (G1StressConcRegionFreeing) { 557 if (!_secondary_free_list.is_empty()) { 558 if (G1ConcRegionFreeingVerbose) { 559 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 560 "forced to look at the secondary_free_list"); 561 } 562 res = new_region_try_secondary_free_list(is_old); 563 if (res != NULL) { 564 return res; 565 } 566 } 567 } 568 569 res = _free_list.remove_region(is_old); 570 571 if (res == NULL) { 572 if (G1ConcRegionFreeingVerbose) { 573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 574 "res == NULL, trying the secondary_free_list"); 575 } 576 res = new_region_try_secondary_free_list(is_old); 577 } 578 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 579 // Currently, only attempts to allocate GC alloc regions set 580 // do_expand to true. So, we should only reach here during a 581 // safepoint. If this assumption changes we might have to 582 // reconsider the use of _expand_heap_after_alloc_failure. 583 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 584 585 ergo_verbose1(ErgoHeapSizing, 586 "attempt heap expansion", 587 ergo_format_reason("region allocation request failed") 588 ergo_format_byte("allocation request"), 589 word_size * HeapWordSize); 590 if (expand(word_size * HeapWordSize)) { 591 // Given that expand() succeeded in expanding the heap, and we 592 // always expand the heap by an amount aligned to the heap 593 // region size, the free list should in theory not be empty. 594 // In either case remove_region() will check for NULL. 595 res = _free_list.remove_region(is_old); 596 } else { 597 _expand_heap_after_alloc_failure = false; 598 } 599 } 600 return res; 601 } 602 603 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 604 size_t word_size) { 605 assert(isHumongous(word_size), "word_size should be humongous"); 606 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 607 608 uint first = G1_NULL_HRS_INDEX; 609 if (num_regions == 1) { 610 // Only one region to allocate, no need to go through the slower 611 // path. The caller will attempt the expansion if this fails, so 612 // let's not try to expand here too. 613 HeapRegion* hr = new_region(word_size, true /* is_old */, false /* do_expand */); 614 if (hr != NULL) { 615 first = hr->hrs_index(); 616 } else { 617 first = G1_NULL_HRS_INDEX; 618 } 619 } else { 620 // We can't allocate humongous regions while cleanupComplete() is 621 // running, since some of the regions we find to be empty might not 622 // yet be added to the free list and it is not straightforward to 623 // know which list they are on so that we can remove them. Note 624 // that we only need to do this if we need to allocate more than 625 // one region to satisfy the current humongous allocation 626 // request. If we are only allocating one region we use the common 627 // region allocation code (see above). 628 wait_while_free_regions_coming(); 629 append_secondary_free_list_if_not_empty_with_lock(); 630 631 if (free_regions() >= num_regions) { 632 first = _hrs.find_contiguous(num_regions); 633 if (first != G1_NULL_HRS_INDEX) { 634 for (uint i = first; i < first + num_regions; ++i) { 635 HeapRegion* hr = region_at(i); 636 assert(hr->is_empty(), "sanity"); 637 assert(is_on_master_free_list(hr), "sanity"); 638 hr->set_pending_removal(true); 639 } 640 _free_list.remove_all_pending(num_regions); 641 } 642 } 643 } 644 return first; 645 } 646 647 HeapWord* 648 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 649 uint num_regions, 650 size_t word_size) { 651 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 652 assert(isHumongous(word_size), "word_size should be humongous"); 653 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 654 655 // Index of last region in the series + 1. 656 uint last = first + num_regions; 657 658 // We need to initialize the region(s) we just discovered. This is 659 // a bit tricky given that it can happen concurrently with 660 // refinement threads refining cards on these regions and 661 // potentially wanting to refine the BOT as they are scanning 662 // those cards (this can happen shortly after a cleanup; see CR 663 // 6991377). So we have to set up the region(s) carefully and in 664 // a specific order. 665 666 // The word size sum of all the regions we will allocate. 667 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 668 assert(word_size <= word_size_sum, "sanity"); 669 670 // This will be the "starts humongous" region. 671 HeapRegion* first_hr = region_at(first); 672 // The header of the new object will be placed at the bottom of 673 // the first region. 674 HeapWord* new_obj = first_hr->bottom(); 675 // This will be the new end of the first region in the series that 676 // should also match the end of the last region in the series. 677 HeapWord* new_end = new_obj + word_size_sum; 678 // This will be the new top of the first region that will reflect 679 // this allocation. 680 HeapWord* new_top = new_obj + word_size; 681 682 // First, we need to zero the header of the space that we will be 683 // allocating. When we update top further down, some refinement 684 // threads might try to scan the region. By zeroing the header we 685 // ensure that any thread that will try to scan the region will 686 // come across the zero klass word and bail out. 687 // 688 // NOTE: It would not have been correct to have used 689 // CollectedHeap::fill_with_object() and make the space look like 690 // an int array. The thread that is doing the allocation will 691 // later update the object header to a potentially different array 692 // type and, for a very short period of time, the klass and length 693 // fields will be inconsistent. This could cause a refinement 694 // thread to calculate the object size incorrectly. 695 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 696 697 // We will set up the first region as "starts humongous". This 698 // will also update the BOT covering all the regions to reflect 699 // that there is a single object that starts at the bottom of the 700 // first region. 701 first_hr->set_startsHumongous(new_top, new_end); 702 703 // Then, if there are any, we will set up the "continues 704 // humongous" regions. 705 HeapRegion* hr = NULL; 706 for (uint i = first + 1; i < last; ++i) { 707 hr = region_at(i); 708 hr->set_continuesHumongous(first_hr); 709 } 710 // If we have "continues humongous" regions (hr != NULL), then the 711 // end of the last one should match new_end. 712 assert(hr == NULL || hr->end() == new_end, "sanity"); 713 714 // Up to this point no concurrent thread would have been able to 715 // do any scanning on any region in this series. All the top 716 // fields still point to bottom, so the intersection between 717 // [bottom,top] and [card_start,card_end] will be empty. Before we 718 // update the top fields, we'll do a storestore to make sure that 719 // no thread sees the update to top before the zeroing of the 720 // object header and the BOT initialization. 721 OrderAccess::storestore(); 722 723 // Now that the BOT and the object header have been initialized, 724 // we can update top of the "starts humongous" region. 725 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 726 "new_top should be in this region"); 727 first_hr->set_top(new_top); 728 if (_hr_printer.is_active()) { 729 HeapWord* bottom = first_hr->bottom(); 730 HeapWord* end = first_hr->orig_end(); 731 if ((first + 1) == last) { 732 // the series has a single humongous region 733 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 734 } else { 735 // the series has more than one humongous regions 736 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 737 } 738 } 739 740 // Now, we will update the top fields of the "continues humongous" 741 // regions. The reason we need to do this is that, otherwise, 742 // these regions would look empty and this will confuse parts of 743 // G1. For example, the code that looks for a consecutive number 744 // of empty regions will consider them empty and try to 745 // re-allocate them. We can extend is_empty() to also include 746 // !continuesHumongous(), but it is easier to just update the top 747 // fields here. The way we set top for all regions (i.e., top == 748 // end for all regions but the last one, top == new_top for the 749 // last one) is actually used when we will free up the humongous 750 // region in free_humongous_region(). 751 hr = NULL; 752 for (uint i = first + 1; i < last; ++i) { 753 hr = region_at(i); 754 if ((i + 1) == last) { 755 // last continues humongous region 756 assert(hr->bottom() < new_top && new_top <= hr->end(), 757 "new_top should fall on this region"); 758 hr->set_top(new_top); 759 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 760 } else { 761 // not last one 762 assert(new_top > hr->end(), "new_top should be above this region"); 763 hr->set_top(hr->end()); 764 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 765 } 766 } 767 // If we have continues humongous regions (hr != NULL), then the 768 // end of the last one should match new_end and its top should 769 // match new_top. 770 assert(hr == NULL || 771 (hr->end() == new_end && hr->top() == new_top), "sanity"); 772 773 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 774 _summary_bytes_used += first_hr->used(); 775 _humongous_set.add(first_hr); 776 777 return new_obj; 778 } 779 780 // If could fit into free regions w/o expansion, try. 781 // Otherwise, if can expand, do so. 782 // Otherwise, if using ex regions might help, try with ex given back. 783 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 784 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 785 786 verify_region_sets_optional(); 787 788 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 789 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 790 uint x_num = expansion_regions(); 791 uint fs = _hrs.free_suffix(); 792 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 793 if (first == G1_NULL_HRS_INDEX) { 794 // The only thing we can do now is attempt expansion. 795 if (fs + x_num >= num_regions) { 796 // If the number of regions we're trying to allocate for this 797 // object is at most the number of regions in the free suffix, 798 // then the call to humongous_obj_allocate_find_first() above 799 // should have succeeded and we wouldn't be here. 800 // 801 // We should only be trying to expand when the free suffix is 802 // not sufficient for the object _and_ we have some expansion 803 // room available. 804 assert(num_regions > fs, "earlier allocation should have succeeded"); 805 806 ergo_verbose1(ErgoHeapSizing, 807 "attempt heap expansion", 808 ergo_format_reason("humongous allocation request failed") 809 ergo_format_byte("allocation request"), 810 word_size * HeapWordSize); 811 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 812 // Even though the heap was expanded, it might not have 813 // reached the desired size. So, we cannot assume that the 814 // allocation will succeed. 815 first = humongous_obj_allocate_find_first(num_regions, word_size); 816 } 817 } 818 } 819 820 HeapWord* result = NULL; 821 if (first != G1_NULL_HRS_INDEX) { 822 result = 823 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 824 assert(result != NULL, "it should always return a valid result"); 825 826 // A successful humongous object allocation changes the used space 827 // information of the old generation so we need to recalculate the 828 // sizes and update the jstat counters here. 829 g1mm()->update_sizes(); 830 } 831 832 verify_region_sets_optional(); 833 834 return result; 835 } 836 837 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 838 assert_heap_not_locked_and_not_at_safepoint(); 839 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 840 841 unsigned int dummy_gc_count_before; 842 int dummy_gclocker_retry_count = 0; 843 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count); 844 } 845 846 HeapWord* 847 G1CollectedHeap::mem_allocate(size_t word_size, 848 bool* gc_overhead_limit_was_exceeded) { 849 assert_heap_not_locked_and_not_at_safepoint(); 850 851 // Loop until the allocation is satisfied, or unsatisfied after GC. 852 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) { 853 unsigned int gc_count_before; 854 855 HeapWord* result = NULL; 856 if (!isHumongous(word_size)) { 857 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count); 858 } else { 859 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count); 860 } 861 if (result != NULL) { 862 return result; 863 } 864 865 // Create the garbage collection operation... 866 VM_G1CollectForAllocation op(gc_count_before, word_size); 867 // ...and get the VM thread to execute it. 868 VMThread::execute(&op); 869 870 if (op.prologue_succeeded() && op.pause_succeeded()) { 871 // If the operation was successful we'll return the result even 872 // if it is NULL. If the allocation attempt failed immediately 873 // after a Full GC, it's unlikely we'll be able to allocate now. 874 HeapWord* result = op.result(); 875 if (result != NULL && !isHumongous(word_size)) { 876 // Allocations that take place on VM operations do not do any 877 // card dirtying and we have to do it here. We only have to do 878 // this for non-humongous allocations, though. 879 dirty_young_block(result, word_size); 880 } 881 return result; 882 } else { 883 if (gclocker_retry_count > GCLockerRetryAllocationCount) { 884 return NULL; 885 } 886 assert(op.result() == NULL, 887 "the result should be NULL if the VM op did not succeed"); 888 } 889 890 // Give a warning if we seem to be looping forever. 891 if ((QueuedAllocationWarningCount > 0) && 892 (try_count % QueuedAllocationWarningCount == 0)) { 893 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 894 } 895 } 896 897 ShouldNotReachHere(); 898 return NULL; 899 } 900 901 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 902 unsigned int *gc_count_before_ret, 903 int* gclocker_retry_count_ret) { 904 // Make sure you read the note in attempt_allocation_humongous(). 905 906 assert_heap_not_locked_and_not_at_safepoint(); 907 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 908 "be called for humongous allocation requests"); 909 910 // We should only get here after the first-level allocation attempt 911 // (attempt_allocation()) failed to allocate. 912 913 // We will loop until a) we manage to successfully perform the 914 // allocation or b) we successfully schedule a collection which 915 // fails to perform the allocation. b) is the only case when we'll 916 // return NULL. 917 HeapWord* result = NULL; 918 for (int try_count = 1; /* we'll return */; try_count += 1) { 919 bool should_try_gc; 920 unsigned int gc_count_before; 921 922 { 923 MutexLockerEx x(Heap_lock); 924 925 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 926 false /* bot_updates */); 927 if (result != NULL) { 928 return result; 929 } 930 931 // If we reach here, attempt_allocation_locked() above failed to 932 // allocate a new region. So the mutator alloc region should be NULL. 933 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 934 935 if (GC_locker::is_active_and_needs_gc()) { 936 if (g1_policy()->can_expand_young_list()) { 937 // No need for an ergo verbose message here, 938 // can_expand_young_list() does this when it returns true. 939 result = _mutator_alloc_region.attempt_allocation_force(word_size, 940 false /* bot_updates */); 941 if (result != NULL) { 942 return result; 943 } 944 } 945 should_try_gc = false; 946 } else { 947 // The GCLocker may not be active but the GCLocker initiated 948 // GC may not yet have been performed (GCLocker::needs_gc() 949 // returns true). In this case we do not try this GC and 950 // wait until the GCLocker initiated GC is performed, and 951 // then retry the allocation. 952 if (GC_locker::needs_gc()) { 953 should_try_gc = false; 954 } else { 955 // Read the GC count while still holding the Heap_lock. 956 gc_count_before = total_collections(); 957 should_try_gc = true; 958 } 959 } 960 } 961 962 if (should_try_gc) { 963 bool succeeded; 964 result = do_collection_pause(word_size, gc_count_before, &succeeded, 965 GCCause::_g1_inc_collection_pause); 966 if (result != NULL) { 967 assert(succeeded, "only way to get back a non-NULL result"); 968 return result; 969 } 970 971 if (succeeded) { 972 // If we get here we successfully scheduled a collection which 973 // failed to allocate. No point in trying to allocate 974 // further. We'll just return NULL. 975 MutexLockerEx x(Heap_lock); 976 *gc_count_before_ret = total_collections(); 977 return NULL; 978 } 979 } else { 980 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 981 MutexLockerEx x(Heap_lock); 982 *gc_count_before_ret = total_collections(); 983 return NULL; 984 } 985 // The GCLocker is either active or the GCLocker initiated 986 // GC has not yet been performed. Stall until it is and 987 // then retry the allocation. 988 GC_locker::stall_until_clear(); 989 (*gclocker_retry_count_ret) += 1; 990 } 991 992 // We can reach here if we were unsuccessful in scheduling a 993 // collection (because another thread beat us to it) or if we were 994 // stalled due to the GC locker. In either can we should retry the 995 // allocation attempt in case another thread successfully 996 // performed a collection and reclaimed enough space. We do the 997 // first attempt (without holding the Heap_lock) here and the 998 // follow-on attempt will be at the start of the next loop 999 // iteration (after taking the Heap_lock). 1000 result = _mutator_alloc_region.attempt_allocation(word_size, 1001 false /* bot_updates */); 1002 if (result != NULL) { 1003 return result; 1004 } 1005 1006 // Give a warning if we seem to be looping forever. 1007 if ((QueuedAllocationWarningCount > 0) && 1008 (try_count % QueuedAllocationWarningCount == 0)) { 1009 warning("G1CollectedHeap::attempt_allocation_slow() " 1010 "retries %d times", try_count); 1011 } 1012 } 1013 1014 ShouldNotReachHere(); 1015 return NULL; 1016 } 1017 1018 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1019 unsigned int * gc_count_before_ret, 1020 int* gclocker_retry_count_ret) { 1021 // The structure of this method has a lot of similarities to 1022 // attempt_allocation_slow(). The reason these two were not merged 1023 // into a single one is that such a method would require several "if 1024 // allocation is not humongous do this, otherwise do that" 1025 // conditional paths which would obscure its flow. In fact, an early 1026 // version of this code did use a unified method which was harder to 1027 // follow and, as a result, it had subtle bugs that were hard to 1028 // track down. So keeping these two methods separate allows each to 1029 // be more readable. It will be good to keep these two in sync as 1030 // much as possible. 1031 1032 assert_heap_not_locked_and_not_at_safepoint(); 1033 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1034 "should only be called for humongous allocations"); 1035 1036 // Humongous objects can exhaust the heap quickly, so we should check if we 1037 // need to start a marking cycle at each humongous object allocation. We do 1038 // the check before we do the actual allocation. The reason for doing it 1039 // before the allocation is that we avoid having to keep track of the newly 1040 // allocated memory while we do a GC. 1041 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1042 word_size)) { 1043 collect(GCCause::_g1_humongous_allocation); 1044 } 1045 1046 // We will loop until a) we manage to successfully perform the 1047 // allocation or b) we successfully schedule a collection which 1048 // fails to perform the allocation. b) is the only case when we'll 1049 // return NULL. 1050 HeapWord* result = NULL; 1051 for (int try_count = 1; /* we'll return */; try_count += 1) { 1052 bool should_try_gc; 1053 unsigned int gc_count_before; 1054 1055 { 1056 MutexLockerEx x(Heap_lock); 1057 1058 // Given that humongous objects are not allocated in young 1059 // regions, we'll first try to do the allocation without doing a 1060 // collection hoping that there's enough space in the heap. 1061 result = humongous_obj_allocate(word_size); 1062 if (result != NULL) { 1063 return result; 1064 } 1065 1066 if (GC_locker::is_active_and_needs_gc()) { 1067 should_try_gc = false; 1068 } else { 1069 // The GCLocker may not be active but the GCLocker initiated 1070 // GC may not yet have been performed (GCLocker::needs_gc() 1071 // returns true). In this case we do not try this GC and 1072 // wait until the GCLocker initiated GC is performed, and 1073 // then retry the allocation. 1074 if (GC_locker::needs_gc()) { 1075 should_try_gc = false; 1076 } else { 1077 // Read the GC count while still holding the Heap_lock. 1078 gc_count_before = total_collections(); 1079 should_try_gc = true; 1080 } 1081 } 1082 } 1083 1084 if (should_try_gc) { 1085 // If we failed to allocate the humongous object, we should try to 1086 // do a collection pause (if we're allowed) in case it reclaims 1087 // enough space for the allocation to succeed after the pause. 1088 1089 bool succeeded; 1090 result = do_collection_pause(word_size, gc_count_before, &succeeded, 1091 GCCause::_g1_humongous_allocation); 1092 if (result != NULL) { 1093 assert(succeeded, "only way to get back a non-NULL result"); 1094 return result; 1095 } 1096 1097 if (succeeded) { 1098 // If we get here we successfully scheduled a collection which 1099 // failed to allocate. No point in trying to allocate 1100 // further. We'll just return NULL. 1101 MutexLockerEx x(Heap_lock); 1102 *gc_count_before_ret = total_collections(); 1103 return NULL; 1104 } 1105 } else { 1106 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1107 MutexLockerEx x(Heap_lock); 1108 *gc_count_before_ret = total_collections(); 1109 return NULL; 1110 } 1111 // The GCLocker is either active or the GCLocker initiated 1112 // GC has not yet been performed. Stall until it is and 1113 // then retry the allocation. 1114 GC_locker::stall_until_clear(); 1115 (*gclocker_retry_count_ret) += 1; 1116 } 1117 1118 // We can reach here if we were unsuccessful in scheduling a 1119 // collection (because another thread beat us to it) or if we were 1120 // stalled due to the GC locker. In either can we should retry the 1121 // allocation attempt in case another thread successfully 1122 // performed a collection and reclaimed enough space. Give a 1123 // warning if we seem to be looping forever. 1124 1125 if ((QueuedAllocationWarningCount > 0) && 1126 (try_count % QueuedAllocationWarningCount == 0)) { 1127 warning("G1CollectedHeap::attempt_allocation_humongous() " 1128 "retries %d times", try_count); 1129 } 1130 } 1131 1132 ShouldNotReachHere(); 1133 return NULL; 1134 } 1135 1136 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1137 bool expect_null_mutator_alloc_region) { 1138 assert_at_safepoint(true /* should_be_vm_thread */); 1139 assert(_mutator_alloc_region.get() == NULL || 1140 !expect_null_mutator_alloc_region, 1141 "the current alloc region was unexpectedly found to be non-NULL"); 1142 1143 if (!isHumongous(word_size)) { 1144 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1145 false /* bot_updates */); 1146 } else { 1147 HeapWord* result = humongous_obj_allocate(word_size); 1148 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1149 g1_policy()->set_initiate_conc_mark_if_possible(); 1150 } 1151 return result; 1152 } 1153 1154 ShouldNotReachHere(); 1155 } 1156 1157 class PostMCRemSetClearClosure: public HeapRegionClosure { 1158 G1CollectedHeap* _g1h; 1159 ModRefBarrierSet* _mr_bs; 1160 public: 1161 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1162 _g1h(g1h), _mr_bs(mr_bs) {} 1163 1164 bool doHeapRegion(HeapRegion* r) { 1165 HeapRegionRemSet* hrrs = r->rem_set(); 1166 1167 if (r->continuesHumongous()) { 1168 // We'll assert that the strong code root list and RSet is empty 1169 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 1170 assert(hrrs->occupied() == 0, "RSet should be empty"); 1171 return false; 1172 } 1173 1174 _g1h->reset_gc_time_stamps(r); 1175 hrrs->clear(); 1176 // You might think here that we could clear just the cards 1177 // corresponding to the used region. But no: if we leave a dirty card 1178 // in a region we might allocate into, then it would prevent that card 1179 // from being enqueued, and cause it to be missed. 1180 // Re: the performance cost: we shouldn't be doing full GC anyway! 1181 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1182 1183 return false; 1184 } 1185 }; 1186 1187 void G1CollectedHeap::clear_rsets_post_compaction() { 1188 PostMCRemSetClearClosure rs_clear(this, g1_barrier_set()); 1189 heap_region_iterate(&rs_clear); 1190 } 1191 1192 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1193 G1CollectedHeap* _g1h; 1194 UpdateRSOopClosure _cl; 1195 int _worker_i; 1196 public: 1197 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1198 _cl(g1->g1_rem_set(), worker_i), 1199 _worker_i(worker_i), 1200 _g1h(g1) 1201 { } 1202 1203 bool doHeapRegion(HeapRegion* r) { 1204 if (!r->continuesHumongous()) { 1205 _cl.set_from(r); 1206 r->oop_iterate(&_cl); 1207 } 1208 return false; 1209 } 1210 }; 1211 1212 class ParRebuildRSTask: public AbstractGangTask { 1213 G1CollectedHeap* _g1; 1214 public: 1215 ParRebuildRSTask(G1CollectedHeap* g1) 1216 : AbstractGangTask("ParRebuildRSTask"), 1217 _g1(g1) 1218 { } 1219 1220 void work(uint worker_id) { 1221 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1222 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1223 _g1->workers()->active_workers(), 1224 HeapRegion::RebuildRSClaimValue); 1225 } 1226 }; 1227 1228 class PostCompactionPrinterClosure: public HeapRegionClosure { 1229 private: 1230 G1HRPrinter* _hr_printer; 1231 public: 1232 bool doHeapRegion(HeapRegion* hr) { 1233 assert(!hr->is_young(), "not expecting to find young regions"); 1234 // We only generate output for non-empty regions. 1235 if (!hr->is_empty()) { 1236 if (!hr->isHumongous()) { 1237 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1238 } else if (hr->startsHumongous()) { 1239 if (hr->region_num() == 1) { 1240 // single humongous region 1241 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1242 } else { 1243 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1244 } 1245 } else { 1246 assert(hr->continuesHumongous(), "only way to get here"); 1247 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1248 } 1249 } 1250 return false; 1251 } 1252 1253 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1254 : _hr_printer(hr_printer) { } 1255 }; 1256 1257 void G1CollectedHeap::print_hrs_post_compaction() { 1258 PostCompactionPrinterClosure cl(hr_printer()); 1259 heap_region_iterate(&cl); 1260 } 1261 1262 bool G1CollectedHeap::do_collection(bool explicit_gc, 1263 bool clear_all_soft_refs, 1264 size_t word_size) { 1265 assert_at_safepoint(true /* should_be_vm_thread */); 1266 1267 if (GC_locker::check_active_before_gc()) { 1268 return false; 1269 } 1270 1271 STWGCTimer* gc_timer = G1MarkSweep::gc_timer(); 1272 gc_timer->register_gc_start(); 1273 1274 SerialOldTracer* gc_tracer = G1MarkSweep::gc_tracer(); 1275 gc_tracer->report_gc_start(gc_cause(), gc_timer->gc_start()); 1276 1277 SvcGCMarker sgcm(SvcGCMarker::FULL); 1278 ResourceMark rm; 1279 1280 print_heap_before_gc(); 1281 trace_heap_before_gc(gc_tracer); 1282 1283 size_t metadata_prev_used = MetaspaceAux::used_bytes(); 1284 1285 verify_region_sets_optional(); 1286 1287 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1288 collector_policy()->should_clear_all_soft_refs(); 1289 1290 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1291 1292 { 1293 IsGCActiveMark x; 1294 1295 // Timing 1296 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1297 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1298 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1299 1300 { 1301 GCTraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, NULL, gc_tracer->gc_id()); 1302 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1303 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1304 1305 double start = os::elapsedTime(); 1306 g1_policy()->record_full_collection_start(); 1307 1308 // Note: When we have a more flexible GC logging framework that 1309 // allows us to add optional attributes to a GC log record we 1310 // could consider timing and reporting how long we wait in the 1311 // following two methods. 1312 wait_while_free_regions_coming(); 1313 // If we start the compaction before the CM threads finish 1314 // scanning the root regions we might trip them over as we'll 1315 // be moving objects / updating references. So let's wait until 1316 // they are done. By telling them to abort, they should complete 1317 // early. 