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 size_t heap_alignment = collector_policy()->max_alignment(); 2010 2011 // Ensure that the sizes are properly aligned. 2012 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2013 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2014 Universe::check_alignment(max_byte_size, heap_alignment, "g1 heap"); 2015 2016 _cg1r = new ConcurrentG1Refine(this); 2017 2018 // Reserve the maximum. 2019 2020 // When compressed oops are enabled, the preferred heap base 2021 // is calculated by subtracting the requested size from the 2022 // 32Gb boundary and using the result as the base address for 2023 // heap reservation. If the requested size is not aligned to 2024 // HeapRegion::GrainBytes (i.e. the alignment that is passed 2025 // into the ReservedHeapSpace constructor) then the actual 2026 // base of the reserved heap may end up differing from the 2027 // address that was requested (i.e. the preferred heap base). 2028 // If this happens then we could end up using a non-optimal 2029 // compressed oops mode. 2030 2031 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2032 heap_alignment); 2033 2034 // It is important to do this in a way such that concurrent readers can't 2035 // temporarily think something is in the heap. (I've actually seen this 2036 // happen in asserts: DLD.) 2037 _reserved.set_word_size(0); 2038 _reserved.set_start((HeapWord*)heap_rs.base()); 2039 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2040 2041 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2042 2043 // Create the gen rem set (and barrier set) for the entire reserved region. 2044 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2045 set_barrier_set(rem_set()->bs()); 2046 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2047 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2048 } else { 2049 vm_exit_during_initialization("G1 requires a mod ref bs."); 2050 return JNI_ENOMEM; 2051 } 2052 2053 // Also create a G1 rem set. 2054 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2055 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2056 } else { 2057 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2058 return JNI_ENOMEM; 2059 } 2060 2061 // Carve out the G1 part of the heap. 2062 2063 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2064 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2065 g1_rs.size()/HeapWordSize); 2066 2067 _g1_storage.initialize(g1_rs, 0); 2068 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2069 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2070 (HeapWord*) _g1_reserved.end(), 2071 _expansion_regions); 2072 2073 // Do later initialization work for concurrent refinement. 2074 _cg1r->init(); 2075 2076 // 6843694 - ensure that the maximum region index can fit 2077 // in the remembered set structures. 2078 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2079 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2080 2081 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2082 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2083 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2084 "too many cards per region"); 2085 2086 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2087 2088 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2089 heap_word_size(init_byte_size)); 2090 2091 _g1h = this; 2092 2093 _in_cset_fast_test_length = max_regions(); 2094 _in_cset_fast_test_base = 2095 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2096 2097 // We're biasing _in_cset_fast_test to avoid subtracting the 2098 // beginning of the heap every time we want to index; basically 2099 // it's the same with what we do with the card table. 2100 _in_cset_fast_test = _in_cset_fast_test_base - 2101 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2102 2103 // Clear the _cset_fast_test bitmap in anticipation of adding 2104 // regions to the incremental collection set for the first 2105 // evacuation pause. 2106 clear_cset_fast_test(); 2107 2108 // Create the ConcurrentMark data structure and thread. 2109 // (Must do this late, so that "max_regions" is defined.) 2110 _cm = new ConcurrentMark(this, heap_rs); 2111 if (_cm == NULL || !_cm->completed_initialization()) { 2112 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2113 return JNI_ENOMEM; 2114 } 2115 _cmThread = _cm->cmThread(); 2116 2117 // Initialize the from_card cache structure of HeapRegionRemSet. 2118 HeapRegionRemSet::init_heap(max_regions()); 2119 2120 // Now expand into the initial heap size. 2121 if (!expand(init_byte_size)) { 2122 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2123 return JNI_ENOMEM; 2124 } 2125 2126 // Perform any initialization actions delegated to the policy. 2127 g1_policy()->init(); 2128 2129 _refine_cte_cl = 2130 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2131 g1_rem_set(), 2132 concurrent_g1_refine()); 2133 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2134 2135 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2136 SATB_Q_FL_lock, 2137 G1SATBProcessCompletedThreshold, 2138 Shared_SATB_Q_lock); 2139 2140 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2141 DirtyCardQ_FL_lock, 2142 concurrent_g1_refine()->yellow_zone(), 2143 concurrent_g1_refine()->red_zone(), 2144 Shared_DirtyCardQ_lock); 2145 2146 if (G1DeferredRSUpdate) { 2147 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2148 DirtyCardQ_FL_lock, 2149 -1, // never trigger processing 2150 -1, // no limit on length 2151 Shared_DirtyCardQ_lock, 2152 &JavaThread::dirty_card_queue_set()); 2153 } 2154 2155 // Initialize the card queue set used to hold cards containing 2156 // references into the collection set. 2157 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2158 DirtyCardQ_FL_lock, 2159 -1, // never trigger processing 2160 -1, // no limit on length 2161 Shared_DirtyCardQ_lock, 2162 &JavaThread::dirty_card_queue_set()); 2163 2164 // In case we're keeping closure specialization stats, initialize those 2165 // counts and that mechanism. 2166 SpecializationStats::clear(); 2167 2168 // Here we allocate the dummy full region that is required by the 2169 // G1AllocRegion class. If we don't pass an address in the reserved 2170 // space here, lots of asserts fire. 2171 2172 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2173 _g1_reserved.start()); 2174 // We'll re-use the same region whether the alloc region will 2175 // require BOT updates or not and, if it doesn't, then a non-young 2176 // region will complain that it cannot support allocations without 2177 // BOT updates. So we'll tag the dummy region as young to avoid that. 2178 dummy_region->set_young(); 2179 // Make sure it's full. 2180 dummy_region->set_top(dummy_region->end()); 2181 G1AllocRegion::setup(this, dummy_region); 2182 2183 init_mutator_alloc_region(); 2184 2185 // Do create of the monitoring and management support so that 2186 // values in the heap have been properly initialized. 2187 _g1mm = new G1MonitoringSupport(this); 2188 2189 return JNI_OK; 2190 } 2191 2192 size_t G1CollectedHeap::max_heap_alignment() { 2193 return HeapRegion::max_heap_alignment(); 2194 } 2195 2196 void G1CollectedHeap::ref_processing_init() { 2197 // Reference processing in G1 currently works as follows: 2198 // 2199 // * There are two reference processor instances. One is 2200 // used to record and process discovered references 2201 // during concurrent marking; the other is used to 2202 // record and process references during STW pauses 2203 // (both full and incremental). 2204 // * Both ref processors need to 'span' the entire heap as 2205 // the regions in the collection set may be dotted around. 2206 // 2207 // * For the concurrent marking ref processor: 2208 // * Reference discovery is enabled at initial marking. 2209 // * Reference discovery is disabled and the discovered 2210 // references processed etc during remarking. 2211 // * Reference discovery is MT (see below). 2212 // * Reference discovery requires a barrier (see below). 2213 // * Reference processing may or may not be MT 2214 // (depending on the value of ParallelRefProcEnabled 2215 // and ParallelGCThreads). 2216 // * A full GC disables reference discovery by the CM 2217 // ref processor and abandons any entries on it's 2218 // discovered lists. 2219 // 2220 // * For the STW processor: 2221 // * Non MT discovery is enabled at the start of a full GC. 2222 // * Processing and enqueueing during a full GC is non-MT. 2223 // * During a full GC, references are processed after marking. 2224 // 2225 // * Discovery (may or may not be MT) is enabled at the start 2226 // of an incremental evacuation pause. 2227 // * References are processed near the end of a STW evacuation pause. 2228 // * For both types of GC: 2229 // * Discovery is atomic - i.e. not concurrent. 2230 // * Reference discovery will not need a barrier. 2231 2232 SharedHeap::ref_processing_init(); 2233 MemRegion mr = reserved_region(); 2234 2235 // Concurrent Mark ref processor 2236 _ref_processor_cm = 2237 new ReferenceProcessor(mr, // span 2238 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2239 // mt processing 2240 (int) ParallelGCThreads, 2241 // degree of mt processing 2242 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2243 // mt discovery 2244 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2245 // degree of mt discovery 2246 false, 2247 // Reference discovery is not atomic 2248 &_is_alive_closure_cm, 2249 // is alive closure 2250 // (for efficiency/performance) 2251 true); 2252 // Setting next fields of discovered 2253 // lists requires a barrier. 2254 2255 // STW ref processor 2256 _ref_processor_stw = 2257 new ReferenceProcessor(mr, // span 2258 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2259 // mt processing 2260 MAX2((int)ParallelGCThreads, 1), 2261 // degree of mt processing 2262 (ParallelGCThreads > 1), 2263 // mt discovery 2264 MAX2((int)ParallelGCThreads, 1), 2265 // degree of mt discovery 2266 true, 2267 // Reference discovery is atomic 2268 &_is_alive_closure_stw, 2269 // is alive closure 2270 // (for efficiency/performance) 2271 false); 2272 // Setting next fields of discovered 2273 // lists requires a barrier. 2274 } 2275 2276 size_t G1CollectedHeap::capacity() const { 2277 return _g1_committed.byte_size(); 2278 } 2279 2280 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2281 assert(!hr->continuesHumongous(), "pre-condition"); 2282 hr->reset_gc_time_stamp(); 2283 if (hr->startsHumongous()) { 2284 uint first_index = hr->hrs_index() + 1; 2285 uint last_index = hr->last_hc_index(); 2286 for (uint i = first_index; i < last_index; i += 1) { 2287 HeapRegion* chr = region_at(i); 2288 assert(chr->continuesHumongous(), "sanity"); 2289 chr->reset_gc_time_stamp(); 2290 } 2291 } 2292 } 2293 2294 #ifndef PRODUCT 2295 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2296 private: 2297 unsigned _gc_time_stamp; 2298 bool _failures; 2299 2300 public: 2301 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2302 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2303 2304 virtual bool doHeapRegion(HeapRegion* hr) { 2305 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2306 if (_gc_time_stamp != region_gc_time_stamp) { 2307 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2308 "expected %d", HR_FORMAT_PARAMS(hr), 2309 region_gc_time_stamp, _gc_time_stamp); 2310 _failures = true; 2311 } 2312 return false; 2313 } 2314 2315 bool failures() { return _failures; } 2316 }; 2317 2318 void G1CollectedHeap::check_gc_time_stamps() { 2319 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2320 heap_region_iterate(&cl); 2321 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2322 } 2323 #endif // PRODUCT 2324 2325 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2326 DirtyCardQueue* into_cset_dcq, 2327 bool concurrent, 2328 int worker_i) { 2329 // Clean cards in the hot card cache 2330 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 2331 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq); 2332 2333 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2334 int n_completed_buffers = 0; 2335 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2336 n_completed_buffers++; 2337 } 2338 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2339 dcqs.clear_n_completed_buffers(); 2340 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2341 } 2342 2343 2344 // Computes the sum of the storage used by the various regions. 2345 2346 size_t G1CollectedHeap::used() const { 2347 assert(Heap_lock->owner() != NULL, 2348 "Should be owned on this thread's behalf."); 2349 size_t result = _summary_bytes_used; 2350 // Read only once in case it is set to NULL concurrently 2351 HeapRegion* hr = _mutator_alloc_region.get(); 2352 if (hr != NULL) 2353 result += hr->used(); 2354 return result; 2355 } 2356 2357 size_t G1CollectedHeap::used_unlocked() const { 2358 size_t result = _summary_bytes_used; 2359 return result; 2360 } 2361 2362 class SumUsedClosure: public HeapRegionClosure { 2363 size_t _used; 2364 public: 2365 SumUsedClosure() : _used(0) {} 2366 bool doHeapRegion(HeapRegion* r) { 2367 if (!r->continuesHumongous()) { 2368 _used += r->used(); 2369 } 2370 return false; 2371 } 2372 size_t result() { return _used; } 2373 }; 2374 2375 size_t G1CollectedHeap::recalculate_used() const { 2376 SumUsedClosure blk; 2377 heap_region_iterate(&blk); 2378 return blk.result(); 2379 } 2380 2381 size_t G1CollectedHeap::unsafe_max_alloc() { 2382 if (free_regions() > 0) return HeapRegion::GrainBytes; 2383 // otherwise, is there space in the current allocation region? 2384 2385 // We need to store the current allocation region in a local variable 2386 // here. The problem is that this method doesn't take any locks and 2387 // there may be other threads which overwrite the current allocation 2388 // region field. attempt_allocation(), for example, sets it to NULL 2389 // and this can happen *after* the NULL check here but before the call 2390 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2391 // to be a problem in the optimized build, since the two loads of the 2392 // current allocation region field are optimized away. 2393 HeapRegion* hr = _mutator_alloc_region.get(); 2394 if (hr == NULL) { 2395 return 0; 2396 } 2397 return hr->free(); 2398 } 2399 2400 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2401 switch (cause) { 2402 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2403 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2404 case GCCause::_g1_humongous_allocation: return true; 2405 default: return false; 2406 } 2407 } 2408 2409 #ifndef PRODUCT 2410 void G1CollectedHeap::allocate_dummy_regions() { 2411 // Let's fill up most of the region 2412 size_t word_size = HeapRegion::GrainWords - 1024; 2413 // And as a result the region we'll allocate will be humongous. 2414 guarantee(isHumongous(word_size), "sanity"); 2415 2416 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2417 // Let's use the existing mechanism for the allocation 2418 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2419 if (dummy_obj != NULL) { 2420 MemRegion mr(dummy_obj, word_size); 2421 CollectedHeap::fill_with_object(mr); 2422 } else { 2423 // If we can't allocate once, we probably cannot allocate 2424 // again. Let's get out of the loop. 2425 break; 2426 } 2427 } 2428 } 2429 #endif // !PRODUCT 2430 2431 void G1CollectedHeap::increment_old_marking_cycles_started() { 2432 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2433 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2434 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2435 _old_marking_cycles_started, _old_marking_cycles_completed)); 2436 2437 _old_marking_cycles_started++; 2438 } 2439 2440 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2441 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2442 2443 // We assume that if concurrent == true, then the caller is a 2444 // concurrent thread that was joined the Suspendible Thread 2445 // Set. If there's ever a cheap way to check this, we should add an 2446 // assert here. 2447 2448 // Given that this method is called at the end of a Full GC or of a 2449 // concurrent cycle, and those can be nested (i.e., a Full GC can 2450 // interrupt a concurrent cycle), the number of full collections 2451 // completed should be either one (in the case where there was no 2452 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2453 // behind the number of full collections started. 2454 2455 // This is the case for the inner caller, i.e. a Full GC. 2456 assert(concurrent || 2457 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2458 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2459 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2460 "is inconsistent with _old_marking_cycles_completed = %u", 2461 _old_marking_cycles_started, _old_marking_cycles_completed)); 2462 2463 // This is the case for the outer caller, i.e. the concurrent cycle. 2464 assert(!concurrent || 2465 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2466 err_msg("for outer caller (concurrent cycle): " 2467 "_old_marking_cycles_started = %u " 2468 "is inconsistent with _old_marking_cycles_completed = %u", 2469 _old_marking_cycles_started, _old_marking_cycles_completed)); 2470 2471 _old_marking_cycles_completed += 1; 2472 2473 // We need to clear the "in_progress" flag in the CM thread before 2474 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2475 // is set) so that if a waiter requests another System.gc() it doesn't 2476 // incorrectly see that a marking cycle is still in progress. 2477 if (concurrent) { 2478 _cmThread->clear_in_progress(); 2479 } 2480 2481 // This notify_all() will ensure that a thread that called 2482 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2483 // and it's waiting for a full GC to finish will be woken up. It is 2484 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2485 FullGCCount_lock->notify_all(); 2486 } 2487 2488 void G1CollectedHeap::register_concurrent_cycle_start(jlong start_time) { 2489 _concurrent_cycle_started = true; 2490 _gc_timer_cm->register_gc_start(start_time); 2491 2492 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start()); 2493 trace_heap_before_gc(_gc_tracer_cm); 2494 } 2495 2496 void G1CollectedHeap::register_concurrent_cycle_end() { 2497 if (_concurrent_cycle_started) { 2498 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2499 2500 if (_cm->has_aborted()) { 2501 _gc_tracer_cm->report_concurrent_mode_failure(); 2502 } 2503 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2504 2505 _concurrent_cycle_started = false; 2506 } 2507 } 2508 2509 void G1CollectedHeap::trace_heap_after_concurrent_cycle() { 2510 if (_concurrent_cycle_started) { 2511 trace_heap_after_gc(_gc_tracer_cm); 2512 } 2513 } 2514 2515 G1YCType G1CollectedHeap::yc_type() { 2516 bool is_young = g1_policy()->gcs_are_young(); 2517 bool is_initial_mark = g1_policy()->during_initial_mark_pause(); 2518 bool is_during_mark = mark_in_progress(); 2519 2520 if (is_initial_mark) { 2521 return InitialMark; 2522 } else if (is_during_mark) { 2523 return DuringMark; 2524 } else if (is_young) { 2525 return Normal; 2526 } else { 2527 return Mixed; 2528 } 2529 } 2530 2531 void G1CollectedHeap::collect(GCCause::Cause cause) { 2532 assert_heap_not_locked(); 2533 2534 unsigned int gc_count_before; 2535 unsigned int old_marking_count_before; 2536 bool retry_gc; 2537 2538 do { 2539 retry_gc = false; 2540 2541 { 2542 MutexLocker ml(Heap_lock); 2543 2544 // Read the GC count while holding the Heap_lock 2545 gc_count_before = total_collections(); 2546 old_marking_count_before = _old_marking_cycles_started; 2547 } 2548 2549 if (should_do_concurrent_full_gc(cause)) { 2550 // Schedule an initial-mark evacuation pause that will start a 2551 // concurrent cycle. We're setting word_size to 0 which means that 2552 // we are not requesting a post-GC allocation. 2553 VM_G1IncCollectionPause op(gc_count_before, 2554 0, /* word_size */ 2555 true, /* should_initiate_conc_mark */ 2556 g1_policy()->max_pause_time_ms(), 2557 cause); 2558 2559 VMThread::execute(&op); 2560 if (!op.pause_succeeded()) { 2561 if (old_marking_count_before == _old_marking_cycles_started) { 2562 retry_gc = op.should_retry_gc(); 2563 } else { 2564 // A Full GC happened while we were trying to schedule the 2565 // initial-mark GC. No point in starting a new cycle given 2566 // that the whole heap was collected anyway. 2567 } 2568 2569 if (retry_gc) { 2570 if (GC_locker::is_active_and_needs_gc()) { 2571 GC_locker::stall_until_clear(); 2572 } 2573 } 2574 } 2575 } else { 2576 if (cause == GCCause::_gc_locker 2577 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2578 2579 // Schedule a standard evacuation pause. We're setting word_size 2580 // to 0 which means that we are not requesting a post-GC allocation. 