1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "code/codeCache.hpp" 27 #include "code/icBuffer.hpp" 28 #include "gc_implementation/g1/bufferingOopClosure.hpp" 29 #include "gc_implementation/g1/concurrentG1Refine.hpp" 30 #include "gc_implementation/g1/concurrentG1RefineThread.hpp" 31 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 32 #include "gc_implementation/g1/g1AllocRegion.inline.hpp" 33 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 34 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 35 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 36 #include "gc_implementation/g1/g1EvacFailure.hpp" 37 #include "gc_implementation/g1/g1GCPhaseTimes.hpp" 38 #include "gc_implementation/g1/g1Log.hpp" 39 #include "gc_implementation/g1/g1MarkSweep.hpp" 40 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 41 #include "gc_implementation/g1/g1RemSet.inline.hpp" 42 #include "gc_implementation/g1/g1YCTypes.hpp" 43 #include "gc_implementation/g1/heapRegion.inline.hpp" 44 #include "gc_implementation/g1/heapRegionRemSet.hpp" 45 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 46 #include "gc_implementation/g1/vm_operations_g1.hpp" 47 #include "gc_implementation/shared/gcHeapSummary.hpp" 48 #include "gc_implementation/shared/gcTimer.hpp" 49 #include "gc_implementation/shared/gcTrace.hpp" 50 #include "gc_implementation/shared/gcTraceTime.hpp" 51 #include "gc_implementation/shared/isGCActiveMark.hpp" 52 #include "memory/gcLocker.inline.hpp" 53 #include "memory/genOopClosures.inline.hpp" 54 #include "memory/generationSpec.hpp" 55 #include "memory/referenceProcessor.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "oops/oop.pcgc.inline.hpp" 58 #include "runtime/vmThread.hpp" 59 60 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 61 62 // turn it on so that the contents of the young list (scan-only / 63 // to-be-collected) are printed at "strategic" points before / during 64 // / after the collection --- this is useful for debugging 65 #define YOUNG_LIST_VERBOSE 0 66 // CURRENT STATUS 67 // This file is under construction. Search for "FIXME". 68 69 // INVARIANTS/NOTES 70 // 71 // All allocation activity covered by the G1CollectedHeap interface is 72 // serialized by acquiring the HeapLock. This happens in mem_allocate 73 // and allocate_new_tlab, which are the "entry" points to the 74 // allocation code from the rest of the JVM. (Note that this does not 75 // apply to TLAB allocation, which is not part of this interface: it 76 // is done by clients of this interface.) 77 78 // Notes on implementation of parallelism in different tasks. 79 // 80 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 81 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 82 // It does use run_task() which sets _n_workers in the task. 83 // G1ParTask executes g1_process_strong_roots() -> 84 // SharedHeap::process_strong_roots() which calls eventually to 85 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 86 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 87 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 88 // 89 90 // Local to this file. 91 92 class RefineCardTableEntryClosure: public CardTableEntryClosure { 93 SuspendibleThreadSet* _sts; 94 G1RemSet* _g1rs; 95 ConcurrentG1Refine* _cg1r; 96 bool _concurrent; 97 public: 98 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 99 G1RemSet* g1rs, 100 ConcurrentG1Refine* cg1r) : 101 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 102 {} 103 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 104 bool oops_into_cset = _g1rs->refine_card(card_ptr, worker_i, false); 105 // This path is executed by the concurrent refine or mutator threads, 106 // concurrently, and so we do not care if card_ptr contains references 107 // that point into the collection set. 108 assert(!oops_into_cset, "should be"); 109 110 if (_concurrent && _sts->should_yield()) { 111 // Caller will actually yield. 112 return false; 113 } 114 // Otherwise, we finished successfully; return true. 115 return true; 116 } 117 void set_concurrent(bool b) { _concurrent = b; } 118 }; 119 120 121 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 122 int _calls; 123 G1CollectedHeap* _g1h; 124 CardTableModRefBS* _ctbs; 125 int _histo[256]; 126 public: 127 ClearLoggedCardTableEntryClosure() : 128 _calls(0), _g1h(G1CollectedHeap::heap()), _ctbs(_g1h->g1_barrier_set()) 129 { 130 for (int i = 0; i < 256; i++) _histo[i] = 0; 131 } 132 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 133 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 134 _calls++; 135 unsigned char* ujb = (unsigned char*)card_ptr; 136 int ind = (int)(*ujb); 137 _histo[ind]++; 138 *card_ptr = -1; 139 } 140 return true; 141 } 142 int calls() { return _calls; } 143 void print_histo() { 144 gclog_or_tty->print_cr("Card table value histogram:"); 145 for (int i = 0; i < 256; i++) { 146 if (_histo[i] != 0) { 147 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 148 } 149 } 150 } 151 }; 152 153 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 154 int _calls; 155 G1CollectedHeap* _g1h; 156 CardTableModRefBS* _ctbs; 157 public: 158 RedirtyLoggedCardTableEntryClosure() : 159 _calls(0), _g1h(G1CollectedHeap::heap()), _ctbs(_g1h->g1_barrier_set()) {} 160 161 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 163 _calls++; 164 *card_ptr = 0; 165 } 166 return true; 167 } 168 int calls() { return _calls; } 169 }; 170 171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 172 public: 173 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 174 *card_ptr = CardTableModRefBS::dirty_card_val(); 175 return true; 176 } 177 }; 178 179 YoungList::YoungList(G1CollectedHeap* g1h) : 180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 182 guarantee(check_list_empty(false), "just making sure..."); 183 } 184 185 void YoungList::push_region(HeapRegion *hr) { 186 assert(!hr->is_young(), "should not already be young"); 187 assert(hr->get_next_young_region() == NULL, "cause it should!"); 188 189 hr->set_next_young_region(_head); 190 _head = hr; 191 192 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 193 ++_length; 194 } 195 196 void YoungList::add_survivor_region(HeapRegion* hr) { 197 assert(hr->is_survivor(), "should be flagged as survivor region"); 198 assert(hr->get_next_young_region() == NULL, "cause it should!"); 199 200 hr->set_next_young_region(_survivor_head); 201 if (_survivor_head == NULL) { 202 _survivor_tail = hr; 203 } 204 _survivor_head = hr; 205 ++_survivor_length; 206 } 207 208 void YoungList::empty_list(HeapRegion* list) { 209 while (list != NULL) { 210 HeapRegion* next = list->get_next_young_region(); 211 list->set_next_young_region(NULL); 212 list->uninstall_surv_rate_group(); 213 list->set_not_young(); 214 list = next; 215 } 216 } 217 218 void YoungList::empty_list() { 219 assert(check_list_well_formed(), "young list should be well formed"); 220 221 empty_list(_head); 222 _head = NULL; 223 _length = 0; 224 225 empty_list(_survivor_head); 226 _survivor_head = NULL; 227 _survivor_tail = NULL; 228 _survivor_length = 0; 229 230 _last_sampled_rs_lengths = 0; 231 232 assert(check_list_empty(false), "just making sure..."); 233 } 234 235 bool YoungList::check_list_well_formed() { 236 bool ret = true; 237 238 uint length = 0; 239 HeapRegion* curr = _head; 240 HeapRegion* last = NULL; 241 while (curr != NULL) { 242 if (!curr->is_young()) { 243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 244 "incorrectly tagged (y: %d, surv: %d)", 245 curr->bottom(), curr->end(), 246 curr->is_young(), curr->is_survivor()); 247 ret = false; 248 } 249 ++length; 250 last = curr; 251 curr = curr->get_next_young_region(); 252 } 253 ret = ret && (length == _length); 254 255 if (!ret) { 256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 257 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 258 length, _length); 259 } 260 261 return ret; 262 } 263 264 bool YoungList::check_list_empty(bool check_sample) { 265 bool ret = true; 266 267 if (_length != 0) { 268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 269 _length); 270 ret = false; 271 } 272 if (check_sample && _last_sampled_rs_lengths != 0) { 273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 274 ret = false; 275 } 276 if (_head != NULL) { 277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 278 ret = false; 279 } 280 if (!ret) { 281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 282 } 283 284 return ret; 285 } 286 287 void 288 YoungList::rs_length_sampling_init() { 289 _sampled_rs_lengths = 0; 290 _curr = _head; 291 } 292 293 bool 294 YoungList::rs_length_sampling_more() { 295 return _curr != NULL; 296 } 297 298 void 299 YoungList::rs_length_sampling_next() { 300 assert( _curr != NULL, "invariant" ); 301 size_t rs_length = _curr->rem_set()->occupied(); 302 303 _sampled_rs_lengths += rs_length; 304 305 // The current region may not yet have been added to the 306 // incremental collection set (it gets added when it is 307 // retired as the current allocation region). 308 if (_curr->in_collection_set()) { 309 // Update the collection set policy information for this region 310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 311 } 312 313 _curr = _curr->get_next_young_region(); 314 if (_curr == NULL) { 315 _last_sampled_rs_lengths = _sampled_rs_lengths; 316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 317 } 318 } 319 320 void 321 YoungList::reset_auxilary_lists() { 322 guarantee( is_empty(), "young list should be empty" ); 323 assert(check_list_well_formed(), "young list should be well formed"); 324 325 // Add survivor regions to SurvRateGroup. 326 _g1h->g1_policy()->note_start_adding_survivor_regions(); 327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 328 329 int young_index_in_cset = 0; 330 for (HeapRegion* curr = _survivor_head; 331 curr != NULL; 332 curr = curr->get_next_young_region()) { 333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 334 335 // The region is a non-empty survivor so let's add it to 336 // the incremental collection set for the next evacuation 337 // pause. 338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 339 young_index_in_cset += 1; 340 } 341 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 342 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 343 344 _head = _survivor_head; 345 _length = _survivor_length; 346 if (_survivor_head != NULL) { 347 assert(_survivor_tail != NULL, "cause it shouldn't be"); 348 assert(_survivor_length > 0, "invariant"); 349 _survivor_tail->set_next_young_region(NULL); 350 } 351 352 // Don't clear the survivor list handles until the start of 353 // the next evacuation pause - we need it in order to re-tag 354 // the survivor regions from this evacuation pause as 'young' 355 // at the start of the next. 356 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 358 359 assert(check_list_well_formed(), "young list should be well formed"); 360 } 361 362 void YoungList::print() { 363 HeapRegion* lists[] = {_head, _survivor_head}; 364 const char* names[] = {"YOUNG", "SURVIVOR"}; 365 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 368 HeapRegion *curr = lists[list]; 369 if (curr == NULL) 370 gclog_or_tty->print_cr(" empty"); 371 while (curr != NULL) { 372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 373 HR_FORMAT_PARAMS(curr), 374 curr->prev_top_at_mark_start(), 375 curr->next_top_at_mark_start(), 376 curr->age_in_surv_rate_group_cond()); 377 curr = curr->get_next_young_region(); 378 } 379 } 380 381 gclog_or_tty->print_cr(""); 382 } 383 384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 385 { 386 // Claim the right to put the region on the dirty cards region list 387 // by installing a self pointer. 388 HeapRegion* next = hr->get_next_dirty_cards_region(); 389 if (next == NULL) { 390 HeapRegion* res = (HeapRegion*) 391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 392 NULL); 393 if (res == NULL) { 394 HeapRegion* head; 395 do { 396 // Put the region to the dirty cards region list. 397 head = _dirty_cards_region_list; 398 next = (HeapRegion*) 399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 400 if (next == head) { 401 assert(hr->get_next_dirty_cards_region() == hr, 402 "hr->get_next_dirty_cards_region() != hr"); 403 if (next == NULL) { 404 // The last region in the list points to itself. 405 hr->set_next_dirty_cards_region(hr); 406 } else { 407 hr->set_next_dirty_cards_region(next); 408 } 409 } 410 } while (next != head); 411 } 412 } 413 } 414 415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 416 { 417 HeapRegion* head; 418 HeapRegion* hr; 419 do { 420 head = _dirty_cards_region_list; 421 if (head == NULL) { 422 return NULL; 423 } 424 HeapRegion* new_head = head->get_next_dirty_cards_region(); 425 if (head == new_head) { 426 // The last region. 427 new_head = NULL; 428 } 429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 430 head); 431 } while (hr != head); 432 assert(hr != NULL, "invariant"); 433 hr->set_next_dirty_cards_region(NULL); 434 return hr; 435 } 436 437 void G1CollectedHeap::stop_conc_gc_threads() { 438 _cg1r->stop(); 439 _cmThread->stop(); 440 } 441 442 #ifdef ASSERT 443 // A region is added to the collection set as it is retired 444 // so an address p can point to a region which will be in the 445 // collection set but has not yet been retired. This method 446 // therefore is only accurate during a GC pause after all 447 // regions have been retired. It is used for debugging 448 // to check if an nmethod has references to objects that can 449 // be move during a partial collection. Though it can be 450 // inaccurate, it is sufficient for G1 because the conservative 451 // implementation of is_scavengable() for G1 will indicate that 452 // all nmethods must be scanned during a partial collection. 453 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 454 HeapRegion* hr = heap_region_containing(p); 455 return hr != NULL && hr->in_collection_set(); 456 } 457 #endif 458 459 // Returns true if the reference points to an object that 460 // can move in an incremental collection. 461 bool G1CollectedHeap::is_scavengable(const void* p) { 462 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 463 G1CollectorPolicy* g1p = g1h->g1_policy(); 464 HeapRegion* hr = heap_region_containing(p); 465 if (hr == NULL) { 466 // null 467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p)); 468 return false; 469 } else { 470 return !hr->isHumongous(); 471 } 472 } 473 474 void G1CollectedHeap::check_ct_logs_at_safepoint() { 475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 476 CardTableModRefBS* ct_bs = g1_barrier_set(); 477 478 // Count the dirty cards at the start. 479 CountNonCleanMemRegionClosure count1(this); 480 ct_bs->mod_card_iterate(&count1); 481 int orig_count = count1.n(); 482 483 // First clear the logged cards. 484 ClearLoggedCardTableEntryClosure clear; 485 dcqs.set_closure(&clear); 486 dcqs.apply_closure_to_all_completed_buffers(); 487 dcqs.iterate_closure_all_threads(false); 488 clear.print_histo(); 489 490 // Now ensure that there's no dirty cards. 491 CountNonCleanMemRegionClosure count2(this); 492 ct_bs->mod_card_iterate(&count2); 493 if (count2.n() != 0) { 494 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 495 count2.n(), orig_count); 496 } 497 guarantee(count2.n() == 0, "Card table should be clean."); 498 499 RedirtyLoggedCardTableEntryClosure redirty; 500 JavaThread::dirty_card_queue_set().set_closure(&redirty); 501 dcqs.apply_closure_to_all_completed_buffers(); 502 dcqs.iterate_closure_all_threads(false); 503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 504 clear.calls(), orig_count); 505 guarantee(redirty.calls() == clear.calls(), 506 "Or else mechanism is broken."); 507 508 CountNonCleanMemRegionClosure count3(this); 509 ct_bs->mod_card_iterate(&count3); 510 if (count3.n() != orig_count) { 511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 512 orig_count, count3.n()); 513 guarantee(count3.n() >= orig_count, "Should have restored them all."); 514 } 515 516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 517 } 518 519 // Private class members. 520 521 G1CollectedHeap* G1CollectedHeap::_g1h; 522 523 // Private methods. 524 525 HeapRegion* 526 G1CollectedHeap::new_region_try_secondary_free_list() { 527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 528 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 529 if (!_secondary_free_list.is_empty()) { 530 if (G1ConcRegionFreeingVerbose) { 531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 532 "secondary_free_list has %u entries", 533 _secondary_free_list.length()); 534 } 535 // It looks as if there are free regions available on the 536 // secondary_free_list. Let's move them to the free_list and try 537 // again to allocate from it. 538 append_secondary_free_list(); 539 540 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 541 "empty we should have moved at least one entry to the free_list"); 542 HeapRegion* res = _free_list.remove_head(); 543 if (G1ConcRegionFreeingVerbose) { 544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 545 "allocated "HR_FORMAT" from secondary_free_list", 546 HR_FORMAT_PARAMS(res)); 547 } 548 return res; 549 } 550 551 // Wait here until we get notified either when (a) there are no 552 // more free regions coming or (b) some regions have been moved on 553 // the secondary_free_list. 554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 555 } 556 557 if (G1ConcRegionFreeingVerbose) { 558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 559 "could not allocate from secondary_free_list"); 560 } 561 return NULL; 562 } 563 564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 566 "the only time we use this to allocate a humongous region is " 567 "when we are allocating a single humongous region"); 568 569 HeapRegion* res; 570 if (G1StressConcRegionFreeing) { 571 if (!_secondary_free_list.is_empty()) { 572 if (G1ConcRegionFreeingVerbose) { 573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 574 "forced to look at the secondary_free_list"); 575 } 576 res = new_region_try_secondary_free_list(); 577 if (res != NULL) { 578 return res; 579 } 580 } 581 } 582 res = _free_list.remove_head_or_null(); 583 if (res == NULL) { 584 if (G1ConcRegionFreeingVerbose) { 585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 586 "res == NULL, trying the secondary_free_list"); 587 } 588 res = new_region_try_secondary_free_list(); 589 } 590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 591 // Currently, only attempts to allocate GC alloc regions set 592 // do_expand to true. So, we should only reach here during a 593 // safepoint. If this assumption changes we might have to 594 // reconsider the use of _expand_heap_after_alloc_failure. 595 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 596 597 ergo_verbose1(ErgoHeapSizing, 598 "attempt heap expansion", 599 ergo_format_reason("region allocation request failed") 600 ergo_format_byte("allocation request"), 601 word_size * HeapWordSize); 602 if (expand(word_size * HeapWordSize)) { 603 // Given that expand() succeeded in expanding the heap, and we 604 // always expand the heap by an amount aligned to the heap 605 // region size, the free list should in theory not be empty. So 606 // it would probably be OK to use remove_head(). But the extra 607 // check for NULL is unlikely to be a performance issue here (we 608 // just expanded the heap!) so let's just be conservative and 609 // use remove_head_or_null(). 610 res = _free_list.remove_head_or_null(); 611 } else { 612 _expand_heap_after_alloc_failure = false; 613 } 614 } 615 return res; 616 } 617 618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 619 size_t word_size) { 620 assert(isHumongous(word_size), "word_size should be humongous"); 621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 622 623 uint first = G1_NULL_HRS_INDEX; 624 if (num_regions == 1) { 625 // Only one region to allocate, no need to go through the slower 626 // path. The caller will attempt the expansion if this fails, so 627 // let's not try to expand here too. 628 HeapRegion* hr = new_region(word_size, false /* do_expand */); 629 if (hr != NULL) { 630 first = hr->hrs_index(); 631 } else { 632 first = G1_NULL_HRS_INDEX; 633 } 634 } else { 635 // We can't allocate humongous regions while cleanupComplete() is 636 // running, since some of the regions we find to be empty might not 637 // yet be added to the free list and it is not straightforward to 638 // know which list they are on so that we can remove them. Note 639 // that we only need to do this if we need to allocate more than 640 // one region to satisfy the current humongous allocation 641 // request. If we are only allocating one region we use the common 642 // region allocation code (see above). 643 wait_while_free_regions_coming(); 644 append_secondary_free_list_if_not_empty_with_lock(); 645 646 if (free_regions() >= num_regions) { 647 first = _hrs.find_contiguous(num_regions); 648 if (first != G1_NULL_HRS_INDEX) { 649 for (uint i = first; i < first + num_regions; ++i) { 650 HeapRegion* hr = region_at(i); 651 assert(hr->is_empty(), "sanity"); 652 assert(is_on_master_free_list(hr), "sanity"); 653 hr->set_pending_removal(true); 654 } 655 _free_list.remove_all_pending(num_regions); 656 } 657 } 658 } 659 return first; 660 } 661 662 HeapWord* 663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 664 uint num_regions, 665 size_t word_size) { 666 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 667 assert(isHumongous(word_size), "word_size should be humongous"); 668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 669 670 // Index of last region in the series + 1. 671 uint last = first + num_regions; 672 673 // We need to initialize the region(s) we just discovered. This is 674 // a bit tricky given that it can happen concurrently with 675 // refinement threads refining cards on these regions and 676 // potentially wanting to refine the BOT as they are scanning 677 // those cards (this can happen shortly after a cleanup; see CR 678 // 6991377). So we have to set up the region(s) carefully and in 679 // a specific order. 680 681 // The word size sum of all the regions we will allocate. 682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 683 assert(word_size <= word_size_sum, "sanity"); 684 685 // This will be the "starts humongous" region. 686 HeapRegion* first_hr = region_at(first); 687 // The header of the new object will be placed at the bottom of 688 // the first region. 689 HeapWord* new_obj = first_hr->bottom(); 690 // This will be the new end of the first region in the series that 691 // should also match the end of the last region in the series. 692 HeapWord* new_end = new_obj + word_size_sum; 693 // This will be the new top of the first region that will reflect 694 // this allocation. 695 HeapWord* new_top = new_obj + word_size; 696 697 // First, we need to zero the header of the space that we will be 698 // allocating. When we update top further down, some refinement 699 // threads might try to scan the region. By zeroing the header we 700 // ensure that any thread that will try to scan the region will 701 // come across the zero klass word and bail out. 702 // 703 // NOTE: It would not have been correct to have used 704 // CollectedHeap::fill_with_object() and make the space look like 705 // an int array. The thread that is doing the allocation will 706 // later update the object header to a potentially different array 707 // type and, for a very short period of time, the klass and length 708 // fields will be inconsistent. This could cause a refinement 709 // thread to calculate the object size incorrectly. 710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 711 712 // We will set up the first region as "starts humongous". This 713 // will also update the BOT covering all the regions to reflect 714 // that there is a single object that starts at the bottom of the 715 // first region. 716 first_hr->set_startsHumongous(new_top, new_end); 717 718 // Then, if there are any, we will set up the "continues 719 // humongous" regions. 720 HeapRegion* hr = NULL; 721 for (uint i = first + 1; i < last; ++i) { 722 hr = region_at(i); 723 hr->set_continuesHumongous(first_hr); 724 } 725 // If we have "continues humongous" regions (hr != NULL), then the 726 // end of the last one should match new_end. 727 assert(hr == NULL || hr->end() == new_end, "sanity"); 728 729 // Up to this point no concurrent thread would have been able to 730 // do any scanning on any region in this series. All the top 731 // fields still point to bottom, so the intersection between 732 // [bottom,top] and [card_start,card_end] will be empty. Before we 733 // update the top fields, we'll do a storestore to make sure that 734 // no thread sees the update to top before the zeroing of the 735 // object header and the BOT initialization. 736 OrderAccess::storestore(); 737 738 // Now that the BOT and the object header have been initialized, 739 // we can update top of the "starts humongous" region. 740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 741 "new_top should be in this region"); 742 first_hr->set_top(new_top); 743 if (_hr_printer.is_active()) { 744 HeapWord* bottom = first_hr->bottom(); 745 HeapWord* end = first_hr->orig_end(); 746 if ((first + 1) == last) { 747 // the series has a single humongous region 748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 749 } else { 750 // the series has more than one humongous regions 751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 752 } 753 } 754 755 // Now, we will update the top fields of the "continues humongous" 756 // regions. The reason we need to do this is that, otherwise, 757 // these regions would look empty and this will confuse parts of 758 // G1. For example, the code that looks for a consecutive number 759 // of empty regions will consider them empty and try to 760 // re-allocate them. We can extend is_empty() to also include 761 // !continuesHumongous(), but it is easier to just update the top 762 // fields here. The way we set top for all regions (i.e., top == 763 // end for all regions but the last one, top == new_top for the 764 // last one) is actually used when we will free up the humongous 765 // region in free_humongous_region(). 766 hr = NULL; 767 for (uint i = first + 1; i < last; ++i) { 768 hr = region_at(i); 769 if ((i + 1) == last) { 770 // last continues humongous region 771 assert(hr->bottom() < new_top && new_top <= hr->end(), 772 "new_top should fall on this region"); 773 hr->set_top(new_top); 774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 775 } else { 776 // not last one 777 assert(new_top > hr->end(), "new_top should be above this region"); 778 hr->set_top(hr->end()); 779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 780 } 781 } 782 // If we have continues humongous regions (hr != NULL), then the 783 // end of the last one should match new_end and its top should 784 // match new_top. 785 assert(hr == NULL || 786 (hr->end() == new_end && hr->top() == new_top), "sanity"); 787 788 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 789 _summary_bytes_used += first_hr->used(); 790 _humongous_set.add(first_hr); 791 792 return new_obj; 793 } 794 795 // If could fit into free regions w/o expansion, try. 796 // Otherwise, if can expand, do so. 797 // Otherwise, if using ex regions might help, try with ex given back. 798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 800 801 verify_region_sets_optional(); 802 803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 805 uint x_num = expansion_regions(); 806 uint fs = _hrs.free_suffix(); 807 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 808 if (first == G1_NULL_HRS_INDEX) { 809 // The only thing we can do now is attempt expansion. 810 if (fs + x_num >= num_regions) { 811 // If the number of regions we're trying to allocate for this 812 // object is at most the number of regions in the free suffix, 813 // then the call to humongous_obj_allocate_find_first() above 814 // should have succeeded and we wouldn't be here. 815 // 816 // We should only be trying to expand when the free suffix is 817 // not sufficient for the object _and_ we have some expansion 818 // room available. 819 assert(num_regions > fs, "earlier allocation should have succeeded"); 820 821 ergo_verbose1(ErgoHeapSizing, 822 "attempt heap expansion", 823 ergo_format_reason("humongous allocation request failed") 824 ergo_format_byte("allocation request"), 825 word_size * HeapWordSize); 826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 827 // Even though the heap was expanded, it might not have 828 // reached the desired size. So, we cannot assume that the 829 // allocation will succeed. 830 first = humongous_obj_allocate_find_first(num_regions, word_size); 831 } 832 } 833 } 834 835 HeapWord* result = NULL; 836 if (first != G1_NULL_HRS_INDEX) { 837 result = 838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 839 assert(result != NULL, "it should always return a valid result"); 840 841 // A successful humongous object allocation changes the used space 842 // information of the old generation so we need to recalculate the 843 // sizes and update the jstat counters here. 844 g1mm()->update_sizes(); 845 } 846 847 verify_region_sets_optional(); 848 849 return result; 850 } 851 852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 853 assert_heap_not_locked_and_not_at_safepoint(); 854 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 855 856 unsigned int dummy_gc_count_before; 857 int dummy_gclocker_retry_count = 0; 858 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count); 859 } 860 861 HeapWord* 862 G1CollectedHeap::mem_allocate(size_t word_size, 863 bool* gc_overhead_limit_was_exceeded) { 864 assert_heap_not_locked_and_not_at_safepoint(); 865 866 // Loop until the allocation is satisfied, or unsatisfied after GC. 867 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) { 868 unsigned int gc_count_before; 869 870 HeapWord* result = NULL; 871 if (!isHumongous(word_size)) { 872 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count); 873 } else { 874 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count); 875 } 876 if (result != NULL) { 877 return result; 878 } 879 880 // Create the garbage collection operation... 