1318 _cm->root_regions()->abort(); 1319 _cm->root_regions()->wait_until_scan_finished(); 1320 append_secondary_free_list_if_not_empty_with_lock(); 1321 1322 gc_prologue(true); 1323 increment_total_collections(true /* full gc */); 1324 increment_old_marking_cycles_started(); 1325 1326 assert(used() == recalculate_used(), "Should be equal"); 1327 1328 verify_before_gc(); 1329 1330 pre_full_gc_dump(gc_timer); 1331 1332 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1333 1334 // Disable discovery and empty the discovered lists 1335 // for the CM ref processor. 1336 ref_processor_cm()->disable_discovery(); 1337 ref_processor_cm()->abandon_partial_discovery(); 1338 ref_processor_cm()->verify_no_references_recorded(); 1339 1340 // Abandon current iterations of concurrent marking and concurrent 1341 // refinement, if any are in progress. We have to do this before 1342 // wait_until_scan_finished() below. 1343 concurrent_mark()->abort(); 1344 1345 // Make sure we'll choose a new allocation region afterwards. 1346 release_mutator_alloc_region(); 1347 abandon_gc_alloc_regions(); 1348 g1_rem_set()->cleanupHRRS(); 1349 1350 // We should call this after we retire any currently active alloc 1351 // regions so that all the ALLOC / RETIRE events are generated 1352 // before the start GC event. 1353 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1354 1355 // We may have added regions to the current incremental collection 1356 // set between the last GC or pause and now. We need to clear the 1357 // incremental collection set and then start rebuilding it afresh 1358 // after this full GC. 1359 abandon_collection_set(g1_policy()->inc_cset_head()); 1360 g1_policy()->clear_incremental_cset(); 1361 g1_policy()->stop_incremental_cset_building(); 1362 1363 tear_down_region_sets(false /* free_list_only */); 1364 g1_policy()->set_gcs_are_young(true); 1365 1366 // See the comments in g1CollectedHeap.hpp and 1367 // G1CollectedHeap::ref_processing_init() about 1368 // how reference processing currently works in G1. 1369 1370 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1371 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1372 1373 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1374 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1375 1376 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1377 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1378 1379 // Do collection work 1380 { 1381 HandleMark hm; // Discard invalid handles created during gc 1382 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1383 } 1384 1385 assert(free_regions() == 0, "we should not have added any free regions"); 1386 rebuild_region_sets(false /* free_list_only */); 1387 1388 // Enqueue any discovered reference objects that have 1389 // not been removed from the discovered lists. 1390 ref_processor_stw()->enqueue_discovered_references(); 1391 1392 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1393 1394 MemoryService::track_memory_usage(); 1395 1396 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1397 ref_processor_stw()->verify_no_references_recorded(); 1398 1399 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1400 ClassLoaderDataGraph::purge(); 1401 MetaspaceAux::verify_metrics(); 1402 1403 // Note: since we've just done a full GC, concurrent 1404 // marking is no longer active. Therefore we need not 1405 // re-enable reference discovery for the CM ref processor. 1406 // That will be done at the start of the next marking cycle. 1407 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1408 ref_processor_cm()->verify_no_references_recorded(); 1409 1410 reset_gc_time_stamp(); 1411 // Since everything potentially moved, we will clear all remembered 1412 // sets, and clear all cards. Later we will rebuild remembered 1413 // sets. We will also reset the GC time stamps of the regions. 1414 clear_rsets_post_compaction(); 1415 check_gc_time_stamps(); 1416 1417 // Resize the heap if necessary. 1418 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1419 1420 if (_hr_printer.is_active()) { 1421 // We should do this after we potentially resize the heap so 1422 // that all the COMMIT / UNCOMMIT events are generated before 1423 // the end GC event. 1424 1425 print_hrs_post_compaction(); 1426 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1427 } 1428 1429 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 1430 if (hot_card_cache->use_cache()) { 1431 hot_card_cache->reset_card_counts(); 1432 hot_card_cache->reset_hot_cache(); 1433 } 1434 1435 // Rebuild remembered sets of all regions. 1436 if (G1CollectedHeap::use_parallel_gc_threads()) { 1437 uint n_workers = 1438 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1439 workers()->active_workers(), 1440 Threads::number_of_non_daemon_threads()); 1441 assert(UseDynamicNumberOfGCThreads || 1442 n_workers == workers()->total_workers(), 1443 "If not dynamic should be using all the workers"); 1444 workers()->set_active_workers(n_workers); 1445 // Set parallel threads in the heap (_n_par_threads) only 1446 // before a parallel phase and always reset it to 0 after 1447 // the phase so that the number of parallel threads does 1448 // no get carried forward to a serial phase where there 1449 // may be code that is "possibly_parallel". 1450 set_par_threads(n_workers); 1451 1452 ParRebuildRSTask rebuild_rs_task(this); 1453 assert(check_heap_region_claim_values( 1454 HeapRegion::InitialClaimValue), "sanity check"); 1455 assert(UseDynamicNumberOfGCThreads || 1456 workers()->active_workers() == workers()->total_workers(), 1457 "Unless dynamic should use total workers"); 1458 // Use the most recent number of active workers 1459 assert(workers()->active_workers() > 0, 1460 "Active workers not properly set"); 1461 set_par_threads(workers()->active_workers()); 1462 workers()->run_task(&rebuild_rs_task); 1463 set_par_threads(0); 1464 assert(check_heap_region_claim_values( 1465 HeapRegion::RebuildRSClaimValue), "sanity check"); 1466 reset_heap_region_claim_values(); 1467 } else { 1468 RebuildRSOutOfRegionClosure rebuild_rs(this); 1469 heap_region_iterate(&rebuild_rs); 1470 } 1471 1472 // Rebuild the strong code root lists for each region 1473 rebuild_strong_code_roots(); 1474 1475 if (true) { // FIXME 1476 MetaspaceGC::compute_new_size(); 1477 } 1478 1479 #ifdef TRACESPINNING 1480 ParallelTaskTerminator::print_termination_counts(); 1481 #endif 1482 1483 // Discard all rset updates 1484 JavaThread::dirty_card_queue_set().abandon_logs(); 1485 assert(!G1DeferredRSUpdate 1486 || (G1DeferredRSUpdate && 1487 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1488 1489 _young_list->reset_sampled_info(); 1490 // At this point there should be no regions in the 1491 // entire heap tagged as young. 1492 assert(check_young_list_empty(true /* check_heap */), 1493 "young list should be empty at this point"); 1494 1495 // Update the number of full collections that have been completed. 1496 increment_old_marking_cycles_completed(false /* concurrent */); 1497 1498 _hrs.verify_optional(); 1499 verify_region_sets_optional(); 1500 1501 verify_after_gc(); 1502 1503 // Start a new incremental collection set for the next pause 1504 assert(g1_policy()->collection_set() == NULL, "must be"); 1505 g1_policy()->start_incremental_cset_building(); 1506 1507 clear_cset_fast_test(); 1508 1509 init_mutator_alloc_region(); 1510 1511 double end = os::elapsedTime(); 1512 g1_policy()->record_full_collection_end(); 1513 1514 if (G1Log::fine()) { 1515 g1_policy()->print_heap_transition(); 1516 } 1517 1518 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1519 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1520 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1521 // before any GC notifications are raised. 1522 g1mm()->update_sizes(); 1523 1524 gc_epilogue(true); 1525 } 1526 1527 if (G1Log::finer()) { 1528 g1_policy()->print_detailed_heap_transition(true /* full */); 1529 } 1530 1531 print_heap_after_gc(); 1532 trace_heap_after_gc(gc_tracer); 1533 1534 post_full_gc_dump(gc_timer); 1535 1536 gc_timer->register_gc_end(); 1537 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1538 } 1539 1540 return true; 1541 } 1542 1543 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1544 // do_collection() will return whether it succeeded in performing 1545 // the GC. Currently, there is no facility on the 1546 // do_full_collection() API to notify the caller than the collection 1547 // did not succeed (e.g., because it was locked out by the GC 1548 // locker). So, right now, we'll ignore the return value. 1549 bool dummy = do_collection(true, /* explicit_gc */ 1550 clear_all_soft_refs, 1551 0 /* word_size */); 1552 } 1553 1554 // This code is mostly copied from TenuredGeneration. 1555 void 1556 G1CollectedHeap:: 1557 resize_if_necessary_after_full_collection(size_t word_size) { 1558 // Include the current allocation, if any, and bytes that will be 1559 // pre-allocated to support collections, as "used". 1560 const size_t used_after_gc = used(); 1561 const size_t capacity_after_gc = capacity(); 1562 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1563 1564 // This is enforced in arguments.cpp. 1565 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1566 "otherwise the code below doesn't make sense"); 1567 1568 // We don't have floating point command-line arguments 1569 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1570 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1571 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1572 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1573 1574 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1575 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1576 1577 // We have to be careful here as these two calculations can overflow 1578 // 32-bit size_t's. 1579 double used_after_gc_d = (double) used_after_gc; 1580 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1581 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1582 1583 // Let's make sure that they are both under the max heap size, which 1584 // by default will make them fit into a size_t. 1585 double desired_capacity_upper_bound = (double) max_heap_size; 1586 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1587 desired_capacity_upper_bound); 1588 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1589 desired_capacity_upper_bound); 1590 1591 // We can now safely turn them into size_t's. 1592 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1593 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1594 1595 // This assert only makes sense here, before we adjust them 1596 // with respect to the min and max heap size. 1597 assert(minimum_desired_capacity <= maximum_desired_capacity, 1598 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1599 "maximum_desired_capacity = "SIZE_FORMAT, 1600 minimum_desired_capacity, maximum_desired_capacity)); 1601 1602 // Should not be greater than the heap max size. No need to adjust 1603 // it with respect to the heap min size as it's a lower bound (i.e., 1604 // we'll try to make the capacity larger than it, not smaller). 1605 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1606 // Should not be less than the heap min size. No need to adjust it 1607 // with respect to the heap max size as it's an upper bound (i.e., 1608 // we'll try to make the capacity smaller than it, not greater). 1609 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1610 1611 if (capacity_after_gc < minimum_desired_capacity) { 1612 // Don't expand unless it's significant 1613 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1614 ergo_verbose4(ErgoHeapSizing, 1615 "attempt heap expansion", 1616 ergo_format_reason("capacity lower than " 1617 "min desired capacity after Full GC") 1618 ergo_format_byte("capacity") 1619 ergo_format_byte("occupancy") 1620 ergo_format_byte_perc("min desired capacity"), 1621 capacity_after_gc, used_after_gc, 1622 minimum_desired_capacity, (double) MinHeapFreeRatio); 1623 expand(expand_bytes); 1624 1625 // No expansion, now see if we want to shrink 1626 } else if (capacity_after_gc > maximum_desired_capacity) { 1627 // Capacity too large, compute shrinking size 1628 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1629 ergo_verbose4(ErgoHeapSizing, 1630 "attempt heap shrinking", 1631 ergo_format_reason("capacity higher than " 1632 "max desired capacity after Full GC") 1633 ergo_format_byte("capacity") 1634 ergo_format_byte("occupancy") 1635 ergo_format_byte_perc("max desired capacity"), 1636 capacity_after_gc, used_after_gc, 1637 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1638 shrink(shrink_bytes); 1639 } 1640 } 1641 1642 1643 HeapWord* 1644 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1645 bool* succeeded) { 1646 assert_at_safepoint(true /* should_be_vm_thread */); 1647 1648 *succeeded = true; 1649 // Let's attempt the allocation first. 1650 HeapWord* result = 1651 attempt_allocation_at_safepoint(word_size, 1652 false /* expect_null_mutator_alloc_region */); 1653 if (result != NULL) { 1654 assert(*succeeded, "sanity"); 1655 return result; 1656 } 1657 1658 // In a G1 heap, we're supposed to keep allocation from failing by 1659 // incremental pauses. Therefore, at least for now, we'll favor 1660 // expansion over collection. (This might change in the future if we can 1661 // do something smarter than full collection to satisfy a failed alloc.) 1662 result = expand_and_allocate(word_size); 1663 if (result != NULL) { 1664 assert(*succeeded, "sanity"); 1665 return result; 1666 } 1667 1668 // Expansion didn't work, we'll try to do a Full GC. 1669 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1670 false, /* clear_all_soft_refs */ 1671 word_size); 1672 if (!gc_succeeded) { 1673 *succeeded = false; 1674 return NULL; 1675 } 1676 1677 // Retry the allocation 1678 result = attempt_allocation_at_safepoint(word_size, 1679 true /* expect_null_mutator_alloc_region */); 1680 if (result != NULL) { 1681 assert(*succeeded, "sanity"); 1682 return result; 1683 } 1684 1685 // Then, try a Full GC that will collect all soft references. 1686 gc_succeeded = do_collection(false, /* explicit_gc */ 1687 true, /* clear_all_soft_refs */ 1688 word_size); 1689 if (!gc_succeeded) { 1690 *succeeded = false; 1691 return NULL; 1692 } 1693 1694 // Retry the allocation once more 1695 result = attempt_allocation_at_safepoint(word_size, 1696 true /* expect_null_mutator_alloc_region */); 1697 if (result != NULL) { 1698 assert(*succeeded, "sanity"); 1699 return result; 1700 } 1701 1702 assert(!collector_policy()->should_clear_all_soft_refs(), 1703 "Flag should have been handled and cleared prior to this point"); 1704 1705 // What else? We might try synchronous finalization later. If the total 1706 // space available is large enough for the allocation, then a more 1707 // complete compaction phase than we've tried so far might be 1708 // appropriate. 1709 assert(*succeeded, "sanity"); 1710 return NULL; 1711 } 1712 1713 // Attempting to expand the heap sufficiently 1714 // to support an allocation of the given "word_size". If 1715 // successful, perform the allocation and return the address of the 1716 // allocated block, or else "NULL". 1717 1718 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1719 assert_at_safepoint(true /* should_be_vm_thread */); 1720 1721 verify_region_sets_optional(); 1722 1723 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1724 ergo_verbose1(ErgoHeapSizing, 1725 "attempt heap expansion", 1726 ergo_format_reason("allocation request failed") 1727 ergo_format_byte("allocation request"), 1728 word_size * HeapWordSize); 1729 if (expand(expand_bytes)) { 1730 _hrs.verify_optional(); 1731 verify_region_sets_optional(); 1732 return attempt_allocation_at_safepoint(word_size, 1733 false /* expect_null_mutator_alloc_region */); 1734 } 1735 return NULL; 1736 } 1737 1738 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1739 HeapWord* new_end) { 1740 assert(old_end != new_end, "don't call this otherwise"); 1741 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1742 1743 // Update the committed mem region. 1744 _g1_committed.set_end(new_end); 1745 // Tell the card table about the update. 1746 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1747 // Tell the BOT about the update. 1748 _bot_shared->resize(_g1_committed.word_size()); 1749 // Tell the hot card cache about the update 1750 _cg1r->hot_card_cache()->resize_card_counts(capacity()); 1751 } 1752 1753 bool G1CollectedHeap::expand(size_t expand_bytes) { 1754 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1755 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1756 HeapRegion::GrainBytes); 1757 ergo_verbose2(ErgoHeapSizing, 1758 "expand the heap", 1759 ergo_format_byte("requested expansion amount") 1760 ergo_format_byte("attempted expansion amount"), 1761 expand_bytes, aligned_expand_bytes); 1762 1763 if (_g1_storage.uncommitted_size() == 0) { 1764 ergo_verbose0(ErgoHeapSizing, 1765 "did not expand the heap", 1766 ergo_format_reason("heap already fully expanded")); 1767 return false; 1768 } 1769 1770 // First commit the memory. 1771 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1772 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1773 if (successful) { 1774 // Then propagate this update to the necessary data structures. 1775 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1776 update_committed_space(old_end, new_end); 1777 1778 FreeRegionList expansion_list("Local Expansion List"); 1779 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1780 assert(mr.start() == old_end, "post-condition"); 1781 // mr might be a smaller region than what was requested if 1782 // expand_by() was unable to allocate the HeapRegion instances 1783 assert(mr.end() <= new_end, "post-condition"); 1784 1785 size_t actual_expand_bytes = mr.byte_size(); 1786 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1787 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1788 "post-condition"); 1789 if (actual_expand_bytes < aligned_expand_bytes) { 1790 // We could not expand _hrs to the desired size. In this case we 1791 // need to shrink the committed space accordingly. 1792 assert(mr.end() < new_end, "invariant"); 1793 1794 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1795 // First uncommit the memory. 1796 _g1_storage.shrink_by(diff_bytes); 1797 // Then propagate this update to the necessary data structures. 1798 update_committed_space(new_end, mr.end()); 1799 } 1800 _free_list.add_as_tail(&expansion_list); 1801 1802 if (_hr_printer.is_active()) { 1803 HeapWord* curr = mr.start(); 1804 while (curr < mr.end()) { 1805 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1806 _hr_printer.commit(curr, curr_end); 1807 curr = curr_end; 1808 } 1809 assert(curr == mr.end(), "post-condition"); 1810 } 1811 g1_policy()->record_new_heap_size(n_regions()); 1812 } else { 1813 ergo_verbose0(ErgoHeapSizing, 1814 "did not expand the heap", 1815 ergo_format_reason("heap expansion operation failed")); 1816 // The expansion of the virtual storage space was unsuccessful. 1817 // Let's see if it was because we ran out of swap. 1818 if (G1ExitOnExpansionFailure && 1819 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1820 // We had head room... 1821 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion"); 1822 } 1823 } 1824 return successful; 1825 } 1826 1827 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1828 size_t aligned_shrink_bytes = 1829 ReservedSpace::page_align_size_down(shrink_bytes); 1830 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1831 HeapRegion::GrainBytes); 1832 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes); 1833 1834 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove); 1835 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1836 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes; 1837 1838 ergo_verbose3(ErgoHeapSizing, 1839 "shrink the heap", 1840 ergo_format_byte("requested shrinking amount") 1841 ergo_format_byte("aligned shrinking amount") 1842 ergo_format_byte("attempted shrinking amount"), 1843 shrink_bytes, aligned_shrink_bytes, shrunk_bytes); 1844 if (num_regions_removed > 0) { 1845 _g1_storage.shrink_by(shrunk_bytes); 1846 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1847 1848 if (_hr_printer.is_active()) { 1849 HeapWord* curr = old_end; 1850 while (curr > new_end) { 1851 HeapWord* curr_end = curr; 1852 curr -= HeapRegion::GrainWords; 1853 _hr_printer.uncommit(curr, curr_end); 1854 } 1855 } 1856 1857 _expansion_regions += num_regions_removed; 1858 update_committed_space(old_end, new_end); 1859 HeapRegionRemSet::shrink_heap(n_regions()); 1860 g1_policy()->record_new_heap_size(n_regions()); 1861 } else { 1862 ergo_verbose0(ErgoHeapSizing, 1863 "did not shrink the heap", 1864 ergo_format_reason("heap shrinking operation failed")); 1865 } 1866 } 1867 1868 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1869 verify_region_sets_optional(); 1870 1871 // We should only reach here at the end of a Full GC which means we 1872 // should not not be holding to any GC alloc regions. The method 1873 // below will make sure of that and do any remaining clean up. 1874 abandon_gc_alloc_regions(); 1875 1876 // Instead of tearing down / rebuilding the free lists here, we 1877 // could instead use the remove_all_pending() method on free_list to 1878 // remove only the ones that we need to remove. 1879 tear_down_region_sets(true /* free_list_only */); 1880 shrink_helper(shrink_bytes); 1881 rebuild_region_sets(true /* free_list_only */); 1882 1883 _hrs.verify_optional(); 1884 verify_region_sets_optional(); 1885 } 1886 1887 // Public methods. 1888 1889 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1890 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1891 #endif // _MSC_VER 1892 1893 1894 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1895 SharedHeap(policy_), 1896 _g1_policy(policy_), 1897 _dirty_card_queue_set(false), 1898 _into_cset_dirty_card_queue_set(false), 1899 _is_alive_closure_cm(this), 1900 _is_alive_closure_stw(this), 1901 _ref_processor_cm(NULL), 1902 _ref_processor_stw(NULL), 1903 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1904 _bot_shared(NULL), 1905 _evac_failure_scan_stack(NULL), 1906 _mark_in_progress(false), 1907 _cg1r(NULL), _summary_bytes_used(0), 1908 _g1mm(NULL), 1909 _refine_cte_cl(NULL), 1910 _full_collection(false), 1911 _free_list("Master Free List", new MasterFreeRegionListMtSafeChecker()), 1912 _secondary_free_list("Secondary Free List", new SecondaryFreeRegionListMtSafeChecker()), 1913 _old_set("Old Set", false /* humongous */, new OldRegionSetMtSafeChecker()), 1914 _humongous_set("Master Humongous Set", true /* humongous */, new HumongousRegionSetMtSafeChecker()), 1915 _free_regions_coming(false), 1916 _young_list(new YoungList(this)), 1917 _gc_time_stamp(0), 1918 _retained_old_gc_alloc_region(NULL), 1919 _survivor_plab_stats(YoungPLABSize, PLABWeight), 1920 _old_plab_stats(OldPLABSize, PLABWeight), 1921 _expand_heap_after_alloc_failure(true), 1922 _surviving_young_words(NULL), 1923 _old_marking_cycles_started(0), 1924 _old_marking_cycles_completed(0), 1925 _concurrent_cycle_started(false), 1926 _in_cset_fast_test(), 1927 _dirty_cards_region_list(NULL), 1928 _worker_cset_start_region(NULL), 1929 _worker_cset_start_region_time_stamp(NULL), 1930 _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()), 1931 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 1932 _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()), 1933 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()) { 1934 1935 _g1h = this; 1936 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1937 vm_exit_during_initialization("Failed necessary allocation."); 1938 } 1939 1940 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1941 1942 int n_queues = MAX2((int)ParallelGCThreads, 1); 1943 _task_queues = new RefToScanQueueSet(n_queues); 1944 1945 uint n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1946 assert(n_rem_sets > 0, "Invariant."); 1947 1948 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1949 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1950 _evacuation_failed_info_array = NEW_C_HEAP_ARRAY(EvacuationFailedInfo, n_queues, mtGC); 1951 1952 for (int i = 0; i < n_queues; i++) { 1953 RefToScanQueue* q = new RefToScanQueue(); 1954 q->initialize(); 1955 _task_queues->register_queue(i, q); 1956 ::new (&_evacuation_failed_info_array[i]) EvacuationFailedInfo(); 1957 } 1958 clear_cset_start_regions(); 1959 1960 // Initialize the G1EvacuationFailureALot counters and flags. 1961 NOT_PRODUCT(reset_evacuation_should_fail();) 1962 1963 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1964 } 1965 1966 jint G1CollectedHeap::initialize() { 1967 CollectedHeap::pre_initialize(); 1968 os::enable_vtime(); 1969 1970 G1Log::init(); 1971 1972 // Necessary to satisfy locking discipline assertions. 1973 1974 MutexLocker x(Heap_lock); 1975 1976 // We have to initialize the printer before committing the heap, as 1977 // it will be used then. 1978 _hr_printer.set_active(G1PrintHeapRegions); 1979 1980 // While there are no constraints in the GC code that HeapWordSize 1981 // be any particular value, there are multiple other areas in the 1982 // system which believe this to be true (e.g. oop->object_size in some 1983 // cases incorrectly returns the size in wordSize units rather than 1984 // HeapWordSize). 1985 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1986 1987 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1988 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1989 size_t heap_alignment = collector_policy()->heap_alignment(); 1990 1991 // Ensure that the sizes are properly aligned. 1992 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1993 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1994 Universe::check_alignment(max_byte_size, heap_alignment, "g1 heap"); 1995 1996 _refine_cte_cl = new RefineCardTableEntryClosure(); 1997 1998 _cg1r = new ConcurrentG1Refine(this, _refine_cte_cl); 1999 2000 // Reserve the maximum. 2001 2002 // When compressed oops are enabled, the preferred heap base 2003 // is calculated by subtracting the requested size from the 2004 // 32Gb boundary and using the result as the base address for 2005 // heap reservation. If the requested size is not aligned to 2006 // HeapRegion::GrainBytes (i.e. the alignment that is passed 2007 // into the ReservedHeapSpace constructor) then the actual 2008 // base of the reserved heap may end up differing from the 2009 // address that was requested (i.e. the preferred heap base). 2010 // If this happens then we could end up using a non-optimal 2011 // compressed oops mode. 2012 2013 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2014 heap_alignment); 2015 2016 // It is important to do this in a way such that concurrent readers can't 2017 // temporarily think something is in the heap. (I've actually seen this 2018 // happen in asserts: DLD.) 2019 _reserved.set_word_size(0); 2020 _reserved.set_start((HeapWord*)heap_rs.base()); 2021 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2022 2023 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2024 2025 // Create the gen rem set (and barrier set) for the entire reserved region. 2026 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2027 set_barrier_set(rem_set()->bs()); 2028 if (!barrier_set()->is_a(BarrierSet::G1SATBCTLogging)) { 2029 vm_exit_during_initialization("G1 requires a G1SATBLoggingCardTableModRefBS"); 2030 return JNI_ENOMEM; 2031 } 2032 2033 // Also create a G1 rem set. 2034 _g1_rem_set = new G1RemSet(this, g1_barrier_set()); 2035 2036 // Carve out the G1 part of the heap. 2037 2038 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2039 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2040 g1_rs.size()/HeapWordSize); 2041 2042 _g1_storage.initialize(g1_rs, 0); 2043 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2044 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2045 (HeapWord*) _g1_reserved.end()); 2046 assert(_hrs.max_length() == _expansion_regions, 2047 err_msg("max length: %u expansion regions: %u", 2048 _hrs.max_length(), _expansion_regions)); 2049 2050 // Do later initialization work for concurrent refinement. 2051 _cg1r->init(); 2052 2053 // 6843694 - ensure that the maximum region index can fit 2054 // in the remembered set structures. 