2581 VM_G1IncCollectionPause op(gc_count_before, 2582 0, /* word_size */ 2583 false, /* should_initiate_conc_mark */ 2584 g1_policy()->max_pause_time_ms(), 2585 cause); 2586 VMThread::execute(&op); 2587 } else { 2588 // Schedule a Full GC. 2589 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2590 VMThread::execute(&op); 2591 } 2592 } 2593 } while (retry_gc); 2594 } 2595 2596 bool G1CollectedHeap::is_in(const void* p) const { 2597 if (_g1_committed.contains(p)) { 2598 // Given that we know that p is in the committed space, 2599 // heap_region_containing_raw() should successfully 2600 // return the containing region. 2601 HeapRegion* hr = heap_region_containing_raw(p); 2602 return hr->is_in(p); 2603 } else { 2604 return false; 2605 } 2606 } 2607 2608 // Iteration functions. 2609 2610 // Iterates an OopClosure over all ref-containing fields of objects 2611 // within a HeapRegion. 2612 2613 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2614 MemRegion _mr; 2615 ExtendedOopClosure* _cl; 2616 public: 2617 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2618 : _mr(mr), _cl(cl) {} 2619 bool doHeapRegion(HeapRegion* r) { 2620 if (!r->continuesHumongous()) { 2621 r->oop_iterate(_cl); 2622 } 2623 return false; 2624 } 2625 }; 2626 2627 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2628 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2629 heap_region_iterate(&blk); 2630 } 2631 2632 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2633 IterateOopClosureRegionClosure blk(mr, cl); 2634 heap_region_iterate(&blk); 2635 } 2636 2637 // Iterates an ObjectClosure over all objects within a HeapRegion. 2638 2639 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2640 ObjectClosure* _cl; 2641 public: 2642 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2643 bool doHeapRegion(HeapRegion* r) { 2644 if (! r->continuesHumongous()) { 2645 r->object_iterate(_cl); 2646 } 2647 return false; 2648 } 2649 }; 2650 2651 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2652 IterateObjectClosureRegionClosure blk(cl); 2653 heap_region_iterate(&blk); 2654 } 2655 2656 // Calls a SpaceClosure on a HeapRegion. 2657 2658 class SpaceClosureRegionClosure: public HeapRegionClosure { 2659 SpaceClosure* _cl; 2660 public: 2661 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2662 bool doHeapRegion(HeapRegion* r) { 2663 _cl->do_space(r); 2664 return false; 2665 } 2666 }; 2667 2668 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2669 SpaceClosureRegionClosure blk(cl); 2670 heap_region_iterate(&blk); 2671 } 2672 2673 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2674 _hrs.iterate(cl); 2675 } 2676 2677 void 2678 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2679 uint worker_id, 2680 uint no_of_par_workers, 2681 jint claim_value) { 2682 const uint regions = n_regions(); 2683 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2684 no_of_par_workers : 2685 1); 2686 assert(UseDynamicNumberOfGCThreads || 2687 no_of_par_workers == workers()->total_workers(), 2688 "Non dynamic should use fixed number of workers"); 2689 // try to spread out the starting points of the workers 2690 const HeapRegion* start_hr = 2691 start_region_for_worker(worker_id, no_of_par_workers); 2692 const uint start_index = start_hr->hrs_index(); 2693 2694 // each worker will actually look at all regions 2695 for (uint count = 0; count < regions; ++count) { 2696 const uint index = (start_index + count) % regions; 2697 assert(0 <= index && index < regions, "sanity"); 2698 HeapRegion* r = region_at(index); 2699 // we'll ignore "continues humongous" regions (we'll process them 2700 // when we come across their corresponding "start humongous" 2701 // region) and regions already claimed 2702 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2703 continue; 2704 } 2705 // OK, try to claim it 2706 if (r->claimHeapRegion(claim_value)) { 2707 // success! 2708 assert(!r->continuesHumongous(), "sanity"); 2709 if (r->startsHumongous()) { 2710 // If the region is "starts humongous" we'll iterate over its 2711 // "continues humongous" first; in fact we'll do them 2712 // first. The order is important. In on case, calling the 2713 // closure on the "starts humongous" region might de-allocate 2714 // and clear all its "continues humongous" regions and, as a 2715 // result, we might end up processing them twice. So, we'll do 2716 // them first (notice: most closures will ignore them anyway) and 2717 // then we'll do the "starts humongous" region. 2718 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2719 HeapRegion* chr = region_at(ch_index); 2720 2721 // if the region has already been claimed or it's not 2722 // "continues humongous" we're done 2723 if (chr->claim_value() == claim_value || 2724 !chr->continuesHumongous()) { 2725 break; 2726 } 2727 2728 // No one should have claimed it directly. We can given 2729 // that we claimed its "starts humongous" region. 2730 assert(chr->claim_value() != claim_value, "sanity"); 2731 assert(chr->humongous_start_region() == r, "sanity"); 2732 2733 if (chr->claimHeapRegion(claim_value)) { 2734 // we should always be able to claim it; no one else should 2735 // be trying to claim this region 2736 2737 bool res2 = cl->doHeapRegion(chr); 2738 assert(!res2, "Should not abort"); 2739 2740 // Right now, this holds (i.e., no closure that actually 2741 // does something with "continues humongous" regions 2742 // clears them). We might have to weaken it in the future, 2743 // but let's leave these two asserts here for extra safety. 2744 assert(chr->continuesHumongous(), "should still be the case"); 2745 assert(chr->humongous_start_region() == r, "sanity"); 2746 } else { 2747 guarantee(false, "we should not reach here"); 2748 } 2749 } 2750 } 2751 2752 assert(!r->continuesHumongous(), "sanity"); 2753 bool res = cl->doHeapRegion(r); 2754 assert(!res, "Should not abort"); 2755 } 2756 } 2757 } 2758 2759 class ResetClaimValuesClosure: public HeapRegionClosure { 2760 public: 2761 bool doHeapRegion(HeapRegion* r) { 2762 r->set_claim_value(HeapRegion::InitialClaimValue); 2763 return false; 2764 } 2765 }; 2766 2767 void G1CollectedHeap::reset_heap_region_claim_values() { 2768 ResetClaimValuesClosure blk; 2769 heap_region_iterate(&blk); 2770 } 2771 2772 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2773 ResetClaimValuesClosure blk; 2774 collection_set_iterate(&blk); 2775 } 2776 2777 #ifdef ASSERT 2778 // This checks whether all regions in the heap have the correct claim 2779 // value. I also piggy-backed on this a check to ensure that the 2780 // humongous_start_region() information on "continues humongous" 2781 // regions is correct. 2782 2783 class CheckClaimValuesClosure : public HeapRegionClosure { 2784 private: 2785 jint _claim_value; 2786 uint _failures; 2787 HeapRegion* _sh_region; 2788 2789 public: 2790 CheckClaimValuesClosure(jint claim_value) : 2791 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2792 bool doHeapRegion(HeapRegion* r) { 2793 if (r->claim_value() != _claim_value) { 2794 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2795 "claim value = %d, should be %d", 2796 HR_FORMAT_PARAMS(r), 2797 r->claim_value(), _claim_value); 2798 ++_failures; 2799 } 2800 if (!r->isHumongous()) { 2801 _sh_region = NULL; 2802 } else if (r->startsHumongous()) { 2803 _sh_region = r; 2804 } else if (r->continuesHumongous()) { 2805 if (r->humongous_start_region() != _sh_region) { 2806 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2807 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2808 HR_FORMAT_PARAMS(r), 2809 r->humongous_start_region(), 2810 _sh_region); 2811 ++_failures; 2812 } 2813 } 2814 return false; 2815 } 2816 uint failures() { return _failures; } 2817 }; 2818 2819 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2820 CheckClaimValuesClosure cl(claim_value); 2821 heap_region_iterate(&cl); 2822 return cl.failures() == 0; 2823 } 2824 2825 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2826 private: 2827 jint _claim_value; 2828 uint _failures; 2829 2830 public: 2831 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2832 _claim_value(claim_value), _failures(0) { } 2833 2834 uint failures() { return _failures; } 2835 2836 bool doHeapRegion(HeapRegion* hr) { 2837 assert(hr->in_collection_set(), "how?"); 2838 assert(!hr->isHumongous(), "H-region in CSet"); 2839 if (hr->claim_value() != _claim_value) { 2840 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2841 "claim value = %d, should be %d", 2842 HR_FORMAT_PARAMS(hr), 2843 hr->claim_value(), _claim_value); 2844 _failures += 1; 2845 } 2846 return false; 2847 } 2848 }; 2849 2850 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2851 CheckClaimValuesInCSetHRClosure cl(claim_value); 2852 collection_set_iterate(&cl); 2853 return cl.failures() == 0; 2854 } 2855 #endif // ASSERT 2856 2857 // Clear the cached CSet starting regions and (more importantly) 2858 // the time stamps. Called when we reset the GC time stamp. 2859 void G1CollectedHeap::clear_cset_start_regions() { 2860 assert(_worker_cset_start_region != NULL, "sanity"); 2861 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2862 2863 int n_queues = MAX2((int)ParallelGCThreads, 1); 2864 for (int i = 0; i < n_queues; i++) { 2865 _worker_cset_start_region[i] = NULL; 2866 _worker_cset_start_region_time_stamp[i] = 0; 2867 } 2868 } 2869 2870 // Given the id of a worker, obtain or calculate a suitable 2871 // starting region for iterating over the current collection set. 2872 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2873 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2874 2875 HeapRegion* result = NULL; 2876 unsigned gc_time_stamp = get_gc_time_stamp(); 2877 2878 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2879 // Cached starting region for current worker was set 2880 // during the current pause - so it's valid. 2881 // Note: the cached starting heap region may be NULL 2882 // (when the collection set is empty). 2883 result = _worker_cset_start_region[worker_i]; 2884 assert(result == NULL || result->in_collection_set(), "sanity"); 2885 return result; 2886 } 2887 2888 // The cached entry was not valid so let's calculate 2889 // a suitable starting heap region for this worker. 2890 2891 // We want the parallel threads to start their collection 2892 // set iteration at different collection set regions to 2893 // avoid contention. 2894 // If we have: 2895 // n collection set regions 2896 // p threads 2897 // Then thread t will start at region floor ((t * n) / p) 2898 2899 result = g1_policy()->collection_set(); 2900 if (G1CollectedHeap::use_parallel_gc_threads()) { 2901 uint cs_size = g1_policy()->cset_region_length(); 2902 uint active_workers = workers()->active_workers(); 2903 assert(UseDynamicNumberOfGCThreads || 2904 active_workers == workers()->total_workers(), 2905 "Unless dynamic should use total workers"); 2906 2907 uint end_ind = (cs_size * worker_i) / active_workers; 2908 uint start_ind = 0; 2909 2910 if (worker_i > 0 && 2911 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2912 // Previous workers starting region is valid 2913 // so let's iterate from there 2914 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2915 result = _worker_cset_start_region[worker_i - 1]; 2916 } 2917 2918 for (uint i = start_ind; i < end_ind; i++) { 2919 result = result->next_in_collection_set(); 2920 } 2921 } 2922 2923 // Note: the calculated starting heap region may be NULL 2924 // (when the collection set is empty). 2925 assert(result == NULL || result->in_collection_set(), "sanity"); 2926 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2927 "should be updated only once per pause"); 2928 _worker_cset_start_region[worker_i] = result; 2929 OrderAccess::storestore(); 2930 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2931 return result; 2932 } 2933 2934 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2935 uint no_of_par_workers) { 2936 uint worker_num = 2937 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2938 assert(UseDynamicNumberOfGCThreads || 2939 no_of_par_workers == workers()->total_workers(), 2940 "Non dynamic should use fixed number of workers"); 2941 const uint start_index = n_regions() * worker_i / worker_num; 2942 return region_at(start_index); 2943 } 2944 2945 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2946 HeapRegion* r = g1_policy()->collection_set(); 2947 while (r != NULL) { 2948 HeapRegion* next = r->next_in_collection_set(); 2949 if (cl->doHeapRegion(r)) { 2950 cl->incomplete(); 2951 return; 2952 } 2953 r = next; 2954 } 2955 } 2956 2957 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2958 HeapRegionClosure *cl) { 2959 if (r == NULL) { 2960 // The CSet is empty so there's nothing to do. 2961 return; 2962 } 2963 2964 assert(r->in_collection_set(), 2965 "Start region must be a member of the collection set."); 2966 HeapRegion* cur = r; 2967 while (cur != NULL) { 2968 HeapRegion* next = cur->next_in_collection_set(); 2969 if (cl->doHeapRegion(cur) && false) { 2970 cl->incomplete(); 2971 return; 2972 } 2973 cur = next; 2974 } 2975 cur = g1_policy()->collection_set(); 2976 while (cur != r) { 2977 HeapRegion* next = cur->next_in_collection_set(); 2978 if (cl->doHeapRegion(cur) && false) { 2979 cl->incomplete(); 2980 return; 2981 } 2982 cur = next; 2983 } 2984 } 2985 2986 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2987 return n_regions() > 0 ? region_at(0) : NULL; 2988 } 2989 2990 2991 Space* G1CollectedHeap::space_containing(const void* addr) const { 2992 Space* res = heap_region_containing(addr); 2993 return res; 2994 } 2995 2996 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2997 Space* sp = space_containing(addr); 2998 if (sp != NULL) { 2999 return sp->block_start(addr); 3000 } 3001 return NULL; 3002 } 3003 3004 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 3005 Space* sp = space_containing(addr); 3006 assert(sp != NULL, "block_size of address outside of heap"); 3007 return sp->block_size(addr); 3008 } 3009 3010 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 3011 Space* sp = space_containing(addr); 3012 return sp->block_is_obj(addr); 3013 } 3014 3015 bool G1CollectedHeap::supports_tlab_allocation() const { 3016 return true; 3017 } 3018 3019 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 3020 return HeapRegion::GrainBytes; 3021 } 3022 3023 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 3024 // Return the remaining space in the cur alloc region, but not less than 3025 // the min TLAB size. 3026 3027 // Also, this value can be at most the humongous object threshold, 3028 // since we can't allow tlabs to grow big enough to accommodate 3029 // humongous objects. 3030 3031 HeapRegion* hr = _mutator_alloc_region.get(); 3032 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 3033 if (hr == NULL) { 3034 return max_tlab_size; 3035 } else { 3036 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 3037 } 3038 } 3039 3040 size_t G1CollectedHeap::max_capacity() const { 3041 return _g1_reserved.byte_size(); 3042 } 3043 3044 jlong G1CollectedHeap::millis_since_last_gc() { 3045 // assert(false, "NYI"); 3046 return 0; 3047 } 3048 3049 void G1CollectedHeap::prepare_for_verify() { 3050 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3051 ensure_parsability(false); 3052 } 3053 g1_rem_set()->prepare_for_verify(); 3054 } 3055 3056 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3057 VerifyOption vo) { 3058 switch (vo) { 3059 case VerifyOption_G1UsePrevMarking: 3060 return hr->obj_allocated_since_prev_marking(obj); 3061 case VerifyOption_G1UseNextMarking: 3062 return hr->obj_allocated_since_next_marking(obj); 3063 case VerifyOption_G1UseMarkWord: 3064 return false; 3065 default: 3066 ShouldNotReachHere(); 3067 } 3068 return false; // keep some compilers happy 3069 } 3070 3071 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3072 switch (vo) { 3073 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3074 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3075 case VerifyOption_G1UseMarkWord: return NULL; 3076 default: ShouldNotReachHere(); 3077 } 3078 return NULL; // keep some compilers happy 3079 } 3080 3081 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3082 switch (vo) { 3083 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3084 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3085 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3086 default: ShouldNotReachHere(); 3087 } 3088 return false; // keep some compilers happy 3089 } 3090 3091 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3092 switch (vo) { 3093 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3094 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3095 case VerifyOption_G1UseMarkWord: return "NONE"; 3096 default: ShouldNotReachHere(); 3097 } 3098 return NULL; // keep some compilers happy 3099 } 3100 3101 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can 3102 // pass it as the perm_blk to SharedHeap::process_strong_roots. 3103 // When process_strong_roots stop calling perm_blk->younger_refs_iterate 3104 // we can change this closure to extend the simpler OopClosure. 3105 class VerifyRootsClosure: public OopsInGenClosure { 3106 private: 3107 G1CollectedHeap* _g1h; 3108 VerifyOption _vo; 3109 bool _failures; 3110 public: 3111 // _vo == UsePrevMarking -> use "prev" marking information, 3112 // _vo == UseNextMarking -> use "next" marking information, 3113 // _vo == UseMarkWord -> use mark word from object header. 3114 VerifyRootsClosure(VerifyOption vo) : 3115 _g1h(G1CollectedHeap::heap()), 3116 _vo(vo), 3117 _failures(false) { } 3118 3119 bool failures() { return _failures; } 3120 3121 template <class T> void do_oop_nv(T* p) { 3122 T heap_oop = oopDesc::load_heap_oop(p); 3123 if (!oopDesc::is_null(heap_oop)) { 3124 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3125 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3126 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3127 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3128 if (_vo == VerifyOption_G1UseMarkWord) { 3129 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3130 } 3131 obj->print_on(gclog_or_tty); 3132 _failures = true; 3133 } 3134 } 3135 } 3136 3137 void do_oop(oop* p) { do_oop_nv(p); } 3138 void do_oop(narrowOop* p) { do_oop_nv(p); } 3139 }; 3140 3141 class G1VerifyCodeRootOopClosure: public OopsInGenClosure { 3142 G1CollectedHeap* _g1h; 3143 OopClosure* _root_cl; 3144 nmethod* _nm; 3145 VerifyOption _vo; 3146 bool _failures; 3147 3148 template <class T> void do_oop_work(T* p) { 3149 // First verify that this root is live 3150 _root_cl->do_oop(p); 3151 3152 if (!G1VerifyHeapRegionCodeRoots) { 3153 // We're not verifying the code roots attached to heap region. 3154 return; 3155 } 3156 3157 // Don't check the code roots during marking verification in a full GC 3158 if (_vo == VerifyOption_G1UseMarkWord) { 3159 return; 3160 } 3161 3162 // Now verify that the current nmethod (which contains p) is 3163 // in the code root list of the heap region containing the 3164 // object referenced by p. 3165 3166 T heap_oop = oopDesc::load_heap_oop(p); 3167 if (!oopDesc::is_null(heap_oop)) { 3168 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3169 3170 // Now fetch the region containing the object 3171 HeapRegion* hr = _g1h->heap_region_containing(obj); 3172 HeapRegionRemSet* hrrs = hr->rem_set(); 3173 // Verify that the strong code root list for this region 3174 // contains the nmethod 3175 if (!hrrs->strong_code_roots_list_contains(_nm)) { 3176 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" " 3177 "from nmethod "PTR_FORMAT" not in strong " 3178 "code roots for region ["PTR_FORMAT","PTR_FORMAT")", 3179 p, _nm, hr->bottom(), hr->end()); 3180 _failures = true; 3181 } 3182 } 3183 } 3184 3185 public: 3186 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo): 3187 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {} 3188 3189 void do_oop(oop* p) { do_oop_work(p); } 3190 void do_oop(narrowOop* p) { do_oop_work(p); } 3191 3192 void set_nmethod(nmethod* nm) { _nm = nm; } 3193 bool failures() { return _failures; } 3194 }; 3195 3196 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure { 3197 G1VerifyCodeRootOopClosure* _oop_cl; 3198 3199 public: 3200 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl): 3201 _oop_cl(oop_cl) {} 3202 3203 void do_code_blob(CodeBlob* cb) { 3204 nmethod* nm = cb->as_nmethod_or_null(); 3205 if (nm != NULL) { 3206 _oop_cl->set_nmethod(nm); 3207 nm->oops_do(_oop_cl); 3208 } 3209 } 3210 }; 3211 3212 class YoungRefCounterClosure : public OopClosure { 3213 G1CollectedHeap* _g1h; 3214 int _count; 3215 public: 3216 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3217 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3218 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3219 3220 int count() { return _count; } 3221 void reset_count() { _count = 0; }; 3222 }; 3223 3224 class VerifyKlassClosure: public KlassClosure { 3225 YoungRefCounterClosure _young_ref_counter_closure; 3226 OopClosure *_oop_closure; 3227 public: 3228 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3229 void do_klass(Klass* k) { 3230 k->oops_do(_oop_closure); 3231 3232 _young_ref_counter_closure.reset_count(); 3233 k->oops_do(&_young_ref_counter_closure); 3234 if (_young_ref_counter_closure.count() > 0) { 3235 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3236 } 3237 } 3238 }; 3239 3240 class VerifyLivenessOopClosure: public OopClosure { 3241 G1CollectedHeap* _g1h; 3242 VerifyOption _vo; 3243 public: 3244 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3245 _g1h(g1h), _vo(vo) 3246 { } 3247 void do_oop(narrowOop *p) { do_oop_work(p); } 3248 void do_oop( oop *p) { do_oop_work(p); } 3249 3250 template <class T> void do_oop_work(T *p) { 3251 oop obj = oopDesc::load_decode_heap_oop(p); 3252 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3253 "Dead object referenced by a not dead object"); 3254 } 3255 }; 3256 3257 class VerifyObjsInRegionClosure: public ObjectClosure { 3258 private: 3259 G1CollectedHeap* _g1h; 3260 size_t _live_bytes; 3261 HeapRegion *_hr; 3262 VerifyOption _vo; 3263 public: 3264 // _vo == UsePrevMarking -> use "prev" marking information, 3265 // _vo == UseNextMarking -> use "next" marking information, 3266 // _vo == UseMarkWord -> use mark word from object header. 