881 VM_G1CollectForAllocation op(gc_count_before, word_size); 882 // ...and get the VM thread to execute it. 883 VMThread::execute(&op); 884 885 if (op.prologue_succeeded() && op.pause_succeeded()) { 886 // If the operation was successful we'll return the result even 887 // if it is NULL. If the allocation attempt failed immediately 888 // after a Full GC, it's unlikely we'll be able to allocate now. 889 HeapWord* result = op.result(); 890 if (result != NULL && !isHumongous(word_size)) { 891 // Allocations that take place on VM operations do not do any 892 // card dirtying and we have to do it here. We only have to do 893 // this for non-humongous allocations, though. 894 dirty_young_block(result, word_size); 895 } 896 return result; 897 } else { 898 if (gclocker_retry_count > GCLockerRetryAllocationCount) { 899 return NULL; 900 } 901 assert(op.result() == NULL, 902 "the result should be NULL if the VM op did not succeed"); 903 } 904 905 // Give a warning if we seem to be looping forever. 906 if ((QueuedAllocationWarningCount > 0) && 907 (try_count % QueuedAllocationWarningCount == 0)) { 908 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 909 } 910 } 911 912 ShouldNotReachHere(); 913 return NULL; 914 } 915 916 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 917 unsigned int *gc_count_before_ret, 918 int* gclocker_retry_count_ret) { 919 // Make sure you read the note in attempt_allocation_humongous(). 920 921 assert_heap_not_locked_and_not_at_safepoint(); 922 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 923 "be called for humongous allocation requests"); 924 925 // We should only get here after the first-level allocation attempt 926 // (attempt_allocation()) failed to allocate. 927 928 // We will loop until a) we manage to successfully perform the 929 // allocation or b) we successfully schedule a collection which 930 // fails to perform the allocation. b) is the only case when we'll 931 // return NULL. 932 HeapWord* result = NULL; 933 for (int try_count = 1; /* we'll return */; try_count += 1) { 934 bool should_try_gc; 935 unsigned int gc_count_before; 936 937 { 938 MutexLockerEx x(Heap_lock); 939 940 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 941 false /* bot_updates */); 942 if (result != NULL) { 943 return result; 944 } 945 946 // If we reach here, attempt_allocation_locked() above failed to 947 // allocate a new region. So the mutator alloc region should be NULL. 948 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 949 950 if (GC_locker::is_active_and_needs_gc()) { 951 if (g1_policy()->can_expand_young_list()) { 952 // No need for an ergo verbose message here, 953 // can_expand_young_list() does this when it returns true. 954 result = _mutator_alloc_region.attempt_allocation_force(word_size, 955 false /* bot_updates */); 956 if (result != NULL) { 957 return result; 958 } 959 } 960 should_try_gc = false; 961 } else { 962 // The GCLocker may not be active but the GCLocker initiated 963 // GC may not yet have been performed (GCLocker::needs_gc() 964 // returns true). In this case we do not try this GC and 965 // wait until the GCLocker initiated GC is performed, and 966 // then retry the allocation. 967 if (GC_locker::needs_gc()) { 968 should_try_gc = false; 969 } else { 970 // Read the GC count while still holding the Heap_lock. 971 gc_count_before = total_collections(); 972 should_try_gc = true; 973 } 974 } 975 } 976 977 if (should_try_gc) { 978 bool succeeded; 979 result = do_collection_pause(word_size, gc_count_before, &succeeded, 980 GCCause::_g1_inc_collection_pause); 981 if (result != NULL) { 982 assert(succeeded, "only way to get back a non-NULL result"); 983 return result; 984 } 985 986 if (succeeded) { 987 // If we get here we successfully scheduled a collection which 988 // failed to allocate. No point in trying to allocate 989 // further. We'll just return NULL. 990 MutexLockerEx x(Heap_lock); 991 *gc_count_before_ret = total_collections(); 992 return NULL; 993 } 994 } else { 995 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 996 MutexLockerEx x(Heap_lock); 997 *gc_count_before_ret = total_collections(); 998 return NULL; 999 } 1000 // The GCLocker is either active or the GCLocker initiated 1001 // GC has not yet been performed. Stall until it is and 1002 // then retry the allocation. 1003 GC_locker::stall_until_clear(); 1004 (*gclocker_retry_count_ret) += 1; 1005 } 1006 1007 // We can reach here if we were unsuccessful in scheduling a 1008 // collection (because another thread beat us to it) or if we were 1009 // stalled due to the GC locker. In either can we should retry the 1010 // allocation attempt in case another thread successfully 1011 // performed a collection and reclaimed enough space. We do the 1012 // first attempt (without holding the Heap_lock) here and the 1013 // follow-on attempt will be at the start of the next loop 1014 // iteration (after taking the Heap_lock). 1015 result = _mutator_alloc_region.attempt_allocation(word_size, 1016 false /* bot_updates */); 1017 if (result != NULL) { 1018 return result; 1019 } 1020 1021 // Give a warning if we seem to be looping forever. 1022 if ((QueuedAllocationWarningCount > 0) && 1023 (try_count % QueuedAllocationWarningCount == 0)) { 1024 warning("G1CollectedHeap::attempt_allocation_slow() " 1025 "retries %d times", try_count); 1026 } 1027 } 1028 1029 ShouldNotReachHere(); 1030 return NULL; 1031 } 1032 1033 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1034 unsigned int * gc_count_before_ret, 1035 int* gclocker_retry_count_ret) { 1036 // The structure of this method has a lot of similarities to 1037 // attempt_allocation_slow(). The reason these two were not merged 1038 // into a single one is that such a method would require several "if 1039 // allocation is not humongous do this, otherwise do that" 1040 // conditional paths which would obscure its flow. In fact, an early 1041 // version of this code did use a unified method which was harder to 1042 // follow and, as a result, it had subtle bugs that were hard to 1043 // track down. So keeping these two methods separate allows each to 1044 // be more readable. It will be good to keep these two in sync as 1045 // much as possible. 1046 1047 assert_heap_not_locked_and_not_at_safepoint(); 1048 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1049 "should only be called for humongous allocations"); 1050 1051 // Humongous objects can exhaust the heap quickly, so we should check if we 1052 // need to start a marking cycle at each humongous object allocation. We do 1053 // the check before we do the actual allocation. The reason for doing it 1054 // before the allocation is that we avoid having to keep track of the newly 1055 // allocated memory while we do a GC. 1056 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1057 word_size)) { 1058 collect(GCCause::_g1_humongous_allocation); 1059 } 1060 1061 // We will loop until a) we manage to successfully perform the 1062 // allocation or b) we successfully schedule a collection which 1063 // fails to perform the allocation. b) is the only case when we'll 1064 // return NULL. 1065 HeapWord* result = NULL; 1066 for (int try_count = 1; /* we'll return */; try_count += 1) { 1067 bool should_try_gc; 1068 unsigned int gc_count_before; 1069 1070 { 1071 MutexLockerEx x(Heap_lock); 1072 1073 // Given that humongous objects are not allocated in young 1074 // regions, we'll first try to do the allocation without doing a 1075 // collection hoping that there's enough space in the heap. 1076 result = humongous_obj_allocate(word_size); 1077 if (result != NULL) { 1078 return result; 1079 } 1080 1081 if (GC_locker::is_active_and_needs_gc()) { 1082 should_try_gc = false; 1083 } else { 1084 // The GCLocker may not be active but the GCLocker initiated 1085 // GC may not yet have been performed (GCLocker::needs_gc() 1086 // returns true). In this case we do not try this GC and 1087 // wait until the GCLocker initiated GC is performed, and 1088 // then retry the allocation. 1089 if (GC_locker::needs_gc()) { 1090 should_try_gc = false; 1091 } else { 1092 // Read the GC count while still holding the Heap_lock. 1093 gc_count_before = total_collections(); 1094 should_try_gc = true; 1095 } 1096 } 1097 } 1098 1099 if (should_try_gc) { 1100 // If we failed to allocate the humongous object, we should try to 1101 // do a collection pause (if we're allowed) in case it reclaims 1102 // enough space for the allocation to succeed after the pause. 1103 1104 bool succeeded; 1105 result = do_collection_pause(word_size, gc_count_before, &succeeded, 1106 GCCause::_g1_humongous_allocation); 1107 if (result != NULL) { 1108 assert(succeeded, "only way to get back a non-NULL result"); 1109 return result; 1110 } 1111 1112 if (succeeded) { 1113 // If we get here we successfully scheduled a collection which 1114 // failed to allocate. No point in trying to allocate 1115 // further. We'll just return NULL. 1116 MutexLockerEx x(Heap_lock); 1117 *gc_count_before_ret = total_collections(); 1118 return NULL; 1119 } 1120 } else { 1121 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1122 MutexLockerEx x(Heap_lock); 1123 *gc_count_before_ret = total_collections(); 1124 return NULL; 1125 } 1126 // The GCLocker is either active or the GCLocker initiated 1127 // GC has not yet been performed. Stall until it is and 1128 // then retry the allocation. 1129 GC_locker::stall_until_clear(); 1130 (*gclocker_retry_count_ret) += 1; 1131 } 1132 1133 // We can reach here if we were unsuccessful in scheduling a 1134 // collection (because another thread beat us to it) or if we were 1135 // stalled due to the GC locker. In either can we should retry the 1136 // allocation attempt in case another thread successfully 1137 // performed a collection and reclaimed enough space. Give a 1138 // warning if we seem to be looping forever. 1139 1140 if ((QueuedAllocationWarningCount > 0) && 1141 (try_count % QueuedAllocationWarningCount == 0)) { 1142 warning("G1CollectedHeap::attempt_allocation_humongous() " 1143 "retries %d times", try_count); 1144 } 1145 } 1146 1147 ShouldNotReachHere(); 1148 return NULL; 1149 } 1150 1151 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1152 bool expect_null_mutator_alloc_region) { 1153 assert_at_safepoint(true /* should_be_vm_thread */); 1154 assert(_mutator_alloc_region.get() == NULL || 1155 !expect_null_mutator_alloc_region, 1156 "the current alloc region was unexpectedly found to be non-NULL"); 1157 1158 if (!isHumongous(word_size)) { 1159 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1160 false /* bot_updates */); 1161 } else { 1162 HeapWord* result = humongous_obj_allocate(word_size); 1163 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1164 g1_policy()->set_initiate_conc_mark_if_possible(); 1165 } 1166 return result; 1167 } 1168 1169 ShouldNotReachHere(); 1170 } 1171 1172 class PostMCRemSetClearClosure: public HeapRegionClosure { 1173 G1CollectedHeap* _g1h; 1174 ModRefBarrierSet* _mr_bs; 1175 public: 1176 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1177 _g1h(g1h), _mr_bs(mr_bs) {} 1178 1179 bool doHeapRegion(HeapRegion* r) { 1180 HeapRegionRemSet* hrrs = r->rem_set(); 1181 1182 if (r->continuesHumongous()) { 1183 // We'll assert that the strong code root list and RSet is empty 1184 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 1185 assert(hrrs->occupied() == 0, "RSet should be empty"); 1186 return false; 1187 } 1188 1189 _g1h->reset_gc_time_stamps(r); 1190 hrrs->clear(); 1191 // You might think here that we could clear just the cards 1192 // corresponding to the used region. But no: if we leave a dirty card 1193 // in a region we might allocate into, then it would prevent that card 1194 // from being enqueued, and cause it to be missed. 1195 // Re: the performance cost: we shouldn't be doing full GC anyway! 1196 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1197 1198 return false; 1199 } 1200 }; 1201 1202 void G1CollectedHeap::clear_rsets_post_compaction() { 1203 PostMCRemSetClearClosure rs_clear(this, g1_barrier_set()); 1204 heap_region_iterate(&rs_clear); 1205 } 1206 1207 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1208 G1CollectedHeap* _g1h; 1209 UpdateRSOopClosure _cl; 1210 int _worker_i; 1211 public: 1212 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1213 _cl(g1->g1_rem_set(), worker_i), 1214 _worker_i(worker_i), 1215 _g1h(g1) 1216 { } 1217 1218 bool doHeapRegion(HeapRegion* r) { 1219 if (!r->continuesHumongous()) { 1220 _cl.set_from(r); 1221 r->oop_iterate(&_cl); 1222 } 1223 return false; 1224 } 1225 }; 1226 1227 class ParRebuildRSTask: public AbstractGangTask { 1228 G1CollectedHeap* _g1; 1229 public: 1230 ParRebuildRSTask(G1CollectedHeap* g1) 1231 : AbstractGangTask("ParRebuildRSTask"), 1232 _g1(g1) 1233 { } 1234 1235 void work(uint worker_id) { 1236 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1237 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1238 _g1->workers()->active_workers(), 1239 HeapRegion::RebuildRSClaimValue); 1240 } 1241 }; 1242 1243 class PostCompactionPrinterClosure: public HeapRegionClosure { 1244 private: 1245 G1HRPrinter* _hr_printer; 1246 public: 1247 bool doHeapRegion(HeapRegion* hr) { 1248 assert(!hr->is_young(), "not expecting to find young regions"); 1249 // We only generate output for non-empty regions. 1250 if (!hr->is_empty()) { 1251 if (!hr->isHumongous()) { 1252 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1253 } else if (hr->startsHumongous()) { 1254 if (hr->region_num() == 1) { 1255 // single humongous region 1256 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1257 } else { 1258 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1259 } 1260 } else { 1261 assert(hr->continuesHumongous(), "only way to get here"); 1262 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1263 } 1264 } 1265 return false; 1266 } 1267 1268 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1269 : _hr_printer(hr_printer) { } 1270 }; 1271 1272 void G1CollectedHeap::print_hrs_post_compaction() { 1273 PostCompactionPrinterClosure cl(hr_printer()); 1274 heap_region_iterate(&cl); 1275 } 1276 1277 bool G1CollectedHeap::do_collection(bool explicit_gc, 1278 bool clear_all_soft_refs, 1279 size_t word_size) { 1280 assert_at_safepoint(true /* should_be_vm_thread */); 1281 1282 if (GC_locker::check_active_before_gc()) { 1283 return false; 1284 } 1285 1286 STWGCTimer* gc_timer = G1MarkSweep::gc_timer(); 1287 gc_timer->register_gc_start(os::elapsed_counter()); 1288 1289 SerialOldTracer* gc_tracer = G1MarkSweep::gc_tracer(); 1290 gc_tracer->report_gc_start(gc_cause(), gc_timer->gc_start()); 1291 1292 SvcGCMarker sgcm(SvcGCMarker::FULL); 1293 ResourceMark rm; 1294 1295 print_heap_before_gc(); 1296 trace_heap_before_gc(gc_tracer); 1297 1298 size_t metadata_prev_used = MetaspaceAux::allocated_used_bytes(); 1299 1300 HRSPhaseSetter x(HRSPhaseFullGC); 1301 verify_region_sets_optional(); 1302 1303 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1304 collector_policy()->should_clear_all_soft_refs(); 1305 1306 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1307 1308 { 1309 IsGCActiveMark x; 1310 1311 // Timing 1312 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1313 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1314 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1315 1316 { 1317 GCTraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, NULL); 1318 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1319 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1320 1321 double start = os::elapsedTime(); 1322 g1_policy()->record_full_collection_start(); 1323 1324 // Note: When we have a more flexible GC logging framework that 1325 // allows us to add optional attributes to a GC log record we 1326 // could consider timing and reporting how long we wait in the 1327 // following two methods. 1328 wait_while_free_regions_coming(); 1329 // If we start the compaction before the CM threads finish 1330 // scanning the root regions we might trip them over as we'll 1331 // be moving objects / updating references. So let's wait until 1332 // they are done. By telling them to abort, they should complete 1333 // early. 1334 _cm->root_regions()->abort(); 1335 _cm->root_regions()->wait_until_scan_finished(); 1336 append_secondary_free_list_if_not_empty_with_lock(); 1337 1338 gc_prologue(true); 1339 increment_total_collections(true /* full gc */); 1340 increment_old_marking_cycles_started(); 1341 1342 assert(used() == recalculate_used(), "Should be equal"); 1343 1344 verify_before_gc(); 1345 1346 pre_full_gc_dump(gc_timer); 1347 1348 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1349 1350 // Disable discovery and empty the discovered lists 1351 // for the CM ref processor. 1352 ref_processor_cm()->disable_discovery(); 1353 ref_processor_cm()->abandon_partial_discovery(); 1354 ref_processor_cm()->verify_no_references_recorded(); 1355 1356 // Abandon current iterations of concurrent marking and concurrent 1357 // refinement, if any are in progress. We have to do this before 1358 // wait_until_scan_finished() below. 1359 concurrent_mark()->abort(); 1360 1361 // Make sure we'll choose a new allocation region afterwards. 1362 release_mutator_alloc_region(); 1363 abandon_gc_alloc_regions(); 1364 g1_rem_set()->cleanupHRRS(); 1365 1366 // We should call this after we retire any currently active alloc 1367 // regions so that all the ALLOC / RETIRE events are generated 1368 // before the start GC event. 1369 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1370 1371 // We may have added regions to the current incremental collection 1372 // set between the last GC or pause and now. We need to clear the 1373 // incremental collection set and then start rebuilding it afresh 1374 // after this full GC. 1375 abandon_collection_set(g1_policy()->inc_cset_head()); 1376 g1_policy()->clear_incremental_cset(); 1377 g1_policy()->stop_incremental_cset_building(); 1378 1379 tear_down_region_sets(false /* free_list_only */); 1380 g1_policy()->set_gcs_are_young(true); 1381 1382 // See the comments in g1CollectedHeap.hpp and 1383 // G1CollectedHeap::ref_processing_init() about 1384 // how reference processing currently works in G1. 1385 1386 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1387 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1388 1389 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1390 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1391 1392 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1393 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1394 1395 // Do collection work 1396 { 1397 HandleMark hm; // Discard invalid handles created during gc 1398 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1399 } 1400 1401 assert(free_regions() == 0, "we should not have added any free regions"); 1402 rebuild_region_sets(false /* free_list_only */); 1403 1404 // Enqueue any discovered reference objects that have 1405 // not been removed from the discovered lists. 1406 ref_processor_stw()->enqueue_discovered_references(); 1407 1408 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1409 1410 MemoryService::track_memory_usage(); 1411 1412 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1413 ref_processor_stw()->verify_no_references_recorded(); 1414 1415 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1416 ClassLoaderDataGraph::purge(); 1417 MetaspaceAux::verify_metrics(); 1418 1419 // Note: since we've just done a full GC, concurrent 1420 // marking is no longer active. Therefore we need not 1421 // re-enable reference discovery for the CM ref processor. 1422 // That will be done at the start of the next marking cycle. 1423 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1424 ref_processor_cm()->verify_no_references_recorded(); 1425 1426 reset_gc_time_stamp(); 1427 // Since everything potentially moved, we will clear all remembered 1428 // sets, and clear all cards. Later we will rebuild remembered 1429 // sets. We will also reset the GC time stamps of the regions. 1430 clear_rsets_post_compaction(); 1431 check_gc_time_stamps(); 1432 1433 // Resize the heap if necessary. 1434 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1435 1436 if (_hr_printer.is_active()) { 1437 // We should do this after we potentially resize the heap so 1438 // that all the COMMIT / UNCOMMIT events are generated before 1439 // the end GC event. 1440 1441 print_hrs_post_compaction(); 1442 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1443 } 1444 1445 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 1446 if (hot_card_cache->use_cache()) { 1447 hot_card_cache->reset_card_counts(); 1448 hot_card_cache->reset_hot_cache(); 1449 } 1450 1451 // Rebuild remembered sets of all regions. 1452 if (G1CollectedHeap::use_parallel_gc_threads()) { 1453 uint n_workers = 1454 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1455 workers()->active_workers(), 1456 Threads::number_of_non_daemon_threads()); 1457 assert(UseDynamicNumberOfGCThreads || 1458 n_workers == workers()->total_workers(), 1459 "If not dynamic should be using all the workers"); 1460 workers()->set_active_workers(n_workers); 1461 // Set parallel threads in the heap (_n_par_threads) only 1462 // before a parallel phase and always reset it to 0 after 1463 // the phase so that the number of parallel threads does 1464 // no get carried forward to a serial phase where there 1465 // may be code that is "possibly_parallel". 1466 set_par_threads(n_workers); 1467 1468 ParRebuildRSTask rebuild_rs_task(this); 1469 assert(check_heap_region_claim_values( 1470 HeapRegion::InitialClaimValue), "sanity check"); 1471 assert(UseDynamicNumberOfGCThreads || 1472 workers()->active_workers() == workers()->total_workers(), 1473 "Unless dynamic should use total workers"); 1474 // Use the most recent number of active workers 1475 assert(workers()->active_workers() > 0, 1476 "Active workers not properly set"); 1477 set_par_threads(workers()->active_workers()); 1478 workers()->run_task(&rebuild_rs_task); 1479 set_par_threads(0); 1480 assert(check_heap_region_claim_values( 1481 HeapRegion::RebuildRSClaimValue), "sanity check"); 1482 reset_heap_region_claim_values(); 1483 } else { 1484 RebuildRSOutOfRegionClosure rebuild_rs(this); 1485 heap_region_iterate(&rebuild_rs); 1486 } 1487 1488 // Rebuild the strong code root lists for each region 1489 rebuild_strong_code_roots(); 1490 1491 if (true) { // FIXME 1492 MetaspaceGC::compute_new_size(); 1493 } 1494 1495 #ifdef TRACESPINNING 1496 ParallelTaskTerminator::print_termination_counts(); 1497 #endif 1498 1499 // Discard all rset updates 1500 JavaThread::dirty_card_queue_set().abandon_logs(); 1501 assert(!G1DeferredRSUpdate 1502 || (G1DeferredRSUpdate && 1503 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1504 1505 _young_list->reset_sampled_info(); 1506 // At this point there should be no regions in the 1507 // entire heap tagged as young. 1508 assert(check_young_list_empty(true /* check_heap */), 1509 "young list should be empty at this point"); 1510 1511 // Update the number of full collections that have been completed. 1512 increment_old_marking_cycles_completed(false /* concurrent */); 1513 1514 _hrs.verify_optional(); 1515 verify_region_sets_optional(); 1516 1517 verify_after_gc(); 1518 1519 // Start a new incremental collection set for the next pause 1520 assert(g1_policy()->collection_set() == NULL, "must be"); 1521 g1_policy()->start_incremental_cset_building(); 1522 1523 // Clear the _cset_fast_test bitmap in anticipation of adding 1524 // regions to the incremental collection set for the next 1525 // evacuation pause. 1526 clear_cset_fast_test(); 1527 1528 init_mutator_alloc_region(); 1529 1530 double end = os::elapsedTime(); 1531 g1_policy()->record_full_collection_end(); 1532 1533 if (G1Log::fine()) { 1534 g1_policy()->print_heap_transition(); 1535 } 1536 1537 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1538 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1539 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1540 // before any GC notifications are raised. 1541 g1mm()->update_sizes(); 1542 1543 gc_epilogue(true); 1544 } 1545 1546 if (G1Log::finer()) { 1547 g1_policy()->print_detailed_heap_transition(true /* full */); 1548 } 1549 1550 print_heap_after_gc(); 1551 trace_heap_after_gc(gc_tracer); 1552 1553 post_full_gc_dump(gc_timer); 1554 1555 gc_timer->register_gc_end(os::elapsed_counter()); 1556 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1557 } 1558 1559 return true; 1560 } 1561 1562 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1563 // do_collection() will return whether it succeeded in performing 1564 // the GC. Currently, there is no facility on the 1565 // do_full_collection() API to notify the caller than the collection 1566 // did not succeed (e.g., because it was locked out by the GC 1567 // locker). So, right now, we'll ignore the return value. 1568 bool dummy = do_collection(true, /* explicit_gc */ 1569 clear_all_soft_refs, 1570 0 /* word_size */); 1571 } 1572 1573 // This code is mostly copied from TenuredGeneration. 1574 void 1575 G1CollectedHeap:: 1576 resize_if_necessary_after_full_collection(size_t word_size) { 1577 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1578 1579 // Include the current allocation, if any, and bytes that will be 1580 // pre-allocated to support collections, as "used". 1581 const size_t used_after_gc = used(); 1582 const size_t capacity_after_gc = capacity(); 1583 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1584 1585 // This is enforced in arguments.cpp. 1586 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1587 "otherwise the code below doesn't make sense"); 1588 1589 // We don't have floating point command-line arguments 1590 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1591 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1592 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1593 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1594 1595 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1596 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1597 1598 // We have to be careful here as these two calculations can overflow 1599 // 32-bit size_t's. 1600 double used_after_gc_d = (double) used_after_gc; 1601 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1602 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1603 1604 // Let's make sure that they are both under the max heap size, which 1605 // by default will make them fit into a size_t. 1606 double desired_capacity_upper_bound = (double) max_heap_size; 1607 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1608 desired_capacity_upper_bound); 1609 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1610 desired_capacity_upper_bound); 1611 1612 // We can now safely turn them into size_t's. 1613 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1614 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1615 1616 // This assert only makes sense here, before we adjust them 1617 // with respect to the min and max heap size. 1618 assert(minimum_desired_capacity <= maximum_desired_capacity, 1619 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1620 "maximum_desired_capacity = "SIZE_FORMAT, 1621 minimum_desired_capacity, maximum_desired_capacity)); 1622 1623 // Should not be greater than the heap max size. No need to adjust 1624 // it with respect to the heap min size as it's a lower bound (i.e., 1625 // we'll try to make the capacity larger than it, not smaller). 1626 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1627 // Should not be less than the heap min size. No need to adjust it 1628 // with respect to the heap max size as it's an upper bound (i.e., 1629 // we'll try to make the capacity smaller than it, not greater). 1630 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1631 1632 if (capacity_after_gc < minimum_desired_capacity) { 1633 // Don't expand unless it's significant 1634 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1635 ergo_verbose4(ErgoHeapSizing, 1636 "attempt heap expansion", 1637 ergo_format_reason("capacity lower than " 1638 "min desired capacity after Full GC") 1639 ergo_format_byte("capacity") 1640 ergo_format_byte("occupancy") 1641 ergo_format_byte_perc("min desired capacity"), 1642 capacity_after_gc, used_after_gc, 1643 minimum_desired_capacity, (double) MinHeapFreeRatio); 1644 expand(expand_bytes); 1645 1646 // No expansion, now see if we want to shrink 1647 } else if (capacity_after_gc > maximum_desired_capacity) { 1648 // Capacity too large, compute shrinking size 1649 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1650 ergo_verbose4(ErgoHeapSizing, 1651 "attempt heap shrinking", 1652 ergo_format_reason("capacity higher than " 1653 "max desired capacity after Full GC") 1654 ergo_format_byte("capacity") 1655 ergo_format_byte("occupancy") 1656 ergo_format_byte_perc("max desired capacity"), 1657 capacity_after_gc, used_after_gc, 1658 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1659 shrink(shrink_bytes); 1660 } 1661 } 1662 1663 1664 HeapWord* 1665 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1666 bool* succeeded) { 1667 assert_at_safepoint(true /* should_be_vm_thread */); 1668 1669 *succeeded = true; 1670 // Let's attempt the allocation first. 1671 HeapWord* result = 1672 attempt_allocation_at_safepoint(word_size, 1673 false /* expect_null_mutator_alloc_region */); 1674 if (result != NULL) { 1675 assert(*succeeded, "sanity"); 1676 return result; 1677 } 1678 1679 // In a G1 heap, we're supposed to keep allocation from failing by 1680 // incremental pauses. Therefore, at least for now, we'll favor 1681 // expansion over collection. (This might change in the future if we can 1682 // do something smarter than full collection to satisfy a failed alloc.) 