2055 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2056 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2057 2058 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2059 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2060 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2061 "too many cards per region"); 2062 2063 FreeRegionList::set_unrealistically_long_length(max_regions() + 1); 2064 2065 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2066 heap_word_size(init_byte_size)); 2067 2068 _g1h = this; 2069 2070 _in_cset_fast_test.initialize(_g1_reserved.start(), _g1_reserved.end(), HeapRegion::GrainBytes); 2071 2072 // Create the ConcurrentMark data structure and thread. 2073 // (Must do this late, so that "max_regions" is defined.) 2074 _cm = new ConcurrentMark(this, heap_rs); 2075 if (_cm == NULL || !_cm->completed_initialization()) { 2076 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2077 return JNI_ENOMEM; 2078 } 2079 _cmThread = _cm->cmThread(); 2080 2081 // Initialize the from_card cache structure of HeapRegionRemSet. 2082 HeapRegionRemSet::init_heap(max_regions()); 2083 2084 // Now expand into the initial heap size. 2085 if (!expand(init_byte_size)) { 2086 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2087 return JNI_ENOMEM; 2088 } 2089 2090 // Perform any initialization actions delegated to the policy. 2091 g1_policy()->init(); 2092 2093 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2094 SATB_Q_FL_lock, 2095 G1SATBProcessCompletedThreshold, 2096 Shared_SATB_Q_lock); 2097 2098 JavaThread::dirty_card_queue_set().initialize(_refine_cte_cl, 2099 DirtyCardQ_CBL_mon, 2100 DirtyCardQ_FL_lock, 2101 concurrent_g1_refine()->yellow_zone(), 2102 concurrent_g1_refine()->red_zone(), 2103 Shared_DirtyCardQ_lock); 2104 2105 if (G1DeferredRSUpdate) { 2106 dirty_card_queue_set().initialize(NULL, // Should never be called by the Java code 2107 DirtyCardQ_CBL_mon, 2108 DirtyCardQ_FL_lock, 2109 -1, // never trigger processing 2110 -1, // no limit on length 2111 Shared_DirtyCardQ_lock, 2112 &JavaThread::dirty_card_queue_set()); 2113 } 2114 2115 // Initialize the card queue set used to hold cards containing 2116 // references into the collection set. 2117 _into_cset_dirty_card_queue_set.initialize(NULL, // Should never be called by the Java code 2118 DirtyCardQ_CBL_mon, 2119 DirtyCardQ_FL_lock, 2120 -1, // never trigger processing 2121 -1, // no limit on length 2122 Shared_DirtyCardQ_lock, 2123 &JavaThread::dirty_card_queue_set()); 2124 2125 // In case we're keeping closure specialization stats, initialize those 2126 // counts and that mechanism. 2127 SpecializationStats::clear(); 2128 2129 // Here we allocate the dummy full region that is required by the 2130 // G1AllocRegion class. If we don't pass an address in the reserved 2131 // space here, lots of asserts fire. 2132 2133 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2134 _g1_reserved.start()); 2135 // We'll re-use the same region whether the alloc region will 2136 // require BOT updates or not and, if it doesn't, then a non-young 2137 // region will complain that it cannot support allocations without 2138 // BOT updates. So we'll tag the dummy region as young to avoid that. 2139 dummy_region->set_young(); 2140 // Make sure it's full. 2141 dummy_region->set_top(dummy_region->end()); 2142 G1AllocRegion::setup(this, dummy_region); 2143 2144 init_mutator_alloc_region(); 2145 2146 // Do create of the monitoring and management support so that 2147 // values in the heap have been properly initialized. 2148 _g1mm = new G1MonitoringSupport(this); 2149 2150 G1StringDedup::initialize(); 2151 2152 return JNI_OK; 2153 } 2154 2155 void G1CollectedHeap::stop() { 2156 // Stop all concurrent threads. We do this to make sure these threads 2157 // do not continue to execute and access resources (e.g. gclog_or_tty) 2158 // that are destroyed during shutdown. 2159 _cg1r->stop(); 2160 _cmThread->stop(); 2161 if (G1StringDedup::is_enabled()) { 2162 G1StringDedup::stop(); 2163 } 2164 } 2165 2166 size_t G1CollectedHeap::conservative_max_heap_alignment() { 2167 return HeapRegion::max_region_size(); 2168 } 2169 2170 void G1CollectedHeap::ref_processing_init() { 2171 // Reference processing in G1 currently works as follows: 2172 // 2173 // * There are two reference processor instances. One is 2174 // used to record and process discovered references 2175 // during concurrent marking; the other is used to 2176 // record and process references during STW pauses 2177 // (both full and incremental). 2178 // * Both ref processors need to 'span' the entire heap as 2179 // the regions in the collection set may be dotted around. 2180 // 2181 // * For the concurrent marking ref processor: 2182 // * Reference discovery is enabled at initial marking. 2183 // * Reference discovery is disabled and the discovered 2184 // references processed etc during remarking. 2185 // * Reference discovery is MT (see below). 2186 // * Reference discovery requires a barrier (see below). 2187 // * Reference processing may or may not be MT 2188 // (depending on the value of ParallelRefProcEnabled 2189 // and ParallelGCThreads). 2190 // * A full GC disables reference discovery by the CM 2191 // ref processor and abandons any entries on it's 2192 // discovered lists. 2193 // 2194 // * For the STW processor: 2195 // * Non MT discovery is enabled at the start of a full GC. 2196 // * Processing and enqueueing during a full GC is non-MT. 2197 // * During a full GC, references are processed after marking. 2198 // 2199 // * Discovery (may or may not be MT) is enabled at the start 2200 // of an incremental evacuation pause. 2201 // * References are processed near the end of a STW evacuation pause. 2202 // * For both types of GC: 2203 // * Discovery is atomic - i.e. not concurrent. 2204 // * Reference discovery will not need a barrier. 2205 2206 SharedHeap::ref_processing_init(); 2207 MemRegion mr = reserved_region(); 2208 2209 // Concurrent Mark ref processor 2210 _ref_processor_cm = 2211 new ReferenceProcessor(mr, // span 2212 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2213 // mt processing 2214 (int) ParallelGCThreads, 2215 // degree of mt processing 2216 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2217 // mt discovery 2218 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2219 // degree of mt discovery 2220 false, 2221 // Reference discovery is not atomic 2222 &_is_alive_closure_cm); 2223 // is alive closure 2224 // (for efficiency/performance) 2225 2226 // STW ref processor 2227 _ref_processor_stw = 2228 new ReferenceProcessor(mr, // span 2229 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2230 // mt processing 2231 MAX2((int)ParallelGCThreads, 1), 2232 // degree of mt processing 2233 (ParallelGCThreads > 1), 2234 // mt discovery 2235 MAX2((int)ParallelGCThreads, 1), 2236 // degree of mt discovery 2237 true, 2238 // Reference discovery is atomic 2239 &_is_alive_closure_stw); 2240 // is alive closure 2241 // (for efficiency/performance) 2242 } 2243 2244 size_t G1CollectedHeap::capacity() const { 2245 return _g1_committed.byte_size(); 2246 } 2247 2248 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2249 assert(!hr->continuesHumongous(), "pre-condition"); 2250 hr->reset_gc_time_stamp(); 2251 if (hr->startsHumongous()) { 2252 uint first_index = hr->hrs_index() + 1; 2253 uint last_index = hr->last_hc_index(); 2254 for (uint i = first_index; i < last_index; i += 1) { 2255 HeapRegion* chr = region_at(i); 2256 assert(chr->continuesHumongous(), "sanity"); 2257 chr->reset_gc_time_stamp(); 2258 } 2259 } 2260 } 2261 2262 #ifndef PRODUCT 2263 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2264 private: 2265 unsigned _gc_time_stamp; 2266 bool _failures; 2267 2268 public: 2269 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2270 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2271 2272 virtual bool doHeapRegion(HeapRegion* hr) { 2273 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2274 if (_gc_time_stamp != region_gc_time_stamp) { 2275 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2276 "expected %d", HR_FORMAT_PARAMS(hr), 2277 region_gc_time_stamp, _gc_time_stamp); 2278 _failures = true; 2279 } 2280 return false; 2281 } 2282 2283 bool failures() { return _failures; } 2284 }; 2285 2286 void G1CollectedHeap::check_gc_time_stamps() { 2287 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2288 heap_region_iterate(&cl); 2289 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2290 } 2291 #endif // PRODUCT 2292 2293 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2294 DirtyCardQueue* into_cset_dcq, 2295 bool concurrent, 2296 uint worker_i) { 2297 // Clean cards in the hot card cache 2298 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 2299 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq); 2300 2301 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2302 int n_completed_buffers = 0; 2303 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2304 n_completed_buffers++; 2305 } 2306 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2307 dcqs.clear_n_completed_buffers(); 2308 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2309 } 2310 2311 2312 // Computes the sum of the storage used by the various regions. 2313 2314 size_t G1CollectedHeap::used() const { 2315 assert(Heap_lock->owner() != NULL, 2316 "Should be owned on this thread's behalf."); 2317 size_t result = _summary_bytes_used; 2318 // Read only once in case it is set to NULL concurrently 2319 HeapRegion* hr = _mutator_alloc_region.get(); 2320 if (hr != NULL) 2321 result += hr->used(); 2322 return result; 2323 } 2324 2325 size_t G1CollectedHeap::used_unlocked() const { 2326 size_t result = _summary_bytes_used; 2327 return result; 2328 } 2329 2330 class SumUsedClosure: public HeapRegionClosure { 2331 size_t _used; 2332 public: 2333 SumUsedClosure() : _used(0) {} 2334 bool doHeapRegion(HeapRegion* r) { 2335 if (!r->continuesHumongous()) { 2336 _used += r->used(); 2337 } 2338 return false; 2339 } 2340 size_t result() { return _used; } 2341 }; 2342 2343 size_t G1CollectedHeap::recalculate_used() const { 2344 double recalculate_used_start = os::elapsedTime(); 2345 2346 SumUsedClosure blk; 2347 heap_region_iterate(&blk); 2348 2349 g1_policy()->phase_times()->record_evac_fail_recalc_used_time((os::elapsedTime() - recalculate_used_start) * 1000.0); 2350 return blk.result(); 2351 } 2352 2353 size_t G1CollectedHeap::unsafe_max_alloc() { 2354 if (free_regions() > 0) return HeapRegion::GrainBytes; 2355 // otherwise, is there space in the current allocation region? 2356 2357 // We need to store the current allocation region in a local variable 2358 // here. The problem is that this method doesn't take any locks and 2359 // there may be other threads which overwrite the current allocation 2360 // region field. attempt_allocation(), for example, sets it to NULL 2361 // and this can happen *after* the NULL check here but before the call 2362 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2363 // to be a problem in the optimized build, since the two loads of the 2364 // current allocation region field are optimized away. 2365 HeapRegion* hr = _mutator_alloc_region.get(); 2366 if (hr == NULL) { 2367 return 0; 2368 } 2369 return hr->free(); 2370 } 2371 2372 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2373 switch (cause) { 2374 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2375 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2376 case GCCause::_g1_humongous_allocation: return true; 2377 default: return false; 2378 } 2379 } 2380 2381 #ifndef PRODUCT 2382 void G1CollectedHeap::allocate_dummy_regions() { 2383 // Let's fill up most of the region 2384 size_t word_size = HeapRegion::GrainWords - 1024; 2385 // And as a result the region we'll allocate will be humongous. 2386 guarantee(isHumongous(word_size), "sanity"); 2387 2388 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2389 // Let's use the existing mechanism for the allocation 2390 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2391 if (dummy_obj != NULL) { 2392 MemRegion mr(dummy_obj, word_size); 2393 CollectedHeap::fill_with_object(mr); 2394 } else { 2395 // If we can't allocate once, we probably cannot allocate 2396 // again. Let's get out of the loop. 2397 break; 2398 } 2399 } 2400 } 2401 #endif // !PRODUCT 2402 2403 void G1CollectedHeap::increment_old_marking_cycles_started() { 2404 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2405 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2406 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2407 _old_marking_cycles_started, _old_marking_cycles_completed)); 2408 2409 _old_marking_cycles_started++; 2410 } 2411 2412 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2413 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2414 2415 // We assume that if concurrent == true, then the caller is a 2416 // concurrent thread that was joined the Suspendible Thread 2417 // Set. If there's ever a cheap way to check this, we should add an 2418 // assert here. 2419 2420 // Given that this method is called at the end of a Full GC or of a 2421 // concurrent cycle, and those can be nested (i.e., a Full GC can 2422 // interrupt a concurrent cycle), the number of full collections 2423 // completed should be either one (in the case where there was no 2424 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2425 // behind the number of full collections started. 2426 2427 // This is the case for the inner caller, i.e. a Full GC. 2428 assert(concurrent || 2429 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2430 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2431 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2432 "is inconsistent with _old_marking_cycles_completed = %u", 2433 _old_marking_cycles_started, _old_marking_cycles_completed)); 2434 2435 // This is the case for the outer caller, i.e. the concurrent cycle. 2436 assert(!concurrent || 2437 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2438 err_msg("for outer caller (concurrent cycle): " 2439 "_old_marking_cycles_started = %u " 2440 "is inconsistent with _old_marking_cycles_completed = %u", 2441 _old_marking_cycles_started, _old_marking_cycles_completed)); 2442 2443 _old_marking_cycles_completed += 1; 2444 2445 // We need to clear the "in_progress" flag in the CM thread before 2446 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2447 // is set) so that if a waiter requests another System.gc() it doesn't 2448 // incorrectly see that a marking cycle is still in progress. 2449 if (concurrent) { 2450 _cmThread->clear_in_progress(); 2451 } 2452 2453 // This notify_all() will ensure that a thread that called 2454 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2455 // and it's waiting for a full GC to finish will be woken up. It is 2456 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2457 FullGCCount_lock->notify_all(); 2458 } 2459 2460 void G1CollectedHeap::register_concurrent_cycle_start(const Ticks& start_time) { 2461 _concurrent_cycle_started = true; 2462 _gc_timer_cm->register_gc_start(start_time); 2463 2464 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start()); 2465 trace_heap_before_gc(_gc_tracer_cm); 2466 } 2467 2468 void G1CollectedHeap::register_concurrent_cycle_end() { 2469 if (_concurrent_cycle_started) { 2470 if (_cm->has_aborted()) { 2471 _gc_tracer_cm->report_concurrent_mode_failure(); 2472 } 2473 2474 _gc_timer_cm->register_gc_end(); 2475 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2476 2477 _concurrent_cycle_started = false; 2478 } 2479 } 2480 2481 void G1CollectedHeap::trace_heap_after_concurrent_cycle() { 2482 if (_concurrent_cycle_started) { 2483 trace_heap_after_gc(_gc_tracer_cm); 2484 } 2485 } 2486 2487 G1YCType G1CollectedHeap::yc_type() { 2488 bool is_young = g1_policy()->gcs_are_young(); 2489 bool is_initial_mark = g1_policy()->during_initial_mark_pause(); 2490 bool is_during_mark = mark_in_progress(); 2491 2492 if (is_initial_mark) { 2493 return InitialMark; 2494 } else if (is_during_mark) { 2495 return DuringMark; 2496 } else if (is_young) { 2497 return Normal; 2498 } else { 2499 return Mixed; 2500 } 2501 } 2502 2503 void G1CollectedHeap::collect(GCCause::Cause cause) { 2504 assert_heap_not_locked(); 2505 2506 unsigned int gc_count_before; 2507 unsigned int old_marking_count_before; 2508 bool retry_gc; 2509 2510 do { 2511 retry_gc = false; 2512 2513 { 2514 MutexLocker ml(Heap_lock); 2515 2516 // Read the GC count while holding the Heap_lock 2517 gc_count_before = total_collections(); 2518 old_marking_count_before = _old_marking_cycles_started; 2519 } 2520 2521 if (should_do_concurrent_full_gc(cause)) { 2522 // Schedule an initial-mark evacuation pause that will start a 2523 // concurrent cycle. We're setting word_size to 0 which means that 2524 // we are not requesting a post-GC allocation. 2525 VM_G1IncCollectionPause op(gc_count_before, 2526 0, /* word_size */ 2527 true, /* should_initiate_conc_mark */ 2528 g1_policy()->max_pause_time_ms(), 2529 cause); 2530 2531 VMThread::execute(&op); 2532 if (!op.pause_succeeded()) { 2533 if (old_marking_count_before == _old_marking_cycles_started) { 2534 retry_gc = op.should_retry_gc(); 2535 } else { 2536 // A Full GC happened while we were trying to schedule the 2537 // initial-mark GC. No point in starting a new cycle given 2538 // that the whole heap was collected anyway. 2539 } 2540 2541 if (retry_gc) { 2542 if (GC_locker::is_active_and_needs_gc()) { 2543 GC_locker::stall_until_clear(); 2544 } 2545 } 2546 } 2547 } else { 2548 if (cause == GCCause::_gc_locker 2549 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2550 2551 // Schedule a standard evacuation pause. We're setting word_size 2552 // to 0 which means that we are not requesting a post-GC allocation. 2553 VM_G1IncCollectionPause op(gc_count_before, 2554 0, /* word_size */ 2555 false, /* should_initiate_conc_mark */ 2556 g1_policy()->max_pause_time_ms(), 2557 cause); 2558 VMThread::execute(&op); 2559 } else { 2560 // Schedule a Full GC. 2561 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2562 VMThread::execute(&op); 2563 } 2564 } 2565 } while (retry_gc); 2566 } 2567 2568 bool G1CollectedHeap::is_in(const void* p) const { 2569 if (_g1_committed.contains(p)) { 2570 // Given that we know that p is in the committed space, 2571 // heap_region_containing_raw() should successfully 2572 // return the containing region. 2573 HeapRegion* hr = heap_region_containing_raw(p); 2574 return hr->is_in(p); 2575 } else { 2576 return false; 2577 } 2578 } 2579 2580 // Iteration functions. 2581 2582 // Iterates an OopClosure over all ref-containing fields of objects 2583 // within a HeapRegion. 2584 2585 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2586 MemRegion _mr; 2587 ExtendedOopClosure* _cl; 2588 public: 2589 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2590 : _mr(mr), _cl(cl) {} 2591 bool doHeapRegion(HeapRegion* r) { 2592 if (!r->continuesHumongous()) { 2593 r->oop_iterate(_cl); 2594 } 2595 return false; 2596 } 2597 }; 2598 2599 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2600 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2601 heap_region_iterate(&blk); 2602 } 2603 2604 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2605 IterateOopClosureRegionClosure blk(mr, cl); 2606 heap_region_iterate(&blk); 2607 } 2608 2609 // Iterates an ObjectClosure over all objects within a HeapRegion. 2610 2611 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2612 ObjectClosure* _cl; 2613 public: 2614 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2615 bool doHeapRegion(HeapRegion* r) { 2616 if (! r->continuesHumongous()) { 2617 r->object_iterate(_cl); 2618 } 2619 return false; 2620 } 2621 }; 2622 2623 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2624 IterateObjectClosureRegionClosure blk(cl); 2625 heap_region_iterate(&blk); 2626 } 2627 2628 // Calls a SpaceClosure on a HeapRegion. 2629 2630 class SpaceClosureRegionClosure: public HeapRegionClosure { 2631 SpaceClosure* _cl; 2632 public: 2633 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2634 bool doHeapRegion(HeapRegion* r) { 2635 _cl->do_space(r); 2636 return false; 2637 } 2638 }; 2639 2640 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2641 SpaceClosureRegionClosure blk(cl); 2642 heap_region_iterate(&blk); 2643 } 2644 2645 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2646 _hrs.iterate(cl); 2647 } 2648 2649 void 2650 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2651 uint worker_id, 2652 uint no_of_par_workers, 2653 jint claim_value) { 2654 const uint regions = n_regions(); 2655 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2656 no_of_par_workers : 2657 1); 2658 assert(UseDynamicNumberOfGCThreads || 2659 no_of_par_workers == workers()->total_workers(), 2660 "Non dynamic should use fixed number of workers"); 2661 // try to spread out the starting points of the workers 2662 const HeapRegion* start_hr = 2663 start_region_for_worker(worker_id, no_of_par_workers); 2664 const uint start_index = start_hr->hrs_index(); 2665 2666 // each worker will actually look at all regions 2667 for (uint count = 0; count < regions; ++count) { 2668 const uint index = (start_index + count) % regions; 2669 assert(0 <= index && index < regions, "sanity"); 2670 HeapRegion* r = region_at(index); 2671 // we'll ignore "continues humongous" regions (we'll process them 2672 // when we come across their corresponding "start humongous" 2673 // region) and regions already claimed 2674 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2675 continue; 2676 } 2677 // OK, try to claim it 2678 if (r->claimHeapRegion(claim_value)) { 2679 // success! 2680 assert(!r->continuesHumongous(), "sanity"); 2681 if (r->startsHumongous()) { 2682 // If the region is "starts humongous" we'll iterate over its 2683 // "continues humongous" first; in fact we'll do them 2684 // first. The order is important. In on case, calling the 2685 // closure on the "starts humongous" region might de-allocate 2686 // and clear all its "continues humongous" regions and, as a 2687 // result, we might end up processing them twice. So, we'll do 2688 // them first (notice: most closures will ignore them anyway) and 2689 // then we'll do the "starts humongous" region. 2690 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2691 HeapRegion* chr = region_at(ch_index); 2692 2693 // if the region has already been claimed or it's not 2694 // "continues humongous" we're done 2695 if (chr->claim_value() == claim_value || 2696 !chr->continuesHumongous()) { 2697 break; 2698 } 2699 2700 // No one should have claimed it directly. We can given 2701 // that we claimed its "starts humongous" region. 2702 assert(chr->claim_value() != claim_value, "sanity"); 2703 assert(chr->humongous_start_region() == r, "sanity"); 2704 2705 if (chr->claimHeapRegion(claim_value)) { 2706 // we should always be able to claim it; no one else should 2707 // be trying to claim this region 2708 2709 bool res2 = cl->doHeapRegion(chr); 2710 assert(!res2, "Should not abort"); 2711 2712 // Right now, this holds (i.e., no closure that actually 2713 // does something with "continues humongous" regions 2714 // clears them). We might have to weaken it in the future, 2715 // but let's leave these two asserts here for extra safety. 2716 assert(chr->continuesHumongous(), "should still be the case"); 2717 assert(chr->humongous_start_region() == r, "sanity"); 2718 } else { 2719 guarantee(false, "we should not reach here"); 2720 } 2721 } 2722 } 2723 2724 assert(!r->continuesHumongous(), "sanity"); 2725 bool res = cl->doHeapRegion(r); 2726 assert(!res, "Should not abort"); 2727 } 2728 } 2729 } 2730 2731 class ResetClaimValuesClosure: public HeapRegionClosure { 2732 public: 2733 bool doHeapRegion(HeapRegion* r) { 2734 r->set_claim_value(HeapRegion::InitialClaimValue); 2735 return false; 2736 } 2737 }; 2738 2739 void G1CollectedHeap::reset_heap_region_claim_values() { 2740 ResetClaimValuesClosure blk; 2741 heap_region_iterate(&blk); 2742 } 2743 2744 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2745 ResetClaimValuesClosure blk; 2746 collection_set_iterate(&blk); 2747 } 2748 2749 #ifdef ASSERT 2750 // This checks whether all regions in the heap have the correct claim 2751 // value. I also piggy-backed on this a check to ensure that the 2752 // humongous_start_region() information on "continues humongous" 2753 // regions is correct. 2754 2755 class CheckClaimValuesClosure : public HeapRegionClosure { 2756 private: 2757 jint _claim_value; 2758 uint _failures; 2759 HeapRegion* _sh_region; 2760 2761 public: 2762 CheckClaimValuesClosure(jint claim_value) : 2763 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2764 bool doHeapRegion(HeapRegion* r) { 2765 if (r->claim_value() != _claim_value) { 2766 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2767 "claim value = %d, should be %d", 2768 HR_FORMAT_PARAMS(r), 2769 r->claim_value(), _claim_value); 2770 ++_failures; 2771 } 2772 if (!r->isHumongous()) { 2773 _sh_region = NULL; 2774 } else if (r->startsHumongous()) { 2775 _sh_region = r; 2776 } else if (r->continuesHumongous()) { 2777 if (r->humongous_start_region() != _sh_region) { 2778 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2779 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2780 HR_FORMAT_PARAMS(r), 2781 r->humongous_start_region(), 2782 _sh_region); 2783 ++_failures; 2784 } 2785 } 2786 return false; 2787 } 2788 uint failures() { return _failures; } 2789 }; 2790 2791 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2792 CheckClaimValuesClosure cl(claim_value); 2793 heap_region_iterate(&cl); 2794 return cl.failures() == 0; 2795 } 2796 2797 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2798 private: 2799 jint _claim_value; 2800 uint _failures; 2801 2802 public: 2803 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2804 _claim_value(claim_value), _failures(0) { } 2805 2806 uint failures() { return _failures; } 2807 2808 bool doHeapRegion(HeapRegion* hr) { 2809 assert(hr->in_collection_set(), "how?"); 2810 assert(!hr->isHumongous(), "H-region in CSet"); 2811 if (hr->claim_value() != _claim_value) { 2812 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2813 "claim value = %d, should be %d", 2814 HR_FORMAT_PARAMS(hr), 2815 hr->claim_value(), _claim_value); 2816 _failures += 1; 2817 } 2818 return false; 2819 } 2820 }; 2821 2822 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2823 CheckClaimValuesInCSetHRClosure cl(claim_value); 2824 collection_set_iterate(&cl); 2825 return cl.failures() == 0; 2826 } 2827 #endif // ASSERT 2828 2829 // Clear the cached CSet starting regions and (more importantly) 2830 // the time stamps. Called when we reset the GC time stamp. 2831 void G1CollectedHeap::clear_cset_start_regions() { 2832 assert(_worker_cset_start_region != NULL, "sanity"); 2833 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2834 2835 int n_queues = MAX2((int)ParallelGCThreads, 1); 2836 for (int i = 0; i < n_queues; i++) { 2837 _worker_cset_start_region[i] = NULL; 2838 _worker_cset_start_region_time_stamp[i] = 0; 2839 } 2840 } 2841 2842 // Given the id of a worker, obtain or calculate a suitable 2843 // starting region for iterating over the current collection set. 2844 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(uint worker_i) { 2845 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2846 2847 HeapRegion* result = NULL; 2848 unsigned gc_time_stamp = get_gc_time_stamp(); 2849 2850 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2851 // Cached starting region for current worker was set 2852 // during the current pause - so it's valid. 2853 // Note: the cached starting heap region may be NULL 2854 // (when the collection set is empty). 2855 result = _worker_cset_start_region[worker_i]; 2856 assert(result == NULL || result->in_collection_set(), "sanity"); 2857 return result; 2858 } 2859 2860 // The cached entry was not valid so let's calculate 2861 // a suitable starting heap region for this worker. 2862 2863 // We want the parallel threads to start their collection 2864 // set iteration at different collection set regions to 2865 // avoid contention. 2866 // If we have: 2867 // n collection set regions 2868 // p threads 2869 // Then thread t will start at region floor ((t * n) / p) 2870 2871 result = g1_policy()->collection_set(); 2872 if (G1CollectedHeap::use_parallel_gc_threads()) { 2873 uint cs_size = g1_policy()->cset_region_length(); 2874 uint active_workers = workers()->active_workers(); 2875 assert(UseDynamicNumberOfGCThreads || 2876 active_workers == workers()->total_workers(), 2877 "Unless dynamic should use total workers"); 2878 2879 uint end_ind = (cs_size * worker_i) / active_workers; 2880 uint start_ind = 0; 2881 2882 if (worker_i > 0 && 2883 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2884 // Previous workers starting region is valid 2885 // so let's iterate from there 2886 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2887 result = _worker_cset_start_region[worker_i - 1]; 2888 } 2889 2890 for (uint i = start_ind; i < end_ind; i++) { 2891 result = result->next_in_collection_set(); 2892 } 2893 } 2894 2895 // Note: the calculated starting heap region may be NULL 2896 // (when the collection set is empty). 