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) 129 { 130 _g1h = G1CollectedHeap::heap(); 131 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 132 for (int i = 0; i < 256; i++) _histo[i] = 0; 133 } 134 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 135 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 136 _calls++; 137 unsigned char* ujb = (unsigned char*)card_ptr; 138 int ind = (int)(*ujb); 139 _histo[ind]++; 140 *card_ptr = -1; 141 } 142 return true; 143 } 144 int calls() { return _calls; } 145 void print_histo() { 146 gclog_or_tty->print_cr("Card table value histogram:"); 147 for (int i = 0; i < 256; i++) { 148 if (_histo[i] != 0) { 149 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 150 } 151 } 152 } 153 }; 154 155 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 156 int _calls; 157 G1CollectedHeap* _g1h; 158 CardTableModRefBS* _ctbs; 159 public: 160 RedirtyLoggedCardTableEntryClosure() : 161 _calls(0) 162 { 163 _g1h = G1CollectedHeap::heap(); 164 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 165 } 166 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 167 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 168 _calls++; 169 *card_ptr = 0; 170 } 171 return true; 172 } 173 int calls() { return _calls; } 174 }; 175 176 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 177 public: 178 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 179 *card_ptr = CardTableModRefBS::dirty_card_val(); 180 return true; 181 } 182 }; 183 184 YoungList::YoungList(G1CollectedHeap* g1h) : 185 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 186 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 187 guarantee(check_list_empty(false), "just making sure..."); 188 } 189 190 void YoungList::push_region(HeapRegion *hr) { 191 assert(!hr->is_young(), "should not already be young"); 192 assert(hr->get_next_young_region() == NULL, "cause it should!"); 193 194 hr->set_next_young_region(_head); 195 _head = hr; 196 197 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 198 ++_length; 199 } 200 201 void YoungList::add_survivor_region(HeapRegion* hr) { 202 assert(hr->is_survivor(), "should be flagged as survivor region"); 203 assert(hr->get_next_young_region() == NULL, "cause it should!"); 204 205 hr->set_next_young_region(_survivor_head); 206 if (_survivor_head == NULL) { 207 _survivor_tail = hr; 208 } 209 _survivor_head = hr; 210 ++_survivor_length; 211 } 212 213 void YoungList::empty_list(HeapRegion* list) { 214 while (list != NULL) { 215 HeapRegion* next = list->get_next_young_region(); 216 list->set_next_young_region(NULL); 217 list->uninstall_surv_rate_group(); 218 list->set_not_young(); 219 list = next; 220 } 221 } 222 223 void YoungList::empty_list() { 224 assert(check_list_well_formed(), "young list should be well formed"); 225 226 empty_list(_head); 227 _head = NULL; 228 _length = 0; 229 230 empty_list(_survivor_head); 231 _survivor_head = NULL; 232 _survivor_tail = NULL; 233 _survivor_length = 0; 234 235 _last_sampled_rs_lengths = 0; 236 237 assert(check_list_empty(false), "just making sure..."); 238 } 239 240 bool YoungList::check_list_well_formed() { 241 bool ret = true; 242 243 uint length = 0; 244 HeapRegion* curr = _head; 245 HeapRegion* last = NULL; 246 while (curr != NULL) { 247 if (!curr->is_young()) { 248 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 249 "incorrectly tagged (y: %d, surv: %d)", 250 curr->bottom(), curr->end(), 251 curr->is_young(), curr->is_survivor()); 252 ret = false; 253 } 254 ++length; 255 last = curr; 256 curr = curr->get_next_young_region(); 257 } 258 ret = ret && (length == _length); 259 260 if (!ret) { 261 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 262 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 263 length, _length); 264 } 265 266 return ret; 267 } 268 269 bool YoungList::check_list_empty(bool check_sample) { 270 bool ret = true; 271 272 if (_length != 0) { 273 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 274 _length); 275 ret = false; 276 } 277 if (check_sample && _last_sampled_rs_lengths != 0) { 278 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 279 ret = false; 280 } 281 if (_head != NULL) { 282 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 283 ret = false; 284 } 285 if (!ret) { 286 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 287 } 288 289 return ret; 290 } 291 292 void 293 YoungList::rs_length_sampling_init() { 294 _sampled_rs_lengths = 0; 295 _curr = _head; 296 } 297 298 bool 299 YoungList::rs_length_sampling_more() { 300 return _curr != NULL; 301 } 302 303 void 304 YoungList::rs_length_sampling_next() { 305 assert( _curr != NULL, "invariant" ); 306 size_t rs_length = _curr->rem_set()->occupied(); 307 308 _sampled_rs_lengths += rs_length; 309 310 // The current region may not yet have been added to the 311 // incremental collection set (it gets added when it is 312 // retired as the current allocation region). 313 if (_curr->in_collection_set()) { 314 // Update the collection set policy information for this region 315 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 316 } 317 318 _curr = _curr->get_next_young_region(); 319 if (_curr == NULL) { 320 _last_sampled_rs_lengths = _sampled_rs_lengths; 321 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 322 } 323 } 324 325 void 326 YoungList::reset_auxilary_lists() { 327 guarantee( is_empty(), "young list should be empty" ); 328 assert(check_list_well_formed(), "young list should be well formed"); 329 330 // Add survivor regions to SurvRateGroup. 331 _g1h->g1_policy()->note_start_adding_survivor_regions(); 332 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 333 334 int young_index_in_cset = 0; 335 for (HeapRegion* curr = _survivor_head; 336 curr != NULL; 337 curr = curr->get_next_young_region()) { 338 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 339 340 // The region is a non-empty survivor so let's add it to 341 // the incremental collection set for the next evacuation 342 // pause. 343 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 344 young_index_in_cset += 1; 345 } 346 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 347 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 348 349 _head = _survivor_head; 350 _length = _survivor_length; 351 if (_survivor_head != NULL) { 352 assert(_survivor_tail != NULL, "cause it shouldn't be"); 353 assert(_survivor_length > 0, "invariant"); 354 _survivor_tail->set_next_young_region(NULL); 355 } 356 357 // Don't clear the survivor list handles until the start of 358 // the next evacuation pause - we need it in order to re-tag 359 // the survivor regions from this evacuation pause as 'young' 360 // at the start of the next. 361 362 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 363 364 assert(check_list_well_formed(), "young list should be well formed"); 365 } 366 367 void YoungList::print() { 368 HeapRegion* lists[] = {_head, _survivor_head}; 369 const char* names[] = {"YOUNG", "SURVIVOR"}; 370 371 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 372 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 373 HeapRegion *curr = lists[list]; 374 if (curr == NULL) 375 gclog_or_tty->print_cr(" empty"); 376 while (curr != NULL) { 377 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 378 HR_FORMAT_PARAMS(curr), 379 curr->prev_top_at_mark_start(), 380 curr->next_top_at_mark_start(), 381 curr->age_in_surv_rate_group_cond()); 382 curr = curr->get_next_young_region(); 383 } 384 } 385 386 gclog_or_tty->print_cr(""); 387 } 388 389 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 390 { 391 // Claim the right to put the region on the dirty cards region list 392 // by installing a self pointer. 393 HeapRegion* next = hr->get_next_dirty_cards_region(); 394 if (next == NULL) { 395 HeapRegion* res = (HeapRegion*) 396 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 397 NULL); 398 if (res == NULL) { 399 HeapRegion* head; 400 do { 401 // Put the region to the dirty cards region list. 402 head = _dirty_cards_region_list; 403 next = (HeapRegion*) 404 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 405 if (next == head) { 406 assert(hr->get_next_dirty_cards_region() == hr, 407 "hr->get_next_dirty_cards_region() != hr"); 408 if (next == NULL) { 409 // The last region in the list points to itself. 410 hr->set_next_dirty_cards_region(hr); 411 } else { 412 hr->set_next_dirty_cards_region(next); 413 } 414 } 415 } while (next != head); 416 } 417 } 418 } 419 420 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 421 { 422 HeapRegion* head; 423 HeapRegion* hr; 424 do { 425 head = _dirty_cards_region_list; 426 if (head == NULL) { 427 return NULL; 428 } 429 HeapRegion* new_head = head->get_next_dirty_cards_region(); 430 if (head == new_head) { 431 // The last region. 432 new_head = NULL; 433 } 434 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 435 head); 436 } while (hr != head); 437 assert(hr != NULL, "invariant"); 438 hr->set_next_dirty_cards_region(NULL); 439 return hr; 440 } 441 442 void G1CollectedHeap::stop_conc_gc_threads() { 443 _cg1r->stop(); 444 _cmThread->stop(); 445 } 446 447 #ifdef ASSERT 448 // A region is added to the collection set as it is retired 449 // so an address p can point to a region which will be in the 450 // collection set but has not yet been retired. This method 451 // therefore is only accurate during a GC pause after all 452 // regions have been retired. It is used for debugging 453 // to check if an nmethod has references to objects that can 454 // be move during a partial collection. Though it can be 455 // inaccurate, it is sufficient for G1 because the conservative 456 // implementation of is_scavengable() for G1 will indicate that 457 // all nmethods must be scanned during a partial collection. 458 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 459 HeapRegion* hr = heap_region_containing(p); 460 return hr != NULL && hr->in_collection_set(); 461 } 462 #endif 463 464 // Returns true if the reference points to an object that 465 // can move in an incremental collection. 466 bool G1CollectedHeap::is_scavengable(const void* p) { 467 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 468 G1CollectorPolicy* g1p = g1h->g1_policy(); 469 HeapRegion* hr = heap_region_containing(p); 470 if (hr == NULL) { 471 // null 472 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p)); 473 return false; 474 } else { 475 return !hr->isHumongous(); 476 } 477 } 478 479 void G1CollectedHeap::check_ct_logs_at_safepoint() { 480 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 481 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 482 483 // Count the dirty cards at the start. 484 CountNonCleanMemRegionClosure count1(this); 485 ct_bs->mod_card_iterate(&count1); 486 int orig_count = count1.n(); 487 488 // First clear the logged cards. 489 ClearLoggedCardTableEntryClosure clear; 490 dcqs.set_closure(&clear); 491 dcqs.apply_closure_to_all_completed_buffers(); 492 dcqs.iterate_closure_all_threads(false); 493 clear.print_histo(); 494 495 // Now ensure that there's no dirty cards. 496 CountNonCleanMemRegionClosure count2(this); 497 ct_bs->mod_card_iterate(&count2); 498 if (count2.n() != 0) { 499 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 500 count2.n(), orig_count); 501 } 502 guarantee(count2.n() == 0, "Card table should be clean."); 503 504 RedirtyLoggedCardTableEntryClosure redirty; 505 JavaThread::dirty_card_queue_set().set_closure(&redirty); 506 dcqs.apply_closure_to_all_completed_buffers(); 507 dcqs.iterate_closure_all_threads(false); 508 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 509 clear.calls(), orig_count); 510 guarantee(redirty.calls() == clear.calls(), 511 "Or else mechanism is broken."); 512 513 CountNonCleanMemRegionClosure count3(this); 514 ct_bs->mod_card_iterate(&count3); 515 if (count3.n() != orig_count) { 516 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 517 orig_count, count3.n()); 518 guarantee(count3.n() >= orig_count, "Should have restored them all."); 519 } 520 521 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 522 } 523 524 // Private class members. 525 526 G1CollectedHeap* G1CollectedHeap::_g1h; 527 528 // Private methods. 529 530 HeapRegion* 531 G1CollectedHeap::new_region_try_secondary_free_list() { 532 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 533 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 534 if (!_secondary_free_list.is_empty()) { 535 if (G1ConcRegionFreeingVerbose) { 536 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 537 "secondary_free_list has %u entries", 538 _secondary_free_list.length()); 539 } 540 // It looks as if there are free regions available on the 541 // secondary_free_list. Let's move them to the free_list and try 542 // again to allocate from it. 543 append_secondary_free_list(); 544 545 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 546 "empty we should have moved at least one entry to the free_list"); 547 HeapRegion* res = _free_list.remove_head(); 548 if (G1ConcRegionFreeingVerbose) { 549 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 550 "allocated "HR_FORMAT" from secondary_free_list", 551 HR_FORMAT_PARAMS(res)); 552 } 553 return res; 554 } 555 556 // Wait here until we get notified either when (a) there are no 557 // more free regions coming or (b) some regions have been moved on 558 // the secondary_free_list. 559 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 560 } 561 562 if (G1ConcRegionFreeingVerbose) { 563 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 564 "could not allocate from secondary_free_list"); 565 } 566 return NULL; 567 } 568 569 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 570 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 571 "the only time we use this to allocate a humongous region is " 572 "when we are allocating a single humongous region"); 573 574 HeapRegion* res; 575 if (G1StressConcRegionFreeing) { 576 if (!_secondary_free_list.is_empty()) { 577 if (G1ConcRegionFreeingVerbose) { 578 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 579 "forced to look at the secondary_free_list"); 580 } 581 res = new_region_try_secondary_free_list(); 582 if (res != NULL) { 583 return res; 584 } 585 } 586 } 587 res = _free_list.remove_head_or_null(); 588 if (res == NULL) { 589 if (G1ConcRegionFreeingVerbose) { 590 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 591 "res == NULL, trying the secondary_free_list"); 592 } 593 res = new_region_try_secondary_free_list(); 594 } 595 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 596 // Currently, only attempts to allocate GC alloc regions set 597 // do_expand to true. So, we should only reach here during a 598 // safepoint. If this assumption changes we might have to 599 // reconsider the use of _expand_heap_after_alloc_failure. 600 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 601 602 ergo_verbose1(ErgoHeapSizing, 603 "attempt heap expansion", 604 ergo_format_reason("region allocation request failed") 605 ergo_format_byte("allocation request"), 606 word_size * HeapWordSize); 607 if (expand(word_size * HeapWordSize)) { 608 // Given that expand() succeeded in expanding the heap, and we 609 // always expand the heap by an amount aligned to the heap 610 // region size, the free list should in theory not be empty. So 611 // it would probably be OK to use remove_head(). But the extra 612 // check for NULL is unlikely to be a performance issue here (we 613 // just expanded the heap!) so let's just be conservative and 614 // use remove_head_or_null(). 615 res = _free_list.remove_head_or_null(); 616 } else { 617 _expand_heap_after_alloc_failure = false; 618 } 619 } 620 return res; 621 } 622 623 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 624 size_t word_size) { 625 assert(isHumongous(word_size), "word_size should be humongous"); 626 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 627 628 uint first = G1_NULL_HRS_INDEX; 629 if (num_regions == 1) { 630 // Only one region to allocate, no need to go through the slower 631 // path. The caller will attempt the expansion if this fails, so 632 // let's not try to expand here too. 633 HeapRegion* hr = new_region(word_size, false /* do_expand */); 634 if (hr != NULL) { 635 first = hr->hrs_index(); 636 } else { 637 first = G1_NULL_HRS_INDEX; 638 } 639 } else { 640 // We can't allocate humongous regions while cleanupComplete() is 641 // running, since some of the regions we find to be empty might not 642 // yet be added to the free list and it is not straightforward to 643 // know which list they are on so that we can remove them. Note 644 // that we only need to do this if we need to allocate more than 645 // one region to satisfy the current humongous allocation 646 // request. If we are only allocating one region we use the common 647 // region allocation code (see above). 648 wait_while_free_regions_coming(); 649 append_secondary_free_list_if_not_empty_with_lock(); 650 651 if (free_regions() >= num_regions) { 652 first = _hrs.find_contiguous(num_regions); 653 if (first != G1_NULL_HRS_INDEX) { 654 for (uint i = first; i < first + num_regions; ++i) { 655 HeapRegion* hr = region_at(i); 656 assert(hr->is_empty(), "sanity"); 657 assert(is_on_master_free_list(hr), "sanity"); 658 hr->set_pending_removal(true); 659 } 660 _free_list.remove_all_pending(num_regions); 661 } 662 } 663 } 664 return first; 665 } 666 667 HeapWord* 668 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 669 uint num_regions, 670 size_t word_size) { 671 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 672 assert(isHumongous(word_size), "word_size should be humongous"); 673 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 674 675 // Index of last region in the series + 1. 676 uint last = first + num_regions; 677 678 // We need to initialize the region(s) we just discovered. This is 679 // a bit tricky given that it can happen concurrently with 680 // refinement threads refining cards on these regions and 681 // potentially wanting to refine the BOT as they are scanning 682 // those cards (this can happen shortly after a cleanup; see CR 683 // 6991377). So we have to set up the region(s) carefully and in 684 // a specific order. 685 686 // The word size sum of all the regions we will allocate. 687 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 688 assert(word_size <= word_size_sum, "sanity"); 689 690 // This will be the "starts humongous" region. 691 HeapRegion* first_hr = region_at(first); 692 // The header of the new object will be placed at the bottom of 693 // the first region. 694 HeapWord* new_obj = first_hr->bottom(); 695 // This will be the new end of the first region in the series that 696 // should also match the end of the last region in the series. 697 HeapWord* new_end = new_obj + word_size_sum; 698 // This will be the new top of the first region that will reflect 699 // this allocation. 700 HeapWord* new_top = new_obj + word_size; 701 702 // First, we need to zero the header of the space that we will be 703 // allocating. When we update top further down, some refinement 704 // threads might try to scan the region. By zeroing the header we 705 // ensure that any thread that will try to scan the region will 706 // come across the zero klass word and bail out. 707 // 708 // NOTE: It would not have been correct to have used 709 // CollectedHeap::fill_with_object() and make the space look like 710 // an int array. The thread that is doing the allocation will 711 // later update the object header to a potentially different array 712 // type and, for a very short period of time, the klass and length 713 // fields will be inconsistent. This could cause a refinement 714 // thread to calculate the object size incorrectly. 715 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 716 717 // We will set up the first region as "starts humongous". This 718 // will also update the BOT covering all the regions to reflect 719 // that there is a single object that starts at the bottom of the 720 // first region. 721 first_hr->set_startsHumongous(new_top, new_end); 722 723 // Then, if there are any, we will set up the "continues 724 // humongous" regions. 725 HeapRegion* hr = NULL; 726 for (uint i = first + 1; i < last; ++i) { 727 hr = region_at(i); 728 hr->set_continuesHumongous(first_hr); 729 } 730 // If we have "continues humongous" regions (hr != NULL), then the 731 // end of the last one should match new_end. 732 assert(hr == NULL || hr->end() == new_end, "sanity"); 733 734 // Up to this point no concurrent thread would have been able to 735 // do any scanning on any region in this series. All the top 736 // fields still point to bottom, so the intersection between 737 // [bottom,top] and [card_start,card_end] will be empty. Before we 738 // update the top fields, we'll do a storestore to make sure that 739 // no thread sees the update to top before the zeroing of the 740 // object header and the BOT initialization. 741 OrderAccess::storestore(); 742 743 // Now that the BOT and the object header have been initialized, 744 // we can update top of the "starts humongous" region. 745 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 746 "new_top should be in this region"); 747 first_hr->set_top(new_top); 748 if (_hr_printer.is_active()) { 749 HeapWord* bottom = first_hr->bottom(); 750 HeapWord* end = first_hr->orig_end(); 751 if ((first + 1) == last) { 752 // the series has a single humongous region 753 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 754 } else { 755 // the series has more than one humongous regions 756 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 757 } 758 } 759 760 // Now, we will update the top fields of the "continues humongous" 761 // regions. The reason we need to do this is that, otherwise, 762 // these regions would look empty and this will confuse parts of 763 // G1. For example, the code that looks for a consecutive number 764 // of empty regions will consider them empty and try to 765 // re-allocate them. We can extend is_empty() to also include 766 // !continuesHumongous(), but it is easier to just update the top 767 // fields here. The way we set top for all regions (i.e., top == 768 // end for all regions but the last one, top == new_top for the 769 // last one) is actually used when we will free up the humongous 770 // region in free_humongous_region(). 771 hr = NULL; 772 for (uint i = first + 1; i < last; ++i) { 773 hr = region_at(i); 774 if ((i + 1) == last) { 775 // last continues humongous region 776 assert(hr->bottom() < new_top && new_top <= hr->end(), 777 "new_top should fall on this region"); 778 hr->set_top(new_top); 779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 780 } else { 781 // not last one 782 assert(new_top > hr->end(), "new_top should be above this region"); 783 hr->set_top(hr->end()); 784 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 785 } 786 } 787 // If we have continues humongous regions (hr != NULL), then the 788 // end of the last one should match new_end and its top should 789 // match new_top. 790 assert(hr == NULL || 791 (hr->end() == new_end && hr->top() == new_top), "sanity"); 792 793 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 794 _summary_bytes_used += first_hr->used(); 795 _humongous_set.add(first_hr); 796 797 return new_obj; 798 } 799 800 // If could fit into free regions w/o expansion, try. 801 // Otherwise, if can expand, do so. 802 // Otherwise, if using ex regions might help, try with ex given back. 803 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 804 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 805 806 verify_region_sets_optional(); 807 808 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 809 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 810 uint x_num = expansion_regions(); 811 uint fs = _hrs.free_suffix(); 812 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 813 if (first == G1_NULL_HRS_INDEX) { 814 // The only thing we can do now is attempt expansion. 815 if (fs + x_num >= num_regions) { 816 // If the number of regions we're trying to allocate for this 817 // object is at most the number of regions in the free suffix, 818 // then the call to humongous_obj_allocate_find_first() above 819 // should have succeeded and we wouldn't be here. 820 // 821 // We should only be trying to expand when the free suffix is 822 // not sufficient for the object _and_ we have some expansion 823 // room available. 824 assert(num_regions > fs, "earlier allocation should have succeeded"); 825 826 ergo_verbose1(ErgoHeapSizing, 827 "attempt heap expansion", 828 ergo_format_reason("humongous allocation request failed") 829 ergo_format_byte("allocation request"), 830 word_size * HeapWordSize); 831 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 832 // Even though the heap was expanded, it might not have 833 // reached the desired size. So, we cannot assume that the 834 // allocation will succeed. 835 first = humongous_obj_allocate_find_first(num_regions, word_size); 836 } 837 } 838 } 839 840 HeapWord* result = NULL; 841 if (first != G1_NULL_HRS_INDEX) { 842 result = 843 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 844 assert(result != NULL, "it should always return a valid result"); 845 846 // A successful humongous object allocation changes the used space 847 // information of the old generation so we need to recalculate the 848 // sizes and update the jstat counters here. 849 g1mm()->update_sizes(); 850 } 851 852 verify_region_sets_optional(); 853 854 return result; 855 } 856 857 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 858 assert_heap_not_locked_and_not_at_safepoint(); 859 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 860 861 unsigned int dummy_gc_count_before; 862 int dummy_gclocker_retry_count = 0; 863 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count); 864 } 865 866 HeapWord* 867 G1CollectedHeap::mem_allocate(size_t word_size, 868 bool* gc_overhead_limit_was_exceeded) { 869 assert_heap_not_locked_and_not_at_safepoint(); 870 871 // Loop until the allocation is satisfied, or unsatisfied after GC. 872 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) { 873 unsigned int gc_count_before; 874 875 HeapWord* result = NULL; 876 if (!isHumongous(word_size)) { 877 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count); 878 } else { 879 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count); 880 } 881 if (result != NULL) { 882 return result; 883 } 884 885 // Create the garbage collection operation... 886 VM_G1CollectForAllocation op(gc_count_before, word_size); 887 // ...and get the VM thread to execute it. 888 VMThread::execute(&op); 889 890 if (op.prologue_succeeded() && op.pause_succeeded()) { 891 // If the operation was successful we'll return the result even 892 // if it is NULL. If the allocation attempt failed immediately 893 // after a Full GC, it's unlikely we'll be able to allocate now. 894 HeapWord* result = op.result(); 895 if (result != NULL && !isHumongous(word_size)) { 896 // Allocations that take place on VM operations do not do any 897 // card dirtying and we have to do it here. We only have to do 898 // this for non-humongous allocations, though. 899 dirty_young_block(result, word_size); 900 } 901 return result; 902 } else { 903 if (gclocker_retry_count > GCLockerRetryAllocationCount) { 904 return NULL; 905 } 906 assert(op.result() == NULL, 907 "the result should be NULL if the VM op did not succeed"); 908 } 909 910 // Give a warning if we seem to be looping forever. 911 if ((QueuedAllocationWarningCount > 0) && 912 (try_count % QueuedAllocationWarningCount == 0)) { 913 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 914 } 915 } 916 917 ShouldNotReachHere(); 918 return NULL; 919 } 920 921 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 922 unsigned int *gc_count_before_ret, 923 int* gclocker_retry_count_ret) { 924 // Make sure you read the note in attempt_allocation_humongous(). 925 926 assert_heap_not_locked_and_not_at_safepoint(); 927 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 928 "be called for humongous allocation requests"); 929 930 // We should only get here after the first-level allocation attempt 931 // (attempt_allocation()) failed to allocate. 932 933 // We will loop until a) we manage to successfully perform the 934 // allocation or b) we successfully schedule a collection which 935 // fails to perform the allocation. b) is the only case when we'll 936 // return NULL. 937 HeapWord* result = NULL; 938 for (int try_count = 1; /* we'll return */; try_count += 1) { 939 bool should_try_gc; 940 unsigned int gc_count_before; 941 942 { 943 MutexLockerEx x(Heap_lock); 944 945 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 946 false /* bot_updates */); 947 if (result != NULL) { 948 return result; 949 } 950 951 // If we reach here, attempt_allocation_locked() above failed to 952 // allocate a new region. So the mutator alloc region should be NULL. 953 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 954 955 if (GC_locker::is_active_and_needs_gc()) { 956 if (g1_policy()->can_expand_young_list()) { 957 // No need for an ergo verbose message here, 958 // can_expand_young_list() does this when it returns true. 959 result = _mutator_alloc_region.attempt_allocation_force(word_size, 960 false /* bot_updates */); 961 if (result != NULL) { 962 return result; 963 } 964 } 965 should_try_gc = false; 966 } else { 967 // The GCLocker may not be active but the GCLocker initiated 968 // GC may not yet have been performed (GCLocker::needs_gc() 969 // returns true). In this case we do not try this GC and 970 // wait until the GCLocker initiated GC is performed, and 971 // then retry the allocation. 