3267 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3268 : _live_bytes(0), _hr(hr), _vo(vo) { 3269 _g1h = G1CollectedHeap::heap(); 3270 } 3271 void do_object(oop o) { 3272 VerifyLivenessOopClosure isLive(_g1h, _vo); 3273 assert(o != NULL, "Huh?"); 3274 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3275 // If the object is alive according to the mark word, 3276 // then verify that the marking information agrees. 3277 // Note we can't verify the contra-positive of the 3278 // above: if the object is dead (according to the mark 3279 // word), it may not be marked, or may have been marked 3280 // but has since became dead, or may have been allocated 3281 // since the last marking. 3282 if (_vo == VerifyOption_G1UseMarkWord) { 3283 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3284 } 3285 3286 o->oop_iterate_no_header(&isLive); 3287 if (!_hr->obj_allocated_since_prev_marking(o)) { 3288 size_t obj_size = o->size(); // Make sure we don't overflow 3289 _live_bytes += (obj_size * HeapWordSize); 3290 } 3291 } 3292 } 3293 size_t live_bytes() { return _live_bytes; } 3294 }; 3295 3296 class PrintObjsInRegionClosure : public ObjectClosure { 3297 HeapRegion *_hr; 3298 G1CollectedHeap *_g1; 3299 public: 3300 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3301 _g1 = G1CollectedHeap::heap(); 3302 }; 3303 3304 void do_object(oop o) { 3305 if (o != NULL) { 3306 HeapWord *start = (HeapWord *) o; 3307 size_t word_sz = o->size(); 3308 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3309 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3310 (void*) o, word_sz, 3311 _g1->isMarkedPrev(o), 3312 _g1->isMarkedNext(o), 3313 _hr->obj_allocated_since_prev_marking(o)); 3314 HeapWord *end = start + word_sz; 3315 HeapWord *cur; 3316 int *val; 3317 for (cur = start; cur < end; cur++) { 3318 val = (int *) cur; 3319 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3320 } 3321 } 3322 } 3323 }; 3324 3325 class VerifyRegionClosure: public HeapRegionClosure { 3326 private: 3327 bool _par; 3328 VerifyOption _vo; 3329 bool _failures; 3330 public: 3331 // _vo == UsePrevMarking -> use "prev" marking information, 3332 // _vo == UseNextMarking -> use "next" marking information, 3333 // _vo == UseMarkWord -> use mark word from object header. 3334 VerifyRegionClosure(bool par, VerifyOption vo) 3335 : _par(par), 3336 _vo(vo), 3337 _failures(false) {} 3338 3339 bool failures() { 3340 return _failures; 3341 } 3342 3343 bool doHeapRegion(HeapRegion* r) { 3344 if (!r->continuesHumongous()) { 3345 bool failures = false; 3346 r->verify(_vo, &failures); 3347 if (failures) { 3348 _failures = true; 3349 } else { 3350 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3351 r->object_iterate(¬_dead_yet_cl); 3352 if (_vo != VerifyOption_G1UseNextMarking) { 3353 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3354 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3355 "max_live_bytes "SIZE_FORMAT" " 3356 "< calculated "SIZE_FORMAT, 3357 r->bottom(), r->end(), 3358 r->max_live_bytes(), 3359 not_dead_yet_cl.live_bytes()); 3360 _failures = true; 3361 } 3362 } else { 3363 // When vo == UseNextMarking we cannot currently do a sanity 3364 // check on the live bytes as the calculation has not been 3365 // finalized yet. 3366 } 3367 } 3368 } 3369 return false; // stop the region iteration if we hit a failure 3370 } 3371 }; 3372 3373 // This is the task used for parallel verification of the heap regions 3374 3375 class G1ParVerifyTask: public AbstractGangTask { 3376 private: 3377 G1CollectedHeap* _g1h; 3378 VerifyOption _vo; 3379 bool _failures; 3380 3381 public: 3382 // _vo == UsePrevMarking -> use "prev" marking information, 3383 // _vo == UseNextMarking -> use "next" marking information, 3384 // _vo == UseMarkWord -> use mark word from object header. 3385 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3386 AbstractGangTask("Parallel verify task"), 3387 _g1h(g1h), 3388 _vo(vo), 3389 _failures(false) { } 3390 3391 bool failures() { 3392 return _failures; 3393 } 3394 3395 void work(uint worker_id) { 3396 HandleMark hm; 3397 VerifyRegionClosure blk(true, _vo); 3398 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3399 _g1h->workers()->active_workers(), 3400 HeapRegion::ParVerifyClaimValue); 3401 if (blk.failures()) { 3402 _failures = true; 3403 } 3404 } 3405 }; 3406 3407 void G1CollectedHeap::verify(bool silent, VerifyOption vo) { 3408 if (SafepointSynchronize::is_at_safepoint()) { 3409 assert(Thread::current()->is_VM_thread(), 3410 "Expected to be executed serially by the VM thread at this point"); 3411 3412 if (!silent) { gclog_or_tty->print("Roots "); } 3413 VerifyRootsClosure rootsCl(vo); 3414 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo); 3415 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl); 3416 VerifyKlassClosure klassCl(this, &rootsCl); 3417 3418 // We apply the relevant closures to all the oops in the 3419 // system dictionary, the string table and the code cache. 3420 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3421 3422 // Need cleared claim bits for the strong roots processing 3423 ClassLoaderDataGraph::clear_claimed_marks(); 3424 3425 process_strong_roots(true, // activate StrongRootsScope 3426 false, // we set "is scavenging" to false, 3427 // so we don't reset the dirty cards. 3428 ScanningOption(so), // roots scanning options 3429 &rootsCl, 3430 &blobsCl, 3431 &klassCl 3432 ); 3433 3434 bool failures = rootsCl.failures() || codeRootsCl.failures(); 3435 3436 if (vo != VerifyOption_G1UseMarkWord) { 3437 // If we're verifying during a full GC then the region sets 3438 // will have been torn down at the start of the GC. Therefore 3439 // verifying the region sets will fail. So we only verify 3440 // the region sets when not in a full GC. 3441 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3442 verify_region_sets(); 3443 } 3444 3445 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3446 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3447 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3448 "sanity check"); 3449 3450 G1ParVerifyTask task(this, vo); 3451 assert(UseDynamicNumberOfGCThreads || 3452 workers()->active_workers() == workers()->total_workers(), 3453 "If not dynamic should be using all the workers"); 3454 int n_workers = workers()->active_workers(); 3455 set_par_threads(n_workers); 3456 workers()->run_task(&task); 3457 set_par_threads(0); 3458 if (task.failures()) { 3459 failures = true; 3460 } 3461 3462 // Checks that the expected amount of parallel work was done. 3463 // The implication is that n_workers is > 0. 3464 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3465 "sanity check"); 3466 3467 reset_heap_region_claim_values(); 3468 3469 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3470 "sanity check"); 3471 } else { 3472 VerifyRegionClosure blk(false, vo); 3473 heap_region_iterate(&blk); 3474 if (blk.failures()) { 3475 failures = true; 3476 } 3477 } 3478 if (!silent) gclog_or_tty->print("RemSet "); 3479 rem_set()->verify(); 3480 3481 if (failures) { 3482 gclog_or_tty->print_cr("Heap:"); 3483 // It helps to have the per-region information in the output to 3484 // help us track down what went wrong. This is why we call 3485 // print_extended_on() instead of print_on(). 3486 print_extended_on(gclog_or_tty); 3487 gclog_or_tty->print_cr(""); 3488 #ifndef PRODUCT 3489 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3490 concurrent_mark()->print_reachable("at-verification-failure", 3491 vo, false /* all */); 3492 } 3493 #endif 3494 gclog_or_tty->flush(); 3495 } 3496 guarantee(!failures, "there should not have been any failures"); 3497 } else { 3498 if (!silent) 3499 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) "); 3500 } 3501 } 3502 3503 void G1CollectedHeap::verify(bool silent) { 3504 verify(silent, VerifyOption_G1UsePrevMarking); 3505 } 3506 3507 double G1CollectedHeap::verify(bool guard, const char* msg) { 3508 double verify_time_ms = 0.0; 3509 3510 if (guard && total_collections() >= VerifyGCStartAt) { 3511 double verify_start = os::elapsedTime(); 3512 HandleMark hm; // Discard invalid handles created during verification 3513 prepare_for_verify(); 3514 Universe::verify(VerifyOption_G1UsePrevMarking, msg); 3515 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 3516 } 3517 3518 return verify_time_ms; 3519 } 3520 3521 void G1CollectedHeap::verify_before_gc() { 3522 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 3523 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 3524 } 3525 3526 void G1CollectedHeap::verify_after_gc() { 3527 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 3528 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 3529 } 3530 3531 class PrintRegionClosure: public HeapRegionClosure { 3532 outputStream* _st; 3533 public: 3534 PrintRegionClosure(outputStream* st) : _st(st) {} 3535 bool doHeapRegion(HeapRegion* r) { 3536 r->print_on(_st); 3537 return false; 3538 } 3539 }; 3540 3541 void G1CollectedHeap::print_on(outputStream* st) const { 3542 st->print(" %-20s", "garbage-first heap"); 3543 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3544 capacity()/K, used_unlocked()/K); 3545 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3546 _g1_storage.low_boundary(), 3547 _g1_storage.high(), 3548 _g1_storage.high_boundary()); 3549 st->cr(); 3550 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3551 uint young_regions = _young_list->length(); 3552 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3553 (size_t) young_regions * HeapRegion::GrainBytes / K); 3554 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3555 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3556 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3557 st->cr(); 3558 MetaspaceAux::print_on(st); 3559 } 3560 3561 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3562 print_on(st); 3563 3564 // Print the per-region information. 3565 st->cr(); 3566 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3567 "HS=humongous(starts), HC=humongous(continues), " 3568 "CS=collection set, F=free, TS=gc time stamp, " 3569 "PTAMS=previous top-at-mark-start, " 3570 "NTAMS=next top-at-mark-start)"); 3571 PrintRegionClosure blk(st); 3572 heap_region_iterate(&blk); 3573 } 3574 3575 void G1CollectedHeap::print_on_error(outputStream* st) const { 3576 this->CollectedHeap::print_on_error(st); 3577 3578 if (_cm != NULL) { 3579 st->cr(); 3580 _cm->print_on_error(st); 3581 } 3582 } 3583 3584 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3585 if (G1CollectedHeap::use_parallel_gc_threads()) { 3586 workers()->print_worker_threads_on(st); 3587 } 3588 _cmThread->print_on(st); 3589 st->cr(); 3590 _cm->print_worker_threads_on(st); 3591 _cg1r->print_worker_threads_on(st); 3592 } 3593 3594 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3595 if (G1CollectedHeap::use_parallel_gc_threads()) { 3596 workers()->threads_do(tc); 3597 } 3598 tc->do_thread(_cmThread); 3599 _cg1r->threads_do(tc); 3600 } 3601 3602 void G1CollectedHeap::print_tracing_info() const { 3603 // We'll overload this to mean "trace GC pause statistics." 3604 if (TraceGen0Time || TraceGen1Time) { 3605 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3606 // to that. 3607 g1_policy()->print_tracing_info(); 3608 } 3609 if (G1SummarizeRSetStats) { 3610 g1_rem_set()->print_summary_info(); 3611 } 3612 if (G1SummarizeConcMark) { 3613 concurrent_mark()->print_summary_info(); 3614 } 3615 g1_policy()->print_yg_surv_rate_info(); 3616 SpecializationStats::print(); 3617 } 3618 3619 #ifndef PRODUCT 3620 // Helpful for debugging RSet issues. 3621 3622 class PrintRSetsClosure : public HeapRegionClosure { 3623 private: 3624 const char* _msg; 3625 size_t _occupied_sum; 3626 3627 public: 3628 bool doHeapRegion(HeapRegion* r) { 3629 HeapRegionRemSet* hrrs = r->rem_set(); 3630 size_t occupied = hrrs->occupied(); 3631 _occupied_sum += occupied; 3632 3633 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3634 HR_FORMAT_PARAMS(r)); 3635 if (occupied == 0) { 3636 gclog_or_tty->print_cr(" RSet is empty"); 3637 } else { 3638 hrrs->print(); 3639 } 3640 gclog_or_tty->print_cr("----------"); 3641 return false; 3642 } 3643 3644 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3645 gclog_or_tty->cr(); 3646 gclog_or_tty->print_cr("========================================"); 3647 gclog_or_tty->print_cr(msg); 3648 gclog_or_tty->cr(); 3649 } 3650 3651 ~PrintRSetsClosure() { 3652 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3653 gclog_or_tty->print_cr("========================================"); 3654 gclog_or_tty->cr(); 3655 } 3656 }; 3657 3658 void G1CollectedHeap::print_cset_rsets() { 3659 PrintRSetsClosure cl("Printing CSet RSets"); 3660 collection_set_iterate(&cl); 3661 } 3662 3663 void G1CollectedHeap::print_all_rsets() { 3664 PrintRSetsClosure cl("Printing All RSets");; 3665 heap_region_iterate(&cl); 3666 } 3667 #endif // PRODUCT 3668 3669 G1CollectedHeap* G1CollectedHeap::heap() { 3670 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3671 "not a garbage-first heap"); 3672 return _g1h; 3673 } 3674 3675 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3676 // always_do_update_barrier = false; 3677 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3678 // Fill TLAB's and such 3679 ensure_parsability(true); 3680 } 3681 3682 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3683 3684 if (G1SummarizeRSetStats && 3685 (G1SummarizeRSetStatsPeriod > 0) && 3686 // we are at the end of the GC. Total collections has already been increased. 3687 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) { 3688 g1_rem_set()->print_periodic_summary_info(); 3689 } 3690 3691 // FIXME: what is this about? 3692 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3693 // is set. 3694 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3695 "derived pointer present")); 3696 // always_do_update_barrier = true; 3697 3698 // We have just completed a GC. Update the soft reference 3699 // policy with the new heap occupancy 3700 Universe::update_heap_info_at_gc(); 3701 } 3702 3703 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3704 unsigned int gc_count_before, 3705 bool* succeeded) { 3706 assert_heap_not_locked_and_not_at_safepoint(); 3707 g1_policy()->record_stop_world_start(); 3708 VM_G1IncCollectionPause op(gc_count_before, 3709 word_size, 3710 false, /* should_initiate_conc_mark */ 3711 g1_policy()->max_pause_time_ms(), 3712 GCCause::_g1_inc_collection_pause); 3713 VMThread::execute(&op); 3714 3715 HeapWord* result = op.result(); 3716 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3717 assert(result == NULL || ret_succeeded, 3718 "the result should be NULL if the VM did not succeed"); 3719 *succeeded = ret_succeeded; 3720 3721 assert_heap_not_locked(); 3722 return result; 3723 } 3724 3725 void 3726 G1CollectedHeap::doConcurrentMark() { 3727 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3728 if (!_cmThread->in_progress()) { 3729 _cmThread->set_started(); 3730 CGC_lock->notify(); 3731 } 3732 } 3733 3734 size_t G1CollectedHeap::pending_card_num() { 3735 size_t extra_cards = 0; 3736 JavaThread *curr = Threads::first(); 3737 while (curr != NULL) { 3738 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3739 extra_cards += dcq.size(); 3740 curr = curr->next(); 3741 } 3742 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3743 size_t buffer_size = dcqs.buffer_size(); 3744 size_t buffer_num = dcqs.completed_buffers_num(); 3745 3746 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3747 // in bytes - not the number of 'entries'. We need to convert 3748 // into a number of cards. 3749 return (buffer_size * buffer_num + extra_cards) / oopSize; 3750 } 3751 3752 size_t G1CollectedHeap::cards_scanned() { 3753 return g1_rem_set()->cardsScanned(); 3754 } 3755 3756 void 3757 G1CollectedHeap::setup_surviving_young_words() { 3758 assert(_surviving_young_words == NULL, "pre-condition"); 3759 uint array_length = g1_policy()->young_cset_region_length(); 3760 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3761 if (_surviving_young_words == NULL) { 3762 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR, 3763 "Not enough space for young surv words summary."); 3764 } 3765 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3766 #ifdef ASSERT 3767 for (uint i = 0; i < array_length; ++i) { 3768 assert( _surviving_young_words[i] == 0, "memset above" ); 3769 } 3770 #endif // !ASSERT 3771 } 3772 3773 void 3774 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3775 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3776 uint array_length = g1_policy()->young_cset_region_length(); 3777 for (uint i = 0; i < array_length; ++i) { 3778 _surviving_young_words[i] += surv_young_words[i]; 3779 } 3780 } 3781 3782 void 3783 G1CollectedHeap::cleanup_surviving_young_words() { 3784 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3785 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3786 _surviving_young_words = NULL; 3787 } 3788 3789 #ifdef ASSERT 3790 class VerifyCSetClosure: public HeapRegionClosure { 3791 public: 3792 bool doHeapRegion(HeapRegion* hr) { 3793 // Here we check that the CSet region's RSet is ready for parallel 3794 // iteration. The fields that we'll verify are only manipulated 3795 // when the region is part of a CSet and is collected. Afterwards, 3796 // we reset these fields when we clear the region's RSet (when the 3797 // region is freed) so they are ready when the region is 3798 // re-allocated. The only exception to this is if there's an 3799 // evacuation failure and instead of freeing the region we leave 3800 // it in the heap. In that case, we reset these fields during 3801 // evacuation failure handling. 3802 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3803 3804 // Here's a good place to add any other checks we'd like to 3805 // perform on CSet regions. 3806 return false; 3807 } 3808 }; 3809 #endif // ASSERT 3810 3811 #if TASKQUEUE_STATS 3812 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3813 st->print_raw_cr("GC Task Stats"); 3814 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3815 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3816 } 3817 3818 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3819 print_taskqueue_stats_hdr(st); 3820 3821 TaskQueueStats totals; 3822 const int n = workers() != NULL ? workers()->total_workers() : 1; 3823 for (int i = 0; i < n; ++i) { 3824 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3825 totals += task_queue(i)->stats; 3826 } 3827 st->print_raw("tot "); totals.print(st); st->cr(); 3828 3829 DEBUG_ONLY(totals.verify()); 3830 } 3831 3832 void G1CollectedHeap::reset_taskqueue_stats() { 3833 const int n = workers() != NULL ? workers()->total_workers() : 1; 3834 for (int i = 0; i < n; ++i) { 3835 task_queue(i)->stats.reset(); 3836 } 3837 } 3838 #endif // TASKQUEUE_STATS 3839 3840 void G1CollectedHeap::log_gc_header() { 3841 if (!G1Log::fine()) { 3842 return; 3843 } 3844 3845 gclog_or_tty->date_stamp(PrintGCDateStamps); 3846 gclog_or_tty->stamp(PrintGCTimeStamps); 3847 3848 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3849 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3850 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3851 3852 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3853 } 3854 3855 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3856 if (!G1Log::fine()) { 3857 return; 3858 } 3859 3860 if (G1Log::finer()) { 3861 if (evacuation_failed()) { 3862 gclog_or_tty->print(" (to-space exhausted)"); 3863 } 3864 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3865 g1_policy()->phase_times()->note_gc_end(); 3866 g1_policy()->phase_times()->print(pause_time_sec); 3867 g1_policy()->print_detailed_heap_transition(); 3868 } else { 3869 if (evacuation_failed()) { 3870 gclog_or_tty->print("--"); 3871 } 3872 g1_policy()->print_heap_transition(); 3873 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3874 } 3875 gclog_or_tty->flush(); 3876 } 3877 3878 bool 3879 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3880 assert_at_safepoint(true /* should_be_vm_thread */); 3881 guarantee(!is_gc_active(), "collection is not reentrant"); 3882 3883 if (GC_locker::check_active_before_gc()) { 3884 return false; 3885 } 3886 3887 _gc_timer_stw->register_gc_start(os::elapsed_counter()); 3888 3889 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start()); 3890 3891 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3892 ResourceMark rm; 3893 3894 print_heap_before_gc(); 3895 trace_heap_before_gc(_gc_tracer_stw); 3896 3897 HRSPhaseSetter x(HRSPhaseEvacuation); 3898 verify_region_sets_optional(); 3899 verify_dirty_young_regions(); 3900 3901 // This call will decide whether this pause is an initial-mark 3902 // pause. If it is, during_initial_mark_pause() will return true 3903 // for the duration of this pause. 