2897 assert(result == NULL || result->in_collection_set(), "sanity"); 2898 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2899 "should be updated only once per pause"); 2900 _worker_cset_start_region[worker_i] = result; 2901 OrderAccess::storestore(); 2902 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2903 return result; 2904 } 2905 2906 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2907 uint no_of_par_workers) { 2908 uint worker_num = 2909 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2910 assert(UseDynamicNumberOfGCThreads || 2911 no_of_par_workers == workers()->total_workers(), 2912 "Non dynamic should use fixed number of workers"); 2913 const uint start_index = n_regions() * worker_i / worker_num; 2914 return region_at(start_index); 2915 } 2916 2917 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2918 HeapRegion* r = g1_policy()->collection_set(); 2919 while (r != NULL) { 2920 HeapRegion* next = r->next_in_collection_set(); 2921 if (cl->doHeapRegion(r)) { 2922 cl->incomplete(); 2923 return; 2924 } 2925 r = next; 2926 } 2927 } 2928 2929 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2930 HeapRegionClosure *cl) { 2931 if (r == NULL) { 2932 // The CSet is empty so there's nothing to do. 2933 return; 2934 } 2935 2936 assert(r->in_collection_set(), 2937 "Start region must be a member of the collection set."); 2938 HeapRegion* cur = r; 2939 while (cur != NULL) { 2940 HeapRegion* next = cur->next_in_collection_set(); 2941 if (cl->doHeapRegion(cur) && false) { 2942 cl->incomplete(); 2943 return; 2944 } 2945 cur = next; 2946 } 2947 cur = g1_policy()->collection_set(); 2948 while (cur != r) { 2949 HeapRegion* next = cur->next_in_collection_set(); 2950 if (cl->doHeapRegion(cur) && false) { 2951 cl->incomplete(); 2952 return; 2953 } 2954 cur = next; 2955 } 2956 } 2957 2958 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2959 return n_regions() > 0 ? region_at(0) : NULL; 2960 } 2961 2962 2963 Space* G1CollectedHeap::space_containing(const void* addr) const { 2964 Space* res = heap_region_containing(addr); 2965 return res; 2966 } 2967 2968 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2969 Space* sp = space_containing(addr); 2970 if (sp != NULL) { 2971 return sp->block_start(addr); 2972 } 2973 return NULL; 2974 } 2975 2976 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2977 Space* sp = space_containing(addr); 2978 assert(sp != NULL, "block_size of address outside of heap"); 2979 return sp->block_size(addr); 2980 } 2981 2982 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2983 Space* sp = space_containing(addr); 2984 return sp->block_is_obj(addr); 2985 } 2986 2987 bool G1CollectedHeap::supports_tlab_allocation() const { 2988 return true; 2989 } 2990 2991 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2992 return (_g1_policy->young_list_target_length() - young_list()->survivor_length()) * HeapRegion::GrainBytes; 2993 } 2994 2995 size_t G1CollectedHeap::tlab_used(Thread* ignored) const { 2996 return young_list()->eden_used_bytes(); 2997 } 2998 2999 // For G1 TLABs should not contain humongous objects, so the maximum TLAB size 3000 // must be smaller than the humongous object limit. 3001 size_t G1CollectedHeap::max_tlab_size() const { 3002 return align_size_down(_humongous_object_threshold_in_words - 1, MinObjAlignment); 3003 } 3004 3005 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 3006 // Return the remaining space in the cur alloc region, but not less than 3007 // the min TLAB size. 3008 3009 // Also, this value can be at most the humongous object threshold, 3010 // since we can't allow tlabs to grow big enough to accommodate 3011 // humongous objects. 3012 3013 HeapRegion* hr = _mutator_alloc_region.get(); 3014 size_t max_tlab = max_tlab_size() * wordSize; 3015 if (hr == NULL) { 3016 return max_tlab; 3017 } else { 3018 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab); 3019 } 3020 } 3021 3022 size_t G1CollectedHeap::max_capacity() const { 3023 return _g1_reserved.byte_size(); 3024 } 3025 3026 jlong G1CollectedHeap::millis_since_last_gc() { 3027 // assert(false, "NYI"); 3028 return 0; 3029 } 3030 3031 void G1CollectedHeap::prepare_for_verify() { 3032 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3033 ensure_parsability(false); 3034 } 3035 g1_rem_set()->prepare_for_verify(); 3036 } 3037 3038 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3039 VerifyOption vo) { 3040 switch (vo) { 3041 case VerifyOption_G1UsePrevMarking: 3042 return hr->obj_allocated_since_prev_marking(obj); 3043 case VerifyOption_G1UseNextMarking: 3044 return hr->obj_allocated_since_next_marking(obj); 3045 case VerifyOption_G1UseMarkWord: 3046 return false; 3047 default: 3048 ShouldNotReachHere(); 3049 } 3050 return false; // keep some compilers happy 3051 } 3052 3053 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3054 switch (vo) { 3055 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3056 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3057 case VerifyOption_G1UseMarkWord: return NULL; 3058 default: ShouldNotReachHere(); 3059 } 3060 return NULL; // keep some compilers happy 3061 } 3062 3063 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3064 switch (vo) { 3065 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3066 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3067 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3068 default: ShouldNotReachHere(); 3069 } 3070 return false; // keep some compilers happy 3071 } 3072 3073 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3074 switch (vo) { 3075 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3076 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3077 case VerifyOption_G1UseMarkWord: return "NONE"; 3078 default: ShouldNotReachHere(); 3079 } 3080 return NULL; // keep some compilers happy 3081 } 3082 3083 class VerifyRootsClosure: public OopClosure { 3084 private: 3085 G1CollectedHeap* _g1h; 3086 VerifyOption _vo; 3087 bool _failures; 3088 public: 3089 // _vo == UsePrevMarking -> use "prev" marking information, 3090 // _vo == UseNextMarking -> use "next" marking information, 3091 // _vo == UseMarkWord -> use mark word from object header. 3092 VerifyRootsClosure(VerifyOption vo) : 3093 _g1h(G1CollectedHeap::heap()), 3094 _vo(vo), 3095 _failures(false) { } 3096 3097 bool failures() { return _failures; } 3098 3099 template <class T> void do_oop_nv(T* p) { 3100 T heap_oop = oopDesc::load_heap_oop(p); 3101 if (!oopDesc::is_null(heap_oop)) { 3102 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3103 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3104 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3105 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3106 if (_vo == VerifyOption_G1UseMarkWord) { 3107 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3108 } 3109 obj->print_on(gclog_or_tty); 3110 _failures = true; 3111 } 3112 } 3113 } 3114 3115 void do_oop(oop* p) { do_oop_nv(p); } 3116 void do_oop(narrowOop* p) { do_oop_nv(p); } 3117 }; 3118 3119 class G1VerifyCodeRootOopClosure: public OopClosure { 3120 G1CollectedHeap* _g1h; 3121 OopClosure* _root_cl; 3122 nmethod* _nm; 3123 VerifyOption _vo; 3124 bool _failures; 3125 3126 template <class T> void do_oop_work(T* p) { 3127 // First verify that this root is live 3128 _root_cl->do_oop(p); 3129 3130 if (!G1VerifyHeapRegionCodeRoots) { 3131 // We're not verifying the code roots attached to heap region. 3132 return; 3133 } 3134 3135 // Don't check the code roots during marking verification in a full GC 3136 if (_vo == VerifyOption_G1UseMarkWord) { 3137 return; 3138 } 3139 3140 // Now verify that the current nmethod (which contains p) is 3141 // in the code root list of the heap region containing the 3142 // object referenced by p. 3143 3144 T heap_oop = oopDesc::load_heap_oop(p); 3145 if (!oopDesc::is_null(heap_oop)) { 3146 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3147 3148 // Now fetch the region containing the object 3149 HeapRegion* hr = _g1h->heap_region_containing(obj); 3150 HeapRegionRemSet* hrrs = hr->rem_set(); 3151 // Verify that the strong code root list for this region 3152 // contains the nmethod 3153 if (!hrrs->strong_code_roots_list_contains(_nm)) { 3154 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" " 3155 "from nmethod "PTR_FORMAT" not in strong " 3156 "code roots for region ["PTR_FORMAT","PTR_FORMAT")", 3157 p, _nm, hr->bottom(), hr->end()); 3158 _failures = true; 3159 } 3160 } 3161 } 3162 3163 public: 3164 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo): 3165 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {} 3166 3167 void do_oop(oop* p) { do_oop_work(p); } 3168 void do_oop(narrowOop* p) { do_oop_work(p); } 3169 3170 void set_nmethod(nmethod* nm) { _nm = nm; } 3171 bool failures() { return _failures; } 3172 }; 3173 3174 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure { 3175 G1VerifyCodeRootOopClosure* _oop_cl; 3176 3177 public: 3178 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl): 3179 _oop_cl(oop_cl) {} 3180 3181 void do_code_blob(CodeBlob* cb) { 3182 nmethod* nm = cb->as_nmethod_or_null(); 3183 if (nm != NULL) { 3184 _oop_cl->set_nmethod(nm); 3185 nm->oops_do(_oop_cl); 3186 } 3187 } 3188 }; 3189 3190 class YoungRefCounterClosure : public OopClosure { 3191 G1CollectedHeap* _g1h; 3192 int _count; 3193 public: 3194 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3195 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3196 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3197 3198 int count() { return _count; } 3199 void reset_count() { _count = 0; }; 3200 }; 3201 3202 class VerifyKlassClosure: public KlassClosure { 3203 YoungRefCounterClosure _young_ref_counter_closure; 3204 OopClosure *_oop_closure; 3205 public: 3206 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3207 void do_klass(Klass* k) { 3208 k->oops_do(_oop_closure); 3209 3210 _young_ref_counter_closure.reset_count(); 3211 k->oops_do(&_young_ref_counter_closure); 3212 if (_young_ref_counter_closure.count() > 0) { 3213 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3214 } 3215 } 3216 }; 3217 3218 class VerifyLivenessOopClosure: public OopClosure { 3219 G1CollectedHeap* _g1h; 3220 VerifyOption _vo; 3221 public: 3222 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3223 _g1h(g1h), _vo(vo) 3224 { } 3225 void do_oop(narrowOop *p) { do_oop_work(p); } 3226 void do_oop( oop *p) { do_oop_work(p); } 3227 3228 template <class T> void do_oop_work(T *p) { 3229 oop obj = oopDesc::load_decode_heap_oop(p); 3230 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3231 "Dead object referenced by a not dead object"); 3232 } 3233 }; 3234 3235 class VerifyObjsInRegionClosure: public ObjectClosure { 3236 private: 3237 G1CollectedHeap* _g1h; 3238 size_t _live_bytes; 3239 HeapRegion *_hr; 3240 VerifyOption _vo; 3241 public: 3242 // _vo == UsePrevMarking -> use "prev" marking information, 3243 // _vo == UseNextMarking -> use "next" marking information, 3244 // _vo == UseMarkWord -> use mark word from object header. 3245 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3246 : _live_bytes(0), _hr(hr), _vo(vo) { 3247 _g1h = G1CollectedHeap::heap(); 3248 } 3249 void do_object(oop o) { 3250 VerifyLivenessOopClosure isLive(_g1h, _vo); 3251 assert(o != NULL, "Huh?"); 3252 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3253 // If the object is alive according to the mark word, 3254 // then verify that the marking information agrees. 3255 // Note we can't verify the contra-positive of the 3256 // above: if the object is dead (according to the mark 3257 // word), it may not be marked, or may have been marked 3258 // but has since became dead, or may have been allocated 3259 // since the last marking. 3260 if (_vo == VerifyOption_G1UseMarkWord) { 3261 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3262 } 3263 3264 o->oop_iterate_no_header(&isLive); 3265 if (!_hr->obj_allocated_since_prev_marking(o)) { 3266 size_t obj_size = o->size(); // Make sure we don't overflow 3267 _live_bytes += (obj_size * HeapWordSize); 3268 } 3269 } 3270 } 3271 size_t live_bytes() { return _live_bytes; } 3272 }; 3273 3274 class PrintObjsInRegionClosure : public ObjectClosure { 3275 HeapRegion *_hr; 3276 G1CollectedHeap *_g1; 3277 public: 3278 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3279 _g1 = G1CollectedHeap::heap(); 3280 }; 3281 3282 void do_object(oop o) { 3283 if (o != NULL) { 3284 HeapWord *start = (HeapWord *) o; 3285 size_t word_sz = o->size(); 3286 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3287 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3288 (void*) o, word_sz, 3289 _g1->isMarkedPrev(o), 3290 _g1->isMarkedNext(o), 3291 _hr->obj_allocated_since_prev_marking(o)); 3292 HeapWord *end = start + word_sz; 3293 HeapWord *cur; 3294 int *val; 3295 for (cur = start; cur < end; cur++) { 3296 val = (int *) cur; 3297 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3298 } 3299 } 3300 } 3301 }; 3302 3303 class VerifyRegionClosure: public HeapRegionClosure { 3304 private: 3305 bool _par; 3306 VerifyOption _vo; 3307 bool _failures; 3308 public: 3309 // _vo == UsePrevMarking -> use "prev" marking information, 3310 // _vo == UseNextMarking -> use "next" marking information, 3311 // _vo == UseMarkWord -> use mark word from object header. 3312 VerifyRegionClosure(bool par, VerifyOption vo) 3313 : _par(par), 3314 _vo(vo), 3315 _failures(false) {} 3316 3317 bool failures() { 3318 return _failures; 3319 } 3320 3321 bool doHeapRegion(HeapRegion* r) { 3322 if (!r->continuesHumongous()) { 3323 bool failures = false; 3324 r->verify(_vo, &failures); 3325 if (failures) { 3326 _failures = true; 3327 } else { 3328 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3329 r->object_iterate(¬_dead_yet_cl); 3330 if (_vo != VerifyOption_G1UseNextMarking) { 3331 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3332 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3333 "max_live_bytes "SIZE_FORMAT" " 3334 "< calculated "SIZE_FORMAT, 3335 r->bottom(), r->end(), 3336 r->max_live_bytes(), 3337 not_dead_yet_cl.live_bytes()); 3338 _failures = true; 3339 } 3340 } else { 3341 // When vo == UseNextMarking we cannot currently do a sanity 3342 // check on the live bytes as the calculation has not been 3343 // finalized yet. 3344 } 3345 } 3346 } 3347 return false; // stop the region iteration if we hit a failure 3348 } 3349 }; 3350 3351 // This is the task used for parallel verification of the heap regions 3352 3353 class G1ParVerifyTask: public AbstractGangTask { 3354 private: 3355 G1CollectedHeap* _g1h; 3356 VerifyOption _vo; 3357 bool _failures; 3358 3359 public: 3360 // _vo == UsePrevMarking -> use "prev" marking information, 3361 // _vo == UseNextMarking -> use "next" marking information, 3362 // _vo == UseMarkWord -> use mark word from object header. 3363 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3364 AbstractGangTask("Parallel verify task"), 3365 _g1h(g1h), 3366 _vo(vo), 3367 _failures(false) { } 3368 3369 bool failures() { 3370 return _failures; 3371 } 3372 3373 void work(uint worker_id) { 3374 HandleMark hm; 3375 VerifyRegionClosure blk(true, _vo); 3376 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3377 _g1h->workers()->active_workers(), 3378 HeapRegion::ParVerifyClaimValue); 3379 if (blk.failures()) { 3380 _failures = true; 3381 } 3382 } 3383 }; 3384 3385 void G1CollectedHeap::verify(bool silent, VerifyOption vo) { 3386 if (SafepointSynchronize::is_at_safepoint()) { 3387 assert(Thread::current()->is_VM_thread(), 3388 "Expected to be executed serially by the VM thread at this point"); 3389 3390 if (!silent) { gclog_or_tty->print("Roots "); } 3391 VerifyRootsClosure rootsCl(vo); 3392 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo); 3393 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl); 3394 VerifyKlassClosure klassCl(this, &rootsCl); 3395 3396 // We apply the relevant closures to all the oops in the 3397 // system dictionary, the string table and the code cache. 3398 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3399 3400 // Need cleared claim bits for the strong roots processing 3401 ClassLoaderDataGraph::clear_claimed_marks(); 3402 3403 process_strong_roots(true, // activate StrongRootsScope 3404 false, // we set "is scavenging" to false, 3405 // so we don't reset the dirty cards. 3406 ScanningOption(so), // roots scanning options 3407 &rootsCl, 3408 &blobsCl, 3409 &klassCl 3410 ); 3411 3412 bool failures = rootsCl.failures() || codeRootsCl.failures(); 3413 3414 if (vo != VerifyOption_G1UseMarkWord) { 3415 // If we're verifying during a full GC then the region sets 3416 // will have been torn down at the start of the GC. Therefore 3417 // verifying the region sets will fail. So we only verify 3418 // the region sets when not in a full GC. 3419 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3420 verify_region_sets(); 3421 } 3422 3423 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3424 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3425 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3426 "sanity check"); 3427 3428 G1ParVerifyTask task(this, vo); 3429 assert(UseDynamicNumberOfGCThreads || 3430 workers()->active_workers() == workers()->total_workers(), 3431 "If not dynamic should be using all the workers"); 3432 int n_workers = workers()->active_workers(); 3433 set_par_threads(n_workers); 3434 workers()->run_task(&task); 3435 set_par_threads(0); 3436 if (task.failures()) { 3437 failures = true; 3438 } 3439 3440 // Checks that the expected amount of parallel work was done. 3441 // The implication is that n_workers is > 0. 3442 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3443 "sanity check"); 3444 3445 reset_heap_region_claim_values(); 3446 3447 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3448 "sanity check"); 3449 } else { 3450 VerifyRegionClosure blk(false, vo); 3451 heap_region_iterate(&blk); 3452 if (blk.failures()) { 3453 failures = true; 3454 } 3455 } 3456 if (!silent) gclog_or_tty->print("RemSet "); 3457 rem_set()->verify(); 3458 3459 if (G1StringDedup::is_enabled()) { 3460 if (!silent) gclog_or_tty->print("StrDedup "); 3461 G1StringDedup::verify(); 3462 } 3463 3464 if (failures) { 3465 gclog_or_tty->print_cr("Heap:"); 3466 // It helps to have the per-region information in the output to 3467 // help us track down what went wrong. This is why we call 3468 // print_extended_on() instead of print_on(). 3469 print_extended_on(gclog_or_tty); 3470 gclog_or_tty->cr(); 3471 #ifndef PRODUCT 3472 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3473 concurrent_mark()->print_reachable("at-verification-failure", 3474 vo, false /* all */); 3475 } 3476 #endif 3477 gclog_or_tty->flush(); 3478 } 3479 guarantee(!failures, "there should not have been any failures"); 3480 } else { 3481 if (!silent) { 3482 gclog_or_tty->print("(SKIPPING Roots, HeapRegionSets, HeapRegions, RemSet"); 3483 if (G1StringDedup::is_enabled()) { 3484 gclog_or_tty->print(", StrDedup"); 3485 } 3486 gclog_or_tty->print(") "); 3487 } 3488 } 3489 } 3490 3491 void G1CollectedHeap::verify(bool silent) { 3492 verify(silent, VerifyOption_G1UsePrevMarking); 3493 } 3494 3495 double G1CollectedHeap::verify(bool guard, const char* msg) { 3496 double verify_time_ms = 0.0; 3497 3498 if (guard && total_collections() >= VerifyGCStartAt) { 3499 double verify_start = os::elapsedTime(); 3500 HandleMark hm; // Discard invalid handles created during verification 3501 prepare_for_verify(); 3502 Universe::verify(VerifyOption_G1UsePrevMarking, msg); 3503 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 3504 } 3505 3506 return verify_time_ms; 3507 } 3508 3509 void G1CollectedHeap::verify_before_gc() { 3510 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 3511 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 3512 } 3513 3514 void G1CollectedHeap::verify_after_gc() { 3515 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 3516 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 3517 } 3518 3519 class PrintRegionClosure: public HeapRegionClosure { 3520 outputStream* _st; 3521 public: 3522 PrintRegionClosure(outputStream* st) : _st(st) {} 3523 bool doHeapRegion(HeapRegion* r) { 3524 r->print_on(_st); 3525 return false; 3526 } 3527 }; 3528 3529 bool G1CollectedHeap::is_obj_dead_cond(const oop obj, 3530 const HeapRegion* hr, 3531 const VerifyOption vo) const { 3532 switch (vo) { 3533 case VerifyOption_G1UsePrevMarking: return is_obj_dead(obj, hr); 3534 case VerifyOption_G1UseNextMarking: return is_obj_ill(obj, hr); 3535 case VerifyOption_G1UseMarkWord: return !obj->is_gc_marked(); 3536 default: ShouldNotReachHere(); 3537 } 3538 return false; // keep some compilers happy 3539 } 3540 3541 bool G1CollectedHeap::is_obj_dead_cond(const oop obj, 3542 const VerifyOption vo) const { 3543 switch (vo) { 3544 case VerifyOption_G1UsePrevMarking: return is_obj_dead(obj); 3545 case VerifyOption_G1UseNextMarking: return is_obj_ill(obj); 3546 case VerifyOption_G1UseMarkWord: return !obj->is_gc_marked(); 3547 default: ShouldNotReachHere(); 3548 } 3549 return false; // keep some compilers happy 3550 } 3551 3552 void G1CollectedHeap::print_on(outputStream* st) const { 3553 st->print(" %-20s", "garbage-first heap"); 3554 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3555 capacity()/K, used_unlocked()/K); 3556 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3557 _g1_storage.low_boundary(), 3558 _g1_storage.high(), 3559 _g1_storage.high_boundary()); 3560 st->cr(); 3561 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3562 uint young_regions = _young_list->length(); 3563 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3564 (size_t) young_regions * HeapRegion::GrainBytes / K); 3565 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3566 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3567 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3568 st->cr(); 3569 MetaspaceAux::print_on(st); 3570 } 3571 3572 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3573 print_on(st); 3574 3575 // Print the per-region information. 3576 st->cr(); 3577 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3578 "HS=humongous(starts), HC=humongous(continues), " 3579 "CS=collection set, F=free, TS=gc time stamp, " 3580 "PTAMS=previous top-at-mark-start, " 3581 "NTAMS=next top-at-mark-start)"); 3582 PrintRegionClosure blk(st); 3583 heap_region_iterate(&blk); 3584 } 3585 3586 void G1CollectedHeap::print_on_error(outputStream* st) const { 3587 this->CollectedHeap::print_on_error(st); 3588 3589 if (_cm != NULL) { 3590 st->cr(); 3591 _cm->print_on_error(st); 3592 } 3593 } 3594 3595 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3596 if (G1CollectedHeap::use_parallel_gc_threads()) { 3597 workers()->print_worker_threads_on(st); 3598 } 3599 _cmThread->print_on(st); 3600 st->cr(); 3601 _cm->print_worker_threads_on(st); 3602 _cg1r->print_worker_threads_on(st); 3603 if (G1StringDedup::is_enabled()) { 3604 G1StringDedup::print_worker_threads_on(st); 3605 } 3606 } 3607 3608 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3609 if (G1CollectedHeap::use_parallel_gc_threads()) { 3610 workers()->threads_do(tc); 3611 } 3612 tc->do_thread(_cmThread); 3613 _cg1r->threads_do(tc); 3614 if (G1StringDedup::is_enabled()) { 3615 G1StringDedup::threads_do(tc); 3616 } 3617 } 3618 3619 void G1CollectedHeap::print_tracing_info() const { 3620 // We'll overload this to mean "trace GC pause statistics." 3621 if (TraceGen0Time || TraceGen1Time) { 3622 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3623 // to that. 3624 g1_policy()->print_tracing_info(); 3625 } 3626 if (G1SummarizeRSetStats) { 3627 g1_rem_set()->print_summary_info(); 3628 } 3629 if (G1SummarizeConcMark) { 3630 concurrent_mark()->print_summary_info(); 3631 } 3632 g1_policy()->print_yg_surv_rate_info(); 3633 SpecializationStats::print(); 3634 } 3635 3636 #ifndef PRODUCT 3637 // Helpful for debugging RSet issues. 3638 3639 class PrintRSetsClosure : public HeapRegionClosure { 3640 private: 3641 const char* _msg; 3642 size_t _occupied_sum; 3643 3644 public: 3645 bool doHeapRegion(HeapRegion* r) { 3646 HeapRegionRemSet* hrrs = r->rem_set(); 3647 size_t occupied = hrrs->occupied(); 3648 _occupied_sum += occupied; 3649 3650 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3651 HR_FORMAT_PARAMS(r)); 3652 if (occupied == 0) { 3653 gclog_or_tty->print_cr(" RSet is empty"); 3654 } else { 3655 hrrs->print(); 3656 } 3657 gclog_or_tty->print_cr("----------"); 3658 return false; 3659 } 3660 3661 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3662 gclog_or_tty->cr(); 3663 gclog_or_tty->print_cr("========================================"); 3664 gclog_or_tty->print_cr("%s", msg); 3665 gclog_or_tty->cr(); 3666 } 3667 3668 ~PrintRSetsClosure() { 3669 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3670 gclog_or_tty->print_cr("========================================"); 3671 gclog_or_tty->cr(); 3672 } 3673 }; 3674 3675 void G1CollectedHeap::print_cset_rsets() { 3676 PrintRSetsClosure cl("Printing CSet RSets"); 3677 collection_set_iterate(&cl); 3678 } 3679 3680 void G1CollectedHeap::print_all_rsets() { 3681 PrintRSetsClosure cl("Printing All RSets");; 3682 heap_region_iterate(&cl); 3683 } 3684 #endif // PRODUCT 3685 3686 G1CollectedHeap* G1CollectedHeap::heap() { 3687 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3688 "not a garbage-first heap"); 3689 return _g1h; 3690 } 3691 3692 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3693 // always_do_update_barrier = false; 3694 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3695 // Fill TLAB's and such 3696 accumulate_statistics_all_tlabs(); 3697 ensure_parsability(true); 3698 3699 if (G1SummarizeRSetStats && (G1SummarizeRSetStatsPeriod > 0) && 3700 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3701 g1_rem_set()->print_periodic_summary_info("Before GC RS summary"); 3702 } 3703 } 3704 3705 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3706 3707 if (G1SummarizeRSetStats && 3708 (G1SummarizeRSetStatsPeriod > 0) && 3709 // we are at the end of the GC. Total collections has already been increased. 3710 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) { 3711 g1_rem_set()->print_periodic_summary_info("After GC RS summary"); 3712 } 3713 3714 // FIXME: what is this about? 3715 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3716 // is set. 3717 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3718 "derived pointer present")); 3719 // always_do_update_barrier = true; 3720 3721 resize_all_tlabs(); 3722 3723 // We have just completed a GC. Update the soft reference 3724 // policy with the new heap occupancy 3725 Universe::update_heap_info_at_gc(); 3726 } 3727 3728 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3729 unsigned int gc_count_before, 3730 bool* succeeded, 3731 GCCause::Cause gc_cause) { 3732 assert_heap_not_locked_and_not_at_safepoint(); 3733 g1_policy()->record_stop_world_start(); 3734 VM_G1IncCollectionPause op(gc_count_before, 3735 word_size, 3736 false, /* should_initiate_conc_mark */ 3737 g1_policy()->max_pause_time_ms(), 3738 gc_cause); 3739 VMThread::execute(&op); 3740 3741 HeapWord* result = op.result(); 3742 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3743 assert(result == NULL || ret_succeeded, 3744 "the result should be NULL if the VM did not succeed"); 3745 *succeeded = ret_succeeded; 3746 3747 assert_heap_not_locked(); 3748 return result; 3749 } 3750 3751 void 3752 G1CollectedHeap::doConcurrentMark() { 3753 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3754 if (!_cmThread->in_progress()) { 3755 _cmThread->set_started(); 3756 CGC_lock->notify(); 3757 } 3758 } 3759 3760 size_t G1CollectedHeap::pending_card_num() { 3761 size_t extra_cards = 0; 3762 JavaThread *curr = Threads::first(); 3763 while (curr != NULL) { 3764 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3765 extra_cards += dcq.size(); 3766 curr = curr->next(); 3767 } 3768 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3769 size_t buffer_size = dcqs.buffer_size(); 3770 size_t buffer_num = dcqs.completed_buffers_num(); 3771 3772 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3773 // in bytes - not the number of 'entries'. We need to convert 3774 // into a number of cards. 