1683 result = expand_and_allocate(word_size); 1684 if (result != NULL) { 1685 assert(*succeeded, "sanity"); 1686 return result; 1687 } 1688 1689 // Expansion didn't work, we'll try to do a Full GC. 1690 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1691 false, /* clear_all_soft_refs */ 1692 word_size); 1693 if (!gc_succeeded) { 1694 *succeeded = false; 1695 return NULL; 1696 } 1697 1698 // Retry the allocation 1699 result = attempt_allocation_at_safepoint(word_size, 1700 true /* expect_null_mutator_alloc_region */); 1701 if (result != NULL) { 1702 assert(*succeeded, "sanity"); 1703 return result; 1704 } 1705 1706 // Then, try a Full GC that will collect all soft references. 1707 gc_succeeded = do_collection(false, /* explicit_gc */ 1708 true, /* clear_all_soft_refs */ 1709 word_size); 1710 if (!gc_succeeded) { 1711 *succeeded = false; 1712 return NULL; 1713 } 1714 1715 // Retry the allocation once more 1716 result = attempt_allocation_at_safepoint(word_size, 1717 true /* expect_null_mutator_alloc_region */); 1718 if (result != NULL) { 1719 assert(*succeeded, "sanity"); 1720 return result; 1721 } 1722 1723 assert(!collector_policy()->should_clear_all_soft_refs(), 1724 "Flag should have been handled and cleared prior to this point"); 1725 1726 // What else? We might try synchronous finalization later. If the total 1727 // space available is large enough for the allocation, then a more 1728 // complete compaction phase than we've tried so far might be 1729 // appropriate. 1730 assert(*succeeded, "sanity"); 1731 return NULL; 1732 } 1733 1734 // Attempting to expand the heap sufficiently 1735 // to support an allocation of the given "word_size". If 1736 // successful, perform the allocation and return the address of the 1737 // allocated block, or else "NULL". 1738 1739 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1740 assert_at_safepoint(true /* should_be_vm_thread */); 1741 1742 verify_region_sets_optional(); 1743 1744 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1745 ergo_verbose1(ErgoHeapSizing, 1746 "attempt heap expansion", 1747 ergo_format_reason("allocation request failed") 1748 ergo_format_byte("allocation request"), 1749 word_size * HeapWordSize); 1750 if (expand(expand_bytes)) { 1751 _hrs.verify_optional(); 1752 verify_region_sets_optional(); 1753 return attempt_allocation_at_safepoint(word_size, 1754 false /* expect_null_mutator_alloc_region */); 1755 } 1756 return NULL; 1757 } 1758 1759 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1760 HeapWord* new_end) { 1761 assert(old_end != new_end, "don't call this otherwise"); 1762 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1763 1764 // Update the committed mem region. 1765 _g1_committed.set_end(new_end); 1766 // Tell the card table about the update. 1767 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1768 // Tell the BOT about the update. 1769 _bot_shared->resize(_g1_committed.word_size()); 1770 // Tell the hot card cache about the update 1771 _cg1r->hot_card_cache()->resize_card_counts(capacity()); 1772 } 1773 1774 bool G1CollectedHeap::expand(size_t expand_bytes) { 1775 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1776 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1777 HeapRegion::GrainBytes); 1778 ergo_verbose2(ErgoHeapSizing, 1779 "expand the heap", 1780 ergo_format_byte("requested expansion amount") 1781 ergo_format_byte("attempted expansion amount"), 1782 expand_bytes, aligned_expand_bytes); 1783 1784 if (_g1_storage.uncommitted_size() == 0) { 1785 ergo_verbose0(ErgoHeapSizing, 1786 "did not expand the heap", 1787 ergo_format_reason("heap already fully expanded")); 1788 return false; 1789 } 1790 1791 // First commit the memory. 1792 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1793 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1794 if (successful) { 1795 // Then propagate this update to the necessary data structures. 1796 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1797 update_committed_space(old_end, new_end); 1798 1799 FreeRegionList expansion_list("Local Expansion List"); 1800 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1801 assert(mr.start() == old_end, "post-condition"); 1802 // mr might be a smaller region than what was requested if 1803 // expand_by() was unable to allocate the HeapRegion instances 1804 assert(mr.end() <= new_end, "post-condition"); 1805 1806 size_t actual_expand_bytes = mr.byte_size(); 1807 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1808 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1809 "post-condition"); 1810 if (actual_expand_bytes < aligned_expand_bytes) { 1811 // We could not expand _hrs to the desired size. In this case we 1812 // need to shrink the committed space accordingly. 1813 assert(mr.end() < new_end, "invariant"); 1814 1815 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1816 // First uncommit the memory. 1817 _g1_storage.shrink_by(diff_bytes); 1818 // Then propagate this update to the necessary data structures. 1819 update_committed_space(new_end, mr.end()); 1820 } 1821 _free_list.add_as_tail(&expansion_list); 1822 1823 if (_hr_printer.is_active()) { 1824 HeapWord* curr = mr.start(); 1825 while (curr < mr.end()) { 1826 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1827 _hr_printer.commit(curr, curr_end); 1828 curr = curr_end; 1829 } 1830 assert(curr == mr.end(), "post-condition"); 1831 } 1832 g1_policy()->record_new_heap_size(n_regions()); 1833 } else { 1834 ergo_verbose0(ErgoHeapSizing, 1835 "did not expand the heap", 1836 ergo_format_reason("heap expansion operation failed")); 1837 // The expansion of the virtual storage space was unsuccessful. 1838 // Let's see if it was because we ran out of swap. 1839 if (G1ExitOnExpansionFailure && 1840 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1841 // We had head room... 1842 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion"); 1843 } 1844 } 1845 return successful; 1846 } 1847 1848 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1849 size_t aligned_shrink_bytes = 1850 ReservedSpace::page_align_size_down(shrink_bytes); 1851 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1852 HeapRegion::GrainBytes); 1853 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes); 1854 1855 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove); 1856 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1857 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes; 1858 1859 ergo_verbose3(ErgoHeapSizing, 1860 "shrink the heap", 1861 ergo_format_byte("requested shrinking amount") 1862 ergo_format_byte("aligned shrinking amount") 1863 ergo_format_byte("attempted shrinking amount"), 1864 shrink_bytes, aligned_shrink_bytes, shrunk_bytes); 1865 if (num_regions_removed > 0) { 1866 _g1_storage.shrink_by(shrunk_bytes); 1867 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1868 1869 if (_hr_printer.is_active()) { 1870 HeapWord* curr = old_end; 1871 while (curr > new_end) { 1872 HeapWord* curr_end = curr; 1873 curr -= HeapRegion::GrainWords; 1874 _hr_printer.uncommit(curr, curr_end); 1875 } 1876 } 1877 1878 _expansion_regions += num_regions_removed; 1879 update_committed_space(old_end, new_end); 1880 HeapRegionRemSet::shrink_heap(n_regions()); 1881 g1_policy()->record_new_heap_size(n_regions()); 1882 } else { 1883 ergo_verbose0(ErgoHeapSizing, 1884 "did not shrink the heap", 1885 ergo_format_reason("heap shrinking operation failed")); 1886 } 1887 } 1888 1889 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1890 verify_region_sets_optional(); 1891 1892 // We should only reach here at the end of a Full GC which means we 1893 // should not not be holding to any GC alloc regions. The method 1894 // below will make sure of that and do any remaining clean up. 1895 abandon_gc_alloc_regions(); 1896 1897 // Instead of tearing down / rebuilding the free lists here, we 1898 // could instead use the remove_all_pending() method on free_list to 1899 // remove only the ones that we need to remove. 1900 tear_down_region_sets(true /* free_list_only */); 1901 shrink_helper(shrink_bytes); 1902 rebuild_region_sets(true /* free_list_only */); 1903 1904 _hrs.verify_optional(); 1905 verify_region_sets_optional(); 1906 } 1907 1908 // Public methods. 1909 1910 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1911 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1912 #endif // _MSC_VER 1913 1914 1915 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1916 SharedHeap(policy_), 1917 _g1_policy(policy_), 1918 _dirty_card_queue_set(false), 1919 _into_cset_dirty_card_queue_set(false), 1920 _is_alive_closure_cm(this), 1921 _is_alive_closure_stw(this), 1922 _ref_processor_cm(NULL), 1923 _ref_processor_stw(NULL), 1924 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1925 _bot_shared(NULL), 1926 _evac_failure_scan_stack(NULL), 1927 _mark_in_progress(false), 1928 _cg1r(NULL), _summary_bytes_used(0), 1929 _g1mm(NULL), 1930 _refine_cte_cl(NULL), 1931 _full_collection(false), 1932 _free_list("Master Free List"), 1933 _secondary_free_list("Secondary Free List"), 1934 _old_set("Old Set"), 1935 _humongous_set("Master Humongous Set"), 1936 _free_regions_coming(false), 1937 _young_list(new YoungList(this)), 1938 _gc_time_stamp(0), 1939 _retained_old_gc_alloc_region(NULL), 1940 _survivor_plab_stats(YoungPLABSize, PLABWeight), 1941 _old_plab_stats(OldPLABSize, PLABWeight), 1942 _expand_heap_after_alloc_failure(true), 1943 _surviving_young_words(NULL), 1944 _old_marking_cycles_started(0), 1945 _old_marking_cycles_completed(0), 1946 _concurrent_cycle_started(false), 1947 _in_cset_fast_test(NULL), 1948 _in_cset_fast_test_base(NULL), 1949 _dirty_cards_region_list(NULL), 1950 _worker_cset_start_region(NULL), 1951 _worker_cset_start_region_time_stamp(NULL), 1952 _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()), 1953 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 1954 _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()), 1955 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()) { 1956 1957 _g1h = this; 1958 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1959 vm_exit_during_initialization("Failed necessary allocation."); 1960 } 1961 1962 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1963 1964 int n_queues = MAX2((int)ParallelGCThreads, 1); 1965 _task_queues = new RefToScanQueueSet(n_queues); 1966 1967 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1968 assert(n_rem_sets > 0, "Invariant."); 1969 1970 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1971 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1972 _evacuation_failed_info_array = NEW_C_HEAP_ARRAY(EvacuationFailedInfo, n_queues, mtGC); 1973 1974 for (int i = 0; i < n_queues; i++) { 1975 RefToScanQueue* q = new RefToScanQueue(); 1976 q->initialize(); 1977 _task_queues->register_queue(i, q); 1978 ::new (&_evacuation_failed_info_array[i]) EvacuationFailedInfo(); 1979 } 1980 clear_cset_start_regions(); 1981 1982 // Initialize the G1EvacuationFailureALot counters and flags. 1983 NOT_PRODUCT(reset_evacuation_should_fail();) 1984 1985 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1986 } 1987 1988 jint G1CollectedHeap::initialize() { 1989 CollectedHeap::pre_initialize(); 1990 os::enable_vtime(); 1991 1992 G1Log::init(); 1993 1994 // Necessary to satisfy locking discipline assertions. 1995 1996 MutexLocker x(Heap_lock); 1997 1998 // We have to initialize the printer before committing the heap, as 1999 // it will be used then. 2000 _hr_printer.set_active(G1PrintHeapRegions); 2001 2002 // While there are no constraints in the GC code that HeapWordSize 2003 // be any particular value, there are multiple other areas in the 2004 // system which believe this to be true (e.g. oop->object_size in some 2005 // cases incorrectly returns the size in wordSize units rather than 2006 // HeapWordSize). 2007 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 2008 2009 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 2010 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 2011 size_t heap_alignment = collector_policy()->heap_alignment(); 2012 2013 // Ensure that the sizes are properly aligned. 2014 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2015 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2016 Universe::check_alignment(max_byte_size, heap_alignment, "g1 heap"); 2017 2018 _cg1r = new ConcurrentG1Refine(this); 2019 2020 // Reserve the maximum. 2021 2022 // When compressed oops are enabled, the preferred heap base 2023 // is calculated by subtracting the requested size from the 2024 // 32Gb boundary and using the result as the base address for 2025 // heap reservation. If the requested size is not aligned to 2026 // HeapRegion::GrainBytes (i.e. the alignment that is passed 2027 // into the ReservedHeapSpace constructor) then the actual 2028 // base of the reserved heap may end up differing from the 2029 // address that was requested (i.e. the preferred heap base). 2030 // If this happens then we could end up using a non-optimal 2031 // compressed oops mode. 2032 2033 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2034 heap_alignment); 2035 2036 // It is important to do this in a way such that concurrent readers can't 2037 // temporarily think something is in the heap. (I've actually seen this 2038 // happen in asserts: DLD.) 2039 _reserved.set_word_size(0); 2040 _reserved.set_start((HeapWord*)heap_rs.base()); 2041 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2042 2043 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2044 2045 // Create the gen rem set (and barrier set) for the entire reserved region. 2046 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2047 set_barrier_set(rem_set()->bs()); 2048 if (!barrier_set()->is_a(BarrierSet::G1SATBCTLogging)) { 2049 vm_exit_during_initialization("G1 requires a G1SATBLoggingCardTableModRefBS"); 2050 return JNI_ENOMEM; 2051 } 2052 2053 // Also create a G1 rem set. 2054 _g1_rem_set = new G1RemSet(this, g1_barrier_set()); 2055 2056 // Carve out the G1 part of the heap. 2057 2058 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2059 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2060 g1_rs.size()/HeapWordSize); 2061 2062 _g1_storage.initialize(g1_rs, 0); 2063 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2064 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2065 (HeapWord*) _g1_reserved.end()); 2066 assert(_hrs.max_length() == _expansion_regions, 2067 err_msg("max length: %u expansion regions: %u", 2068 _hrs.max_length(), _expansion_regions)); 2069 2070 // Do later initialization work for concurrent refinement. 2071 _cg1r->init(); 2072 2073 // 6843694 - ensure that the maximum region index can fit 2074 // in the remembered set structures. 2075 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2076 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2077 2078 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2079 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2080 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2081 "too many cards per region"); 2082 2083 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2084 2085 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2086 heap_word_size(init_byte_size)); 2087 2088 _g1h = this; 2089 2090 _in_cset_fast_test_length = max_regions(); 2091 _in_cset_fast_test_base = 2092 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2093 2094 // We're biasing _in_cset_fast_test to avoid subtracting the 2095 // beginning of the heap every time we want to index; basically 2096 // it's the same with what we do with the card table. 2097 _in_cset_fast_test = _in_cset_fast_test_base - 2098 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2099 2100 // Clear the _cset_fast_test bitmap in anticipation of adding 2101 // regions to the incremental collection set for the first 2102 // evacuation pause. 2103 clear_cset_fast_test(); 2104 2105 // Create the ConcurrentMark data structure and thread. 2106 // (Must do this late, so that "max_regions" is defined.) 2107 _cm = new ConcurrentMark(this, heap_rs); 2108 if (_cm == NULL || !_cm->completed_initialization()) { 2109 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2110 return JNI_ENOMEM; 2111 } 2112 _cmThread = _cm->cmThread(); 2113 2114 // Initialize the from_card cache structure of HeapRegionRemSet. 2115 HeapRegionRemSet::init_heap(max_regions()); 2116 2117 // Now expand into the initial heap size. 2118 if (!expand(init_byte_size)) { 2119 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2120 return JNI_ENOMEM; 2121 } 2122 2123 // Perform any initialization actions delegated to the policy. 2124 g1_policy()->init(); 2125 2126 _refine_cte_cl = 2127 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2128 g1_rem_set(), 2129 concurrent_g1_refine()); 2130 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2131 2132 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2133 SATB_Q_FL_lock, 2134 G1SATBProcessCompletedThreshold, 2135 Shared_SATB_Q_lock); 2136 2137 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2138 DirtyCardQ_FL_lock, 2139 concurrent_g1_refine()->yellow_zone(), 2140 concurrent_g1_refine()->red_zone(), 2141 Shared_DirtyCardQ_lock); 2142 2143 if (G1DeferredRSUpdate) { 2144 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2145 DirtyCardQ_FL_lock, 2146 -1, // never trigger processing 2147 -1, // no limit on length 2148 Shared_DirtyCardQ_lock, 2149 &JavaThread::dirty_card_queue_set()); 2150 } 2151 2152 // Initialize the card queue set used to hold cards containing 2153 // references into the collection set. 2154 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2155 DirtyCardQ_FL_lock, 2156 -1, // never trigger processing 2157 -1, // no limit on length 2158 Shared_DirtyCardQ_lock, 2159 &JavaThread::dirty_card_queue_set()); 2160 2161 // In case we're keeping closure specialization stats, initialize those 2162 // counts and that mechanism. 2163 SpecializationStats::clear(); 2164 2165 // Here we allocate the dummy full region that is required by the 2166 // G1AllocRegion class. If we don't pass an address in the reserved 2167 // space here, lots of asserts fire. 2168 2169 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2170 _g1_reserved.start()); 2171 // We'll re-use the same region whether the alloc region will 2172 // require BOT updates or not and, if it doesn't, then a non-young 2173 // region will complain that it cannot support allocations without 2174 // BOT updates. So we'll tag the dummy region as young to avoid that. 2175 dummy_region->set_young(); 2176 // Make sure it's full. 2177 dummy_region->set_top(dummy_region->end()); 2178 G1AllocRegion::setup(this, dummy_region); 2179 2180 init_mutator_alloc_region(); 2181 2182 // Do create of the monitoring and management support so that 2183 // values in the heap have been properly initialized. 2184 _g1mm = new G1MonitoringSupport(this); 2185 2186 return JNI_OK; 2187 } 2188 2189 size_t G1CollectedHeap::conservative_max_heap_alignment() { 2190 return HeapRegion::max_region_size(); 2191 } 2192 2193 void G1CollectedHeap::ref_processing_init() { 2194 // Reference processing in G1 currently works as follows: 2195 // 2196 // * There are two reference processor instances. One is 2197 // used to record and process discovered references 2198 // during concurrent marking; the other is used to 2199 // record and process references during STW pauses 2200 // (both full and incremental). 2201 // * Both ref processors need to 'span' the entire heap as 2202 // the regions in the collection set may be dotted around. 2203 // 2204 // * For the concurrent marking ref processor: 2205 // * Reference discovery is enabled at initial marking. 2206 // * Reference discovery is disabled and the discovered 2207 // references processed etc during remarking. 2208 // * Reference discovery is MT (see below). 2209 // * Reference discovery requires a barrier (see below). 2210 // * Reference processing may or may not be MT 2211 // (depending on the value of ParallelRefProcEnabled 2212 // and ParallelGCThreads). 2213 // * A full GC disables reference discovery by the CM 2214 // ref processor and abandons any entries on it's 2215 // discovered lists. 2216 // 2217 // * For the STW processor: 2218 // * Non MT discovery is enabled at the start of a full GC. 2219 // * Processing and enqueueing during a full GC is non-MT. 2220 // * During a full GC, references are processed after marking. 2221 // 2222 // * Discovery (may or may not be MT) is enabled at the start 2223 // of an incremental evacuation pause. 2224 // * References are processed near the end of a STW evacuation pause. 2225 // * For both types of GC: 2226 // * Discovery is atomic - i.e. not concurrent. 2227 // * Reference discovery will not need a barrier. 2228 2229 SharedHeap::ref_processing_init(); 2230 MemRegion mr = reserved_region(); 2231 2232 // Concurrent Mark ref processor 2233 _ref_processor_cm = 2234 new ReferenceProcessor(mr, // span 2235 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2236 // mt processing 2237 (int) ParallelGCThreads, 2238 // degree of mt processing 2239 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2240 // mt discovery 2241 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2242 // degree of mt discovery 2243 false, 2244 // Reference discovery is not atomic 2245 &_is_alive_closure_cm, 2246 // is alive closure 2247 // (for efficiency/performance) 2248 true); 2249 // Setting next fields of discovered 2250 // lists requires a barrier. 2251 2252 // STW ref processor 2253 _ref_processor_stw = 2254 new ReferenceProcessor(mr, // span 2255 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2256 // mt processing 2257 MAX2((int)ParallelGCThreads, 1), 2258 // degree of mt processing 2259 (ParallelGCThreads > 1), 2260 // mt discovery 2261 MAX2((int)ParallelGCThreads, 1), 2262 // degree of mt discovery 2263 true, 2264 // Reference discovery is atomic 2265 &_is_alive_closure_stw, 2266 // is alive closure 2267 // (for efficiency/performance) 2268 false); 2269 // Setting next fields of discovered 2270 // lists requires a barrier. 2271 } 2272 2273 size_t G1CollectedHeap::capacity() const { 2274 return _g1_committed.byte_size(); 2275 } 2276 2277 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2278 assert(!hr->continuesHumongous(), "pre-condition"); 2279 hr->reset_gc_time_stamp(); 2280 if (hr->startsHumongous()) { 2281 uint first_index = hr->hrs_index() + 1; 2282 uint last_index = hr->last_hc_index(); 2283 for (uint i = first_index; i < last_index; i += 1) { 2284 HeapRegion* chr = region_at(i); 2285 assert(chr->continuesHumongous(), "sanity"); 2286 chr->reset_gc_time_stamp(); 2287 } 2288 } 2289 } 2290 2291 #ifndef PRODUCT 2292 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2293 private: 2294 unsigned _gc_time_stamp; 2295 bool _failures; 2296 2297 public: 2298 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2299 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2300 2301 virtual bool doHeapRegion(HeapRegion* hr) { 2302 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2303 if (_gc_time_stamp != region_gc_time_stamp) { 2304 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2305 "expected %d", HR_FORMAT_PARAMS(hr), 2306 region_gc_time_stamp, _gc_time_stamp); 2307 _failures = true; 2308 } 2309 return false; 2310 } 2311 2312 bool failures() { return _failures; } 2313 }; 2314 2315 void G1CollectedHeap::check_gc_time_stamps() { 2316 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2317 heap_region_iterate(&cl); 2318 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2319 } 2320 #endif // PRODUCT 2321 2322 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2323 DirtyCardQueue* into_cset_dcq, 2324 bool concurrent, 2325 int worker_i) { 2326 // Clean cards in the hot card cache 2327 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 2328 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq); 2329 2330 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2331 int n_completed_buffers = 0; 2332 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2333 n_completed_buffers++; 2334 } 2335 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2336 dcqs.clear_n_completed_buffers(); 2337 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2338 } 2339 2340 2341 // Computes the sum of the storage used by the various regions. 2342 2343 size_t G1CollectedHeap::used() const { 2344 assert(Heap_lock->owner() != NULL, 2345 "Should be owned on this thread's behalf."); 2346 size_t result = _summary_bytes_used; 2347 // Read only once in case it is set to NULL concurrently 2348 HeapRegion* hr = _mutator_alloc_region.get(); 2349 if (hr != NULL) 2350 result += hr->used(); 2351 return result; 2352 } 2353 2354 size_t G1CollectedHeap::used_unlocked() const { 2355 size_t result = _summary_bytes_used; 2356 return result; 2357 } 2358 2359 class SumUsedClosure: public HeapRegionClosure { 2360 size_t _used; 2361 public: 2362 SumUsedClosure() : _used(0) {} 2363 bool doHeapRegion(HeapRegion* r) { 2364 if (!r->continuesHumongous()) { 2365 _used += r->used(); 2366 } 2367 return false; 2368 } 2369 size_t result() { return _used; } 2370 }; 2371 2372 size_t G1CollectedHeap::recalculate_used() const { 2373 SumUsedClosure blk; 2374 heap_region_iterate(&blk); 2375 return blk.result(); 2376 } 2377 2378 size_t G1CollectedHeap::unsafe_max_alloc() { 2379 if (free_regions() > 0) return HeapRegion::GrainBytes; 2380 // otherwise, is there space in the current allocation region? 2381 2382 // We need to store the current allocation region in a local variable 2383 // here. The problem is that this method doesn't take any locks and 2384 // there may be other threads which overwrite the current allocation 2385 // region field. attempt_allocation(), for example, sets it to NULL 2386 // and this can happen *after* the NULL check here but before the call 2387 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2388 // to be a problem in the optimized build, since the two loads of the 2389 // current allocation region field are optimized away. 2390 HeapRegion* hr = _mutator_alloc_region.get(); 2391 if (hr == NULL) { 2392 return 0; 2393 } 2394 return hr->free(); 2395 } 2396 2397 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2398 switch (cause) { 2399 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2400 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2401 case GCCause::_g1_humongous_allocation: return true; 2402 default: return false; 2403 } 2404 } 2405 2406 #ifndef PRODUCT 2407 void G1CollectedHeap::allocate_dummy_regions() { 2408 // Let's fill up most of the region 2409 size_t word_size = HeapRegion::GrainWords - 1024; 2410 // And as a result the region we'll allocate will be humongous. 2411 guarantee(isHumongous(word_size), "sanity"); 2412 2413 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2414 // Let's use the existing mechanism for the allocation 2415 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2416 if (dummy_obj != NULL) { 2417 MemRegion mr(dummy_obj, word_size); 2418 CollectedHeap::fill_with_object(mr); 2419 } else { 2420 // If we can't allocate once, we probably cannot allocate 2421 // again. Let's get out of the loop. 2422 break; 2423 } 2424 } 2425 } 2426 #endif // !PRODUCT 2427 2428 void G1CollectedHeap::increment_old_marking_cycles_started() { 2429 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2430 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2431 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2432 _old_marking_cycles_started, _old_marking_cycles_completed)); 2433 2434 _old_marking_cycles_started++; 2435 } 2436 2437 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2438 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2439 2440 // We assume that if concurrent == true, then the caller is a 2441 // concurrent thread that was joined the Suspendible Thread 2442 // Set. If there's ever a cheap way to check this, we should add an 2443 // assert here. 2444 2445 // Given that this method is called at the end of a Full GC or of a 2446 // concurrent cycle, and those can be nested (i.e., a Full GC can 2447 // interrupt a concurrent cycle), the number of full collections 2448 // completed should be either one (in the case where there was no 2449 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2450 // behind the number of full collections started. 2451 2452 // This is the case for the inner caller, i.e. a Full GC. 2453 assert(concurrent || 2454 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2455 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2456 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2457 "is inconsistent with _old_marking_cycles_completed = %u", 2458 _old_marking_cycles_started, _old_marking_cycles_completed)); 2459 2460 // This is the case for the outer caller, i.e. the concurrent cycle. 2461 assert(!concurrent || 2462 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2463 err_msg("for outer caller (concurrent cycle): " 2464 "_old_marking_cycles_started = %u " 2465 "is inconsistent with _old_marking_cycles_completed = %u", 2466 _old_marking_cycles_started, _old_marking_cycles_completed)); 2467 2468 _old_marking_cycles_completed += 1; 2469 2470 // We need to clear the "in_progress" flag in the CM thread before 2471 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2472 // is set) so that if a waiter requests another System.gc() it doesn't 2473 // incorrectly see that a marking cycle is still in progress. 