3904 g1_policy()->decide_on_conc_mark_initiation(); 3905 3906 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3907 assert(!g1_policy()->during_initial_mark_pause() || 3908 g1_policy()->gcs_are_young(), "sanity"); 3909 3910 // We also do not allow mixed GCs during marking. 3911 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3912 3913 // Record whether this pause is an initial mark. When the current 3914 // thread has completed its logging output and it's safe to signal 3915 // the CM thread, the flag's value in the policy has been reset. 3916 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3917 3918 // Inner scope for scope based logging, timers, and stats collection 3919 { 3920 EvacuationInfo evacuation_info; 3921 3922 if (g1_policy()->during_initial_mark_pause()) { 3923 // We are about to start a marking cycle, so we increment the 3924 // full collection counter. 3925 increment_old_marking_cycles_started(); 3926 register_concurrent_cycle_start(_gc_timer_stw->gc_start()); 3927 } 3928 3929 _gc_tracer_stw->report_yc_type(yc_type()); 3930 3931 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3932 3933 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3934 workers()->active_workers() : 1); 3935 double pause_start_sec = os::elapsedTime(); 3936 g1_policy()->phase_times()->note_gc_start(active_workers); 3937 log_gc_header(); 3938 3939 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3940 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3941 3942 // If the secondary_free_list is not empty, append it to the 3943 // free_list. No need to wait for the cleanup operation to finish; 3944 // the region allocation code will check the secondary_free_list 3945 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3946 // set, skip this step so that the region allocation code has to 3947 // get entries from the secondary_free_list. 3948 if (!G1StressConcRegionFreeing) { 3949 append_secondary_free_list_if_not_empty_with_lock(); 3950 } 3951 3952 assert(check_young_list_well_formed(), "young list should be well formed"); 3953 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3954 "sanity check"); 3955 3956 // Don't dynamically change the number of GC threads this early. A value of 3957 // 0 is used to indicate serial work. When parallel work is done, 3958 // it will be set. 3959 3960 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3961 IsGCActiveMark x; 3962 3963 gc_prologue(false); 3964 increment_total_collections(false /* full gc */); 3965 increment_gc_time_stamp(); 3966 3967 verify_before_gc(); 3968 3969 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3970 3971 // Please see comment in g1CollectedHeap.hpp and 3972 // G1CollectedHeap::ref_processing_init() to see how 3973 // reference processing currently works in G1. 3974 3975 // Enable discovery in the STW reference processor 3976 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3977 true /*verify_no_refs*/); 3978 3979 { 3980 // We want to temporarily turn off discovery by the 3981 // CM ref processor, if necessary, and turn it back on 3982 // on again later if we do. Using a scoped 3983 // NoRefDiscovery object will do this. 3984 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3985 3986 // Forget the current alloc region (we might even choose it to be part 3987 // of the collection set!). 3988 release_mutator_alloc_region(); 3989 3990 // We should call this after we retire the mutator alloc 3991 // region(s) so that all the ALLOC / RETIRE events are generated 3992 // before the start GC event. 3993 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3994 3995 // This timing is only used by the ergonomics to handle our pause target. 3996 // It is unclear why this should not include the full pause. We will 3997 // investigate this in CR 7178365. 3998 // 3999 // Preserving the old comment here if that helps the investigation: 4000 // 4001 // The elapsed time induced by the start time below deliberately elides 4002 // the possible verification above. 4003 double sample_start_time_sec = os::elapsedTime(); 4004 4005 #if YOUNG_LIST_VERBOSE 4006 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 4007 _young_list->print(); 4008 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4009 #endif // YOUNG_LIST_VERBOSE 4010 4011 g1_policy()->record_collection_pause_start(sample_start_time_sec); 4012 4013 double scan_wait_start = os::elapsedTime(); 4014 // We have to wait until the CM threads finish scanning the 4015 // root regions as it's the only way to ensure that all the 4016 // objects on them have been correctly scanned before we start 4017 // moving them during the GC. 4018 bool waited = _cm->root_regions()->wait_until_scan_finished(); 4019 double wait_time_ms = 0.0; 4020 if (waited) { 4021 double scan_wait_end = os::elapsedTime(); 4022 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 4023 } 4024 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 4025 4026 #if YOUNG_LIST_VERBOSE 4027 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 4028 _young_list->print(); 4029 #endif // YOUNG_LIST_VERBOSE 4030 4031 if (g1_policy()->during_initial_mark_pause()) { 4032 concurrent_mark()->checkpointRootsInitialPre(); 4033 } 4034 4035 #if YOUNG_LIST_VERBOSE 4036 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 4037 _young_list->print(); 4038 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4039 #endif // YOUNG_LIST_VERBOSE 4040 4041 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); 4042 4043 _cm->note_start_of_gc(); 4044 // We should not verify the per-thread SATB buffers given that 4045 // we have not filtered them yet (we'll do so during the 4046 // GC). We also call this after finalize_cset() to 4047 // ensure that the CSet has been finalized. 4048 _cm->verify_no_cset_oops(true /* verify_stacks */, 4049 true /* verify_enqueued_buffers */, 4050 false /* verify_thread_buffers */, 4051 true /* verify_fingers */); 4052 4053 if (_hr_printer.is_active()) { 4054 HeapRegion* hr = g1_policy()->collection_set(); 4055 while (hr != NULL) { 4056 G1HRPrinter::RegionType type; 4057 if (!hr->is_young()) { 4058 type = G1HRPrinter::Old; 4059 } else if (hr->is_survivor()) { 4060 type = G1HRPrinter::Survivor; 4061 } else { 4062 type = G1HRPrinter::Eden; 4063 } 4064 _hr_printer.cset(hr); 4065 hr = hr->next_in_collection_set(); 4066 } 4067 } 4068 4069 #ifdef ASSERT 4070 VerifyCSetClosure cl; 4071 collection_set_iterate(&cl); 4072 #endif // ASSERT 4073 4074 setup_surviving_young_words(); 4075 4076 // Initialize the GC alloc regions. 4077 init_gc_alloc_regions(evacuation_info); 4078 4079 // Actually do the work... 4080 evacuate_collection_set(evacuation_info); 4081 4082 // We do this to mainly verify the per-thread SATB buffers 4083 // (which have been filtered by now) since we didn't verify 4084 // them earlier. No point in re-checking the stacks / enqueued 4085 // buffers given that the CSet has not changed since last time 4086 // we checked. 4087 _cm->verify_no_cset_oops(false /* verify_stacks */, 4088 false /* verify_enqueued_buffers */, 4089 true /* verify_thread_buffers */, 4090 true /* verify_fingers */); 4091 4092 free_collection_set(g1_policy()->collection_set(), evacuation_info); 4093 g1_policy()->clear_collection_set(); 4094 4095 cleanup_surviving_young_words(); 4096 4097 // Start a new incremental collection set for the next pause. 4098 g1_policy()->start_incremental_cset_building(); 4099 4100 // Clear the _cset_fast_test bitmap in anticipation of adding 4101 // regions to the incremental collection set for the next 4102 // evacuation pause. 4103 clear_cset_fast_test(); 4104 4105 _young_list->reset_sampled_info(); 4106 4107 // Don't check the whole heap at this point as the 4108 // GC alloc regions from this pause have been tagged 4109 // as survivors and moved on to the survivor list. 4110 // Survivor regions will fail the !is_young() check. 4111 assert(check_young_list_empty(false /* check_heap */), 4112 "young list should be empty"); 4113 4114 #if YOUNG_LIST_VERBOSE 4115 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 4116 _young_list->print(); 4117 #endif // YOUNG_LIST_VERBOSE 4118 4119 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 4120 _young_list->first_survivor_region(), 4121 _young_list->last_survivor_region()); 4122 4123 _young_list->reset_auxilary_lists(); 4124 4125 if (evacuation_failed()) { 4126 _summary_bytes_used = recalculate_used(); 4127 uint n_queues = MAX2((int)ParallelGCThreads, 1); 4128 for (uint i = 0; i < n_queues; i++) { 4129 if (_evacuation_failed_info_array[i].has_failed()) { 4130 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]); 4131 } 4132 } 4133 } else { 4134 // The "used" of the the collection set have already been subtracted 4135 // when they were freed. Add in the bytes evacuated. 4136 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 4137 } 4138 4139 if (g1_policy()->during_initial_mark_pause()) { 4140 // We have to do this before we notify the CM threads that 4141 // they can start working to make sure that all the 4142 // appropriate initialization is done on the CM object. 4143 concurrent_mark()->checkpointRootsInitialPost(); 4144 set_marking_started(); 4145 // Note that we don't actually trigger the CM thread at 4146 // this point. We do that later when we're sure that 4147 // the current thread has completed its logging output. 4148 } 4149 4150 allocate_dummy_regions(); 4151 4152 #if YOUNG_LIST_VERBOSE 4153 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 4154 _young_list->print(); 4155 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4156 #endif // YOUNG_LIST_VERBOSE 4157 4158 init_mutator_alloc_region(); 4159 4160 { 4161 size_t expand_bytes = g1_policy()->expansion_amount(); 4162 if (expand_bytes > 0) { 4163 size_t bytes_before = capacity(); 4164 // No need for an ergo verbose message here, 4165 // expansion_amount() does this when it returns a value > 0. 4166 if (!expand(expand_bytes)) { 4167 // We failed to expand the heap so let's verify that 4168 // committed/uncommitted amount match the backing store 4169 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 4170 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 4171 } 4172 } 4173 } 4174 4175 // We redo the verification but now wrt to the new CSet which 4176 // has just got initialized after the previous CSet was freed. 4177 _cm->verify_no_cset_oops(true /* verify_stacks */, 4178 true /* verify_enqueued_buffers */, 4179 true /* verify_thread_buffers */, 4180 true /* verify_fingers */); 4181 _cm->note_end_of_gc(); 4182 4183 // This timing is only used by the ergonomics to handle our pause target. 4184 // It is unclear why this should not include the full pause. We will 4185 // investigate this in CR 7178365. 4186 double sample_end_time_sec = os::elapsedTime(); 4187 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 4188 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info); 4189 4190 MemoryService::track_memory_usage(); 4191 4192 // In prepare_for_verify() below we'll need to scan the deferred 4193 // update buffers to bring the RSets up-to-date if 4194 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4195 // the update buffers we'll probably need to scan cards on the 4196 // regions we just allocated to (i.e., the GC alloc 4197 // regions). However, during the last GC we called 4198 // set_saved_mark() on all the GC alloc regions, so card 4199 // scanning might skip the [saved_mark_word()...top()] area of 4200 // those regions (i.e., the area we allocated objects into 4201 // during the last GC). But it shouldn't. Given that 4202 // saved_mark_word() is conditional on whether the GC time stamp 4203 // on the region is current or not, by incrementing the GC time 4204 // stamp here we invalidate all the GC time stamps on all the 4205 // regions and saved_mark_word() will simply return top() for 4206 // all the regions. This is a nicer way of ensuring this rather 4207 // than iterating over the regions and fixing them. In fact, the 4208 // GC time stamp increment here also ensures that 4209 // saved_mark_word() will return top() between pauses, i.e., 4210 // during concurrent refinement. So we don't need the 4211 // is_gc_active() check to decided which top to use when 4212 // scanning cards (see CR 7039627). 4213 increment_gc_time_stamp(); 4214 4215 verify_after_gc(); 4216 4217 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4218 ref_processor_stw()->verify_no_references_recorded(); 4219 4220 // CM reference discovery will be re-enabled if necessary. 4221 } 4222 4223 // We should do this after we potentially expand the heap so 4224 // that all the COMMIT events are generated before the end GC 4225 // event, and after we retire the GC alloc regions so that all 4226 // RETIRE events are generated before the end GC event. 4227 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4228 4229 if (mark_in_progress()) { 4230 concurrent_mark()->update_g1_committed(); 4231 } 4232 4233 #ifdef TRACESPINNING 4234 ParallelTaskTerminator::print_termination_counts(); 4235 #endif 4236 4237 gc_epilogue(false); 4238 } 4239 4240 // Print the remainder of the GC log output. 4241 log_gc_footer(os::elapsedTime() - pause_start_sec); 4242 4243 // It is not yet to safe to tell the concurrent mark to 4244 // start as we have some optional output below. We don't want the 4245 // output from the concurrent mark thread interfering with this 4246 // logging output either. 4247 4248 _hrs.verify_optional(); 4249 verify_region_sets_optional(); 4250 4251 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4252 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4253 4254 print_heap_after_gc(); 4255 trace_heap_after_gc(_gc_tracer_stw); 4256 4257 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4258 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4259 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4260 // before any GC notifications are raised. 4261 g1mm()->update_sizes(); 4262 4263 _gc_tracer_stw->report_evacuation_info(&evacuation_info); 4264 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold()); 4265 _gc_timer_stw->register_gc_end(os::elapsed_counter()); 4266 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions()); 4267 } 4268 // It should now be safe to tell the concurrent mark thread to start 4269 // without its logging output interfering with the logging output 4270 // that came from the pause. 4271 4272 if (should_start_conc_mark) { 4273 // CAUTION: after the doConcurrentMark() call below, 4274 // the concurrent marking thread(s) could be running 4275 // concurrently with us. Make sure that anything after 4276 // this point does not assume that we are the only GC thread 4277 // running. Note: of course, the actual marking work will 4278 // not start until the safepoint itself is released in 4279 // ConcurrentGCThread::safepoint_desynchronize(). 4280 doConcurrentMark(); 4281 } 4282 4283 return true; 4284 } 4285 4286 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4287 { 4288 size_t gclab_word_size; 4289 switch (purpose) { 4290 case GCAllocForSurvived: 4291 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4292 break; 4293 case GCAllocForTenured: 4294 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4295 break; 4296 default: 4297 assert(false, "unknown GCAllocPurpose"); 4298 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4299 break; 4300 } 4301 4302 // Prevent humongous PLAB sizes for two reasons: 4303 // * PLABs are allocated using a similar paths as oops, but should 4304 // never be in a humongous region 4305 // * Allowing humongous PLABs needlessly churns the region free lists 4306 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4307 } 4308 4309 void G1CollectedHeap::init_mutator_alloc_region() { 4310 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4311 _mutator_alloc_region.init(); 4312 } 4313 4314 void G1CollectedHeap::release_mutator_alloc_region() { 4315 _mutator_alloc_region.release(); 4316 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4317 } 4318 4319 void G1CollectedHeap::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { 4320 assert_at_safepoint(true /* should_be_vm_thread */); 4321 4322 _survivor_gc_alloc_region.init(); 4323 _old_gc_alloc_region.init(); 4324 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4325 _retained_old_gc_alloc_region = NULL; 4326 4327 // We will discard the current GC alloc region if: 4328 // a) it's in the collection set (it can happen!), 4329 // b) it's already full (no point in using it), 4330 // c) it's empty (this means that it was emptied during 4331 // a cleanup and it should be on the free list now), or 4332 // d) it's humongous (this means that it was emptied 4333 // during a cleanup and was added to the free list, but 4334 // has been subsequently used to allocate a humongous 4335 // object that may be less than the region size). 4336 if (retained_region != NULL && 4337 !retained_region->in_collection_set() && 4338 !(retained_region->top() == retained_region->end()) && 4339 !retained_region->is_empty() && 4340 !retained_region->isHumongous()) { 4341 retained_region->set_saved_mark(); 4342 // The retained region was added to the old region set when it was 4343 // retired. We have to remove it now, since we don't allow regions 4344 // we allocate to in the region sets. We'll re-add it later, when 4345 // it's retired again. 4346 _old_set.remove(retained_region); 4347 bool during_im = g1_policy()->during_initial_mark_pause(); 4348 retained_region->note_start_of_copying(during_im); 4349 _old_gc_alloc_region.set(retained_region); 4350 _hr_printer.reuse(retained_region); 4351 evacuation_info.set_alloc_regions_used_before(retained_region->used()); 4352 } 4353 } 4354 4355 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { 4356 evacuation_info.set_allocation_regions(_survivor_gc_alloc_region.count() + 4357 _old_gc_alloc_region.count()); 4358 _survivor_gc_alloc_region.release(); 4359 // If we have an old GC alloc region to release, we'll save it in 4360 // _retained_old_gc_alloc_region. If we don't 4361 // _retained_old_gc_alloc_region will become NULL. This is what we 4362 // want either way so no reason to check explicitly for either 4363 // condition. 4364 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4365 4366 if (ResizePLAB) { 4367 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4368 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4369 } 4370 } 4371 4372 void G1CollectedHeap::abandon_gc_alloc_regions() { 4373 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4374 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4375 _retained_old_gc_alloc_region = NULL; 4376 } 4377 4378 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4379 _drain_in_progress = false; 4380 set_evac_failure_closure(cl); 4381 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4382 } 4383 4384 void G1CollectedHeap::finalize_for_evac_failure() { 4385 assert(_evac_failure_scan_stack != NULL && 4386 _evac_failure_scan_stack->length() == 0, 4387 "Postcondition"); 4388 assert(!_drain_in_progress, "Postcondition"); 4389 delete _evac_failure_scan_stack; 4390 _evac_failure_scan_stack = NULL; 4391 } 4392 4393 void G1CollectedHeap::remove_self_forwarding_pointers() { 4394 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4395 4396 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4397 4398 if (G1CollectedHeap::use_parallel_gc_threads()) { 4399 set_par_threads(); 4400 workers()->run_task(&rsfp_task); 4401 set_par_threads(0); 4402 } else { 4403 rsfp_task.work(0); 4404 } 4405 4406 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4407 4408 // Reset the claim values in the regions in the collection set. 4409 reset_cset_heap_region_claim_values(); 4410 4411 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4412 4413 // Now restore saved marks, if any. 4414 assert(_objs_with_preserved_marks.size() == 4415 _preserved_marks_of_objs.size(), "Both or none."); 4416 while (!