3775 return (buffer_size * buffer_num + extra_cards) / oopSize; 3776 } 3777 3778 size_t G1CollectedHeap::cards_scanned() { 3779 return g1_rem_set()->cardsScanned(); 3780 } 3781 3782 void 3783 G1CollectedHeap::setup_surviving_young_words() { 3784 assert(_surviving_young_words == NULL, "pre-condition"); 3785 uint array_length = g1_policy()->young_cset_region_length(); 3786 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3787 if (_surviving_young_words == NULL) { 3788 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR, 3789 "Not enough space for young surv words summary."); 3790 } 3791 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3792 #ifdef ASSERT 3793 for (uint i = 0; i < array_length; ++i) { 3794 assert( _surviving_young_words[i] == 0, "memset above" ); 3795 } 3796 #endif // !ASSERT 3797 } 3798 3799 void 3800 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3801 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3802 uint array_length = g1_policy()->young_cset_region_length(); 3803 for (uint i = 0; i < array_length; ++i) { 3804 _surviving_young_words[i] += surv_young_words[i]; 3805 } 3806 } 3807 3808 void 3809 G1CollectedHeap::cleanup_surviving_young_words() { 3810 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3811 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3812 _surviving_young_words = NULL; 3813 } 3814 3815 #ifdef ASSERT 3816 class VerifyCSetClosure: public HeapRegionClosure { 3817 public: 3818 bool doHeapRegion(HeapRegion* hr) { 3819 // Here we check that the CSet region's RSet is ready for parallel 3820 // iteration. The fields that we'll verify are only manipulated 3821 // when the region is part of a CSet and is collected. Afterwards, 3822 // we reset these fields when we clear the region's RSet (when the 3823 // region is freed) so they are ready when the region is 3824 // re-allocated. The only exception to this is if there's an 3825 // evacuation failure and instead of freeing the region we leave 3826 // it in the heap. In that case, we reset these fields during 3827 // evacuation failure handling. 3828 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3829 3830 // Here's a good place to add any other checks we'd like to 3831 // perform on CSet regions. 3832 return false; 3833 } 3834 }; 3835 #endif // ASSERT 3836 3837 #if TASKQUEUE_STATS 3838 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3839 st->print_raw_cr("GC Task Stats"); 3840 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3841 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3842 } 3843 3844 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3845 print_taskqueue_stats_hdr(st); 3846 3847 TaskQueueStats totals; 3848 const int n = workers() != NULL ? workers()->total_workers() : 1; 3849 for (int i = 0; i < n; ++i) { 3850 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3851 totals += task_queue(i)->stats; 3852 } 3853 st->print_raw("tot "); totals.print(st); st->cr(); 3854 3855 DEBUG_ONLY(totals.verify()); 3856 } 3857 3858 void G1CollectedHeap::reset_taskqueue_stats() { 3859 const int n = workers() != NULL ? workers()->total_workers() : 1; 3860 for (int i = 0; i < n; ++i) { 3861 task_queue(i)->stats.reset(); 3862 } 3863 } 3864 #endif // TASKQUEUE_STATS 3865 3866 void G1CollectedHeap::log_gc_header() { 3867 if (!G1Log::fine()) { 3868 return; 3869 } 3870 3871 gclog_or_tty->gclog_stamp(_gc_tracer_stw->gc_id()); 3872 3873 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3874 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3875 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3876 3877 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3878 } 3879 3880 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3881 if (!G1Log::fine()) { 3882 return; 3883 } 3884 3885 if (G1Log::finer()) { 3886 if (evacuation_failed()) { 3887 gclog_or_tty->print(" (to-space exhausted)"); 3888 } 3889 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3890 g1_policy()->phase_times()->note_gc_end(); 3891 g1_policy()->phase_times()->print(pause_time_sec); 3892 g1_policy()->print_detailed_heap_transition(); 3893 } else { 3894 if (evacuation_failed()) { 3895 gclog_or_tty->print("--"); 3896 } 3897 g1_policy()->print_heap_transition(); 3898 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3899 } 3900 gclog_or_tty->flush(); 3901 } 3902 3903 bool 3904 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3905 assert_at_safepoint(true /* should_be_vm_thread */); 3906 guarantee(!is_gc_active(), "collection is not reentrant"); 3907 3908 if (GC_locker::check_active_before_gc()) { 3909 return false; 3910 } 3911 3912 _gc_timer_stw->register_gc_start(); 3913 3914 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start()); 3915 3916 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3917 ResourceMark rm; 3918 3919 print_heap_before_gc(); 3920 trace_heap_before_gc(_gc_tracer_stw); 3921 3922 verify_region_sets_optional(); 3923 verify_dirty_young_regions(); 3924 3925 // This call will decide whether this pause is an initial-mark 3926 // pause. If it is, during_initial_mark_pause() will return true 3927 // for the duration of this pause. 3928 g1_policy()->decide_on_conc_mark_initiation(); 3929 3930 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3931 assert(!g1_policy()->during_initial_mark_pause() || 3932 g1_policy()->gcs_are_young(), "sanity"); 3933 3934 // We also do not allow mixed GCs during marking. 3935 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3936 3937 // Record whether this pause is an initial mark. When the current 3938 // thread has completed its logging output and it's safe to signal 3939 // the CM thread, the flag's value in the policy has been reset. 3940 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3941 3942 // Inner scope for scope based logging, timers, and stats collection 3943 { 3944 EvacuationInfo evacuation_info; 3945 3946 if (g1_policy()->during_initial_mark_pause()) { 3947 // We are about to start a marking cycle, so we increment the 3948 // full collection counter. 3949 increment_old_marking_cycles_started(); 3950 register_concurrent_cycle_start(_gc_timer_stw->gc_start()); 3951 } 3952 3953 _gc_tracer_stw->report_yc_type(yc_type()); 3954 3955 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3956 3957 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3958 workers()->active_workers() : 1); 3959 double pause_start_sec = os::elapsedTime(); 3960 g1_policy()->phase_times()->note_gc_start(active_workers); 3961 log_gc_header(); 3962 3963 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3964 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3965 3966 // If the secondary_free_list is not empty, append it to the 3967 // free_list. No need to wait for the cleanup operation to finish; 3968 // the region allocation code will check the secondary_free_list 3969 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3970 // set, skip this step so that the region allocation code has to 3971 // get entries from the secondary_free_list. 3972 if (!G1StressConcRegionFreeing) { 3973 append_secondary_free_list_if_not_empty_with_lock(); 3974 } 3975 3976 assert(check_young_list_well_formed(), "young list should be well formed"); 3977 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3978 "sanity check"); 3979 3980 // Don't dynamically change the number of GC threads this early. A value of 3981 // 0 is used to indicate serial work. When parallel work is done, 3982 // it will be set. 3983 3984 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3985 IsGCActiveMark x; 3986 3987 gc_prologue(false); 3988 increment_total_collections(false /* full gc */); 3989 increment_gc_time_stamp(); 3990 3991 verify_before_gc(); 3992 3993 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3994 3995 // Please see comment in g1CollectedHeap.hpp and 3996 // G1CollectedHeap::ref_processing_init() to see how 3997 // reference processing currently works in G1. 3998 3999 // Enable discovery in the STW reference processor 4000 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 4001 true /*verify_no_refs*/); 4002 4003 { 4004 // We want to temporarily turn off discovery by the 4005 // CM ref processor, if necessary, and turn it back on 4006 // on again later if we do. Using a scoped 4007 // NoRefDiscovery object will do this. 4008 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 4009 4010 // Forget the current alloc region (we might even choose it to be part 4011 // of the collection set!). 4012 release_mutator_alloc_region(); 4013 4014 // We should call this after we retire the mutator alloc 4015 // region(s) so that all the ALLOC / RETIRE events are generated 4016 // before the start GC event. 4017 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 4018 4019 // This timing is only used by the ergonomics to handle our pause target. 4020 // It is unclear why this should not include the full pause. We will 4021 // investigate this in CR 7178365. 4022 // 4023 // Preserving the old comment here if that helps the investigation: 4024 // 4025 // The elapsed time induced by the start time below deliberately elides 4026 // the possible verification above. 4027 double sample_start_time_sec = os::elapsedTime(); 4028 4029 #if YOUNG_LIST_VERBOSE 4030 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 4031 _young_list->print(); 4032 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4033 #endif // YOUNG_LIST_VERBOSE 4034 4035 g1_policy()->record_collection_pause_start(sample_start_time_sec); 4036 4037 double scan_wait_start = os::elapsedTime(); 4038 // We have to wait until the CM threads finish scanning the 4039 // root regions as it's the only way to ensure that all the 4040 // objects on them have been correctly scanned before we start 4041 // moving them during the GC. 4042 bool waited = _cm->root_regions()->wait_until_scan_finished(); 4043 double wait_time_ms = 0.0; 4044 if (waited) { 4045 double scan_wait_end = os::elapsedTime(); 4046 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 4047 } 4048 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 4049 4050 #if YOUNG_LIST_VERBOSE 4051 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 4052 _young_list->print(); 4053 #endif // YOUNG_LIST_VERBOSE 4054 4055 if (g1_policy()->during_initial_mark_pause()) { 4056 concurrent_mark()->checkpointRootsInitialPre(); 4057 } 4058 4059 #if YOUNG_LIST_VERBOSE 4060 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 4061 _young_list->print(); 4062 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4063 #endif // YOUNG_LIST_VERBOSE 4064 4065 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); 4066 4067 _cm->note_start_of_gc(); 4068 // We should not verify the per-thread SATB buffers given that 4069 // we have not filtered them yet (we'll do so during the 4070 // GC). We also call this after finalize_cset() to 4071 // ensure that the CSet has been finalized. 4072 _cm->verify_no_cset_oops(true /* verify_stacks */, 4073 true /* verify_enqueued_buffers */, 4074 false /* verify_thread_buffers */, 4075 true /* verify_fingers */); 4076 4077 if (_hr_printer.is_active()) { 4078 HeapRegion* hr = g1_policy()->collection_set(); 4079 while (hr != NULL) { 4080 G1HRPrinter::RegionType type; 4081 if (!hr->is_young()) { 4082 type = G1HRPrinter::Old; 4083 } else if (hr->is_survivor()) { 4084 type = G1HRPrinter::Survivor; 4085 } else { 4086 type = G1HRPrinter::Eden; 4087 } 4088 _hr_printer.cset(hr); 4089 hr = hr->next_in_collection_set(); 4090 } 4091 } 4092 4093 #ifdef ASSERT 4094 VerifyCSetClosure cl; 4095 collection_set_iterate(&cl); 4096 #endif // ASSERT 4097 4098 setup_surviving_young_words(); 4099 4100 // Initialize the GC alloc regions. 4101 init_gc_alloc_regions(evacuation_info); 4102 4103 // Actually do the work... 4104 evacuate_collection_set(evacuation_info); 4105 4106 // We do this to mainly verify the per-thread SATB buffers 4107 // (which have been filtered by now) since we didn't verify 4108 // them earlier. No point in re-checking the stacks / enqueued 4109 // buffers given that the CSet has not changed since last time 4110 // we checked. 4111 _cm->verify_no_cset_oops(false /* verify_stacks */, 4112 false /* verify_enqueued_buffers */, 4113 true /* verify_thread_buffers */, 4114 true /* verify_fingers */); 4115 4116 free_collection_set(g1_policy()->collection_set(), evacuation_info); 4117 g1_policy()->clear_collection_set(); 4118 4119 cleanup_surviving_young_words(); 4120 4121 // Start a new incremental collection set for the next pause. 4122 g1_policy()->start_incremental_cset_building(); 4123 4124 clear_cset_fast_test(); 4125 4126 _young_list->reset_sampled_info(); 4127 4128 // Don't check the whole heap at this point as the 4129 // GC alloc regions from this pause have been tagged 4130 // as survivors and moved on to the survivor list. 4131 // Survivor regions will fail the !is_young() check. 4132 assert(check_young_list_empty(false /* check_heap */), 4133 "young list should be empty"); 4134 4135 #if YOUNG_LIST_VERBOSE 4136 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 4137 _young_list->print(); 4138 #endif // YOUNG_LIST_VERBOSE 4139 4140 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 4141 _young_list->first_survivor_region(), 4142 _young_list->last_survivor_region()); 4143 4144 _young_list->reset_auxilary_lists(); 4145 4146 if (evacuation_failed()) { 4147 _summary_bytes_used = recalculate_used(); 4148 uint n_queues = MAX2((int)ParallelGCThreads, 1); 4149 for (uint i = 0; i < n_queues; i++) { 4150 if (_evacuation_failed_info_array[i].has_failed()) { 4151 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]); 4152 } 4153 } 4154 } else { 4155 // The "used" of the the collection set have already been subtracted 4156 // when they were freed. Add in the bytes evacuated. 4157 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 4158 } 4159 4160 if (g1_policy()->during_initial_mark_pause()) { 4161 // We have to do this before we notify the CM threads that 4162 // they can start working to make sure that all the 4163 // appropriate initialization is done on the CM object. 4164 concurrent_mark()->checkpointRootsInitialPost(); 4165 set_marking_started(); 4166 // Note that we don't actually trigger the CM thread at 4167 // this point. We do that later when we're sure that 4168 // the current thread has completed its logging output. 4169 } 4170 4171 allocate_dummy_regions(); 4172 4173 #if YOUNG_LIST_VERBOSE 4174 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 4175 _young_list->print(); 4176 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4177 #endif // YOUNG_LIST_VERBOSE 4178 4179 init_mutator_alloc_region(); 4180 4181 { 4182 size_t expand_bytes = g1_policy()->expansion_amount(); 4183 if (expand_bytes > 0) { 4184 size_t bytes_before = capacity(); 4185 // No need for an ergo verbose message here, 4186 // expansion_amount() does this when it returns a value > 0. 4187 if (!expand(expand_bytes)) { 4188 // We failed to expand the heap so let's verify that 4189 // committed/uncommitted amount match the backing store 4190 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 4191 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 4192 } 4193 } 4194 } 4195 4196 // We redo the verification but now wrt to the new CSet which 4197 // has just got initialized after the previous CSet was freed. 4198 _cm->verify_no_cset_oops(true /* verify_stacks */, 4199 true /* verify_enqueued_buffers */, 4200 true /* verify_thread_buffers */, 4201 true /* verify_fingers */); 4202 _cm->note_end_of_gc(); 4203 4204 // This timing is only used by the ergonomics to handle our pause target. 4205 // It is unclear why this should not include the full pause. We will 4206 // investigate this in CR 7178365. 4207 double sample_end_time_sec = os::elapsedTime(); 4208 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 4209 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info); 4210 4211 MemoryService::track_memory_usage(); 4212 4213 // In prepare_for_verify() below we'll need to scan the deferred 4214 // update buffers to bring the RSets up-to-date if 4215 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4216 // the update buffers we'll probably need to scan cards on the 4217 // regions we just allocated to (i.e., the GC alloc 4218 // regions). However, during the last GC we called 4219 // set_saved_mark() on all the GC alloc regions, so card 4220 // scanning might skip the [saved_mark_word()...top()] area of 4221 // those regions (i.e., the area we allocated objects into 4222 // during the last GC). But it shouldn't. Given that 4223 // saved_mark_word() is conditional on whether the GC time stamp 4224 // on the region is current or not, by incrementing the GC time 4225 // stamp here we invalidate all the GC time stamps on all the 4226 // regions and saved_mark_word() will simply return top() for 4227 // all the regions. This is a nicer way of ensuring this rather 4228 // than iterating over the regions and fixing them. In fact, the 4229 // GC time stamp increment here also ensures that 4230 // saved_mark_word() will return top() between pauses, i.e., 4231 // during concurrent refinement. So we don't need the 4232 // is_gc_active() check to decided which top to use when 4233 // scanning cards (see CR 7039627). 4234 increment_gc_time_stamp(); 4235 4236 verify_after_gc(); 4237 4238 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4239 ref_processor_stw()->verify_no_references_recorded(); 4240 4241 // CM reference discovery will be re-enabled if necessary. 4242 } 4243 4244 // We should do this after we potentially expand the heap so 4245 // that all the COMMIT events are generated before the end GC 4246 // event, and after we retire the GC alloc regions so that all 4247 // RETIRE events are generated before the end GC event. 4248 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4249 4250 if (mark_in_progress()) { 4251 concurrent_mark()->update_g1_committed(); 4252 } 4253 4254 #ifdef TRACESPINNING 4255 ParallelTaskTerminator::print_termination_counts(); 4256 #endif 4257 4258 gc_epilogue(false); 4259 } 4260 4261 // Print the remainder of the GC log output. 4262 log_gc_footer(os::elapsedTime() - pause_start_sec); 4263 4264 // It is not yet to safe to tell the concurrent mark to 4265 // start as we have some optional output below. We don't want the 4266 // output from the concurrent mark thread interfering with this 4267 // logging output either. 4268 4269 _hrs.verify_optional(); 4270 verify_region_sets_optional(); 4271 4272 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4273 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4274 4275 print_heap_after_gc(); 4276 trace_heap_after_gc(_gc_tracer_stw); 4277 4278 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4279 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4280 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4281 // before any GC notifications are raised. 4282 g1mm()->update_sizes(); 4283 4284 _gc_tracer_stw->report_evacuation_info(&evacuation_info); 4285 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold()); 4286 _gc_timer_stw->register_gc_end(); 4287 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions()); 4288 } 4289 // It should now be safe to tell the concurrent mark thread to start 4290 // without its logging output interfering with the logging output 4291 // that came from the pause. 4292 4293 if (should_start_conc_mark) { 4294 // CAUTION: after the doConcurrentMark() call below, 4295 // the concurrent marking thread(s) could be running 4296 // concurrently with us. Make sure that anything after 4297 // this point does not assume that we are the only GC thread 4298 // running. Note: of course, the actual marking work will 4299 // not start until the safepoint itself is released in 4300 // SuspendibleThreadSet::desynchronize(). 4301 doConcurrentMark(); 4302 } 4303 4304 return true; 4305 } 4306 4307 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4308 { 4309 size_t gclab_word_size; 4310 switch (purpose) { 4311 case GCAllocForSurvived: 4312 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4313 break; 4314 case GCAllocForTenured: 4315 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4316 break; 4317 default: 4318 assert(false, "unknown GCAllocPurpose"); 4319 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4320 break; 4321 } 4322 4323 // Prevent humongous PLAB sizes for two reasons: 4324 // * PLABs are allocated using a similar paths as oops, but should 4325 // never be in a humongous region 4326 // * Allowing humongous PLABs needlessly churns the region free lists 4327 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4328 } 4329 4330 void G1CollectedHeap::init_mutator_alloc_region() { 4331 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4332 _mutator_alloc_region.init(); 4333 } 4334 4335 void G1CollectedHeap::release_mutator_alloc_region() { 4336 _mutator_alloc_region.release(); 4337 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4338 } 4339 4340 void G1CollectedHeap::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { 4341 assert_at_safepoint(true /* should_be_vm_thread */); 4342 4343 _survivor_gc_alloc_region.init(); 4344 _old_gc_alloc_region.init(); 4345 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4346 _retained_old_gc_alloc_region = NULL; 4347 4348 // We will discard the current GC alloc region if: 4349 // a) it's in the collection set (it can happen!), 4350 // b) it's already full (no point in using it), 4351 // c) it's empty (this means that it was emptied during 4352 // a cleanup and it should be on the free list now), or 4353 // d) it's humongous (this means that it was emptied 4354 // during a cleanup and was added to the free list, but 4355 // has been subsequently used to allocate a humongous 4356 // object that may be less than the region size). 4357 if (retained_region != NULL && 4358 !retained_region->in_collection_set() && 4359 !(retained_region->top() == retained_region->end()) && 4360 !retained_region->is_empty() && 4361 !retained_region->isHumongous()) { 4362 retained_region->set_saved_mark(); 4363 // The retained region was added to the old region set when it was 4364 // retired. We have to remove it now, since we don't allow regions 4365 // we allocate to in the region sets. We'll re-add it later, when 4366 // it's retired again. 4367 _old_set.remove(retained_region); 4368 bool during_im = g1_policy()->during_initial_mark_pause(); 4369 retained_region->note_start_of_copying(during_im); 4370 _old_gc_alloc_region.set(retained_region); 4371 _hr_printer.reuse(retained_region); 4372 evacuation_info.set_alloc_regions_used_before(retained_region->used()); 4373 } 4374 } 4375 4376 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { 4377 evacuation_info.set_allocation_regions(_survivor_gc_alloc_region.count() + 4378 _old_gc_alloc_region.count()); 4379 _survivor_gc_alloc_region.release(); 4380 // If we have an old GC alloc region to release, we'll save it in 4381 // _retained_old_gc_alloc_region. If we don't 4382 // _retained_old_gc_alloc_region will become NULL. This is what we 4383 // want either way so no reason to check explicitly for either 4384 // condition. 4385 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4386 4387 if (ResizePLAB) { 4388 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4389 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4390 } 4391 } 4392 4393 void G1CollectedHeap::abandon_gc_alloc_regions() { 4394 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4395 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4396 _retained_old_gc_alloc_region = NULL; 4397 } 4398 4399 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4400 _drain_in_progress = false; 4401 set_evac_failure_closure(cl); 4402 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4403 } 4404 4405 void G1CollectedHeap::finalize_for_evac_failure() { 4406 assert(_evac_failure_scan_stack != NULL && 4407 _evac_failure_scan_stack->length() == 0, 4408 "Postcondition"); 4409 assert(!_drain_in_progress, "Postcondition"); 4410 delete _evac_failure_scan_stack; 4411 _evac_failure_scan_stack = NULL; 4412 } 4413 4414 void G1CollectedHeap::remove_self_forwarding_pointers() { 4415 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4416 4417 double remove_self_forwards_start = os::elapsedTime(); 4418 4419 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4420 4421 if (G1CollectedHeap::use_parallel_gc_threads()) { 4422 set_par_threads(); 4423 workers()->run_task(&rsfp_task); 4424 set_par_threads(0); 4425 } else { 4426 rsfp_task.work(0); 4427 } 4428 4429 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4430 4431 // Reset the claim values in the regions in the collection set. 4432 reset_cset_heap_region_claim_values(); 4433 4434 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4435 4436 // Now restore saved marks, if any. 4437 assert(_objs_with_preserved_marks.size() == 4438 _preserved_marks_of_objs.size(), "Both or none."); 4439 while (!_objs_with_preserved_marks.is_empty()) { 4440 oop obj = _objs_with_preserved_marks.pop(); 4441 markOop m = _preserved_marks_of_objs.pop(); 4442 obj->set_mark(m); 4443 } 4444 _objs_with_preserved_marks.clear(true); 4445 _preserved_marks_of_objs.clear(true); 4446 4447 g1_policy()->phase_times()->record_evac_fail_remove_self_forwards((os::elapsedTime() - remove_self_forwards_start) * 1000.0); 4448 } 4449 4450 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4451 _evac_failure_scan_stack->push(obj); 4452 } 4453 4454 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4455 assert(_evac_failure_scan_stack != NULL, "precondition"); 4456 4457 while (_evac_failure_scan_stack->length() > 0) { 4458 oop obj = _evac_failure_scan_stack->pop(); 4459 _evac_failure_closure->set_region(heap_region_containing(obj)); 4460 obj->oop_iterate_backwards(_evac_failure_closure); 4461 } 4462 } 4463 4464 oop 4465 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state, 4466 oop old) { 4467 assert(obj_in_cs(old), 4468 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4469 (HeapWord*) old)); 4470 markOop m = old->mark(); 4471 oop forward_ptr = old->forward_to_atomic(old); 4472 if (forward_ptr == NULL) { 4473 // Forward-to-self succeeded. 4474 assert(_par_scan_state != NULL, "par scan state"); 4475 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4476 uint queue_num = _par_scan_state->queue_num(); 4477 4478 _evacuation_failed = true; 4479 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size()); 4480 if (_evac_failure_closure != cl) { 4481 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4482 assert(!_drain_in_progress, 4483 "Should only be true while someone holds the lock."); 4484 // Set the global evac-failure closure to the current thread's. 4485 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4486 set_evac_failure_closure(cl); 4487 // Now do the common part. 4488 handle_evacuation_failure_common(old, m); 4489 // Reset to NULL. 4490 set_evac_failure_closure(NULL); 4491 } else { 4492 // The lock is already held, and this is recursive. 4493 assert(_drain_in_progress, "This should only be the recursive case."); 4494 handle_evacuation_failure_common(old, m); 4495 } 4496 return old; 4497 } else { 4498 // Forward-to-self failed. Either someone else managed to allocate 4499 // space for this object (old != forward_ptr) or they beat us in 4500 // self-forwarding it (old == forward_ptr). 4501 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4502 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4503 "should not be in the CSet", 4504 (HeapWord*) old, (HeapWord*) forward_ptr)); 4505 return forward_ptr; 4506 } 4507 } 4508 4509 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4510 preserve_mark_if_necessary(old, m); 4511 4512 HeapRegion* r = heap_region_containing(old); 4513 if (!r->evacuation_failed()) { 4514 r->set_evacuation_failed(true); 4515 _hr_printer.evac_failure(r); 4516 } 4517 4518 push_on_evac_failure_scan_stack(old); 4519 4520 if (!_drain_in_progress) { 4521 // prevent recursion in copy_to_survivor_space() 4522 _drain_in_progress = true; 4523 drain_evac_failure_scan_stack(); 4524 _drain_in_progress = false; 4525 } 4526 } 4527 4528 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4529 assert(evacuation_failed(), "Oversaving!"); 4530 // We want to call the "for_promotion_failure" version only in the 4531 // case of a promotion failure. 4532 if (m->must_be_preserved_for_promotion_failure(obj)) { 4533 _objs_with_preserved_marks.push(obj); 4534 _preserved_marks_of_objs.push(m); 4535 } 4536 } 4537 4538 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4539 size_t word_size) { 4540 if (purpose == GCAllocForSurvived) { 4541 HeapWord* result = survivor_attempt_allocation(word_size); 4542 if (result != NULL) { 4543 return result; 4544 } else { 4545 // Let's try to allocate in the old gen in case we can fit the 4546 // object there. 