2474 if (concurrent) { 2475 _cmThread->clear_in_progress(); 2476 } 2477 2478 // This notify_all() will ensure that a thread that called 2479 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2480 // and it's waiting for a full GC to finish will be woken up. It is 2481 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2482 FullGCCount_lock->notify_all(); 2483 } 2484 2485 void G1CollectedHeap::register_concurrent_cycle_start(jlong start_time) { 2486 _concurrent_cycle_started = true; 2487 _gc_timer_cm->register_gc_start(start_time); 2488 2489 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start()); 2490 trace_heap_before_gc(_gc_tracer_cm); 2491 } 2492 2493 void G1CollectedHeap::register_concurrent_cycle_end() { 2494 if (_concurrent_cycle_started) { 2495 if (_cm->has_aborted()) { 2496 _gc_tracer_cm->report_concurrent_mode_failure(); 2497 } 2498 2499 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2500 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2501 2502 _concurrent_cycle_started = false; 2503 } 2504 } 2505 2506 void G1CollectedHeap::trace_heap_after_concurrent_cycle() { 2507 if (_concurrent_cycle_started) { 2508 trace_heap_after_gc(_gc_tracer_cm); 2509 } 2510 } 2511 2512 G1YCType G1CollectedHeap::yc_type() { 2513 bool is_young = g1_policy()->gcs_are_young(); 2514 bool is_initial_mark = g1_policy()->during_initial_mark_pause(); 2515 bool is_during_mark = mark_in_progress(); 2516 2517 if (is_initial_mark) { 2518 return InitialMark; 2519 } else if (is_during_mark) { 2520 return DuringMark; 2521 } else if (is_young) { 2522 return Normal; 2523 } else { 2524 return Mixed; 2525 } 2526 } 2527 2528 void G1CollectedHeap::collect(GCCause::Cause cause) { 2529 assert_heap_not_locked(); 2530 2531 unsigned int gc_count_before; 2532 unsigned int old_marking_count_before; 2533 bool retry_gc; 2534 2535 do { 2536 retry_gc = false; 2537 2538 { 2539 MutexLocker ml(Heap_lock); 2540 2541 // Read the GC count while holding the Heap_lock 2542 gc_count_before = total_collections(); 2543 old_marking_count_before = _old_marking_cycles_started; 2544 } 2545 2546 if (should_do_concurrent_full_gc(cause)) { 2547 // Schedule an initial-mark evacuation pause that will start a 2548 // concurrent cycle. We're setting word_size to 0 which means that 2549 // we are not requesting a post-GC allocation. 2550 VM_G1IncCollectionPause op(gc_count_before, 2551 0, /* word_size */ 2552 true, /* should_initiate_conc_mark */ 2553 g1_policy()->max_pause_time_ms(), 2554 cause); 2555 2556 VMThread::execute(&op); 2557 if (!op.pause_succeeded()) { 2558 if (old_marking_count_before == _old_marking_cycles_started) { 2559 retry_gc = op.should_retry_gc(); 2560 } else { 2561 // A Full GC happened while we were trying to schedule the 2562 // initial-mark GC. No point in starting a new cycle given 2563 // that the whole heap was collected anyway. 2564 } 2565 2566 if (retry_gc) { 2567 if (GC_locker::is_active_and_needs_gc()) { 2568 GC_locker::stall_until_clear(); 2569 } 2570 } 2571 } 2572 } else { 2573 if (cause == GCCause::_gc_locker 2574 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2575 2576 // Schedule a standard evacuation pause. We're setting word_size 2577 // to 0 which means that we are not requesting a post-GC allocation. 2578 VM_G1IncCollectionPause op(gc_count_before, 2579 0, /* word_size */ 2580 false, /* should_initiate_conc_mark */ 2581 g1_policy()->max_pause_time_ms(), 2582 cause); 2583 VMThread::execute(&op); 2584 } else { 2585 // Schedule a Full GC. 2586 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2587 VMThread::execute(&op); 2588 } 2589 } 2590 } while (retry_gc); 2591 } 2592 2593 bool G1CollectedHeap::is_in(const void* p) const { 2594 if (_g1_committed.contains(p)) { 2595 // Given that we know that p is in the committed space, 2596 // heap_region_containing_raw() should successfully 2597 // return the containing region. 2598 HeapRegion* hr = heap_region_containing_raw(p); 2599 return hr->is_in(p); 2600 } else { 2601 return false; 2602 } 2603 } 2604 2605 // Iteration functions. 2606 2607 // Iterates an OopClosure over all ref-containing fields of objects 2608 // within a HeapRegion. 2609 2610 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2611 MemRegion _mr; 2612 ExtendedOopClosure* _cl; 2613 public: 2614 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2615 : _mr(mr), _cl(cl) {} 2616 bool doHeapRegion(HeapRegion* r) { 2617 if (!r->continuesHumongous()) { 2618 r->oop_iterate(_cl); 2619 } 2620 return false; 2621 } 2622 }; 2623 2624 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2625 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2626 heap_region_iterate(&blk); 2627 } 2628 2629 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2630 IterateOopClosureRegionClosure blk(mr, cl); 2631 heap_region_iterate(&blk); 2632 } 2633 2634 // Iterates an ObjectClosure over all objects within a HeapRegion. 2635 2636 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2637 ObjectClosure* _cl; 2638 public: 2639 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2640 bool doHeapRegion(HeapRegion* r) { 2641 if (! r->continuesHumongous()) { 2642 r->object_iterate(_cl); 2643 } 2644 return false; 2645 } 2646 }; 2647 2648 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2649 IterateObjectClosureRegionClosure blk(cl); 2650 heap_region_iterate(&blk); 2651 } 2652 2653 // Calls a SpaceClosure on a HeapRegion. 2654 2655 class SpaceClosureRegionClosure: public HeapRegionClosure { 2656 SpaceClosure* _cl; 2657 public: 2658 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2659 bool doHeapRegion(HeapRegion* r) { 2660 _cl->do_space(r); 2661 return false; 2662 } 2663 }; 2664 2665 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2666 SpaceClosureRegionClosure blk(cl); 2667 heap_region_iterate(&blk); 2668 } 2669 2670 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2671 _hrs.iterate(cl); 2672 } 2673 2674 void 2675 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2676 uint worker_id, 2677 uint no_of_par_workers, 2678 jint claim_value) { 2679 const uint regions = n_regions(); 2680 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2681 no_of_par_workers : 2682 1); 2683 assert(UseDynamicNumberOfGCThreads || 2684 no_of_par_workers == workers()->total_workers(), 2685 "Non dynamic should use fixed number of workers"); 2686 // try to spread out the starting points of the workers 2687 const HeapRegion* start_hr = 2688 start_region_for_worker(worker_id, no_of_par_workers); 2689 const uint start_index = start_hr->hrs_index(); 2690 2691 // each worker will actually look at all regions 2692 for (uint count = 0; count < regions; ++count) { 2693 const uint index = (start_index + count) % regions; 2694 assert(0 <= index && index < regions, "sanity"); 2695 HeapRegion* r = region_at(index); 2696 // we'll ignore "continues humongous" regions (we'll process them 2697 // when we come across their corresponding "start humongous" 2698 // region) and regions already claimed 2699 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2700 continue; 2701 } 2702 // OK, try to claim it 2703 if (r->claimHeapRegion(claim_value)) { 2704 // success! 2705 assert(!r->continuesHumongous(), "sanity"); 2706 if (r->startsHumongous()) { 2707 // If the region is "starts humongous" we'll iterate over its 2708 // "continues humongous" first; in fact we'll do them 2709 // first. The order is important. In on case, calling the 2710 // closure on the "starts humongous" region might de-allocate 2711 // and clear all its "continues humongous" regions and, as a 2712 // result, we might end up processing them twice. So, we'll do 2713 // them first (notice: most closures will ignore them anyway) and 2714 // then we'll do the "starts humongous" region. 2715 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2716 HeapRegion* chr = region_at(ch_index); 2717 2718 // if the region has already been claimed or it's not 2719 // "continues humongous" we're done 2720 if (chr->claim_value() == claim_value || 2721 !chr->continuesHumongous()) { 2722 break; 2723 } 2724 2725 // No one should have claimed it directly. We can given 2726 // that we claimed its "starts humongous" region. 2727 assert(chr->claim_value() != claim_value, "sanity"); 2728 assert(chr->humongous_start_region() == r, "sanity"); 2729 2730 if (chr->claimHeapRegion(claim_value)) { 2731 // we should always be able to claim it; no one else should 2732 // be trying to claim this region 2733 2734 bool res2 = cl->doHeapRegion(chr); 2735 assert(!res2, "Should not abort"); 2736 2737 // Right now, this holds (i.e., no closure that actually 2738 // does something with "continues humongous" regions 2739 // clears them). We might have to weaken it in the future, 2740 // but let's leave these two asserts here for extra safety. 2741 assert(chr->continuesHumongous(), "should still be the case"); 2742 assert(chr->humongous_start_region() == r, "sanity"); 2743 } else { 2744 guarantee(false, "we should not reach here"); 2745 } 2746 } 2747 } 2748 2749 assert(!r->continuesHumongous(), "sanity"); 2750 bool res = cl->doHeapRegion(r); 2751 assert(!res, "Should not abort"); 2752 } 2753 } 2754 } 2755 2756 class ResetClaimValuesClosure: public HeapRegionClosure { 2757 public: 2758 bool doHeapRegion(HeapRegion* r) { 2759 r->set_claim_value(HeapRegion::InitialClaimValue); 2760 return false; 2761 } 2762 }; 2763 2764 void G1CollectedHeap::reset_heap_region_claim_values() { 2765 ResetClaimValuesClosure blk; 2766 heap_region_iterate(&blk); 2767 } 2768 2769 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2770 ResetClaimValuesClosure blk; 2771 collection_set_iterate(&blk); 2772 } 2773 2774 #ifdef ASSERT 2775 // This checks whether all regions in the heap have the correct claim 2776 // value. I also piggy-backed on this a check to ensure that the 2777 // humongous_start_region() information on "continues humongous" 2778 // regions is correct. 2779 2780 class CheckClaimValuesClosure : public HeapRegionClosure { 2781 private: 2782 jint _claim_value; 2783 uint _failures; 2784 HeapRegion* _sh_region; 2785 2786 public: 2787 CheckClaimValuesClosure(jint claim_value) : 2788 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2789 bool doHeapRegion(HeapRegion* r) { 2790 if (r->claim_value() != _claim_value) { 2791 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2792 "claim value = %d, should be %d", 2793 HR_FORMAT_PARAMS(r), 2794 r->claim_value(), _claim_value); 2795 ++_failures; 2796 } 2797 if (!r->isHumongous()) { 2798 _sh_region = NULL; 2799 } else if (r->startsHumongous()) { 2800 _sh_region = r; 2801 } else if (r->continuesHumongous()) { 2802 if (r->humongous_start_region() != _sh_region) { 2803 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2804 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2805 HR_FORMAT_PARAMS(r), 2806 r->humongous_start_region(), 2807 _sh_region); 2808 ++_failures; 2809 } 2810 } 2811 return false; 2812 } 2813 uint failures() { return _failures; } 2814 }; 2815 2816 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2817 CheckClaimValuesClosure cl(claim_value); 2818 heap_region_iterate(&cl); 2819 return cl.failures() == 0; 2820 } 2821 2822 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2823 private: 2824 jint _claim_value; 2825 uint _failures; 2826 2827 public: 2828 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2829 _claim_value(claim_value), _failures(0) { } 2830 2831 uint failures() { return _failures; } 2832 2833 bool doHeapRegion(HeapRegion* hr) { 2834 assert(hr->in_collection_set(), "how?"); 2835 assert(!hr->isHumongous(), "H-region in CSet"); 2836 if (hr->claim_value() != _claim_value) { 2837 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2838 "claim value = %d, should be %d", 2839 HR_FORMAT_PARAMS(hr), 2840 hr->claim_value(), _claim_value); 2841 _failures += 1; 2842 } 2843 return false; 2844 } 2845 }; 2846 2847 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2848 CheckClaimValuesInCSetHRClosure cl(claim_value); 2849 collection_set_iterate(&cl); 2850 return cl.failures() == 0; 2851 } 2852 #endif // ASSERT 2853 2854 // Clear the cached CSet starting regions and (more importantly) 2855 // the time stamps. Called when we reset the GC time stamp. 2856 void G1CollectedHeap::clear_cset_start_regions() { 2857 assert(_worker_cset_start_region != NULL, "sanity"); 2858 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2859 2860 int n_queues = MAX2((int)ParallelGCThreads, 1); 2861 for (int i = 0; i < n_queues; i++) { 2862 _worker_cset_start_region[i] = NULL; 2863 _worker_cset_start_region_time_stamp[i] = 0; 2864 } 2865 } 2866 2867 // Given the id of a worker, obtain or calculate a suitable 2868 // starting region for iterating over the current collection set. 2869 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2870 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2871 2872 HeapRegion* result = NULL; 2873 unsigned gc_time_stamp = get_gc_time_stamp(); 2874 2875 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2876 // Cached starting region for current worker was set 2877 // during the current pause - so it's valid. 2878 // Note: the cached starting heap region may be NULL 2879 // (when the collection set is empty). 2880 result = _worker_cset_start_region[worker_i]; 2881 assert(result == NULL || result->in_collection_set(), "sanity"); 2882 return result; 2883 } 2884 2885 // The cached entry was not valid so let's calculate 2886 // a suitable starting heap region for this worker. 2887 2888 // We want the parallel threads to start their collection 2889 // set iteration at different collection set regions to 2890 // avoid contention. 2891 // If we have: 2892 // n collection set regions 2893 // p threads 2894 // Then thread t will start at region floor ((t * n) / p) 2895 2896 result = g1_policy()->collection_set(); 2897 if (G1CollectedHeap::use_parallel_gc_threads()) { 2898 uint cs_size = g1_policy()->cset_region_length(); 2899 uint active_workers = workers()->active_workers(); 2900 assert(UseDynamicNumberOfGCThreads || 2901 active_workers == workers()->total_workers(), 2902 "Unless dynamic should use total workers"); 2903 2904 uint end_ind = (cs_size * worker_i) / active_workers; 2905 uint start_ind = 0; 2906 2907 if (worker_i > 0 && 2908 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2909 // Previous workers starting region is valid 2910 // so let's iterate from there 2911 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2912 result = _worker_cset_start_region[worker_i - 1]; 2913 } 2914 2915 for (uint i = start_ind; i < end_ind; i++) { 2916 result = result->next_in_collection_set(); 2917 } 2918 } 2919 2920 // Note: the calculated starting heap region may be NULL 2921 // (when the collection set is empty). 2922 assert(result == NULL || result->in_collection_set(), "sanity"); 2923 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2924 "should be updated only once per pause"); 2925 _worker_cset_start_region[worker_i] = result; 2926 OrderAccess::storestore(); 2927 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2928 return result; 2929 } 2930 2931 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2932 uint no_of_par_workers) { 2933 uint worker_num = 2934 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2935 assert(UseDynamicNumberOfGCThreads || 2936 no_of_par_workers == workers()->total_workers(), 2937 "Non dynamic should use fixed number of workers"); 2938 const uint start_index = n_regions() * worker_i / worker_num; 2939 return region_at(start_index); 2940 } 2941 2942 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2943 HeapRegion* r = g1_policy()->collection_set(); 2944 while (r != NULL) { 2945 HeapRegion* next = r->next_in_collection_set(); 2946 if (cl->doHeapRegion(r)) { 2947 cl->incomplete(); 2948 return; 2949 } 2950 r = next; 2951 } 2952 } 2953 2954 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2955 HeapRegionClosure *cl) { 2956 if (r == NULL) { 2957 // The CSet is empty so there's nothing to do. 2958 return; 2959 } 2960 2961 assert(r->in_collection_set(), 2962 "Start region must be a member of the collection set."); 2963 HeapRegion* cur = r; 2964 while (cur != NULL) { 2965 HeapRegion* next = cur->next_in_collection_set(); 2966 if (cl->doHeapRegion(cur) && false) { 2967 cl->incomplete(); 2968 return; 2969 } 2970 cur = next; 2971 } 2972 cur = g1_policy()->collection_set(); 2973 while (cur != r) { 2974 HeapRegion* next = cur->next_in_collection_set(); 2975 if (cl->doHeapRegion(cur) && false) { 2976 cl->incomplete(); 2977 return; 2978 } 2979 cur = next; 2980 } 2981 } 2982 2983 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2984 return n_regions() > 0 ? region_at(0) : NULL; 2985 } 2986 2987 2988 Space* G1CollectedHeap::space_containing(const void* addr) const { 2989 Space* res = heap_region_containing(addr); 2990 return res; 2991 } 2992 2993 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2994 Space* sp = space_containing(addr); 2995 if (sp != NULL) { 2996 return sp->block_start(addr); 2997 } 2998 return NULL; 2999 } 3000 3001 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 3002 Space* sp = space_containing(addr); 3003 assert(sp != NULL, "block_size of address outside of heap"); 3004 return sp->block_size(addr); 3005 } 3006 3007 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 3008 Space* sp = space_containing(addr); 3009 return sp->block_is_obj(addr); 3010 } 3011 3012 bool G1CollectedHeap::supports_tlab_allocation() const { 3013 return true; 3014 } 3015 3016 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 3017 return HeapRegion::GrainBytes; 3018 } 3019 3020 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 3021 // Return the remaining space in the cur alloc region, but not less than 3022 // the min TLAB size. 3023 3024 // Also, this value can be at most the humongous object threshold, 3025 // since we can't allow tlabs to grow big enough to accommodate 3026 // humongous objects. 3027 3028 HeapRegion* hr = _mutator_alloc_region.get(); 3029 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 3030 if (hr == NULL) { 3031 return max_tlab_size; 3032 } else { 3033 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 3034 } 3035 } 3036 3037 size_t G1CollectedHeap::max_capacity() const { 3038 return _g1_reserved.byte_size(); 3039 } 3040 3041 jlong G1CollectedHeap::millis_since_last_gc() { 3042 // assert(false, "NYI"); 3043 return 0; 3044 } 3045 3046 void G1CollectedHeap::prepare_for_verify() { 3047 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3048 ensure_parsability(false); 3049 } 3050 g1_rem_set()->prepare_for_verify(); 3051 } 3052 3053 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3054 VerifyOption vo) { 3055 switch (vo) { 3056 case VerifyOption_G1UsePrevMarking: 3057 return hr->obj_allocated_since_prev_marking(obj); 3058 case VerifyOption_G1UseNextMarking: 3059 return hr->obj_allocated_since_next_marking(obj); 3060 case VerifyOption_G1UseMarkWord: 3061 return false; 3062 default: 3063 ShouldNotReachHere(); 3064 } 3065 return false; // keep some compilers happy 3066 } 3067 3068 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3069 switch (vo) { 3070 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3071 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3072 case VerifyOption_G1UseMarkWord: return NULL; 3073 default: ShouldNotReachHere(); 3074 } 3075 return NULL; // keep some compilers happy 3076 } 3077 3078 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3079 switch (vo) { 3080 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3081 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3082 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3083 default: ShouldNotReachHere(); 3084 } 3085 return false; // keep some compilers happy 3086 } 3087 3088 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3089 switch (vo) { 3090 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3091 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3092 case VerifyOption_G1UseMarkWord: return "NONE"; 3093 default: ShouldNotReachHere(); 3094 } 3095 return NULL; // keep some compilers happy 3096 } 3097 3098 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can 3099 // pass it as the perm_blk to SharedHeap::process_strong_roots. 3100 // When process_strong_roots stop calling perm_blk->younger_refs_iterate 3101 // we can change this closure to extend the simpler OopClosure. 3102 class VerifyRootsClosure: public OopsInGenClosure { 3103 private: 3104 G1CollectedHeap* _g1h; 3105 VerifyOption _vo; 3106 bool _failures; 3107 public: 3108 // _vo == UsePrevMarking -> use "prev" marking information, 3109 // _vo == UseNextMarking -> use "next" marking information, 3110 // _vo == UseMarkWord -> use mark word from object header. 3111 VerifyRootsClosure(VerifyOption vo) : 3112 _g1h(G1CollectedHeap::heap()), 3113 _vo(vo), 3114 _failures(false) { } 3115 3116 bool failures() { return _failures; } 3117 3118 template <class T> void do_oop_nv(T* p) { 3119 T heap_oop = oopDesc::load_heap_oop(p); 3120 if (!oopDesc::is_null(heap_oop)) { 3121 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3122 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3123 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3124 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3125 if (_vo == VerifyOption_G1UseMarkWord) { 3126 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3127 } 3128 obj->print_on(gclog_or_tty); 3129 _failures = true; 3130 } 3131 } 3132 } 3133 3134 void do_oop(oop* p) { do_oop_nv(p); } 3135 void do_oop(narrowOop* p) { do_oop_nv(p); } 3136 }; 3137 3138 class G1VerifyCodeRootOopClosure: public OopsInGenClosure { 3139 G1CollectedHeap* _g1h; 3140 OopClosure* _root_cl; 3141 nmethod* _nm; 3142 VerifyOption _vo; 3143 bool _failures; 3144 3145 template <class T> void do_oop_work(T* p) { 3146 // First verify that this root is live 3147 _root_cl->do_oop(p); 3148 3149 if (!G1VerifyHeapRegionCodeRoots) { 3150 // We're not verifying the code roots attached to heap region. 3151 return; 3152 } 3153 3154 // Don't check the code roots during marking verification in a full GC 3155 if (_vo == VerifyOption_G1UseMarkWord) { 3156 return; 3157 } 3158 3159 // Now verify that the current nmethod (which contains p) is 3160 // in the code root list of the heap region containing the 3161 // object referenced by p. 3162 3163 T heap_oop = oopDesc::load_heap_oop(p); 3164 if (!oopDesc::is_null(heap_oop)) { 3165 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3166 3167 // Now fetch the region containing the object 3168 HeapRegion* hr = _g1h->heap_region_containing(obj); 3169 HeapRegionRemSet* hrrs = hr->rem_set(); 3170 // Verify that the strong code root list for this region 3171 // contains the nmethod 3172 if (!hrrs->strong_code_roots_list_contains(_nm)) { 3173 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" " 3174 "from nmethod "PTR_FORMAT" not in strong " 3175 "code roots for region ["PTR_FORMAT","PTR_FORMAT")", 3176 p, _nm, hr->bottom(), hr->end()); 3177 _failures = true; 3178 } 3179 } 3180 } 3181 3182 public: 3183 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo): 3184 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {} 3185 3186 void do_oop(oop* p) { do_oop_work(p); } 3187 void do_oop(narrowOop* p) { do_oop_work(p); } 3188 3189 void set_nmethod(nmethod* nm) { _nm = nm; } 3190 bool failures() { return _failures; } 3191 }; 3192 3193 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure { 3194 G1VerifyCodeRootOopClosure* _oop_cl; 3195 3196 public: 3197 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl): 3198 _oop_cl(oop_cl) {} 3199 3200 void do_code_blob(CodeBlob* cb) { 3201 nmethod* nm = cb->as_nmethod_or_null(); 3202 if (nm != NULL) { 3203 _oop_cl->set_nmethod(nm); 3204 nm->oops_do(_oop_cl); 3205 } 3206 } 3207 }; 3208 3209 class YoungRefCounterClosure : public OopClosure { 3210 G1CollectedHeap* _g1h; 3211 int _count; 3212 public: 3213 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3214 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3215 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3216 3217 int count() { return _count; } 3218 void reset_count() { _count = 0; }; 3219 }; 3220 3221 class VerifyKlassClosure: public KlassClosure { 3222 YoungRefCounterClosure _young_ref_counter_closure; 3223 OopClosure *_oop_closure; 3224 public: 3225 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3226 void do_klass(Klass* k) { 3227 k->oops_do(_oop_closure); 3228 3229 _young_ref_counter_closure.reset_count(); 3230 k->oops_do(&_young_ref_counter_closure); 3231 if (_young_ref_counter_closure.count() > 0) { 3232 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3233 } 3234 } 3235 }; 3236 3237 class VerifyLivenessOopClosure: public OopClosure { 3238 G1CollectedHeap* _g1h; 3239 VerifyOption _vo; 3240 public: 3241 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3242 _g1h(g1h), _vo(vo) 3243 { } 3244 void do_oop(narrowOop *p) { do_oop_work(p); } 3245 void do_oop( oop *p) { do_oop_work(p); } 3246 3247 template <class T> void do_oop_work(T *p) { 3248 oop obj = oopDesc::load_decode_heap_oop(p); 3249 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3250 "Dead object referenced by a not dead object"); 3251 } 3252 }; 3253 3254 class VerifyObjsInRegionClosure: public ObjectClosure { 3255 private: 3256 G1CollectedHeap* _g1h; 3257 size_t _live_bytes; 3258 HeapRegion *_hr; 3259 VerifyOption _vo; 3260 public: 3261 // _vo == UsePrevMarking -> use "prev" marking information, 3262 // _vo == UseNextMarking -> use "next" marking information, 3263 // _vo == UseMarkWord -> use mark word from object header. 3264 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3265 : _live_bytes(0), _hr(hr), _vo(vo) { 3266 _g1h = G1CollectedHeap::heap(); 3267 } 3268 void do_object(oop o) { 3269 VerifyLivenessOopClosure isLive(_g1h, _vo); 3270 assert(o != NULL, "Huh?"); 3271 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3272 // If the object is alive according to the mark word, 3273 // then verify that the marking information agrees. 3274 // Note we can't verify the contra-positive of the 3275 // above: if the object is dead (according to the mark 3276 // word), it may not be marked, or may have been marked 3277 // but has since became dead, or may have been allocated 3278 // since the last marking. 3279 if (_vo == VerifyOption_G1UseMarkWord) { 3280 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3281 } 3282 3283 o->oop_iterate_no_header(&isLive); 3284 if (!_hr->obj_allocated_since_prev_marking(o)) { 3285 size_t obj_size = o->size(); // Make sure we don't overflow 3286 _live_bytes += (obj_size * HeapWordSize); 3287 } 3288 } 3289 } 3290 size_t live_bytes() { return _live_bytes; } 3291 }; 3292 3293 class PrintObjsInRegionClosure : public ObjectClosure { 3294 HeapRegion *_hr; 3295 G1CollectedHeap *_g1; 3296 public: 3297 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3298 _g1 = G1CollectedHeap::heap(); 3299 }; 3300 3301 void do_object(oop o) { 3302 if (o != NULL) { 3303 HeapWord *start = (HeapWord *) o; 3304 size_t word_sz = o->size(); 3305 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3306 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3307 (void*) o, word_sz, 3308 _g1->isMarkedPrev(o), 3309 _g1->isMarkedNext(o), 3310 _hr->obj_allocated_since_prev_marking(o)); 3311 HeapWord *end = start + word_sz; 3312 HeapWord *cur; 3313 int *val; 3314 for (cur = start; cur < end; cur++) { 3315 val = (int *) cur; 3316 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3317 } 3318 } 3319 } 3320 }; 3321 3322 class VerifyRegionClosure: public HeapRegionClosure { 3323 private: 3324 bool _par; 3325 VerifyOption _vo; 3326 bool _failures; 3327 public: 3328 // _vo == UsePrevMarking -> use "prev" marking information, 3329 // _vo == UseNextMarking -> use "next" marking information, 3330 // _vo == UseMarkWord -> use mark word from object header. 3331 VerifyRegionClosure(bool par, VerifyOption vo) 3332 : _par(par), 3333 _vo(vo), 3334 _failures(false) {} 3335 3336 bool failures() { 3337 return _failures; 3338 } 3339 3340 bool doHeapRegion(HeapRegion* r) { 3341 if (!r->continuesHumongous()) { 3342 bool failures = false; 3343 r->verify(_vo, &failures); 3344 if (failures) { 3345 _failures = true; 3346 } else { 3347 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3348 r->object_iterate(¬_dead_yet_cl); 3349 if (_vo != VerifyOption_G1UseNextMarking) { 3350 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3351 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3352 "max_live_bytes "SIZE_FORMAT" " 3353 "< calculated "SIZE_FORMAT, 3354 r->bottom(), r->end(), 3355 r->max_live_bytes(), 3356 not_dead_yet_cl.live_bytes()); 3357 _failures = true; 3358 } 3359 } else { 3360 // When vo == UseNextMarking we cannot currently do a sanity 3361 // check on the live bytes as the calculation has not been 3362 // finalized yet. 3363 } 3364 } 3365 } 3366 return false; // stop the region iteration if we hit a failure 3367 } 3368 }; 3369 3370 // This is the task used for parallel verification of the heap regions 3371 3372 class G1ParVerifyTask: public AbstractGangTask { 3373 private: 3374 G1CollectedHeap* _g1h; 3375 VerifyOption _vo; 3376 bool _failures; 3377 3378 public: 3379 // _vo == UsePrevMarking -> use "prev" marking information, 3380 // _vo == UseNextMarking -> use "next" marking information, 3381 // _vo == UseMarkWord -> use mark word from object header. 3382 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3383 AbstractGangTask("Parallel verify task"), 3384 _g1h(g1h), 3385 _vo(vo), 3386 _failures(false) { } 3387 3388 bool failures() { 3389 return _failures; 3390 } 3391 3392 void work(uint worker_id) { 3393 HandleMark hm; 3394 VerifyRegionClosure blk(true, _vo); 3395 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3396 _g1h->workers()->active_workers(), 3397 HeapRegion::ParVerifyClaimValue); 3398 if (blk.failures()) { 3399 _failures = true; 3400 } 3401 } 3402 }; 3403 3404 void G1CollectedHeap::verify(bool silent, VerifyOption vo) { 3405 if (SafepointSynchronize::is_at_safepoint()) { 3406 assert(Thread::current()->is_VM_thread(), 3407 "Expected to be executed serially by the VM thread at this point"); 3408 3409 if (!silent) { gclog_or_tty->print("Roots "); } 3410 VerifyRootsClosure rootsCl(vo); 3411 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo); 3412 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl); 3413 VerifyKlassClosure klassCl(this, &rootsCl); 3414 3415 // We apply the relevant closures to all the oops in the 3416 // system dictionary, the string table and the code cache. 3417 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3418 3419 // Need cleared claim bits for the strong roots processing 3420 ClassLoaderDataGraph::clear_claimed_marks(); 3421 3422 process_strong_roots(true, // activate StrongRootsScope 3423 false, // we set "is scavenging" to false, 3424 // so we don't reset the dirty cards. 