_objs_with_preserved_marks.is_empty()) { 4417 oop obj = _objs_with_preserved_marks.pop(); 4418 markOop m = _preserved_marks_of_objs.pop(); 4419 obj->set_mark(m); 4420 } 4421 _objs_with_preserved_marks.clear(true); 4422 _preserved_marks_of_objs.clear(true); 4423 } 4424 4425 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4426 _evac_failure_scan_stack->push(obj); 4427 } 4428 4429 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4430 assert(_evac_failure_scan_stack != NULL, "precondition"); 4431 4432 while (_evac_failure_scan_stack->length() > 0) { 4433 oop obj = _evac_failure_scan_stack->pop(); 4434 _evac_failure_closure->set_region(heap_region_containing(obj)); 4435 obj->oop_iterate_backwards(_evac_failure_closure); 4436 } 4437 } 4438 4439 oop 4440 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state, 4441 oop old) { 4442 assert(obj_in_cs(old), 4443 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4444 (HeapWord*) old)); 4445 markOop m = old->mark(); 4446 oop forward_ptr = old->forward_to_atomic(old); 4447 if (forward_ptr == NULL) { 4448 // Forward-to-self succeeded. 4449 assert(_par_scan_state != NULL, "par scan state"); 4450 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4451 uint queue_num = _par_scan_state->queue_num(); 4452 4453 _evacuation_failed = true; 4454 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size()); 4455 if (_evac_failure_closure != cl) { 4456 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4457 assert(!_drain_in_progress, 4458 "Should only be true while someone holds the lock."); 4459 // Set the global evac-failure closure to the current thread's. 4460 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4461 set_evac_failure_closure(cl); 4462 // Now do the common part. 4463 handle_evacuation_failure_common(old, m); 4464 // Reset to NULL. 4465 set_evac_failure_closure(NULL); 4466 } else { 4467 // The lock is already held, and this is recursive. 4468 assert(_drain_in_progress, "This should only be the recursive case."); 4469 handle_evacuation_failure_common(old, m); 4470 } 4471 return old; 4472 } else { 4473 // Forward-to-self failed. Either someone else managed to allocate 4474 // space for this object (old != forward_ptr) or they beat us in 4475 // self-forwarding it (old == forward_ptr). 4476 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4477 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4478 "should not be in the CSet", 4479 (HeapWord*) old, (HeapWord*) forward_ptr)); 4480 return forward_ptr; 4481 } 4482 } 4483 4484 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4485 preserve_mark_if_necessary(old, m); 4486 4487 HeapRegion* r = heap_region_containing(old); 4488 if (!r->evacuation_failed()) { 4489 r->set_evacuation_failed(true); 4490 _hr_printer.evac_failure(r); 4491 } 4492 4493 push_on_evac_failure_scan_stack(old); 4494 4495 if (!_drain_in_progress) { 4496 // prevent recursion in copy_to_survivor_space() 4497 _drain_in_progress = true; 4498 drain_evac_failure_scan_stack(); 4499 _drain_in_progress = false; 4500 } 4501 } 4502 4503 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4504 assert(evacuation_failed(), "Oversaving!"); 4505 // We want to call the "for_promotion_failure" version only in the 4506 // case of a promotion failure. 4507 if (m->must_be_preserved_for_promotion_failure(obj)) { 4508 _objs_with_preserved_marks.push(obj); 4509 _preserved_marks_of_objs.push(m); 4510 } 4511 } 4512 4513 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4514 size_t word_size) { 4515 if (purpose == GCAllocForSurvived) { 4516 HeapWord* result = survivor_attempt_allocation(word_size); 4517 if (result != NULL) { 4518 return result; 4519 } else { 4520 // Let's try to allocate in the old gen in case we can fit the 4521 // object there. 4522 return old_attempt_allocation(word_size); 4523 } 4524 } else { 4525 assert(purpose == GCAllocForTenured, "sanity"); 4526 HeapWord* result = old_attempt_allocation(word_size); 4527 if (result != NULL) { 4528 return result; 4529 } else { 4530 // Let's try to allocate in the survivors in case we can fit the 4531 // object there. 4532 return survivor_attempt_allocation(word_size); 4533 } 4534 } 4535 4536 ShouldNotReachHere(); 4537 // Trying to keep some compilers happy. 4538 return NULL; 4539 } 4540 4541 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4542 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4543 4544 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4545 : _g1h(g1h), 4546 _refs(g1h->task_queue(queue_num)), 4547 _dcq(&g1h->dirty_card_queue_set()), 4548 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4549 _g1_rem(g1h->g1_rem_set()), 4550 _hash_seed(17), _queue_num(queue_num), 4551 _term_attempts(0), 4552 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4553 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4554 _age_table(false), 4555 _strong_roots_time(0), _term_time(0), 4556 _alloc_buffer_waste(0), _undo_waste(0) { 4557 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4558 // we "sacrifice" entry 0 to keep track of surviving bytes for 4559 // non-young regions (where the age is -1) 4560 // We also add a few elements at the beginning and at the end in 4561 // an attempt to eliminate cache contention 4562 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4563 uint array_length = PADDING_ELEM_NUM + 4564 real_length + 4565 PADDING_ELEM_NUM; 4566 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4567 if (_surviving_young_words_base == NULL) 4568 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR, 4569 "Not enough space for young surv histo."); 4570 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4571 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4572 4573 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4574 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4575 4576 _start = os::elapsedTime(); 4577 } 4578 4579 void 4580 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4581 { 4582 st->print_raw_cr("GC Termination Stats"); 4583 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4584 " ------waste (KiB)------"); 4585 st->print_raw_cr("thr ms ms % ms % attempts" 4586 " total alloc undo"); 4587 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4588 " ------- ------- -------"); 4589 } 4590 4591 void 4592 G1ParScanThreadState::print_termination_stats(int i, 4593 outputStream* const st) const 4594 { 4595 const double elapsed_ms = elapsed_time() * 1000.0; 4596 const double s_roots_ms = strong_roots_time() * 1000.0; 4597 const double term_ms = term_time() * 1000.0; 4598 st->print_cr("%3d %9.2f %9.2f %6.2f " 4599 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4600 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4601 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4602 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4603 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4604 alloc_buffer_waste() * HeapWordSize / K, 4605 undo_waste() * HeapWordSize / K); 4606 } 4607 4608 #ifdef ASSERT 4609 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4610 assert(ref != NULL, "invariant"); 4611 assert(UseCompressedOops, "sanity"); 4612 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4613 oop p = oopDesc::load_decode_heap_oop(ref); 4614 assert(_g1h->is_in_g1_reserved(p), 4615 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4616 return true; 4617 } 4618 4619 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4620 assert(ref != NULL, "invariant"); 4621 if (has_partial_array_mask(ref)) { 4622 // Must be in the collection set--it's already been copied. 4623 oop p = clear_partial_array_mask(ref); 4624 assert(_g1h->obj_in_cs(p), 4625 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4626 } else { 4627 oop p = oopDesc::load_decode_heap_oop(ref); 4628 assert(_g1h->is_in_g1_reserved(p), 4629 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4630 } 4631 return true; 4632 } 4633 4634 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4635 if (ref.is_narrow()) { 4636 return verify_ref((narrowOop*) ref); 4637 } else { 4638 return verify_ref((oop*) ref); 4639 } 4640 } 4641 #endif // ASSERT 4642 4643 void G1ParScanThreadState::trim_queue() { 4644 assert(_evac_cl != NULL, "not set"); 4645 assert(_evac_failure_cl != NULL, "not set"); 4646 assert(_partial_scan_cl != NULL, "not set"); 4647 4648 StarTask ref; 4649 do { 4650 // Drain the overflow stack first, so other threads can steal. 4651 while (refs()->pop_overflow(ref)) { 4652 deal_with_reference(ref); 4653 } 4654 4655 while (refs()->pop_local(ref)) { 4656 deal_with_reference(ref); 4657 } 4658 } while (!refs()->is_empty()); 4659 } 4660 4661 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4662 G1ParScanThreadState* par_scan_state) : 4663 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4664 _par_scan_state(par_scan_state), 4665 _worker_id(par_scan_state->queue_num()), 4666 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4667 _mark_in_progress(_g1->mark_in_progress()) { } 4668 4669 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4670 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4671 #ifdef ASSERT 4672 HeapRegion* hr = _g1->heap_region_containing(obj); 4673 assert(hr != NULL, "sanity"); 4674 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4675 #endif // ASSERT 4676 4677 // We know that the object is not moving so it's safe to read its size. 4678 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4679 } 4680 4681 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4682 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4683 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4684 #ifdef ASSERT 4685 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4686 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4687 assert(from_obj != to_obj, "should not be self-forwarded"); 4688 4689 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4690 assert(from_hr != NULL, "sanity"); 4691 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4692 4693 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4694 assert(to_hr != NULL, "sanity"); 4695 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4696 #endif // ASSERT 4697 4698 // The object might be in the process of being copied by another 4699 // worker so we cannot trust that its to-space image is 4700 // well-formed. So we have to read its size from its from-space 4701 // image which we know should not be changing. 4702 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4703 } 4704 4705 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4706 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4707 ::copy_to_survivor_space(oop old) { 4708 size_t word_sz = old->size(); 4709 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4710 // +1 to make the -1 indexes valid... 4711 int young_index = from_region->young_index_in_cset()+1; 4712 assert( (from_region->is_young() && young_index > 0) || 4713 (!from_region->is_young() && young_index == 0), "invariant" ); 4714 G1CollectorPolicy* g1p = _g1->g1_policy(); 4715 markOop m = old->mark(); 4716 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4717 : m->age(); 4718 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4719 word_sz); 4720 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4721 #ifndef PRODUCT 4722 // Should this evacuation fail? 4723 if (_g1->evacuation_should_fail()) { 4724 if (obj_ptr != NULL) { 4725 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4726 obj_ptr = NULL; 4727 } 4728 } 4729 #endif // !PRODUCT 4730 4731 if (obj_ptr == NULL) { 4732 // This will either forward-to-self, or detect that someone else has 4733 // installed a forwarding pointer. 4734 return _g1->handle_evacuation_failure_par(_par_scan_state, old); 4735 } 4736 4737 oop obj = oop(obj_ptr); 4738 4739 // We're going to allocate linearly, so might as well prefetch ahead. 4740 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4741 4742 oop forward_ptr = old->forward_to_atomic(obj); 4743 if (forward_ptr == NULL) { 4744 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4745 if (g1p->track_object_age(alloc_purpose)) { 4746 // We could simply do obj->incr_age(). However, this causes a 4747 // performance issue. obj->incr_age() will first check whether 4748 // the object has a displaced mark by checking its mark word; 4749 // getting the mark word from the new location of the object 4750 // stalls. So, given that we already have the mark word and we 4751 // are about to install it anyway, it's better to increase the 4752 // age on the mark word, when the object does not have a 4753 // displaced mark word. We're not expecting many objects to have 4754 // a displaced marked word, so that case is not optimized 4755 // further (it could be...) and we simply call obj->incr_age(). 4756 4757 if (m->has_displaced_mark_helper()) { 4758 // in this case, we have to install the mark word first, 4759 // otherwise obj looks to be forwarded (the old mark word, 4760 // which contains the forward pointer, was copied) 4761 obj->set_mark(m); 4762 obj->incr_age(); 4763 } else { 4764 m = m->incr_age(); 4765 obj->set_mark(m); 4766 } 4767 _par_scan_state->age_table()->add(obj, word_sz); 4768 } else { 4769 obj->set_mark(m); 4770 } 4771 4772 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4773 surv_young_words[young_index] += word_sz; 4774 4775 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4776 // We keep track of the next start index in the length field of 4777 // the to-space object. The actual length can be found in the 4778 // length field of the from-space object. 4779 arrayOop(obj)->set_length(0); 4780 oop* old_p = set_partial_array_mask(old); 4781 _par_scan_state->push_on_queue(old_p); 4782 } else { 4783 // No point in using the slower heap_region_containing() method, 4784 // given that we know obj is in the heap. 4785 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4786 obj->oop_iterate_backwards(&_scanner); 4787 } 4788 } else { 4789 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4790 obj = forward_ptr; 4791 } 4792 return obj; 4793 } 4794 4795 template <class T> 4796 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4797 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4798 _scanned_klass->record_modified_oops(); 4799 } 4800 } 4801 4802 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4803 template <class T> 4804 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4805 ::do_oop_work(T* p) { 4806 oop obj = oopDesc::load_decode_heap_oop(p); 4807 assert(barrier != G1BarrierRS || obj != NULL, 4808 "Precondition: G1BarrierRS implies obj is non-NULL"); 4809 4810 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4811 4812 // here the null check is implicit in the cset_fast_test() test 4813 if (_g1->in_cset_fast_test(obj)) { 4814 oop forwardee; 4815 if (obj->is_forwarded()) { 4816 forwardee = obj->forwardee(); 4817 } else { 4818 forwardee = copy_to_survivor_space(obj); 4819 } 4820 assert(forwardee != NULL, "forwardee should not be NULL"); 4821 oopDesc::encode_store_heap_oop(p, forwardee); 4822 if (do_mark_object && forwardee != obj) { 4823 // If the object is self-forwarded we don't need to explicitly 4824 // mark it, the evacuation failure protocol will do so. 4825 mark_forwarded_object(obj, forwardee); 4826 } 4827 4828 // When scanning the RS, we only care about objs in CS. 4829 if (barrier == G1BarrierRS) { 4830 _par_scan_state->update_rs(_from, p, _worker_id); 4831 } else if (barrier == G1BarrierKlass) { 4832 do_klass_barrier(p, forwardee); 4833 } 4834 } else { 4835 // The object is not in collection set. If we're a root scanning 4836 // closure during an initial mark pause (i.e. do_mark_object will 4837 // be true) then attempt to mark the object. 4838 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4839 mark_object(obj); 4840 } 4841 } 4842 4843 if (barrier == G1BarrierEvac && obj != NULL) { 4844 _par_scan_state->update_rs(_from, p, _worker_id); 4845 } 4846 4847 if (do_gen_barrier && obj != NULL) { 4848 par_do_barrier(p); 4849 } 4850 } 4851 4852 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4853 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4854 4855 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4856 assert(has_partial_array_mask(p), "invariant"); 4857 oop from_obj = clear_partial_array_mask(p); 4858 4859 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4860 assert(from_obj->is_objArray(), "must be obj array"); 4861 objArrayOop from_obj_array = objArrayOop(from_obj); 4862 // The from-space object contains the real length. 4863 int length = from_obj_array->length(); 4864 4865 assert(from_obj->is_forwarded(), "must be forwarded"); 4866 oop to_obj = from_obj->forwardee(); 4867 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4868 objArrayOop to_obj_array = objArrayOop(to_obj); 4869 // We keep track of the next start index in the length field of the 4870 // to-space object. 4871 int next_index = to_obj_array->length(); 4872 assert(0 <= next_index && next_index < length, 4873 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4874 4875 int start = next_index; 4876 int end = length; 4877 int remainder = end - start; 4878 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4879 if (remainder > 2 * ParGCArrayScanChunk) { 4880 end = start + ParGCArrayScanChunk; 4881 to_obj_array->set_length(end); 4882 // Push the remainder before we process the range in case another 4883 // worker has run out of things to do and can steal it. 4884 oop* from_obj_p = set_partial_array_mask(from_obj); 4885 _par_scan_state->push_on_queue(from_obj_p); 4886 } else { 4887 assert(length == end, "sanity"); 4888 // We'll process the final range for this object. Restore the length 4889 // so that the heap remains parsable in case of evacuation failure. 4890 to_obj_array->set_length(end); 4891 } 4892 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4893 // Process indexes [start,end). It will also process the header 4894 // along with the first chunk (i.e., the chunk with start == 0). 4895 // Note that at this point the length field of to_obj_array is not 4896 // correct given that we are using it to keep track of the next 4897 // start index. oop_iterate_range() (thankfully!) ignores the length 4898 // field and only relies on the start / end parameters. It does 4899 // however return the size of the object which will be incorrect. So 4900 // we have to ignore it even if we wanted to use it. 4901 to_obj_array->oop_iterate_range(&_scanner, start, end); 4902 } 4903 4904 class G1ParEvacuateFollowersClosure : public VoidClosure { 4905 protected: 4906 G1CollectedHeap* _g1h; 4907 G1ParScanThreadState* _par_scan_state; 4908 RefToScanQueueSet* _queues; 4909 ParallelTaskTerminator* _terminator; 4910 4911 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4912 RefToScanQueueSet* queues() { return _queues; } 4913 ParallelTaskTerminator* terminator() { return _terminator; } 4914 4915 public: 4916 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4917 G1ParScanThreadState* par_scan_state, 4918 RefToScanQueueSet* queues, 4919 ParallelTaskTerminator* terminator) 4920 : _g1h(g1h), _par_scan_state(par_scan_state), 4921 _queues(queues), _terminator(terminator) {} 4922 4923 void do_void(); 4924 4925 private: 4926 inline bool offer_termination(); 4927 }; 4928 4929 bool G1ParEvacuateFollowersClosure::offer_termination() { 4930 G1ParScanThreadState* const pss = par_scan_state(); 4931 pss->start_term_time(); 4932 const bool res = terminator()->offer_termination(); 4933 pss->end_term_time(); 4934 return res; 4935 } 4936 4937 void G1ParEvacuateFollowersClosure::do_void() { 4938 StarTask stolen_task; 4939 G1ParScanThreadState* const pss = par_scan_state(); 4940 pss->trim_queue(); 4941 4942 do { 4943 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4944 assert(pss->verify_task(stolen_task), "sanity"); 4945 if (stolen_task.is_narrow()) { 4946 pss->deal_with_reference((narrowOop*) stolen_task); 4947 } else { 4948 pss->deal_with_reference((oop*) stolen_task); 4949 } 4950 4951 // We've just processed a reference and we might have made 4952 // available new entries on the queues. So we have to make sure 4953 // we drain the queues as necessary. 4954 pss->trim_queue(); 4955 } 4956 } while (!offer_termination()); 4957 4958 pss->retire_alloc_buffers(); 4959 } 4960 4961 class G1KlassScanClosure : public KlassClosure { 4962 G1ParCopyHelper* _closure; 4963 bool _process_only_dirty; 4964 int _count; 4965 public: 4966 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4967 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4968 void do_klass(Klass* klass) { 4969 // If the klass has not been dirtied we know that there's 4970 // no references into the young gen and we can skip it. 4971 if (!