4547 return old_attempt_allocation(word_size); 4548 } 4549 } else { 4550 assert(purpose == GCAllocForTenured, "sanity"); 4551 HeapWord* result = old_attempt_allocation(word_size); 4552 if (result != NULL) { 4553 return result; 4554 } else { 4555 // Let's try to allocate in the survivors in case we can fit the 4556 // object there. 4557 return survivor_attempt_allocation(word_size); 4558 } 4559 } 4560 4561 ShouldNotReachHere(); 4562 // Trying to keep some compilers happy. 4563 return NULL; 4564 } 4565 4566 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4567 ParGCAllocBuffer(gclab_word_size), _retired(true) { } 4568 4569 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, ReferenceProcessor* rp) 4570 : _g1h(g1h), 4571 _refs(g1h->task_queue(queue_num)), 4572 _dcq(&g1h->dirty_card_queue_set()), 4573 _ct_bs(g1h->g1_barrier_set()), 4574 _g1_rem(g1h->g1_rem_set()), 4575 _hash_seed(17), _queue_num(queue_num), 4576 _term_attempts(0), 4577 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4578 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4579 _age_table(false), _scanner(g1h, this, rp), 4580 _strong_roots_time(0), _term_time(0), 4581 _alloc_buffer_waste(0), _undo_waste(0) { 4582 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4583 // we "sacrifice" entry 0 to keep track of surviving bytes for 4584 // non-young regions (where the age is -1) 4585 // We also add a few elements at the beginning and at the end in 4586 // an attempt to eliminate cache contention 4587 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4588 uint array_length = PADDING_ELEM_NUM + 4589 real_length + 4590 PADDING_ELEM_NUM; 4591 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4592 if (_surviving_young_words_base == NULL) 4593 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR, 4594 "Not enough space for young surv histo."); 4595 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4596 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4597 4598 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4599 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4600 4601 _start = os::elapsedTime(); 4602 } 4603 4604 void 4605 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4606 { 4607 st->print_raw_cr("GC Termination Stats"); 4608 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4609 " ------waste (KiB)------"); 4610 st->print_raw_cr("thr ms ms % ms % attempts" 4611 " total alloc undo"); 4612 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4613 " ------- ------- -------"); 4614 } 4615 4616 void 4617 G1ParScanThreadState::print_termination_stats(int i, 4618 outputStream* const st) const 4619 { 4620 const double elapsed_ms = elapsed_time() * 1000.0; 4621 const double s_roots_ms = strong_roots_time() * 1000.0; 4622 const double term_ms = term_time() * 1000.0; 4623 st->print_cr("%3d %9.2f %9.2f %6.2f " 4624 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4625 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4626 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4627 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4628 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4629 alloc_buffer_waste() * HeapWordSize / K, 4630 undo_waste() * HeapWordSize / K); 4631 } 4632 4633 #ifdef ASSERT 4634 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4635 assert(ref != NULL, "invariant"); 4636 assert(UseCompressedOops, "sanity"); 4637 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4638 oop p = oopDesc::load_decode_heap_oop(ref); 4639 assert(_g1h->is_in_g1_reserved(p), 4640 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4641 return true; 4642 } 4643 4644 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4645 assert(ref != NULL, "invariant"); 4646 if (has_partial_array_mask(ref)) { 4647 // Must be in the collection set--it's already been copied. 4648 oop p = clear_partial_array_mask(ref); 4649 assert(_g1h->obj_in_cs(p), 4650 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4651 } else { 4652 oop p = oopDesc::load_decode_heap_oop(ref); 4653 assert(_g1h->is_in_g1_reserved(p), 4654 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4655 } 4656 return true; 4657 } 4658 4659 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4660 if (ref.is_narrow()) { 4661 return verify_ref((narrowOop*) ref); 4662 } else { 4663 return verify_ref((oop*) ref); 4664 } 4665 } 4666 #endif // ASSERT 4667 4668 void G1ParScanThreadState::trim_queue() { 4669 assert(_evac_failure_cl != NULL, "not set"); 4670 4671 StarTask ref; 4672 do { 4673 // Drain the overflow stack first, so other threads can steal. 4674 while (refs()->pop_overflow(ref)) { 4675 deal_with_reference(ref); 4676 } 4677 4678 while (refs()->pop_local(ref)) { 4679 deal_with_reference(ref); 4680 } 4681 } while (!refs()->is_empty()); 4682 } 4683 4684 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4685 G1ParScanThreadState* par_scan_state) : 4686 _g1(g1), _par_scan_state(par_scan_state), 4687 _worker_id(par_scan_state->queue_num()) { } 4688 4689 void G1ParCopyHelper::mark_object(oop obj) { 4690 #ifdef ASSERT 4691 HeapRegion* hr = _g1->heap_region_containing(obj); 4692 assert(hr != NULL, "sanity"); 4693 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4694 #endif // ASSERT 4695 4696 // We know that the object is not moving so it's safe to read its size. 4697 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4698 } 4699 4700 void G1ParCopyHelper::mark_forwarded_object(oop from_obj, oop to_obj) { 4701 #ifdef ASSERT 4702 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4703 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4704 assert(from_obj != to_obj, "should not be self-forwarded"); 4705 4706 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4707 assert(from_hr != NULL, "sanity"); 4708 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4709 4710 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4711 assert(to_hr != NULL, "sanity"); 4712 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4713 #endif // ASSERT 4714 4715 // The object might be in the process of being copied by another 4716 // worker so we cannot trust that its to-space image is 4717 // well-formed. So we have to read its size from its from-space 4718 // image which we know should not be changing. 4719 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4720 } 4721 4722 oop G1ParScanThreadState::copy_to_survivor_space(oop const old) { 4723 size_t word_sz = old->size(); 4724 HeapRegion* from_region = _g1h->heap_region_containing_raw(old); 4725 // +1 to make the -1 indexes valid... 4726 int young_index = from_region->young_index_in_cset()+1; 4727 assert( (from_region->is_young() && young_index > 0) || 4728 (!from_region->is_young() && young_index == 0), "invariant" ); 4729 G1CollectorPolicy* g1p = _g1h->g1_policy(); 4730 markOop m = old->mark(); 4731 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4732 : m->age(); 4733 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4734 word_sz); 4735 HeapWord* obj_ptr = allocate(alloc_purpose, word_sz); 4736 #ifndef PRODUCT 4737 // Should this evacuation fail? 4738 if (_g1h->evacuation_should_fail()) { 4739 if (obj_ptr != NULL) { 4740 undo_allocation(alloc_purpose, obj_ptr, word_sz); 4741 obj_ptr = NULL; 4742 } 4743 } 4744 #endif // !PRODUCT 4745 4746 if (obj_ptr == NULL) { 4747 // This will either forward-to-self, or detect that someone else has 4748 // installed a forwarding pointer. 4749 return _g1h->handle_evacuation_failure_par(this, old); 4750 } 4751 4752 oop obj = oop(obj_ptr); 4753 4754 // We're going to allocate linearly, so might as well prefetch ahead. 4755 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4756 4757 oop forward_ptr = old->forward_to_atomic(obj); 4758 if (forward_ptr == NULL) { 4759 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4760 4761 // alloc_purpose is just a hint to allocate() above, recheck the type of region 4762 // we actually allocated from and update alloc_purpose accordingly 4763 HeapRegion* to_region = _g1h->heap_region_containing_raw(obj_ptr); 4764 alloc_purpose = to_region->is_young() ? GCAllocForSurvived : GCAllocForTenured; 4765 4766 if (g1p->track_object_age(alloc_purpose)) { 4767 // We could simply do obj->incr_age(). However, this causes a 4768 // performance issue. obj->incr_age() will first check whether 4769 // the object has a displaced mark by checking its mark word; 4770 // getting the mark word from the new location of the object 4771 // stalls. So, given that we already have the mark word and we 4772 // are about to install it anyway, it's better to increase the 4773 // age on the mark word, when the object does not have a 4774 // displaced mark word. We're not expecting many objects to have 4775 // a displaced marked word, so that case is not optimized 4776 // further (it could be...) and we simply call obj->incr_age(). 4777 4778 if (m->has_displaced_mark_helper()) { 4779 // in this case, we have to install the mark word first, 4780 // otherwise obj looks to be forwarded (the old mark word, 4781 // which contains the forward pointer, was copied) 4782 obj->set_mark(m); 4783 obj->incr_age(); 4784 } else { 4785 m = m->incr_age(); 4786 obj->set_mark(m); 4787 } 4788 age_table()->add(obj, word_sz); 4789 } else { 4790 obj->set_mark(m); 4791 } 4792 4793 if (G1StringDedup::is_enabled()) { 4794 G1StringDedup::enqueue_from_evacuation(from_region->is_young(), 4795 to_region->is_young(), 4796 queue_num(), 4797 obj); 4798 } 4799 4800 size_t* surv_young_words = surviving_young_words(); 4801 surv_young_words[young_index] += word_sz; 4802 4803 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4804 // We keep track of the next start index in the length field of 4805 // the to-space object. The actual length can be found in the 4806 // length field of the from-space object. 4807 arrayOop(obj)->set_length(0); 4808 oop* old_p = set_partial_array_mask(old); 4809 push_on_queue(old_p); 4810 } else { 4811 // No point in using the slower heap_region_containing() method, 4812 // given that we know obj is in the heap. 4813 _scanner.set_region(_g1h->heap_region_containing_raw(obj)); 4814 obj->oop_iterate_backwards(&_scanner); 4815 } 4816 } else { 4817 undo_allocation(alloc_purpose, obj_ptr, word_sz); 4818 obj = forward_ptr; 4819 } 4820 return obj; 4821 } 4822 4823 template <class T> 4824 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4825 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4826 _scanned_klass->record_modified_oops(); 4827 } 4828 } 4829 4830 template <G1Barrier barrier, bool do_mark_object> 4831 template <class T> 4832 void G1ParCopyClosure<barrier, do_mark_object>::do_oop_work(T* p) { 4833 T heap_oop = oopDesc::load_heap_oop(p); 4834 4835 if (oopDesc::is_null(heap_oop)) { 4836 return; 4837 } 4838 4839 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 4840 4841 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4842 4843 if (_g1->in_cset_fast_test(obj)) { 4844 oop forwardee; 4845 if (obj->is_forwarded()) { 4846 forwardee = obj->forwardee(); 4847 } else { 4848 forwardee = _par_scan_state->copy_to_survivor_space(obj); 4849 } 4850 assert(forwardee != NULL, "forwardee should not be NULL"); 4851 oopDesc::encode_store_heap_oop(p, forwardee); 4852 if (do_mark_object && forwardee != obj) { 4853 // If the object is self-forwarded we don't need to explicitly 4854 // mark it, the evacuation failure protocol will do so. 4855 mark_forwarded_object(obj, forwardee); 4856 } 4857 4858 if (barrier == G1BarrierKlass) { 4859 do_klass_barrier(p, forwardee); 4860 } 4861 } else { 4862 // The object is not in collection set. If we're a root scanning 4863 // closure during an initial mark pause (i.e. do_mark_object will 4864 // be true) then attempt to mark the object. 4865 if (do_mark_object) { 4866 mark_object(obj); 4867 } 4868 } 4869 4870 if (barrier == G1BarrierEvac) { 4871 _par_scan_state->update_rs(_from, p, _worker_id); 4872 } 4873 } 4874 4875 template void G1ParCopyClosure<G1BarrierEvac, false>::do_oop_work(oop* p); 4876 template void G1ParCopyClosure<G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4877 4878 class G1ParEvacuateFollowersClosure : public VoidClosure { 4879 protected: 4880 G1CollectedHeap* _g1h; 4881 G1ParScanThreadState* _par_scan_state; 4882 RefToScanQueueSet* _queues; 4883 ParallelTaskTerminator* _terminator; 4884 4885 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4886 RefToScanQueueSet* queues() { return _queues; } 4887 ParallelTaskTerminator* terminator() { return _terminator; } 4888 4889 public: 4890 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4891 G1ParScanThreadState* par_scan_state, 4892 RefToScanQueueSet* queues, 4893 ParallelTaskTerminator* terminator) 4894 : _g1h(g1h), _par_scan_state(par_scan_state), 4895 _queues(queues), _terminator(terminator) {} 4896 4897 void do_void(); 4898 4899 private: 4900 inline bool offer_termination(); 4901 }; 4902 4903 bool G1ParEvacuateFollowersClosure::offer_termination() { 4904 G1ParScanThreadState* const pss = par_scan_state(); 4905 pss->start_term_time(); 4906 const bool res = terminator()->offer_termination(); 4907 pss->end_term_time(); 4908 return res; 4909 } 4910 4911 void G1ParEvacuateFollowersClosure::do_void() { 4912 StarTask stolen_task; 4913 G1ParScanThreadState* const pss = par_scan_state(); 4914 pss->trim_queue(); 4915 4916 do { 4917 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4918 assert(pss->verify_task(stolen_task), "sanity"); 4919 if (stolen_task.is_narrow()) { 4920 pss->deal_with_reference((narrowOop*) stolen_task); 4921 } else { 4922 pss->deal_with_reference((oop*) stolen_task); 4923 } 4924 4925 // We've just processed a reference and we might have made 4926 // available new entries on the queues. So we have to make sure 4927 // we drain the queues as necessary. 4928 pss->trim_queue(); 4929 } 4930 } while (!offer_termination()); 4931 } 4932 4933 class G1KlassScanClosure : public KlassClosure { 4934 G1ParCopyHelper* _closure; 4935 bool _process_only_dirty; 4936 int _count; 4937 public: 4938 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4939 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4940 void do_klass(Klass* klass) { 4941 // If the klass has not been dirtied we know that there's 4942 // no references into the young gen and we can skip it. 4943 if (!_process_only_dirty || klass->has_modified_oops()) { 4944 // Clean the klass since we're going to scavenge all the metadata. 4945 klass->clear_modified_oops(); 4946 4947 // Tell the closure that this klass is the Klass to scavenge 4948 // and is the one to dirty if oops are left pointing into the young gen. 4949 _closure->set_scanned_klass(klass); 4950 4951 klass->oops_do(_closure); 4952 4953 _closure->set_scanned_klass(NULL); 4954 } 4955 _count++; 4956 } 4957 }; 4958 4959 class G1ParTask : public AbstractGangTask { 4960 protected: 4961 G1CollectedHeap* _g1h; 4962 RefToScanQueueSet *_queues; 4963 ParallelTaskTerminator _terminator; 4964 uint _n_workers; 4965 4966 Mutex _stats_lock; 4967 Mutex* stats_lock() { return &_stats_lock; } 4968 4969 size_t getNCards() { 4970 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4971 / G1BlockOffsetSharedArray::N_bytes; 4972 } 4973 4974 public: 4975 G1ParTask(G1CollectedHeap* g1h, 4976 RefToScanQueueSet *task_queues) 4977 : AbstractGangTask("G1 collection"), 4978 _g1h(g1h), 4979 _queues(task_queues), 4980 _terminator(0, _queues), 4981 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4982 {} 4983 4984 RefToScanQueueSet* queues() { return _queues; } 4985 4986 RefToScanQueue *work_queue(int i) { 4987 return queues()->queue(i); 4988 } 4989 4990 ParallelTaskTerminator* terminator() { return &_terminator; } 4991 4992 virtual void set_for_termination(int active_workers) { 4993 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4994 // in the young space (_par_seq_tasks) in the G1 heap 4995 // for SequentialSubTasksDone. 4996 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4997 // both of which need setting by set_n_termination(). 4998 _g1h->SharedHeap::set_n_termination(active_workers); 4999 _g1h->set_n_termination(active_workers); 5000 terminator()->reset_for_reuse(active_workers); 5001 _n_workers = active_workers; 5002 } 5003 5004 void work(uint worker_id) { 5005 if (worker_id >= _n_workers) return; // no work needed this round 5006 5007 double start_time_ms = os::elapsedTime() * 1000.0; 5008 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 5009 5010 { 5011 ResourceMark rm; 5012 HandleMark hm; 5013 5014 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 5015 5016 G1ParScanThreadState pss(_g1h, worker_id, rp); 5017 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 5018 5019 pss.set_evac_failure_closure(&evac_failure_cl); 5020 5021 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 5022 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 5023 5024 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 5025 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 5026 5027 bool only_young = _g1h->g1_policy()->gcs_are_young(); 5028 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 5029 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 5030 5031 OopClosure* scan_root_cl = &only_scan_root_cl; 5032 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 5033 5034 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5035 // We also need to mark copied objects. 5036 scan_root_cl = &scan_mark_root_cl; 5037 scan_klasses_cl = &scan_mark_klasses_cl_s; 5038 } 5039 5040 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 5041 5042 // Don't scan the scavengable methods in the code cache as part 5043 // of strong root scanning. The code roots that point into a 5044 // region in the collection set are scanned when we scan the 5045 // region's RSet. 5046 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings; 5047 5048 pss.start_strong_roots(); 5049 _g1h->g1_process_strong_roots(/* is scavenging */ true, 5050 SharedHeap::ScanningOption(so), 5051 scan_root_cl, 5052 &push_heap_rs_cl, 5053 scan_klasses_cl, 5054 worker_id); 5055 pss.end_strong_roots(); 5056 5057 { 5058 double start = os::elapsedTime(); 5059 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 5060 evac.do_void(); 5061 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 5062 double term_ms = pss.term_time()*1000.0; 5063 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 5064 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 5065 } 5066 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 5067 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 5068 5069 if (ParallelGCVerbose) { 5070 MutexLocker x(stats_lock()); 5071 pss.print_termination_stats(worker_id); 5072 } 5073 5074 assert(pss.refs()->is_empty(), "should be empty"); 5075 5076 // Close the inner scope so that the ResourceMark and HandleMark 5077 // destructors are executed here and are included as part of the 5078 // "GC Worker Time". 5079 } 5080 5081 double end_time_ms = os::elapsedTime() * 1000.0; 5082 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 5083 } 5084 }; 5085 5086 // *** Common G1 Evacuation Stuff 5087 5088 // This method is run in a GC worker. 5089 5090 void 5091 G1CollectedHeap:: 5092 g1_process_strong_roots(bool is_scavenging, 5093 ScanningOption so, 5094 OopClosure* scan_non_heap_roots, 5095 OopsInHeapRegionClosure* scan_rs, 5096 G1KlassScanClosure* scan_klasses, 5097 uint worker_i) { 5098 5099 // First scan the strong roots 5100 double ext_roots_start = os::elapsedTime(); 5101 double closure_app_time_sec = 0.0; 5102 5103 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 5104 5105 assert(so & SO_CodeCache || scan_rs != NULL, "must scan code roots somehow"); 5106 // Walk the code cache/strong code roots w/o buffering, because StarTask 5107 // cannot handle unaligned oop locations. 5108 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, true /* do_marking */); 5109 5110 process_strong_roots(false, // no scoping; this is parallel code 5111 is_scavenging, so, 5112 &buf_scan_non_heap_roots, 5113 &eager_scan_code_roots, 5114 scan_klasses 5115 ); 5116 5117 // Now the CM ref_processor roots. 5118 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5119 // We need to treat the discovered reference lists of the 5120 // concurrent mark ref processor as roots and keep entries 5121 // (which are added by the marking threads) on them live 5122 // until they can be processed at the end of marking. 5123 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5124 } 5125 5126 // Finish up any enqueued closure apps (attributed as object copy time). 5127 buf_scan_non_heap_roots.done(); 5128 5129 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5130 5131 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5132 5133 double ext_root_time_ms = 5134 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5135 5136 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5137 5138 // During conc marking we have to filter the per-thread SATB buffers 5139 // to make sure we remove any oops into the CSet (which will show up 5140 // as implicitly live). 5141 double satb_filtering_ms = 0.0; 5142 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5143 if (mark_in_progress()) { 5144 double satb_filter_start = os::elapsedTime(); 5145 5146 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5147 5148 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5149 } 5150 } 5151 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5152 5153 // If this is an initial mark pause, and we're not scanning 5154 // the entire code cache, we need to mark the oops in the 5155 // strong code root lists for the regions that are not in 5156 // the collection set. 5157 // Note all threads participate in this set of root tasks. 5158 double mark_strong_code_roots_ms = 0.0; 5159 if (g1_policy()->during_initial_mark_pause() && !(so & SO_CodeCache)) { 5160 double mark_strong_roots_start = os::elapsedTime(); 5161 mark_strong_code_roots(worker_i); 5162 mark_strong_code_roots_ms = (os::elapsedTime() - mark_strong_roots_start) * 1000.0; 5163 } 5164 g1_policy()->phase_times()->record_strong_code_root_mark_time(worker_i, mark_strong_code_roots_ms); 5165 5166 // Now scan the complement of the collection set. 5167 if (scan_rs != NULL) { 5168 g1_rem_set()->oops_into_collection_set_do(scan_rs, &eager_scan_code_roots, worker_i); 5169 } 5170 _process_strong_tasks->all_tasks_completed(); 5171 } 5172 5173 void 5174 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) { 5175 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5176 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs); 5177 } 5178 5179 class G1StringSymbolTableUnlinkTask : public AbstractGangTask { 5180 private: 5181 BoolObjectClosure* _is_alive; 5182 int _initial_string_table_size; 5183 int _initial_symbol_table_size; 5184 5185 bool _process_strings; 5186 int _strings_processed; 5187 int _strings_removed; 5188 5189 bool _process_symbols; 5190 int _symbols_processed; 5191 int _symbols_removed; 5192 5193 bool _do_in_parallel; 5194 public: 5195 G1StringSymbolTableUnlinkTask(BoolObjectClosure* is_alive, bool process_strings, bool process_symbols) : 5196 AbstractGangTask("Par String/Symbol table unlink"), _is_alive(is_alive), 5197 _do_in_parallel(G1CollectedHeap::use_parallel_gc_threads()), 5198 _process_strings(process_strings), _strings_processed(0), _strings_removed(0), 5199 _process_symbols(process_symbols), _symbols_processed(0), _symbols_removed(0) { 5200 5201 _initial_string_table_size = StringTable::the_table()->table_size(); 5202 _initial_symbol_table_size = SymbolTable::the_table()->table_size(); 5203 if (process_strings) { 5204 StringTable::clear_parallel_claimed_index(); 5205 } 5206 if (process_symbols) { 5207 SymbolTable::clear_parallel_claimed_index(); 5208 } 5209 } 5210 5211 ~G1StringSymbolTableUnlinkTask() { 5212 guarantee(!_process_strings || !_do_in_parallel || StringTable::parallel_claimed_index() >= _initial_string_table_size, 5213 err_msg("claim value "INT32_FORMAT" after unlink less than initial string table size "INT32_FORMAT, 5214 StringTable::parallel_claimed_index(), _initial_string_table_size)); 5215 guarantee(!_process_symbols || !_do_in_parallel || SymbolTable::parallel_claimed_index() >= _initial_symbol_table_size, 5216 err_msg("claim value "INT32_FORMAT" after unlink less than initial symbol table size "INT32_FORMAT, 5217 SymbolTable::parallel_claimed_index(), _initial_symbol_table_size)); 5218 } 5219 5220 void work(uint worker_id) { 5221 if (_do_in_parallel) { 5222 int strings_processed = 0; 5223 int strings_removed = 0; 5224 int symbols_processed = 0; 5225 int symbols_removed = 0; 5226 if (_process_strings) { 5227 StringTable::possibly_parallel_unlink(_is_alive, &strings_processed, &strings_removed); 5228 Atomic::add(strings_processed, &_strings_processed); 5229 Atomic::add(strings_removed, &_strings_removed); 5230 } 5231 if (_process_symbols) { 5232 SymbolTable::possibly_parallel_unlink(&symbols_processed, &symbols_removed); 5233 Atomic::add(symbols_processed, &_symbols_processed); 5234 Atomic::add(symbols_removed, &_symbols_removed); 5235 } 5236 } else { 5237 if (_process_strings) { 5238 StringTable::unlink(_is_alive, &_strings_processed, &_strings_removed); 5239 } 5240 if (_process_symbols) { 5241 SymbolTable::unlink(&_symbols_processed, &_symbols_removed); 5242 } 5243 } 5244 } 5245 5246 size_t strings_processed() const { return (size_t)_strings_processed; } 5247 size_t strings_removed() const { return (size_t)_strings_removed; } 5248 5249 size_t symbols_processed() const { return (size_t)_symbols_processed; } 5250 size_t symbols_removed() const { return (size_t)_symbols_removed; } 5251 }; 5252 5253 void G1CollectedHeap::unlink_string_and_symbol_table(BoolObjectClosure* is_alive, 5254 bool process_strings, bool process_symbols) { 5255 uint n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5256 _g1h->workers()->active_workers() : 1); 5257 5258 G1StringSymbolTableUnlinkTask g1_unlink_task(is_alive, process_strings, process_symbols); 5259 if (G1CollectedHeap::use_parallel_gc_threads()) { 5260 set_par_threads(n_workers); 5261 workers()->run_task(&g1_unlink_task); 5262 set_par_threads(0); 5263 } else { 5264 g1_unlink_task.work(0); 5265 } 5266 if (G1TraceStringSymbolTableScrubbing) { 5267 gclog_or_tty->print_cr("Cleaned string and symbol table, " 5268 "strings: "SIZE_FORMAT" processed, "SIZE_FORMAT" removed, " 5269 "symbols: "SIZE_FORMAT" processed, "SIZE_FORMAT" removed", 5270 g1_unlink_task.strings_processed(), g1_unlink_task.strings_removed(), 5271 g1_unlink_task.symbols_processed(), g1_unlink_task.symbols_removed()); 5272 } 5273 5274 if (G1StringDedup::is_enabled()) { 5275 G1StringDedup::unlink(is_alive); 5276 } 5277 } 5278 5279 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 5280 private: 5281 size_t _num_processed; 5282 5283 public: 5284 RedirtyLoggedCardTableEntryFastClosure() : CardTableEntryClosure(), _num_processed(0) { } 5285 5286 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 5287 *card_ptr = CardTableModRefBS::dirty_card_val(); 5288 _num_processed++; 5289 return true; 5290 } 5291 5292 size_t num_processed() const { return _num_processed; } 5293 }; 5294 5295 class G1RedirtyLoggedCardsTask : public AbstractGangTask { 5296 private: 5297 DirtyCardQueueSet* _queue; 5298 public: 5299 G1RedirtyLoggedCardsTask(DirtyCardQueueSet* queue) : AbstractGangTask("Redirty Cards"), _queue(queue) { } 5300 5301 virtual void work(uint worker_id) { 5302 double start_time = os::elapsedTime(); 5303 5304 RedirtyLoggedCardTableEntryFastClosure cl; 5305 if (G1CollectedHeap::heap()->use_parallel_gc_threads()) { 5306 _queue->par_apply_closure_to_all_completed_buffers(&cl); 5307 } else { 5308 _queue->apply_closure_to_all_completed_buffers(&cl); 5309 } 5310 5311 G1GCPhaseTimes* timer = G1CollectedHeap::heap()->g1_policy()->phase_times(); 5312 timer->record_redirty_logged_cards_time_ms(worker_id, (os::elapsedTime() - start_time) * 1000.0); 5313 timer->record_redirty_logged_cards_processed_cards(worker_id, cl.num_processed()); 5314 } 5315 }; 5316 5317 void G1CollectedHeap::redirty_logged_cards() { 5318 guarantee(G1DeferredRSUpdate, "Must only be called when using deferred RS updates."); 5319 double redirty_logged_cards_start = os::elapsedTime(); 5320 5321 uint n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5322 _g1h->workers()->active_workers() : 1); 5323 5324 G1RedirtyLoggedCardsTask redirty_task(&dirty_card_queue_set()); 5325 dirty_card_queue_set().reset_for_par_iteration(); 5326 if (use_parallel_gc_threads()) { 5327 set_par_threads(n_workers); 5328 workers()->run_task(&redirty_task); 5329 set_par_threads(0); 5330 } else { 5331 redirty_task.work(0); 5332 } 5333 5334 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5335 dcq.merge_bufferlists(&dirty_card_queue_set()); 5336 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5337 5338 g1_policy()->phase_times()->record_redirty_logged_cards_time_ms((os::elapsedTime() - redirty_logged_cards_start) * 1000.0); 5339 } 5340 5341 // Weak Reference Processing support 5342 5343 // An always "is_alive" closure that is used to preserve referents. 