3425 ScanningOption(so), // roots scanning options 3426 &rootsCl, 3427 &blobsCl, 3428 &klassCl 3429 ); 3430 3431 bool failures = rootsCl.failures() || codeRootsCl.failures(); 3432 3433 if (vo != VerifyOption_G1UseMarkWord) { 3434 // If we're verifying during a full GC then the region sets 3435 // will have been torn down at the start of the GC. Therefore 3436 // verifying the region sets will fail. So we only verify 3437 // the region sets when not in a full GC. 3438 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3439 verify_region_sets(); 3440 } 3441 3442 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3443 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3444 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3445 "sanity check"); 3446 3447 G1ParVerifyTask task(this, vo); 3448 assert(UseDynamicNumberOfGCThreads || 3449 workers()->active_workers() == workers()->total_workers(), 3450 "If not dynamic should be using all the workers"); 3451 int n_workers = workers()->active_workers(); 3452 set_par_threads(n_workers); 3453 workers()->run_task(&task); 3454 set_par_threads(0); 3455 if (task.failures()) { 3456 failures = true; 3457 } 3458 3459 // Checks that the expected amount of parallel work was done. 3460 // The implication is that n_workers is > 0. 3461 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3462 "sanity check"); 3463 3464 reset_heap_region_claim_values(); 3465 3466 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3467 "sanity check"); 3468 } else { 3469 VerifyRegionClosure blk(false, vo); 3470 heap_region_iterate(&blk); 3471 if (blk.failures()) { 3472 failures = true; 3473 } 3474 } 3475 if (!silent) gclog_or_tty->print("RemSet "); 3476 rem_set()->verify(); 3477 3478 if (failures) { 3479 gclog_or_tty->print_cr("Heap:"); 3480 // It helps to have the per-region information in the output to 3481 // help us track down what went wrong. This is why we call 3482 // print_extended_on() instead of print_on(). 3483 print_extended_on(gclog_or_tty); 3484 gclog_or_tty->print_cr(""); 3485 #ifndef PRODUCT 3486 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3487 concurrent_mark()->print_reachable("at-verification-failure", 3488 vo, false /* all */); 3489 } 3490 #endif 3491 gclog_or_tty->flush(); 3492 } 3493 guarantee(!failures, "there should not have been any failures"); 3494 } else { 3495 if (!silent) 3496 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) "); 3497 } 3498 } 3499 3500 void G1CollectedHeap::verify(bool silent) { 3501 verify(silent, VerifyOption_G1UsePrevMarking); 3502 } 3503 3504 double G1CollectedHeap::verify(bool guard, const char* msg) { 3505 double verify_time_ms = 0.0; 3506 3507 if (guard && total_collections() >= VerifyGCStartAt) { 3508 double verify_start = os::elapsedTime(); 3509 HandleMark hm; // Discard invalid handles created during verification 3510 prepare_for_verify(); 3511 Universe::verify(VerifyOption_G1UsePrevMarking, msg); 3512 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 3513 } 3514 3515 return verify_time_ms; 3516 } 3517 3518 void G1CollectedHeap::verify_before_gc() { 3519 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 3520 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 3521 } 3522 3523 void G1CollectedHeap::verify_after_gc() { 3524 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 3525 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 3526 } 3527 3528 class PrintRegionClosure: public HeapRegionClosure { 3529 outputStream* _st; 3530 public: 3531 PrintRegionClosure(outputStream* st) : _st(st) {} 3532 bool doHeapRegion(HeapRegion* r) { 3533 r->print_on(_st); 3534 return false; 3535 } 3536 }; 3537 3538 void G1CollectedHeap::print_on(outputStream* st) const { 3539 st->print(" %-20s", "garbage-first heap"); 3540 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3541 capacity()/K, used_unlocked()/K); 3542 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3543 _g1_storage.low_boundary(), 3544 _g1_storage.high(), 3545 _g1_storage.high_boundary()); 3546 st->cr(); 3547 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3548 uint young_regions = _young_list->length(); 3549 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3550 (size_t) young_regions * HeapRegion::GrainBytes / K); 3551 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3552 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3553 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3554 st->cr(); 3555 MetaspaceAux::print_on(st); 3556 } 3557 3558 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3559 print_on(st); 3560 3561 // Print the per-region information. 3562 st->cr(); 3563 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3564 "HS=humongous(starts), HC=humongous(continues), " 3565 "CS=collection set, F=free, TS=gc time stamp, " 3566 "PTAMS=previous top-at-mark-start, " 3567 "NTAMS=next top-at-mark-start)"); 3568 PrintRegionClosure blk(st); 3569 heap_region_iterate(&blk); 3570 } 3571 3572 void G1CollectedHeap::print_on_error(outputStream* st) const { 3573 this->CollectedHeap::print_on_error(st); 3574 3575 if (_cm != NULL) { 3576 st->cr(); 3577 _cm->print_on_error(st); 3578 } 3579 } 3580 3581 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3582 if (G1CollectedHeap::use_parallel_gc_threads()) { 3583 workers()->print_worker_threads_on(st); 3584 } 3585 _cmThread->print_on(st); 3586 st->cr(); 3587 _cm->print_worker_threads_on(st); 3588 _cg1r->print_worker_threads_on(st); 3589 } 3590 3591 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3592 if (G1CollectedHeap::use_parallel_gc_threads()) { 3593 workers()->threads_do(tc); 3594 } 3595 tc->do_thread(_cmThread); 3596 _cg1r->threads_do(tc); 3597 } 3598 3599 void G1CollectedHeap::print_tracing_info() const { 3600 // We'll overload this to mean "trace GC pause statistics." 3601 if (TraceGen0Time || TraceGen1Time) { 3602 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3603 // to that. 3604 g1_policy()->print_tracing_info(); 3605 } 3606 if (G1SummarizeRSetStats) { 3607 g1_rem_set()->print_summary_info(); 3608 } 3609 if (G1SummarizeConcMark) { 3610 concurrent_mark()->print_summary_info(); 3611 } 3612 g1_policy()->print_yg_surv_rate_info(); 3613 SpecializationStats::print(); 3614 } 3615 3616 #ifndef PRODUCT 3617 // Helpful for debugging RSet issues. 3618 3619 class PrintRSetsClosure : public HeapRegionClosure { 3620 private: 3621 const char* _msg; 3622 size_t _occupied_sum; 3623 3624 public: 3625 bool doHeapRegion(HeapRegion* r) { 3626 HeapRegionRemSet* hrrs = r->rem_set(); 3627 size_t occupied = hrrs->occupied(); 3628 _occupied_sum += occupied; 3629 3630 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3631 HR_FORMAT_PARAMS(r)); 3632 if (occupied == 0) { 3633 gclog_or_tty->print_cr(" RSet is empty"); 3634 } else { 3635 hrrs->print(); 3636 } 3637 gclog_or_tty->print_cr("----------"); 3638 return false; 3639 } 3640 3641 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3642 gclog_or_tty->cr(); 3643 gclog_or_tty->print_cr("========================================"); 3644 gclog_or_tty->print_cr(msg); 3645 gclog_or_tty->cr(); 3646 } 3647 3648 ~PrintRSetsClosure() { 3649 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3650 gclog_or_tty->print_cr("========================================"); 3651 gclog_or_tty->cr(); 3652 } 3653 }; 3654 3655 void G1CollectedHeap::print_cset_rsets() { 3656 PrintRSetsClosure cl("Printing CSet RSets"); 3657 collection_set_iterate(&cl); 3658 } 3659 3660 void G1CollectedHeap::print_all_rsets() { 3661 PrintRSetsClosure cl("Printing All RSets");; 3662 heap_region_iterate(&cl); 3663 } 3664 #endif // PRODUCT 3665 3666 G1CollectedHeap* G1CollectedHeap::heap() { 3667 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3668 "not a garbage-first heap"); 3669 return _g1h; 3670 } 3671 3672 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3673 // always_do_update_barrier = false; 3674 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3675 // Fill TLAB's and such 3676 ensure_parsability(true); 3677 3678 if (G1SummarizeRSetStats && (G1SummarizeRSetStatsPeriod > 0) && 3679 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3680 g1_rem_set()->print_periodic_summary_info("Before GC RS summary"); 3681 } 3682 } 3683 3684 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3685 3686 if (G1SummarizeRSetStats && 3687 (G1SummarizeRSetStatsPeriod > 0) && 3688 // we are at the end of the GC. Total collections has already been increased. 3689 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) { 3690 g1_rem_set()->print_periodic_summary_info("After GC RS summary"); 3691 } 3692 3693 // FIXME: what is this about? 3694 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3695 // is set. 3696 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3697 "derived pointer present")); 3698 // always_do_update_barrier = true; 3699 3700 // We have just completed a GC. Update the soft reference 3701 // policy with the new heap occupancy 3702 Universe::update_heap_info_at_gc(); 3703 } 3704 3705 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3706 unsigned int gc_count_before, 3707 bool* succeeded, 3708 GCCause::Cause gc_cause) { 3709 assert_heap_not_locked_and_not_at_safepoint(); 3710 g1_policy()->record_stop_world_start(); 3711 VM_G1IncCollectionPause op(gc_count_before, 3712 word_size, 3713 false, /* should_initiate_conc_mark */ 3714 g1_policy()->max_pause_time_ms(), 3715 gc_cause); 3716 VMThread::execute(&op); 3717 3718 HeapWord* result = op.result(); 3719 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3720 assert(result == NULL || ret_succeeded, 3721 "the result should be NULL if the VM did not succeed"); 3722 *succeeded = ret_succeeded; 3723 3724 assert_heap_not_locked(); 3725 return result; 3726 } 3727 3728 void 3729 G1CollectedHeap::doConcurrentMark() { 3730 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3731 if (!_cmThread->in_progress()) { 3732 _cmThread->set_started(); 3733 CGC_lock->notify(); 3734 } 3735 } 3736 3737 size_t G1CollectedHeap::pending_card_num() { 3738 size_t extra_cards = 0; 3739 JavaThread *curr = Threads::first(); 3740 while (curr != NULL) { 3741 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3742 extra_cards += dcq.size(); 3743 curr = curr->next(); 3744 } 3745 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3746 size_t buffer_size = dcqs.buffer_size(); 3747 size_t buffer_num = dcqs.completed_buffers_num(); 3748 3749 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3750 // in bytes - not the number of 'entries'. We need to convert 3751 // into a number of cards. 3752 return (buffer_size * buffer_num + extra_cards) / oopSize; 3753 } 3754 3755 size_t G1CollectedHeap::cards_scanned() { 3756 return g1_rem_set()->cardsScanned(); 3757 } 3758 3759 void 3760 G1CollectedHeap::setup_surviving_young_words() { 3761 assert(_surviving_young_words == NULL, "pre-condition"); 3762 uint array_length = g1_policy()->young_cset_region_length(); 3763 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3764 if (_surviving_young_words == NULL) { 3765 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR, 3766 "Not enough space for young surv words summary."); 3767 } 3768 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3769 #ifdef ASSERT 3770 for (uint i = 0; i < array_length; ++i) { 3771 assert( _surviving_young_words[i] == 0, "memset above" ); 3772 } 3773 #endif // !ASSERT 3774 } 3775 3776 void 3777 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3778 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3779 uint array_length = g1_policy()->young_cset_region_length(); 3780 for (uint i = 0; i < array_length; ++i) { 3781 _surviving_young_words[i] += surv_young_words[i]; 3782 } 3783 } 3784 3785 void 3786 G1CollectedHeap::cleanup_surviving_young_words() { 3787 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3788 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3789 _surviving_young_words = NULL; 3790 } 3791 3792 #ifdef ASSERT 3793 class VerifyCSetClosure: public HeapRegionClosure { 3794 public: 3795 bool doHeapRegion(HeapRegion* hr) { 3796 // Here we check that the CSet region's RSet is ready for parallel 3797 // iteration. The fields that we'll verify are only manipulated 3798 // when the region is part of a CSet and is collected. Afterwards, 3799 // we reset these fields when we clear the region's RSet (when the 3800 // region is freed) so they are ready when the region is 3801 // re-allocated. The only exception to this is if there's an 3802 // evacuation failure and instead of freeing the region we leave 3803 // it in the heap. In that case, we reset these fields during 3804 // evacuation failure handling. 3805 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3806 3807 // Here's a good place to add any other checks we'd like to 3808 // perform on CSet regions. 3809 return false; 3810 } 3811 }; 3812 #endif // ASSERT 3813 3814 #if TASKQUEUE_STATS 3815 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3816 st->print_raw_cr("GC Task Stats"); 3817 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3818 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3819 } 3820 3821 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3822 print_taskqueue_stats_hdr(st); 3823 3824 TaskQueueStats totals; 3825 const int n = workers() != NULL ? workers()->total_workers() : 1; 3826 for (int i = 0; i < n; ++i) { 3827 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3828 totals += task_queue(i)->stats; 3829 } 3830 st->print_raw("tot "); totals.print(st); st->cr(); 3831 3832 DEBUG_ONLY(totals.verify()); 3833 } 3834 3835 void G1CollectedHeap::reset_taskqueue_stats() { 3836 const int n = workers() != NULL ? workers()->total_workers() : 1; 3837 for (int i = 0; i < n; ++i) { 3838 task_queue(i)->stats.reset(); 3839 } 3840 } 3841 #endif // TASKQUEUE_STATS 3842 3843 void G1CollectedHeap::log_gc_header() { 3844 if (!G1Log::fine()) { 3845 return; 3846 } 3847 3848 gclog_or_tty->date_stamp(PrintGCDateStamps); 3849 gclog_or_tty->stamp(PrintGCTimeStamps); 3850 3851 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3852 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3853 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3854 3855 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3856 } 3857 3858 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3859 if (!G1Log::fine()) { 3860 return; 3861 } 3862 3863 if (G1Log::finer()) { 3864 if (evacuation_failed()) { 3865 gclog_or_tty->print(" (to-space exhausted)"); 3866 } 3867 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3868 g1_policy()->phase_times()->note_gc_end(); 3869 g1_policy()->phase_times()->print(pause_time_sec); 3870 g1_policy()->print_detailed_heap_transition(); 3871 } else { 3872 if (evacuation_failed()) { 3873 gclog_or_tty->print("--"); 3874 } 3875 g1_policy()->print_heap_transition(); 3876 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3877 } 3878 gclog_or_tty->flush(); 3879 } 3880 3881 bool 3882 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3883 assert_at_safepoint(true /* should_be_vm_thread */); 3884 guarantee(!is_gc_active(), "collection is not reentrant"); 3885 3886 if (GC_locker::check_active_before_gc()) { 3887 return false; 3888 } 3889 3890 _gc_timer_stw->register_gc_start(os::elapsed_counter()); 3891 3892 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start()); 3893 3894 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3895 ResourceMark rm; 3896 3897 print_heap_before_gc(); 3898 trace_heap_before_gc(_gc_tracer_stw); 3899 3900 HRSPhaseSetter x(HRSPhaseEvacuation); 3901 verify_region_sets_optional(); 3902 verify_dirty_young_regions(); 3903 3904 // This call will decide whether this pause is an initial-mark 3905 // pause. If it is, during_initial_mark_pause() will return true 3906 // for the duration of this pause. 3907 g1_policy()->decide_on_conc_mark_initiation(); 3908 3909 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3910 assert(!g1_policy()->during_initial_mark_pause() || 3911 g1_policy()->gcs_are_young(), "sanity"); 3912 3913 // We also do not allow mixed GCs during marking. 3914 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3915 3916 // Record whether this pause is an initial mark. When the current 3917 // thread has completed its logging output and it's safe to signal 3918 // the CM thread, the flag's value in the policy has been reset. 3919 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3920 3921 // Inner scope for scope based logging, timers, and stats collection 3922 { 3923 EvacuationInfo evacuation_info; 3924 3925 if (g1_policy()->during_initial_mark_pause()) { 3926 // We are about to start a marking cycle, so we increment the 3927 // full collection counter. 3928 increment_old_marking_cycles_started(); 3929 register_concurrent_cycle_start(_gc_timer_stw->gc_start()); 3930 } 3931 3932 _gc_tracer_stw->report_yc_type(yc_type()); 3933 3934 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3935 3936 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3937 workers()->active_workers() : 1); 3938 double pause_start_sec = os::elapsedTime(); 3939 g1_policy()->phase_times()->note_gc_start(active_workers); 3940 log_gc_header(); 3941 3942 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3943 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3944 3945 // If the secondary_free_list is not empty, append it to the 3946 // free_list. No need to wait for the cleanup operation to finish; 3947 // the region allocation code will check the secondary_free_list 3948 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3949 // set, skip this step so that the region allocation code has to 3950 // get entries from the secondary_free_list. 3951 if (!G1StressConcRegionFreeing) { 3952 append_secondary_free_list_if_not_empty_with_lock(); 3953 } 3954 3955 assert(check_young_list_well_formed(), "young list should be well formed"); 3956 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3957 "sanity check"); 3958 3959 // Don't dynamically change the number of GC threads this early. A value of 3960 // 0 is used to indicate serial work. When parallel work is done, 3961 // it will be set. 3962 3963 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3964 IsGCActiveMark x; 3965 3966 gc_prologue(false); 3967 increment_total_collections(false /* full gc */); 3968 increment_gc_time_stamp(); 3969 3970 verify_before_gc(); 3971 3972 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3973 3974 // Please see comment in g1CollectedHeap.hpp and 3975 // G1CollectedHeap::ref_processing_init() to see how 3976 // reference processing currently works in G1. 3977 3978 // Enable discovery in the STW reference processor 3979 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3980 true /*verify_no_refs*/); 3981 3982 { 3983 // We want to temporarily turn off discovery by the 3984 // CM ref processor, if necessary, and turn it back on 3985 // on again later if we do. Using a scoped 3986 // NoRefDiscovery object will do this. 3987 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3988 3989 // Forget the current alloc region (we might even choose it to be part 3990 // of the collection set!). 3991 release_mutator_alloc_region(); 3992 3993 // We should call this after we retire the mutator alloc 3994 // region(s) so that all the ALLOC / RETIRE events are generated 3995 // before the start GC event. 3996 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3997 3998 // This timing is only used by the ergonomics to handle our pause target. 3999 // It is unclear why this should not include the full pause. We will 4000 // investigate this in CR 7178365. 4001 // 4002 // Preserving the old comment here if that helps the investigation: 4003 // 4004 // The elapsed time induced by the start time below deliberately elides 4005 // the possible verification above. 4006 double sample_start_time_sec = os::elapsedTime(); 4007 4008 #if YOUNG_LIST_VERBOSE 4009 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 4010 _young_list->print(); 4011 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4012 #endif // YOUNG_LIST_VERBOSE 4013 4014 g1_policy()->record_collection_pause_start(sample_start_time_sec); 4015 4016 double scan_wait_start = os::elapsedTime(); 4017 // We have to wait until the CM threads finish scanning the 4018 // root regions as it's the only way to ensure that all the 4019 // objects on them have been correctly scanned before we start 4020 // moving them during the GC. 4021 bool waited = _cm->root_regions()->wait_until_scan_finished(); 4022 double wait_time_ms = 0.0; 4023 if (waited) { 4024 double scan_wait_end = os::elapsedTime(); 4025 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 4026 } 4027 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 4028 4029 #if YOUNG_LIST_VERBOSE 4030 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 4031 _young_list->print(); 4032 #endif // YOUNG_LIST_VERBOSE 4033 4034 if (g1_policy()->during_initial_mark_pause()) { 4035 concurrent_mark()->checkpointRootsInitialPre(); 4036 } 4037 4038 #if YOUNG_LIST_VERBOSE 4039 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 4040 _young_list->print(); 4041 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4042 #endif // YOUNG_LIST_VERBOSE 4043 4044 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); 4045 4046 _cm->note_start_of_gc(); 4047 // We should not verify the per-thread SATB buffers given that 4048 // we have not filtered them yet (we'll do so during the 4049 // GC). We also call this after finalize_cset() to 4050 // ensure that the CSet has been finalized. 4051 _cm->verify_no_cset_oops(true /* verify_stacks */, 4052 true /* verify_enqueued_buffers */, 4053 false /* verify_thread_buffers */, 4054 true /* verify_fingers */); 4055 4056 if (_hr_printer.is_active()) { 4057 HeapRegion* hr = g1_policy()->collection_set(); 4058 while (hr != NULL) { 4059 G1HRPrinter::RegionType type; 4060 if (!hr->is_young()) { 4061 type = G1HRPrinter::Old; 4062 } else if (hr->is_survivor()) { 4063 type = G1HRPrinter::Survivor; 4064 } else { 4065 type = G1HRPrinter::Eden; 4066 } 4067 _hr_printer.cset(hr); 4068 hr = hr->next_in_collection_set(); 4069 } 4070 } 4071 4072 #ifdef ASSERT 4073 VerifyCSetClosure cl; 4074 collection_set_iterate(&cl); 4075 #endif // ASSERT 4076 4077 setup_surviving_young_words(); 4078 4079 // Initialize the GC alloc regions. 4080 init_gc_alloc_regions(evacuation_info); 4081 4082 // Actually do the work... 4083 evacuate_collection_set(evacuation_info); 4084 4085 // We do this to mainly verify the per-thread SATB buffers 4086 // (which have been filtered by now) since we didn't verify 4087 // them earlier. No point in re-checking the stacks / enqueued 4088 // buffers given that the CSet has not changed since last time 4089 // we checked. 4090 _cm->verify_no_cset_oops(false /* verify_stacks */, 4091 false /* verify_enqueued_buffers */, 4092 true /* verify_thread_buffers */, 4093 true /* verify_fingers */); 4094 4095 free_collection_set(g1_policy()->collection_set(), evacuation_info); 4096 g1_policy()->clear_collection_set(); 4097 4098 cleanup_surviving_young_words(); 4099 4100 // Start a new incremental collection set for the next pause. 4101 g1_policy()->start_incremental_cset_building(); 4102 4103 // Clear the _cset_fast_test bitmap in anticipation of adding 4104 // regions to the incremental collection set for the next 4105 // evacuation pause. 4106 clear_cset_fast_test(); 4107 4108 _young_list->reset_sampled_info(); 4109 4110 // Don't check the whole heap at this point as the 4111 // GC alloc regions from this pause have been tagged 4112 // as survivors and moved on to the survivor list. 4113 // Survivor regions will fail the !is_young() check. 4114 assert(check_young_list_empty(false /* check_heap */), 4115 "young list should be empty"); 4116 4117 #if YOUNG_LIST_VERBOSE 4118 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 4119 _young_list->print(); 4120 #endif // YOUNG_LIST_VERBOSE 4121 4122 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 4123 _young_list->first_survivor_region(), 4124 _young_list->last_survivor_region()); 4125 4126 _young_list->reset_auxilary_lists(); 4127 4128 if (evacuation_failed()) { 4129 _summary_bytes_used = recalculate_used(); 4130 uint n_queues = MAX2((int)ParallelGCThreads, 1); 4131 for (uint i = 0; i < n_queues; i++) { 4132 if (_evacuation_failed_info_array[i].has_failed()) { 4133 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]); 4134 } 4135 } 4136 } else { 4137 // The "used" of the the collection set have already been subtracted 4138 // when they were freed. Add in the bytes evacuated. 4139 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 4140 } 4141 4142 if (g1_policy()->during_initial_mark_pause()) { 4143 // We have to do this before we notify the CM threads that 4144 // they can start working to make sure that all the 4145 // appropriate initialization is done on the CM object. 4146 concurrent_mark()->checkpointRootsInitialPost(); 4147 set_marking_started(); 4148 // Note that we don't actually trigger the CM thread at 4149 // this point. We do that later when we're sure that 4150 // the current thread has completed its logging output. 4151 } 4152 4153 allocate_dummy_regions(); 4154 4155 #if YOUNG_LIST_VERBOSE 4156 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 4157 _young_list->print(); 4158 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4159 #endif // YOUNG_LIST_VERBOSE 4160 4161 init_mutator_alloc_region(); 4162 4163 { 4164 size_t expand_bytes = g1_policy()->expansion_amount(); 4165 if (expand_bytes > 0) { 4166 size_t bytes_before = capacity(); 4167 // No need for an ergo verbose message here, 4168 // expansion_amount() does this when it returns a value > 0. 4169 if (!expand(expand_bytes)) { 4170 // We failed to expand the heap so let's verify that 4171 // committed/uncommitted amount match the backing store 4172 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 4173 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 4174 } 4175 } 4176 } 4177 4178 // We redo the verification but now wrt to the new CSet which 4179 // has just got initialized after the previous CSet was freed. 4180 _cm->verify_no_cset_oops(true /* verify_stacks */, 4181 true /* verify_enqueued_buffers */, 4182 true /* verify_thread_buffers */, 4183 true /* verify_fingers */); 4184 _cm->note_end_of_gc(); 4185 4186 // This timing is only used by the ergonomics to handle our pause target. 4187 // It is unclear why this should not include the full pause. We will 4188 // investigate this in CR 7178365. 4189 double sample_end_time_sec = os::elapsedTime(); 4190 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 4191 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info); 4192 4193 MemoryService::track_memory_usage(); 4194 4195 // In prepare_for_verify() below we'll need to scan the deferred 4196 // update buffers to bring the RSets up-to-date if 4197 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4198 // the update buffers we'll probably need to scan cards on the 4199 // regions we just allocated to (i.e., the GC alloc 4200 // regions). However, during the last GC we called 4201 // set_saved_mark() on all the GC alloc regions, so card 4202 // scanning might skip the [saved_mark_word()...top()] area of 4203 // those regions (i.e., the area we allocated objects into 4204 // during the last GC). But it shouldn't. Given that 4205 // saved_mark_word() is conditional on whether the GC time stamp 4206 // on the region is current or not, by incrementing the GC time 4207 // stamp here we invalidate all the GC time stamps on all the 4208 // regions and saved_mark_word() will simply return top() for 4209 // all the regions. This is a nicer way of ensuring this rather 4210 // than iterating over the regions and fixing them. In fact, the 4211 // GC time stamp increment here also ensures that 4212 // saved_mark_word() will return top() between pauses, i.e., 4213 // during concurrent refinement. So we don't need the 4214 // is_gc_active() check to decided which top to use when 4215 // scanning cards (see CR 7039627). 4216 increment_gc_time_stamp(); 4217 4218 verify_after_gc(); 4219 4220 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4221 ref_processor_stw()->verify_no_references_recorded(); 4222 4223 // CM reference discovery will be re-enabled if necessary. 4224 } 4225 4226 // We should do this after we potentially expand the heap so 4227 // that all the COMMIT events are generated before the end GC 4228 // event, and after we retire the GC alloc regions so that all 4229 // RETIRE events are generated before the end GC event. 4230 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4231 4232 if (mark_in_progress()) { 4233 concurrent_mark()->update_g1_committed(); 4234 } 4235 4236 #ifdef TRACESPINNING 4237 ParallelTaskTerminator::print_termination_counts(); 4238 #endif 4239 4240 gc_epilogue(false); 4241 } 4242 4243 // Print the remainder of the GC log output. 4244 log_gc_footer(os::elapsedTime() - pause_start_sec); 4245 4246 // It is not yet to safe to tell the concurrent mark to 4247 // start as we have some optional output below. We don't want the 4248 // output from the concurrent mark thread interfering with this 4249 // logging output either. 4250 4251 _hrs.verify_optional(); 4252 verify_region_sets_optional(); 4253 4254 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4255 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4256 4257 print_heap_after_gc(); 4258 trace_heap_after_gc(_gc_tracer_stw); 4259 4260 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4261 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4262 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4263 // before any GC notifications are raised. 4264 g1mm()->update_sizes(); 4265 4266 _gc_tracer_stw->report_evacuation_info(&evacuation_info); 4267 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold()); 4268 _gc_timer_stw->register_gc_end(os::elapsed_counter()); 4269 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions()); 4270 } 4271 // It should now be safe to tell the concurrent mark thread to start 4272 // without its logging output interfering with the logging output 4273 // that came from the pause. 