972 if (GC_locker::needs_gc()) { 973 should_try_gc = false; 974 } else { 975 // Read the GC count while still holding the Heap_lock. 976 gc_count_before = total_collections(); 977 should_try_gc = true; 978 } 979 } 980 } 981 982 if (should_try_gc) { 983 bool succeeded; 984 result = do_collection_pause(word_size, gc_count_before, &succeeded); 985 if (result != NULL) { 986 assert(succeeded, "only way to get back a non-NULL result"); 987 return result; 988 } 989 990 if (succeeded) { 991 // If we get here we successfully scheduled a collection which 992 // failed to allocate. No point in trying to allocate 993 // further. We'll just return NULL. 994 MutexLockerEx x(Heap_lock); 995 *gc_count_before_ret = total_collections(); 996 return NULL; 997 } 998 } else { 999 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1000 MutexLockerEx x(Heap_lock); 1001 *gc_count_before_ret = total_collections(); 1002 return NULL; 1003 } 1004 // The GCLocker is either active or the GCLocker initiated 1005 // GC has not yet been performed. Stall until it is and 1006 // then retry the allocation. 1007 GC_locker::stall_until_clear(); 1008 (*gclocker_retry_count_ret) += 1; 1009 } 1010 1011 // We can reach here if we were unsuccessful in scheduling a 1012 // collection (because another thread beat us to it) or if we were 1013 // stalled due to the GC locker. In either can we should retry the 1014 // allocation attempt in case another thread successfully 1015 // performed a collection and reclaimed enough space. We do the 1016 // first attempt (without holding the Heap_lock) here and the 1017 // follow-on attempt will be at the start of the next loop 1018 // iteration (after taking the Heap_lock). 1019 result = _mutator_alloc_region.attempt_allocation(word_size, 1020 false /* bot_updates */); 1021 if (result != NULL) { 1022 return result; 1023 } 1024 1025 // Give a warning if we seem to be looping forever. 1026 if ((QueuedAllocationWarningCount > 0) && 1027 (try_count % QueuedAllocationWarningCount == 0)) { 1028 warning("G1CollectedHeap::attempt_allocation_slow() " 1029 "retries %d times", try_count); 1030 } 1031 } 1032 1033 ShouldNotReachHere(); 1034 return NULL; 1035 } 1036 1037 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1038 unsigned int * gc_count_before_ret, 1039 int* gclocker_retry_count_ret) { 1040 // The structure of this method has a lot of similarities to 1041 // attempt_allocation_slow(). The reason these two were not merged 1042 // into a single one is that such a method would require several "if 1043 // allocation is not humongous do this, otherwise do that" 1044 // conditional paths which would obscure its flow. In fact, an early 1045 // version of this code did use a unified method which was harder to 1046 // follow and, as a result, it had subtle bugs that were hard to 1047 // track down. So keeping these two methods separate allows each to 1048 // be more readable. It will be good to keep these two in sync as 1049 // much as possible. 1050 1051 assert_heap_not_locked_and_not_at_safepoint(); 1052 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1053 "should only be called for humongous allocations"); 1054 1055 // Humongous objects can exhaust the heap quickly, so we should check if we 1056 // need to start a marking cycle at each humongous object allocation. We do 1057 // the check before we do the actual allocation. The reason for doing it 1058 // before the allocation is that we avoid having to keep track of the newly 1059 // allocated memory while we do a GC. 1060 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1061 word_size)) { 1062 collect(GCCause::_g1_humongous_allocation); 1063 } 1064 1065 // We will loop until a) we manage to successfully perform the 1066 // allocation or b) we successfully schedule a collection which 1067 // fails to perform the allocation. b) is the only case when we'll 1068 // return NULL. 1069 HeapWord* result = NULL; 1070 for (int try_count = 1; /* we'll return */; try_count += 1) { 1071 bool should_try_gc; 1072 unsigned int gc_count_before; 1073 1074 { 1075 MutexLockerEx x(Heap_lock); 1076 1077 // Given that humongous objects are not allocated in young 1078 // regions, we'll first try to do the allocation without doing a 1079 // collection hoping that there's enough space in the heap. 1080 result = humongous_obj_allocate(word_size); 1081 if (result != NULL) { 1082 return result; 1083 } 1084 1085 if (GC_locker::is_active_and_needs_gc()) { 1086 should_try_gc = false; 1087 } else { 1088 // The GCLocker may not be active but the GCLocker initiated 1089 // GC may not yet have been performed (GCLocker::needs_gc() 1090 // returns true). In this case we do not try this GC and 1091 // wait until the GCLocker initiated GC is performed, and 1092 // then retry the allocation. 1093 if (GC_locker::needs_gc()) { 1094 should_try_gc = false; 1095 } else { 1096 // Read the GC count while still holding the Heap_lock. 1097 gc_count_before = total_collections(); 1098 should_try_gc = true; 1099 } 1100 } 1101 } 1102 1103 if (should_try_gc) { 1104 // If we failed to allocate the humongous object, we should try to 1105 // do a collection pause (if we're allowed) in case it reclaims 1106 // enough space for the allocation to succeed after the pause. 1107 1108 bool succeeded; 1109 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1110 if (result != NULL) { 1111 assert(succeeded, "only way to get back a non-NULL result"); 1112 return result; 1113 } 1114 1115 if (succeeded) { 1116 // If we get here we successfully scheduled a collection which 1117 // failed to allocate. No point in trying to allocate 1118 // further. We'll just return NULL. 1119 MutexLockerEx x(Heap_lock); 1120 *gc_count_before_ret = total_collections(); 1121 return NULL; 1122 } 1123 } else { 1124 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1125 MutexLockerEx x(Heap_lock); 1126 *gc_count_before_ret = total_collections(); 1127 return NULL; 1128 } 1129 // The GCLocker is either active or the GCLocker initiated 1130 // GC has not yet been performed. Stall until it is and 1131 // then retry the allocation. 1132 GC_locker::stall_until_clear(); 1133 (*gclocker_retry_count_ret) += 1; 1134 } 1135 1136 // We can reach here if we were unsuccessful in scheduling a 1137 // collection (because another thread beat us to it) or if we were 1138 // stalled due to the GC locker. In either can we should retry the 1139 // allocation attempt in case another thread successfully 1140 // performed a collection and reclaimed enough space. Give a 1141 // warning if we seem to be looping forever. 1142 1143 if ((QueuedAllocationWarningCount > 0) && 1144 (try_count % QueuedAllocationWarningCount == 0)) { 1145 warning("G1CollectedHeap::attempt_allocation_humongous() " 1146 "retries %d times", try_count); 1147 } 1148 } 1149 1150 ShouldNotReachHere(); 1151 return NULL; 1152 } 1153 1154 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1155 bool expect_null_mutator_alloc_region) { 1156 assert_at_safepoint(true /* should_be_vm_thread */); 1157 assert(_mutator_alloc_region.get() == NULL || 1158 !expect_null_mutator_alloc_region, 1159 "the current alloc region was unexpectedly found to be non-NULL"); 1160 1161 if (!isHumongous(word_size)) { 1162 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1163 false /* bot_updates */); 1164 } else { 1165 HeapWord* result = humongous_obj_allocate(word_size); 1166 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1167 g1_policy()->set_initiate_conc_mark_if_possible(); 1168 } 1169 return result; 1170 } 1171 1172 ShouldNotReachHere(); 1173 } 1174 1175 class PostMCRemSetClearClosure: public HeapRegionClosure { 1176 G1CollectedHeap* _g1h; 1177 ModRefBarrierSet* _mr_bs; 1178 public: 1179 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1180 _g1h(g1h), _mr_bs(mr_bs) {} 1181 1182 bool doHeapRegion(HeapRegion* r) { 1183 HeapRegionRemSet* hrrs = r->rem_set(); 1184 1185 if (r->continuesHumongous()) { 1186 // We'll assert that the strong code root list and RSet is empty 1187 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 1188 assert(hrrs->occupied() == 0, "RSet should be empty"); 1189 return false; 1190 } 1191 1192 _g1h->reset_gc_time_stamps(r); 1193 hrrs->clear(); 1194 // You might think here that we could clear just the cards 1195 // corresponding to the used region. But no: if we leave a dirty card 1196 // in a region we might allocate into, then it would prevent that card 1197 // from being enqueued, and cause it to be missed. 1198 // Re: the performance cost: we shouldn't be doing full GC anyway! 1199 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1200 1201 return false; 1202 } 1203 }; 1204 1205 void G1CollectedHeap::clear_rsets_post_compaction() { 1206 PostMCRemSetClearClosure rs_clear(this, mr_bs()); 1207 heap_region_iterate(&rs_clear); 1208 } 1209 1210 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1211 G1CollectedHeap* _g1h; 1212 UpdateRSOopClosure _cl; 1213 int _worker_i; 1214 public: 1215 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1216 _cl(g1->g1_rem_set(), worker_i), 1217 _worker_i(worker_i), 1218 _g1h(g1) 1219 { } 1220 1221 bool doHeapRegion(HeapRegion* r) { 1222 if (!r->continuesHumongous()) { 1223 _cl.set_from(r); 1224 r->oop_iterate(&_cl); 1225 } 1226 return false; 1227 } 1228 }; 1229 1230 class ParRebuildRSTask: public AbstractGangTask { 1231 G1CollectedHeap* _g1; 1232 public: 1233 ParRebuildRSTask(G1CollectedHeap* g1) 1234 : AbstractGangTask("ParRebuildRSTask"), 1235 _g1(g1) 1236 { } 1237 1238 void work(uint worker_id) { 1239 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1240 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1241 _g1->workers()->active_workers(), 1242 HeapRegion::RebuildRSClaimValue); 1243 } 1244 }; 1245 1246 class PostCompactionPrinterClosure: public HeapRegionClosure { 1247 private: 1248 G1HRPrinter* _hr_printer; 1249 public: 1250 bool doHeapRegion(HeapRegion* hr) { 1251 assert(!hr->is_young(), "not expecting to find young regions"); 1252 // We only generate output for non-empty regions. 1253 if (!hr->is_empty()) { 1254 if (!hr->isHumongous()) { 1255 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1256 } else if (hr->startsHumongous()) { 1257 if (hr->region_num() == 1) { 1258 // single humongous region 1259 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1260 } else { 1261 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1262 } 1263 } else { 1264 assert(hr->continuesHumongous(), "only way to get here"); 1265 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1266 } 1267 } 1268 return false; 1269 } 1270 1271 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1272 : _hr_printer(hr_printer) { } 1273 }; 1274 1275 void G1CollectedHeap::print_hrs_post_compaction() { 1276 PostCompactionPrinterClosure cl(hr_printer()); 1277 heap_region_iterate(&cl); 1278 } 1279 1280 bool G1CollectedHeap::do_collection(bool explicit_gc, 1281 bool clear_all_soft_refs, 1282 size_t word_size) { 1283 assert_at_safepoint(true /* should_be_vm_thread */); 1284 1285 if (GC_locker::check_active_before_gc()) { 1286 return false; 1287 } 1288 1289 STWGCTimer* gc_timer = G1MarkSweep::gc_timer(); 1290 gc_timer->register_gc_start(os::elapsed_counter()); 1291 1292 SerialOldTracer* gc_tracer = G1MarkSweep::gc_tracer(); 1293 gc_tracer->report_gc_start(gc_cause(), gc_timer->gc_start()); 1294 1295 SvcGCMarker sgcm(SvcGCMarker::FULL); 1296 ResourceMark rm; 1297 1298 print_heap_before_gc(); 1299 trace_heap_before_gc(gc_tracer); 1300 1301 size_t metadata_prev_used = MetaspaceAux::allocated_used_bytes(); 1302 1303 HRSPhaseSetter x(HRSPhaseFullGC); 1304 verify_region_sets_optional(); 1305 1306 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1307 collector_policy()->should_clear_all_soft_refs(); 1308 1309 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1310 1311 { 1312 IsGCActiveMark x; 1313 1314 // Timing 1315 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1316 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1317 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1318 1319 { 1320 GCTraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, NULL); 1321 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1322 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1323 1324 double start = os::elapsedTime(); 1325 g1_policy()->record_full_collection_start(); 1326 1327 // Note: When we have a more flexible GC logging framework that 1328 // allows us to add optional attributes to a GC log record we 1329 // could consider timing and reporting how long we wait in the 1330 // following two methods. 1331 wait_while_free_regions_coming(); 1332 // If we start the compaction before the CM threads finish 1333 // scanning the root regions we might trip them over as we'll 1334 // be moving objects / updating references. So let's wait until 1335 // they are done. By telling them to abort, they should complete 1336 // early. 1337 _cm->root_regions()->abort(); 1338 _cm->root_regions()->wait_until_scan_finished(); 1339 append_secondary_free_list_if_not_empty_with_lock(); 1340 1341 gc_prologue(true); 1342 increment_total_collections(true /* full gc */); 1343 increment_old_marking_cycles_started(); 1344 1345 assert(used() == recalculate_used(), "Should be equal"); 1346 1347 verify_before_gc(); 1348 1349 pre_full_gc_dump(gc_timer); 1350 1351 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1352 1353 // Disable discovery and empty the discovered lists 1354 // for the CM ref processor. 1355 ref_processor_cm()->disable_discovery(); 1356 ref_processor_cm()->abandon_partial_discovery(); 1357 ref_processor_cm()->verify_no_references_recorded(); 1358 1359 // Abandon current iterations of concurrent marking and concurrent 1360 // refinement, if any are in progress. We have to do this before 1361 // wait_until_scan_finished() below. 1362 concurrent_mark()->abort(); 1363 1364 // Make sure we'll choose a new allocation region afterwards. 1365 release_mutator_alloc_region(); 1366 abandon_gc_alloc_regions(); 1367 g1_rem_set()->cleanupHRRS(); 1368 1369 // We should call this after we retire any currently active alloc 1370 // regions so that all the ALLOC / RETIRE events are generated 1371 // before the start GC event. 1372 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1373 1374 // We may have added regions to the current incremental collection 1375 // set between the last GC or pause and now. We need to clear the 1376 // incremental collection set and then start rebuilding it afresh 1377 // after this full GC. 1378 abandon_collection_set(g1_policy()->inc_cset_head()); 1379 g1_policy()->clear_incremental_cset(); 1380 g1_policy()->stop_incremental_cset_building(); 1381 1382 tear_down_region_sets(false /* free_list_only */); 1383 g1_policy()->set_gcs_are_young(true); 1384 1385 // See the comments in g1CollectedHeap.hpp and 1386 // G1CollectedHeap::ref_processing_init() about 1387 // how reference processing currently works in G1. 1388 1389 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1390 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1391 1392 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1393 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1394 1395 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1396 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1397 1398 // Do collection work 1399 { 1400 HandleMark hm; // Discard invalid handles created during gc 1401 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1402 } 1403 1404 assert(free_regions() == 0, "we should not have added any free regions"); 1405 rebuild_region_sets(false /* free_list_only */); 1406 1407 // Enqueue any discovered reference objects that have 1408 // not been removed from the discovered lists. 1409 ref_processor_stw()->enqueue_discovered_references(); 1410 1411 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1412 1413 MemoryService::track_memory_usage(); 1414 1415 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1416 ref_processor_stw()->verify_no_references_recorded(); 1417 1418 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1419 ClassLoaderDataGraph::purge(); 1420 MetaspaceAux::verify_metrics(); 1421 1422 // Note: since we've just done a full GC, concurrent 1423 // marking is no longer active. Therefore we need not 1424 // re-enable reference discovery for the CM ref processor. 1425 // That will be done at the start of the next marking cycle. 1426 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1427 ref_processor_cm()->verify_no_references_recorded(); 1428 1429 reset_gc_time_stamp(); 1430 // Since everything potentially moved, we will clear all remembered 1431 // sets, and clear all cards. Later we will rebuild remembered 1432 // sets. We will also reset the GC time stamps of the regions. 1433 clear_rsets_post_compaction(); 1434 check_gc_time_stamps(); 1435 1436 // Resize the heap if necessary. 1437 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1438 1439 if (_hr_printer.is_active()) { 1440 // We should do this after we potentially resize the heap so 1441 // that all the COMMIT / UNCOMMIT events are generated before 1442 // the end GC event. 1443 1444 print_hrs_post_compaction(); 1445 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1446 } 1447 1448 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 1449 if (hot_card_cache->use_cache()) { 1450 hot_card_cache->reset_card_counts(); 1451 hot_card_cache->reset_hot_cache(); 1452 } 1453 1454 // Rebuild remembered sets of all regions. 1455 if (G1CollectedHeap::use_parallel_gc_threads()) { 1456 uint n_workers = 1457 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1458 workers()->active_workers(), 1459 Threads::number_of_non_daemon_threads()); 1460 assert(UseDynamicNumberOfGCThreads || 1461 n_workers == workers()->total_workers(), 1462 "If not dynamic should be using all the workers"); 1463 workers()->set_active_workers(n_workers); 1464 // Set parallel threads in the heap (_n_par_threads) only 1465 // before a parallel phase and always reset it to 0 after 1466 // the phase so that the number of parallel threads does 1467 // no get carried forward to a serial phase where there 1468 // may be code that is "possibly_parallel". 1469 set_par_threads(n_workers); 1470 1471 ParRebuildRSTask rebuild_rs_task(this); 1472 assert(check_heap_region_claim_values( 1473 HeapRegion::InitialClaimValue), "sanity check"); 1474 assert(UseDynamicNumberOfGCThreads || 1475 workers()->active_workers() == workers()->total_workers(), 1476 "Unless dynamic should use total workers"); 1477 // Use the most recent number of active workers 1478 assert(workers()->active_workers() > 0, 1479 "Active workers not properly set"); 1480 set_par_threads(workers()->active_workers()); 1481 workers()->run_task(&rebuild_rs_task); 1482 set_par_threads(0); 1483 assert(check_heap_region_claim_values( 1484 HeapRegion::RebuildRSClaimValue), "sanity check"); 1485 reset_heap_region_claim_values(); 1486 } else { 1487 RebuildRSOutOfRegionClosure rebuild_rs(this); 1488 heap_region_iterate(&rebuild_rs); 1489 } 1490 1491 // Rebuild the strong code root lists for each region 1492 rebuild_strong_code_roots(); 1493 1494 if (true) { // FIXME 1495 MetaspaceGC::compute_new_size(); 1496 } 1497 1498 #ifdef TRACESPINNING 1499 ParallelTaskTerminator::print_termination_counts(); 1500 #endif 1501 1502 // Discard all rset updates 1503 JavaThread::dirty_card_queue_set().abandon_logs(); 1504 assert(!G1DeferredRSUpdate 1505 || (G1DeferredRSUpdate && 1506 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1507 1508 _young_list->reset_sampled_info(); 1509 // At this point there should be no regions in the 1510 // entire heap tagged as young. 1511 assert(check_young_list_empty(true /* check_heap */), 1512 "young list should be empty at this point"); 1513 1514 // Update the number of full collections that have been completed. 1515 increment_old_marking_cycles_completed(false /* concurrent */); 1516 1517 _hrs.verify_optional(); 1518 verify_region_sets_optional(); 1519 1520 verify_after_gc(); 1521 1522 // Start a new incremental collection set for the next pause 1523 assert(g1_policy()->collection_set() == NULL, "must be"); 1524 g1_policy()->start_incremental_cset_building(); 1525 1526 // Clear the _cset_fast_test bitmap in anticipation of adding 1527 // regions to the incremental collection set for the next 1528 // evacuation pause. 1529 clear_cset_fast_test(); 1530 1531 init_mutator_alloc_region(); 1532 1533 double end = os::elapsedTime(); 1534 g1_policy()->record_full_collection_end(); 1535 1536 if (G1Log::fine()) { 1537 g1_policy()->print_heap_transition(); 1538 } 1539 1540 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1541 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1542 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1543 // before any GC notifications are raised. 1544 g1mm()->update_sizes(); 1545 1546 gc_epilogue(true); 1547 } 1548 1549 if (G1Log::finer()) { 1550 g1_policy()->print_detailed_heap_transition(true /* full */); 1551 } 1552 1553 print_heap_after_gc(); 1554 trace_heap_after_gc(gc_tracer); 1555 1556 post_full_gc_dump(gc_timer); 1557 1558 gc_timer->register_gc_end(os::elapsed_counter()); 1559 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1560 } 1561 1562 return true; 1563 } 1564 1565 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1566 // do_collection() will return whether it succeeded in performing 1567 // the GC. Currently, there is no facility on the 1568 // do_full_collection() API to notify the caller than the collection 1569 // did not succeed (e.g., because it was locked out by the GC 1570 // locker). So, right now, we'll ignore the return value. 1571 bool dummy = do_collection(true, /* explicit_gc */ 1572 clear_all_soft_refs, 1573 0 /* word_size */); 1574 } 1575 1576 // This code is mostly copied from TenuredGeneration. 1577 void 1578 G1CollectedHeap:: 1579 resize_if_necessary_after_full_collection(size_t word_size) { 1580 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1581 1582 // Include the current allocation, if any, and bytes that will be 1583 // pre-allocated to support collections, as "used". 1584 const size_t used_after_gc = used(); 1585 const size_t capacity_after_gc = capacity(); 1586 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1587 1588 // This is enforced in arguments.cpp. 1589 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1590 "otherwise the code below doesn't make sense"); 1591 1592 // We don't have floating point command-line arguments 1593 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1594 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1595 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1596 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1597 1598 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1599 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1600 1601 // We have to be careful here as these two calculations can overflow 1602 // 32-bit size_t's. 1603 double used_after_gc_d = (double) used_after_gc; 1604 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1605 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1606 1607 // Let's make sure that they are both under the max heap size, which 1608 // by default will make them fit into a size_t. 1609 double desired_capacity_upper_bound = (double) max_heap_size; 1610 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1611 desired_capacity_upper_bound); 1612 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1613 desired_capacity_upper_bound); 1614 1615 // We can now safely turn them into size_t's. 1616 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1617 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1618 1619 // This assert only makes sense here, before we adjust them 1620 // with respect to the min and max heap size. 1621 assert(minimum_desired_capacity <= maximum_desired_capacity, 1622 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1623 "maximum_desired_capacity = "SIZE_FORMAT, 1624 minimum_desired_capacity, maximum_desired_capacity)); 1625 1626 // Should not be greater than the heap max size. No need to adjust 1627 // it with respect to the heap min size as it's a lower bound (i.e., 1628 // we'll try to make the capacity larger than it, not smaller). 1629 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1630 // Should not be less than the heap min size. No need to adjust it 1631 // with respect to the heap max size as it's an upper bound (i.e., 1632 // we'll try to make the capacity smaller than it, not greater). 1633 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1634 1635 if (capacity_after_gc < minimum_desired_capacity) { 1636 // Don't expand unless it's significant 1637 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1638 ergo_verbose4(ErgoHeapSizing, 1639 "attempt heap expansion", 1640 ergo_format_reason("capacity lower than " 1641 "min desired capacity after Full GC") 1642 ergo_format_byte("capacity") 1643 ergo_format_byte("occupancy") 1644 ergo_format_byte_perc("min desired capacity"), 1645 capacity_after_gc, used_after_gc, 1646 minimum_desired_capacity, (double) MinHeapFreeRatio); 1647 expand(expand_bytes); 1648 1649 // No expansion, now see if we want to shrink 1650 } else if (capacity_after_gc > maximum_desired_capacity) { 1651 // Capacity too large, compute shrinking size 1652 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1653 ergo_verbose4(ErgoHeapSizing, 1654 "attempt heap shrinking", 1655 ergo_format_reason("capacity higher than " 1656 "max desired capacity after Full GC") 1657 ergo_format_byte("capacity") 1658 ergo_format_byte("occupancy") 1659 ergo_format_byte_perc("max desired capacity"), 1660 capacity_after_gc, used_after_gc, 1661 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1662 shrink(shrink_bytes); 1663 } 1664 } 1665 1666 1667 HeapWord* 1668 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1669 bool* succeeded) { 1670 assert_at_safepoint(true /* should_be_vm_thread */); 1671 1672 *succeeded = true; 1673 // Let's attempt the allocation first. 1674 HeapWord* result = 1675 attempt_allocation_at_safepoint(word_size, 1676 false /* expect_null_mutator_alloc_region */); 1677 if (result != NULL) { 1678 assert(*succeeded, "sanity"); 1679 return result; 1680 } 1681 1682 // In a G1 heap, we're supposed to keep allocation from failing by 1683 // incremental pauses. Therefore, at least for now, we'll favor 1684 // expansion over collection. (This might change in the future if we can 1685 // do something smarter than full collection to satisfy a failed alloc.) 1686 result = expand_and_allocate(word_size); 1687 if (result != NULL) { 1688 assert(*succeeded, "sanity"); 1689 return result; 1690 } 1691 1692 // Expansion didn't work, we'll try to do a Full GC. 1693 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1694 false, /* clear_all_soft_refs */ 1695 word_size); 1696 if (!gc_succeeded) { 1697 *succeeded = false; 1698 return NULL; 1699 } 1700 1701 // Retry the allocation 1702 result = attempt_allocation_at_safepoint(word_size, 1703 true /* expect_null_mutator_alloc_region */); 1704 if (result != NULL) { 1705 assert(*succeeded, "sanity"); 1706 return result; 1707 } 1708 1709 // Then, try a Full GC that will collect all soft references. 1710 gc_succeeded = do_collection(false, /* explicit_gc */ 1711 true, /* clear_all_soft_refs */ 1712 word_size); 1713 if (!gc_succeeded) { 1714 *succeeded = false; 1715 return NULL; 1716 } 1717 1718 // Retry the allocation once more 1719 result = attempt_allocation_at_safepoint(word_size, 1720 true /* expect_null_mutator_alloc_region */); 1721 if (result != NULL) { 1722 assert(*succeeded, "sanity"); 1723 return result; 1724 } 1725 1726 assert(!collector_policy()->should_clear_all_soft_refs(), 1727 "Flag should have been handled and cleared prior to this point"); 1728 1729 // What else? We might try synchronous finalization later. If the total 1730 // space available is large enough for the allocation, then a more 1731 // complete compaction phase than we've tried so far might be 1732 // appropriate. 1733 assert(*succeeded, "sanity"); 1734 return NULL; 1735 } 1736 1737 // Attempting to expand the heap sufficiently 1738 // to support an allocation of the given "word_size". If 1739 // successful, perform the allocation and return the address of the 1740 // allocated block, or else "NULL". 1741 1742 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1743 assert_at_safepoint(true /* should_be_vm_thread */); 1744 1745 verify_region_sets_optional(); 1746 1747 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1748 ergo_verbose1(ErgoHeapSizing, 1749 "attempt heap expansion", 1750 ergo_format_reason("allocation request failed") 1751 ergo_format_byte("allocation request"), 1752 word_size * HeapWordSize); 1753 if (expand(expand_bytes)) { 1754 _hrs.verify_optional(); 1755 verify_region_sets_optional(); 1756 return attempt_allocation_at_safepoint(word_size, 1757 false /* expect_null_mutator_alloc_region */); 1758 } 1759 return NULL; 1760 } 1761 1762 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1763 HeapWord* new_end) { 1764 assert(old_end != new_end, "don't call this otherwise"); 1765 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1766 1767 // Update the committed mem region. 1768 _g1_committed.set_end(new_end); 1769 // Tell the card table about the update. 1770 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1771 // Tell the BOT about the update. 