5344 // If the object is non-null then it's alive. Used in the preservation 5345 // of referent objects that are pointed to by reference objects 5346 // discovered by the CM ref processor. 5347 class G1AlwaysAliveClosure: public BoolObjectClosure { 5348 G1CollectedHeap* _g1; 5349 public: 5350 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5351 bool do_object_b(oop p) { 5352 if (p != NULL) { 5353 return true; 5354 } 5355 return false; 5356 } 5357 }; 5358 5359 bool G1STWIsAliveClosure::do_object_b(oop p) { 5360 // An object is reachable if it is outside the collection set, 5361 // or is inside and copied. 5362 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5363 } 5364 5365 // Non Copying Keep Alive closure 5366 class G1KeepAliveClosure: public OopClosure { 5367 G1CollectedHeap* _g1; 5368 public: 5369 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5370 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5371 void do_oop( oop* p) { 5372 oop obj = *p; 5373 5374 if (_g1->obj_in_cs(obj)) { 5375 assert( obj->is_forwarded(), "invariant" ); 5376 *p = obj->forwardee(); 5377 } 5378 } 5379 }; 5380 5381 // Copying Keep Alive closure - can be called from both 5382 // serial and parallel code as long as different worker 5383 // threads utilize different G1ParScanThreadState instances 5384 // and different queues. 5385 5386 class G1CopyingKeepAliveClosure: public OopClosure { 5387 G1CollectedHeap* _g1h; 5388 OopClosure* _copy_non_heap_obj_cl; 5389 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5390 G1ParScanThreadState* _par_scan_state; 5391 5392 public: 5393 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5394 OopClosure* non_heap_obj_cl, 5395 OopsInHeapRegionClosure* metadata_obj_cl, 5396 G1ParScanThreadState* pss): 5397 _g1h(g1h), 5398 _copy_non_heap_obj_cl(non_heap_obj_cl), 5399 _copy_metadata_obj_cl(metadata_obj_cl), 5400 _par_scan_state(pss) 5401 {} 5402 5403 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5404 virtual void do_oop( oop* p) { do_oop_work(p); } 5405 5406 template <class T> void do_oop_work(T* p) { 5407 oop obj = oopDesc::load_decode_heap_oop(p); 5408 5409 if (_g1h->obj_in_cs(obj)) { 5410 // If the referent object has been forwarded (either copied 5411 // to a new location or to itself in the event of an 5412 // evacuation failure) then we need to update the reference 5413 // field and, if both reference and referent are in the G1 5414 // heap, update the RSet for the referent. 5415 // 5416 // If the referent has not been forwarded then we have to keep 5417 // it alive by policy. Therefore we have copy the referent. 5418 // 5419 // If the reference field is in the G1 heap then we can push 5420 // on the PSS queue. When the queue is drained (after each 5421 // phase of reference processing) the object and it's followers 5422 // will be copied, the reference field set to point to the 5423 // new location, and the RSet updated. Otherwise we need to 5424 // use the the non-heap or metadata closures directly to copy 5425 // the referent object and update the pointer, while avoiding 5426 // updating the RSet. 5427 5428 if (_g1h->is_in_g1_reserved(p)) { 5429 _par_scan_state->push_on_queue(p); 5430 } else { 5431 assert(!Metaspace::contains((const void*)p), 5432 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5433 PTR_FORMAT, p)); 5434 _copy_non_heap_obj_cl->do_oop(p); 5435 } 5436 } 5437 } 5438 }; 5439 5440 // Serial drain queue closure. Called as the 'complete_gc' 5441 // closure for each discovered list in some of the 5442 // reference processing phases. 5443 5444 class G1STWDrainQueueClosure: public VoidClosure { 5445 protected: 5446 G1CollectedHeap* _g1h; 5447 G1ParScanThreadState* _par_scan_state; 5448 5449 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5450 5451 public: 5452 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5453 _g1h(g1h), 5454 _par_scan_state(pss) 5455 { } 5456 5457 void do_void() { 5458 G1ParScanThreadState* const pss = par_scan_state(); 5459 pss->trim_queue(); 5460 } 5461 }; 5462 5463 // Parallel Reference Processing closures 5464 5465 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5466 // processing during G1 evacuation pauses. 5467 5468 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5469 private: 5470 G1CollectedHeap* _g1h; 5471 RefToScanQueueSet* _queues; 5472 FlexibleWorkGang* _workers; 5473 int _active_workers; 5474 5475 public: 5476 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5477 FlexibleWorkGang* workers, 5478 RefToScanQueueSet *task_queues, 5479 int n_workers) : 5480 _g1h(g1h), 5481 _queues(task_queues), 5482 _workers(workers), 5483 _active_workers(n_workers) 5484 { 5485 assert(n_workers > 0, "shouldn't call this otherwise"); 5486 } 5487 5488 // Executes the given task using concurrent marking worker threads. 5489 virtual void execute(ProcessTask& task); 5490 virtual void execute(EnqueueTask& task); 5491 }; 5492 5493 // Gang task for possibly parallel reference processing 5494 5495 class G1STWRefProcTaskProxy: public AbstractGangTask { 5496 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5497 ProcessTask& _proc_task; 5498 G1CollectedHeap* _g1h; 5499 RefToScanQueueSet *_task_queues; 5500 ParallelTaskTerminator* _terminator; 5501 5502 public: 5503 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5504 G1CollectedHeap* g1h, 5505 RefToScanQueueSet *task_queues, 5506 ParallelTaskTerminator* terminator) : 5507 AbstractGangTask("Process reference objects in parallel"), 5508 _proc_task(proc_task), 5509 _g1h(g1h), 5510 _task_queues(task_queues), 5511 _terminator(terminator) 5512 {} 5513 5514 virtual void work(uint worker_id) { 5515 // The reference processing task executed by a single worker. 5516 ResourceMark rm; 5517 HandleMark hm; 5518 5519 G1STWIsAliveClosure is_alive(_g1h); 5520 5521 G1ParScanThreadState pss(_g1h, worker_id, NULL); 5522 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5523 5524 pss.set_evac_failure_closure(&evac_failure_cl); 5525 5526 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5527 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5528 5529 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5530 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5531 5532 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5533 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5534 5535 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5536 // We also need to mark copied objects. 5537 copy_non_heap_cl = ©_mark_non_heap_cl; 5538 copy_metadata_cl = ©_mark_metadata_cl; 5539 } 5540 5541 // Keep alive closure. 5542 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5543 5544 // Complete GC closure 5545 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5546 5547 // Call the reference processing task's work routine. 5548 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5549 5550 // Note we cannot assert that the refs array is empty here as not all 5551 // of the processing tasks (specifically phase2 - pp2_work) execute 5552 // the complete_gc closure (which ordinarily would drain the queue) so 5553 // the queue may not be empty. 5554 } 5555 }; 5556 5557 // Driver routine for parallel reference processing. 5558 // Creates an instance of the ref processing gang 5559 // task and has the worker threads execute it. 5560 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5561 assert(_workers != NULL, "Need parallel worker threads."); 5562 5563 ParallelTaskTerminator terminator(_active_workers, _queues); 5564 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5565 5566 _g1h->set_par_threads(_active_workers); 5567 _workers->run_task(&proc_task_proxy); 5568 _g1h->set_par_threads(0); 5569 } 5570 5571 // Gang task for parallel reference enqueueing. 5572 5573 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5574 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5575 EnqueueTask& _enq_task; 5576 5577 public: 5578 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5579 AbstractGangTask("Enqueue reference objects in parallel"), 5580 _enq_task(enq_task) 5581 { } 5582 5583 virtual void work(uint worker_id) { 5584 _enq_task.work(worker_id); 5585 } 5586 }; 5587 5588 // Driver routine for parallel reference enqueueing. 5589 // Creates an instance of the ref enqueueing gang 5590 // task and has the worker threads execute it. 5591 5592 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5593 assert(_workers != NULL, "Need parallel worker threads."); 5594 5595 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5596 5597 _g1h->set_par_threads(_active_workers); 5598 _workers->run_task(&enq_task_proxy); 5599 _g1h->set_par_threads(0); 5600 } 5601 5602 // End of weak reference support closures 5603 5604 // Abstract task used to preserve (i.e. copy) any referent objects 5605 // that are in the collection set and are pointed to by reference 5606 // objects discovered by the CM ref processor. 5607 5608 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5609 protected: 5610 G1CollectedHeap* _g1h; 5611 RefToScanQueueSet *_queues; 5612 ParallelTaskTerminator _terminator; 5613 uint _n_workers; 5614 5615 public: 5616 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5617 AbstractGangTask("ParPreserveCMReferents"), 5618 _g1h(g1h), 5619 _queues(task_queues), 5620 _terminator(workers, _queues), 5621 _n_workers(workers) 5622 { } 5623 5624 void work(uint worker_id) { 5625 ResourceMark rm; 5626 HandleMark hm; 5627 5628 G1ParScanThreadState pss(_g1h, worker_id, NULL); 5629 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5630 5631 pss.set_evac_failure_closure(&evac_failure_cl); 5632 5633 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5634 5635 5636 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5637 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5638 5639 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5640 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5641 5642 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5643 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5644 5645 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5646 // We also need to mark copied objects. 5647 copy_non_heap_cl = ©_mark_non_heap_cl; 5648 copy_metadata_cl = ©_mark_metadata_cl; 5649 } 5650 5651 // Is alive closure 5652 G1AlwaysAliveClosure always_alive(_g1h); 5653 5654 // Copying keep alive closure. Applied to referent objects that need 5655 // to be copied. 5656 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5657 5658 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5659 5660 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5661 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5662 5663 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5664 // So this must be true - but assert just in case someone decides to 5665 // change the worker ids. 5666 assert(0 <= worker_id && worker_id < limit, "sanity"); 5667 assert(!rp->discovery_is_atomic(), "check this code"); 5668 5669 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5670 for (uint idx = worker_id; idx < limit; idx += stride) { 5671 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5672 5673 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5674 while (iter.has_next()) { 5675 // Since discovery is not atomic for the CM ref processor, we 5676 // can see some null referent objects. 5677 iter.load_ptrs(DEBUG_ONLY(true)); 5678 oop ref = iter.obj(); 5679 5680 // This will filter nulls. 5681 if (iter.is_referent_alive()) { 5682 iter.make_referent_alive(); 5683 } 5684 iter.move_to_next(); 5685 } 5686 } 5687 5688 // Drain the queue - which may cause stealing 5689 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5690 drain_queue.do_void(); 5691 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5692 assert(pss.refs()->is_empty(), "should be"); 5693 } 5694 }; 5695 5696 // Weak Reference processing during an evacuation pause (part 1). 5697 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5698 double ref_proc_start = os::elapsedTime(); 5699 5700 ReferenceProcessor* rp = _ref_processor_stw; 5701 assert(rp->discovery_enabled(), "should have been enabled"); 5702 5703 // Any reference objects, in the collection set, that were 'discovered' 5704 // by the CM ref processor should have already been copied (either by 5705 // applying the external root copy closure to the discovered lists, or 5706 // by following an RSet entry). 5707 // 5708 // But some of the referents, that are in the collection set, that these 5709 // reference objects point to may not have been copied: the STW ref 5710 // processor would have seen that the reference object had already 5711 // been 'discovered' and would have skipped discovering the reference, 5712 // but would not have treated the reference object as a regular oop. 5713 // As a result the copy closure would not have been applied to the 5714 // referent object. 5715 // 5716 // We need to explicitly copy these referent objects - the references 5717 // will be processed at the end of remarking. 5718 // 5719 // We also need to do this copying before we process the reference 5720 // objects discovered by the STW ref processor in case one of these 5721 // referents points to another object which is also referenced by an 5722 // object discovered by the STW ref processor. 5723 5724 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5725 no_of_gc_workers == workers()->active_workers(), 5726 "Need to reset active GC workers"); 5727 5728 set_par_threads(no_of_gc_workers); 5729 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5730 no_of_gc_workers, 5731 _task_queues); 5732 5733 if (G1CollectedHeap::use_parallel_gc_threads()) { 5734 workers()->run_task(&keep_cm_referents); 5735 } else { 5736 keep_cm_referents.work(0); 5737 } 5738 5739 set_par_threads(0); 5740 5741 // Closure to test whether a referent is alive. 5742 G1STWIsAliveClosure is_alive(this); 5743 5744 // Even when parallel reference processing is enabled, the processing 5745 // of JNI refs is serial and performed serially by the current thread 5746 // rather than by a worker. The following PSS will be used for processing 5747 // JNI refs. 5748 5749 // Use only a single queue for this PSS. 5750 G1ParScanThreadState pss(this, 0, NULL); 5751 5752 // We do not embed a reference processor in the copying/scanning 5753 // closures while we're actually processing the discovered 5754 // reference objects. 5755 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5756 5757 pss.set_evac_failure_closure(&evac_failure_cl); 5758 5759 assert(pss.refs()->is_empty(), "pre-condition"); 5760 5761 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5762 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5763 5764 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5765 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5766 5767 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5768 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5769 5770 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5771 // We also need to mark copied objects. 5772 copy_non_heap_cl = ©_mark_non_heap_cl; 5773 copy_metadata_cl = ©_mark_metadata_cl; 5774 } 5775 5776 // Keep alive closure. 5777 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5778 5779 // Serial Complete GC closure 5780 G1STWDrainQueueClosure drain_queue(this, &pss); 5781 5782 // Setup the soft refs policy... 5783 rp->setup_policy(false); 5784 5785 ReferenceProcessorStats stats; 5786 if (!rp->processing_is_mt()) { 5787 // Serial reference processing... 5788 stats = rp->process_discovered_references(&is_alive, 5789 &keep_alive, 5790 &drain_queue, 5791 NULL, 5792 _gc_timer_stw, 5793 _gc_tracer_stw->gc_id()); 5794 } else { 5795 // Parallel reference processing 5796 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5797 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5798 5799 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5800 stats = rp->process_discovered_references(&is_alive, 5801 &keep_alive, 5802 &drain_queue, 5803 &par_task_executor, 5804 _gc_timer_stw, 5805 _gc_tracer_stw->gc_id()); 5806 } 5807 5808 _gc_tracer_stw->report_gc_reference_stats(stats); 5809 5810 // We have completed copying any necessary live referent objects. 5811 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5812 5813 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5814 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5815 } 5816 5817 // Weak Reference processing during an evacuation pause (part 2). 5818 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5819 double ref_enq_start = os::elapsedTime(); 5820 5821 ReferenceProcessor* rp = _ref_processor_stw; 5822 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5823 5824 // Now enqueue any remaining on the discovered lists on to 5825 // the pending list. 5826 if (!rp->processing_is_mt()) { 5827 // Serial reference processing... 5828 rp->enqueue_discovered_references(); 5829 } else { 5830 // Parallel reference enqueueing 5831 5832 assert(no_of_gc_workers == workers()->active_workers(), 5833 "Need to reset active workers"); 5834 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5835 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5836 5837 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5838 rp->enqueue_discovered_references(&par_task_executor); 5839 } 5840 5841 rp->verify_no_references_recorded(); 5842 assert(!rp->discovery_enabled(), "should have been disabled"); 5843 5844 // FIXME 5845 // CM's reference processing also cleans up the string and symbol tables. 5846 // Should we do that here also? We could, but it is a serial operation 5847 // and could significantly increase the pause time. 5848 5849 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5850 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5851 } 5852 5853 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { 5854 _expand_heap_after_alloc_failure = true; 5855 _evacuation_failed = false; 5856 5857 // Should G1EvacuationFailureALot be in effect for this GC? 5858 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5859 5860 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5861 5862 // Disable the hot card cache. 5863 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 5864 hot_card_cache->reset_hot_cache_claimed_index(); 5865 hot_card_cache->set_use_cache(false); 5866 5867 uint n_workers; 5868 if (G1CollectedHeap::use_parallel_gc_threads()) { 5869 n_workers = 5870 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5871 workers()->active_workers(), 5872 Threads::number_of_non_daemon_threads()); 5873 assert(UseDynamicNumberOfGCThreads || 5874 n_workers == workers()->total_workers(), 5875 "If not dynamic should be using all the workers"); 5876 workers()->set_active_workers(n_workers); 5877 set_par_threads(n_workers); 5878 } else { 5879 assert(n_par_threads() == 0, 5880 "Should be the original non-parallel value"); 5881 n_workers = 1; 5882 } 5883 5884 G1ParTask g1_par_task(this, _task_queues); 5885 5886 init_for_evac_failure(NULL); 5887 5888 rem_set()->prepare_for_younger_refs_iterate(true); 5889 5890 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5891 double start_par_time_sec = os::elapsedTime(); 5892 double end_par_time_sec; 5893 5894 { 5895 StrongRootsScope srs(this); 5896 5897 if (G1CollectedHeap::use_parallel_gc_threads()) { 5898 // The individual threads will set their evac-failure closures. 5899 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5900 // These tasks use ShareHeap::_process_strong_tasks 5901 assert(UseDynamicNumberOfGCThreads || 5902 workers()->active_workers() == workers()->total_workers(), 5903 "If not dynamic should be using all the workers"); 5904 workers()->run_task(&g1_par_task); 5905 } else { 5906 g1_par_task.set_for_termination(n_workers); 5907 g1_par_task.work(0); 5908 } 5909 end_par_time_sec = os::elapsedTime(); 5910 5911 // Closing the inner scope will execute the destructor 5912 // for the StrongRootsScope object. We record the current 5913 // elapsed time before closing the scope so that time 5914 // taken for the SRS destructor is NOT included in the 5915 // reported parallel time. 5916 } 5917 5918 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5919 g1_policy()->phase_times()->record_par_time(par_time_ms); 5920 5921 double code_root_fixup_time_ms = 5922 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5923 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5924 5925 set_par_threads(0); 5926 5927 // Process any discovered reference objects - we have 5928 // to do this _before_ we retire the GC alloc regions 5929 // as we may have to copy some 'reachable' referent 5930 // objects (and their reachable sub-graphs) that were 5931 // not copied during the pause. 5932 process_discovered_references(n_workers); 5933 5934 // Weak root processing. 5935 { 5936 G1STWIsAliveClosure is_alive(this); 5937 G1KeepAliveClosure keep_alive(this); 5938 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5939 if (G1StringDedup::is_enabled()) { 5940 G1StringDedup::unlink_or_oops_do(&is_alive, &keep_alive); 5941 } 5942 } 5943 5944 release_gc_alloc_regions(n_workers, evacuation_info); 5945 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5946 5947 // Reset and re-enable the hot card cache. 5948 // Note the counts for the cards in the regions in the 5949 // collection set are reset when the collection set is freed. 5950 hot_card_cache->reset_hot_cache(); 5951 hot_card_cache->set_use_cache(true); 5952 5953 // Migrate the strong code roots attached to each region in 5954 // the collection set. Ideally we would like to do this 5955 // after we have finished the scanning/evacuation of the 5956 // strong code roots for a particular heap region. 5957 migrate_strong_code_roots(); 5958 5959 purge_code_root_memory(); 5960 5961 if (g1_policy()->during_initial_mark_pause()) { 5962 // Reset the claim values set during marking the strong code roots 5963 reset_heap_region_claim_values(); 5964 } 5965 5966 finalize_for_evac_failure(); 5967 5968 if (evacuation_failed()) { 5969 remove_self_forwarding_pointers(); 5970 5971 // Reset the G1EvacuationFailureALot counters and flags 5972 // Note: the values are reset only when an actual 5973 // evacuation failure occurs. 5974 NOT_PRODUCT(reset_evacuation_should_fail();) 5975 } 5976 5977 // Enqueue any remaining references remaining on the STW 5978 // reference processor's discovered lists. We need to do 5979 // this after the card table is cleaned (and verified) as 5980 // the act of enqueueing entries on to the pending list 5981 // will log these updates (and dirty their associated 5982 // cards). We need these updates logged to update any 5983 // RSets. 5984 enqueue_discovered_references(n_workers); 5985 5986 if (G1DeferredRSUpdate) { 5987 redirty_logged_cards(); 5988 } 5989 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5990 } 5991 5992 void G1CollectedHeap::free_region(HeapRegion* hr, 5993 FreeRegionList* free_list, 5994 bool par, 5995 bool locked) { 5996 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5997 assert(!hr->is_empty(), "the region should not be empty"); 5998 assert(free_list != NULL, "pre-condition"); 5999 6000 // Clear the card counts for this region. 6001 // Note: we only need to do this if the region is not young 6002 // (since we don't refine cards in young regions). 6003 if (!hr->is_young()) { 6004 _cg1r->hot_card_cache()->reset_card_counts(hr); 6005 } 6006 hr->hr_clear(par, true /* clear_space */, locked /* locked */); 6007 free_list->add_ordered(hr); 6008 } 6009 6010 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 6011 FreeRegionList* free_list, 6012 bool par) { 6013 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 6014 assert(free_list != NULL, "pre-condition"); 6015 6016 size_t hr_capacity = hr->capacity(); 6017 // We need to read this before we make the region non-humongous, 6018 // otherwise the information will be gone. 6019 uint last_index = hr->last_hc_index(); 6020 hr->set_notHumongous(); 6021 free_region(hr, free_list, par); 6022 6023 uint i = hr->hrs_index() + 1; 6024 while (i < last_index) { 6025 HeapRegion* curr_hr = region_at(i); 6026 assert(curr_hr->continuesHumongous(), "invariant"); 6027 curr_hr->set_notHumongous(); 6028 free_region(curr_hr, free_list, par); 6029 i += 1; 6030 } 6031 } 6032 6033 void G1CollectedHeap::remove_from_old_sets(const HeapRegionSetCount& old_regions_removed, 6034 const HeapRegionSetCount& humongous_regions_removed) { 6035 if (old_regions_removed.length() > 0 || humongous_regions_removed.length() > 0) { 6036 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 6037 _old_set.bulk_remove(old_regions_removed); 6038 _humongous_set.bulk_remove(humongous_regions_removed); 6039 } 6040 6041 } 6042 6043 void G1CollectedHeap::prepend_to_freelist(FreeRegionList* list) { 6044 assert(list != NULL, "list can't be null"); 6045 if (!list->is_empty()) { 6046 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 6047 _free_list.add_ordered(list); 6048 } 6049 } 6050 6051 void G1CollectedHeap::decrement_summary_bytes(size_t bytes) { 6052 assert(_summary_bytes_used >= bytes, 6053 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" should be >= bytes: "SIZE_FORMAT, 6054 _summary_bytes_used, bytes)); 6055 _summary_bytes_used -= bytes; 6056 } 6057 6058 class G1ParCleanupCTTask : public AbstractGangTask { 6059 G1SATBCardTableModRefBS* _ct_bs; 6060 G1CollectedHeap* _g1h; 6061 HeapRegion* volatile _su_head; 6062 public: 6063 G1ParCleanupCTTask(G1SATBCardTableModRefBS* ct_bs, 6064 G1CollectedHeap* g1h) : 6065 AbstractGangTask("G1 Par Cleanup CT Task"), 6066 _ct_bs(ct_bs), _g1h(g1h) { } 6067 6068 void work(uint worker_id) { 6069 HeapRegion* r; 6070 while (r = _g1h->pop_dirty_cards_region()) { 6071 clear_cards(r); 6072 } 6073 } 6074 6075 void clear_cards(HeapRegion* r) { 6076 // Cards of the survivors should have already been dirtied. 6077 if (!r->is_survivor()) { 6078 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 6079 } 6080 } 6081 }; 6082 6083 #ifndef PRODUCT 6084 class G1VerifyCardTableCleanup: public HeapRegionClosure { 6085 G1CollectedHeap* _g1h; 6086 G1SATBCardTableModRefBS* _ct_bs; 6087 public: 6088 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, G1SATBCardTableModRefBS* ct_bs) 6089 : _g1h(g1h), _ct_bs(ct_bs) { } 6090 virtual bool doHeapRegion(HeapRegion* r) { 6091 if (r->is_survivor()) { 6092 _g1h->verify_dirty_region(r); 6093 } else { 6094 _g1h->verify_not_dirty_region(r); 6095 } 6096 return false; 6097 } 6098 }; 6099 6100 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 6101 // All of the region should be clean. 6102 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6103 MemRegion mr(hr->bottom(), hr->end()); 6104 ct_bs->verify_not_dirty_region(mr); 6105 } 6106 6107 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 6108 // We cannot guarantee that [bottom(),end()] is dirty. Threads 6109 // dirty allocated blocks as they allocate them. The thread that 6110 // retires each region and replaces it with a new one will do a 6111 // maximal allocation to fill in [pre_dummy_top(),end()] but will 6112 // not dirty that area (one less thing to have to do while holding 6113 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 6114 // is dirty. 6115 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6116 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 6117 if (hr->is_young()) { 6118 ct_bs->verify_g1_young_region(mr); 6119 } else { 6120 ct_bs->verify_dirty_region(mr); 6121 } 6122 } 6123 6124 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 6125 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6126 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 6127 verify_dirty_region(hr); 6128 } 6129 } 6130 6131 void G1CollectedHeap::verify_dirty_young_regions() { 6132 verify_dirty_young_list(_young_list->first_region()); 6133 } 6134 #endif 6135 6136 void G1CollectedHeap::cleanUpCardTable() { 6137 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6138 double start = os::elapsedTime(); 6139 6140 { 6141 // Iterate over the dirty cards region list. 6142 G1ParCleanupCTTask cleanup_task(ct_bs, this); 6143 6144 if (G1CollectedHeap::use_parallel_gc_threads()) { 6145 set_par_threads(); 6146 workers()->run_task(&cleanup_task); 6147 set_par_threads(0); 6148 } else { 6149 while (_dirty_cards_region_list) { 6150 HeapRegion* r = _dirty_cards_region_list; 6151 cleanup_task.clear_cards(r); 6152 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 6153 if (_dirty_cards_region_list == r) { 6154 // The last region. 