_process_only_dirty || klass->has_modified_oops()) { 4972 // Clean the klass since we're going to scavenge all the metadata. 4973 klass->clear_modified_oops(); 4974 4975 // Tell the closure that this klass is the Klass to scavenge 4976 // and is the one to dirty if oops are left pointing into the young gen. 4977 _closure->set_scanned_klass(klass); 4978 4979 klass->oops_do(_closure); 4980 4981 _closure->set_scanned_klass(NULL); 4982 } 4983 _count++; 4984 } 4985 }; 4986 4987 class G1ParTask : public AbstractGangTask { 4988 protected: 4989 G1CollectedHeap* _g1h; 4990 RefToScanQueueSet *_queues; 4991 ParallelTaskTerminator _terminator; 4992 uint _n_workers; 4993 4994 Mutex _stats_lock; 4995 Mutex* stats_lock() { return &_stats_lock; } 4996 4997 size_t getNCards() { 4998 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4999 / G1BlockOffsetSharedArray::N_bytes; 5000 } 5001 5002 public: 5003 G1ParTask(G1CollectedHeap* g1h, 5004 RefToScanQueueSet *task_queues) 5005 : AbstractGangTask("G1 collection"), 5006 _g1h(g1h), 5007 _queues(task_queues), 5008 _terminator(0, _queues), 5009 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 5010 {} 5011 5012 RefToScanQueueSet* queues() { return _queues; } 5013 5014 RefToScanQueue *work_queue(int i) { 5015 return queues()->queue(i); 5016 } 5017 5018 ParallelTaskTerminator* terminator() { return &_terminator; } 5019 5020 virtual void set_for_termination(int active_workers) { 5021 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 5022 // in the young space (_par_seq_tasks) in the G1 heap 5023 // for SequentialSubTasksDone. 5024 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 5025 // both of which need setting by set_n_termination(). 5026 _g1h->SharedHeap::set_n_termination(active_workers); 5027 _g1h->set_n_termination(active_workers); 5028 terminator()->reset_for_reuse(active_workers); 5029 _n_workers = active_workers; 5030 } 5031 5032 void work(uint worker_id) { 5033 if (worker_id >= _n_workers) return; // no work needed this round 5034 5035 double start_time_ms = os::elapsedTime() * 1000.0; 5036 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 5037 5038 { 5039 ResourceMark rm; 5040 HandleMark hm; 5041 5042 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 5043 5044 G1ParScanThreadState pss(_g1h, worker_id); 5045 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 5046 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 5047 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 5048 5049 pss.set_evac_closure(&scan_evac_cl); 5050 pss.set_evac_failure_closure(&evac_failure_cl); 5051 pss.set_partial_scan_closure(&partial_scan_cl); 5052 5053 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 5054 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 5055 5056 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 5057 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 5058 5059 bool only_young = _g1h->g1_policy()->gcs_are_young(); 5060 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 5061 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 5062 5063 OopClosure* scan_root_cl = &only_scan_root_cl; 5064 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 5065 5066 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5067 // We also need to mark copied objects. 5068 scan_root_cl = &scan_mark_root_cl; 5069 scan_klasses_cl = &scan_mark_klasses_cl_s; 5070 } 5071 5072 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 5073 5074 // Don't scan the scavengable methods in the code cache as part 5075 // of strong root scanning. The code roots that point into a 5076 // region in the collection set are scanned when we scan the 5077 // region's RSet. 5078 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings; 5079 5080 pss.start_strong_roots(); 5081 _g1h->g1_process_strong_roots(/* is scavenging */ true, 5082 SharedHeap::ScanningOption(so), 5083 scan_root_cl, 5084 &push_heap_rs_cl, 5085 scan_klasses_cl, 5086 worker_id); 5087 pss.end_strong_roots(); 5088 5089 { 5090 double start = os::elapsedTime(); 5091 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 5092 evac.do_void(); 5093 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 5094 double term_ms = pss.term_time()*1000.0; 5095 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 5096 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 5097 } 5098 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 5099 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 5100 5101 if (ParallelGCVerbose) { 5102 MutexLocker x(stats_lock()); 5103 pss.print_termination_stats(worker_id); 5104 } 5105 5106 assert(pss.refs()->is_empty(), "should be empty"); 5107 5108 // Close the inner scope so that the ResourceMark and HandleMark 5109 // destructors are executed here and are included as part of the 5110 // "GC Worker Time". 5111 } 5112 5113 double end_time_ms = os::elapsedTime() * 1000.0; 5114 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 5115 } 5116 }; 5117 5118 // *** Common G1 Evacuation Stuff 5119 5120 // This method is run in a GC worker. 5121 5122 void 5123 G1CollectedHeap:: 5124 g1_process_strong_roots(bool is_scavenging, 5125 ScanningOption so, 5126 OopClosure* scan_non_heap_roots, 5127 OopsInHeapRegionClosure* scan_rs, 5128 G1KlassScanClosure* scan_klasses, 5129 int worker_i) { 5130 5131 // First scan the strong roots 5132 double ext_roots_start = os::elapsedTime(); 5133 double closure_app_time_sec = 0.0; 5134 5135 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 5136 5137 assert(so & SO_CodeCache || scan_rs != NULL, "must scan code roots somehow"); 5138 // Walk the code cache/strong code roots w/o buffering, because StarTask 5139 // cannot handle unaligned oop locations. 5140 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, true /* do_marking */); 5141 5142 process_strong_roots(false, // no scoping; this is parallel code 5143 is_scavenging, so, 5144 &buf_scan_non_heap_roots, 5145 &eager_scan_code_roots, 5146 scan_klasses 5147 ); 5148 5149 // Now the CM ref_processor roots. 5150 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5151 // We need to treat the discovered reference lists of the 5152 // concurrent mark ref processor as roots and keep entries 5153 // (which are added by the marking threads) on them live 5154 // until they can be processed at the end of marking. 5155 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5156 } 5157 5158 // Finish up any enqueued closure apps (attributed as object copy time). 5159 buf_scan_non_heap_roots.done(); 5160 5161 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5162 5163 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5164 5165 double ext_root_time_ms = 5166 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5167 5168 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5169 5170 // During conc marking we have to filter the per-thread SATB buffers 5171 // to make sure we remove any oops into the CSet (which will show up 5172 // as implicitly live). 5173 double satb_filtering_ms = 0.0; 5174 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5175 if (mark_in_progress()) { 5176 double satb_filter_start = os::elapsedTime(); 5177 5178 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5179 5180 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5181 } 5182 } 5183 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5184 5185 // If this is an initial mark pause, and we're not scanning 5186 // the entire code cache, we need to mark the oops in the 5187 // strong code root lists for the regions that are not in 5188 // the collection set. 5189 // Note all threads participate in this set of root tasks. 5190 double mark_strong_code_roots_ms = 0.0; 5191 if (g1_policy()->during_initial_mark_pause() && !(so & SO_CodeCache)) { 5192 double mark_strong_roots_start = os::elapsedTime(); 5193 mark_strong_code_roots(worker_i); 5194 mark_strong_code_roots_ms = (os::elapsedTime() - mark_strong_roots_start) * 1000.0; 5195 } 5196 g1_policy()->phase_times()->record_strong_code_root_mark_time(worker_i, mark_strong_code_roots_ms); 5197 5198 // Now scan the complement of the collection set. 5199 if (scan_rs != NULL) { 5200 g1_rem_set()->oops_into_collection_set_do(scan_rs, &eager_scan_code_roots, worker_i); 5201 } 5202 _process_strong_tasks->all_tasks_completed(); 5203 } 5204 5205 void 5206 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) { 5207 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5208 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs); 5209 } 5210 5211 // Weak Reference Processing support 5212 5213 // An always "is_alive" closure that is used to preserve referents. 5214 // If the object is non-null then it's alive. Used in the preservation 5215 // of referent objects that are pointed to by reference objects 5216 // discovered by the CM ref processor. 5217 class G1AlwaysAliveClosure: public BoolObjectClosure { 5218 G1CollectedHeap* _g1; 5219 public: 5220 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5221 bool do_object_b(oop p) { 5222 if (p != NULL) { 5223 return true; 5224 } 5225 return false; 5226 } 5227 }; 5228 5229 bool G1STWIsAliveClosure::do_object_b(oop p) { 5230 // An object is reachable if it is outside the collection set, 5231 // or is inside and copied. 5232 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5233 } 5234 5235 // Non Copying Keep Alive closure 5236 class G1KeepAliveClosure: public OopClosure { 5237 G1CollectedHeap* _g1; 5238 public: 5239 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5240 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5241 void do_oop( oop* p) { 5242 oop obj = *p; 5243 5244 if (_g1->obj_in_cs(obj)) { 5245 assert( obj->is_forwarded(), "invariant" ); 5246 *p = obj->forwardee(); 5247 } 5248 } 5249 }; 5250 5251 // Copying Keep Alive closure - can be called from both 5252 // serial and parallel code as long as different worker 5253 // threads utilize different G1ParScanThreadState instances 5254 // and different queues. 5255 5256 class G1CopyingKeepAliveClosure: public OopClosure { 5257 G1CollectedHeap* _g1h; 5258 OopClosure* _copy_non_heap_obj_cl; 5259 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5260 G1ParScanThreadState* _par_scan_state; 5261 5262 public: 5263 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5264 OopClosure* non_heap_obj_cl, 5265 OopsInHeapRegionClosure* metadata_obj_cl, 5266 G1ParScanThreadState* pss): 5267 _g1h(g1h), 5268 _copy_non_heap_obj_cl(non_heap_obj_cl), 5269 _copy_metadata_obj_cl(metadata_obj_cl), 5270 _par_scan_state(pss) 5271 {} 5272 5273 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5274 virtual void do_oop( oop* p) { do_oop_work(p); } 5275 5276 template <class T> void do_oop_work(T* p) { 5277 oop obj = oopDesc::load_decode_heap_oop(p); 5278 5279 if (_g1h->obj_in_cs(obj)) { 5280 // If the referent object has been forwarded (either copied 5281 // to a new location or to itself in the event of an 5282 // evacuation failure) then we need to update the reference 5283 // field and, if both reference and referent are in the G1 5284 // heap, update the RSet for the referent. 5285 // 5286 // If the referent has not been forwarded then we have to keep 5287 // it alive by policy. Therefore we have copy the referent. 5288 // 5289 // If the reference field is in the G1 heap then we can push 5290 // on the PSS queue. When the queue is drained (after each 5291 // phase of reference processing) the object and it's followers 5292 // will be copied, the reference field set to point to the 5293 // new location, and the RSet updated. Otherwise we need to 5294 // use the the non-heap or metadata closures directly to copy 5295 // the referent object and update the pointer, while avoiding 5296 // updating the RSet. 5297 5298 if (_g1h->is_in_g1_reserved(p)) { 5299 _par_scan_state->push_on_queue(p); 5300 } else { 5301 assert(!ClassLoaderDataGraph::contains((address)p), 5302 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5303 PTR_FORMAT, p)); 5304 _copy_non_heap_obj_cl->do_oop(p); 5305 } 5306 } 5307 } 5308 }; 5309 5310 // Serial drain queue closure. Called as the 'complete_gc' 5311 // closure for each discovered list in some of the 5312 // reference processing phases. 5313 5314 class G1STWDrainQueueClosure: public VoidClosure { 5315 protected: 5316 G1CollectedHeap* _g1h; 5317 G1ParScanThreadState* _par_scan_state; 5318 5319 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5320 5321 public: 5322 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5323 _g1h(g1h), 5324 _par_scan_state(pss) 5325 { } 5326 5327 void do_void() { 5328 G1ParScanThreadState* const pss = par_scan_state(); 5329 pss->trim_queue(); 5330 } 5331 }; 5332 5333 // Parallel Reference Processing closures 5334 5335 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5336 // processing during G1 evacuation pauses. 5337 5338 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5339 private: 5340 G1CollectedHeap* _g1h; 5341 RefToScanQueueSet* _queues; 5342 FlexibleWorkGang* _workers; 5343 int _active_workers; 5344 5345 public: 5346 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5347 FlexibleWorkGang* workers, 5348 RefToScanQueueSet *task_queues, 5349 int n_workers) : 5350 _g1h(g1h), 5351 _queues(task_queues), 5352 _workers(workers), 5353 _active_workers(n_workers) 5354 { 5355 assert(n_workers > 0, "shouldn't call this otherwise"); 5356 } 5357 5358 // Executes the given task using concurrent marking worker threads. 5359 virtual void execute(ProcessTask& task); 5360 virtual void execute(EnqueueTask& task); 5361 }; 5362 5363 // Gang task for possibly parallel reference processing 5364 5365 class G1STWRefProcTaskProxy: public AbstractGangTask { 5366 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5367 ProcessTask& _proc_task; 5368 G1CollectedHeap* _g1h; 5369 RefToScanQueueSet *_task_queues; 5370 ParallelTaskTerminator* _terminator; 5371 5372 public: 5373 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5374 G1CollectedHeap* g1h, 5375 RefToScanQueueSet *task_queues, 5376 ParallelTaskTerminator* terminator) : 5377 AbstractGangTask("Process reference objects in parallel"), 5378 _proc_task(proc_task), 5379 _g1h(g1h), 5380 _task_queues(task_queues), 5381 _terminator(terminator) 5382 {} 5383 5384 virtual void work(uint worker_id) { 5385 // The reference processing task executed by a single worker. 5386 ResourceMark rm; 5387 HandleMark hm; 5388 5389 G1STWIsAliveClosure is_alive(_g1h); 5390 5391 G1ParScanThreadState pss(_g1h, worker_id); 5392 5393 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5394 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5395 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5396 5397 pss.set_evac_closure(&scan_evac_cl); 5398 pss.set_evac_failure_closure(&evac_failure_cl); 5399 pss.set_partial_scan_closure(&partial_scan_cl); 5400 5401 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5402 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5403 5404 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5405 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5406 5407 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5408 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5409 5410 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5411 // We also need to mark copied objects. 5412 copy_non_heap_cl = ©_mark_non_heap_cl; 5413 copy_metadata_cl = ©_mark_metadata_cl; 5414 } 5415 5416 // Keep alive closure. 5417 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5418 5419 // Complete GC closure 5420 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5421 5422 // Call the reference processing task's work routine. 5423 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5424 5425 // Note we cannot assert that the refs array is empty here as not all 5426 // of the processing tasks (specifically phase2 - pp2_work) execute 5427 // the complete_gc closure (which ordinarily would drain the queue) so 5428 // the queue may not be empty. 5429 } 5430 }; 5431 5432 // Driver routine for parallel reference processing. 5433 // Creates an instance of the ref processing gang 5434 // task and has the worker threads execute it. 5435 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5436 assert(_workers != NULL, "Need parallel worker threads."); 5437 5438 ParallelTaskTerminator terminator(_active_workers, _queues); 5439 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5440 5441 _g1h->set_par_threads(_active_workers); 5442 _workers->run_task(&proc_task_proxy); 5443 _g1h->set_par_threads(0); 5444 } 5445 5446 // Gang task for parallel reference enqueueing. 5447 5448 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5449 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5450 EnqueueTask& _enq_task; 5451 5452 public: 5453 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5454 AbstractGangTask("Enqueue reference objects in parallel"), 5455 _enq_task(enq_task) 5456 { } 5457 5458 virtual void work(uint worker_id) { 5459 _enq_task.work(worker_id); 5460 } 5461 }; 5462 5463 // Driver routine for parallel reference enqueueing. 5464 // Creates an instance of the ref enqueueing gang 5465 // task and has the worker threads execute it. 5466 5467 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5468 assert(_workers != NULL, "Need parallel worker threads."); 5469 5470 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5471 5472 _g1h->set_par_threads(_active_workers); 5473 _workers->run_task(&enq_task_proxy); 5474 _g1h->set_par_threads(0); 5475 } 5476 5477 // End of weak reference support closures 5478 5479 // Abstract task used to preserve (i.e. copy) any referent objects 5480 // that are in the collection set and are pointed to by reference 5481 // objects discovered by the CM ref processor. 5482 5483 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5484 protected: 5485 G1CollectedHeap* _g1h; 5486 RefToScanQueueSet *_queues; 5487 ParallelTaskTerminator _terminator; 5488 uint _n_workers; 5489 5490 public: 5491 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5492 AbstractGangTask("ParPreserveCMReferents"), 5493 _g1h(g1h), 5494 _queues(task_queues), 5495 _terminator(workers, _queues), 5496 _n_workers(workers) 5497 { } 5498 5499 void work(uint worker_id) { 5500 ResourceMark rm; 5501 HandleMark hm; 5502 5503 G1ParScanThreadState pss(_g1h, worker_id); 5504 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5505 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5506 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5507 5508 pss.set_evac_closure(&scan_evac_cl); 5509 pss.set_evac_failure_closure(&evac_failure_cl); 5510 pss.set_partial_scan_closure(&partial_scan_cl); 5511 5512 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5513 5514 5515 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5516 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5517 5518 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5519 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5520 5521 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5522 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5523 5524 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5525 // We also need to mark copied objects. 5526 copy_non_heap_cl = ©_mark_non_heap_cl; 5527 copy_metadata_cl = ©_mark_metadata_cl; 5528 } 5529 5530 // Is alive closure 5531 G1AlwaysAliveClosure always_alive(_g1h); 5532 5533 // Copying keep alive closure. Applied to referent objects that need 5534 // to be copied. 5535 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5536 5537 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5538 5539 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5540 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5541 5542 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5543 // So this must be true - but assert just in case someone decides to 5544 // change the worker ids. 5545 assert(0 <= worker_id && worker_id < limit, "sanity"); 5546 assert(!rp->discovery_is_atomic(), "check this code"); 5547 5548 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5549 for (uint idx = worker_id; idx < limit; idx += stride) { 5550 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5551 5552 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5553 while (iter.has_next()) { 5554 // Since discovery is not atomic for the CM ref processor, we 5555 // can see some null referent objects. 5556 iter.load_ptrs(DEBUG_ONLY(true)); 5557 oop ref = iter.obj(); 5558 5559 // This will filter nulls. 5560 if (iter.is_referent_alive()) { 5561 iter.make_referent_alive(); 5562 } 5563 iter.move_to_next(); 5564 } 5565 } 5566 5567 // Drain the queue - which may cause stealing 5568 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5569 drain_queue.do_void(); 5570 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5571 assert(pss.refs()->is_empty(), "should be"); 5572 } 5573 }; 5574 5575 // Weak Reference processing during an evacuation pause (part 1). 