1772 _bot_shared->resize(_g1_committed.word_size()); 1773 // Tell the hot card cache about the update 1774 _cg1r->hot_card_cache()->resize_card_counts(capacity()); 1775 } 1776 1777 bool G1CollectedHeap::expand(size_t expand_bytes) { 1778 size_t old_mem_size = _g1_storage.committed_size(); 1779 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1780 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1781 HeapRegion::GrainBytes); 1782 ergo_verbose2(ErgoHeapSizing, 1783 "expand the heap", 1784 ergo_format_byte("requested expansion amount") 1785 ergo_format_byte("attempted expansion amount"), 1786 expand_bytes, aligned_expand_bytes); 1787 1788 // First commit the memory. 1789 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1790 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1791 if (successful) { 1792 // Then propagate this update to the necessary data structures. 1793 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1794 update_committed_space(old_end, new_end); 1795 1796 FreeRegionList expansion_list("Local Expansion List"); 1797 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1798 assert(mr.start() == old_end, "post-condition"); 1799 // mr might be a smaller region than what was requested if 1800 // expand_by() was unable to allocate the HeapRegion instances 1801 assert(mr.end() <= new_end, "post-condition"); 1802 1803 size_t actual_expand_bytes = mr.byte_size(); 1804 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1805 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1806 "post-condition"); 1807 if (actual_expand_bytes < aligned_expand_bytes) { 1808 // We could not expand _hrs to the desired size. In this case we 1809 // need to shrink the committed space accordingly. 1810 assert(mr.end() < new_end, "invariant"); 1811 1812 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1813 // First uncommit the memory. 1814 _g1_storage.shrink_by(diff_bytes); 1815 // Then propagate this update to the necessary data structures. 1816 update_committed_space(new_end, mr.end()); 1817 } 1818 _free_list.add_as_tail(&expansion_list); 1819 1820 if (_hr_printer.is_active()) { 1821 HeapWord* curr = mr.start(); 1822 while (curr < mr.end()) { 1823 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1824 _hr_printer.commit(curr, curr_end); 1825 curr = curr_end; 1826 } 1827 assert(curr == mr.end(), "post-condition"); 1828 } 1829 g1_policy()->record_new_heap_size(n_regions()); 1830 } else { 1831 ergo_verbose0(ErgoHeapSizing, 1832 "did not expand the heap", 1833 ergo_format_reason("heap expansion operation failed")); 1834 // The expansion of the virtual storage space was unsuccessful. 1835 // Let's see if it was because we ran out of swap. 1836 if (G1ExitOnExpansionFailure && 1837 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1838 // We had head room... 1839 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion"); 1840 } 1841 } 1842 return successful; 1843 } 1844 1845 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1846 size_t old_mem_size = _g1_storage.committed_size(); 1847 size_t aligned_shrink_bytes = 1848 ReservedSpace::page_align_size_down(shrink_bytes); 1849 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1850 HeapRegion::GrainBytes); 1851 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes); 1852 1853 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove); 1854 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1855 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes; 1856 1857 ergo_verbose3(ErgoHeapSizing, 1858 "shrink the heap", 1859 ergo_format_byte("requested shrinking amount") 1860 ergo_format_byte("aligned shrinking amount") 1861 ergo_format_byte("attempted shrinking amount"), 1862 shrink_bytes, aligned_shrink_bytes, shrunk_bytes); 1863 if (num_regions_removed > 0) { 1864 _g1_storage.shrink_by(shrunk_bytes); 1865 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1866 1867 if (_hr_printer.is_active()) { 1868 HeapWord* curr = old_end; 1869 while (curr > new_end) { 1870 HeapWord* curr_end = curr; 1871 curr -= HeapRegion::GrainWords; 1872 _hr_printer.uncommit(curr, curr_end); 1873 } 1874 } 1875 1876 _expansion_regions += num_regions_removed; 1877 update_committed_space(old_end, new_end); 1878 HeapRegionRemSet::shrink_heap(n_regions()); 1879 g1_policy()->record_new_heap_size(n_regions()); 1880 } else { 1881 ergo_verbose0(ErgoHeapSizing, 1882 "did not shrink the heap", 1883 ergo_format_reason("heap shrinking operation failed")); 1884 } 1885 } 1886 1887 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1888 verify_region_sets_optional(); 1889 1890 // We should only reach here at the end of a Full GC which means we 1891 // should not not be holding to any GC alloc regions. The method 1892 // below will make sure of that and do any remaining clean up. 1893 abandon_gc_alloc_regions(); 1894 1895 // Instead of tearing down / rebuilding the free lists here, we 1896 // could instead use the remove_all_pending() method on free_list to 1897 // remove only the ones that we need to remove. 1898 tear_down_region_sets(true /* free_list_only */); 1899 shrink_helper(shrink_bytes); 1900 rebuild_region_sets(true /* free_list_only */); 1901 1902 _hrs.verify_optional(); 1903 verify_region_sets_optional(); 1904 } 1905 1906 // Public methods. 1907 1908 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1909 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1910 #endif // _MSC_VER 1911 1912 1913 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1914 SharedHeap(policy_), 1915 _g1_policy(policy_), 1916 _dirty_card_queue_set(false), 1917 _into_cset_dirty_card_queue_set(false), 1918 _is_alive_closure_cm(this), 1919 _is_alive_closure_stw(this), 1920 _ref_processor_cm(NULL), 1921 _ref_processor_stw(NULL), 1922 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1923 _bot_shared(NULL), 1924 _evac_failure_scan_stack(NULL), 1925 _mark_in_progress(false), 1926 _cg1r(NULL), _summary_bytes_used(0), 1927 _g1mm(NULL), 1928 _refine_cte_cl(NULL), 1929 _full_collection(false), 1930 _free_list("Master Free List"), 1931 _secondary_free_list("Secondary Free List"), 1932 _old_set("Old Set"), 1933 _humongous_set("Master Humongous Set"), 1934 _free_regions_coming(false), 1935 _young_list(new YoungList(this)), 1936 _gc_time_stamp(0), 1937 _retained_old_gc_alloc_region(NULL), 1938 _survivor_plab_stats(YoungPLABSize, PLABWeight), 1939 _old_plab_stats(OldPLABSize, PLABWeight), 1940 _expand_heap_after_alloc_failure(true), 1941 _surviving_young_words(NULL), 1942 _old_marking_cycles_started(0), 1943 _old_marking_cycles_completed(0), 1944 _concurrent_cycle_started(false), 1945 _in_cset_fast_test(NULL), 1946 _in_cset_fast_test_base(NULL), 1947 _dirty_cards_region_list(NULL), 1948 _worker_cset_start_region(NULL), 1949 _worker_cset_start_region_time_stamp(NULL), 1950 _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()), 1951 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 1952 _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()), 1953 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()) { 1954 1955 _g1h = this; 1956 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1957 vm_exit_during_initialization("Failed necessary allocation."); 1958 } 1959 1960 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1961 1962 int n_queues = MAX2((int)ParallelGCThreads, 1); 1963 _task_queues = new RefToScanQueueSet(n_queues); 1964 1965 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1966 assert(n_rem_sets > 0, "Invariant."); 1967 1968 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1969 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1970 _evacuation_failed_info_array = NEW_C_HEAP_ARRAY(EvacuationFailedInfo, n_queues, mtGC); 1971 1972 for (int i = 0; i < n_queues; i++) { 1973 RefToScanQueue* q = new RefToScanQueue(); 1974 q->initialize(); 1975 _task_queues->register_queue(i, q); 1976 ::new (&_evacuation_failed_info_array[i]) EvacuationFailedInfo(); 1977 } 1978 clear_cset_start_regions(); 1979 1980 // Initialize the G1EvacuationFailureALot counters and flags. 1981 NOT_PRODUCT(reset_evacuation_should_fail();) 1982 1983 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1984 } 1985 1986 jint G1CollectedHeap::initialize() { 1987 CollectedHeap::pre_initialize(); 1988 os::enable_vtime(); 1989 1990 G1Log::init(); 1991 1992 // Necessary to satisfy locking discipline assertions. 1993 1994 MutexLocker x(Heap_lock); 1995 1996 // We have to initialize the printer before committing the heap, as 1997 // it will be used then. 1998 _hr_printer.set_active(G1PrintHeapRegions); 1999 2000 // While there are no constraints in the GC code that HeapWordSize 2001 // be any particular value, there are multiple other areas in the 2002 // system which believe this to be true (e.g. oop->object_size in some 2003 // cases incorrectly returns the size in wordSize units rather than 2004 // HeapWordSize). 2005 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 2006 2007 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 2008 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 2009 2010 // Ensure that the sizes are properly aligned. 2011 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2012 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2013 2014 _cg1r = new ConcurrentG1Refine(this); 2015 2016 // Reserve the maximum. 2017 2018 // When compressed oops are enabled, the preferred heap base 2019 // is calculated by subtracting the requested size from the 2020 // 32Gb boundary and using the result as the base address for 2021 // heap reservation. If the requested size is not aligned to 2022 // HeapRegion::GrainBytes (i.e. the alignment that is passed 2023 // into the ReservedHeapSpace constructor) then the actual 2024 // base of the reserved heap may end up differing from the 2025 // address that was requested (i.e. the preferred heap base). 2026 // If this happens then we could end up using a non-optimal 2027 // compressed oops mode. 2028 2029 // Since max_byte_size is aligned to the size of a heap region (checked 2030 // above). 2031 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2032 2033 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2034 HeapRegion::GrainBytes); 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::ModRef)) { 2049 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2050 } else { 2051 vm_exit_during_initialization("G1 requires a mod ref bs."); 2052 return JNI_ENOMEM; 2053 } 2054 2055 // Also create a G1 rem set. 2056 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2057 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2058 } else { 2059 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2060 return JNI_ENOMEM; 2061 } 2062 2063 // Carve out the G1 part of the heap. 2064 2065 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2066 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2067 g1_rs.size()/HeapWordSize); 2068 2069 _g1_storage.initialize(g1_rs, 0); 2070 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2071 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2072 (HeapWord*) _g1_reserved.end(), 2073 _expansion_regions); 2074 2075 // Do later initialization work for concurrent refinement. 2076 _cg1r->init(); 2077 2078 // 6843694 - ensure that the maximum region index can fit 2079 // in the remembered set structures. 2080 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2081 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2082 2083 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2084 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2085 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2086 "too many cards per region"); 2087 2088 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2089 2090 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2091 heap_word_size(init_byte_size)); 2092 2093 _g1h = this; 2094 2095 _in_cset_fast_test_length = max_regions(); 2096 _in_cset_fast_test_base = 2097 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2098 2099 // We're biasing _in_cset_fast_test to avoid subtracting the 2100 // beginning of the heap every time we want to index; basically 2101 // it's the same with what we do with the card table. 2102 _in_cset_fast_test = _in_cset_fast_test_base - 2103 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2104 2105 // Clear the _cset_fast_test bitmap in anticipation of adding 2106 // regions to the incremental collection set for the first 2107 // evacuation pause. 2108 clear_cset_fast_test(); 2109 2110 // Create the ConcurrentMark data structure and thread. 2111 // (Must do this late, so that "max_regions" is defined.) 2112 _cm = new ConcurrentMark(this, heap_rs); 2113 if (_cm == NULL || !_cm->completed_initialization()) { 2114 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2115 return JNI_ENOMEM; 2116 } 2117 _cmThread = _cm->cmThread(); 2118 2119 // Initialize the from_card cache structure of HeapRegionRemSet. 2120 HeapRegionRemSet::init_heap(max_regions()); 2121 2122 // Now expand into the initial heap size. 2123 if (!expand(init_byte_size)) { 2124 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2125 return JNI_ENOMEM; 2126 } 2127 2128 // Perform any initialization actions delegated to the policy. 2129 g1_policy()->init(); 2130 2131 _refine_cte_cl = 2132 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2133 g1_rem_set(), 2134 concurrent_g1_refine()); 2135 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2136 2137 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2138 SATB_Q_FL_lock, 2139 G1SATBProcessCompletedThreshold, 2140 Shared_SATB_Q_lock); 2141 2142 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2143 DirtyCardQ_FL_lock, 2144 concurrent_g1_refine()->yellow_zone(), 2145 concurrent_g1_refine()->red_zone(), 2146 Shared_DirtyCardQ_lock); 2147 2148 if (G1DeferredRSUpdate) { 2149 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2150 DirtyCardQ_FL_lock, 2151 -1, // never trigger processing 2152 -1, // no limit on length 2153 Shared_DirtyCardQ_lock, 2154 &JavaThread::dirty_card_queue_set()); 2155 } 2156 2157 // Initialize the card queue set used to hold cards containing 2158 // references into the collection set. 2159 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2160 DirtyCardQ_FL_lock, 2161 -1, // never trigger processing 2162 -1, // no limit on length 2163 Shared_DirtyCardQ_lock, 2164 &JavaThread::dirty_card_queue_set()); 2165 2166 // In case we're keeping closure specialization stats, initialize those 2167 // counts and that mechanism. 2168 SpecializationStats::clear(); 2169 2170 // Here we allocate the dummy full region that is required by the 2171 // G1AllocRegion class. If we don't pass an address in the reserved 2172 // space here, lots of asserts fire. 2173 2174 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2175 _g1_reserved.start()); 2176 // We'll re-use the same region whether the alloc region will 2177 // require BOT updates or not and, if it doesn't, then a non-young 2178 // region will complain that it cannot support allocations without 2179 // BOT updates. So we'll tag the dummy region as young to avoid that. 2180 dummy_region->set_young(); 2181 // Make sure it's full. 2182 dummy_region->set_top(dummy_region->end()); 2183 G1AllocRegion::setup(this, dummy_region); 2184 2185 init_mutator_alloc_region(); 2186 2187 // Do create of the monitoring and management support so that 2188 // values in the heap have been properly initialized. 2189 _g1mm = new G1MonitoringSupport(this); 2190 2191 return JNI_OK; 2192 } 2193 2194 void G1CollectedHeap::ref_processing_init() { 2195 // Reference processing in G1 currently works as follows: 2196 // 2197 // * There are two reference processor instances. One is 2198 // used to record and process discovered references 2199 // during concurrent marking; the other is used to 2200 // record and process references during STW pauses 2201 // (both full and incremental). 2202 // * Both ref processors need to 'span' the entire heap as 2203 // the regions in the collection set may be dotted around. 2204 // 2205 // * For the concurrent marking ref processor: 2206 // * Reference discovery is enabled at initial marking. 2207 // * Reference discovery is disabled and the discovered 2208 // references processed etc during remarking. 2209 // * Reference discovery is MT (see below). 2210 // * Reference discovery requires a barrier (see below). 2211 // * Reference processing may or may not be MT 2212 // (depending on the value of ParallelRefProcEnabled 2213 // and ParallelGCThreads). 2214 // * A full GC disables reference discovery by the CM 2215 // ref processor and abandons any entries on it's 2216 // discovered lists. 2217 // 2218 // * For the STW processor: 2219 // * Non MT discovery is enabled at the start of a full GC. 2220 // * Processing and enqueueing during a full GC is non-MT. 2221 // * During a full GC, references are processed after marking. 2222 // 2223 // * Discovery (may or may not be MT) is enabled at the start 2224 // of an incremental evacuation pause. 2225 // * References are processed near the end of a STW evacuation pause. 2226 // * For both types of GC: 2227 // * Discovery is atomic - i.e. not concurrent. 2228 // * Reference discovery will not need a barrier. 2229 2230 SharedHeap::ref_processing_init(); 2231 MemRegion mr = reserved_region(); 2232 2233 // Concurrent Mark ref processor 2234 _ref_processor_cm = 2235 new ReferenceProcessor(mr, // span 2236 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2237 // mt processing 2238 (int) ParallelGCThreads, 2239 // degree of mt processing 2240 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2241 // mt discovery 2242 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2243 // degree of mt discovery 2244 false, 2245 // Reference discovery is not atomic 2246 &_is_alive_closure_cm, 2247 // is alive closure 2248 // (for efficiency/performance) 2249 true); 2250 // Setting next fields of discovered 2251 // lists requires a barrier. 2252 2253 // STW ref processor 2254 _ref_processor_stw = 2255 new ReferenceProcessor(mr, // span 2256 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2257 // mt processing 2258 MAX2((int)ParallelGCThreads, 1), 2259 // degree of mt processing 2260 (ParallelGCThreads > 1), 2261 // mt discovery 2262 MAX2((int)ParallelGCThreads, 1), 2263 // degree of mt discovery 2264 true, 2265 // Reference discovery is atomic 2266 &_is_alive_closure_stw, 2267 // is alive closure 2268 // (for efficiency/performance) 2269 false); 2270 // Setting next fields of discovered 2271 // lists requires a barrier. 2272 } 2273 2274 size_t G1CollectedHeap::capacity() const { 2275 return _g1_committed.byte_size(); 2276 } 2277 2278 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2279 assert(!hr->continuesHumongous(), "pre-condition"); 2280 hr->reset_gc_time_stamp(); 2281 if (hr->startsHumongous()) { 2282 uint first_index = hr->hrs_index() + 1; 2283 uint last_index = hr->last_hc_index(); 2284 for (uint i = first_index; i < last_index; i += 1) { 2285 HeapRegion* chr = region_at(i); 2286 assert(chr->continuesHumongous(), "sanity"); 2287 chr->reset_gc_time_stamp(); 2288 } 2289 } 2290 } 2291 2292 #ifndef PRODUCT 2293 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2294 private: 2295 unsigned _gc_time_stamp; 2296 bool _failures; 2297 2298 public: 2299 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2300 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2301 2302 virtual bool doHeapRegion(HeapRegion* hr) { 2303 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2304 if (_gc_time_stamp != region_gc_time_stamp) { 2305 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2306 "expected %d", HR_FORMAT_PARAMS(hr), 2307 region_gc_time_stamp, _gc_time_stamp); 2308 _failures = true; 2309 } 2310 return false; 2311 } 2312 2313 bool failures() { return _failures; } 2314 }; 2315 2316 void G1CollectedHeap::check_gc_time_stamps() { 2317 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2318 heap_region_iterate(&cl); 2319 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2320 } 2321 #endif // PRODUCT 2322 2323 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2324 DirtyCardQueue* into_cset_dcq, 2325 bool concurrent, 2326 int worker_i) { 2327 // Clean cards in the hot card cache 2328 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 2329 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq); 2330 2331 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2332 int n_completed_buffers = 0; 2333 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2334 n_completed_buffers++; 2335 } 2336 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2337 dcqs.clear_n_completed_buffers(); 2338 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2339 } 2340 2341 2342 // Computes the sum of the storage used by the various regions. 2343 2344 size_t G1CollectedHeap::used() const { 2345 assert(Heap_lock->owner() != NULL, 2346 "Should be owned on this thread's behalf."); 2347 size_t result = _summary_bytes_used; 2348 // Read only once in case it is set to NULL concurrently 2349 HeapRegion* hr = _mutator_alloc_region.get(); 2350 if (hr != NULL) 2351 result += hr->used(); 2352 return result; 2353 } 2354 2355 size_t G1CollectedHeap::used_unlocked() const { 2356 size_t result = _summary_bytes_used; 2357 return result; 2358 } 2359 2360 class SumUsedClosure: public HeapRegionClosure { 2361 size_t _used; 2362 public: 2363 SumUsedClosure() : _used(0) {} 2364 bool doHeapRegion(HeapRegion* r) { 2365 if (!r->continuesHumongous()) { 2366 _used += r->used(); 2367 } 2368 return false; 2369 } 2370 size_t result() { return _used; } 2371 }; 2372 2373 size_t G1CollectedHeap::recalculate_used() const { 2374 SumUsedClosure blk; 2375 heap_region_iterate(&blk); 2376 return blk.result(); 2377 } 2378 2379 size_t G1CollectedHeap::unsafe_max_alloc() { 2380 if (free_regions() > 0) return HeapRegion::GrainBytes; 2381 // otherwise, is there space in the current allocation region? 2382 2383 // We need to store the current allocation region in a local variable 2384 // here. The problem is that this method doesn't take any locks and 2385 // there may be other threads which overwrite the current allocation 2386 // region field. attempt_allocation(), for example, sets it to NULL 2387 // and this can happen *after* the NULL check here but before the call 2388 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2389 // to be a problem in the optimized build, since the two loads of the 2390 // current allocation region field are optimized away. 2391 HeapRegion* hr = _mutator_alloc_region.get(); 2392 if (hr == NULL) { 2393 return 0; 2394 } 2395 return hr->free(); 2396 } 2397 2398 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2399 switch (cause) { 2400 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2401 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2402 case GCCause::_g1_humongous_allocation: return true; 2403 default: return false; 2404 } 2405 } 2406 2407 #ifndef PRODUCT 2408 void G1CollectedHeap::allocate_dummy_regions() { 2409 // Let's fill up most of the region 2410 size_t word_size = HeapRegion::GrainWords - 1024; 2411 // And as a result the region we'll allocate will be humongous. 2412 guarantee(isHumongous(word_size), "sanity"); 2413 2414 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2415 // Let's use the existing mechanism for the allocation 2416 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2417 if (dummy_obj != NULL) { 2418 MemRegion mr(dummy_obj, word_size); 2419 CollectedHeap::fill_with_object(mr); 2420 } else { 2421 // If we can't allocate once, we probably cannot allocate 2422 // again. Let's get out of the loop. 2423 break; 2424 } 2425 } 2426 } 2427 #endif // !PRODUCT 2428 2429 void G1CollectedHeap::increment_old_marking_cycles_started() { 2430 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2431 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2432 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2433 _old_marking_cycles_started, _old_marking_cycles_completed)); 2434 2435 _old_marking_cycles_started++; 2436 } 2437 2438 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2439 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2440 2441 // We assume that if concurrent == true, then the caller is a 2442 // concurrent thread that was joined the Suspendible Thread 2443 // Set. If there's ever a cheap way to check this, we should add an 2444 // assert here. 2445 2446 // Given that this method is called at the end of a Full GC or of a 2447 // concurrent cycle, and those can be nested (i.e., a Full GC can 2448 // interrupt a concurrent cycle), the number of full collections 2449 // completed should be either one (in the case where there was no 2450 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2451 // behind the number of full collections started. 2452 2453 // This is the case for the inner caller, i.e. a Full GC. 2454 assert(concurrent || 2455 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2456 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2457 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2458 "is inconsistent with _old_marking_cycles_completed = %u", 2459 _old_marking_cycles_started, _old_marking_cycles_completed)); 2460 2461 // This is the case for the outer caller, i.e. the concurrent cycle. 2462 assert(!concurrent || 2463 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2464 err_msg("for outer caller (concurrent cycle): " 2465 "_old_marking_cycles_started = %u " 2466 "is inconsistent with _old_marking_cycles_completed = %u", 2467 _old_marking_cycles_started, _old_marking_cycles_completed)); 2468 2469 _old_marking_cycles_completed += 1; 2470 2471 // We need to clear the "in_progress" flag in the CM thread before 2472 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2473 // is set) so that if a waiter requests another System.gc() it doesn't 2474 // incorrectly see that a marking cycle is still in progress. 2475 if (concurrent) { 2476 _cmThread->clear_in_progress(); 2477 } 2478 2479 // This notify_all() will ensure that a thread that called 2480 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2481 // and it's waiting for a full GC to finish will be woken up. It is 2482 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2483 FullGCCount_lock->notify_all(); 2484 } 2485 2486 void G1CollectedHeap::register_concurrent_cycle_start(jlong start_time) { 2487 _concurrent_cycle_started = true; 2488 _gc_timer_cm->register_gc_start(start_time); 2489 2490 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start()); 2491 trace_heap_before_gc(_gc_tracer_cm); 2492 } 2493 2494 void G1CollectedHeap::register_concurrent_cycle_end() { 2495 if (_concurrent_cycle_started) { 2496 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2497 2498 if (_cm->has_aborted()) { 2499 _gc_tracer_cm->report_concurrent_mode_failure(); 2500 } 2501 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2502 2503 _concurrent_cycle_started = false; 2504 } 2505 } 2506 2507 void G1CollectedHeap::trace_heap_after_concurrent_cycle() { 2508 if (_concurrent_cycle_started) { 2509 trace_heap_after_gc(_gc_tracer_cm); 2510 } 2511 } 2512 2513 G1YCType G1CollectedHeap::yc_type() { 2514 bool is_young = g1_policy()->gcs_are_young(); 2515 bool is_initial_mark = g1_policy()->during_initial_mark_pause(); 2516 bool is_during_mark = mark_in_progress(); 2517 2518 if (is_initial_mark) { 2519 return InitialMark; 2520 } else if (is_during_mark) { 2521 return DuringMark; 2522 } else if (is_young) { 2523 return Normal; 2524 } else { 2525 return Mixed; 2526 } 2527 } 2528 2529 void G1CollectedHeap::collect(GCCause::Cause cause) { 2530 assert_heap_not_locked(); 2531 2532 unsigned int gc_count_before; 2533 unsigned int old_marking_count_before; 2534 bool retry_gc; 2535 2536 do { 2537 retry_gc = false; 2538 2539 { 2540 MutexLocker ml(Heap_lock); 2541 2542 // Read the GC count while holding the Heap_lock 2543 gc_count_before = total_collections(); 2544 old_marking_count_before = _old_marking_cycles_started; 2545 } 2546 2547 if (should_do_concurrent_full_gc(cause)) { 2548 // Schedule an initial-mark evacuation pause that will start a 2549 // concurrent cycle. We're setting word_size to 0 which means that 2550 // we are not requesting a post-GC allocation. 2551 VM_G1IncCollectionPause op(gc_count_before, 2552 0, /* word_size */ 2553 true, /* should_initiate_conc_mark */ 2554 g1_policy()->max_pause_time_ms(), 2555 cause); 2556 2557 VMThread::execute(&op); 2558 if (!op.pause_succeeded()) { 2559 if (old_marking_count_before == _old_marking_cycles_started) { 2560 retry_gc = op.should_retry_gc(); 2561 } else { 2562 // A Full GC happened while we were trying to schedule the 2563 // initial-mark GC. No point in starting a new cycle given 2564 // that the whole heap was collected anyway. 2565 } 2566 2567 if (retry_gc) { 2568 if (GC_locker::is_active_and_needs_gc()) { 2569 GC_locker::stall_until_clear(); 2570 } 2571 } 2572 } 2573 } else { 2574 if (cause == GCCause::_gc_locker 2575 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2576 2577 // Schedule a standard evacuation pause. We're setting word_size 2578 // to 0 which means that we are not requesting a post-GC allocation. 2579 VM_G1IncCollectionPause op(gc_count_before, 2580 0, /* word_size */ 2581 false, /* should_initiate_conc_mark */ 2582 g1_policy()->max_pause_time_ms(), 2583 cause); 2584 VMThread::execute(&op); 2585 } else { 2586 // Schedule a Full GC. 2587 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2588 VMThread::execute(&op); 2589 } 2590 } 2591 } while (retry_gc); 2592 } 2593 2594 bool G1CollectedHeap::is_in(const void* p) const { 2595 if (_g1_committed.contains(p)) { 2596 // Given that we know that p is in the committed space, 2597 // heap_region_containing_raw() should successfully 2598 // return the containing region. 