4274 4275 if (should_start_conc_mark) { 4276 // CAUTION: after the doConcurrentMark() call below, 4277 // the concurrent marking thread(s) could be running 4278 // concurrently with us. Make sure that anything after 4279 // this point does not assume that we are the only GC thread 4280 // running. Note: of course, the actual marking work will 4281 // not start until the safepoint itself is released in 4282 // ConcurrentGCThread::safepoint_desynchronize(). 4283 doConcurrentMark(); 4284 } 4285 4286 return true; 4287 } 4288 4289 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4290 { 4291 size_t gclab_word_size; 4292 switch (purpose) { 4293 case GCAllocForSurvived: 4294 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4295 break; 4296 case GCAllocForTenured: 4297 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4298 break; 4299 default: 4300 assert(false, "unknown GCAllocPurpose"); 4301 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4302 break; 4303 } 4304 4305 // Prevent humongous PLAB sizes for two reasons: 4306 // * PLABs are allocated using a similar paths as oops, but should 4307 // never be in a humongous region 4308 // * Allowing humongous PLABs needlessly churns the region free lists 4309 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4310 } 4311 4312 void G1CollectedHeap::init_mutator_alloc_region() { 4313 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4314 _mutator_alloc_region.init(); 4315 } 4316 4317 void G1CollectedHeap::release_mutator_alloc_region() { 4318 _mutator_alloc_region.release(); 4319 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4320 } 4321 4322 void G1CollectedHeap::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { 4323 assert_at_safepoint(true /* should_be_vm_thread */); 4324 4325 _survivor_gc_alloc_region.init(); 4326 _old_gc_alloc_region.init(); 4327 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4328 _retained_old_gc_alloc_region = NULL; 4329 4330 // We will discard the current GC alloc region if: 4331 // a) it's in the collection set (it can happen!), 4332 // b) it's already full (no point in using it), 4333 // c) it's empty (this means that it was emptied during 4334 // a cleanup and it should be on the free list now), or 4335 // d) it's humongous (this means that it was emptied 4336 // during a cleanup and was added to the free list, but 4337 // has been subsequently used to allocate a humongous 4338 // object that may be less than the region size). 4339 if (retained_region != NULL && 4340 !retained_region->in_collection_set() && 4341 !(retained_region->top() == retained_region->end()) && 4342 !retained_region->is_empty() && 4343 !retained_region->isHumongous()) { 4344 retained_region->set_saved_mark(); 4345 // The retained region was added to the old region set when it was 4346 // retired. We have to remove it now, since we don't allow regions 4347 // we allocate to in the region sets. We'll re-add it later, when 4348 // it's retired again. 4349 _old_set.remove(retained_region); 4350 bool during_im = g1_policy()->during_initial_mark_pause(); 4351 retained_region->note_start_of_copying(during_im); 4352 _old_gc_alloc_region.set(retained_region); 4353 _hr_printer.reuse(retained_region); 4354 evacuation_info.set_alloc_regions_used_before(retained_region->used()); 4355 } 4356 } 4357 4358 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { 4359 evacuation_info.set_allocation_regions(_survivor_gc_alloc_region.count() + 4360 _old_gc_alloc_region.count()); 4361 _survivor_gc_alloc_region.release(); 4362 // If we have an old GC alloc region to release, we'll save it in 4363 // _retained_old_gc_alloc_region. If we don't 4364 // _retained_old_gc_alloc_region will become NULL. This is what we 4365 // want either way so no reason to check explicitly for either 4366 // condition. 4367 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4368 4369 if (ResizePLAB) { 4370 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4371 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4372 } 4373 } 4374 4375 void G1CollectedHeap::abandon_gc_alloc_regions() { 4376 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4377 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4378 _retained_old_gc_alloc_region = NULL; 4379 } 4380 4381 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4382 _drain_in_progress = false; 4383 set_evac_failure_closure(cl); 4384 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4385 } 4386 4387 void G1CollectedHeap::finalize_for_evac_failure() { 4388 assert(_evac_failure_scan_stack != NULL && 4389 _evac_failure_scan_stack->length() == 0, 4390 "Postcondition"); 4391 assert(!_drain_in_progress, "Postcondition"); 4392 delete _evac_failure_scan_stack; 4393 _evac_failure_scan_stack = NULL; 4394 } 4395 4396 void G1CollectedHeap::remove_self_forwarding_pointers() { 4397 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4398 4399 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4400 4401 if (G1CollectedHeap::use_parallel_gc_threads()) { 4402 set_par_threads(); 4403 workers()->run_task(&rsfp_task); 4404 set_par_threads(0); 4405 } else { 4406 rsfp_task.work(0); 4407 } 4408 4409 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4410 4411 // Reset the claim values in the regions in the collection set. 4412 reset_cset_heap_region_claim_values(); 4413 4414 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4415 4416 // Now restore saved marks, if any. 4417 assert(_objs_with_preserved_marks.size() == 4418 _preserved_marks_of_objs.size(), "Both or none."); 4419 while (!_objs_with_preserved_marks.is_empty()) { 4420 oop obj = _objs_with_preserved_marks.pop(); 4421 markOop m = _preserved_marks_of_objs.pop(); 4422 obj->set_mark(m); 4423 } 4424 _objs_with_preserved_marks.clear(true); 4425 _preserved_marks_of_objs.clear(true); 4426 } 4427 4428 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4429 _evac_failure_scan_stack->push(obj); 4430 } 4431 4432 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4433 assert(_evac_failure_scan_stack != NULL, "precondition"); 4434 4435 while (_evac_failure_scan_stack->length() > 0) { 4436 oop obj = _evac_failure_scan_stack->pop(); 4437 _evac_failure_closure->set_region(heap_region_containing(obj)); 4438 obj->oop_iterate_backwards(_evac_failure_closure); 4439 } 4440 } 4441 4442 oop 4443 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state, 4444 oop old) { 4445 assert(obj_in_cs(old), 4446 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4447 (HeapWord*) old)); 4448 markOop m = old->mark(); 4449 oop forward_ptr = old->forward_to_atomic(old); 4450 if (forward_ptr == NULL) { 4451 // Forward-to-self succeeded. 4452 assert(_par_scan_state != NULL, "par scan state"); 4453 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4454 uint queue_num = _par_scan_state->queue_num(); 4455 4456 _evacuation_failed = true; 4457 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size()); 4458 if (_evac_failure_closure != cl) { 4459 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4460 assert(!_drain_in_progress, 4461 "Should only be true while someone holds the lock."); 4462 // Set the global evac-failure closure to the current thread's. 4463 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4464 set_evac_failure_closure(cl); 4465 // Now do the common part. 4466 handle_evacuation_failure_common(old, m); 4467 // Reset to NULL. 4468 set_evac_failure_closure(NULL); 4469 } else { 4470 // The lock is already held, and this is recursive. 4471 assert(_drain_in_progress, "This should only be the recursive case."); 4472 handle_evacuation_failure_common(old, m); 4473 } 4474 return old; 4475 } else { 4476 // Forward-to-self failed. Either someone else managed to allocate 4477 // space for this object (old != forward_ptr) or they beat us in 4478 // self-forwarding it (old == forward_ptr). 4479 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4480 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4481 "should not be in the CSet", 4482 (HeapWord*) old, (HeapWord*) forward_ptr)); 4483 return forward_ptr; 4484 } 4485 } 4486 4487 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4488 preserve_mark_if_necessary(old, m); 4489 4490 HeapRegion* r = heap_region_containing(old); 4491 if (!r->evacuation_failed()) { 4492 r->set_evacuation_failed(true); 4493 _hr_printer.evac_failure(r); 4494 } 4495 4496 push_on_evac_failure_scan_stack(old); 4497 4498 if (!_drain_in_progress) { 4499 // prevent recursion in copy_to_survivor_space() 4500 _drain_in_progress = true; 4501 drain_evac_failure_scan_stack(); 4502 _drain_in_progress = false; 4503 } 4504 } 4505 4506 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4507 assert(evacuation_failed(), "Oversaving!"); 4508 // We want to call the "for_promotion_failure" version only in the 4509 // case of a promotion failure. 4510 if (m->must_be_preserved_for_promotion_failure(obj)) { 4511 _objs_with_preserved_marks.push(obj); 4512 _preserved_marks_of_objs.push(m); 4513 } 4514 } 4515 4516 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4517 size_t word_size) { 4518 if (purpose == GCAllocForSurvived) { 4519 HeapWord* result = survivor_attempt_allocation(word_size); 4520 if (result != NULL) { 4521 return result; 4522 } else { 4523 // Let's try to allocate in the old gen in case we can fit the 4524 // object there. 4525 return old_attempt_allocation(word_size); 4526 } 4527 } else { 4528 assert(purpose == GCAllocForTenured, "sanity"); 4529 HeapWord* result = old_attempt_allocation(word_size); 4530 if (result != NULL) { 4531 return result; 4532 } else { 4533 // Let's try to allocate in the survivors in case we can fit the 4534 // object there. 4535 return survivor_attempt_allocation(word_size); 4536 } 4537 } 4538 4539 ShouldNotReachHere(); 4540 // Trying to keep some compilers happy. 4541 return NULL; 4542 } 4543 4544 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4545 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4546 4547 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4548 : _g1h(g1h), 4549 _refs(g1h->task_queue(queue_num)), 4550 _dcq(&g1h->dirty_card_queue_set()), 4551 _ct_bs(g1h->g1_barrier_set()), 4552 _g1_rem(g1h->g1_rem_set()), 4553 _hash_seed(17), _queue_num(queue_num), 4554 _term_attempts(0), 4555 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4556 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4557 _age_table(false), 4558 _strong_roots_time(0), _term_time(0), 4559 _alloc_buffer_waste(0), _undo_waste(0) { 4560 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4561 // we "sacrifice" entry 0 to keep track of surviving bytes for 4562 // non-young regions (where the age is -1) 4563 // We also add a few elements at the beginning and at the end in 4564 // an attempt to eliminate cache contention 4565 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4566 uint array_length = PADDING_ELEM_NUM + 4567 real_length + 4568 PADDING_ELEM_NUM; 4569 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4570 if (_surviving_young_words_base == NULL) 4571 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR, 4572 "Not enough space for young surv histo."); 4573 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4574 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4575 4576 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4577 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4578 4579 _start = os::elapsedTime(); 4580 } 4581 4582 void 4583 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4584 { 4585 st->print_raw_cr("GC Termination Stats"); 4586 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4587 " ------waste (KiB)------"); 4588 st->print_raw_cr("thr ms ms % ms % attempts" 4589 " total alloc undo"); 4590 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4591 " ------- ------- -------"); 4592 } 4593 4594 void 4595 G1ParScanThreadState::print_termination_stats(int i, 4596 outputStream* const st) const 4597 { 4598 const double elapsed_ms = elapsed_time() * 1000.0; 4599 const double s_roots_ms = strong_roots_time() * 1000.0; 4600 const double term_ms = term_time() * 1000.0; 4601 st->print_cr("%3d %9.2f %9.2f %6.2f " 4602 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4603 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4604 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4605 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4606 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4607 alloc_buffer_waste() * HeapWordSize / K, 4608 undo_waste() * HeapWordSize / K); 4609 } 4610 4611 #ifdef ASSERT 4612 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4613 assert(ref != NULL, "invariant"); 4614 assert(UseCompressedOops, "sanity"); 4615 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4616 oop p = oopDesc::load_decode_heap_oop(ref); 4617 assert(_g1h->is_in_g1_reserved(p), 4618 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4619 return true; 4620 } 4621 4622 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4623 assert(ref != NULL, "invariant"); 4624 if (has_partial_array_mask(ref)) { 4625 // Must be in the collection set--it's already been copied. 4626 oop p = clear_partial_array_mask(ref); 4627 assert(_g1h->obj_in_cs(p), 4628 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4629 } else { 4630 oop p = oopDesc::load_decode_heap_oop(ref); 4631 assert(_g1h->is_in_g1_reserved(p), 4632 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, (void *)p)); 4633 } 4634 return true; 4635 } 4636 4637 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4638 if (ref.is_narrow()) { 4639 return verify_ref((narrowOop*) ref); 4640 } else { 4641 return verify_ref((oop*) ref); 4642 } 4643 } 4644 #endif // ASSERT 4645 4646 void G1ParScanThreadState::trim_queue() { 4647 assert(_evac_cl != NULL, "not set"); 4648 assert(_evac_failure_cl != NULL, "not set"); 4649 assert(_partial_scan_cl != NULL, "not set"); 4650 4651 StarTask ref; 4652 do { 4653 // Drain the overflow stack first, so other threads can steal. 4654 while (refs()->pop_overflow(ref)) { 4655 deal_with_reference(ref); 4656 } 4657 4658 while (refs()->pop_local(ref)) { 4659 deal_with_reference(ref); 4660 } 4661 } while (!refs()->is_empty()); 4662 } 4663 4664 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4665 G1ParScanThreadState* par_scan_state) : 4666 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4667 _par_scan_state(par_scan_state), 4668 _worker_id(par_scan_state->queue_num()), 4669 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4670 _mark_in_progress(_g1->mark_in_progress()) { } 4671 4672 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4673 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4674 #ifdef ASSERT 4675 HeapRegion* hr = _g1->heap_region_containing(obj); 4676 assert(hr != NULL, "sanity"); 4677 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4678 #endif // ASSERT 4679 4680 // We know that the object is not moving so it's safe to read its size. 4681 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4682 } 4683 4684 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4685 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4686 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4687 #ifdef ASSERT 4688 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4689 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4690 assert(from_obj != to_obj, "should not be self-forwarded"); 4691 4692 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4693 assert(from_hr != NULL, "sanity"); 4694 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4695 4696 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4697 assert(to_hr != NULL, "sanity"); 4698 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4699 #endif // ASSERT 4700 4701 // The object might be in the process of being copied by another 4702 // worker so we cannot trust that its to-space image is 4703 // well-formed. So we have to read its size from its from-space 4704 // image which we know should not be changing. 4705 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4706 } 4707 4708 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4709 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4710 ::copy_to_survivor_space(oop old) { 4711 size_t word_sz = old->size(); 4712 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4713 // +1 to make the -1 indexes valid... 4714 int young_index = from_region->young_index_in_cset()+1; 4715 assert( (from_region->is_young() && young_index > 0) || 4716 (!from_region->is_young() && young_index == 0), "invariant" ); 4717 G1CollectorPolicy* g1p = _g1->g1_policy(); 4718 markOop m = old->mark(); 4719 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4720 : m->age(); 4721 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4722 word_sz); 4723 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4724 #ifndef PRODUCT 4725 // Should this evacuation fail? 4726 if (_g1->evacuation_should_fail()) { 4727 if (obj_ptr != NULL) { 4728 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4729 obj_ptr = NULL; 4730 } 4731 } 4732 #endif // !PRODUCT 4733 4734 if (obj_ptr == NULL) { 4735 // This will either forward-to-self, or detect that someone else has 4736 // installed a forwarding pointer. 4737 return _g1->handle_evacuation_failure_par(_par_scan_state, old); 4738 } 4739 4740 oop obj = oop(obj_ptr); 4741 4742 // We're going to allocate linearly, so might as well prefetch ahead. 4743 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4744 4745 oop forward_ptr = old->forward_to_atomic(obj); 4746 if (forward_ptr == NULL) { 4747 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4748 if (g1p->track_object_age(alloc_purpose)) { 4749 // We could simply do obj->incr_age(). However, this causes a 4750 // performance issue. obj->incr_age() will first check whether 4751 // the object has a displaced mark by checking its mark word; 4752 // getting the mark word from the new location of the object 4753 // stalls. So, given that we already have the mark word and we 4754 // are about to install it anyway, it's better to increase the 4755 // age on the mark word, when the object does not have a 4756 // displaced mark word. We're not expecting many objects to have 4757 // a displaced marked word, so that case is not optimized 4758 // further (it could be...) and we simply call obj->incr_age(). 4759 4760 if (m->has_displaced_mark_helper()) { 4761 // in this case, we have to install the mark word first, 4762 // otherwise obj looks to be forwarded (the old mark word, 4763 // which contains the forward pointer, was copied) 4764 obj->set_mark(m); 4765 obj->incr_age(); 4766 } else { 4767 m = m->incr_age(); 4768 obj->set_mark(m); 4769 } 4770 _par_scan_state->age_table()->add(obj, word_sz); 4771 } else { 4772 obj->set_mark(m); 4773 } 4774 4775 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4776 surv_young_words[young_index] += word_sz; 4777 4778 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4779 // We keep track of the next start index in the length field of 4780 // the to-space object. The actual length can be found in the 4781 // length field of the from-space object. 4782 arrayOop(obj)->set_length(0); 4783 oop* old_p = set_partial_array_mask(old); 4784 _par_scan_state->push_on_queue(old_p); 4785 } else { 4786 // No point in using the slower heap_region_containing() method, 4787 // given that we know obj is in the heap. 4788 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4789 obj->oop_iterate_backwards(&_scanner); 4790 } 4791 } else { 4792 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4793 obj = forward_ptr; 4794 } 4795 return obj; 4796 } 4797 4798 template <class T> 4799 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4800 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4801 _scanned_klass->record_modified_oops(); 4802 } 4803 } 4804 4805 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4806 template <class T> 4807 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4808 ::do_oop_work(T* p) { 4809 oop obj = oopDesc::load_decode_heap_oop(p); 4810 assert(barrier != G1BarrierRS || obj != NULL, 4811 "Precondition: G1BarrierRS implies obj is non-NULL"); 4812 4813 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4814 4815 // here the null check is implicit in the cset_fast_test() test 4816 if (_g1->in_cset_fast_test(obj)) { 4817 oop forwardee; 4818 if (obj->is_forwarded()) { 4819 forwardee = obj->forwardee(); 4820 } else { 4821 forwardee = copy_to_survivor_space(obj); 4822 } 4823 assert(forwardee != NULL, "forwardee should not be NULL"); 4824 oopDesc::encode_store_heap_oop(p, forwardee); 4825 if (do_mark_object && forwardee != obj) { 4826 // If the object is self-forwarded we don't need to explicitly 4827 // mark it, the evacuation failure protocol will do so. 4828 mark_forwarded_object(obj, forwardee); 4829 } 4830 4831 // When scanning the RS, we only care about objs in CS. 4832 if (barrier == G1BarrierRS) { 4833 _par_scan_state->update_rs(_from, p, _worker_id); 4834 } else if (barrier == G1BarrierKlass) { 4835 do_klass_barrier(p, forwardee); 4836 } 4837 } else { 4838 // The object is not in collection set. If we're a root scanning 4839 // closure during an initial mark pause (i.e. do_mark_object will 4840 // be true) then attempt to mark the object. 4841 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4842 mark_object(obj); 4843 } 4844 } 4845 4846 if (barrier == G1BarrierEvac && obj != NULL) { 4847 _par_scan_state->update_rs(_from, p, _worker_id); 4848 } 4849 4850 if (do_gen_barrier && obj != NULL) { 4851 par_do_barrier(p); 4852 } 4853 } 4854 4855 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4856 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4857 4858 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4859 assert(has_partial_array_mask(p), "invariant"); 4860 oop from_obj = clear_partial_array_mask(p); 4861 4862 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4863 assert(from_obj->is_objArray(), "must be obj array"); 4864 objArrayOop from_obj_array = objArrayOop(from_obj); 4865 // The from-space object contains the real length. 4866 int length = from_obj_array->length(); 4867 4868 assert(from_obj->is_forwarded(), "must be forwarded"); 4869 oop to_obj = from_obj->forwardee(); 4870 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4871 objArrayOop to_obj_array = objArrayOop(to_obj); 4872 // We keep track of the next start index in the length field of the 4873 // to-space object. 4874 int next_index = to_obj_array->length(); 4875 assert(0 <= next_index && next_index < length, 4876 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4877 4878 int start = next_index; 4879 int end = length; 4880 int remainder = end - start; 4881 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4882 if (remainder > 2 * ParGCArrayScanChunk) { 4883 end = start + ParGCArrayScanChunk; 4884 to_obj_array->set_length(end); 4885 // Push the remainder before we process the range in case another 4886 // worker has run out of things to do and can steal it. 4887 oop* from_obj_p = set_partial_array_mask(from_obj); 4888 _par_scan_state->push_on_queue(from_obj_p); 4889 } else { 4890 assert(length == end, "sanity"); 4891 // We'll process the final range for this object. Restore the length 4892 // so that the heap remains parsable in case of evacuation failure. 4893 to_obj_array->set_length(end); 4894 } 4895 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4896 // Process indexes [start,end). It will also process the header 4897 // along with the first chunk (i.e., the chunk with start == 0). 4898 // Note that at this point the length field of to_obj_array is not 4899 // correct given that we are using it to keep track of the next 4900 // start index. oop_iterate_range() (thankfully!) ignores the length 4901 // field and only relies on the start / end parameters. It does 4902 // however return the size of the object which will be incorrect. So 4903 // we have to ignore it even if we wanted to use it. 4904 to_obj_array->oop_iterate_range(&_scanner, start, end); 4905 } 4906 4907 class G1ParEvacuateFollowersClosure : public VoidClosure { 4908 protected: 4909 G1CollectedHeap* _g1h; 4910 G1ParScanThreadState* _par_scan_state; 4911 RefToScanQueueSet* _queues; 4912 ParallelTaskTerminator* _terminator; 4913 4914 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4915 RefToScanQueueSet* queues() { return _queues; } 4916 ParallelTaskTerminator* terminator() { return _terminator; } 4917 4918 public: 4919 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4920 G1ParScanThreadState* par_scan_state, 4921 RefToScanQueueSet* queues, 4922 ParallelTaskTerminator* terminator) 4923 : _g1h(g1h), _par_scan_state(par_scan_state), 4924 _queues(queues), _terminator(terminator) {} 4925 4926 void do_void(); 4927 4928 private: 4929 inline bool offer_termination(); 4930 }; 4931 4932 bool G1ParEvacuateFollowersClosure::offer_termination() { 4933 G1ParScanThreadState* const pss = par_scan_state(); 4934 pss->start_term_time(); 4935 const bool res = terminator()->offer_termination(); 4936 pss->end_term_time(); 4937 return res; 4938 } 4939 4940 void G1ParEvacuateFollowersClosure::do_void() { 4941 StarTask stolen_task; 4942 G1ParScanThreadState* const pss = par_scan_state(); 4943 pss->trim_queue(); 4944 4945 do { 4946 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4947 assert(pss->verify_task(stolen_task), "sanity"); 4948 if (stolen_task.is_narrow()) { 4949 pss->deal_with_reference((narrowOop*) stolen_task); 4950 } else { 4951 pss->deal_with_reference((oop*) stolen_task); 4952 } 4953 4954 // We've just processed a reference and we might have made 4955 // available new entries on the queues. So we have to make sure 4956 // we drain the queues as necessary. 4957 pss->trim_queue(); 4958 } 4959 } while (!offer_termination()); 4960 4961 pss->retire_alloc_buffers(); 4962 } 4963 4964 class G1KlassScanClosure : public KlassClosure { 4965 G1ParCopyHelper* _closure; 4966 bool _process_only_dirty; 4967 int _count; 4968 public: 4969 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4970 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4971 void do_klass(Klass* klass) { 4972 // If the klass has not been dirtied we know that there's 4973 // no references into the young gen and we can skip it. 4974 if (!_process_only_dirty || klass->has_modified_oops()) { 4975 // Clean the klass since we're going to scavenge all the metadata. 4976 klass->clear_modified_oops(); 4977 4978 // Tell the closure that this klass is the Klass to scavenge 4979 // and is the one to dirty if oops are left pointing into the young gen. 4980 _closure->set_scanned_klass(klass); 4981 4982 klass->oops_do(_closure); 4983 4984 _closure->set_scanned_klass(NULL); 4985 } 4986 _count++; 4987 } 4988 }; 4989 4990 class G1ParTask : public AbstractGangTask { 4991 protected: 4992 G1CollectedHeap* _g1h; 4993 RefToScanQueueSet *_queues; 4994 ParallelTaskTerminator _terminator; 4995 uint _n_workers; 4996 4997 Mutex _stats_lock; 4998 Mutex* stats_lock() { return &_stats_lock; } 4999 5000 size_t getNCards() { 5001 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 5002 / G1BlockOffsetSharedArray::N_bytes; 5003 } 5004 5005 public: 5006 G1ParTask(G1CollectedHeap* g1h, 5007 RefToScanQueueSet *task_queues) 5008 : AbstractGangTask("G1 collection"), 5009 _g1h(g1h), 5010 _queues(task_queues), 5011 _terminator(0, _queues), 5012 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 5013 {} 5014 5015 RefToScanQueueSet* queues() { return _queues; } 5016 5017 RefToScanQueue *work_queue(int i) { 5018 return queues()->queue(i); 5019 } 5020 5021 ParallelTaskTerminator* terminator() { return &_terminator; } 5022 5023 virtual void set_for_termination(int active_workers) { 5024 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 5025 // in the young space (_par_seq_tasks) in the G1 heap 5026 // for SequentialSubTasksDone. 5027 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 5028 // both of which need setting by set_n_termination(). 5029 _g1h->SharedHeap::set_n_termination(active_workers); 5030 _g1h->set_n_termination(active_workers); 5031 terminator()->reset_for_reuse(active_workers); 5032 _n_workers = active_workers; 5033 } 5034 5035 void work(uint worker_id) { 5036 if (worker_id >= _n_workers) return; // no work needed this round 5037 5038 double start_time_ms = os::elapsedTime() * 1000.0; 5039 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 5040 5041 { 5042 ResourceMark rm; 5043 HandleMark hm; 5044 5045 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 5046 5047 G1ParScanThreadState pss(_g1h, worker_id); 5048 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 5049 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 5050 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 5051 5052 pss.set_evac_closure(&scan_evac_cl); 5053 pss.set_evac_failure_closure(&evac_failure_cl); 5054 pss.set_partial_scan_closure(&partial_scan_cl); 5055 5056 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 5057 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 5058 5059 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 5060 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 5061 5062 bool only_young = _g1h->g1_policy()->gcs_are_young(); 5063 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 5064 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 5065 5066 OopClosure* scan_root_cl = &only_scan_root_cl; 5067 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 5068 5069 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5070 // We also need to mark copied objects. 