5576 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5577 double ref_proc_start = os::elapsedTime(); 5578 5579 ReferenceProcessor* rp = _ref_processor_stw; 5580 assert(rp->discovery_enabled(), "should have been enabled"); 5581 5582 // Any reference objects, in the collection set, that were 'discovered' 5583 // by the CM ref processor should have already been copied (either by 5584 // applying the external root copy closure to the discovered lists, or 5585 // by following an RSet entry). 5586 // 5587 // But some of the referents, that are in the collection set, that these 5588 // reference objects point to may not have been copied: the STW ref 5589 // processor would have seen that the reference object had already 5590 // been 'discovered' and would have skipped discovering the reference, 5591 // but would not have treated the reference object as a regular oop. 5592 // As a result the copy closure would not have been applied to the 5593 // referent object. 5594 // 5595 // We need to explicitly copy these referent objects - the references 5596 // will be processed at the end of remarking. 5597 // 5598 // We also need to do this copying before we process the reference 5599 // objects discovered by the STW ref processor in case one of these 5600 // referents points to another object which is also referenced by an 5601 // object discovered by the STW ref processor. 5602 5603 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5604 no_of_gc_workers == workers()->active_workers(), 5605 "Need to reset active GC workers"); 5606 5607 set_par_threads(no_of_gc_workers); 5608 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5609 no_of_gc_workers, 5610 _task_queues); 5611 5612 if (G1CollectedHeap::use_parallel_gc_threads()) { 5613 workers()->run_task(&keep_cm_referents); 5614 } else { 5615 keep_cm_referents.work(0); 5616 } 5617 5618 set_par_threads(0); 5619 5620 // Closure to test whether a referent is alive. 5621 G1STWIsAliveClosure is_alive(this); 5622 5623 // Even when parallel reference processing is enabled, the processing 5624 // of JNI refs is serial and performed serially by the current thread 5625 // rather than by a worker. The following PSS will be used for processing 5626 // JNI refs. 5627 5628 // Use only a single queue for this PSS. 5629 G1ParScanThreadState pss(this, 0); 5630 5631 // We do not embed a reference processor in the copying/scanning 5632 // closures while we're actually processing the discovered 5633 // reference objects. 5634 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5635 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5636 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5637 5638 pss.set_evac_closure(&scan_evac_cl); 5639 pss.set_evac_failure_closure(&evac_failure_cl); 5640 pss.set_partial_scan_closure(&partial_scan_cl); 5641 5642 assert(pss.refs()->is_empty(), "pre-condition"); 5643 5644 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5645 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5646 5647 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5648 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5649 5650 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5651 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5652 5653 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5654 // We also need to mark copied objects. 5655 copy_non_heap_cl = ©_mark_non_heap_cl; 5656 copy_metadata_cl = ©_mark_metadata_cl; 5657 } 5658 5659 // Keep alive closure. 5660 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5661 5662 // Serial Complete GC closure 5663 G1STWDrainQueueClosure drain_queue(this, &pss); 5664 5665 // Setup the soft refs policy... 5666 rp->setup_policy(false); 5667 5668 ReferenceProcessorStats stats; 5669 if (!rp->processing_is_mt()) { 5670 // Serial reference processing... 5671 stats = rp->process_discovered_references(&is_alive, 5672 &keep_alive, 5673 &drain_queue, 5674 NULL, 5675 _gc_timer_stw); 5676 } else { 5677 // Parallel reference processing 5678 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5679 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5680 5681 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5682 stats = rp->process_discovered_references(&is_alive, 5683 &keep_alive, 5684 &drain_queue, 5685 &par_task_executor, 5686 _gc_timer_stw); 5687 } 5688 5689 _gc_tracer_stw->report_gc_reference_stats(stats); 5690 // We have completed copying any necessary live referent objects 5691 // (that were not copied during the actual pause) so we can 5692 // retire any active alloc buffers 5693 pss.retire_alloc_buffers(); 5694 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5695 5696 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5697 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5698 } 5699 5700 // Weak Reference processing during an evacuation pause (part 2). 5701 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5702 double ref_enq_start = os::elapsedTime(); 5703 5704 ReferenceProcessor* rp = _ref_processor_stw; 5705 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5706 5707 // Now enqueue any remaining on the discovered lists on to 5708 // the pending list. 5709 if (!rp->processing_is_mt()) { 5710 // Serial reference processing... 5711 rp->enqueue_discovered_references(); 5712 } else { 5713 // Parallel reference enqueueing 5714 5715 assert(no_of_gc_workers == workers()->active_workers(), 5716 "Need to reset active workers"); 5717 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5718 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5719 5720 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5721 rp->enqueue_discovered_references(&par_task_executor); 5722 } 5723 5724 rp->verify_no_references_recorded(); 5725 assert(!rp->discovery_enabled(), "should have been disabled"); 5726 5727 // FIXME 5728 // CM's reference processing also cleans up the string and symbol tables. 5729 // Should we do that here also? We could, but it is a serial operation 5730 // and could significantly increase the pause time. 5731 5732 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5733 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5734 } 5735 5736 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { 5737 _expand_heap_after_alloc_failure = true; 5738 _evacuation_failed = false; 5739 5740 // Should G1EvacuationFailureALot be in effect for this GC? 5741 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5742 5743 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5744 5745 // Disable the hot card cache. 5746 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 5747 hot_card_cache->reset_hot_cache_claimed_index(); 5748 hot_card_cache->set_use_cache(false); 5749 5750 uint n_workers; 5751 if (G1CollectedHeap::use_parallel_gc_threads()) { 5752 n_workers = 5753 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5754 workers()->active_workers(), 5755 Threads::number_of_non_daemon_threads()); 5756 assert(UseDynamicNumberOfGCThreads || 5757 n_workers == workers()->total_workers(), 5758 "If not dynamic should be using all the workers"); 5759 workers()->set_active_workers(n_workers); 5760 set_par_threads(n_workers); 5761 } else { 5762 assert(n_par_threads() == 0, 5763 "Should be the original non-parallel value"); 5764 n_workers = 1; 5765 } 5766 5767 G1ParTask g1_par_task(this, _task_queues); 5768 5769 init_for_evac_failure(NULL); 5770 5771 rem_set()->prepare_for_younger_refs_iterate(true); 5772 5773 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5774 double start_par_time_sec = os::elapsedTime(); 5775 double end_par_time_sec; 5776 5777 { 5778 StrongRootsScope srs(this); 5779 5780 if (G1CollectedHeap::use_parallel_gc_threads()) { 5781 // The individual threads will set their evac-failure closures. 5782 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5783 // These tasks use ShareHeap::_process_strong_tasks 5784 assert(UseDynamicNumberOfGCThreads || 5785 workers()->active_workers() == workers()->total_workers(), 5786 "If not dynamic should be using all the workers"); 5787 workers()->run_task(&g1_par_task); 5788 } else { 5789 g1_par_task.set_for_termination(n_workers); 5790 g1_par_task.work(0); 5791 } 5792 end_par_time_sec = os::elapsedTime(); 5793 5794 // Closing the inner scope will execute the destructor 5795 // for the StrongRootsScope object. We record the current 5796 // elapsed time before closing the scope so that time 5797 // taken for the SRS destructor is NOT included in the 5798 // reported parallel time. 5799 } 5800 5801 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5802 g1_policy()->phase_times()->record_par_time(par_time_ms); 5803 5804 double code_root_fixup_time_ms = 5805 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5806 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5807 5808 set_par_threads(0); 5809 5810 // Process any discovered reference objects - we have 5811 // to do this _before_ we retire the GC alloc regions 5812 // as we may have to copy some 'reachable' referent 5813 // objects (and their reachable sub-graphs) that were 5814 // not copied during the pause. 5815 process_discovered_references(n_workers); 5816 5817 // Weak root processing. 5818 { 5819 G1STWIsAliveClosure is_alive(this); 5820 G1KeepAliveClosure keep_alive(this); 5821 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5822 } 5823 5824 release_gc_alloc_regions(n_workers, evacuation_info); 5825 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5826 5827 // Reset and re-enable the hot card cache. 5828 // Note the counts for the cards in the regions in the 5829 // collection set are reset when the collection set is freed. 5830 hot_card_cache->reset_hot_cache(); 5831 hot_card_cache->set_use_cache(true); 5832 5833 // Migrate the strong code roots attached to each region in 5834 // the collection set. Ideally we would like to do this 5835 // after we have finished the scanning/evacuation of the 5836 // strong code roots for a particular heap region. 5837 migrate_strong_code_roots(); 5838 5839 if (g1_policy()->during_initial_mark_pause()) { 5840 // Reset the claim values set during marking the strong code roots 5841 reset_heap_region_claim_values(); 5842 } 5843 5844 finalize_for_evac_failure(); 5845 5846 if (evacuation_failed()) { 5847 remove_self_forwarding_pointers(); 5848 5849 // Reset the G1EvacuationFailureALot counters and flags 5850 // Note: the values are reset only when an actual 5851 // evacuation failure occurs. 5852 NOT_PRODUCT(reset_evacuation_should_fail();) 5853 } 5854 5855 // Enqueue any remaining references remaining on the STW 5856 // reference processor's discovered lists. We need to do 5857 // this after the card table is cleaned (and verified) as 5858 // the act of enqueueing entries on to the pending list 5859 // will log these updates (and dirty their associated 5860 // cards). We need these updates logged to update any 5861 // RSets. 5862 enqueue_discovered_references(n_workers); 5863 5864 if (G1DeferredRSUpdate) { 5865 RedirtyLoggedCardTableEntryFastClosure redirty; 5866 dirty_card_queue_set().set_closure(&redirty); 5867 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5868 5869 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5870 dcq.merge_bufferlists(&dirty_card_queue_set()); 5871 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5872 } 5873 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5874 } 5875 5876 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5877 size_t* pre_used, 5878 FreeRegionList* free_list, 5879 OldRegionSet* old_proxy_set, 5880 HumongousRegionSet* humongous_proxy_set, 5881 HRRSCleanupTask* hrrs_cleanup_task, 5882 bool par) { 5883 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5884 if (hr->isHumongous()) { 5885 assert(hr->startsHumongous(), "we should only see starts humongous"); 5886 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5887 } else { 5888 _old_set.remove_with_proxy(hr, old_proxy_set); 5889 free_region(hr, pre_used, free_list, par); 5890 } 5891 } else { 5892 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5893 } 5894 } 5895 5896 void G1CollectedHeap::free_region(HeapRegion* hr, 5897 size_t* pre_used, 5898 FreeRegionList* free_list, 5899 bool par) { 5900 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5901 assert(!hr->is_empty(), "the region should not be empty"); 5902 assert(free_list != NULL, "pre-condition"); 5903 5904 // Clear the card counts for this region. 5905 // Note: we only need to do this if the region is not young 5906 // (since we don't refine cards in young regions). 5907 if (!hr->is_young()) { 5908 _cg1r->hot_card_cache()->reset_card_counts(hr); 5909 } 5910 *pre_used += hr->used(); 5911 hr->hr_clear(par, true /* clear_space */); 5912 free_list->add_as_head(hr); 5913 } 5914 5915 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5916 size_t* pre_used, 5917 FreeRegionList* free_list, 5918 HumongousRegionSet* humongous_proxy_set, 5919 bool par) { 5920 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5921 assert(free_list != NULL, "pre-condition"); 5922 assert(humongous_proxy_set != NULL, "pre-condition"); 5923 5924 size_t hr_used = hr->used(); 5925 size_t hr_capacity = hr->capacity(); 5926 size_t hr_pre_used = 0; 5927 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5928 // We need to read this before we make the region non-humongous, 5929 // otherwise the information will be gone. 5930 uint last_index = hr->last_hc_index(); 5931 hr->set_notHumongous(); 5932 free_region(hr, &hr_pre_used, free_list, par); 5933 5934 uint i = hr->hrs_index() + 1; 5935 while (i < last_index) { 5936 HeapRegion* curr_hr = region_at(i); 5937 assert(curr_hr->continuesHumongous(), "invariant"); 5938 curr_hr->set_notHumongous(); 5939 free_region(curr_hr, &hr_pre_used, free_list, par); 5940 i += 1; 5941 } 5942 assert(hr_pre_used == hr_used, 5943 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5944 "should be the same", hr_pre_used, hr_used)); 5945 *pre_used += hr_pre_used; 5946 } 5947 5948 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5949 FreeRegionList* free_list, 5950 OldRegionSet* old_proxy_set, 5951 HumongousRegionSet* humongous_proxy_set, 5952 bool par) { 5953 if (pre_used > 0) { 5954 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5955 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5956 assert(_summary_bytes_used >= pre_used, 5957 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5958 "should be >= pre_used: "SIZE_FORMAT, 5959 _summary_bytes_used, pre_used)); 5960 _summary_bytes_used -= pre_used; 5961 } 5962 if (free_list != NULL && !free_list->is_empty()) { 5963 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5964 _free_list.add_as_head(free_list); 5965 } 5966 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5967 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5968 _old_set.update_from_proxy(old_proxy_set); 5969 } 5970 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5971 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5972 _humongous_set.update_from_proxy(humongous_proxy_set); 5973 } 5974 } 5975 5976 class G1ParCleanupCTTask : public AbstractGangTask { 5977 CardTableModRefBS* _ct_bs; 5978 G1CollectedHeap* _g1h; 5979 HeapRegion* volatile _su_head; 5980 public: 5981 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5982 G1CollectedHeap* g1h) : 5983 AbstractGangTask("G1 Par Cleanup CT Task"), 5984 _ct_bs(ct_bs), _g1h(g1h) { } 5985 5986 void work(uint worker_id) { 5987 HeapRegion* r; 5988 while (r = _g1h->pop_dirty_cards_region()) { 5989 clear_cards(r); 5990 } 5991 } 5992 5993 void clear_cards(HeapRegion* r) { 5994 // Cards of the survivors should have already been dirtied. 5995 if (!r->is_survivor()) { 5996 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5997 } 5998 } 5999 }; 6000 6001 #ifndef PRODUCT 6002 class G1VerifyCardTableCleanup: public HeapRegionClosure { 6003 G1CollectedHeap* _g1h; 6004 CardTableModRefBS* _ct_bs; 6005 public: 6006 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 6007 : _g1h(g1h), _ct_bs(ct_bs) { } 6008 virtual bool doHeapRegion(HeapRegion* r) { 6009 if (r->is_survivor()) { 6010 _g1h->verify_dirty_region(r); 6011 } else { 6012 _g1h->verify_not_dirty_region(r); 6013 } 6014 return false; 6015 } 6016 }; 6017 6018 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 6019 // All of the region should be clean. 6020 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 6021 MemRegion mr(hr->bottom(), hr->end()); 6022 ct_bs->verify_not_dirty_region(mr); 6023 } 6024 6025 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 6026 // We cannot guarantee that [bottom(),end()] is dirty. Threads 6027 // dirty allocated blocks as they allocate them. The thread that 6028 // retires each region and replaces it with a new one will do a 6029 // maximal allocation to fill in [pre_dummy_top(),end()] but will 6030 // not dirty that area (one less thing to have to do while holding 6031 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 6032 // is dirty. 6033 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6034 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 6035 ct_bs->verify_dirty_region(mr); 6036 } 6037 6038 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 6039 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6040 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 6041 verify_dirty_region(hr); 6042 } 6043 } 6044 6045 void G1CollectedHeap::verify_dirty_young_regions() { 6046 verify_dirty_young_list(_young_list->first_region()); 6047 } 6048 #endif 6049 6050 void G1CollectedHeap::cleanUpCardTable() { 6051 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 6052 double start = os::elapsedTime(); 6053 6054 { 6055 // Iterate over the dirty cards region list. 6056 G1ParCleanupCTTask cleanup_task(ct_bs, this); 6057 6058 if (G1CollectedHeap::use_parallel_gc_threads()) { 6059 set_par_threads(); 6060 workers()->run_task(&cleanup_task); 6061 set_par_threads(0); 6062 } else { 6063 while (_dirty_cards_region_list) { 6064 HeapRegion* r = _dirty_cards_region_list; 6065 cleanup_task.clear_cards(r); 6066 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 6067 if (_dirty_cards_region_list == r) { 6068 // The last region. 6069 _dirty_cards_region_list = NULL; 6070 } 6071 r->set_next_dirty_cards_region(NULL); 6072 } 6073 } 6074 #ifndef PRODUCT 6075 if (G1VerifyCTCleanup || VerifyAfterGC) { 6076 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 6077 heap_region_iterate(&cleanup_verifier); 6078 } 6079 #endif 6080 } 6081 6082 double elapsed = os::elapsedTime() - start; 6083 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 6084 } 6085 6086 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) { 6087 size_t pre_used = 0; 6088 FreeRegionList local_free_list("Local List for CSet Freeing"); 6089 6090 double young_time_ms = 0.0; 6091 double non_young_time_ms = 0.0; 6092 6093 // Since the collection set is a superset of the the young list, 6094 // all we need to do to clear the young list is clear its 6095 // head and length, and unlink any young regions in the code below 6096 _young_list->clear(); 6097 6098 G1CollectorPolicy* policy = g1_policy(); 6099 6100 double start_sec = os::elapsedTime(); 6101 bool non_young = true; 6102 6103 HeapRegion* cur = cs_head; 6104 int age_bound = -1; 6105 size_t rs_lengths = 0; 6106 6107 while (cur != NULL) { 6108 assert(!is_on_master_free_list(cur), "sanity"); 6109 if (non_young) { 6110 if (cur->is_young()) { 6111 double end_sec = os::elapsedTime(); 6112 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6113 non_young_time_ms += elapsed_ms; 6114 6115 start_sec = os::elapsedTime(); 6116 non_young = false; 6117 } 6118 } else { 6119 if (!cur->is_young()) { 6120 double end_sec = os::elapsedTime(); 6121 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6122 young_time_ms += elapsed_ms; 6123 6124 start_sec = os::elapsedTime(); 6125 non_young = true; 6126 } 6127 } 6128 6129 rs_lengths += cur->rem_set()->occupied(); 6130 6131 HeapRegion* next = cur->next_in_collection_set(); 6132 assert(cur->in_collection_set(), "bad CS"); 6133 cur->set_next_in_collection_set(NULL); 6134 cur->set_in_collection_set(false); 6135 6136 if (cur->is_young()) { 6137 int index = cur->young_index_in_cset(); 6138 assert(index != -1, "invariant"); 6139 assert((uint) index < policy->young_cset_region_length(), "invariant"); 6140 size_t words_survived = _surviving_young_words[index]; 6141 cur->record_surv_words_in_group(words_survived); 6142 6143 // At this point the we have 'popped' cur from the collection set 6144 // (linked via next_in_collection_set()) but it is still in the 6145 // young list (linked via next_young_region()). Clear the 6146 // _next_young_region field. 6147 cur->set_next_young_region(NULL); 6148 } else { 6149 int index = cur->young_index_in_cset(); 6150 assert(index == -1, "invariant"); 6151 } 6152 6153 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 6154 (!cur->is_young() && cur->young_index_in_cset() == -1), 6155 "invariant" ); 6156 6157 if (!cur->evacuation_failed()) { 6158 MemRegion used_mr = cur->used_region(); 6159 6160 // And the region is empty. 