2599 HeapRegion* hr = heap_region_containing_raw(p); 2600 return hr->is_in(p); 2601 } else { 2602 return false; 2603 } 2604 } 2605 2606 // Iteration functions. 2607 2608 // Iterates an OopClosure over all ref-containing fields of objects 2609 // within a HeapRegion. 2610 2611 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2612 MemRegion _mr; 2613 ExtendedOopClosure* _cl; 2614 public: 2615 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2616 : _mr(mr), _cl(cl) {} 2617 bool doHeapRegion(HeapRegion* r) { 2618 if (!r->continuesHumongous()) { 2619 r->oop_iterate(_cl); 2620 } 2621 return false; 2622 } 2623 }; 2624 2625 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2626 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2627 heap_region_iterate(&blk); 2628 } 2629 2630 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2631 IterateOopClosureRegionClosure blk(mr, cl); 2632 heap_region_iterate(&blk); 2633 } 2634 2635 // Iterates an ObjectClosure over all objects within a HeapRegion. 2636 2637 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2638 ObjectClosure* _cl; 2639 public: 2640 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2641 bool doHeapRegion(HeapRegion* r) { 2642 if (! r->continuesHumongous()) { 2643 r->object_iterate(_cl); 2644 } 2645 return false; 2646 } 2647 }; 2648 2649 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2650 IterateObjectClosureRegionClosure blk(cl); 2651 heap_region_iterate(&blk); 2652 } 2653 2654 // Calls a SpaceClosure on a HeapRegion. 2655 2656 class SpaceClosureRegionClosure: public HeapRegionClosure { 2657 SpaceClosure* _cl; 2658 public: 2659 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2660 bool doHeapRegion(HeapRegion* r) { 2661 _cl->do_space(r); 2662 return false; 2663 } 2664 }; 2665 2666 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2667 SpaceClosureRegionClosure blk(cl); 2668 heap_region_iterate(&blk); 2669 } 2670 2671 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2672 _hrs.iterate(cl); 2673 } 2674 2675 void 2676 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2677 uint worker_id, 2678 uint no_of_par_workers, 2679 jint claim_value) { 2680 const uint regions = n_regions(); 2681 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2682 no_of_par_workers : 2683 1); 2684 assert(UseDynamicNumberOfGCThreads || 2685 no_of_par_workers == workers()->total_workers(), 2686 "Non dynamic should use fixed number of workers"); 2687 // try to spread out the starting points of the workers 2688 const HeapRegion* start_hr = 2689 start_region_for_worker(worker_id, no_of_par_workers); 2690 const uint start_index = start_hr->hrs_index(); 2691 2692 // each worker will actually look at all regions 2693 for (uint count = 0; count < regions; ++count) { 2694 const uint index = (start_index + count) % regions; 2695 assert(0 <= index && index < regions, "sanity"); 2696 HeapRegion* r = region_at(index); 2697 // we'll ignore "continues humongous" regions (we'll process them 2698 // when we come across their corresponding "start humongous" 2699 // region) and regions already claimed 2700 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2701 continue; 2702 } 2703 // OK, try to claim it 2704 if (r->claimHeapRegion(claim_value)) { 2705 // success! 2706 assert(!r->continuesHumongous(), "sanity"); 2707 if (r->startsHumongous()) { 2708 // If the region is "starts humongous" we'll iterate over its 2709 // "continues humongous" first; in fact we'll do them 2710 // first. The order is important. In on case, calling the 2711 // closure on the "starts humongous" region might de-allocate 2712 // and clear all its "continues humongous" regions and, as a 2713 // result, we might end up processing them twice. So, we'll do 2714 // them first (notice: most closures will ignore them anyway) and 2715 // then we'll do the "starts humongous" region. 2716 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2717 HeapRegion* chr = region_at(ch_index); 2718 2719 // if the region has already been claimed or it's not 2720 // "continues humongous" we're done 2721 if (chr->claim_value() == claim_value || 2722 !chr->continuesHumongous()) { 2723 break; 2724 } 2725 2726 // No one should have claimed it directly. We can given 2727 // that we claimed its "starts humongous" region. 2728 assert(chr->claim_value() != claim_value, "sanity"); 2729 assert(chr->humongous_start_region() == r, "sanity"); 2730 2731 if (chr->claimHeapRegion(claim_value)) { 2732 // we should always be able to claim it; no one else should 2733 // be trying to claim this region 2734 2735 bool res2 = cl->doHeapRegion(chr); 2736 assert(!res2, "Should not abort"); 2737 2738 // Right now, this holds (i.e., no closure that actually 2739 // does something with "continues humongous" regions 2740 // clears them). We might have to weaken it in the future, 2741 // but let's leave these two asserts here for extra safety. 2742 assert(chr->continuesHumongous(), "should still be the case"); 2743 assert(chr->humongous_start_region() == r, "sanity"); 2744 } else { 2745 guarantee(false, "we should not reach here"); 2746 } 2747 } 2748 } 2749 2750 assert(!r->continuesHumongous(), "sanity"); 2751 bool res = cl->doHeapRegion(r); 2752 assert(!res, "Should not abort"); 2753 } 2754 } 2755 } 2756 2757 class ResetClaimValuesClosure: public HeapRegionClosure { 2758 public: 2759 bool doHeapRegion(HeapRegion* r) { 2760 r->set_claim_value(HeapRegion::InitialClaimValue); 2761 return false; 2762 } 2763 }; 2764 2765 void G1CollectedHeap::reset_heap_region_claim_values() { 2766 ResetClaimValuesClosure blk; 2767 heap_region_iterate(&blk); 2768 } 2769 2770 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2771 ResetClaimValuesClosure blk; 2772 collection_set_iterate(&blk); 2773 } 2774 2775 #ifdef ASSERT 2776 // This checks whether all regions in the heap have the correct claim 2777 // value. I also piggy-backed on this a check to ensure that the 2778 // humongous_start_region() information on "continues humongous" 2779 // regions is correct. 2780 2781 class CheckClaimValuesClosure : public HeapRegionClosure { 2782 private: 2783 jint _claim_value; 2784 uint _failures; 2785 HeapRegion* _sh_region; 2786 2787 public: 2788 CheckClaimValuesClosure(jint claim_value) : 2789 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2790 bool doHeapRegion(HeapRegion* r) { 2791 if (r->claim_value() != _claim_value) { 2792 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2793 "claim value = %d, should be %d", 2794 HR_FORMAT_PARAMS(r), 2795 r->claim_value(), _claim_value); 2796 ++_failures; 2797 } 2798 if (!r->isHumongous()) { 2799 _sh_region = NULL; 2800 } else if (r->startsHumongous()) { 2801 _sh_region = r; 2802 } else if (r->continuesHumongous()) { 2803 if (r->humongous_start_region() != _sh_region) { 2804 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2805 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2806 HR_FORMAT_PARAMS(r), 2807 r->humongous_start_region(), 2808 _sh_region); 2809 ++_failures; 2810 } 2811 } 2812 return false; 2813 } 2814 uint failures() { return _failures; } 2815 }; 2816 2817 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2818 CheckClaimValuesClosure cl(claim_value); 2819 heap_region_iterate(&cl); 2820 return cl.failures() == 0; 2821 } 2822 2823 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2824 private: 2825 jint _claim_value; 2826 uint _failures; 2827 2828 public: 2829 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2830 _claim_value(claim_value), _failures(0) { } 2831 2832 uint failures() { return _failures; } 2833 2834 bool doHeapRegion(HeapRegion* hr) { 2835 assert(hr->in_collection_set(), "how?"); 2836 assert(!hr->isHumongous(), "H-region in CSet"); 2837 if (hr->claim_value() != _claim_value) { 2838 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2839 "claim value = %d, should be %d", 2840 HR_FORMAT_PARAMS(hr), 2841 hr->claim_value(), _claim_value); 2842 _failures += 1; 2843 } 2844 return false; 2845 } 2846 }; 2847 2848 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2849 CheckClaimValuesInCSetHRClosure cl(claim_value); 2850 collection_set_iterate(&cl); 2851 return cl.failures() == 0; 2852 } 2853 #endif // ASSERT 2854 2855 // Clear the cached CSet starting regions and (more importantly) 2856 // the time stamps. Called when we reset the GC time stamp. 2857 void G1CollectedHeap::clear_cset_start_regions() { 2858 assert(_worker_cset_start_region != NULL, "sanity"); 2859 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2860 2861 int n_queues = MAX2((int)ParallelGCThreads, 1); 2862 for (int i = 0; i < n_queues; i++) { 2863 _worker_cset_start_region[i] = NULL; 2864 _worker_cset_start_region_time_stamp[i] = 0; 2865 } 2866 } 2867 2868 // Given the id of a worker, obtain or calculate a suitable 2869 // starting region for iterating over the current collection set. 2870 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2871 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2872 2873 HeapRegion* result = NULL; 2874 unsigned gc_time_stamp = get_gc_time_stamp(); 2875 2876 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2877 // Cached starting region for current worker was set 2878 // during the current pause - so it's valid. 2879 // Note: the cached starting heap region may be NULL 2880 // (when the collection set is empty). 2881 result = _worker_cset_start_region[worker_i]; 2882 assert(result == NULL || result->in_collection_set(), "sanity"); 2883 return result; 2884 } 2885 2886 // The cached entry was not valid so let's calculate 2887 // a suitable starting heap region for this worker. 2888 2889 // We want the parallel threads to start their collection 2890 // set iteration at different collection set regions to 2891 // avoid contention. 2892 // If we have: 2893 // n collection set regions 2894 // p threads 2895 // Then thread t will start at region floor ((t * n) / p) 2896 2897 result = g1_policy()->collection_set(); 2898 if (G1CollectedHeap::use_parallel_gc_threads()) { 2899 uint cs_size = g1_policy()->cset_region_length(); 2900 uint active_workers = workers()->active_workers(); 2901 assert(UseDynamicNumberOfGCThreads || 2902 active_workers == workers()->total_workers(), 2903 "Unless dynamic should use total workers"); 2904 2905 uint end_ind = (cs_size * worker_i) / active_workers; 2906 uint start_ind = 0; 2907 2908 if (worker_i > 0 && 2909 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2910 // Previous workers starting region is valid 2911 // so let's iterate from there 2912 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2913 result = _worker_cset_start_region[worker_i - 1]; 2914 } 2915 2916 for (uint i = start_ind; i < end_ind; i++) { 2917 result = result->next_in_collection_set(); 2918 } 2919 } 2920 2921 // Note: the calculated starting heap region may be NULL 2922 // (when the collection set is empty). 2923 assert(result == NULL || result->in_collection_set(), "sanity"); 2924 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2925 "should be updated only once per pause"); 2926 _worker_cset_start_region[worker_i] = result; 2927 OrderAccess::storestore(); 2928 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2929 return result; 2930 } 2931 2932 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2933 uint no_of_par_workers) { 2934 uint worker_num = 2935 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2936 assert(UseDynamicNumberOfGCThreads || 2937 no_of_par_workers == workers()->total_workers(), 2938 "Non dynamic should use fixed number of workers"); 2939 const uint start_index = n_regions() * worker_i / worker_num; 2940 return region_at(start_index); 2941 } 2942 2943 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2944 HeapRegion* r = g1_policy()->collection_set(); 2945 while (r != NULL) { 2946 HeapRegion* next = r->next_in_collection_set(); 2947 if (cl->doHeapRegion(r)) { 2948 cl->incomplete(); 2949 return; 2950 } 2951 r = next; 2952 } 2953 } 2954 2955 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2956 HeapRegionClosure *cl) { 2957 if (r == NULL) { 2958 // The CSet is empty so there's nothing to do. 2959 return; 2960 } 2961 2962 assert(r->in_collection_set(), 2963 "Start region must be a member of the collection set."); 2964 HeapRegion* cur = r; 2965 while (cur != NULL) { 2966 HeapRegion* next = cur->next_in_collection_set(); 2967 if (cl->doHeapRegion(cur) && false) { 2968 cl->incomplete(); 2969 return; 2970 } 2971 cur = next; 2972 } 2973 cur = g1_policy()->collection_set(); 2974 while (cur != r) { 2975 HeapRegion* next = cur->next_in_collection_set(); 2976 if (cl->doHeapRegion(cur) && false) { 2977 cl->incomplete(); 2978 return; 2979 } 2980 cur = next; 2981 } 2982 } 2983 2984 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2985 return n_regions() > 0 ? region_at(0) : NULL; 2986 } 2987 2988 2989 Space* G1CollectedHeap::space_containing(const void* addr) const { 2990 Space* res = heap_region_containing(addr); 2991 return res; 2992 } 2993 2994 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2995 Space* sp = space_containing(addr); 2996 if (sp != NULL) { 2997 return sp->block_start(addr); 2998 } 2999 return NULL; 3000 } 3001 3002 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 3003 Space* sp = space_containing(addr); 3004 assert(sp != NULL, "block_size of address outside of heap"); 3005 return sp->block_size(addr); 3006 } 3007 3008 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 3009 Space* sp = space_containing(addr); 3010 return sp->block_is_obj(addr); 3011 } 3012 3013 bool G1CollectedHeap::supports_tlab_allocation() const { 3014 return true; 3015 } 3016 3017 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 3018 return HeapRegion::GrainBytes; 3019 } 3020 3021 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 3022 // Return the remaining space in the cur alloc region, but not less than 3023 // the min TLAB size. 3024 3025 // Also, this value can be at most the humongous object threshold, 3026 // since we can't allow tlabs to grow big enough to accommodate 3027 // humongous objects. 3028 3029 HeapRegion* hr = _mutator_alloc_region.get(); 3030 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 3031 if (hr == NULL) { 3032 return max_tlab_size; 3033 } else { 3034 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 3035 } 3036 } 3037 3038 size_t G1CollectedHeap::max_capacity() const { 3039 return _g1_reserved.byte_size(); 3040 } 3041 3042 jlong G1CollectedHeap::millis_since_last_gc() { 3043 // assert(false, "NYI"); 3044 return 0; 3045 } 3046 3047 void G1CollectedHeap::prepare_for_verify() { 3048 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3049 ensure_parsability(false); 3050 } 3051 g1_rem_set()->prepare_for_verify(); 3052 } 3053 3054 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3055 VerifyOption vo) { 3056 switch (vo) { 3057 case VerifyOption_G1UsePrevMarking: 3058 return hr->obj_allocated_since_prev_marking(obj); 3059 case VerifyOption_G1UseNextMarking: 3060 return hr->obj_allocated_since_next_marking(obj); 3061 case VerifyOption_G1UseMarkWord: 3062 return false; 3063 default: 3064 ShouldNotReachHere(); 3065 } 3066 return false; // keep some compilers happy 3067 } 3068 3069 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3070 switch (vo) { 3071 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3072 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3073 case VerifyOption_G1UseMarkWord: return NULL; 3074 default: ShouldNotReachHere(); 3075 } 3076 return NULL; // keep some compilers happy 3077 } 3078 3079 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3080 switch (vo) { 3081 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3082 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3083 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3084 default: ShouldNotReachHere(); 3085 } 3086 return false; // keep some compilers happy 3087 } 3088 3089 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3090 switch (vo) { 3091 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3092 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3093 case VerifyOption_G1UseMarkWord: return "NONE"; 3094 default: ShouldNotReachHere(); 3095 } 3096 return NULL; // keep some compilers happy 3097 } 3098 3099 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can 3100 // pass it as the perm_blk to SharedHeap::process_strong_roots. 3101 // When process_strong_roots stop calling perm_blk->younger_refs_iterate 3102 // we can change this closure to extend the simpler OopClosure. 3103 class VerifyRootsClosure: public OopsInGenClosure { 3104 private: 3105 G1CollectedHeap* _g1h; 3106 VerifyOption _vo; 3107 bool _failures; 3108 public: 3109 // _vo == UsePrevMarking -> use "prev" marking information, 3110 // _vo == UseNextMarking -> use "next" marking information, 3111 // _vo == UseMarkWord -> use mark word from object header. 3112 VerifyRootsClosure(VerifyOption vo) : 3113 _g1h(G1CollectedHeap::heap()), 3114 _vo(vo), 3115 _failures(false) { } 3116 3117 bool failures() { return _failures; } 3118 3119 template <class T> void do_oop_nv(T* p) { 3120 T heap_oop = oopDesc::load_heap_oop(p); 3121 if (!oopDesc::is_null(heap_oop)) { 3122 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3123 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3124 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3125 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3126 if (_vo == VerifyOption_G1UseMarkWord) { 3127 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3128 } 3129 obj->print_on(gclog_or_tty); 3130 _failures = true; 3131 } 3132 } 3133 } 3134 3135 void do_oop(oop* p) { do_oop_nv(p); } 3136 void do_oop(narrowOop* p) { do_oop_nv(p); } 3137 }; 3138 3139 class G1VerifyCodeRootOopClosure: public OopsInGenClosure { 3140 G1CollectedHeap* _g1h; 3141 OopClosure* _root_cl; 3142 nmethod* _nm; 3143 VerifyOption _vo; 3144 bool _failures; 3145 3146 template <class T> void do_oop_work(T* p) { 3147 // First verify that this root is live 3148 _root_cl->do_oop(p); 3149 3150 if (!G1VerifyHeapRegionCodeRoots) { 3151 // We're not verifying the code roots attached to heap region. 3152 return; 3153 } 3154 3155 // Don't check the code roots during marking verification in a full GC 3156 if (_vo == VerifyOption_G1UseMarkWord) { 3157 return; 3158 } 3159 3160 // Now verify that the current nmethod (which contains p) is 3161 // in the code root list of the heap region containing the 3162 // object referenced by p. 3163 3164 T heap_oop = oopDesc::load_heap_oop(p); 3165 if (!oopDesc::is_null(heap_oop)) { 3166 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3167 3168 // Now fetch the region containing the object 3169 HeapRegion* hr = _g1h->heap_region_containing(obj); 3170 HeapRegionRemSet* hrrs = hr->rem_set(); 3171 // Verify that the strong code root list for this region 3172 // contains the nmethod 3173 if (!hrrs->strong_code_roots_list_contains(_nm)) { 3174 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" " 3175 "from nmethod "PTR_FORMAT" not in strong " 3176 "code roots for region ["PTR_FORMAT","PTR_FORMAT")", 3177 p, _nm, hr->bottom(), hr->end()); 3178 _failures = true; 3179 } 3180 } 3181 } 3182 3183 public: 3184 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo): 3185 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {} 3186 3187 void do_oop(oop* p) { do_oop_work(p); } 3188 void do_oop(narrowOop* p) { do_oop_work(p); } 3189 3190 void set_nmethod(nmethod* nm) { _nm = nm; } 3191 bool failures() { return _failures; } 3192 }; 3193 3194 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure { 3195 G1VerifyCodeRootOopClosure* _oop_cl; 3196 3197 public: 3198 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl): 3199 _oop_cl(oop_cl) {} 3200 3201 void do_code_blob(CodeBlob* cb) { 3202 nmethod* nm = cb->as_nmethod_or_null(); 3203 if (nm != NULL) { 3204 _oop_cl->set_nmethod(nm); 3205 nm->oops_do(_oop_cl); 3206 } 3207 } 3208 }; 3209 3210 class YoungRefCounterClosure : public OopClosure { 3211 G1CollectedHeap* _g1h; 3212 int _count; 3213 public: 3214 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3215 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3216 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3217 3218 int count() { return _count; } 3219 void reset_count() { _count = 0; }; 3220 }; 3221 3222 class VerifyKlassClosure: public KlassClosure { 3223 YoungRefCounterClosure _young_ref_counter_closure; 3224 OopClosure *_oop_closure; 3225 public: 3226 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3227 void do_klass(Klass* k) { 3228 k->oops_do(_oop_closure); 3229 3230 _young_ref_counter_closure.reset_count(); 3231 k->oops_do(&_young_ref_counter_closure); 3232 if (_young_ref_counter_closure.count() > 0) { 3233 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3234 } 3235 } 3236 }; 3237 3238 class VerifyLivenessOopClosure: public OopClosure { 3239 G1CollectedHeap* _g1h; 3240 VerifyOption _vo; 3241 public: 3242 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3243 _g1h(g1h), _vo(vo) 3244 { } 3245 void do_oop(narrowOop *p) { do_oop_work(p); } 3246 void do_oop( oop *p) { do_oop_work(p); } 3247 3248 template <class T> void do_oop_work(T *p) { 3249 oop obj = oopDesc::load_decode_heap_oop(p); 3250 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3251 "Dead object referenced by a not dead object"); 3252 } 3253 }; 3254 3255 class VerifyObjsInRegionClosure: public ObjectClosure { 3256 private: 3257 G1CollectedHeap* _g1h; 3258 size_t _live_bytes; 3259 HeapRegion *_hr; 3260 VerifyOption _vo; 3261 public: 3262 // _vo == UsePrevMarking -> use "prev" marking information, 3263 // _vo == UseNextMarking -> use "next" marking information, 3264 // _vo == UseMarkWord -> use mark word from object header. 3265 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3266 : _live_bytes(0), _hr(hr), _vo(vo) { 3267 _g1h = G1CollectedHeap::heap(); 3268 } 3269 void do_object(oop o) { 3270 VerifyLivenessOopClosure isLive(_g1h, _vo); 3271 assert(o != NULL, "Huh?"); 3272 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3273 // If the object is alive according to the mark word, 3274 // then verify that the marking information agrees. 3275 // Note we can't verify the contra-positive of the 3276 // above: if the object is dead (according to the mark 3277 // word), it may not be marked, or may have been marked 3278 // but has since became dead, or may have been allocated 3279 // since the last marking. 3280 if (_vo == VerifyOption_G1UseMarkWord) { 3281 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3282 } 3283 3284 o->oop_iterate_no_header(&isLive); 3285 if (!_hr->obj_allocated_since_prev_marking(o)) { 3286 size_t obj_size = o->size(); // Make sure we don't overflow 3287 _live_bytes += (obj_size * HeapWordSize); 3288 } 3289 } 3290 } 3291 size_t live_bytes() { return _live_bytes; } 3292 }; 3293 3294 class PrintObjsInRegionClosure : public ObjectClosure { 3295 HeapRegion *_hr; 3296 G1CollectedHeap *_g1; 3297 public: 3298 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3299 _g1 = G1CollectedHeap::heap(); 3300 }; 3301 3302 void do_object(oop o) { 3303 if (o != NULL) { 3304 HeapWord *start = (HeapWord *) o; 3305 size_t word_sz = o->size(); 3306 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3307 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3308 (void*) o, word_sz, 3309 _g1->isMarkedPrev(o), 3310 _g1->isMarkedNext(o), 3311 _hr->obj_allocated_since_prev_marking(o)); 3312 HeapWord *end = start + word_sz; 3313 HeapWord *cur; 3314 int *val; 3315 for (cur = start; cur < end; cur++) { 3316 val = (int *) cur; 3317 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3318 } 3319 } 3320 } 3321 }; 3322 3323 class VerifyRegionClosure: public HeapRegionClosure { 3324 private: 3325 bool _par; 3326 VerifyOption _vo; 3327 bool _failures; 3328 public: 3329 // _vo == UsePrevMarking -> use "prev" marking information, 3330 // _vo == UseNextMarking -> use "next" marking information, 3331 // _vo == UseMarkWord -> use mark word from object header. 3332 VerifyRegionClosure(bool par, VerifyOption vo) 3333 : _par(par), 3334 _vo(vo), 3335 _failures(false) {} 3336 3337 bool failures() { 3338 return _failures; 3339 } 3340 3341 bool doHeapRegion(HeapRegion* r) { 3342 if (!r->continuesHumongous()) { 3343 bool failures = false; 3344 r->verify(_vo, &failures); 3345 if (failures) { 3346 _failures = true; 3347 } else { 3348 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3349 r->object_iterate(¬_dead_yet_cl); 3350 if (_vo != VerifyOption_G1UseNextMarking) { 3351 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3352 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3353 "max_live_bytes "SIZE_FORMAT" " 3354 "< calculated "SIZE_FORMAT, 3355 r->bottom(), r->end(), 3356 r->max_live_bytes(), 3357 not_dead_yet_cl.live_bytes()); 3358 _failures = true; 3359 } 3360 } else { 3361 // When vo == UseNextMarking we cannot currently do a sanity 3362 // check on the live bytes as the calculation has not been 3363 // finalized yet. 3364 } 3365 } 3366 } 3367 return false; // stop the region iteration if we hit a failure 3368 } 3369 }; 3370 3371 // This is the task used for parallel verification of the heap regions 3372 3373 class G1ParVerifyTask: public AbstractGangTask { 3374 private: 3375 G1CollectedHeap* _g1h; 3376 VerifyOption _vo; 3377 bool _failures; 3378 3379 public: 3380 // _vo == UsePrevMarking -> use "prev" marking information, 3381 // _vo == UseNextMarking -> use "next" marking information, 3382 // _vo == UseMarkWord -> use mark word from object header. 3383 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3384 AbstractGangTask("Parallel verify task"), 3385 _g1h(g1h), 3386 _vo(vo), 3387 _failures(false) { } 3388 3389 bool failures() { 3390 return _failures; 3391 } 3392 3393 void work(uint worker_id) { 3394 HandleMark hm; 3395 VerifyRegionClosure blk(true, _vo); 3396 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3397 _g1h->workers()->active_workers(), 3398 HeapRegion::ParVerifyClaimValue); 3399 if (blk.failures()) { 3400 _failures = true; 3401 } 3402 } 3403 }; 3404 3405 void G1CollectedHeap::verify(bool silent, VerifyOption vo) { 3406 if (SafepointSynchronize::is_at_safepoint()) { 3407 assert(Thread::current()->is_VM_thread(), 3408 "Expected to be executed serially by the VM thread at this point"); 3409 3410 if (!silent) { gclog_or_tty->print("Roots "); } 3411 VerifyRootsClosure rootsCl(vo); 3412 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo); 3413 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl); 3414 VerifyKlassClosure klassCl(this, &rootsCl); 3415 3416 // We apply the relevant closures to all the oops in the 3417 // system dictionary, the string table and the code cache. 3418 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3419 3420 // Need cleared claim bits for the strong roots processing 3421 ClassLoaderDataGraph::clear_claimed_marks(); 3422 3423 process_strong_roots(true, // activate StrongRootsScope 3424 false, // we set "is scavenging" to false, 3425 // so we don't reset the dirty cards. 3426 ScanningOption(so), // roots scanning options 3427 &rootsCl, 3428 &blobsCl, 3429 &klassCl 3430 ); 3431 3432 bool failures = rootsCl.failures() || codeRootsCl.failures(); 3433 3434 if (vo != VerifyOption_G1UseMarkWord) { 3435 // If we're verifying during a full GC then the region sets 3436 // will have been torn down at the start of the GC. Therefore 3437 // verifying the region sets will fail. So we only verify 3438 // the region sets when not in a full GC. 3439 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3440 verify_region_sets(); 3441 } 3442 3443 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3444 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3445 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3446 "sanity check"); 3447 3448 G1ParVerifyTask task(this, vo); 3449 assert(UseDynamicNumberOfGCThreads || 3450 workers()->active_workers() == workers()->total_workers(), 3451 "If not dynamic should be using all the workers"); 3452 int n_workers = workers()->active_workers(); 3453 set_par_threads(n_workers); 3454 workers()->run_task(&task); 3455 set_par_threads(0); 3456 if (task.failures()) { 3457 failures = true; 3458 } 3459 3460 // Checks that the expected amount of parallel work was done. 3461 // The implication is that n_workers is > 0. 3462 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3463 "sanity check"); 3464 3465 reset_heap_region_claim_values(); 3466 3467 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3468 "sanity check"); 3469 } else { 3470 VerifyRegionClosure blk(false, vo); 3471 heap_region_iterate(&blk); 3472 if (blk.failures()) { 3473 failures = true; 3474 } 3475 } 3476 if (!silent) gclog_or_tty->print("RemSet "); 3477 rem_set()->verify(); 3478 3479 if (failures) { 3480 gclog_or_tty->print_cr("Heap:"); 3481 // It helps to have the per-region information in the output to 3482 // help us track down what went wrong. This is why we call 3483 // print_extended_on() instead of print_on(). 