5071 scan_root_cl = &scan_mark_root_cl; 5072 scan_klasses_cl = &scan_mark_klasses_cl_s; 5073 } 5074 5075 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 5076 5077 // Don't scan the scavengable methods in the code cache as part 5078 // of strong root scanning. The code roots that point into a 5079 // region in the collection set are scanned when we scan the 5080 // region's RSet. 5081 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings; 5082 5083 pss.start_strong_roots(); 5084 _g1h->g1_process_strong_roots(/* is scavenging */ true, 5085 SharedHeap::ScanningOption(so), 5086 scan_root_cl, 5087 &push_heap_rs_cl, 5088 scan_klasses_cl, 5089 worker_id); 5090 pss.end_strong_roots(); 5091 5092 { 5093 double start = os::elapsedTime(); 5094 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 5095 evac.do_void(); 5096 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 5097 double term_ms = pss.term_time()*1000.0; 5098 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 5099 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 5100 } 5101 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 5102 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 5103 5104 if (ParallelGCVerbose) { 5105 MutexLocker x(stats_lock()); 5106 pss.print_termination_stats(worker_id); 5107 } 5108 5109 assert(pss.refs()->is_empty(), "should be empty"); 5110 5111 // Close the inner scope so that the ResourceMark and HandleMark 5112 // destructors are executed here and are included as part of the 5113 // "GC Worker Time". 5114 } 5115 5116 double end_time_ms = os::elapsedTime() * 1000.0; 5117 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 5118 } 5119 }; 5120 5121 // *** Common G1 Evacuation Stuff 5122 5123 // This method is run in a GC worker. 5124 5125 void 5126 G1CollectedHeap:: 5127 g1_process_strong_roots(bool is_scavenging, 5128 ScanningOption so, 5129 OopClosure* scan_non_heap_roots, 5130 OopsInHeapRegionClosure* scan_rs, 5131 G1KlassScanClosure* scan_klasses, 5132 int worker_i) { 5133 5134 // First scan the strong roots 5135 double ext_roots_start = os::elapsedTime(); 5136 double closure_app_time_sec = 0.0; 5137 5138 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 5139 5140 assert(so & SO_CodeCache || scan_rs != NULL, "must scan code roots somehow"); 5141 // Walk the code cache/strong code roots w/o buffering, because StarTask 5142 // cannot handle unaligned oop locations. 5143 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, true /* do_marking */); 5144 5145 process_strong_roots(false, // no scoping; this is parallel code 5146 is_scavenging, so, 5147 &buf_scan_non_heap_roots, 5148 &eager_scan_code_roots, 5149 scan_klasses 5150 ); 5151 5152 // Now the CM ref_processor roots. 5153 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5154 // We need to treat the discovered reference lists of the 5155 // concurrent mark ref processor as roots and keep entries 5156 // (which are added by the marking threads) on them live 5157 // until they can be processed at the end of marking. 5158 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5159 } 5160 5161 // Finish up any enqueued closure apps (attributed as object copy time). 5162 buf_scan_non_heap_roots.done(); 5163 5164 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5165 5166 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5167 5168 double ext_root_time_ms = 5169 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5170 5171 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5172 5173 // During conc marking we have to filter the per-thread SATB buffers 5174 // to make sure we remove any oops into the CSet (which will show up 5175 // as implicitly live). 5176 double satb_filtering_ms = 0.0; 5177 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5178 if (mark_in_progress()) { 5179 double satb_filter_start = os::elapsedTime(); 5180 5181 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5182 5183 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5184 } 5185 } 5186 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5187 5188 // If this is an initial mark pause, and we're not scanning 5189 // the entire code cache, we need to mark the oops in the 5190 // strong code root lists for the regions that are not in 5191 // the collection set. 5192 // Note all threads participate in this set of root tasks. 5193 double mark_strong_code_roots_ms = 0.0; 5194 if (g1_policy()->during_initial_mark_pause() && !(so & SO_CodeCache)) { 5195 double mark_strong_roots_start = os::elapsedTime(); 5196 mark_strong_code_roots(worker_i); 5197 mark_strong_code_roots_ms = (os::elapsedTime() - mark_strong_roots_start) * 1000.0; 5198 } 5199 g1_policy()->phase_times()->record_strong_code_root_mark_time(worker_i, mark_strong_code_roots_ms); 5200 5201 // Now scan the complement of the collection set. 5202 if (scan_rs != NULL) { 5203 g1_rem_set()->oops_into_collection_set_do(scan_rs, &eager_scan_code_roots, worker_i); 5204 } 5205 _process_strong_tasks->all_tasks_completed(); 5206 } 5207 5208 void 5209 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) { 5210 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5211 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs); 5212 } 5213 5214 // Weak Reference Processing support 5215 5216 // An always "is_alive" closure that is used to preserve referents. 5217 // If the object is non-null then it's alive. Used in the preservation 5218 // of referent objects that are pointed to by reference objects 5219 // discovered by the CM ref processor. 5220 class G1AlwaysAliveClosure: public BoolObjectClosure { 5221 G1CollectedHeap* _g1; 5222 public: 5223 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5224 bool do_object_b(oop p) { 5225 if (p != NULL) { 5226 return true; 5227 } 5228 return false; 5229 } 5230 }; 5231 5232 bool G1STWIsAliveClosure::do_object_b(oop p) { 5233 // An object is reachable if it is outside the collection set, 5234 // or is inside and copied. 5235 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5236 } 5237 5238 // Non Copying Keep Alive closure 5239 class G1KeepAliveClosure: public OopClosure { 5240 G1CollectedHeap* _g1; 5241 public: 5242 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5243 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5244 void do_oop( oop* p) { 5245 oop obj = *p; 5246 5247 if (_g1->obj_in_cs(obj)) { 5248 assert( obj->is_forwarded(), "invariant" ); 5249 *p = obj->forwardee(); 5250 } 5251 } 5252 }; 5253 5254 // Copying Keep Alive closure - can be called from both 5255 // serial and parallel code as long as different worker 5256 // threads utilize different G1ParScanThreadState instances 5257 // and different queues. 5258 5259 class G1CopyingKeepAliveClosure: public OopClosure { 5260 G1CollectedHeap* _g1h; 5261 OopClosure* _copy_non_heap_obj_cl; 5262 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5263 G1ParScanThreadState* _par_scan_state; 5264 5265 public: 5266 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5267 OopClosure* non_heap_obj_cl, 5268 OopsInHeapRegionClosure* metadata_obj_cl, 5269 G1ParScanThreadState* pss): 5270 _g1h(g1h), 5271 _copy_non_heap_obj_cl(non_heap_obj_cl), 5272 _copy_metadata_obj_cl(metadata_obj_cl), 5273 _par_scan_state(pss) 5274 {} 5275 5276 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5277 virtual void do_oop( oop* p) { do_oop_work(p); } 5278 5279 template <class T> void do_oop_work(T* p) { 5280 oop obj = oopDesc::load_decode_heap_oop(p); 5281 5282 if (_g1h->obj_in_cs(obj)) { 5283 // If the referent object has been forwarded (either copied 5284 // to a new location or to itself in the event of an 5285 // evacuation failure) then we need to update the reference 5286 // field and, if both reference and referent are in the G1 5287 // heap, update the RSet for the referent. 5288 // 5289 // If the referent has not been forwarded then we have to keep 5290 // it alive by policy. Therefore we have copy the referent. 5291 // 5292 // If the reference field is in the G1 heap then we can push 5293 // on the PSS queue. When the queue is drained (after each 5294 // phase of reference processing) the object and it's followers 5295 // will be copied, the reference field set to point to the 5296 // new location, and the RSet updated. Otherwise we need to 5297 // use the the non-heap or metadata closures directly to copy 5298 // the referent object and update the pointer, while avoiding 5299 // updating the RSet. 5300 5301 if (_g1h->is_in_g1_reserved(p)) { 5302 _par_scan_state->push_on_queue(p); 5303 } else { 5304 assert(!ClassLoaderDataGraph::contains((address)p), 5305 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5306 PTR_FORMAT, p)); 5307 _copy_non_heap_obj_cl->do_oop(p); 5308 } 5309 } 5310 } 5311 }; 5312 5313 // Serial drain queue closure. Called as the 'complete_gc' 5314 // closure for each discovered list in some of the 5315 // reference processing phases. 5316 5317 class G1STWDrainQueueClosure: public VoidClosure { 5318 protected: 5319 G1CollectedHeap* _g1h; 5320 G1ParScanThreadState* _par_scan_state; 5321 5322 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5323 5324 public: 5325 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5326 _g1h(g1h), 5327 _par_scan_state(pss) 5328 { } 5329 5330 void do_void() { 5331 G1ParScanThreadState* const pss = par_scan_state(); 5332 pss->trim_queue(); 5333 } 5334 }; 5335 5336 // Parallel Reference Processing closures 5337 5338 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5339 // processing during G1 evacuation pauses. 5340 5341 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5342 private: 5343 G1CollectedHeap* _g1h; 5344 RefToScanQueueSet* _queues; 5345 FlexibleWorkGang* _workers; 5346 int _active_workers; 5347 5348 public: 5349 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5350 FlexibleWorkGang* workers, 5351 RefToScanQueueSet *task_queues, 5352 int n_workers) : 5353 _g1h(g1h), 5354 _queues(task_queues), 5355 _workers(workers), 5356 _active_workers(n_workers) 5357 { 5358 assert(n_workers > 0, "shouldn't call this otherwise"); 5359 } 5360 5361 // Executes the given task using concurrent marking worker threads. 5362 virtual void execute(ProcessTask& task); 5363 virtual void execute(EnqueueTask& task); 5364 }; 5365 5366 // Gang task for possibly parallel reference processing 5367 5368 class G1STWRefProcTaskProxy: public AbstractGangTask { 5369 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5370 ProcessTask& _proc_task; 5371 G1CollectedHeap* _g1h; 5372 RefToScanQueueSet *_task_queues; 5373 ParallelTaskTerminator* _terminator; 5374 5375 public: 5376 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5377 G1CollectedHeap* g1h, 5378 RefToScanQueueSet *task_queues, 5379 ParallelTaskTerminator* terminator) : 5380 AbstractGangTask("Process reference objects in parallel"), 5381 _proc_task(proc_task), 5382 _g1h(g1h), 5383 _task_queues(task_queues), 5384 _terminator(terminator) 5385 {} 5386 5387 virtual void work(uint worker_id) { 5388 // The reference processing task executed by a single worker. 5389 ResourceMark rm; 5390 HandleMark hm; 5391 5392 G1STWIsAliveClosure is_alive(_g1h); 5393 5394 G1ParScanThreadState pss(_g1h, worker_id); 5395 5396 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5397 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5398 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5399 5400 pss.set_evac_closure(&scan_evac_cl); 5401 pss.set_evac_failure_closure(&evac_failure_cl); 5402 pss.set_partial_scan_closure(&partial_scan_cl); 5403 5404 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5405 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5406 5407 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5408 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5409 5410 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5411 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5412 5413 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5414 // We also need to mark copied objects. 5415 copy_non_heap_cl = ©_mark_non_heap_cl; 5416 copy_metadata_cl = ©_mark_metadata_cl; 5417 } 5418 5419 // Keep alive closure. 5420 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5421 5422 // Complete GC closure 5423 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5424 5425 // Call the reference processing task's work routine. 5426 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5427 5428 // Note we cannot assert that the refs array is empty here as not all 5429 // of the processing tasks (specifically phase2 - pp2_work) execute 5430 // the complete_gc closure (which ordinarily would drain the queue) so 5431 // the queue may not be empty. 5432 } 5433 }; 5434 5435 // Driver routine for parallel reference processing. 5436 // Creates an instance of the ref processing gang 5437 // task and has the worker threads execute it. 5438 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5439 assert(_workers != NULL, "Need parallel worker threads."); 5440 5441 ParallelTaskTerminator terminator(_active_workers, _queues); 5442 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5443 5444 _g1h->set_par_threads(_active_workers); 5445 _workers->run_task(&proc_task_proxy); 5446 _g1h->set_par_threads(0); 5447 } 5448 5449 // Gang task for parallel reference enqueueing. 5450 5451 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5452 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5453 EnqueueTask& _enq_task; 5454 5455 public: 5456 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5457 AbstractGangTask("Enqueue reference objects in parallel"), 5458 _enq_task(enq_task) 5459 { } 5460 5461 virtual void work(uint worker_id) { 5462 _enq_task.work(worker_id); 5463 } 5464 }; 5465 5466 // Driver routine for parallel reference enqueueing. 5467 // Creates an instance of the ref enqueueing gang 5468 // task and has the worker threads execute it. 5469 5470 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5471 assert(_workers != NULL, "Need parallel worker threads."); 5472 5473 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5474 5475 _g1h->set_par_threads(_active_workers); 5476 _workers->run_task(&enq_task_proxy); 5477 _g1h->set_par_threads(0); 5478 } 5479 5480 // End of weak reference support closures 5481 5482 // Abstract task used to preserve (i.e. copy) any referent objects 5483 // that are in the collection set and are pointed to by reference 5484 // objects discovered by the CM ref processor. 5485 5486 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5487 protected: 5488 G1CollectedHeap* _g1h; 5489 RefToScanQueueSet *_queues; 5490 ParallelTaskTerminator _terminator; 5491 uint _n_workers; 5492 5493 public: 5494 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5495 AbstractGangTask("ParPreserveCMReferents"), 5496 _g1h(g1h), 5497 _queues(task_queues), 5498 _terminator(workers, _queues), 5499 _n_workers(workers) 5500 { } 5501 5502 void work(uint worker_id) { 5503 ResourceMark rm; 5504 HandleMark hm; 5505 5506 G1ParScanThreadState pss(_g1h, worker_id); 5507 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5508 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5509 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5510 5511 pss.set_evac_closure(&scan_evac_cl); 5512 pss.set_evac_failure_closure(&evac_failure_cl); 5513 pss.set_partial_scan_closure(&partial_scan_cl); 5514 5515 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5516 5517 5518 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5519 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5520 5521 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5522 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5523 5524 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5525 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5526 5527 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5528 // We also need to mark copied objects. 5529 copy_non_heap_cl = ©_mark_non_heap_cl; 5530 copy_metadata_cl = ©_mark_metadata_cl; 5531 } 5532 5533 // Is alive closure 5534 G1AlwaysAliveClosure always_alive(_g1h); 5535 5536 // Copying keep alive closure. Applied to referent objects that need 5537 // to be copied. 5538 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5539 5540 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5541 5542 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5543 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5544 5545 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5546 // So this must be true - but assert just in case someone decides to 5547 // change the worker ids. 5548 assert(0 <= worker_id && worker_id < limit, "sanity"); 5549 assert(!rp->discovery_is_atomic(), "check this code"); 5550 5551 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5552 for (uint idx = worker_id; idx < limit; idx += stride) { 5553 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5554 5555 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5556 while (iter.has_next()) { 5557 // Since discovery is not atomic for the CM ref processor, we 5558 // can see some null referent objects. 5559 iter.load_ptrs(DEBUG_ONLY(true)); 5560 oop ref = iter.obj(); 5561 5562 // This will filter nulls. 5563 if (iter.is_referent_alive()) { 5564 iter.make_referent_alive(); 5565 } 5566 iter.move_to_next(); 5567 } 5568 } 5569 5570 // Drain the queue - which may cause stealing 5571 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5572 drain_queue.do_void(); 5573 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5574 assert(pss.refs()->is_empty(), "should be"); 5575 } 5576 }; 5577 5578 // Weak Reference processing during an evacuation pause (part 1). 5579 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5580 double ref_proc_start = os::elapsedTime(); 5581 5582 ReferenceProcessor* rp = _ref_processor_stw; 5583 assert(rp->discovery_enabled(), "should have been enabled"); 5584 5585 // Any reference objects, in the collection set, that were 'discovered' 5586 // by the CM ref processor should have already been copied (either by 5587 // applying the external root copy closure to the discovered lists, or 5588 // by following an RSet entry). 5589 // 5590 // But some of the referents, that are in the collection set, that these 5591 // reference objects point to may not have been copied: the STW ref 5592 // processor would have seen that the reference object had already 5593 // been 'discovered' and would have skipped discovering the reference, 5594 // but would not have treated the reference object as a regular oop. 5595 // As a result the copy closure would not have been applied to the 5596 // referent object. 5597 // 5598 // We need to explicitly copy these referent objects - the references 5599 // will be processed at the end of remarking. 5600 // 5601 // We also need to do this copying before we process the reference 5602 // objects discovered by the STW ref processor in case one of these 5603 // referents points to another object which is also referenced by an 5604 // object discovered by the STW ref processor. 5605 5606 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5607 no_of_gc_workers == workers()->active_workers(), 5608 "Need to reset active GC workers"); 5609 5610 set_par_threads(no_of_gc_workers); 5611 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5612 no_of_gc_workers, 5613 _task_queues); 5614 5615 if (G1CollectedHeap::use_parallel_gc_threads()) { 5616 workers()->run_task(&keep_cm_referents); 5617 } else { 5618 keep_cm_referents.work(0); 5619 } 5620 5621 set_par_threads(0); 5622 5623 // Closure to test whether a referent is alive. 5624 G1STWIsAliveClosure is_alive(this); 5625 5626 // Even when parallel reference processing is enabled, the processing 5627 // of JNI refs is serial and performed serially by the current thread 5628 // rather than by a worker. The following PSS will be used for processing 5629 // JNI refs. 5630 5631 // Use only a single queue for this PSS. 5632 G1ParScanThreadState pss(this, 0); 5633 5634 // We do not embed a reference processor in the copying/scanning 5635 // closures while we're actually processing the discovered 5636 // reference objects. 5637 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5638 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5639 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5640 5641 pss.set_evac_closure(&scan_evac_cl); 5642 pss.set_evac_failure_closure(&evac_failure_cl); 5643 pss.set_partial_scan_closure(&partial_scan_cl); 5644 5645 assert(pss.refs()->is_empty(), "pre-condition"); 5646 5647 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5648 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5649 5650 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5651 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5652 5653 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5654 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5655 5656 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5657 // We also need to mark copied objects. 5658 copy_non_heap_cl = ©_mark_non_heap_cl; 5659 copy_metadata_cl = ©_mark_metadata_cl; 5660 } 5661 5662 // Keep alive closure. 5663 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5664 5665 // Serial Complete GC closure 5666 G1STWDrainQueueClosure drain_queue(this, &pss); 5667 5668 // Setup the soft refs policy... 5669 rp->setup_policy(false); 5670 5671 ReferenceProcessorStats stats; 5672 if (!rp->processing_is_mt()) { 5673 // Serial reference processing... 5674 stats = rp->process_discovered_references(&is_alive, 5675 &keep_alive, 5676 &drain_queue, 5677 NULL, 5678 _gc_timer_stw); 5679 } else { 5680 // Parallel reference processing 5681 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5682 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5683 5684 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5685 stats = rp->process_discovered_references(&is_alive, 5686 &keep_alive, 5687 &drain_queue, 5688 &par_task_executor, 5689 _gc_timer_stw); 5690 } 5691 5692 _gc_tracer_stw->report_gc_reference_stats(stats); 5693 // We have completed copying any necessary live referent objects 5694 // (that were not copied during the actual pause) so we can 5695 // retire any active alloc buffers 5696 pss.retire_alloc_buffers(); 5697 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5698 5699 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5700 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5701 } 5702 5703 // Weak Reference processing during an evacuation pause (part 2). 5704 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5705 double ref_enq_start = os::elapsedTime(); 5706 5707 ReferenceProcessor* rp = _ref_processor_stw; 5708 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5709 5710 // Now enqueue any remaining on the discovered lists on to 5711 // the pending list. 5712 if (!rp->processing_is_mt()) { 5713 // Serial reference processing... 5714 rp->enqueue_discovered_references(); 5715 } else { 5716 // Parallel reference enqueueing 5717 5718 assert(no_of_gc_workers == workers()->active_workers(), 5719 "Need to reset active workers"); 5720 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5721 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5722 5723 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5724 rp->enqueue_discovered_references(&par_task_executor); 5725 } 5726 5727 rp->verify_no_references_recorded(); 5728 assert(!rp->discovery_enabled(), "should have been disabled"); 5729 5730 // FIXME 5731 // CM's reference processing also cleans up the string and symbol tables. 5732 // Should we do that here also? We could, but it is a serial operation 5733 // and could significantly increase the pause time. 5734 5735 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5736 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5737 } 5738 5739 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { 5740 _expand_heap_after_alloc_failure = true; 5741 _evacuation_failed = false; 5742 5743 // Should G1EvacuationFailureALot be in effect for this GC? 5744 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5745 5746 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5747 5748 // Disable the hot card cache. 5749 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 5750 hot_card_cache->reset_hot_cache_claimed_index(); 5751 hot_card_cache->set_use_cache(false); 5752 5753 uint n_workers; 5754 if (G1CollectedHeap::use_parallel_gc_threads()) { 5755 n_workers = 5756 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5757 workers()->active_workers(), 5758 Threads::number_of_non_daemon_threads()); 5759 assert(UseDynamicNumberOfGCThreads || 5760 n_workers == workers()->total_workers(), 5761 "If not dynamic should be using all the workers"); 5762 workers()->set_active_workers(n_workers); 5763 set_par_threads(n_workers); 5764 } else { 5765 assert(n_par_threads() == 0, 5766 "Should be the original non-parallel value"); 5767 n_workers = 1; 5768 } 5769 5770 G1ParTask g1_par_task(this, _task_queues); 5771 5772 init_for_evac_failure(NULL); 5773 5774 rem_set()->prepare_for_younger_refs_iterate(true); 5775 5776 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5777 double start_par_time_sec = os::elapsedTime(); 5778 double end_par_time_sec; 5779 5780 { 5781 StrongRootsScope srs(this); 5782 5783 if (G1CollectedHeap::use_parallel_gc_threads()) { 5784 // The individual threads will set their evac-failure closures. 5785 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5786 // These tasks use ShareHeap::_process_strong_tasks 5787 assert(UseDynamicNumberOfGCThreads || 5788 workers()->active_workers() == workers()->total_workers(), 5789 "If not dynamic should be using all the workers"); 5790 workers()->run_task(&g1_par_task); 5791 } else { 5792 g1_par_task.set_for_termination(n_workers); 5793 g1_par_task.work(0); 5794 } 5795 end_par_time_sec = os::elapsedTime(); 5796 5797 // Closing the inner scope will execute the destructor 5798 // for the StrongRootsScope object. We record the current 5799 // elapsed time before closing the scope so that time 5800 // taken for the SRS destructor is NOT included in the 5801 // reported parallel time. 5802 } 5803 5804 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5805 g1_policy()->phase_times()->record_par_time(par_time_ms); 5806 5807 double code_root_fixup_time_ms = 5808 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5809 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5810 5811 set_par_threads(0); 5812 5813 // Process any discovered reference objects - we have 5814 // to do this _before_ we retire the GC alloc regions 5815 // as we may have to copy some 'reachable' referent 5816 // objects (and their reachable sub-graphs) that were 5817 // not copied during the pause. 5818 process_discovered_references(n_workers); 5819 5820 // Weak root processing. 5821 { 5822 G1STWIsAliveClosure is_alive(this); 5823 G1KeepAliveClosure keep_alive(this); 5824 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5825 } 5826 5827 release_gc_alloc_regions(n_workers, evacuation_info); 5828 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5829 5830 // Reset and re-enable the hot card cache. 5831 // Note the counts for the cards in the regions in the 5832 // collection set are reset when the collection set is freed. 5833 hot_card_cache->reset_hot_cache(); 5834 hot_card_cache->set_use_cache(true); 5835 5836 // Migrate the strong code roots attached to each region in 5837 // the collection set. Ideally we would like to do this 5838 // after we have finished the scanning/evacuation of the 5839 // strong code roots for a particular heap region. 5840 migrate_strong_code_roots(); 5841 5842 if (g1_policy()->during_initial_mark_pause()) { 5843 // Reset the claim values set during marking the strong code roots 5844 reset_heap_region_claim_values(); 5845 } 5846 5847 finalize_for_evac_failure(); 5848 5849 if (evacuation_failed()) { 5850 remove_self_forwarding_pointers(); 5851 5852 // Reset the G1EvacuationFailureALot counters and flags 5853 // Note: the values are reset only when an actual 5854 // evacuation failure occurs. 5855 NOT_PRODUCT(reset_evacuation_should_fail();) 5856 } 5857 5858 // Enqueue any remaining references remaining on the STW 5859 // reference processor's discovered lists. We need to do 5860 // this after the card table is cleaned (and verified) as 5861 // the act of enqueueing entries on to the pending list 5862 // will log these updates (and dirty their associated 5863 // cards). We need these updates logged to update any 5864 // RSets. 5865 enqueue_discovered_references(n_workers); 5866 5867 if (G1DeferredRSUpdate) { 5868 RedirtyLoggedCardTableEntryFastClosure redirty; 5869 dirty_card_queue_set().set_closure(&redirty); 5870 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5871 5872 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5873 dcq.merge_bufferlists(&dirty_card_queue_set()); 5874 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5875 } 5876 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5877 } 5878 5879 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5880 size_t* pre_used, 5881 FreeRegionList* free_list, 5882 OldRegionSet* old_proxy_set, 5883 HumongousRegionSet* humongous_proxy_set, 5884 HRRSCleanupTask* hrrs_cleanup_task, 5885 bool par) { 5886 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5887 if (hr->isHumongous()) { 5888 assert(hr->startsHumongous(), "we should only see starts humongous"); 5889 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5890 } else { 5891 _old_set.remove_with_proxy(hr, old_proxy_set); 5892 free_region(hr, pre_used, free_list, par); 5893 } 5894 } else { 5895 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5896 } 5897 } 5898 5899 void G1CollectedHeap::free_region(HeapRegion* hr, 5900 size_t* pre_used, 5901 FreeRegionList* free_list, 5902 bool par) { 5903 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5904 assert(!hr->is_empty(), "the region should not be empty"); 5905 assert(free_list != NULL, "pre-condition"); 5906 5907 // Clear the card counts for this region. 5908 // Note: we only need to do this if the region is not young 5909 // (since we don't refine cards in young regions). 