6155 _dirty_cards_region_list = NULL; 6156 } 6157 r->set_next_dirty_cards_region(NULL); 6158 } 6159 } 6160 #ifndef PRODUCT 6161 if (G1VerifyCTCleanup || VerifyAfterGC) { 6162 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 6163 heap_region_iterate(&cleanup_verifier); 6164 } 6165 #endif 6166 } 6167 6168 double elapsed = os::elapsedTime() - start; 6169 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 6170 } 6171 6172 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) { 6173 size_t pre_used = 0; 6174 FreeRegionList local_free_list("Local List for CSet Freeing"); 6175 6176 double young_time_ms = 0.0; 6177 double non_young_time_ms = 0.0; 6178 6179 // Since the collection set is a superset of the the young list, 6180 // all we need to do to clear the young list is clear its 6181 // head and length, and unlink any young regions in the code below 6182 _young_list->clear(); 6183 6184 G1CollectorPolicy* policy = g1_policy(); 6185 6186 double start_sec = os::elapsedTime(); 6187 bool non_young = true; 6188 6189 HeapRegion* cur = cs_head; 6190 int age_bound = -1; 6191 size_t rs_lengths = 0; 6192 6193 while (cur != NULL) { 6194 assert(!is_on_master_free_list(cur), "sanity"); 6195 if (non_young) { 6196 if (cur->is_young()) { 6197 double end_sec = os::elapsedTime(); 6198 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6199 non_young_time_ms += elapsed_ms; 6200 6201 start_sec = os::elapsedTime(); 6202 non_young = false; 6203 } 6204 } else { 6205 if (!cur->is_young()) { 6206 double end_sec = os::elapsedTime(); 6207 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6208 young_time_ms += elapsed_ms; 6209 6210 start_sec = os::elapsedTime(); 6211 non_young = true; 6212 } 6213 } 6214 6215 rs_lengths += cur->rem_set()->occupied_locked(); 6216 6217 HeapRegion* next = cur->next_in_collection_set(); 6218 assert(cur->in_collection_set(), "bad CS"); 6219 cur->set_next_in_collection_set(NULL); 6220 cur->set_in_collection_set(false); 6221 6222 if (cur->is_young()) { 6223 int index = cur->young_index_in_cset(); 6224 assert(index != -1, "invariant"); 6225 assert((uint) index < policy->young_cset_region_length(), "invariant"); 6226 size_t words_survived = _surviving_young_words[index]; 6227 cur->record_surv_words_in_group(words_survived); 6228 6229 // At this point the we have 'popped' cur from the collection set 6230 // (linked via next_in_collection_set()) but it is still in the 6231 // young list (linked via next_young_region()). Clear the 6232 // _next_young_region field. 6233 cur->set_next_young_region(NULL); 6234 } else { 6235 int index = cur->young_index_in_cset(); 6236 assert(index == -1, "invariant"); 6237 } 6238 6239 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 6240 (!cur->is_young() && cur->young_index_in_cset() == -1), 6241 "invariant" ); 6242 6243 if (!cur->evacuation_failed()) { 6244 MemRegion used_mr = cur->used_region(); 6245 6246 // And the region is empty. 6247 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 6248 pre_used += cur->used(); 6249 free_region(cur, &local_free_list, false /* par */, true /* locked */); 6250 } else { 6251 cur->uninstall_surv_rate_group(); 6252 if (cur->is_young()) { 6253 cur->set_young_index_in_cset(-1); 6254 } 6255 cur->set_not_young(); 6256 cur->set_evacuation_failed(false); 6257 // The region is now considered to be old. 6258 _old_set.add(cur); 6259 evacuation_info.increment_collectionset_used_after(cur->used()); 6260 } 6261 cur = next; 6262 } 6263 6264 evacuation_info.set_regions_freed(local_free_list.length()); 6265 policy->record_max_rs_lengths(rs_lengths); 6266 policy->cset_regions_freed(); 6267 6268 double end_sec = os::elapsedTime(); 6269 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6270 6271 if (non_young) { 6272 non_young_time_ms += elapsed_ms; 6273 } else { 6274 young_time_ms += elapsed_ms; 6275 } 6276 6277 prepend_to_freelist(&local_free_list); 6278 decrement_summary_bytes(pre_used); 6279 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6280 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6281 } 6282 6283 // This routine is similar to the above but does not record 6284 // any policy statistics or update free lists; we are abandoning 6285 // the current incremental collection set in preparation of a 6286 // full collection. After the full GC we will start to build up 6287 // the incremental collection set again. 6288 // This is only called when we're doing a full collection 6289 // and is immediately followed by the tearing down of the young list. 6290 6291 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6292 HeapRegion* cur = cs_head; 6293 6294 while (cur != NULL) { 6295 HeapRegion* next = cur->next_in_collection_set(); 6296 assert(cur->in_collection_set(), "bad CS"); 6297 cur->set_next_in_collection_set(NULL); 6298 cur->set_in_collection_set(false); 6299 cur->set_young_index_in_cset(-1); 6300 cur = next; 6301 } 6302 } 6303 6304 void G1CollectedHeap::set_free_regions_coming() { 6305 if (G1ConcRegionFreeingVerbose) { 6306 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6307 "setting free regions coming"); 6308 } 6309 6310 assert(!free_regions_coming(), "pre-condition"); 6311 _free_regions_coming = true; 6312 } 6313 6314 void G1CollectedHeap::reset_free_regions_coming() { 6315 assert(free_regions_coming(), "pre-condition"); 6316 6317 { 6318 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6319 _free_regions_coming = false; 6320 SecondaryFreeList_lock->notify_all(); 6321 } 6322 6323 if (G1ConcRegionFreeingVerbose) { 6324 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6325 "reset free regions coming"); 6326 } 6327 } 6328 6329 void G1CollectedHeap::wait_while_free_regions_coming() { 6330 // Most of the time we won't have to wait, so let's do a quick test 6331 // first before we take the lock. 6332 if (!free_regions_coming()) { 6333 return; 6334 } 6335 6336 if (G1ConcRegionFreeingVerbose) { 6337 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6338 "waiting for free regions"); 6339 } 6340 6341 { 6342 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6343 while (free_regions_coming()) { 6344 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6345 } 6346 } 6347 6348 if (G1ConcRegionFreeingVerbose) { 6349 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6350 "done waiting for free regions"); 6351 } 6352 } 6353 6354 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6355 assert(heap_lock_held_for_gc(), 6356 "the heap lock should already be held by or for this thread"); 6357 _young_list->push_region(hr); 6358 } 6359 6360 class NoYoungRegionsClosure: public HeapRegionClosure { 6361 private: 6362 bool _success; 6363 public: 6364 NoYoungRegionsClosure() : _success(true) { } 6365 bool doHeapRegion(HeapRegion* r) { 6366 if (r->is_young()) { 6367 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6368 r->bottom(), r->end()); 6369 _success = false; 6370 } 6371 return false; 6372 } 6373 bool success() { return _success; } 6374 }; 6375 6376 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6377 bool ret = _young_list->check_list_empty(check_sample); 6378 6379 if (check_heap) { 6380 NoYoungRegionsClosure closure; 6381 heap_region_iterate(&closure); 6382 ret = ret && closure.success(); 6383 } 6384 6385 return ret; 6386 } 6387 6388 class TearDownRegionSetsClosure : public HeapRegionClosure { 6389 private: 6390 HeapRegionSet *_old_set; 6391 6392 public: 6393 TearDownRegionSetsClosure(HeapRegionSet* old_set) : _old_set(old_set) { } 6394 6395 bool doHeapRegion(HeapRegion* r) { 6396 if (r->is_empty()) { 6397 // We ignore empty regions, we'll empty the free list afterwards 6398 } else if (r->is_young()) { 6399 // We ignore young regions, we'll empty the young list afterwards 6400 } else if (r->isHumongous()) { 6401 // We ignore humongous regions, we're not tearing down the 6402 // humongous region set 6403 } else { 6404 // The rest should be old 6405 _old_set->remove(r); 6406 } 6407 return false; 6408 } 6409 6410 ~TearDownRegionSetsClosure() { 6411 assert(_old_set->is_empty(), "post-condition"); 6412 } 6413 }; 6414 6415 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6416 assert_at_safepoint(true /* should_be_vm_thread */); 6417 6418 if (!free_list_only) { 6419 TearDownRegionSetsClosure cl(&_old_set); 6420 heap_region_iterate(&cl); 6421 6422 // Note that emptying the _young_list is postponed and instead done as 6423 // the first step when rebuilding the regions sets again. The reason for 6424 // this is that during a full GC string deduplication needs to know if 6425 // a collected region was young or old when the full GC was initiated. 6426 } 6427 _free_list.remove_all(); 6428 } 6429 6430 class RebuildRegionSetsClosure : public HeapRegionClosure { 6431 private: 6432 bool _free_list_only; 6433 HeapRegionSet* _old_set; 6434 FreeRegionList* _free_list; 6435 size_t _total_used; 6436 6437 public: 6438 RebuildRegionSetsClosure(bool free_list_only, 6439 HeapRegionSet* old_set, FreeRegionList* free_list) : 6440 _free_list_only(free_list_only), 6441 _old_set(old_set), _free_list(free_list), _total_used(0) { 6442 assert(_free_list->is_empty(), "pre-condition"); 6443 if (!free_list_only) { 6444 assert(_old_set->is_empty(), "pre-condition"); 6445 } 6446 } 6447 6448 bool doHeapRegion(HeapRegion* r) { 6449 if (r->continuesHumongous()) { 6450 return false; 6451 } 6452 6453 if (r->is_empty()) { 6454 // Add free regions to the free list 6455 _free_list->add_as_tail(r); 6456 } else if (!_free_list_only) { 6457 assert(!r->is_young(), "we should not come across young regions"); 6458 6459 if (r->isHumongous()) { 6460 // We ignore humongous regions, we left the humongous set unchanged 6461 } else { 6462 // The rest should be old, add them to the old set 6463 _old_set->add(r); 6464 } 6465 _total_used += r->used(); 6466 } 6467 6468 return false; 6469 } 6470 6471 size_t total_used() { 6472 return _total_used; 6473 } 6474 }; 6475 6476 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6477 assert_at_safepoint(true /* should_be_vm_thread */); 6478 6479 if (!free_list_only) { 6480 _young_list->empty_list(); 6481 } 6482 6483 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6484 heap_region_iterate(&cl); 6485 6486 if (!free_list_only) { 6487 _summary_bytes_used = cl.total_used(); 6488 } 6489 assert(_summary_bytes_used == recalculate_used(), 6490 err_msg("inconsistent _summary_bytes_used, " 6491 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6492 _summary_bytes_used, recalculate_used())); 6493 } 6494 6495 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6496 _refine_cte_cl->set_concurrent(concurrent); 6497 } 6498 6499 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6500 HeapRegion* hr = heap_region_containing(p); 6501 if (hr == NULL) { 6502 return false; 6503 } else { 6504 return hr->is_in(p); 6505 } 6506 } 6507 6508 // Methods for the mutator alloc region 6509 6510 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6511 bool force) { 6512 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6513 assert(!force || g1_policy()->can_expand_young_list(), 6514 "if force is true we should be able to expand the young list"); 6515 bool young_list_full = g1_policy()->is_young_list_full(); 6516 if (force || !young_list_full) { 6517 HeapRegion* new_alloc_region = new_region(word_size, 6518 false /* is_old */, 6519 false /* do_expand */); 6520 if (new_alloc_region != NULL) { 6521 set_region_short_lived_locked(new_alloc_region); 6522 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6523 return new_alloc_region; 6524 } 6525 } 6526 return NULL; 6527 } 6528 6529 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6530 size_t allocated_bytes) { 6531 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6532 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6533 6534 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6535 _summary_bytes_used += allocated_bytes; 6536 _hr_printer.retire(alloc_region); 6537 // We update the eden sizes here, when the region is retired, 6538 // instead of when it's allocated, since this is the point that its 6539 // used space has been recored in _summary_bytes_used. 6540 g1mm()->update_eden_size(); 6541 } 6542 6543 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6544 bool force) { 6545 return _g1h->new_mutator_alloc_region(word_size, force); 6546 } 6547 6548 void G1CollectedHeap::set_par_threads() { 6549 // Don't change the number of workers. Use the value previously set 6550 // in the workgroup. 6551 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6552 uint n_workers = workers()->active_workers(); 6553 assert(UseDynamicNumberOfGCThreads || 6554 n_workers == workers()->total_workers(), 6555 "Otherwise should be using the total number of workers"); 6556 if (n_workers == 0) { 6557 assert(false, "Should have been set in prior evacuation pause."); 6558 n_workers = ParallelGCThreads; 6559 workers()->set_active_workers(n_workers); 6560 } 6561 set_par_threads(n_workers); 6562 } 6563 6564 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6565 size_t allocated_bytes) { 6566 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6567 } 6568 6569 // Methods for the GC alloc regions 6570 6571 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6572 uint count, 6573 GCAllocPurpose ap) { 6574 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6575 6576 if (count < g1_policy()->max_regions(ap)) { 6577 bool survivor = (ap == GCAllocForSurvived); 6578 HeapRegion* new_alloc_region = new_region(word_size, 6579 !survivor, 6580 true /* do_expand */); 6581 if (new_alloc_region != NULL) { 6582 // We really only need to do this for old regions given that we 6583 // should never scan survivors. But it doesn't hurt to do it 6584 // for survivors too. 6585 new_alloc_region->set_saved_mark(); 6586 if (survivor) { 6587 new_alloc_region->set_survivor(); 6588 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6589 } else { 6590 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6591 } 6592 bool during_im = g1_policy()->during_initial_mark_pause(); 6593 new_alloc_region->note_start_of_copying(during_im); 6594 return new_alloc_region; 6595 } else { 6596 g1_policy()->note_alloc_region_limit_reached(ap); 6597 } 6598 } 6599 return NULL; 6600 } 6601 6602 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6603 size_t allocated_bytes, 6604 GCAllocPurpose ap) { 6605 bool during_im = g1_policy()->during_initial_mark_pause(); 6606 alloc_region->note_end_of_copying(during_im); 6607 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6608 if (ap == GCAllocForSurvived) { 6609 young_list()->add_survivor_region(alloc_region); 6610 } else { 6611 _old_set.add(alloc_region); 6612 } 6613 _hr_printer.retire(alloc_region); 6614 } 6615 6616 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6617 bool force) { 6618 assert(!force, "not supported for GC alloc regions"); 6619 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6620 } 6621 6622 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6623 size_t allocated_bytes) { 6624 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6625 GCAllocForSurvived); 6626 } 6627 6628 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6629 bool force) { 6630 assert(!force, "not supported for GC alloc regions"); 6631 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6632 } 6633 6634 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6635 size_t allocated_bytes) { 6636 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6637 GCAllocForTenured); 6638 } 6639 // Heap region set verification 6640 6641 class VerifyRegionListsClosure : public HeapRegionClosure { 6642 private: 6643 HeapRegionSet* _old_set; 6644 HeapRegionSet* _humongous_set; 6645 FreeRegionList* _free_list; 6646 6647 public: 6648 HeapRegionSetCount _old_count; 6649 HeapRegionSetCount _humongous_count; 6650 HeapRegionSetCount _free_count; 6651 6652 VerifyRegionListsClosure(HeapRegionSet* old_set, 6653 HeapRegionSet* humongous_set, 6654 FreeRegionList* free_list) : 6655 _old_set(old_set), _humongous_set(humongous_set), _free_list(free_list), 6656 _old_count(), _humongous_count(), _free_count(){ } 6657 6658 bool doHeapRegion(HeapRegion* hr) { 6659 if (hr->continuesHumongous()) { 6660 return false; 6661 } 6662 6663 if (hr->is_young()) { 6664 // TODO 6665 } else if (hr->startsHumongous()) { 6666 assert(hr->containing_set() == _humongous_set, err_msg("Heap region %u is starts humongous but not in humongous set.", hr->region_num())); 6667 _humongous_count.increment(1u, hr->capacity()); 6668 } else if (hr->is_empty()) { 6669 assert(hr->containing_set() == _free_list, err_msg("Heap region %u is empty but not on the free list.", hr->region_num())); 6670 _free_count.increment(1u, hr->capacity()); 6671 } else { 6672 assert(hr->containing_set() == _old_set, err_msg("Heap region %u is old but not in the old set.", hr->region_num())); 6673 _old_count.increment(1u, hr->capacity()); 6674 } 6675 return false; 6676 } 6677 6678 void verify_counts(HeapRegionSet* old_set, HeapRegionSet* humongous_set, FreeRegionList* free_list) { 6679 guarantee(old_set->length() == _old_count.length(), err_msg("Old set count mismatch. Expected %u, actual %u.", old_set->length(), _old_count.length())); 6680 guarantee(old_set->total_capacity_bytes() == _old_count.capacity(), err_msg("Old set capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT, 6681 old_set->total_capacity_bytes(), _old_count.capacity())); 6682 6683 guarantee(humongous_set->length() == _humongous_count.length(), err_msg("Hum set count mismatch. Expected %u, actual %u.", humongous_set->length(), _humongous_count.length())); 6684 guarantee(humongous_set->total_capacity_bytes() == _humongous_count.capacity(), err_msg("Hum set capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT, 6685 humongous_set->total_capacity_bytes(), _humongous_count.capacity())); 6686 6687 guarantee(free_list->length() == _free_count.length(), err_msg("Free list count mismatch. Expected %u, actual %u.", free_list->length(), _free_count.length())); 6688 guarantee(free_list->total_capacity_bytes() == _free_count.capacity(), err_msg("Free list capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT, 6689 free_list->total_capacity_bytes(), _free_count.capacity())); 6690 } 6691 }; 6692 6693 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6694 HeapWord* bottom) { 6695 HeapWord* end = bottom + HeapRegion::GrainWords; 6696 MemRegion mr(bottom, end); 6697 assert(_g1_reserved.contains(mr), "invariant"); 6698 // This might return NULL if the allocation fails 6699 return new HeapRegion(hrs_index, _bot_shared, mr); 6700 } 6701 6702 void G1CollectedHeap::verify_region_sets() { 6703 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6704 6705 // First, check the explicit lists. 6706 _free_list.verify_list(); 6707 { 6708 // Given that a concurrent operation might be adding regions to 6709 // the secondary free list we have to take the lock before 6710 // verifying it. 6711 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6712 _secondary_free_list.verify_list(); 6713 } 6714 6715 // If a concurrent region freeing operation is in progress it will 6716 // be difficult to correctly attributed any free regions we come 6717 // across to the correct free list given that they might belong to 6718 // one of several (free_list, secondary_free_list, any local lists, 6719 // etc.). So, if that's the case we will skip the rest of the 6720 // verification operation. Alternatively, waiting for the concurrent 6721 // operation to complete will have a non-trivial effect on the GC's 6722 // operation (no concurrent operation will last longer than the 6723 // interval between two calls to verification) and it might hide 6724 // any issues that we would like to catch during testing. 6725 if (free_regions_coming()) { 6726 return; 6727 } 6728 6729 // Make sure we append the secondary_free_list on the free_list so 6730 // that all free regions we will come across can be safely 6731 // attributed to the free_list. 6732 append_secondary_free_list_if_not_empty_with_lock(); 6733 6734 // Finally, make sure that the region accounting in the lists is 6735 // consistent with what we see in the heap. 6736 6737 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6738 heap_region_iterate(&cl); 6739 cl.verify_counts(&_old_set, &_humongous_set, &_free_list); 6740 } 6741 6742 // Optimized nmethod scanning 6743 6744 class RegisterNMethodOopClosure: public OopClosure { 6745 G1CollectedHeap* _g1h; 6746 nmethod* _nm; 6747 6748 template <class T> void do_oop_work(T* p) { 6749 T heap_oop = oopDesc::load_heap_oop(p); 6750 if (!oopDesc::is_null(heap_oop)) { 6751 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6752 HeapRegion* hr = _g1h->heap_region_containing(obj); 6753 assert(!hr->continuesHumongous(), 6754 err_msg("trying to add code root "PTR_FORMAT" in continuation of humongous region "HR_FORMAT 6755 " starting at "HR_FORMAT, 6756 _nm, HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region()))); 6757 6758 // HeapRegion::add_strong_code_root() avoids adding duplicate 6759 // entries but having duplicates is OK since we "mark" nmethods 6760 // as visited when we scan the strong code root lists during the GC. 6761 hr->add_strong_code_root(_nm); 6762 assert(hr->rem_set()->strong_code_roots_list_contains(_nm), 6763 err_msg("failed to add code root "PTR_FORMAT" to remembered set of region "HR_FORMAT, 6764 _nm, HR_FORMAT_PARAMS(hr))); 6765 } 6766 } 6767 6768 public: 6769 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6770 _g1h(g1h), _nm(nm) {} 6771 6772 void do_oop(oop* p) { do_oop_work(p); } 6773 void do_oop(narrowOop* p) { do_oop_work(p); } 6774 }; 6775 6776 class UnregisterNMethodOopClosure: public OopClosure { 6777 G1CollectedHeap* _g1h; 6778 nmethod* _nm; 6779 6780 template <class T> void do_oop_work(T* p) { 6781 T heap_oop = oopDesc::load_heap_oop(p); 6782 if (!oopDesc::is_null(heap_oop)) { 6783 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6784 HeapRegion* hr = _g1h->heap_region_containing(obj); 6785 assert(!hr->continuesHumongous(), 6786 err_msg("trying to remove code root "PTR_FORMAT" in continuation of humongous region "HR_FORMAT 6787 " starting at "HR_FORMAT, 6788 _nm, HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region()))); 6789 6790 hr->remove_strong_code_root(_nm); 6791 assert(!hr->rem_set()->strong_code_roots_list_contains(_nm), 6792 err_msg("failed to remove code root "PTR_FORMAT" of region "HR_FORMAT, 6793 _nm, HR_FORMAT_PARAMS(hr))); 6794 } 6795 } 6796 6797 public: 6798 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6799 _g1h(g1h), _nm(nm) {} 6800 6801 void do_oop(oop* p) { do_oop_work(p); } 6802 void do_oop(narrowOop* p) { do_oop_work(p); } 6803 }; 6804 6805 void G1CollectedHeap::register_nmethod(nmethod* nm) { 6806 CollectedHeap::register_nmethod(nm); 6807 6808 guarantee(nm != NULL, "sanity"); 6809 RegisterNMethodOopClosure reg_cl(this, nm); 6810 nm->oops_do(®_cl); 6811 } 6812 6813 void G1CollectedHeap::unregister_nmethod(nmethod* nm) { 6814 CollectedHeap::unregister_nmethod(nm); 6815 6816 guarantee(nm != NULL, "sanity"); 6817 UnregisterNMethodOopClosure reg_cl(this, nm); 6818 nm->oops_do(®_cl, true); 6819 } 6820 6821 class MigrateCodeRootsHeapRegionClosure: public HeapRegionClosure { 6822 public: 6823 bool doHeapRegion(HeapRegion *hr) { 6824 assert(!hr->isHumongous(), 6825 err_msg("humongous region "HR_FORMAT" should not have been added to collection set", 6826 HR_FORMAT_PARAMS(hr))); 6827 hr->migrate_strong_code_roots(); 6828 return false; 6829 } 6830 }; 6831 6832 void G1CollectedHeap::migrate_strong_code_roots() { 6833 MigrateCodeRootsHeapRegionClosure cl; 6834 double migrate_start = os::elapsedTime(); 6835 collection_set_iterate(&cl); 6836 double migration_time_ms = (os::elapsedTime() - migrate_start) * 1000.0; 6837 g1_policy()->phase_times()->record_strong_code_root_migration_time(migration_time_ms); 6838 } 6839 6840 void G1CollectedHeap::purge_code_root_memory() { 6841 double purge_start = os::elapsedTime(); 6842 G1CodeRootSet::purge_chunks(G1CodeRootsChunkCacheKeepPercent); 6843 double purge_time_ms = (os::elapsedTime() - purge_start) * 1000.0; 6844 g1_policy()->phase_times()->record_strong_code_root_purge_time(purge_time_ms); 6845 } 6846 6847 // Mark all the code roots that point into regions *not* in the 6848 // collection set. 6849 // 6850 // Note we do not want to use a "marking" CodeBlobToOopClosure while 6851 // walking the the code roots lists of regions not in the collection 6852 // set. Suppose we have an nmethod (M) that points to objects in two 6853 // separate regions - one in the collection set (R1) and one not (R2). 6854 // Using a "marking" CodeBlobToOopClosure here would result in "marking" 6855 // nmethod M when walking the code roots for R1. When we come to scan 6856 // the code roots for R2, we would see that M is already marked and it 6857 // would be skipped and the objects in R2 that are referenced from M 6858 // would not be evacuated. 6859 6860 class MarkStrongCodeRootCodeBlobClosure: public CodeBlobClosure { 6861 6862 class MarkStrongCodeRootOopClosure: public OopClosure { 6863 ConcurrentMark* _cm; 6864 HeapRegion* _hr; 6865 uint _worker_id; 6866 6867 template <class T> void do_oop_work(T* p) { 6868 T heap_oop = oopDesc::load_heap_oop(p); 6869 if (!oopDesc::is_null(heap_oop)) { 6870 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6871 // Only mark objects in the region (which is assumed 6872 // to be not in the collection set). 6873 if (_hr->is_in(obj)) { 6874 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 6875 } 6876 } 6877 } 6878 6879 public: 6880 MarkStrongCodeRootOopClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id) : 6881 _cm(cm), _hr(hr), _worker_id(worker_id) { 6882 assert(!_hr->in_collection_set(), "sanity"); 6883 } 6884 6885 void do_oop(narrowOop* p) { do_oop_work(p); } 6886 void do_oop(oop* p) { do_oop_work(p); } 6887 }; 6888 6889 MarkStrongCodeRootOopClosure _oop_cl; 6890 6891 public: 6892 MarkStrongCodeRootCodeBlobClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id): 6893 _oop_cl(cm, hr, worker_id) {} 6894 6895 void do_code_blob(CodeBlob* cb) { 6896 nmethod* nm = (cb == NULL) ? NULL : cb->as_nmethod_or_null(); 6897 if (nm != NULL) { 6898 nm->oops_do(&_oop_cl); 6899 } 6900 } 6901 }; 6902 6903 class MarkStrongCodeRootsHRClosure: public HeapRegionClosure { 6904 G1CollectedHeap* _g1h; 6905 uint _worker_id; 6906 6907 public: 6908 MarkStrongCodeRootsHRClosure(G1CollectedHeap* g1h, uint worker_id) : 6909 _g1h(g1h), _worker_id(worker_id) {} 6910 6911 bool doHeapRegion(HeapRegion *hr) { 6912 HeapRegionRemSet* hrrs = hr->rem_set(); 6913 if (hr->continuesHumongous()) { 6914 // Code roots should never be attached to a continuation of a humongous region 6915 assert(hrrs->strong_code_roots_list_length() == 0, 6916 err_msg("code roots should never be attached to continuations of humongous region "HR_FORMAT 6917 " starting at "HR_FORMAT", but has "SIZE_FORMAT, 6918 HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region()), 6919 hrrs->strong_code_roots_list_length())); 6920 return false; 6921 } 6922 6923 if (hr->in_collection_set()) { 6924 // Don't mark code roots into regions in the collection set here. 6925 // They will be marked when we scan them. 6926 return false; 6927 } 6928 6929 MarkStrongCodeRootCodeBlobClosure cb_cl(_g1h->concurrent_mark(), hr, _worker_id); 6930 hr->strong_code_roots_do(&cb_cl); 6931 return false; 6932 } 6933 }; 6934 6935 void G1CollectedHeap::mark_strong_code_roots(uint worker_id) { 6936 MarkStrongCodeRootsHRClosure cl(this, worker_id); 6937 if (G1CollectedHeap::use_parallel_gc_threads()) { 6938 heap_region_par_iterate_chunked(&cl, 6939 worker_id, 6940 workers()->active_workers(), 6941 HeapRegion::ParMarkRootClaimValue); 6942 } else { 6943 heap_region_iterate(&cl); 6944 } 6945 } 6946 6947 class RebuildStrongCodeRootClosure: public CodeBlobClosure { 6948 G1CollectedHeap* _g1h; 6949 6950 public: 6951 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) : 6952 _g1h(g1h) {} 6953 6954 void do_code_blob(CodeBlob* cb) { 6955 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL; 6956 if (nm == NULL) { 6957 return; 6958 } 6959 6960 if (ScavengeRootsInCode) { 6961 _g1h->register_nmethod(nm); 6962 } 6963 } 6964 }; 6965 6966 void G1CollectedHeap::rebuild_strong_code_roots() { 6967 RebuildStrongCodeRootClosure blob_cl(this); 6968 CodeCache::blobs_do(&blob_cl); 6969 }