3484 print_extended_on(gclog_or_tty); 3485 gclog_or_tty->print_cr(""); 3486 #ifndef PRODUCT 3487 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3488 concurrent_mark()->print_reachable("at-verification-failure", 3489 vo, false /* all */); 3490 } 3491 #endif 3492 gclog_or_tty->flush(); 3493 } 3494 guarantee(!failures, "there should not have been any failures"); 3495 } else { 3496 if (!silent) 3497 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) "); 3498 } 3499 } 3500 3501 void G1CollectedHeap::verify(bool silent) { 3502 verify(silent, VerifyOption_G1UsePrevMarking); 3503 } 3504 3505 double G1CollectedHeap::verify(bool guard, const char* msg) { 3506 double verify_time_ms = 0.0; 3507 3508 if (guard && total_collections() >= VerifyGCStartAt) { 3509 double verify_start = os::elapsedTime(); 3510 HandleMark hm; // Discard invalid handles created during verification 3511 prepare_for_verify(); 3512 Universe::verify(VerifyOption_G1UsePrevMarking, msg); 3513 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 3514 } 3515 3516 return verify_time_ms; 3517 } 3518 3519 void G1CollectedHeap::verify_before_gc() { 3520 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 3521 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 3522 } 3523 3524 void G1CollectedHeap::verify_after_gc() { 3525 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 3526 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 3527 } 3528 3529 class PrintRegionClosure: public HeapRegionClosure { 3530 outputStream* _st; 3531 public: 3532 PrintRegionClosure(outputStream* st) : _st(st) {} 3533 bool doHeapRegion(HeapRegion* r) { 3534 r->print_on(_st); 3535 return false; 3536 } 3537 }; 3538 3539 void G1CollectedHeap::print_on(outputStream* st) const { 3540 st->print(" %-20s", "garbage-first heap"); 3541 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3542 capacity()/K, used_unlocked()/K); 3543 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3544 _g1_storage.low_boundary(), 3545 _g1_storage.high(), 3546 _g1_storage.high_boundary()); 3547 st->cr(); 3548 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3549 uint young_regions = _young_list->length(); 3550 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3551 (size_t) young_regions * HeapRegion::GrainBytes / K); 3552 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3553 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3554 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3555 st->cr(); 3556 MetaspaceAux::print_on(st); 3557 } 3558 3559 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3560 print_on(st); 3561 3562 // Print the per-region information. 3563 st->cr(); 3564 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3565 "HS=humongous(starts), HC=humongous(continues), " 3566 "CS=collection set, F=free, TS=gc time stamp, " 3567 "PTAMS=previous top-at-mark-start, " 3568 "NTAMS=next top-at-mark-start)"); 3569 PrintRegionClosure blk(st); 3570 heap_region_iterate(&blk); 3571 } 3572 3573 void G1CollectedHeap::print_on_error(outputStream* st) const { 3574 this->CollectedHeap::print_on_error(st); 3575 3576 if (_cm != NULL) { 3577 st->cr(); 3578 _cm->print_on_error(st); 3579 } 3580 } 3581 3582 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3583 if (G1CollectedHeap::use_parallel_gc_threads()) { 3584 workers()->print_worker_threads_on(st); 3585 } 3586 _cmThread->print_on(st); 3587 st->cr(); 3588 _cm->print_worker_threads_on(st); 3589 _cg1r->print_worker_threads_on(st); 3590 } 3591 3592 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3593 if (G1CollectedHeap::use_parallel_gc_threads()) { 3594 workers()->threads_do(tc); 3595 } 3596 tc->do_thread(_cmThread); 3597 _cg1r->threads_do(tc); 3598 } 3599 3600 void G1CollectedHeap::print_tracing_info() const { 3601 // We'll overload this to mean "trace GC pause statistics." 3602 if (TraceGen0Time || TraceGen1Time) { 3603 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3604 // to that. 3605 g1_policy()->print_tracing_info(); 3606 } 3607 if (G1SummarizeRSetStats) { 3608 g1_rem_set()->print_summary_info(); 3609 } 3610 if (G1SummarizeConcMark) { 3611 concurrent_mark()->print_summary_info(); 3612 } 3613 g1_policy()->print_yg_surv_rate_info(); 3614 SpecializationStats::print(); 3615 } 3616 3617 #ifndef PRODUCT 3618 // Helpful for debugging RSet issues. 3619 3620 class PrintRSetsClosure : public HeapRegionClosure { 3621 private: 3622 const char* _msg; 3623 size_t _occupied_sum; 3624 3625 public: 3626 bool doHeapRegion(HeapRegion* r) { 3627 HeapRegionRemSet* hrrs = r->rem_set(); 3628 size_t occupied = hrrs->occupied(); 3629 _occupied_sum += occupied; 3630 3631 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3632 HR_FORMAT_PARAMS(r)); 3633 if (occupied == 0) { 3634 gclog_or_tty->print_cr(" RSet is empty"); 3635 } else { 3636 hrrs->print(); 3637 } 3638 gclog_or_tty->print_cr("----------"); 3639 return false; 3640 } 3641 3642 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3643 gclog_or_tty->cr(); 3644 gclog_or_tty->print_cr("========================================"); 3645 gclog_or_tty->print_cr(msg); 3646 gclog_or_tty->cr(); 3647 } 3648 3649 ~PrintRSetsClosure() { 3650 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3651 gclog_or_tty->print_cr("========================================"); 3652 gclog_or_tty->cr(); 3653 } 3654 }; 3655 3656 void G1CollectedHeap::print_cset_rsets() { 3657 PrintRSetsClosure cl("Printing CSet RSets"); 3658 collection_set_iterate(&cl); 3659 } 3660 3661 void G1CollectedHeap::print_all_rsets() { 3662 PrintRSetsClosure cl("Printing All RSets");; 3663 heap_region_iterate(&cl); 3664 } 3665 #endif // PRODUCT 3666 3667 G1CollectedHeap* G1CollectedHeap::heap() { 3668 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3669 "not a garbage-first heap"); 3670 return _g1h; 3671 } 3672 3673 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3674 // always_do_update_barrier = false; 3675 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3676 // Fill TLAB's and such 3677 ensure_parsability(true); 3678 } 3679 3680 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3681 3682 if (G1SummarizeRSetStats && 3683 (G1SummarizeRSetStatsPeriod > 0) && 3684 // we are at the end of the GC. Total collections has already been increased. 3685 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) { 3686 g1_rem_set()->print_periodic_summary_info(); 3687 } 3688 3689 // FIXME: what is this about? 3690 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3691 // is set. 3692 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3693 "derived pointer present")); 3694 // always_do_update_barrier = true; 3695 3696 // We have just completed a GC. Update the soft reference 3697 // policy with the new heap occupancy 3698 Universe::update_heap_info_at_gc(); 3699 } 3700 3701 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3702 unsigned int gc_count_before, 3703 bool* succeeded) { 3704 assert_heap_not_locked_and_not_at_safepoint(); 3705 g1_policy()->record_stop_world_start(); 3706 VM_G1IncCollectionPause op(gc_count_before, 3707 word_size, 3708 false, /* should_initiate_conc_mark */ 3709 g1_policy()->max_pause_time_ms(), 3710 GCCause::_g1_inc_collection_pause); 3711 VMThread::execute(&op); 3712 3713 HeapWord* result = op.result(); 3714 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3715 assert(result == NULL || ret_succeeded, 3716 "the result should be NULL if the VM did not succeed"); 3717 *succeeded = ret_succeeded; 3718 3719 assert_heap_not_locked(); 3720 return result; 3721 } 3722 3723 void 3724 G1CollectedHeap::doConcurrentMark() { 3725 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3726 if (!_cmThread->in_progress()) { 3727 _cmThread->set_started(); 3728 CGC_lock->notify(); 3729 } 3730 } 3731 3732 size_t G1CollectedHeap::pending_card_num() { 3733 size_t extra_cards = 0; 3734 JavaThread *curr = Threads::first(); 3735 while (curr != NULL) { 3736 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3737 extra_cards += dcq.size(); 3738 curr = curr->next(); 3739 } 3740 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3741 size_t buffer_size = dcqs.buffer_size(); 3742 size_t buffer_num = dcqs.completed_buffers_num(); 3743 3744 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3745 // in bytes - not the number of 'entries'. We need to convert 3746 // into a number of cards. 3747 return (buffer_size * buffer_num + extra_cards) / oopSize; 3748 } 3749 3750 size_t G1CollectedHeap::cards_scanned() { 3751 return g1_rem_set()->cardsScanned(); 3752 } 3753 3754 void 3755 G1CollectedHeap::setup_surviving_young_words() { 3756 assert(_surviving_young_words == NULL, "pre-condition"); 3757 uint array_length = g1_policy()->young_cset_region_length(); 3758 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3759 if (_surviving_young_words == NULL) { 3760 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR, 3761 "Not enough space for young surv words summary."); 3762 } 3763 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3764 #ifdef ASSERT 3765 for (uint i = 0; i < array_length; ++i) { 3766 assert( _surviving_young_words[i] == 0, "memset above" ); 3767 } 3768 #endif // !ASSERT 3769 } 3770 3771 void 3772 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3773 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3774 uint array_length = g1_policy()->young_cset_region_length(); 3775 for (uint i = 0; i < array_length; ++i) { 3776 _surviving_young_words[i] += surv_young_words[i]; 3777 } 3778 } 3779 3780 void 3781 G1CollectedHeap::cleanup_surviving_young_words() { 3782 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3783 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3784 _surviving_young_words = NULL; 3785 } 3786 3787 #ifdef ASSERT 3788 class VerifyCSetClosure: public HeapRegionClosure { 3789 public: 3790 bool doHeapRegion(HeapRegion* hr) { 3791 // Here we check that the CSet region's RSet is ready for parallel 3792 // iteration. The fields that we'll verify are only manipulated 3793 // when the region is part of a CSet and is collected. Afterwards, 3794 // we reset these fields when we clear the region's RSet (when the 3795 // region is freed) so they are ready when the region is 3796 // re-allocated. The only exception to this is if there's an 3797 // evacuation failure and instead of freeing the region we leave 3798 // it in the heap. In that case, we reset these fields during 3799 // evacuation failure handling. 3800 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3801 3802 // Here's a good place to add any other checks we'd like to 3803 // perform on CSet regions. 3804 return false; 3805 } 3806 }; 3807 #endif // ASSERT 3808 3809 #if TASKQUEUE_STATS 3810 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3811 st->print_raw_cr("GC Task Stats"); 3812 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3813 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3814 } 3815 3816 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3817 print_taskqueue_stats_hdr(st); 3818 3819 TaskQueueStats totals; 3820 const int n = workers() != NULL ? workers()->total_workers() : 1; 3821 for (int i = 0; i < n; ++i) { 3822 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3823 totals += task_queue(i)->stats; 3824 } 3825 st->print_raw("tot "); totals.print(st); st->cr(); 3826 3827 DEBUG_ONLY(totals.verify()); 3828 } 3829 3830 void G1CollectedHeap::reset_taskqueue_stats() { 3831 const int n = workers() != NULL ? workers()->total_workers() : 1; 3832 for (int i = 0; i < n; ++i) { 3833 task_queue(i)->stats.reset(); 3834 } 3835 } 3836 #endif // TASKQUEUE_STATS 3837 3838 void G1CollectedHeap::log_gc_header() { 3839 if (!G1Log::fine()) { 3840 return; 3841 } 3842 3843 gclog_or_tty->date_stamp(PrintGCDateStamps); 3844 gclog_or_tty->stamp(PrintGCTimeStamps); 3845 3846 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3847 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3848 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3849 3850 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3851 } 3852 3853 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3854 if (!G1Log::fine()) { 3855 return; 3856 } 3857 3858 if (G1Log::finer()) { 3859 if (evacuation_failed()) { 3860 gclog_or_tty->print(" (to-space exhausted)"); 3861 } 3862 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3863 g1_policy()->phase_times()->note_gc_end(); 3864 g1_policy()->phase_times()->print(pause_time_sec); 3865 g1_policy()->print_detailed_heap_transition(); 3866 } else { 3867 if (evacuation_failed()) { 3868 gclog_or_tty->print("--"); 3869 } 3870 g1_policy()->print_heap_transition(); 3871 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3872 } 3873 gclog_or_tty->flush(); 3874 } 3875 3876 bool 3877 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3878 assert_at_safepoint(true /* should_be_vm_thread */); 3879 guarantee(!is_gc_active(), "collection is not reentrant"); 3880 3881 if (GC_locker::check_active_before_gc()) { 3882 return false; 3883 } 3884 3885 _gc_timer_stw->register_gc_start(os::elapsed_counter()); 3886 3887 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start()); 3888 3889 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3890 ResourceMark rm; 3891 3892 print_heap_before_gc(); 3893 trace_heap_before_gc(_gc_tracer_stw); 3894 3895 HRSPhaseSetter x(HRSPhaseEvacuation); 3896 verify_region_sets_optional(); 3897 verify_dirty_young_regions(); 3898 3899 // This call will decide whether this pause is an initial-mark 3900 // pause. If it is, during_initial_mark_pause() will return true 3901 // for the duration of this pause. 3902 g1_policy()->decide_on_conc_mark_initiation(); 3903 3904 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3905 assert(!g1_policy()->during_initial_mark_pause() || 3906 g1_policy()->gcs_are_young(), "sanity"); 3907 3908 // We also do not allow mixed GCs during marking. 3909 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3910 3911 // Record whether this pause is an initial mark. When the current 3912 // thread has completed its logging output and it's safe to signal 3913 // the CM thread, the flag's value in the policy has been reset. 3914 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3915 3916 // Inner scope for scope based logging, timers, and stats collection 3917 { 3918 EvacuationInfo evacuation_info; 3919 3920 if (g1_policy()->during_initial_mark_pause()) { 3921 // We are about to start a marking cycle, so we increment the 3922 // full collection counter. 3923 increment_old_marking_cycles_started(); 3924 register_concurrent_cycle_start(_gc_timer_stw->gc_start()); 3925 } 3926 3927 _gc_tracer_stw->report_yc_type(yc_type()); 3928 3929 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3930 3931 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3932 workers()->active_workers() : 1); 3933 double pause_start_sec = os::elapsedTime(); 3934 g1_policy()->phase_times()->note_gc_start(active_workers); 3935 log_gc_header(); 3936 3937 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3938 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3939 3940 // If the secondary_free_list is not empty, append it to the 3941 // free_list. No need to wait for the cleanup operation to finish; 3942 // the region allocation code will check the secondary_free_list 3943 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3944 // set, skip this step so that the region allocation code has to 3945 // get entries from the secondary_free_list. 3946 if (!G1StressConcRegionFreeing) { 3947 append_secondary_free_list_if_not_empty_with_lock(); 3948 } 3949 3950 assert(check_young_list_well_formed(), "young list should be well formed"); 3951 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3952 "sanity check"); 3953 3954 // Don't dynamically change the number of GC threads this early. A value of 3955 // 0 is used to indicate serial work. When parallel work is done, 3956 // it will be set. 3957 3958 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3959 IsGCActiveMark x; 3960 3961 gc_prologue(false); 3962 increment_total_collections(false /* full gc */); 3963 increment_gc_time_stamp(); 3964 3965 verify_before_gc(); 3966 3967 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3968 3969 // Please see comment in g1CollectedHeap.hpp and 3970 // G1CollectedHeap::ref_processing_init() to see how 3971 // reference processing currently works in G1. 3972 3973 // Enable discovery in the STW reference processor 3974 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3975 true /*verify_no_refs*/); 3976 3977 { 3978 // We want to temporarily turn off discovery by the 3979 // CM ref processor, if necessary, and turn it back on 3980 // on again later if we do. Using a scoped 3981 // NoRefDiscovery object will do this. 3982 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3983 3984 // Forget the current alloc region (we might even choose it to be part 3985 // of the collection set!). 3986 release_mutator_alloc_region(); 3987 3988 // We should call this after we retire the mutator alloc 3989 // region(s) so that all the ALLOC / RETIRE events are generated 3990 // before the start GC event. 3991 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3992 3993 // This timing is only used by the ergonomics to handle our pause target. 3994 // It is unclear why this should not include the full pause. We will 3995 // investigate this in CR 7178365. 3996 // 3997 // Preserving the old comment here if that helps the investigation: 3998 // 3999 // The elapsed time induced by the start time below deliberately elides 4000 // the possible verification above. 4001 double sample_start_time_sec = os::elapsedTime(); 4002 4003 #if YOUNG_LIST_VERBOSE 4004 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 4005 _young_list->print(); 4006 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4007 #endif // YOUNG_LIST_VERBOSE 4008 4009 g1_policy()->record_collection_pause_start(sample_start_time_sec); 4010 4011 double scan_wait_start = os::elapsedTime(); 4012 // We have to wait until the CM threads finish scanning the 4013 // root regions as it's the only way to ensure that all the 4014 // objects on them have been correctly scanned before we start 4015 // moving them during the GC. 4016 bool waited = _cm->root_regions()->wait_until_scan_finished(); 4017 double wait_time_ms = 0.0; 4018 if (waited) { 4019 double scan_wait_end = os::elapsedTime(); 4020 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 4021 } 4022 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 4023 4024 #if YOUNG_LIST_VERBOSE 4025 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 4026 _young_list->print(); 4027 #endif // YOUNG_LIST_VERBOSE 4028 4029 if (g1_policy()->during_initial_mark_pause()) { 4030 concurrent_mark()->checkpointRootsInitialPre(); 4031 } 4032 4033 #if YOUNG_LIST_VERBOSE 4034 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 4035 _young_list->print(); 4036 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4037 #endif // YOUNG_LIST_VERBOSE 4038 4039 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); 4040 4041 _cm->note_start_of_gc(); 4042 // We should not verify the per-thread SATB buffers given that 4043 // we have not filtered them yet (we'll do so during the 4044 // GC). We also call this after finalize_cset() to 4045 // ensure that the CSet has been finalized. 4046 _cm->verify_no_cset_oops(true /* verify_stacks */, 4047 true /* verify_enqueued_buffers */, 4048 false /* verify_thread_buffers */, 4049 true /* verify_fingers */); 4050 4051 if (_hr_printer.is_active()) { 4052 HeapRegion* hr = g1_policy()->collection_set(); 4053 while (hr != NULL) { 4054 G1HRPrinter::RegionType type; 4055 if (!hr->is_young()) { 4056 type = G1HRPrinter::Old; 4057 } else if (hr->is_survivor()) { 4058 type = G1HRPrinter::Survivor; 4059 } else { 4060 type = G1HRPrinter::Eden; 4061 } 4062 _hr_printer.cset(hr); 4063 hr = hr->next_in_collection_set(); 4064 } 4065 } 4066 4067 #ifdef ASSERT 4068 VerifyCSetClosure cl; 4069 collection_set_iterate(&cl); 4070 #endif // ASSERT 4071 4072 setup_surviving_young_words(); 4073 4074 // Initialize the GC alloc regions. 4075 init_gc_alloc_regions(evacuation_info); 4076 4077 // Actually do the work... 4078 evacuate_collection_set(evacuation_info); 4079 4080 // We do this to mainly verify the per-thread SATB buffers 4081 // (which have been filtered by now) since we didn't verify 4082 // them earlier. No point in re-checking the stacks / enqueued 4083 // buffers given that the CSet has not changed since last time 4084 // we checked. 4085 _cm->verify_no_cset_oops(false /* verify_stacks */, 4086 false /* verify_enqueued_buffers */, 4087 true /* verify_thread_buffers */, 4088 true /* verify_fingers */); 4089 4090 free_collection_set(g1_policy()->collection_set(), evacuation_info); 4091 g1_policy()->clear_collection_set(); 4092 4093 cleanup_surviving_young_words(); 4094 4095 // Start a new incremental collection set for the next pause. 4096 g1_policy()->start_incremental_cset_building(); 4097 4098 // Clear the _cset_fast_test bitmap in anticipation of adding 4099 // regions to the incremental collection set for the next 4100 // evacuation pause. 4101 clear_cset_fast_test(); 4102 4103 _young_list->reset_sampled_info(); 4104 4105 // Don't check the whole heap at this point as the 4106 // GC alloc regions from this pause have been tagged 4107 // as survivors and moved on to the survivor list. 4108 // Survivor regions will fail the !is_young() check. 4109 assert(check_young_list_empty(false /* check_heap */), 4110 "young list should be empty"); 4111 4112 #if YOUNG_LIST_VERBOSE 4113 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 4114 _young_list->print(); 4115 #endif // YOUNG_LIST_VERBOSE 4116 4117 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 4118 _young_list->first_survivor_region(), 4119 _young_list->last_survivor_region()); 4120 4121 _young_list->reset_auxilary_lists(); 4122 4123 if (evacuation_failed()) { 4124 _summary_bytes_used = recalculate_used(); 4125 uint n_queues = MAX2((int)ParallelGCThreads, 1); 4126 for (uint i = 0; i < n_queues; i++) { 4127 if (_evacuation_failed_info_array[i].has_failed()) { 4128 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]); 4129 } 4130 } 4131 } else { 4132 // The "used" of the the collection set have already been subtracted 4133 // when they were freed. Add in the bytes evacuated. 4134 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 4135 } 4136 4137 if (g1_policy()->during_initial_mark_pause()) { 4138 // We have to do this before we notify the CM threads that 4139 // they can start working to make sure that all the 4140 // appropriate initialization is done on the CM object. 4141 concurrent_mark()->checkpointRootsInitialPost(); 4142 set_marking_started(); 4143 // Note that we don't actually trigger the CM thread at 4144 // this point. We do that later when we're sure that 4145 // the current thread has completed its logging output. 4146 } 4147 4148 allocate_dummy_regions(); 4149 4150 #if YOUNG_LIST_VERBOSE 4151 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 4152 _young_list->print(); 4153 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4154 #endif // YOUNG_LIST_VERBOSE 4155 4156 init_mutator_alloc_region(); 4157 4158 { 4159 size_t expand_bytes = g1_policy()->expansion_amount(); 4160 if (expand_bytes > 0) { 4161 size_t bytes_before = capacity(); 4162 // No need for an ergo verbose message here, 4163 // expansion_amount() does this when it returns a value > 0. 4164 if (!expand(expand_bytes)) { 4165 // We failed to expand the heap so let's verify that 4166 // committed/uncommitted amount match the backing store 4167 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 4168 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 4169 } 4170 } 4171 } 4172 4173 // We redo the verification but now wrt to the new CSet which 4174 // has just got initialized after the previous CSet was freed. 4175 _cm->verify_no_cset_oops(true /* verify_stacks */, 4176 true /* verify_enqueued_buffers */, 4177 true /* verify_thread_buffers */, 4178 true /* verify_fingers */); 4179 _cm->note_end_of_gc(); 4180 4181 // This timing is only used by the ergonomics to handle our pause target. 4182 // It is unclear why this should not include the full pause. We will 4183 // investigate this in CR 7178365. 4184 double sample_end_time_sec = os::elapsedTime(); 4185 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 4186 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info); 4187 4188 MemoryService::track_memory_usage(); 4189 4190 // In prepare_for_verify() below we'll need to scan the deferred 4191 // update buffers to bring the RSets up-to-date if 4192 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4193 // the update buffers we'll probably need to scan cards on the 4194 // regions we just allocated to (i.e., the GC alloc 4195 // regions). However, during the last GC we called 4196 // set_saved_mark() on all the GC alloc regions, so card 4197 // scanning might skip the [saved_mark_word()...top()] area of 4198 // those regions (i.e., the area we allocated objects into 4199 // during the last GC). But it shouldn't. Given that 4200 // saved_mark_word() is conditional on whether the GC time stamp 4201 // on the region is current or not, by incrementing the GC time 4202 // stamp here we invalidate all the GC time stamps on all the 4203 // regions and saved_mark_word() will simply return top() for 4204 // all the regions. This is a nicer way of ensuring this rather 4205 // than iterating over the regions and fixing them. In fact, the 4206 // GC time stamp increment here also ensures that 4207 // saved_mark_word() will return top() between pauses, i.e., 4208 // during concurrent refinement. So we don't need the 4209 // is_gc_active() check to decided which top to use when 4210 // scanning cards (see CR 7039627). 4211 increment_gc_time_stamp(); 4212 4213 verify_after_gc(); 4214 4215 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4216 ref_processor_stw()->verify_no_references_recorded(); 4217 4218 // CM reference discovery will be re-enabled if necessary. 4219 } 4220 4221 // We should do this after we potentially expand the heap so 4222 // that all the COMMIT events are generated before the end GC 4223 // event, and after we retire the GC alloc regions so that all 4224 // RETIRE events are generated before the end GC event. 4225 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4226 4227 if (mark_in_progress()) { 4228 concurrent_mark()->update_g1_committed(); 4229 } 4230 4231 #ifdef TRACESPINNING 4232 ParallelTaskTerminator::print_termination_counts(); 4233 #endif 4234 4235 gc_epilogue(false); 4236 } 4237 4238 // Print the remainder of the GC log output. 4239 log_gc_footer(os::elapsedTime() - pause_start_sec); 4240 4241 // It is not yet to safe to tell the concurrent mark to 4242 // start as we have some optional output below. We don't want the 4243 // output from the concurrent mark thread interfering with this 4244 // logging output either. 4245 4246 _hrs.verify_optional(); 4247 verify_region_sets_optional(); 4248 4249 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4250 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4251 4252 print_heap_after_gc(); 4253 trace_heap_after_gc(_gc_tracer_stw); 4254 4255 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4256 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4257 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4258 // before any GC notifications are raised. 4259 g1mm()->update_sizes(); 4260 4261 _gc_tracer_stw->report_evacuation_info(&evacuation_info); 4262 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold()); 4263 _gc_timer_stw->register_gc_end(os::elapsed_counter()); 4264 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions()); 4265 } 4266 // It should now be safe to tell the concurrent mark thread to start 4267 // without its logging output interfering with the logging output 4268 // that came from the pause. 4269 4270 if (should_start_conc_mark) { 4271 // CAUTION: after the doConcurrentMark() call below, 4272 // the concurrent marking thread(s) could be running 4273 // concurrently with us. Make sure that anything after 4274 // this point does not assume that we are the only GC thread 4275 // running. Note: of course, the actual marking work will 4276 // not start until the safepoint itself is released in 4277 // ConcurrentGCThread::safepoint_desynchronize(). 4278 doConcurrentMark(); 4279 } 4280 4281 return true; 4282 } 4283 4284 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4285 { 4286 size_t gclab_word_size; 4287 switch (purpose) { 4288 case GCAllocForSurvived: 4289 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4290 break; 4291 case GCAllocForTenured: 4292 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4293 break; 4294 default: 4295 assert(false, "unknown GCAllocPurpose"); 4296 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4297 break; 4298 } 4299 4300 // Prevent humongous PLAB sizes for two reasons: 4301 // * PLABs are allocated using a similar paths as oops, but should 4302 // never be in a humongous region 4303 // * Allowing humongous PLABs needlessly churns the region free lists 4304 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4305 } 4306 4307 void G1CollectedHeap::init_mutator_alloc_region() { 4308 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4309 _mutator_alloc_region.init(); 4310 } 4311 4312 void G1CollectedHeap::release_mutator_alloc_region() { 4313 _mutator_alloc_region.release(); 4314 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4315 } 4316 4317 void G1CollectedHeap::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { 4318 assert_at_safepoint(true /* should_be_vm_thread */); 4319 4320 _survivor_gc_alloc_region.init(); 4321 _old_gc_alloc_region.init(); 4322 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4323 _retained_old_gc_alloc_region = NULL; 4324 4325 // We will discard the current GC alloc region if: 4326 // a) it's in the collection set (it can happen!), 4327 // b) it's already full (no point in using it), 4328 // c) it's empty (this means that it was emptied during 4329 // a cleanup and it should be on the free list now), or 4330 // d) it's humongous (this means that it was emptied 4331 // during a cleanup and was added to the free list, but 4332 // has been subsequently used to allocate a humongous 4333 // object that may be less than the region size). 