5910 if (!hr->is_young()) { 5911 _cg1r->hot_card_cache()->reset_card_counts(hr); 5912 } 5913 *pre_used += hr->used(); 5914 hr->hr_clear(par, true /* clear_space */); 5915 free_list->add_as_head(hr); 5916 } 5917 5918 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5919 size_t* pre_used, 5920 FreeRegionList* free_list, 5921 HumongousRegionSet* humongous_proxy_set, 5922 bool par) { 5923 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5924 assert(free_list != NULL, "pre-condition"); 5925 assert(humongous_proxy_set != NULL, "pre-condition"); 5926 5927 size_t hr_used = hr->used(); 5928 size_t hr_capacity = hr->capacity(); 5929 size_t hr_pre_used = 0; 5930 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5931 // We need to read this before we make the region non-humongous, 5932 // otherwise the information will be gone. 5933 uint last_index = hr->last_hc_index(); 5934 hr->set_notHumongous(); 5935 free_region(hr, &hr_pre_used, free_list, par); 5936 5937 uint i = hr->hrs_index() + 1; 5938 while (i < last_index) { 5939 HeapRegion* curr_hr = region_at(i); 5940 assert(curr_hr->continuesHumongous(), "invariant"); 5941 curr_hr->set_notHumongous(); 5942 free_region(curr_hr, &hr_pre_used, free_list, par); 5943 i += 1; 5944 } 5945 assert(hr_pre_used == hr_used, 5946 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5947 "should be the same", hr_pre_used, hr_used)); 5948 *pre_used += hr_pre_used; 5949 } 5950 5951 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5952 FreeRegionList* free_list, 5953 OldRegionSet* old_proxy_set, 5954 HumongousRegionSet* humongous_proxy_set, 5955 bool par) { 5956 if (pre_used > 0) { 5957 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5958 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5959 assert(_summary_bytes_used >= pre_used, 5960 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5961 "should be >= pre_used: "SIZE_FORMAT, 5962 _summary_bytes_used, pre_used)); 5963 _summary_bytes_used -= pre_used; 5964 } 5965 if (free_list != NULL && !free_list->is_empty()) { 5966 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5967 _free_list.add_as_head(free_list); 5968 } 5969 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5970 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5971 _old_set.update_from_proxy(old_proxy_set); 5972 } 5973 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5974 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5975 _humongous_set.update_from_proxy(humongous_proxy_set); 5976 } 5977 } 5978 5979 class G1ParCleanupCTTask : public AbstractGangTask { 5980 G1SATBCardTableModRefBS* _ct_bs; 5981 G1CollectedHeap* _g1h; 5982 HeapRegion* volatile _su_head; 5983 public: 5984 G1ParCleanupCTTask(G1SATBCardTableModRefBS* ct_bs, 5985 G1CollectedHeap* g1h) : 5986 AbstractGangTask("G1 Par Cleanup CT Task"), 5987 _ct_bs(ct_bs), _g1h(g1h) { } 5988 5989 void work(uint worker_id) { 5990 HeapRegion* r; 5991 while (r = _g1h->pop_dirty_cards_region()) { 5992 clear_cards(r); 5993 } 5994 } 5995 5996 void clear_cards(HeapRegion* r) { 5997 // Cards of the survivors should have already been dirtied. 5998 if (!r->is_survivor()) { 5999 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 6000 } 6001 } 6002 }; 6003 6004 #ifndef PRODUCT 6005 class G1VerifyCardTableCleanup: public HeapRegionClosure { 6006 G1CollectedHeap* _g1h; 6007 G1SATBCardTableModRefBS* _ct_bs; 6008 public: 6009 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, G1SATBCardTableModRefBS* ct_bs) 6010 : _g1h(g1h), _ct_bs(ct_bs) { } 6011 virtual bool doHeapRegion(HeapRegion* r) { 6012 if (r->is_survivor()) { 6013 _g1h->verify_dirty_region(r); 6014 } else { 6015 _g1h->verify_not_dirty_region(r); 6016 } 6017 return false; 6018 } 6019 }; 6020 6021 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 6022 // All of the region should be clean. 6023 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6024 MemRegion mr(hr->bottom(), hr->end()); 6025 ct_bs->verify_not_dirty_region(mr); 6026 } 6027 6028 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 6029 // We cannot guarantee that [bottom(),end()] is dirty. Threads 6030 // dirty allocated blocks as they allocate them. The thread that 6031 // retires each region and replaces it with a new one will do a 6032 // maximal allocation to fill in [pre_dummy_top(),end()] but will 6033 // not dirty that area (one less thing to have to do while holding 6034 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 6035 // is dirty. 6036 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6037 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 6038 if (hr->is_young()) { 6039 ct_bs->verify_g1_young_region(mr); 6040 } else { 6041 ct_bs->verify_dirty_region(mr); 6042 } 6043 } 6044 6045 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 6046 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6047 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 6048 verify_dirty_region(hr); 6049 } 6050 } 6051 6052 void G1CollectedHeap::verify_dirty_young_regions() { 6053 verify_dirty_young_list(_young_list->first_region()); 6054 } 6055 #endif 6056 6057 void G1CollectedHeap::cleanUpCardTable() { 6058 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set(); 6059 double start = os::elapsedTime(); 6060 6061 { 6062 // Iterate over the dirty cards region list. 6063 G1ParCleanupCTTask cleanup_task(ct_bs, this); 6064 6065 if (G1CollectedHeap::use_parallel_gc_threads()) { 6066 set_par_threads(); 6067 workers()->run_task(&cleanup_task); 6068 set_par_threads(0); 6069 } else { 6070 while (_dirty_cards_region_list) { 6071 HeapRegion* r = _dirty_cards_region_list; 6072 cleanup_task.clear_cards(r); 6073 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 6074 if (_dirty_cards_region_list == r) { 6075 // The last region. 6076 _dirty_cards_region_list = NULL; 6077 } 6078 r->set_next_dirty_cards_region(NULL); 6079 } 6080 } 6081 #ifndef PRODUCT 6082 if (G1VerifyCTCleanup || VerifyAfterGC) { 6083 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 6084 heap_region_iterate(&cleanup_verifier); 6085 } 6086 #endif 6087 } 6088 6089 double elapsed = os::elapsedTime() - start; 6090 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 6091 } 6092 6093 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) { 6094 size_t pre_used = 0; 6095 FreeRegionList local_free_list("Local List for CSet Freeing"); 6096 6097 double young_time_ms = 0.0; 6098 double non_young_time_ms = 0.0; 6099 6100 // Since the collection set is a superset of the the young list, 6101 // all we need to do to clear the young list is clear its 6102 // head and length, and unlink any young regions in the code below 6103 _young_list->clear(); 6104 6105 G1CollectorPolicy* policy = g1_policy(); 6106 6107 double start_sec = os::elapsedTime(); 6108 bool non_young = true; 6109 6110 HeapRegion* cur = cs_head; 6111 int age_bound = -1; 6112 size_t rs_lengths = 0; 6113 6114 while (cur != NULL) { 6115 assert(!is_on_master_free_list(cur), "sanity"); 6116 if (non_young) { 6117 if (cur->is_young()) { 6118 double end_sec = os::elapsedTime(); 6119 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6120 non_young_time_ms += elapsed_ms; 6121 6122 start_sec = os::elapsedTime(); 6123 non_young = false; 6124 } 6125 } else { 6126 if (!cur->is_young()) { 6127 double end_sec = os::elapsedTime(); 6128 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6129 young_time_ms += elapsed_ms; 6130 6131 start_sec = os::elapsedTime(); 6132 non_young = true; 6133 } 6134 } 6135 6136 rs_lengths += cur->rem_set()->occupied(); 6137 6138 HeapRegion* next = cur->next_in_collection_set(); 6139 assert(cur->in_collection_set(), "bad CS"); 6140 cur->set_next_in_collection_set(NULL); 6141 cur->set_in_collection_set(false); 6142 6143 if (cur->is_young()) { 6144 int index = cur->young_index_in_cset(); 6145 assert(index != -1, "invariant"); 6146 assert((uint) index < policy->young_cset_region_length(), "invariant"); 6147 size_t words_survived = _surviving_young_words[index]; 6148 cur->record_surv_words_in_group(words_survived); 6149 6150 // At this point the we have 'popped' cur from the collection set 6151 // (linked via next_in_collection_set()) but it is still in the 6152 // young list (linked via next_young_region()). Clear the 6153 // _next_young_region field. 6154 cur->set_next_young_region(NULL); 6155 } else { 6156 int index = cur->young_index_in_cset(); 6157 assert(index == -1, "invariant"); 6158 } 6159 6160 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 6161 (!cur->is_young() && cur->young_index_in_cset() == -1), 6162 "invariant" ); 6163 6164 if (!cur->evacuation_failed()) { 6165 MemRegion used_mr = cur->used_region(); 6166 6167 // And the region is empty. 6168 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 6169 free_region(cur, &pre_used, &local_free_list, false /* par */); 6170 } else { 6171 cur->uninstall_surv_rate_group(); 6172 if (cur->is_young()) { 6173 cur->set_young_index_in_cset(-1); 6174 } 6175 cur->set_not_young(); 6176 cur->set_evacuation_failed(false); 6177 // The region is now considered to be old. 6178 _old_set.add(cur); 6179 evacuation_info.increment_collectionset_used_after(cur->used()); 6180 } 6181 cur = next; 6182 } 6183 6184 evacuation_info.set_regions_freed(local_free_list.length()); 6185 policy->record_max_rs_lengths(rs_lengths); 6186 policy->cset_regions_freed(); 6187 6188 double end_sec = os::elapsedTime(); 6189 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6190 6191 if (non_young) { 6192 non_young_time_ms += elapsed_ms; 6193 } else { 6194 young_time_ms += elapsed_ms; 6195 } 6196 6197 update_sets_after_freeing_regions(pre_used, &local_free_list, 6198 NULL /* old_proxy_set */, 6199 NULL /* humongous_proxy_set */, 6200 false /* par */); 6201 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6202 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6203 } 6204 6205 // This routine is similar to the above but does not record 6206 // any policy statistics or update free lists; we are abandoning 6207 // the current incremental collection set in preparation of a 6208 // full collection. After the full GC we will start to build up 6209 // the incremental collection set again. 6210 // This is only called when we're doing a full collection 6211 // and is immediately followed by the tearing down of the young list. 6212 6213 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6214 HeapRegion* cur = cs_head; 6215 6216 while (cur != NULL) { 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 cur->set_young_index_in_cset(-1); 6222 cur = next; 6223 } 6224 } 6225 6226 void G1CollectedHeap::set_free_regions_coming() { 6227 if (G1ConcRegionFreeingVerbose) { 6228 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6229 "setting free regions coming"); 6230 } 6231 6232 assert(!free_regions_coming(), "pre-condition"); 6233 _free_regions_coming = true; 6234 } 6235 6236 void G1CollectedHeap::reset_free_regions_coming() { 6237 assert(free_regions_coming(), "pre-condition"); 6238 6239 { 6240 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6241 _free_regions_coming = false; 6242 SecondaryFreeList_lock->notify_all(); 6243 } 6244 6245 if (G1ConcRegionFreeingVerbose) { 6246 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6247 "reset free regions coming"); 6248 } 6249 } 6250 6251 void G1CollectedHeap::wait_while_free_regions_coming() { 6252 // Most of the time we won't have to wait, so let's do a quick test 6253 // first before we take the lock. 6254 if (!free_regions_coming()) { 6255 return; 6256 } 6257 6258 if (G1ConcRegionFreeingVerbose) { 6259 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6260 "waiting for free regions"); 6261 } 6262 6263 { 6264 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6265 while (free_regions_coming()) { 6266 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6267 } 6268 } 6269 6270 if (G1ConcRegionFreeingVerbose) { 6271 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6272 "done waiting for free regions"); 6273 } 6274 } 6275 6276 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6277 assert(heap_lock_held_for_gc(), 6278 "the heap lock should already be held by or for this thread"); 6279 _young_list->push_region(hr); 6280 } 6281 6282 class NoYoungRegionsClosure: public HeapRegionClosure { 6283 private: 6284 bool _success; 6285 public: 6286 NoYoungRegionsClosure() : _success(true) { } 6287 bool doHeapRegion(HeapRegion* r) { 6288 if (r->is_young()) { 6289 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6290 r->bottom(), r->end()); 6291 _success = false; 6292 } 6293 return false; 6294 } 6295 bool success() { return _success; } 6296 }; 6297 6298 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6299 bool ret = _young_list->check_list_empty(check_sample); 6300 6301 if (check_heap) { 6302 NoYoungRegionsClosure closure; 6303 heap_region_iterate(&closure); 6304 ret = ret && closure.success(); 6305 } 6306 6307 return ret; 6308 } 6309 6310 class TearDownRegionSetsClosure : public HeapRegionClosure { 6311 private: 6312 OldRegionSet *_old_set; 6313 6314 public: 6315 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6316 6317 bool doHeapRegion(HeapRegion* r) { 6318 if (r->is_empty()) { 6319 // We ignore empty regions, we'll empty the free list afterwards 6320 } else if (r->is_young()) { 6321 // We ignore young regions, we'll empty the young list afterwards 6322 } else if (r->isHumongous()) { 6323 // We ignore humongous regions, we're not tearing down the 6324 // humongous region set 6325 } else { 6326 // The rest should be old 6327 _old_set->remove(r); 6328 } 6329 return false; 6330 } 6331 6332 ~TearDownRegionSetsClosure() { 6333 assert(_old_set->is_empty(), "post-condition"); 6334 } 6335 }; 6336 6337 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6338 assert_at_safepoint(true /* should_be_vm_thread */); 6339 6340 if (!free_list_only) { 6341 TearDownRegionSetsClosure cl(&_old_set); 6342 heap_region_iterate(&cl); 6343 6344 // Need to do this after the heap iteration to be able to 6345 // recognize the young regions and ignore them during the iteration. 6346 _young_list->empty_list(); 6347 } 6348 _free_list.remove_all(); 6349 } 6350 6351 class RebuildRegionSetsClosure : public HeapRegionClosure { 6352 private: 6353 bool _free_list_only; 6354 OldRegionSet* _old_set; 6355 FreeRegionList* _free_list; 6356 size_t _total_used; 6357 6358 public: 6359 RebuildRegionSetsClosure(bool free_list_only, 6360 OldRegionSet* old_set, FreeRegionList* free_list) : 6361 _free_list_only(free_list_only), 6362 _old_set(old_set), _free_list(free_list), _total_used(0) { 6363 assert(_free_list->is_empty(), "pre-condition"); 6364 if (!free_list_only) { 6365 assert(_old_set->is_empty(), "pre-condition"); 6366 } 6367 } 6368 6369 bool doHeapRegion(HeapRegion* r) { 6370 if (r->continuesHumongous()) { 6371 return false; 6372 } 6373 6374 if (r->is_empty()) { 6375 // Add free regions to the free list 6376 _free_list->add_as_tail(r); 6377 } else if (!_free_list_only) { 6378 assert(!r->is_young(), "we should not come across young regions"); 6379 6380 if (r->isHumongous()) { 6381 // We ignore humongous regions, we left the humongous set unchanged 6382 } else { 6383 // The rest should be old, add them to the old set 6384 _old_set->add(r); 6385 } 6386 _total_used += r->used(); 6387 } 6388 6389 return false; 6390 } 6391 6392 size_t total_used() { 6393 return _total_used; 6394 } 6395 }; 6396 6397 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6398 assert_at_safepoint(true /* should_be_vm_thread */); 6399 6400 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6401 heap_region_iterate(&cl); 6402 6403 if (!free_list_only) { 6404 _summary_bytes_used = cl.total_used(); 6405 } 6406 assert(_summary_bytes_used == recalculate_used(), 6407 err_msg("inconsistent _summary_bytes_used, " 6408 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6409 _summary_bytes_used, recalculate_used())); 6410 } 6411 6412 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6413 _refine_cte_cl->set_concurrent(concurrent); 6414 } 6415 6416 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6417 HeapRegion* hr = heap_region_containing(p); 6418 if (hr == NULL) { 6419 return false; 6420 } else { 6421 return hr->is_in(p); 6422 } 6423 } 6424 6425 // Methods for the mutator alloc region 6426 6427 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6428 bool force) { 6429 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6430 assert(!force || g1_policy()->can_expand_young_list(), 6431 "if force is true we should be able to expand the young list"); 6432 bool young_list_full = g1_policy()->is_young_list_full(); 6433 if (force || !young_list_full) { 6434 HeapRegion* new_alloc_region = new_region(word_size, 6435 false /* do_expand */); 6436 if (new_alloc_region != NULL) { 6437 set_region_short_lived_locked(new_alloc_region); 6438 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6439 return new_alloc_region; 6440 } 6441 } 6442 return NULL; 6443 } 6444 6445 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6446 size_t allocated_bytes) { 6447 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6448 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6449 6450 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6451 _summary_bytes_used += allocated_bytes; 6452 _hr_printer.retire(alloc_region); 6453 // We update the eden sizes here, when the region is retired, 6454 // instead of when it's allocated, since this is the point that its 6455 // used space has been recored in _summary_bytes_used. 6456 g1mm()->update_eden_size(); 6457 } 6458 6459 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6460 bool force) { 6461 return _g1h->new_mutator_alloc_region(word_size, force); 6462 } 6463 6464 void G1CollectedHeap::set_par_threads() { 6465 // Don't change the number of workers. Use the value previously set 6466 // in the workgroup. 6467 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6468 uint n_workers = workers()->active_workers(); 6469 assert(UseDynamicNumberOfGCThreads || 6470 n_workers == workers()->total_workers(), 6471 "Otherwise should be using the total number of workers"); 6472 if (n_workers == 0) { 6473 assert(false, "Should have been set in prior evacuation pause."); 6474 n_workers = ParallelGCThreads; 6475 workers()->set_active_workers(n_workers); 6476 } 6477 set_par_threads(n_workers); 6478 } 6479 6480 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6481 size_t allocated_bytes) { 6482 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6483 } 6484 6485 // Methods for the GC alloc regions 6486 6487 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6488 uint count, 6489 GCAllocPurpose ap) { 6490 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6491 6492 if (count < g1_policy()->max_regions(ap)) { 6493 HeapRegion* new_alloc_region = new_region(word_size, 6494 true /* do_expand */); 6495 if (new_alloc_region != NULL) { 6496 // We really only need to do this for old regions given that we 6497 // should never scan survivors. But it doesn't hurt to do it 6498 // for survivors too. 6499 new_alloc_region->set_saved_mark(); 6500 if (ap == GCAllocForSurvived) { 6501 new_alloc_region->set_survivor(); 6502 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6503 } else { 6504 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6505 } 6506 bool during_im = g1_policy()->during_initial_mark_pause(); 6507 new_alloc_region->note_start_of_copying(during_im); 6508 return new_alloc_region; 6509 } else { 6510 g1_policy()->note_alloc_region_limit_reached(ap); 6511 } 6512 } 6513 return NULL; 6514 } 6515 6516 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6517 size_t allocated_bytes, 6518 GCAllocPurpose ap) { 6519 bool during_im = g1_policy()->during_initial_mark_pause(); 6520 alloc_region->note_end_of_copying(during_im); 6521 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6522 if (ap == GCAllocForSurvived) { 6523 young_list()->add_survivor_region(alloc_region); 6524 } else { 6525 _old_set.add(alloc_region); 6526 } 6527 _hr_printer.retire(alloc_region); 6528 } 6529 6530 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6531 bool force) { 6532 assert(!force, "not supported for GC alloc regions"); 6533 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6534 } 6535 6536 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6537 size_t allocated_bytes) { 6538 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6539 GCAllocForSurvived); 6540 } 6541 6542 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6543 bool force) { 6544 assert(!force, "not supported for GC alloc regions"); 6545 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6546 } 6547 6548 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6549 size_t allocated_bytes) { 6550 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6551 GCAllocForTenured); 6552 } 6553 // Heap region set verification 6554 6555 class VerifyRegionListsClosure : public HeapRegionClosure { 6556 private: 6557 FreeRegionList* _free_list; 6558 OldRegionSet* _old_set; 6559 HumongousRegionSet* _humongous_set; 6560 uint _region_count; 6561 6562 public: 6563 VerifyRegionListsClosure(OldRegionSet* old_set, 6564 HumongousRegionSet* humongous_set, 6565 FreeRegionList* free_list) : 6566 _old_set(old_set), _humongous_set(humongous_set), 6567 _free_list(free_list), _region_count(0) { } 6568 6569 uint region_count() { return _region_count; } 6570 6571 bool doHeapRegion(HeapRegion* hr) { 6572 _region_count += 1; 6573 6574 if (hr->continuesHumongous()) { 6575 return false; 6576 } 6577 6578 if (hr->is_young()) { 6579 // TODO 6580 } else if (hr->startsHumongous()) { 6581 _humongous_set->verify_next_region(hr); 6582 } else if (hr->is_empty()) { 6583 _free_list->verify_next_region(hr); 6584 } else { 6585 _old_set->verify_next_region(hr); 6586 } 6587 return false; 6588 } 6589 }; 6590 6591 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6592 HeapWord* bottom) { 6593 HeapWord* end = bottom + HeapRegion::GrainWords; 6594 MemRegion mr(bottom, end); 6595 assert(_g1_reserved.contains(mr), "invariant"); 6596 // This might return NULL if the allocation fails 6597 return new HeapRegion(hrs_index, _bot_shared, mr); 6598 } 6599 6600 void G1CollectedHeap::verify_region_sets() { 6601 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6602 6603 // First, check the explicit lists. 6604 _free_list.verify(); 6605 { 6606 // Given that a concurrent operation might be adding regions to 6607 // the secondary free list we have to take the lock before 6608 // verifying it. 6609 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6610 _secondary_free_list.verify(); 6611 } 6612 _old_set.verify(); 6613 _humongous_set.verify(); 6614 6615 // If a concurrent region freeing operation is in progress it will 6616 // be difficult to correctly attributed any free regions we come 6617 // across to the correct free list given that they might belong to 6618 // one of several (free_list, secondary_free_list, any local lists, 6619 // etc.). So, if that's the case we will skip the rest of the 6620 // verification operation. Alternatively, waiting for the concurrent 6621 // operation to complete will have a non-trivial effect on the GC's 6622 // operation (no concurrent operation will last longer than the 6623 // interval between two calls to verification) and it might hide 6624 // any issues that we would like to catch during testing. 6625 if (free_regions_coming()) { 6626 return; 6627 } 6628 6629 // Make sure we append the secondary_free_list on the free_list so 6630 // that all free regions we will come across can be safely 6631 // attributed to the free_list. 6632 append_secondary_free_list_if_not_empty_with_lock(); 6633 6634 // Finally, make sure that the region accounting in the lists is 6635 // consistent with what we see in the heap. 6636 _old_set.verify_start(); 6637 _humongous_set.verify_start(); 6638 _free_list.verify_start(); 6639 6640 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6641 heap_region_iterate(&cl); 6642 6643 _old_set.verify_end(); 6644 _humongous_set.verify_end(); 6645 _free_list.verify_end(); 6646 } 6647 6648 // Optimized nmethod scanning 6649 6650 class RegisterNMethodOopClosure: public OopClosure { 6651 G1CollectedHeap* _g1h; 6652 nmethod* _nm; 6653 6654 template <class T> void do_oop_work(T* p) { 6655 T heap_oop = oopDesc::load_heap_oop(p); 6656 if (!oopDesc::is_null(heap_oop)) { 6657 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6658 HeapRegion* hr = _g1h->heap_region_containing(obj); 6659 assert(!hr->isHumongous(), "code root in humongous region?"); 6660 6661 // HeapRegion::add_strong_code_root() avoids adding duplicate 6662 // entries but having duplicates is OK since we "mark" nmethods 6663 // as visited when we scan the strong code root lists during the GC. 6664 hr->add_strong_code_root(_nm); 6665 assert(hr->rem_set()->strong_code_roots_list_contains(_nm), "add failed?"); 6666 } 6667 } 6668 6669 public: 6670 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6671 _g1h(g1h), _nm(nm) {} 6672 6673 void do_oop(oop* p) { do_oop_work(p); } 6674 void do_oop(narrowOop* p) { do_oop_work(p); } 6675 }; 6676 6677 class UnregisterNMethodOopClosure: public OopClosure { 6678 G1CollectedHeap* _g1h; 6679 nmethod* _nm; 6680 6681 template <class T> void do_oop_work(T* p) { 6682 T heap_oop = oopDesc::load_heap_oop(p); 6683 if (!oopDesc::is_null(heap_oop)) { 6684 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6685 HeapRegion* hr = _g1h->heap_region_containing(obj); 6686 assert(!hr->isHumongous(), "code root in humongous region?"); 6687 hr->remove_strong_code_root(_nm); 6688 assert(!hr->rem_set()->strong_code_roots_list_contains(_nm), "remove failed?"); 6689 } 6690 } 6691 6692 public: 6693 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6694 _g1h(g1h), _nm(nm) {} 6695 6696 void do_oop(oop* p) { do_oop_work(p); } 6697 void do_oop(narrowOop* p) { do_oop_work(p); } 6698 }; 6699 6700 void G1CollectedHeap::register_nmethod(nmethod* nm) { 6701 CollectedHeap::register_nmethod(nm); 6702 6703 guarantee(nm != NULL, "sanity"); 6704 RegisterNMethodOopClosure reg_cl(this, nm); 6705 nm->oops_do(®_cl); 6706 } 6707 6708 void G1CollectedHeap::unregister_nmethod(nmethod* nm) { 6709 CollectedHeap::unregister_nmethod(nm); 6710 6711 guarantee(nm != NULL, "sanity"); 6712 UnregisterNMethodOopClosure reg_cl(this, nm); 6713 nm->oops_do(®_cl, true); 6714 } 6715 6716 class MigrateCodeRootsHeapRegionClosure: public HeapRegionClosure { 6717 public: 6718 bool doHeapRegion(HeapRegion *hr) { 6719 assert(!hr->isHumongous(), "humongous region in collection set?"); 6720 hr->migrate_strong_code_roots(); 6721 return false; 6722 } 6723 }; 6724 6725 void G1CollectedHeap::migrate_strong_code_roots() { 6726 MigrateCodeRootsHeapRegionClosure cl; 6727 double migrate_start = os::elapsedTime(); 6728 collection_set_iterate(&cl); 6729 double migration_time_ms = (os::elapsedTime() - migrate_start) * 1000.0; 6730 g1_policy()->phase_times()->record_strong_code_root_migration_time(migration_time_ms); 6731 } 6732 6733 // Mark all the code roots that point into regions *not* in the 6734 // collection set. 6735 // 6736 // Note we do not want to use a "marking" CodeBlobToOopClosure while 6737 // walking the the code roots lists of regions not in the collection 6738 // set. Suppose we have an nmethod (M) that points to objects in two 6739 // separate regions - one in the collection set (R1) and one not (R2). 6740 // Using a "marking" CodeBlobToOopClosure here would result in "marking" 6741 // nmethod M when walking the code roots for R1. When we come to scan 6742 // the code roots for R2, we would see that M is already marked and it 6743 // would be skipped and the objects in R2 that are referenced from M 6744 // would not be evacuated. 6745 6746 class MarkStrongCodeRootCodeBlobClosure: public CodeBlobClosure { 6747 6748 class MarkStrongCodeRootOopClosure: public OopClosure { 6749 ConcurrentMark* _cm; 6750 HeapRegion* _hr; 6751 uint _worker_id; 6752 6753 template <class T> void do_oop_work(T* p) { 6754 T heap_oop = oopDesc::load_heap_oop(p); 6755 if (!oopDesc::is_null(heap_oop)) { 6756 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6757 // Only mark objects in the region (which is assumed 6758 // to be not in the collection set). 6759 if (_hr->is_in(obj)) { 6760 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 6761 } 6762 } 6763 } 6764 6765 public: 6766 MarkStrongCodeRootOopClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id) : 6767 _cm(cm), _hr(hr), _worker_id(worker_id) { 6768 assert(!_hr->in_collection_set(), "sanity"); 6769 } 6770 6771 void do_oop(narrowOop* p) { do_oop_work(p); } 6772 void do_oop(oop* p) { do_oop_work(p); } 6773 }; 6774 6775 MarkStrongCodeRootOopClosure _oop_cl; 6776 6777 public: 6778 MarkStrongCodeRootCodeBlobClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id): 6779 _oop_cl(cm, hr, worker_id) {} 6780 6781 void do_code_blob(CodeBlob* cb) { 6782 nmethod* nm = (cb == NULL) ? NULL : cb->as_nmethod_or_null(); 6783 if (nm != NULL) { 6784 nm->oops_do(&_oop_cl); 6785 } 6786 } 6787 }; 6788 6789 class MarkStrongCodeRootsHRClosure: public HeapRegionClosure { 6790 G1CollectedHeap* _g1h; 6791 uint _worker_id; 6792 6793 public: 6794 MarkStrongCodeRootsHRClosure(G1CollectedHeap* g1h, uint worker_id) : 6795 _g1h(g1h), _worker_id(worker_id) {} 6796 6797 bool doHeapRegion(HeapRegion *hr) { 6798 HeapRegionRemSet* hrrs = hr->rem_set(); 6799 if (hr->isHumongous()) { 6800 // Code roots should never be attached to a humongous region 6801 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 6802 return false; 6803 } 6804 6805 if (hr->in_collection_set()) { 6806 // Don't mark code roots into regions in the collection set here. 6807 // They will be marked when we scan them. 6808 return false; 6809 } 6810 6811 MarkStrongCodeRootCodeBlobClosure cb_cl(_g1h->concurrent_mark(), hr, _worker_id); 6812 hr->strong_code_roots_do(&cb_cl); 6813 return false; 6814 } 6815 }; 6816 6817 void G1CollectedHeap::mark_strong_code_roots(uint worker_id) { 6818 MarkStrongCodeRootsHRClosure cl(this, worker_id); 6819 if (G1CollectedHeap::use_parallel_gc_threads()) { 6820 heap_region_par_iterate_chunked(&cl, 6821 worker_id, 6822 workers()->active_workers(), 6823 HeapRegion::ParMarkRootClaimValue); 6824 } else { 6825 heap_region_iterate(&cl); 6826 } 6827 } 6828 6829 class RebuildStrongCodeRootClosure: public CodeBlobClosure { 6830 G1CollectedHeap* _g1h; 6831 6832 public: 6833 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) : 6834 _g1h(g1h) {} 6835 6836 void do_code_blob(CodeBlob* cb) { 6837 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL; 6838 if (nm == NULL) { 6839 return; 6840 } 6841 6842 if (ScavengeRootsInCode && nm->detect_scavenge_root_oops()) { 6843 _g1h->register_nmethod(nm); 6844 } 6845 } 6846 }; 6847 6848 void G1CollectedHeap::rebuild_strong_code_roots() { 6849 RebuildStrongCodeRootClosure blob_cl(this); 6850 CodeCache::blobs_do(&blob_cl); 6851 }