4334 if (retained_region != NULL && 4335 !retained_region->in_collection_set() && 4336 !(retained_region->top() == retained_region->end()) && 4337 !retained_region->is_empty() && 4338 !retained_region->isHumongous()) { 4339 retained_region->set_saved_mark(); 4340 // The retained region was added to the old region set when it was 4341 // retired. We have to remove it now, since we don't allow regions 4342 // we allocate to in the region sets. We'll re-add it later, when 4343 // it's retired again. 4344 _old_set.remove(retained_region); 4345 bool during_im = g1_policy()->during_initial_mark_pause(); 4346 retained_region->note_start_of_copying(during_im); 4347 _old_gc_alloc_region.set(retained_region); 4348 _hr_printer.reuse(retained_region); 4349 evacuation_info.set_alloc_regions_used_before(retained_region->used()); 4350 } 4351 } 4352 4353 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { 4354 evacuation_info.set_allocation_regions(_survivor_gc_alloc_region.count() + 4355 _old_gc_alloc_region.count()); 4356 _survivor_gc_alloc_region.release(); 4357 // If we have an old GC alloc region to release, we'll save it in 4358 // _retained_old_gc_alloc_region. If we don't 4359 // _retained_old_gc_alloc_region will become NULL. This is what we 4360 // want either way so no reason to check explicitly for either 4361 // condition. 4362 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4363 4364 if (ResizePLAB) { 4365 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4366 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4367 } 4368 } 4369 4370 void G1CollectedHeap::abandon_gc_alloc_regions() { 4371 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4372 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4373 _retained_old_gc_alloc_region = NULL; 4374 } 4375 4376 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4377 _drain_in_progress = false; 4378 set_evac_failure_closure(cl); 4379 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4380 } 4381 4382 void G1CollectedHeap::finalize_for_evac_failure() { 4383 assert(_evac_failure_scan_stack != NULL && 4384 _evac_failure_scan_stack->length() == 0, 4385 "Postcondition"); 4386 assert(!_drain_in_progress, "Postcondition"); 4387 delete _evac_failure_scan_stack; 4388 _evac_failure_scan_stack = NULL; 4389 } 4390 4391 void G1CollectedHeap::remove_self_forwarding_pointers() { 4392 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4393 4394 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4395 4396 if (G1CollectedHeap::use_parallel_gc_threads()) { 4397 set_par_threads(); 4398 workers()->run_task(&rsfp_task); 4399 set_par_threads(0); 4400 } else { 4401 rsfp_task.work(0); 4402 } 4403 4404 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4405 4406 // Reset the claim values in the regions in the collection set. 4407 reset_cset_heap_region_claim_values(); 4408 4409 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4410 4411 // Now restore saved marks, if any. 4412 assert(_objs_with_preserved_marks.size() == 4413 _preserved_marks_of_objs.size(), "Both or none."); 4414 while (!_objs_with_preserved_marks.is_empty()) { 4415 oop obj = _objs_with_preserved_marks.pop(); 4416 markOop m = _preserved_marks_of_objs.pop(); 4417 obj->set_mark(m); 4418 } 4419 _objs_with_preserved_marks.clear(true); 4420 _preserved_marks_of_objs.clear(true); 4421 } 4422 4423 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4424 _evac_failure_scan_stack->push(obj); 4425 } 4426 4427 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4428 assert(_evac_failure_scan_stack != NULL, "precondition"); 4429 4430 while (_evac_failure_scan_stack->length() > 0) { 4431 oop obj = _evac_failure_scan_stack->pop(); 4432 _evac_failure_closure->set_region(heap_region_containing(obj)); 4433 obj->oop_iterate_backwards(_evac_failure_closure); 4434 } 4435 } 4436 4437 oop 4438 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state, 4439 oop old) { 4440 assert(obj_in_cs(old), 4441 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4442 (HeapWord*) old)); 4443 markOop m = old->mark(); 4444 oop forward_ptr = old->forward_to_atomic(old); 4445 if (forward_ptr == NULL) { 4446 // Forward-to-self succeeded. 4447 assert(_par_scan_state != NULL, "par scan state"); 4448 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4449 uint queue_num = _par_scan_state->queue_num(); 4450 4451 _evacuation_failed = true; 4452 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size()); 4453 if (_evac_failure_closure != cl) { 4454 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4455 assert(!_drain_in_progress, 4456 "Should only be true while someone holds the lock."); 4457 // Set the global evac-failure closure to the current thread's. 4458 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4459 set_evac_failure_closure(cl); 4460 // Now do the common part. 4461 handle_evacuation_failure_common(old, m); 4462 // Reset to NULL. 4463 set_evac_failure_closure(NULL); 4464 } else { 4465 // The lock is already held, and this is recursive. 4466 assert(_drain_in_progress, "This should only be the recursive case."); 4467 handle_evacuation_failure_common(old, m); 4468 } 4469 return old; 4470 } else { 4471 // Forward-to-self failed. Either someone else managed to allocate 4472 // space for this object (old != forward_ptr) or they beat us in 4473 // self-forwarding it (old == forward_ptr). 4474 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4475 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4476 "should not be in the CSet", 4477 (HeapWord*) old, (HeapWord*) forward_ptr)); 4478 return forward_ptr; 4479 } 4480 } 4481 4482 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4483 preserve_mark_if_necessary(old, m); 4484 4485 HeapRegion* r = heap_region_containing(old); 4486 if (!r->evacuation_failed()) { 4487 r->set_evacuation_failed(true); 4488 _hr_printer.evac_failure(r); 4489 } 4490 4491 push_on_evac_failure_scan_stack(old); 4492 4493 if (!_drain_in_progress) { 4494 // prevent recursion in copy_to_survivor_space() 4495 _drain_in_progress = true; 4496 drain_evac_failure_scan_stack(); 4497 _drain_in_progress = false; 4498 } 4499 } 4500 4501 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4502 assert(evacuation_failed(), "Oversaving!"); 4503 // We want to call the "for_promotion_failure" version only in the 4504 // case of a promotion failure. 4505 if (m->must_be_preserved_for_promotion_failure(obj)) { 4506 _objs_with_preserved_marks.push(obj); 4507 _preserved_marks_of_objs.push(m); 4508 } 4509 } 4510 4511 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4512 size_t word_size) { 4513 if (purpose == GCAllocForSurvived) { 4514 HeapWord* result = survivor_attempt_allocation(word_size); 4515 if (result != NULL) { 4516 return result; 4517 } else { 4518 // Let's try to allocate in the old gen in case we can fit the 4519 // object there. 4520 return old_attempt_allocation(word_size); 4521 } 4522 } else { 4523 assert(purpose == GCAllocForTenured, "sanity"); 4524 HeapWord* result = old_attempt_allocation(word_size); 4525 if (result != NULL) { 4526 return result; 4527 } else { 4528 // Let's try to allocate in the survivors in case we can fit the 4529 // object there. 4530 return survivor_attempt_allocation(word_size); 4531 } 4532 } 4533 4534 ShouldNotReachHere(); 4535 // Trying to keep some compilers happy. 4536 return NULL; 4537 } 4538 4539 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4540 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4541 4542 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4543 : _g1h(g1h), 4544 _refs(g1h->task_queue(queue_num)), 4545 _dcq(&g1h->dirty_card_queue_set()), 4546 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4547 _g1_rem(g1h->g1_rem_set()), 4548 _hash_seed(17), _queue_num(queue_num), 4549 _term_attempts(0), 4550 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4551 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4552 _age_table(false), 4553 _strong_roots_time(0), _term_time(0), 4554 _alloc_buffer_waste(0), _undo_waste(0) { 4555 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4556 // we "sacrifice" entry 0 to keep track of surviving bytes for 4557 // non-young regions (where the age is -1) 4558 // We also add a few elements at the beginning and at the end in 4559 // an attempt to eliminate cache contention 4560 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4561 uint array_length = PADDING_ELEM_NUM + 4562 real_length + 4563 PADDING_ELEM_NUM; 4564 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4565 if (_surviving_young_words_base == NULL) 4566 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR, 4567 "Not enough space for young surv histo."); 4568 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4569 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4570 4571 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4572 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4573 4574 _start = os::elapsedTime(); 4575 } 4576 4577 void 4578 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4579 { 4580 st->print_raw_cr("GC Termination Stats"); 4581 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4582 " ------waste (KiB)------"); 4583 st->print_raw_cr("thr ms ms % ms % attempts" 4584 " total alloc undo"); 4585 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4586 " ------- ------- -------"); 4587 } 4588 4589 void 4590 G1ParScanThreadState::print_termination_stats(int i, 4591 outputStream* const st) const 4592 { 4593 const double elapsed_ms = elapsed_time() * 1000.0; 4594 const double s_roots_ms = strong_roots_time() * 1000.0; 4595 const double term_ms = term_time() * 1000.0; 4596 st->print_cr("%3d %9.2f %9.2f %6.2f " 4597 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4598 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4599 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4600 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4601 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4602 alloc_buffer_waste() * HeapWordSize / K, 4603 undo_waste() * HeapWordSize / K); 4604 } 4605 4606 #ifdef ASSERT 4607 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4608 assert(ref != NULL, "invariant"); 4609 assert(UseCompressedOops, "sanity"); 4610 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4611 oop p = oopDesc::load_decode_heap_oop(ref); 4612 assert(_g1h->is_in_g1_reserved(p), 4613 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4614 return true; 4615 } 4616 4617 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4618 assert(ref != NULL, "invariant"); 4619 if (has_partial_array_mask(ref)) { 4620 // Must be in the collection set--it's already been copied. 4621 oop p = clear_partial_array_mask(ref); 4622 assert(_g1h->obj_in_cs(p), 4623 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4624 } else { 4625 oop p = oopDesc::load_decode_heap_oop(ref); 4626 assert(_g1h->is_in_g1_reserved(p), 4627 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4628 } 4629 return true; 4630 } 4631 4632 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4633 if (ref.is_narrow()) { 4634 return verify_ref((narrowOop*) ref); 4635 } else { 4636 return verify_ref((oop*) ref); 4637 } 4638 } 4639 #endif // ASSERT 4640 4641 void G1ParScanThreadState::trim_queue() { 4642 assert(_evac_cl != NULL, "not set"); 4643 assert(_evac_failure_cl != NULL, "not set"); 4644 assert(_partial_scan_cl != NULL, "not set"); 4645 4646 StarTask ref; 4647 do { 4648 // Drain the overflow stack first, so other threads can steal. 4649 while (refs()->pop_overflow(ref)) { 4650 deal_with_reference(ref); 4651 } 4652 4653 while (refs()->pop_local(ref)) { 4654 deal_with_reference(ref); 4655 } 4656 } while (!refs()->is_empty()); 4657 } 4658 4659 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4660 G1ParScanThreadState* par_scan_state) : 4661 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4662 _par_scan_state(par_scan_state), 4663 _worker_id(par_scan_state->queue_num()), 4664 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4665 _mark_in_progress(_g1->mark_in_progress()) { } 4666 4667 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4668 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4669 #ifdef ASSERT 4670 HeapRegion* hr = _g1->heap_region_containing(obj); 4671 assert(hr != NULL, "sanity"); 4672 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4673 #endif // ASSERT 4674 4675 // We know that the object is not moving so it's safe to read its size. 4676 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4677 } 4678 4679 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4680 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4681 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4682 #ifdef ASSERT 4683 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4684 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4685 assert(from_obj != to_obj, "should not be self-forwarded"); 4686 4687 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4688 assert(from_hr != NULL, "sanity"); 4689 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4690 4691 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4692 assert(to_hr != NULL, "sanity"); 4693 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4694 #endif // ASSERT 4695 4696 // The object might be in the process of being copied by another 4697 // worker so we cannot trust that its to-space image is 4698 // well-formed. So we have to read its size from its from-space 4699 // image which we know should not be changing. 4700 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4701 } 4702 4703 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4704 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4705 ::copy_to_survivor_space(oop old) { 4706 size_t word_sz = old->size(); 4707 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4708 // +1 to make the -1 indexes valid... 4709 int young_index = from_region->young_index_in_cset()+1; 4710 assert( (from_region->is_young() && young_index > 0) || 4711 (!from_region->is_young() && young_index == 0), "invariant" ); 4712 G1CollectorPolicy* g1p = _g1->g1_policy(); 4713 markOop m = old->mark(); 4714 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4715 : m->age(); 4716 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4717 word_sz); 4718 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4719 #ifndef PRODUCT 4720 // Should this evacuation fail? 4721 if (_g1->evacuation_should_fail()) { 4722 if (obj_ptr != NULL) { 4723 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4724 obj_ptr = NULL; 4725 } 4726 } 4727 #endif // !PRODUCT 4728 4729 if (obj_ptr == NULL) { 4730 // This will either forward-to-self, or detect that someone else has 4731 // installed a forwarding pointer. 4732 return _g1->handle_evacuation_failure_par(_par_scan_state, old); 4733 } 4734 4735 oop obj = oop(obj_ptr); 4736 4737 // We're going to allocate linearly, so might as well prefetch ahead. 4738 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4739 4740 oop forward_ptr = old->forward_to_atomic(obj); 4741 if (forward_ptr == NULL) { 4742 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4743 if (g1p->track_object_age(alloc_purpose)) { 4744 // We could simply do obj->incr_age(). However, this causes a 4745 // performance issue. obj->incr_age() will first check whether 4746 // the object has a displaced mark by checking its mark word; 4747 // getting the mark word from the new location of the object 4748 // stalls. So, given that we already have the mark word and we 4749 // are about to install it anyway, it's better to increase the 4750 // age on the mark word, when the object does not have a 4751 // displaced mark word. We're not expecting many objects to have 4752 // a displaced marked word, so that case is not optimized 4753 // further (it could be...) and we simply call obj->incr_age(). 4754 4755 if (m->has_displaced_mark_helper()) { 4756 // in this case, we have to install the mark word first, 4757 // otherwise obj looks to be forwarded (the old mark word, 4758 // which contains the forward pointer, was copied) 4759 obj->set_mark(m); 4760 obj->incr_age(); 4761 } else { 4762 m = m->incr_age(); 4763 obj->set_mark(m); 4764 } 4765 _par_scan_state->age_table()->add(obj, word_sz); 4766 } else { 4767 obj->set_mark(m); 4768 } 4769 4770 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4771 surv_young_words[young_index] += word_sz; 4772 4773 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4774 // We keep track of the next start index in the length field of 4775 // the to-space object. The actual length can be found in the 4776 // length field of the from-space object. 4777 arrayOop(obj)->set_length(0); 4778 oop* old_p = set_partial_array_mask(old); 4779 _par_scan_state->push_on_queue(old_p); 4780 } else { 4781 // No point in using the slower heap_region_containing() method, 4782 // given that we know obj is in the heap. 4783 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4784 obj->oop_iterate_backwards(&_scanner); 4785 } 4786 } else { 4787 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4788 obj = forward_ptr; 4789 } 4790 return obj; 4791 } 4792 4793 template <class T> 4794 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4795 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4796 _scanned_klass->record_modified_oops(); 4797 } 4798 } 4799 4800 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4801 template <class T> 4802 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4803 ::do_oop_work(T* p) { 4804 oop obj = oopDesc::load_decode_heap_oop(p); 4805 assert(barrier != G1BarrierRS || obj != NULL, 4806 "Precondition: G1BarrierRS implies obj is non-NULL"); 4807 4808 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4809 4810 // here the null check is implicit in the cset_fast_test() test 4811 if (_g1->in_cset_fast_test(obj)) { 4812 oop forwardee; 4813 if (obj->is_forwarded()) { 4814 forwardee = obj->forwardee(); 4815 } else { 4816 forwardee = copy_to_survivor_space(obj); 4817 } 4818 assert(forwardee != NULL, "forwardee should not be NULL"); 4819 oopDesc::encode_store_heap_oop(p, forwardee); 4820 if (do_mark_object && forwardee != obj) { 4821 // If the object is self-forwarded we don't need to explicitly 4822 // mark it, the evacuation failure protocol will do so. 4823 mark_forwarded_object(obj, forwardee); 4824 } 4825 4826 // When scanning the RS, we only care about objs in CS. 4827 if (barrier == G1BarrierRS) { 4828 _par_scan_state->update_rs(_from, p, _worker_id); 4829 } else if (barrier == G1BarrierKlass) { 4830 do_klass_barrier(p, forwardee); 4831 } 4832 } else { 4833 // The object is not in collection set. If we're a root scanning 4834 // closure during an initial mark pause (i.e. do_mark_object will 4835 // be true) then attempt to mark the object. 4836 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4837 mark_object(obj); 4838 } 4839 } 4840 4841 if (barrier == G1BarrierEvac && obj != NULL) { 4842 _par_scan_state->update_rs(_from, p, _worker_id); 4843 } 4844 4845 if (do_gen_barrier && obj != NULL) { 4846 par_do_barrier(p); 4847 } 4848 } 4849 4850 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4851 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4852 4853 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4854 assert(has_partial_array_mask(p), "invariant"); 4855 oop from_obj = clear_partial_array_mask(p); 4856 4857 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4858 assert(from_obj->is_objArray(), "must be obj array"); 4859 objArrayOop from_obj_array = objArrayOop(from_obj); 4860 // The from-space object contains the real length. 4861 int length = from_obj_array->length(); 4862 4863 assert(from_obj->is_forwarded(), "must be forwarded"); 4864 oop to_obj = from_obj->forwardee(); 4865 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4866 objArrayOop to_obj_array = objArrayOop(to_obj); 4867 // We keep track of the next start index in the length field of the 4868 // to-space object. 4869 int next_index = to_obj_array->length(); 4870 assert(0 <= next_index && next_index < length, 4871 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4872 4873 int start = next_index; 4874 int end = length; 4875 int remainder = end - start; 4876 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4877 if (remainder > 2 * ParGCArrayScanChunk) { 4878 end = start + ParGCArrayScanChunk; 4879 to_obj_array->set_length(end); 4880 // Push the remainder before we process the range in case another 4881 // worker has run out of things to do and can steal it. 4882 oop* from_obj_p = set_partial_array_mask(from_obj); 4883 _par_scan_state->push_on_queue(from_obj_p); 4884 } else { 4885 assert(length == end, "sanity"); 4886 // We'll process the final range for this object. Restore the length 4887 // so that the heap remains parsable in case of evacuation failure. 4888 to_obj_array->set_length(end); 4889 } 4890 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4891 // Process indexes [start,end). It will also process the header 4892 // along with the first chunk (i.e., the chunk with start == 0). 4893 // Note that at this point the length field of to_obj_array is not 4894 // correct given that we are using it to keep track of the next 4895 // start index. oop_iterate_range() (thankfully!) ignores the length 4896 // field and only relies on the start / end parameters. It does 4897 // however return the size of the object which will be incorrect. So 4898 // we have to ignore it even if we wanted to use it. 4899 to_obj_array->oop_iterate_range(&_scanner, start, end); 4900 } 4901 4902 class G1ParEvacuateFollowersClosure : public VoidClosure { 4903 protected: 4904 G1CollectedHeap* _g1h; 4905 G1ParScanThreadState* _par_scan_state; 4906 RefToScanQueueSet* _queues; 4907 ParallelTaskTerminator* _terminator; 4908 4909 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4910 RefToScanQueueSet* queues() { return _queues; } 4911 ParallelTaskTerminator* terminator() { return _terminator; } 4912 4913 public: 4914 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4915 G1ParScanThreadState* par_scan_state, 4916 RefToScanQueueSet* queues, 4917 ParallelTaskTerminator* terminator) 4918 : _g1h(g1h), _par_scan_state(par_scan_state), 4919 _queues(queues), _terminator(terminator) {} 4920 4921 void do_void(); 4922 4923 private: 4924 inline bool offer_termination(); 4925 }; 4926 4927 bool G1ParEvacuateFollowersClosure::offer_termination() { 4928 G1ParScanThreadState* const pss = par_scan_state(); 4929 pss->start_term_time(); 4930 const bool res = terminator()->offer_termination(); 4931 pss->end_term_time(); 4932 return res; 4933 } 4934 4935 void G1ParEvacuateFollowersClosure::do_void() { 4936 StarTask stolen_task; 4937 G1ParScanThreadState* const pss = par_scan_state(); 4938 pss->trim_queue(); 4939 4940 do { 4941 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4942 assert(pss->verify_task(stolen_task), "sanity"); 4943 if (stolen_task.is_narrow()) { 4944 pss->deal_with_reference((narrowOop*) stolen_task); 4945 } else { 4946 pss->deal_with_reference((oop*) stolen_task); 4947 } 4948 4949 // We've just processed a reference and we might have made 4950 // available new entries on the queues. So we have to make sure 4951 // we drain the queues as necessary. 4952 pss->trim_queue(); 4953 } 4954 } while (!offer_termination()); 4955 4956 pss->retire_alloc_buffers(); 4957 } 4958 4959 class G1KlassScanClosure : public KlassClosure { 4960 G1ParCopyHelper* _closure; 4961 bool _process_only_dirty; 4962 int _count; 4963 public: 4964 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4965 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4966 void do_klass(Klass* klass) { 4967 // If the klass has not been dirtied we know that there's 4968 // no references into the young gen and we can skip it. 4969 if (!_process_only_dirty || klass->has_modified_oops()) { 4970 // Clean the klass since we're going to scavenge all the metadata. 4971 klass->clear_modified_oops(); 4972 4973 // Tell the closure that this klass is the Klass to scavenge 4974 // and is the one to dirty if oops are left pointing into the young gen. 4975 _closure->set_scanned_klass(klass); 4976 4977 klass->oops_do(_closure); 4978 4979 _closure->set_scanned_klass(NULL); 4980 } 4981 _count++; 4982 } 4983 }; 4984 4985 class G1ParTask : public AbstractGangTask { 4986 protected: 4987 G1CollectedHeap* _g1h; 4988 RefToScanQueueSet *_queues; 4989 ParallelTaskTerminator _terminator; 4990 uint _n_workers; 4991 4992 Mutex _stats_lock; 4993 Mutex* stats_lock() { return &_stats_lock; } 4994 4995 size_t getNCards() { 4996 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4997 / G1BlockOffsetSharedArray::N_bytes; 4998 } 4999 5000 public: 5001 G1ParTask(G1CollectedHeap* g1h, 5002 RefToScanQueueSet *task_queues) 5003 : AbstractGangTask("G1 collection"), 5004 _g1h(g1h), 5005 _queues(task_queues), 5006 _terminator(0, _queues), 5007 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 5008 {} 5009 5010 RefToScanQueueSet* queues() { return _queues; } 5011 5012 RefToScanQueue *work_queue(int i) { 5013 return queues()->queue(i); 5014 } 5015 5016 ParallelTaskTerminator* terminator() { return &_terminator; } 5017 5018 virtual void set_for_termination(int active_workers) { 5019 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 5020 // in the young space (_par_seq_tasks) in the G1 heap 5021 // for SequentialSubTasksDone. 5022 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 5023 // both of which need setting by set_n_termination(). 5024 _g1h->SharedHeap::set_n_termination(active_workers); 5025 _g1h->set_n_termination(active_workers); 5026 terminator()->reset_for_reuse(active_workers); 5027 _n_workers = active_workers; 5028 } 5029 5030 void work(uint worker_id) { 5031 if (worker_id >= _n_workers) return; // no work needed this round 5032 5033 double start_time_ms = os::elapsedTime() * 1000.0; 5034 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 5035 5036 { 5037 ResourceMark rm; 5038 HandleMark hm; 5039 5040 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 5041 5042 G1ParScanThreadState pss(_g1h, worker_id); 5043 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 5044 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 5045 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 5046 5047 pss.set_evac_closure(&scan_evac_cl); 5048 pss.set_evac_failure_closure(&evac_failure_cl); 5049 pss.set_partial_scan_closure(&partial_scan_cl); 5050 5051 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 5052 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 5053 5054 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 5055 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 5056 5057 bool only_young = _g1h->g1_policy()->gcs_are_young(); 5058 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 5059 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 5060 5061 OopClosure* scan_root_cl = &only_scan_root_cl; 5062 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 5063 5064 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5065 // We also need to mark copied objects. 5066 scan_root_cl = &scan_mark_root_cl; 5067 scan_klasses_cl = &scan_mark_klasses_cl_s; 5068 } 5069 5070 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 5071 5072 // Don't scan the scavengable methods in the code cache as part 5073 // of strong root scanning. The code roots that point into a 5074 // region in the collection set are scanned when we scan the 5075 // region's RSet. 5076 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings; 5077 5078 pss.start_strong_roots(); 5079 _g1h->g1_process_strong_roots(/* is scavenging */ true, 5080 SharedHeap::ScanningOption(so), 5081 scan_root_cl, 5082 &push_heap_rs_cl, 5083 scan_klasses_cl, 5084 worker_id); 5085 pss.end_strong_roots(); 5086 5087 { 5088 double start = os::elapsedTime(); 5089 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 5090 evac.do_void(); 5091 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 5092 double term_ms = pss.term_time()*1000.0; 5093 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 5094 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 5095 } 5096 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 5097 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 5098 5099 if (ParallelGCVerbose) { 5100 MutexLocker x(stats_lock()); 5101 pss.print_termination_stats(worker_id); 5102 } 5103 5104 assert(pss.refs()->is_empty(), "should be empty"); 5105 5106 // Close the inner scope so that the ResourceMark and HandleMark 5107 // destructors are executed here and are included as part of the 5108 // "GC Worker Time". 5109 } 5110 5111 double end_time_ms = os::elapsedTime() * 1000.0; 5112 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 5113 } 5114 }; 5115 5116 // *** Common G1 Evacuation Stuff 5117 5118 // This method is run in a GC worker. 5119 5120 void 5121 G1CollectedHeap:: 5122 g1_process_strong_roots(bool is_scavenging, 5123 ScanningOption so, 5124 OopClosure* scan_non_heap_roots, 5125 OopsInHeapRegionClosure* scan_rs, 5126 G1KlassScanClosure* scan_klasses, 5127 int worker_i) { 5128 5129 // First scan the strong roots 5130 double ext_roots_start = os::elapsedTime(); 5131 double closure_app_time_sec = 0.0; 5132 5133 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 5134 5135 assert(so & SO_CodeCache || scan_rs != NULL, "must scan code roots somehow"); 5136 // Walk the code cache/strong code roots w/o buffering, because StarTask 5137 // cannot handle unaligned oop locations. 5138 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, true /* do_marking */); 5139 5140 process_strong_roots(false, // no scoping; this is parallel code 5141 is_scavenging, so, 5142 &buf_scan_non_heap_roots, 5143 &eager_scan_code_roots, 5144 scan_klasses 5145 ); 5146 5147 // Now the CM ref_processor roots. 5148 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5149 // We need to treat the discovered reference lists of the 5150 // concurrent mark ref processor as roots and keep entries 5151 // (which are added by the marking threads) on them live 5152 // until they can be processed at the end of marking. 5153 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5154 } 5155 5156 // Finish up any enqueued closure apps (attributed as object copy time). 5157 buf_scan_non_heap_roots.done(); 5158 5159 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5160 5161 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5162 5163 double ext_root_time_ms = 5164 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5165 5166 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5167 5168 // During conc marking we have to filter the per-thread SATB buffers 5169 // to make sure we remove any oops into the CSet (which will show up 5170 // as implicitly live). 5171 double satb_filtering_ms = 0.0; 5172 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5173 if (mark_in_progress()) { 5174 double satb_filter_start = os::elapsedTime(); 5175 5176 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5177 5178 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5179 } 5180 } 5181 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5182 5183 // If this is an initial mark pause, and we're not scanning 5184 // the entire code cache, we need to mark the oops in the 5185 // strong code root lists for the regions that are not in 5186 // the collection set. 5187 // Note all threads participate in this set of root tasks. 5188 double mark_strong_code_roots_ms = 0.0; 5189 if (g1_policy()->during_initial_mark_pause() && !(so & SO_CodeCache)) { 5190 double mark_strong_roots_start = os::elapsedTime(); 5191 mark_strong_code_roots(worker_i); 5192 mark_strong_code_roots_ms = (os::elapsedTime() - mark_strong_roots_start) * 1000.0; 5193 } 5194 g1_policy()->phase_times()->record_strong_code_root_mark_time(worker_i, mark_strong_code_roots_ms); 5195 5196 // Now scan the complement of the collection set. 5197 if (scan_rs != NULL) { 5198 g1_rem_set()->oops_into_collection_set_do(scan_rs, &eager_scan_code_roots, worker_i); 5199 } 5200 _process_strong_tasks->all_tasks_completed(); 5201 } 5202 5203 void 5204 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) { 5205 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5206 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs); 5207 } 5208 5209 // Weak Reference Processing support 5210 5211 // An always "is_alive" closure that is used to preserve referents. 5212 // If the object is non-null then it's alive. Used in the preservation 5213 // of referent objects that are pointed to by reference objects 5214 // discovered by the CM ref processor. 5215 class G1AlwaysAliveClosure: public BoolObjectClosure { 5216 G1CollectedHeap* _g1; 5217 public: 5218 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5219 bool do_object_b(oop p) { 5220 if (p != NULL) { 5221 return true; 5222 } 5223 return false; 5224 } 5225 }; 5226 5227 bool G1STWIsAliveClosure::do_object_b(oop p) { 5228 // An object is reachable if it is outside the collection set, 5229 // or is inside and copied. 5230 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5231 } 5232 5233 // Non Copying Keep Alive closure 5234 class G1KeepAliveClosure: public OopClosure { 5235 G1CollectedHeap* _g1; 5236 public: 5237 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5238 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5239 void do_oop( oop* p) { 5240 oop obj = *p; 5241 5242 if (_g1->obj_in_cs(obj)) { 5243 assert( obj->is_forwarded(), "invariant" ); 5244 *p = obj->forwardee(); 5245 } 5246 } 5247 }; 5248 5249 // Copying Keep Alive closure - can be called from both 5250 // serial and parallel code as long as different worker 5251 // threads utilize different G1ParScanThreadState instances 5252 // and different queues. 5253 5254 class G1CopyingKeepAliveClosure: public OopClosure { 5255 G1CollectedHeap* _g1h; 5256 OopClosure* _copy_non_heap_obj_cl; 5257 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5258 G1ParScanThreadState* _par_scan_state; 5259 5260 public: 5261 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5262 OopClosure* non_heap_obj_cl, 5263 OopsInHeapRegionClosure* metadata_obj_cl, 5264 G1ParScanThreadState* pss): 5265 _g1h(g1h), 5266 _copy_non_heap_obj_cl(non_heap_obj_cl), 5267 _copy_metadata_obj_cl(metadata_obj_cl), 5268 _par_scan_state(pss) 5269 {} 5270 5271 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5272 virtual void do_oop( oop* p) { do_oop_work(p); } 5273 5274 template <class T> void do_oop_work(T* p) { 5275 oop obj = oopDesc::load_decode_heap_oop(p); 5276 5277 if (_g1h->obj_in_cs(obj)) { 5278 // If the referent object has been forwarded (either copied 5279 // to a new location or to itself in the event of an 5280 // evacuation failure) then we need to update the reference 5281 // field and, if both reference and referent are in the G1 5282 // heap, update the RSet for the referent. 5283 // 5284 // If the referent has not been forwarded then we have to keep 5285 // it alive by policy. Therefore we have copy the referent. 5286 // 5287 // If the reference field is in the G1 heap then we can push 5288 // on the PSS queue. When the queue is drained (after each 5289 // phase of reference processing) the object and it's followers 5290 // will be copied, the reference field set to point to the 5291 // new location, and the RSet updated. Otherwise we need to 5292 // use the the non-heap or metadata closures directly to copy 5293 // the referent object and update the pointer, while avoiding 5294 // updating the RSet. 5295 5296 if (_g1h->is_in_g1_reserved(p)) { 5297 _par_scan_state->push_on_queue(p); 5298 } else { 5299 assert(!ClassLoaderDataGraph::contains((address)p), 5300 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5301 PTR_FORMAT, p)); 5302 _copy_non_heap_obj_cl->do_oop(p); 5303 } 5304 } 5305 } 5306 }; 5307 5308 // Serial drain queue closure. Called as the 'complete_gc' 5309 // closure for each discovered list in some of the 5310 // reference processing phases. 5311 5312 class G1STWDrainQueueClosure: public VoidClosure { 5313 protected: 5314 G1CollectedHeap* _g1h; 5315 G1ParScanThreadState* _par_scan_state; 5316 5317 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5318 5319 public: 5320 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5321 _g1h(g1h), 5322 _par_scan_state(pss) 5323 { } 5324 5325 void do_void() { 5326 G1ParScanThreadState* const pss = par_scan_state(); 5327 pss->trim_queue(); 5328 } 5329 }; 5330 5331 // Parallel Reference Processing closures 5332 5333 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5334 // processing during G1 evacuation pauses. 5335 5336 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5337 private: 5338 G1CollectedHeap* _g1h; 5339 RefToScanQueueSet* _queues; 5340 FlexibleWorkGang* _workers; 5341 int _active_workers; 5342 5343 public: 5344 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5345 FlexibleWorkGang* workers, 5346 RefToScanQueueSet *task_queues, 5347 int n_workers) : 5348 _g1h(g1h), 5349 _queues(task_queues), 5350 _workers(workers), 5351 _active_workers(n_workers) 5352 { 5353 assert(n_workers > 0, "shouldn't call this otherwise"); 5354 } 5355 5356 // Executes the given task using concurrent marking worker threads. 5357 virtual void execute(ProcessTask& task); 5358 virtual void execute(EnqueueTask& task); 5359 }; 5360 5361 // Gang task for possibly parallel reference processing 5362 5363 class G1STWRefProcTaskProxy: public AbstractGangTask { 5364 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5365 ProcessTask& _proc_task; 5366 G1CollectedHeap* _g1h; 5367 RefToScanQueueSet *_task_queues; 5368 ParallelTaskTerminator* _terminator; 5369 5370 public: 5371 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5372 G1CollectedHeap* g1h, 5373 RefToScanQueueSet *task_queues, 5374 ParallelTaskTerminator* terminator) : 5375 AbstractGangTask("Process reference objects in parallel"), 5376 _proc_task(proc_task), 5377 _g1h(g1h), 5378 _task_queues(task_queues), 5379 _terminator(terminator) 5380 {} 5381 5382 virtual void work(uint worker_id) { 5383 // The reference processing task executed by a single worker. 5384 ResourceMark rm; 5385 HandleMark hm; 5386 5387 G1STWIsAliveClosure is_alive(_g1h); 5388 5389 G1ParScanThreadState pss(_g1h, worker_id); 5390 5391 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5392 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5393 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5394 5395 pss.set_evac_closure(&scan_evac_cl); 5396 pss.set_evac_failure_closure(&evac_failure_cl); 5397 pss.set_partial_scan_closure(&partial_scan_cl); 5398 5399 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5400 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5401 5402 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5403 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5404 5405 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5406 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5407 5408 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5409 // We also need to mark copied objects. 5410 copy_non_heap_cl = ©_mark_non_heap_cl; 5411 copy_metadata_cl = ©_mark_metadata_cl; 5412 } 5413 5414 // Keep alive closure. 5415 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5416 5417 // Complete GC closure 5418 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5419 5420 // Call the reference processing task's work routine. 5421 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5422 5423 // Note we cannot assert that the refs array is empty here as not all 5424 // of the processing tasks (specifically phase2 - pp2_work) execute 5425 // the complete_gc closure (which ordinarily would drain the queue) so 5426 // the queue may not be empty. 5427 } 5428 }; 5429 5430 // Driver routine for parallel reference processing. 5431 // Creates an instance of the ref processing gang 5432 // task and has the worker threads execute it. 5433 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5434 assert(_workers != NULL, "Need parallel worker threads."); 5435 5436 ParallelTaskTerminator terminator(_active_workers, _queues); 5437 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5438 5439 _g1h->set_par_threads(_active_workers); 5440 _workers->run_task(&proc_task_proxy); 5441 _g1h->set_par_threads(0); 5442 } 5443 5444 // Gang task for parallel reference enqueueing. 5445 5446 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5447 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5448 EnqueueTask& _enq_task; 5449 5450 public: 5451 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5452 AbstractGangTask("Enqueue reference objects in parallel"), 5453 _enq_task(enq_task) 5454 { } 5455 5456 virtual void work(uint worker_id) { 5457 _enq_task.work(worker_id); 5458 } 5459 }; 5460 5461 // Driver routine for parallel reference enqueueing. 5462 // Creates an instance of the ref enqueueing gang 5463 // task and has the worker threads execute it. 5464 5465 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5466 assert(_workers != NULL, "Need parallel worker threads."); 5467 5468 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5469 5470 _g1h->set_par_threads(_active_workers); 5471 _workers->run_task(&enq_task_proxy); 5472 _g1h->set_par_threads(0); 5473 } 5474 5475 // End of weak reference support closures 5476 5477 // Abstract task used to preserve (i.e. copy) any referent objects 5478 // that are in the collection set and are pointed to by reference 5479 // objects discovered by the CM ref processor. 5480 5481 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5482 protected: 5483 G1CollectedHeap* _g1h; 5484 RefToScanQueueSet *_queues; 5485 ParallelTaskTerminator _terminator; 5486 uint _n_workers; 5487 5488 public: 5489 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5490 AbstractGangTask("ParPreserveCMReferents"), 5491 _g1h(g1h), 5492 _queues(task_queues), 5493 _terminator(workers, _queues), 5494 _n_workers(workers) 5495 { } 5496 5497 void work(uint worker_id) { 5498 ResourceMark rm; 5499 HandleMark hm; 5500 5501 G1ParScanThreadState pss(_g1h, worker_id); 5502 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5503 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5504 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5505 5506 pss.set_evac_closure(&scan_evac_cl); 5507 pss.set_evac_failure_closure(&evac_failure_cl); 5508 pss.set_partial_scan_closure(&partial_scan_cl); 5509 5510 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5511 5512 5513 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5514 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5515 5516 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5517 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5518 5519 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5520 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5521 5522 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5523 // We also need to mark copied objects. 5524 copy_non_heap_cl = ©_mark_non_heap_cl; 5525 copy_metadata_cl = ©_mark_metadata_cl; 5526 } 5527 5528 // Is alive closure 5529 G1AlwaysAliveClosure always_alive(_g1h); 5530 5531 // Copying keep alive closure. Applied to referent objects that need 5532 // to be copied. 5533 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5534 5535 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5536 5537 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5538 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5539 5540 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5541 // So this must be true - but assert just in case someone decides to 5542 // change the worker ids. 5543 assert(0 <= worker_id && worker_id < limit, "sanity"); 5544 assert(!rp->discovery_is_atomic(), "check this code"); 5545 5546 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5547 for (uint idx = worker_id; idx < limit; idx += stride) { 5548 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5549 5550 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5551 while (iter.has_next()) { 5552 // Since discovery is not atomic for the CM ref processor, we 5553 // can see some null referent objects. 5554 iter.load_ptrs(DEBUG_ONLY(true)); 5555 oop ref = iter.obj(); 5556 5557 // This will filter nulls. 5558 if (iter.is_referent_alive()) { 5559 iter.make_referent_alive(); 5560 } 5561 iter.move_to_next(); 5562 } 5563 } 5564 5565 // Drain the queue - which may cause stealing 5566 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5567 drain_queue.do_void(); 5568 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5569 assert(pss.refs()->is_empty(), "should be"); 5570 } 5571 }; 5572 5573 // Weak Reference processing during an evacuation pause (part 1). 5574 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5575 double ref_proc_start = os::elapsedTime(); 5576 5577 ReferenceProcessor* rp = _ref_processor_stw; 5578 assert(rp->discovery_enabled(), "should have been enabled"); 5579 5580 // Any reference objects, in the collection set, that were 'discovered' 5581 // by the CM ref processor should have already been copied (either by 5582 // applying the external root copy closure to the discovered lists, or 5583 // by following an RSet entry). 5584 // 5585 // But some of the referents, that are in the collection set, that these 5586 // reference objects point to may not have been copied: the STW ref 5587 // processor would have seen that the reference object had already 5588 // been 'discovered' and would have skipped discovering the reference, 5589 // but would not have treated the reference object as a regular oop. 5590 // As a result the copy closure would not have been applied to the 5591 // referent object. 5592 // 5593 // We need to explicitly copy these referent objects - the references 5594 // will be processed at the end of remarking. 5595 // 5596 // We also need to do this copying before we process the reference 5597 // objects discovered by the STW ref processor in case one of these 5598 // referents points to another object which is also referenced by an 5599 // object discovered by the STW ref processor. 5600 5601 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5602 no_of_gc_workers == workers()->active_workers(), 5603 "Need to reset active GC workers"); 5604 5605 set_par_threads(no_of_gc_workers); 5606 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5607 no_of_gc_workers, 5608 _task_queues); 5609 5610 if (G1CollectedHeap::use_parallel_gc_threads()) { 5611 workers()->run_task(&keep_cm_referents); 5612 } else { 5613 keep_cm_referents.work(0); 5614 } 5615 5616 set_par_threads(0); 5617 5618 // Closure to test whether a referent is alive. 5619 G1STWIsAliveClosure is_alive(this); 5620 5621 // Even when parallel reference processing is enabled, the processing 5622 // of JNI refs is serial and performed serially by the current thread 5623 // rather than by a worker. The following PSS will be used for processing 5624 // JNI refs. 5625 5626 // Use only a single queue for this PSS. 5627 G1ParScanThreadState pss(this, 0); 5628 5629 // We do not embed a reference processor in the copying/scanning 5630 // closures while we're actually processing the discovered 5631 // reference objects. 5632 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5633 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5634 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5635 5636 pss.set_evac_closure(&scan_evac_cl); 5637 pss.set_evac_failure_closure(&evac_failure_cl); 5638 pss.set_partial_scan_closure(&partial_scan_cl); 5639 5640 assert(pss.refs()->is_empty(), "pre-condition"); 5641 5642 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5643 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5644 5645 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5646 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5647 5648 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5649 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5650 5651 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5652 // We also need to mark copied objects. 5653 copy_non_heap_cl = ©_mark_non_heap_cl; 5654 copy_metadata_cl = ©_mark_metadata_cl; 5655 } 5656 5657 // Keep alive closure. 5658 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5659 5660 // Serial Complete GC closure 5661 G1STWDrainQueueClosure drain_queue(this, &pss); 5662 5663 // Setup the soft refs policy... 5664 rp->setup_policy(false); 5665 5666 ReferenceProcessorStats stats; 5667 if (!rp->processing_is_mt()) { 5668 // Serial reference processing... 5669 stats = rp->process_discovered_references(&is_alive, 5670 &keep_alive, 5671 &drain_queue, 5672 NULL, 5673 _gc_timer_stw); 5674 } else { 5675 // Parallel reference processing 5676 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5677 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5678 5679 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5680 stats = rp->process_discovered_references(&is_alive, 5681 &keep_alive, 5682 &drain_queue, 5683 &par_task_executor, 5684 _gc_timer_stw); 5685 } 5686 5687 _gc_tracer_stw->report_gc_reference_stats(stats); 5688 // We have completed copying any necessary live referent objects 5689 // (that were not copied during the actual pause) so we can 5690 // retire any active alloc buffers 5691 pss.retire_alloc_buffers(); 5692 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5693 5694 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5695 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5696 } 5697 5698 // Weak Reference processing during an evacuation pause (part 2). 5699 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5700 double ref_enq_start = os::elapsedTime(); 5701 5702 ReferenceProcessor* rp = _ref_processor_stw; 5703 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5704 5705 // Now enqueue any remaining on the discovered lists on to 5706 // the pending list. 5707 if (!rp->processing_is_mt()) { 5708 // Serial reference processing... 5709 rp->enqueue_discovered_references(); 5710 } else { 5711 // Parallel reference enqueueing 5712 5713 assert(no_of_gc_workers == workers()->active_workers(), 5714 "Need to reset active workers"); 5715 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5716 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5717 5718 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5719 rp->enqueue_discovered_references(&par_task_executor); 5720 } 5721 5722 rp->verify_no_references_recorded(); 5723 assert(!rp->discovery_enabled(), "should have been disabled"); 5724 5725 // FIXME 5726 // CM's reference processing also cleans up the string and symbol tables. 5727 // Should we do that here also? We could, but it is a serial operation 5728 // and could significantly increase the pause time. 5729 5730 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5731 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5732 } 5733 5734 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { 5735 _expand_heap_after_alloc_failure = true; 5736 _evacuation_failed = false; 5737 5738 // Should G1EvacuationFailureALot be in effect for this GC? 5739 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5740 5741 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5742 5743 // Disable the hot card cache. 5744 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 5745 hot_card_cache->reset_hot_cache_claimed_index(); 5746 hot_card_cache->set_use_cache(false); 5747 5748 uint n_workers; 5749 if (G1CollectedHeap::use_parallel_gc_threads()) { 5750 n_workers = 5751 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5752 workers()->active_workers(), 5753 Threads::number_of_non_daemon_threads()); 5754 assert(UseDynamicNumberOfGCThreads || 5755 n_workers == workers()->total_workers(), 5756 "If not dynamic should be using all the workers"); 5757 workers()->set_active_workers(n_workers); 5758 set_par_threads(n_workers); 5759 } else { 5760 assert(n_par_threads() == 0, 5761 "Should be the original non-parallel value"); 5762 n_workers = 1; 5763 } 5764 5765 G1ParTask g1_par_task(this, _task_queues); 5766 5767 init_for_evac_failure(NULL); 5768 5769 rem_set()->prepare_for_younger_refs_iterate(true); 5770 5771 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5772 double start_par_time_sec = os::elapsedTime(); 5773 double end_par_time_sec; 5774 5775 { 5776 StrongRootsScope srs(this); 5777 5778 if (G1CollectedHeap::use_parallel_gc_threads()) { 5779 // The individual threads will set their evac-failure closures. 5780 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5781 // These tasks use ShareHeap::_process_strong_tasks 5782 assert(UseDynamicNumberOfGCThreads || 5783 workers()->active_workers() == workers()->total_workers(), 5784 "If not dynamic should be using all the workers"); 5785 workers()->run_task(&g1_par_task); 5786 } else { 5787 g1_par_task.set_for_termination(n_workers); 5788 g1_par_task.work(0); 5789 } 5790 end_par_time_sec = os::elapsedTime(); 5791 5792 // Closing the inner scope will execute the destructor 5793 // for the StrongRootsScope object. We record the current 5794 // elapsed time before closing the scope so that time 5795 // taken for the SRS destructor is NOT included in the 5796 // reported parallel time. 5797 } 5798 5799 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5800 g1_policy()->phase_times()->record_par_time(par_time_ms); 5801 5802 double code_root_fixup_time_ms = 5803 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5804 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5805 5806 set_par_threads(0); 5807 5808 // Process any discovered reference objects - we have 5809 // to do this _before_ we retire the GC alloc regions 5810 // as we may have to copy some 'reachable' referent 5811 // objects (and their reachable sub-graphs) that were 5812 // not copied during the pause. 5813 process_discovered_references(n_workers); 5814 5815 // Weak root processing. 5816 { 5817 G1STWIsAliveClosure is_alive(this); 5818 G1KeepAliveClosure keep_alive(this); 5819 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5820 } 5821 5822 release_gc_alloc_regions(n_workers, evacuation_info); 5823 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5824 5825 // Reset and re-enable the hot card cache. 5826 // Note the counts for the cards in the regions in the 5827 // collection set are reset when the collection set is freed. 5828 hot_card_cache->reset_hot_cache(); 5829 hot_card_cache->set_use_cache(true); 5830 5831 // Migrate the strong code roots attached to each region in 5832 // the collection set. Ideally we would like to do this 5833 // after we have finished the scanning/evacuation of the 5834 // strong code roots for a particular heap region. 5835 migrate_strong_code_roots(); 5836 5837 if (g1_policy()->during_initial_mark_pause()) { 5838 // Reset the claim values set during marking the strong code roots 5839 reset_heap_region_claim_values(); 5840 } 5841 5842 finalize_for_evac_failure(); 5843 5844 if (evacuation_failed()) { 5845 remove_self_forwarding_pointers(); 5846 5847 // Reset the G1EvacuationFailureALot counters and flags 5848 // Note: the values are reset only when an actual 5849 // evacuation failure occurs. 5850 NOT_PRODUCT(reset_evacuation_should_fail();) 5851 } 5852 5853 // Enqueue any remaining references remaining on the STW 5854 // reference processor's discovered lists. We need to do 5855 // this after the card table is cleaned (and verified) as 5856 // the act of enqueueing entries on to the pending list 5857 // will log these updates (and dirty their associated 5858 // cards). We need these updates logged to update any 5859 // RSets. 5860 enqueue_discovered_references(n_workers); 5861 5862 if (G1DeferredRSUpdate) { 5863 RedirtyLoggedCardTableEntryFastClosure redirty; 5864 dirty_card_queue_set().set_closure(&redirty); 5865 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5866 5867 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5868 dcq.merge_bufferlists(&dirty_card_queue_set()); 5869 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5870 } 5871 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5872 } 5873 5874 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5875 size_t* pre_used, 5876 FreeRegionList* free_list, 5877 OldRegionSet* old_proxy_set, 5878 HumongousRegionSet* humongous_proxy_set, 5879 HRRSCleanupTask* hrrs_cleanup_task, 5880 bool par) { 5881 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5882 if (hr->isHumongous()) { 5883 assert(hr->startsHumongous(), "we should only see starts humongous"); 5884 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5885 } else { 5886 _old_set.remove_with_proxy(hr, old_proxy_set); 5887 free_region(hr, pre_used, free_list, par); 5888 } 5889 } else { 5890 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5891 } 5892 } 5893 5894 void G1CollectedHeap::free_region(HeapRegion* hr, 5895 size_t* pre_used, 5896 FreeRegionList* free_list, 5897 bool par) { 5898 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5899 assert(!hr->is_empty(), "the region should not be empty"); 5900 assert(free_list != NULL, "pre-condition"); 5901 5902 // Clear the card counts for this region. 5903 // Note: we only need to do this if the region is not young 5904 // (since we don't refine cards in young regions). 5905 if (!hr->is_young()) { 5906 _cg1r->hot_card_cache()->reset_card_counts(hr); 5907 } 5908 *pre_used += hr->used(); 5909 hr->hr_clear(par, true /* clear_space */); 5910 free_list->add_as_head(hr); 5911 } 5912 5913 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5914 size_t* pre_used, 5915 FreeRegionList* free_list, 5916 HumongousRegionSet* humongous_proxy_set, 5917 bool par) { 5918 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5919 assert(free_list != NULL, "pre-condition"); 5920 assert(humongous_proxy_set != NULL, "pre-condition"); 5921 5922 size_t hr_used = hr->used(); 5923 size_t hr_capacity = hr->capacity(); 5924 size_t hr_pre_used = 0; 5925 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5926 // We need to read this before we make the region non-humongous, 5927 // otherwise the information will be gone. 5928 uint last_index = hr->last_hc_index(); 5929 hr->set_notHumongous(); 5930 free_region(hr, &hr_pre_used, free_list, par); 5931 5932 uint i = hr->hrs_index() + 1; 5933 while (i < last_index) { 5934 HeapRegion* curr_hr = region_at(i); 5935 assert(curr_hr->continuesHumongous(), "invariant"); 5936 curr_hr->set_notHumongous(); 5937 free_region(curr_hr, &hr_pre_used, free_list, par); 5938 i += 1; 5939 } 5940 assert(hr_pre_used == hr_used, 5941 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5942 "should be the same", hr_pre_used, hr_used)); 5943 *pre_used += hr_pre_used; 5944 } 5945 5946 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5947 FreeRegionList* free_list, 5948 OldRegionSet* old_proxy_set, 5949 HumongousRegionSet* humongous_proxy_set, 5950 bool par) { 5951 if (pre_used > 0) { 5952 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5953 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5954 assert(_summary_bytes_used >= pre_used, 5955 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5956 "should be >= pre_used: "SIZE_FORMAT, 5957 _summary_bytes_used, pre_used)); 5958 _summary_bytes_used -= pre_used; 5959 } 5960 if (free_list != NULL && !free_list->is_empty()) { 5961 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5962 _free_list.add_as_head(free_list); 5963 } 5964 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5965 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5966 _old_set.update_from_proxy(old_proxy_set); 5967 } 5968 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5969 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5970 _humongous_set.update_from_proxy(humongous_proxy_set); 5971 } 5972 } 5973 5974 class G1ParCleanupCTTask : public AbstractGangTask { 5975 CardTableModRefBS* _ct_bs; 5976 G1CollectedHeap* _g1h; 5977 HeapRegion* volatile _su_head; 5978 public: 5979 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5980 G1CollectedHeap* g1h) : 5981 AbstractGangTask("G1 Par Cleanup CT Task"), 5982 _ct_bs(ct_bs), _g1h(g1h) { } 5983 5984 void work(uint worker_id) { 5985 HeapRegion* r; 5986 while (r = _g1h->pop_dirty_cards_region()) { 5987 clear_cards(r); 5988 } 5989 } 5990 5991 void clear_cards(HeapRegion* r) { 5992 // Cards of the survivors should have already been dirtied. 5993 if (!r->is_survivor()) { 5994 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5995 } 5996 } 5997 }; 5998 5999 #ifndef PRODUCT 6000 class G1VerifyCardTableCleanup: public HeapRegionClosure { 6001 G1CollectedHeap* _g1h; 6002 CardTableModRefBS* _ct_bs; 6003 public: 6004 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 6005 : _g1h(g1h), _ct_bs(ct_bs) { } 6006 virtual bool doHeapRegion(HeapRegion* r) { 6007 if (r->is_survivor()) { 6008 _g1h->verify_dirty_region(r); 6009 } else { 6010 _g1h->verify_not_dirty_region(r); 6011 } 6012 return false; 6013 } 6014 }; 6015 6016 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 6017 // All of the region should be clean. 6018 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 6019 MemRegion mr(hr->bottom(), hr->end()); 6020 ct_bs->verify_not_dirty_region(mr); 6021 } 6022 6023 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 6024 // We cannot guarantee that [bottom(),end()] is dirty. Threads 6025 // dirty allocated blocks as they allocate them. The thread that 6026 // retires each region and replaces it with a new one will do a 6027 // maximal allocation to fill in [pre_dummy_top(),end()] but will 6028 // not dirty that area (one less thing to have to do while holding 6029 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 6030 // is dirty. 6031 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6032 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 6033 ct_bs->verify_dirty_region(mr); 6034 } 6035 6036 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 6037 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6038 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 6039 verify_dirty_region(hr); 6040 } 6041 } 6042 6043 void G1CollectedHeap::verify_dirty_young_regions() { 6044 verify_dirty_young_list(_young_list->first_region()); 6045 } 6046 #endif 6047 6048 void G1CollectedHeap::cleanUpCardTable() { 6049 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 6050 double start = os::elapsedTime(); 6051 6052 { 6053 // Iterate over the dirty cards region list. 6054 G1ParCleanupCTTask cleanup_task(ct_bs, this); 6055 6056 if (G1CollectedHeap::use_parallel_gc_threads()) { 6057 set_par_threads(); 6058 workers()->run_task(&cleanup_task); 6059 set_par_threads(0); 6060 } else { 6061 while (_dirty_cards_region_list) { 6062 HeapRegion* r = _dirty_cards_region_list; 6063 cleanup_task.clear_cards(r); 6064 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 6065 if (_dirty_cards_region_list == r) { 6066 // The last region. 6067 _dirty_cards_region_list = NULL; 6068 } 6069 r->set_next_dirty_cards_region(NULL); 6070 } 6071 } 6072 #ifndef PRODUCT 6073 if (G1VerifyCTCleanup || VerifyAfterGC) { 6074 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 6075 heap_region_iterate(&cleanup_verifier); 6076 } 6077 #endif 6078 } 6079 6080 double elapsed = os::elapsedTime() - start; 6081 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 6082 } 6083 6084 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) { 6085 size_t pre_used = 0; 6086 FreeRegionList local_free_list("Local List for CSet Freeing"); 6087 6088 double young_time_ms = 0.0; 6089 double non_young_time_ms = 0.0; 6090 6091 // Since the collection set is a superset of the the young list, 6092 // all we need to do to clear the young list is clear its 6093 // head and length, and unlink any young regions in the code below 6094 _young_list->clear(); 6095 6096 G1CollectorPolicy* policy = g1_policy(); 6097 6098 double start_sec = os::elapsedTime(); 6099 bool non_young = true; 6100 6101 HeapRegion* cur = cs_head; 6102 int age_bound = -1; 6103 size_t rs_lengths = 0; 6104 6105 while (cur != NULL) { 6106 assert(!is_on_master_free_list(cur), "sanity"); 6107 if (non_young) { 6108 if (cur->is_young()) { 6109 double end_sec = os::elapsedTime(); 6110 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6111 non_young_time_ms += elapsed_ms; 6112 6113 start_sec = os::elapsedTime(); 6114 non_young = false; 6115 } 6116 } else { 6117 if (!cur->is_young()) { 6118 double end_sec = os::elapsedTime(); 6119 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6120 young_time_ms += elapsed_ms; 6121 6122 start_sec = os::elapsedTime(); 6123 non_young = true; 6124 } 6125 } 6126 6127 rs_lengths += cur->rem_set()->occupied(); 6128 6129 HeapRegion* next = cur->next_in_collection_set(); 6130 assert(cur->in_collection_set(), "bad CS"); 6131 cur->set_next_in_collection_set(NULL); 6132 cur->set_in_collection_set(false); 6133 6134 if (cur->is_young()) { 6135 int index = cur->young_index_in_cset(); 6136 assert(index != -1, "invariant"); 6137 assert((uint) index < policy->young_cset_region_length(), "invariant"); 6138 size_t words_survived = _surviving_young_words[index]; 6139 cur->record_surv_words_in_group(words_survived); 6140 6141 // At this point the we have 'popped' cur from the collection set 6142 // (linked via next_in_collection_set()) but it is still in the 6143 // young list (linked via next_young_region()). Clear the 6144 // _next_young_region field. 6145 cur->set_next_young_region(NULL); 6146 } else { 6147 int index = cur->young_index_in_cset(); 6148 assert(index == -1, "invariant"); 6149 } 6150 6151 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 6152 (!cur->is_young() && cur->young_index_in_cset() == -1), 6153 "invariant" ); 6154 6155 if (!cur->evacuation_failed()) { 6156 MemRegion used_mr = cur->used_region(); 6157 6158 // And the region is empty. 6159 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 6160 free_region(cur, &pre_used, &local_free_list, false /* par */); 6161 } else { 6162 cur->uninstall_surv_rate_group(); 6163 if (cur->is_young()) { 6164 cur->set_young_index_in_cset(-1); 6165 } 6166 cur->set_not_young(); 6167 cur->set_evacuation_failed(false); 6168 // The region is now considered to be old. 6169 _old_set.add(cur); 6170 evacuation_info.increment_collectionset_used_after(cur->used()); 6171 } 6172 cur = next; 6173 } 6174 6175 evacuation_info.set_regions_freed(local_free_list.length()); 6176 policy->record_max_rs_lengths(rs_lengths); 6177 policy->cset_regions_freed(); 6178 6179 double end_sec = os::elapsedTime(); 6180 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6181 6182 if (non_young) { 6183 non_young_time_ms += elapsed_ms; 6184 } else { 6185 young_time_ms += elapsed_ms; 6186 } 6187 6188 update_sets_after_freeing_regions(pre_used, &local_free_list, 6189 NULL /* old_proxy_set */, 6190 NULL /* humongous_proxy_set */, 6191 false /* par */); 6192 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6193 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6194 } 6195 6196 // This routine is similar to the above but does not record 6197 // any policy statistics or update free lists; we are abandoning 6198 // the current incremental collection set in preparation of a 6199 // full collection. After the full GC we will start to build up 6200 // the incremental collection set again. 6201 // This is only called when we're doing a full collection 6202 // and is immediately followed by the tearing down of the young list. 6203 6204 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6205 HeapRegion* cur = cs_head; 6206 6207 while (cur != NULL) { 6208 HeapRegion* next = cur->next_in_collection_set(); 6209 assert(cur->in_collection_set(), "bad CS"); 6210 cur->set_next_in_collection_set(NULL); 6211 cur->set_in_collection_set(false); 6212 cur->set_young_index_in_cset(-1); 6213 cur = next; 6214 } 6215 } 6216 6217 void G1CollectedHeap::set_free_regions_coming() { 6218 if (G1ConcRegionFreeingVerbose) { 6219 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6220 "setting free regions coming"); 6221 } 6222 6223 assert(!free_regions_coming(), "pre-condition"); 6224 _free_regions_coming = true; 6225 } 6226 6227 void G1CollectedHeap::reset_free_regions_coming() { 6228 assert(free_regions_coming(), "pre-condition"); 6229 6230 { 6231 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6232 _free_regions_coming = false; 6233 SecondaryFreeList_lock->notify_all(); 6234 } 6235 6236 if (G1ConcRegionFreeingVerbose) { 6237 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6238 "reset free regions coming"); 6239 } 6240 } 6241 6242 void G1CollectedHeap::wait_while_free_regions_coming() { 6243 // Most of the time we won't have to wait, so let's do a quick test 6244 // first before we take the lock. 6245 if (!free_regions_coming()) { 6246 return; 6247 } 6248 6249 if (G1ConcRegionFreeingVerbose) { 6250 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6251 "waiting for free regions"); 6252 } 6253 6254 { 6255 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6256 while (free_regions_coming()) { 6257 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6258 } 6259 } 6260 6261 if (G1ConcRegionFreeingVerbose) { 6262 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6263 "done waiting for free regions"); 6264 } 6265 } 6266 6267 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6268 assert(heap_lock_held_for_gc(), 6269 "the heap lock should already be held by or for this thread"); 6270 _young_list->push_region(hr); 6271 } 6272 6273 class NoYoungRegionsClosure: public HeapRegionClosure { 6274 private: 6275 bool _success; 6276 public: 6277 NoYoungRegionsClosure() : _success(true) { } 6278 bool doHeapRegion(HeapRegion* r) { 6279 if (r->is_young()) { 6280 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6281 r->bottom(), r->end()); 6282 _success = false; 6283 } 6284 return false; 6285 } 6286 bool success() { return _success; } 6287 }; 6288 6289 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6290 bool ret = _young_list->check_list_empty(check_sample); 6291 6292 if (check_heap) { 6293 NoYoungRegionsClosure closure; 6294 heap_region_iterate(&closure); 6295 ret = ret && closure.success(); 6296 } 6297 6298 return ret; 6299 } 6300 6301 class TearDownRegionSetsClosure : public HeapRegionClosure { 6302 private: 6303 OldRegionSet *_old_set; 6304 6305 public: 6306 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6307 6308 bool doHeapRegion(HeapRegion* r) { 6309 if (r->is_empty()) { 6310 // We ignore empty regions, we'll empty the free list afterwards 6311 } else if (r->is_young()) { 6312 // We ignore young regions, we'll empty the young list afterwards 6313 } else if (r->isHumongous()) { 6314 // We ignore humongous regions, we're not tearing down the 6315 // humongous region set 6316 } else { 6317 // The rest should be old 6318 _old_set->remove(r); 6319 } 6320 return false; 6321 } 6322 6323 ~TearDownRegionSetsClosure() { 6324 assert(_old_set->is_empty(), "post-condition"); 6325 } 6326 }; 6327 6328 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6329 assert_at_safepoint(true /* should_be_vm_thread */); 6330 6331 if (!free_list_only) { 6332 TearDownRegionSetsClosure cl(&_old_set); 6333 heap_region_iterate(&cl); 6334 6335 // Need to do this after the heap iteration to be able to 6336 // recognize the young regions and ignore them during the iteration. 6337 _young_list->empty_list(); 6338 } 6339 _free_list.remove_all(); 6340 } 6341 6342 class RebuildRegionSetsClosure : public HeapRegionClosure { 6343 private: 6344 bool _free_list_only; 6345 OldRegionSet* _old_set; 6346 FreeRegionList* _free_list; 6347 size_t _total_used; 6348 6349 public: 6350 RebuildRegionSetsClosure(bool free_list_only, 6351 OldRegionSet* old_set, FreeRegionList* free_list) : 6352 _free_list_only(free_list_only), 6353 _old_set(old_set), _free_list(free_list), _total_used(0) { 6354 assert(_free_list->is_empty(), "pre-condition"); 6355 if (!free_list_only) { 6356 assert(_old_set->is_empty(), "pre-condition"); 6357 } 6358 } 6359 6360 bool doHeapRegion(HeapRegion* r) { 6361 if (r->continuesHumongous()) { 6362 return false; 6363 } 6364 6365 if (r->is_empty()) { 6366 // Add free regions to the free list 6367 _free_list->add_as_tail(r); 6368 } else if (!_free_list_only) { 6369 assert(!r->is_young(), "we should not come across young regions"); 6370 6371 if (r->isHumongous()) { 6372 // We ignore humongous regions, we left the humongous set unchanged 6373 } else { 6374 // The rest should be old, add them to the old set 6375 _old_set->add(r); 6376 } 6377 _total_used += r->used(); 6378 } 6379 6380 return false; 6381 } 6382 6383 size_t total_used() { 6384 return _total_used; 6385 } 6386 }; 6387 6388 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6389 assert_at_safepoint(true /* should_be_vm_thread */); 6390 6391 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6392 heap_region_iterate(&cl); 6393 6394 if (!free_list_only) { 6395 _summary_bytes_used = cl.total_used(); 6396 } 6397 assert(_summary_bytes_used == recalculate_used(), 6398 err_msg("inconsistent _summary_bytes_used, " 6399 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6400 _summary_bytes_used, recalculate_used())); 6401 } 6402 6403 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6404 _refine_cte_cl->set_concurrent(concurrent); 6405 } 6406 6407 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6408 HeapRegion* hr = heap_region_containing(p); 6409 if (hr == NULL) { 6410 return false; 6411 } else { 6412 return hr->is_in(p); 6413 } 6414 } 6415 6416 // Methods for the mutator alloc region 6417 6418 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6419 bool force) { 6420 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6421 assert(!force || g1_policy()->can_expand_young_list(), 6422 "if force is true we should be able to expand the young list"); 6423 bool young_list_full = g1_policy()->is_young_list_full(); 6424 if (force || !young_list_full) { 6425 HeapRegion* new_alloc_region = new_region(word_size, 6426 false /* do_expand */); 6427 if (new_alloc_region != NULL) { 6428 set_region_short_lived_locked(new_alloc_region); 6429 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6430 return new_alloc_region; 6431 } 6432 } 6433 return NULL; 6434 } 6435 6436 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6437 size_t allocated_bytes) { 6438 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6439 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6440 6441 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6442 _summary_bytes_used += allocated_bytes; 6443 _hr_printer.retire(alloc_region); 6444 // We update the eden sizes here, when the region is retired, 6445 // instead of when it's allocated, since this is the point that its 6446 // used space has been recored in _summary_bytes_used. 6447 g1mm()->update_eden_size(); 6448 } 6449 6450 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6451 bool force) { 6452 return _g1h->new_mutator_alloc_region(word_size, force); 6453 } 6454 6455 void G1CollectedHeap::set_par_threads() { 6456 // Don't change the number of workers. Use the value previously set 6457 // in the workgroup. 6458 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6459 uint n_workers = workers()->active_workers(); 6460 assert(UseDynamicNumberOfGCThreads || 6461 n_workers == workers()->total_workers(), 6462 "Otherwise should be using the total number of workers"); 6463 if (n_workers == 0) { 6464 assert(false, "Should have been set in prior evacuation pause."); 6465 n_workers = ParallelGCThreads; 6466 workers()->set_active_workers(n_workers); 6467 } 6468 set_par_threads(n_workers); 6469 } 6470 6471 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6472 size_t allocated_bytes) { 6473 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6474 } 6475 6476 // Methods for the GC alloc regions 6477 6478 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6479 uint count, 6480 GCAllocPurpose ap) { 6481 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6482 6483 if (count < g1_policy()->max_regions(ap)) { 6484 HeapRegion* new_alloc_region = new_region(word_size, 6485 true /* do_expand */); 6486 if (new_alloc_region != NULL) { 6487 // We really only need to do this for old regions given that we 6488 // should never scan survivors. But it doesn't hurt to do it 6489 // for survivors too. 6490 new_alloc_region->set_saved_mark(); 6491 if (ap == GCAllocForSurvived) { 6492 new_alloc_region->set_survivor(); 6493 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6494 } else { 6495 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6496 } 6497 bool during_im = g1_policy()->during_initial_mark_pause(); 6498 new_alloc_region->note_start_of_copying(during_im); 6499 return new_alloc_region; 6500 } else { 6501 g1_policy()->note_alloc_region_limit_reached(ap); 6502 } 6503 } 6504 return NULL; 6505 } 6506 6507 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6508 size_t allocated_bytes, 6509 GCAllocPurpose ap) { 6510 bool during_im = g1_policy()->during_initial_mark_pause(); 6511 alloc_region->note_end_of_copying(during_im); 6512 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6513 if (ap == GCAllocForSurvived) { 6514 young_list()->add_survivor_region(alloc_region); 6515 } else { 6516 _old_set.add(alloc_region); 6517 } 6518 _hr_printer.retire(alloc_region); 6519 } 6520 6521 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6522 bool force) { 6523 assert(!force, "not supported for GC alloc regions"); 6524 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6525 } 6526 6527 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6528 size_t allocated_bytes) { 6529 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6530 GCAllocForSurvived); 6531 } 6532 6533 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6534 bool force) { 6535 assert(!force, "not supported for GC alloc regions"); 6536 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6537 } 6538 6539 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6540 size_t allocated_bytes) { 6541 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6542 GCAllocForTenured); 6543 } 6544 // Heap region set verification 6545 6546 class VerifyRegionListsClosure : public HeapRegionClosure { 6547 private: 6548 FreeRegionList* _free_list; 6549 OldRegionSet* _old_set; 6550 HumongousRegionSet* _humongous_set; 6551 uint _region_count; 6552 6553 public: 6554 VerifyRegionListsClosure(OldRegionSet* old_set, 6555 HumongousRegionSet* humongous_set, 6556 FreeRegionList* free_list) : 6557 _old_set(old_set), _humongous_set(humongous_set), 6558 _free_list(free_list), _region_count(0) { } 6559 6560 uint region_count() { return _region_count; } 6561 6562 bool doHeapRegion(HeapRegion* hr) { 6563 _region_count += 1; 6564 6565 if (hr->continuesHumongous()) { 6566 return false; 6567 } 6568 6569 if (hr->is_young()) { 6570 // TODO 6571 } else if (hr->startsHumongous()) { 6572 _humongous_set->verify_next_region(hr); 6573 } else if (hr->is_empty()) { 6574 _free_list->verify_next_region(hr); 6575 } else { 6576 _old_set->verify_next_region(hr); 6577 } 6578 return false; 6579 } 6580 }; 6581 6582 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6583 HeapWord* bottom) { 6584 HeapWord* end = bottom + HeapRegion::GrainWords; 6585 MemRegion mr(bottom, end); 6586 assert(_g1_reserved.contains(mr), "invariant"); 6587 // This might return NULL if the allocation fails 6588 return new HeapRegion(hrs_index, _bot_shared, mr); 6589 } 6590 6591 void G1CollectedHeap::verify_region_sets() { 6592 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6593 6594 // First, check the explicit lists. 6595 _free_list.verify(); 6596 { 6597 // Given that a concurrent operation might be adding regions to 6598 // the secondary free list we have to take the lock before 6599 // verifying it. 6600 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6601 _secondary_free_list.verify(); 6602 } 6603 _old_set.verify(); 6604 _humongous_set.verify(); 6605 6606 // If a concurrent region freeing operation is in progress it will 6607 // be difficult to correctly attributed any free regions we come 6608 // across to the correct free list given that they might belong to 6609 // one of several (free_list, secondary_free_list, any local lists, 6610 // etc.). So, if that's the case we will skip the rest of the 6611 // verification operation. Alternatively, waiting for the concurrent 6612 // operation to complete will have a non-trivial effect on the GC's 6613 // operation (no concurrent operation will last longer than the 6614 // interval between two calls to verification) and it might hide 6615 // any issues that we would like to catch during testing. 6616 if (free_regions_coming()) { 6617 return; 6618 } 6619 6620 // Make sure we append the secondary_free_list on the free_list so 6621 // that all free regions we will come across can be safely 6622 // attributed to the free_list. 6623 append_secondary_free_list_if_not_empty_with_lock(); 6624 6625 // Finally, make sure that the region accounting in the lists is 6626 // consistent with what we see in the heap. 6627 _old_set.verify_start(); 6628 _humongous_set.verify_start(); 6629 _free_list.verify_start(); 6630 6631 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6632 heap_region_iterate(&cl); 6633 6634 _old_set.verify_end(); 6635 _humongous_set.verify_end(); 6636 _free_list.verify_end(); 6637 } 6638 6639 // Optimized nmethod scanning 6640 6641 class RegisterNMethodOopClosure: public OopClosure { 6642 G1CollectedHeap* _g1h; 6643 nmethod* _nm; 6644 6645 template <class T> void do_oop_work(T* p) { 6646 T heap_oop = oopDesc::load_heap_oop(p); 6647 if (!oopDesc::is_null(heap_oop)) { 6648 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6649 HeapRegion* hr = _g1h->heap_region_containing(obj); 6650 assert(!hr->isHumongous(), "code root in humongous region?"); 6651 6652 // HeapRegion::add_strong_code_root() avoids adding duplicate 6653 // entries but having duplicates is OK since we "mark" nmethods 6654 // as visited when we scan the strong code root lists during the GC. 6655 hr->add_strong_code_root(_nm); 6656 assert(hr->rem_set()->strong_code_roots_list_contains(_nm), "add failed?"); 6657 } 6658 } 6659 6660 public: 6661 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6662 _g1h(g1h), _nm(nm) {} 6663 6664 void do_oop(oop* p) { do_oop_work(p); } 6665 void do_oop(narrowOop* p) { do_oop_work(p); } 6666 }; 6667 6668 class UnregisterNMethodOopClosure: public OopClosure { 6669 G1CollectedHeap* _g1h; 6670 nmethod* _nm; 6671 6672 template <class T> void do_oop_work(T* p) { 6673 T heap_oop = oopDesc::load_heap_oop(p); 6674 if (!oopDesc::is_null(heap_oop)) { 6675 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6676 HeapRegion* hr = _g1h->heap_region_containing(obj); 6677 assert(!hr->isHumongous(), "code root in humongous region?"); 6678 hr->remove_strong_code_root(_nm); 6679 assert(!hr->rem_set()->strong_code_roots_list_contains(_nm), "remove failed?"); 6680 } 6681 } 6682 6683 public: 6684 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6685 _g1h(g1h), _nm(nm) {} 6686 6687 void do_oop(oop* p) { do_oop_work(p); } 6688 void do_oop(narrowOop* p) { do_oop_work(p); } 6689 }; 6690 6691 void G1CollectedHeap::register_nmethod(nmethod* nm) { 6692 CollectedHeap::register_nmethod(nm); 6693 6694 guarantee(nm != NULL, "sanity"); 6695 RegisterNMethodOopClosure reg_cl(this, nm); 6696 nm->oops_do(®_cl); 6697 } 6698 6699 void G1CollectedHeap::unregister_nmethod(nmethod* nm) { 6700 CollectedHeap::unregister_nmethod(nm); 6701 6702 guarantee(nm != NULL, "sanity"); 6703 UnregisterNMethodOopClosure reg_cl(this, nm); 6704 nm->oops_do(®_cl, true); 6705 } 6706 6707 class MigrateCodeRootsHeapRegionClosure: public HeapRegionClosure { 6708 public: 6709 bool doHeapRegion(HeapRegion *hr) { 6710 assert(!hr->isHumongous(), "humongous region in collection set?"); 6711 hr->migrate_strong_code_roots(); 6712 return false; 6713 } 6714 }; 6715 6716 void G1CollectedHeap::migrate_strong_code_roots() { 6717 MigrateCodeRootsHeapRegionClosure cl; 6718 double migrate_start = os::elapsedTime(); 6719 collection_set_iterate(&cl); 6720 double migration_time_ms = (os::elapsedTime() - migrate_start) * 1000.0; 6721 g1_policy()->phase_times()->record_strong_code_root_migration_time(migration_time_ms); 6722 } 6723 6724 // Mark all the code roots that point into regions *not* in the 6725 // collection set. 6726 // 6727 // Note we do not want to use a "marking" CodeBlobToOopClosure while 6728 // walking the the code roots lists of regions not in the collection 6729 // set. Suppose we have an nmethod (M) that points to objects in two 6730 // separate regions - one in the collection set (R1) and one not (R2). 6731 // Using a "marking" CodeBlobToOopClosure here would result in "marking" 6732 // nmethod M when walking the code roots for R1. When we come to scan 6733 // the code roots for R2, we would see that M is already marked and it 6734 // would be skipped and the objects in R2 that are referenced from M 6735 // would not be evacuated. 6736 6737 class MarkStrongCodeRootCodeBlobClosure: public CodeBlobClosure { 6738 6739 class MarkStrongCodeRootOopClosure: public OopClosure { 6740 ConcurrentMark* _cm; 6741 HeapRegion* _hr; 6742 uint _worker_id; 6743 6744 template <class T> void do_oop_work(T* p) { 6745 T heap_oop = oopDesc::load_heap_oop(p); 6746 if (!oopDesc::is_null(heap_oop)) { 6747 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6748 // Only mark objects in the region (which is assumed 6749 // to be not in the collection set). 6750 if (_hr->is_in(obj)) { 6751 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 6752 } 6753 } 6754 } 6755 6756 public: 6757 MarkStrongCodeRootOopClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id) : 6758 _cm(cm), _hr(hr), _worker_id(worker_id) { 6759 assert(!_hr->in_collection_set(), "sanity"); 6760 } 6761 6762 void do_oop(narrowOop* p) { do_oop_work(p); } 6763 void do_oop(oop* p) { do_oop_work(p); } 6764 }; 6765 6766 MarkStrongCodeRootOopClosure _oop_cl; 6767 6768 public: 6769 MarkStrongCodeRootCodeBlobClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id): 6770 _oop_cl(cm, hr, worker_id) {} 6771 6772 void do_code_blob(CodeBlob* cb) { 6773 nmethod* nm = (cb == NULL) ? NULL : cb->as_nmethod_or_null(); 6774 if (nm != NULL) { 6775 nm->oops_do(&_oop_cl); 6776 } 6777 } 6778 }; 6779 6780 class MarkStrongCodeRootsHRClosure: public HeapRegionClosure { 6781 G1CollectedHeap* _g1h; 6782 uint _worker_id; 6783 6784 public: 6785 MarkStrongCodeRootsHRClosure(G1CollectedHeap* g1h, uint worker_id) : 6786 _g1h(g1h), _worker_id(worker_id) {} 6787 6788 bool doHeapRegion(HeapRegion *hr) { 6789 HeapRegionRemSet* hrrs = hr->rem_set(); 6790 if (hr->isHumongous()) { 6791 // Code roots should never be attached to a humongous region 6792 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 6793 return false; 6794 } 6795 6796 if (hr->in_collection_set()) { 6797 // Don't mark code roots into regions in the collection set here. 6798 // They will be marked when we scan them. 6799 return false; 6800 } 6801 6802 MarkStrongCodeRootCodeBlobClosure cb_cl(_g1h->concurrent_mark(), hr, _worker_id); 6803 hr->strong_code_roots_do(&cb_cl); 6804 return false; 6805 } 6806 }; 6807 6808 void G1CollectedHeap::mark_strong_code_roots(uint worker_id) { 6809 MarkStrongCodeRootsHRClosure cl(this, worker_id); 6810 if (G1CollectedHeap::use_parallel_gc_threads()) { 6811 heap_region_par_iterate_chunked(&cl, 6812 worker_id, 6813 workers()->active_workers(), 6814 HeapRegion::ParMarkRootClaimValue); 6815 } else { 6816 heap_region_iterate(&cl); 6817 } 6818 } 6819 6820 class RebuildStrongCodeRootClosure: public CodeBlobClosure { 6821 G1CollectedHeap* _g1h; 6822 6823 public: 6824 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) : 6825 _g1h(g1h) {} 6826 6827 void do_code_blob(CodeBlob* cb) { 6828 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL; 6829 if (nm == NULL) { 6830 return; 6831 } 6832 6833 if (ScavengeRootsInCode && nm->detect_scavenge_root_oops()) { 6834 _g1h->register_nmethod(nm); 6835 } 6836 } 6837 }; 6838 6839 void G1CollectedHeap::rebuild_strong_code_roots() { 6840 RebuildStrongCodeRootClosure blob_cl(this); 6841 CodeCache::blobs_do(&blob_cl); 6842 }