6161 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 6162 free_region(cur, &pre_used, &local_free_list, false /* par */); 6163 } else { 6164 cur->uninstall_surv_rate_group(); 6165 if (cur->is_young()) { 6166 cur->set_young_index_in_cset(-1); 6167 } 6168 cur->set_not_young(); 6169 cur->set_evacuation_failed(false); 6170 // The region is now considered to be old. 6171 _old_set.add(cur); 6172 evacuation_info.increment_collectionset_used_after(cur->used()); 6173 } 6174 cur = next; 6175 } 6176 6177 evacuation_info.set_regions_freed(local_free_list.length()); 6178 policy->record_max_rs_lengths(rs_lengths); 6179 policy->cset_regions_freed(); 6180 6181 double end_sec = os::elapsedTime(); 6182 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6183 6184 if (non_young) { 6185 non_young_time_ms += elapsed_ms; 6186 } else { 6187 young_time_ms += elapsed_ms; 6188 } 6189 6190 update_sets_after_freeing_regions(pre_used, &local_free_list, 6191 NULL /* old_proxy_set */, 6192 NULL /* humongous_proxy_set */, 6193 false /* par */); 6194 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6195 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6196 } 6197 6198 // This routine is similar to the above but does not record 6199 // any policy statistics or update free lists; we are abandoning 6200 // the current incremental collection set in preparation of a 6201 // full collection. After the full GC we will start to build up 6202 // the incremental collection set again. 6203 // This is only called when we're doing a full collection 6204 // and is immediately followed by the tearing down of the young list. 6205 6206 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6207 HeapRegion* cur = cs_head; 6208 6209 while (cur != NULL) { 6210 HeapRegion* next = cur->next_in_collection_set(); 6211 assert(cur->in_collection_set(), "bad CS"); 6212 cur->set_next_in_collection_set(NULL); 6213 cur->set_in_collection_set(false); 6214 cur->set_young_index_in_cset(-1); 6215 cur = next; 6216 } 6217 } 6218 6219 void G1CollectedHeap::set_free_regions_coming() { 6220 if (G1ConcRegionFreeingVerbose) { 6221 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6222 "setting free regions coming"); 6223 } 6224 6225 assert(!free_regions_coming(), "pre-condition"); 6226 _free_regions_coming = true; 6227 } 6228 6229 void G1CollectedHeap::reset_free_regions_coming() { 6230 assert(free_regions_coming(), "pre-condition"); 6231 6232 { 6233 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6234 _free_regions_coming = false; 6235 SecondaryFreeList_lock->notify_all(); 6236 } 6237 6238 if (G1ConcRegionFreeingVerbose) { 6239 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6240 "reset free regions coming"); 6241 } 6242 } 6243 6244 void G1CollectedHeap::wait_while_free_regions_coming() { 6245 // Most of the time we won't have to wait, so let's do a quick test 6246 // first before we take the lock. 6247 if (!free_regions_coming()) { 6248 return; 6249 } 6250 6251 if (G1ConcRegionFreeingVerbose) { 6252 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6253 "waiting for free regions"); 6254 } 6255 6256 { 6257 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6258 while (free_regions_coming()) { 6259 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6260 } 6261 } 6262 6263 if (G1ConcRegionFreeingVerbose) { 6264 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6265 "done waiting for free regions"); 6266 } 6267 } 6268 6269 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6270 assert(heap_lock_held_for_gc(), 6271 "the heap lock should already be held by or for this thread"); 6272 _young_list->push_region(hr); 6273 } 6274 6275 class NoYoungRegionsClosure: public HeapRegionClosure { 6276 private: 6277 bool _success; 6278 public: 6279 NoYoungRegionsClosure() : _success(true) { } 6280 bool doHeapRegion(HeapRegion* r) { 6281 if (r->is_young()) { 6282 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6283 r->bottom(), r->end()); 6284 _success = false; 6285 } 6286 return false; 6287 } 6288 bool success() { return _success; } 6289 }; 6290 6291 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6292 bool ret = _young_list->check_list_empty(check_sample); 6293 6294 if (check_heap) { 6295 NoYoungRegionsClosure closure; 6296 heap_region_iterate(&closure); 6297 ret = ret && closure.success(); 6298 } 6299 6300 return ret; 6301 } 6302 6303 class TearDownRegionSetsClosure : public HeapRegionClosure { 6304 private: 6305 OldRegionSet *_old_set; 6306 6307 public: 6308 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6309 6310 bool doHeapRegion(HeapRegion* r) { 6311 if (r->is_empty()) { 6312 // We ignore empty regions, we'll empty the free list afterwards 6313 } else if (r->is_young()) { 6314 // We ignore young regions, we'll empty the young list afterwards 6315 } else if (r->isHumongous()) { 6316 // We ignore humongous regions, we're not tearing down the 6317 // humongous region set 6318 } else { 6319 // The rest should be old 6320 _old_set->remove(r); 6321 } 6322 return false; 6323 } 6324 6325 ~TearDownRegionSetsClosure() { 6326 assert(_old_set->is_empty(), "post-condition"); 6327 } 6328 }; 6329 6330 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6331 assert_at_safepoint(true /* should_be_vm_thread */); 6332 6333 if (!free_list_only) { 6334 TearDownRegionSetsClosure cl(&_old_set); 6335 heap_region_iterate(&cl); 6336 6337 // Need to do this after the heap iteration to be able to 6338 // recognize the young regions and ignore them during the iteration. 6339 _young_list->empty_list(); 6340 } 6341 _free_list.remove_all(); 6342 } 6343 6344 class RebuildRegionSetsClosure : public HeapRegionClosure { 6345 private: 6346 bool _free_list_only; 6347 OldRegionSet* _old_set; 6348 FreeRegionList* _free_list; 6349 size_t _total_used; 6350 6351 public: 6352 RebuildRegionSetsClosure(bool free_list_only, 6353 OldRegionSet* old_set, FreeRegionList* free_list) : 6354 _free_list_only(free_list_only), 6355 _old_set(old_set), _free_list(free_list), _total_used(0) { 6356 assert(_free_list->is_empty(), "pre-condition"); 6357 if (!free_list_only) { 6358 assert(_old_set->is_empty(), "pre-condition"); 6359 } 6360 } 6361 6362 bool doHeapRegion(HeapRegion* r) { 6363 if (r->continuesHumongous()) { 6364 return false; 6365 } 6366 6367 if (r->is_empty()) { 6368 // Add free regions to the free list 6369 _free_list->add_as_tail(r); 6370 } else if (!_free_list_only) { 6371 assert(!r->is_young(), "we should not come across young regions"); 6372 6373 if (r->isHumongous()) { 6374 // We ignore humongous regions, we left the humongous set unchanged 6375 } else { 6376 // The rest should be old, add them to the old set 6377 _old_set->add(r); 6378 } 6379 _total_used += r->used(); 6380 } 6381 6382 return false; 6383 } 6384 6385 size_t total_used() { 6386 return _total_used; 6387 } 6388 }; 6389 6390 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6391 assert_at_safepoint(true /* should_be_vm_thread */); 6392 6393 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6394 heap_region_iterate(&cl); 6395 6396 if (!free_list_only) { 6397 _summary_bytes_used = cl.total_used(); 6398 } 6399 assert(_summary_bytes_used == recalculate_used(), 6400 err_msg("inconsistent _summary_bytes_used, " 6401 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6402 _summary_bytes_used, recalculate_used())); 6403 } 6404 6405 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6406 _refine_cte_cl->set_concurrent(concurrent); 6407 } 6408 6409 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6410 HeapRegion* hr = heap_region_containing(p); 6411 if (hr == NULL) { 6412 return false; 6413 } else { 6414 return hr->is_in(p); 6415 } 6416 } 6417 6418 // Methods for the mutator alloc region 6419 6420 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6421 bool force) { 6422 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6423 assert(!force || g1_policy()->can_expand_young_list(), 6424 "if force is true we should be able to expand the young list"); 6425 bool young_list_full = g1_policy()->is_young_list_full(); 6426 if (force || !young_list_full) { 6427 HeapRegion* new_alloc_region = new_region(word_size, 6428 false /* do_expand */); 6429 if (new_alloc_region != NULL) { 6430 set_region_short_lived_locked(new_alloc_region); 6431 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6432 return new_alloc_region; 6433 } 6434 } 6435 return NULL; 6436 } 6437 6438 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6439 size_t allocated_bytes) { 6440 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6441 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6442 6443 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6444 _summary_bytes_used += allocated_bytes; 6445 _hr_printer.retire(alloc_region); 6446 // We update the eden sizes here, when the region is retired, 6447 // instead of when it's allocated, since this is the point that its 6448 // used space has been recored in _summary_bytes_used. 6449 g1mm()->update_eden_size(); 6450 } 6451 6452 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6453 bool force) { 6454 return _g1h->new_mutator_alloc_region(word_size, force); 6455 } 6456 6457 void G1CollectedHeap::set_par_threads() { 6458 // Don't change the number of workers. Use the value previously set 6459 // in the workgroup. 6460 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6461 uint n_workers = workers()->active_workers(); 6462 assert(UseDynamicNumberOfGCThreads || 6463 n_workers == workers()->total_workers(), 6464 "Otherwise should be using the total number of workers"); 6465 if (n_workers == 0) { 6466 assert(false, "Should have been set in prior evacuation pause."); 6467 n_workers = ParallelGCThreads; 6468 workers()->set_active_workers(n_workers); 6469 } 6470 set_par_threads(n_workers); 6471 } 6472 6473 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6474 size_t allocated_bytes) { 6475 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6476 } 6477 6478 // Methods for the GC alloc regions 6479 6480 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6481 uint count, 6482 GCAllocPurpose ap) { 6483 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6484 6485 if (count < g1_policy()->max_regions(ap)) { 6486 HeapRegion* new_alloc_region = new_region(word_size, 6487 true /* do_expand */); 6488 if (new_alloc_region != NULL) { 6489 // We really only need to do this for old regions given that we 6490 // should never scan survivors. But it doesn't hurt to do it 6491 // for survivors too. 6492 new_alloc_region->set_saved_mark(); 6493 if (ap == GCAllocForSurvived) { 6494 new_alloc_region->set_survivor(); 6495 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6496 } else { 6497 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6498 } 6499 bool during_im = g1_policy()->during_initial_mark_pause(); 6500 new_alloc_region->note_start_of_copying(during_im); 6501 return new_alloc_region; 6502 } else { 6503 g1_policy()->note_alloc_region_limit_reached(ap); 6504 } 6505 } 6506 return NULL; 6507 } 6508 6509 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6510 size_t allocated_bytes, 6511 GCAllocPurpose ap) { 6512 bool during_im = g1_policy()->during_initial_mark_pause(); 6513 alloc_region->note_end_of_copying(during_im); 6514 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6515 if (ap == GCAllocForSurvived) { 6516 young_list()->add_survivor_region(alloc_region); 6517 } else { 6518 _old_set.add(alloc_region); 6519 } 6520 _hr_printer.retire(alloc_region); 6521 } 6522 6523 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6524 bool force) { 6525 assert(!force, "not supported for GC alloc regions"); 6526 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6527 } 6528 6529 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6530 size_t allocated_bytes) { 6531 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6532 GCAllocForSurvived); 6533 } 6534 6535 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6536 bool force) { 6537 assert(!force, "not supported for GC alloc regions"); 6538 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6539 } 6540 6541 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6542 size_t allocated_bytes) { 6543 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6544 GCAllocForTenured); 6545 } 6546 // Heap region set verification 6547 6548 class VerifyRegionListsClosure : public HeapRegionClosure { 6549 private: 6550 FreeRegionList* _free_list; 6551 OldRegionSet* _old_set; 6552 HumongousRegionSet* _humongous_set; 6553 uint _region_count; 6554 6555 public: 6556 VerifyRegionListsClosure(OldRegionSet* old_set, 6557 HumongousRegionSet* humongous_set, 6558 FreeRegionList* free_list) : 6559 _old_set(old_set), _humongous_set(humongous_set), 6560 _free_list(free_list), _region_count(0) { } 6561 6562 uint region_count() { return _region_count; } 6563 6564 bool doHeapRegion(HeapRegion* hr) { 6565 _region_count += 1; 6566 6567 if (hr->continuesHumongous()) { 6568 return false; 6569 } 6570 6571 if (hr->is_young()) { 6572 // TODO 6573 } else if (hr->startsHumongous()) { 6574 _humongous_set->verify_next_region(hr); 6575 } else if (hr->is_empty()) { 6576 _free_list->verify_next_region(hr); 6577 } else { 6578 _old_set->verify_next_region(hr); 6579 } 6580 return false; 6581 } 6582 }; 6583 6584 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6585 HeapWord* bottom) { 6586 HeapWord* end = bottom + HeapRegion::GrainWords; 6587 MemRegion mr(bottom, end); 6588 assert(_g1_reserved.contains(mr), "invariant"); 6589 // This might return NULL if the allocation fails 6590 return new HeapRegion(hrs_index, _bot_shared, mr); 6591 } 6592 6593 void G1CollectedHeap::verify_region_sets() { 6594 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6595 6596 // First, check the explicit lists. 6597 _free_list.verify(); 6598 { 6599 // Given that a concurrent operation might be adding regions to 6600 // the secondary free list we have to take the lock before 6601 // verifying it. 6602 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6603 _secondary_free_list.verify(); 6604 } 6605 _old_set.verify(); 6606 _humongous_set.verify(); 6607 6608 // If a concurrent region freeing operation is in progress it will 6609 // be difficult to correctly attributed any free regions we come 6610 // across to the correct free list given that they might belong to 6611 // one of several (free_list, secondary_free_list, any local lists, 6612 // etc.). So, if that's the case we will skip the rest of the 6613 // verification operation. Alternatively, waiting for the concurrent 6614 // operation to complete will have a non-trivial effect on the GC's 6615 // operation (no concurrent operation will last longer than the 6616 // interval between two calls to verification) and it might hide 6617 // any issues that we would like to catch during testing. 6618 if (free_regions_coming()) { 6619 return; 6620 } 6621 6622 // Make sure we append the secondary_free_list on the free_list so 6623 // that all free regions we will come across can be safely 6624 // attributed to the free_list. 6625 append_secondary_free_list_if_not_empty_with_lock(); 6626 6627 // Finally, make sure that the region accounting in the lists is 6628 // consistent with what we see in the heap. 6629 _old_set.verify_start(); 6630 _humongous_set.verify_start(); 6631 _free_list.verify_start(); 6632 6633 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6634 heap_region_iterate(&cl); 6635 6636 _old_set.verify_end(); 6637 _humongous_set.verify_end(); 6638 _free_list.verify_end(); 6639 } 6640 6641 // Optimized nmethod scanning 6642 6643 class RegisterNMethodOopClosure: public OopClosure { 6644 G1CollectedHeap* _g1h; 6645 nmethod* _nm; 6646 6647 template <class T> void do_oop_work(T* p) { 6648 T heap_oop = oopDesc::load_heap_oop(p); 6649 if (!oopDesc::is_null(heap_oop)) { 6650 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6651 HeapRegion* hr = _g1h->heap_region_containing(obj); 6652 assert(!hr->isHumongous(), "code root in humongous region?"); 6653 6654 // HeapRegion::add_strong_code_root() avoids adding duplicate 6655 // entries but having duplicates is OK since we "mark" nmethods 6656 // as visited when we scan the strong code root lists during the GC. 6657 hr->add_strong_code_root(_nm); 6658 assert(hr->rem_set()->strong_code_roots_list_contains(_nm), "add failed?"); 6659 } 6660 } 6661 6662 public: 6663 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6664 _g1h(g1h), _nm(nm) {} 6665 6666 void do_oop(oop* p) { do_oop_work(p); } 6667 void do_oop(narrowOop* p) { do_oop_work(p); } 6668 }; 6669 6670 class UnregisterNMethodOopClosure: public OopClosure { 6671 G1CollectedHeap* _g1h; 6672 nmethod* _nm; 6673 6674 template <class T> void do_oop_work(T* p) { 6675 T heap_oop = oopDesc::load_heap_oop(p); 6676 if (!oopDesc::is_null(heap_oop)) { 6677 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6678 HeapRegion* hr = _g1h->heap_region_containing(obj); 6679 assert(!hr->isHumongous(), "code root in humongous region?"); 6680 hr->remove_strong_code_root(_nm); 6681 assert(!hr->rem_set()->strong_code_roots_list_contains(_nm), "remove failed?"); 6682 } 6683 } 6684 6685 public: 6686 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6687 _g1h(g1h), _nm(nm) {} 6688 6689 void do_oop(oop* p) { do_oop_work(p); } 6690 void do_oop(narrowOop* p) { do_oop_work(p); } 6691 }; 6692 6693 void G1CollectedHeap::register_nmethod(nmethod* nm) { 6694 CollectedHeap::register_nmethod(nm); 6695 6696 guarantee(nm != NULL, "sanity"); 6697 RegisterNMethodOopClosure reg_cl(this, nm); 6698 nm->oops_do(®_cl); 6699 } 6700 6701 void G1CollectedHeap::unregister_nmethod(nmethod* nm) { 6702 CollectedHeap::unregister_nmethod(nm); 6703 6704 guarantee(nm != NULL, "sanity"); 6705 UnregisterNMethodOopClosure reg_cl(this, nm); 6706 nm->oops_do(®_cl, true); 6707 } 6708 6709 class MigrateCodeRootsHeapRegionClosure: public HeapRegionClosure { 6710 public: 6711 bool doHeapRegion(HeapRegion *hr) { 6712 assert(!hr->isHumongous(), "humongous region in collection set?"); 6713 hr->migrate_strong_code_roots(); 6714 return false; 6715 } 6716 }; 6717 6718 void G1CollectedHeap::migrate_strong_code_roots() { 6719 MigrateCodeRootsHeapRegionClosure cl; 6720 double migrate_start = os::elapsedTime(); 6721 collection_set_iterate(&cl); 6722 double migration_time_ms = (os::elapsedTime() - migrate_start) * 1000.0; 6723 g1_policy()->phase_times()->record_strong_code_root_migration_time(migration_time_ms); 6724 } 6725 6726 // Mark all the code roots that point into regions *not* in the 6727 // collection set. 6728 // 6729 // Note we do not want to use a "marking" CodeBlobToOopClosure while 6730 // walking the the code roots lists of regions not in the collection 6731 // set. Suppose we have an nmethod (M) that points to objects in two 6732 // separate regions - one in the collection set (R1) and one not (R2). 6733 // Using a "marking" CodeBlobToOopClosure here would result in "marking" 6734 // nmethod M when walking the code roots for R1. When we come to scan 6735 // the code roots for R2, we would see that M is already marked and it 6736 // would be skipped and the objects in R2 that are referenced from M 6737 // would not be evacuated. 6738 6739 class MarkStrongCodeRootCodeBlobClosure: public CodeBlobClosure { 6740 6741 class MarkStrongCodeRootOopClosure: public OopClosure { 6742 ConcurrentMark* _cm; 6743 HeapRegion* _hr; 6744 uint _worker_id; 6745 6746 template <class T> void do_oop_work(T* p) { 6747 T heap_oop = oopDesc::load_heap_oop(p); 6748 if (!oopDesc::is_null(heap_oop)) { 6749 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6750 // Only mark objects in the region (which is assumed 6751 // to be not in the collection set). 6752 if (_hr->is_in(obj)) { 6753 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 6754 } 6755 } 6756 } 6757 6758 public: 6759 MarkStrongCodeRootOopClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id) : 6760 _cm(cm), _hr(hr), _worker_id(worker_id) { 6761 assert(!_hr->in_collection_set(), "sanity"); 6762 } 6763 6764 void do_oop(narrowOop* p) { do_oop_work(p); } 6765 void do_oop(oop* p) { do_oop_work(p); } 6766 }; 6767 6768 MarkStrongCodeRootOopClosure _oop_cl; 6769 6770 public: 6771 MarkStrongCodeRootCodeBlobClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id): 6772 _oop_cl(cm, hr, worker_id) {} 6773 6774 void do_code_blob(CodeBlob* cb) { 6775 nmethod* nm = (cb == NULL) ? NULL : cb->as_nmethod_or_null(); 6776 if (nm != NULL) { 6777 nm->oops_do(&_oop_cl); 6778 } 6779 } 6780 }; 6781 6782 class MarkStrongCodeRootsHRClosure: public HeapRegionClosure { 6783 G1CollectedHeap* _g1h; 6784 uint _worker_id; 6785 6786 public: 6787 MarkStrongCodeRootsHRClosure(G1CollectedHeap* g1h, uint worker_id) : 6788 _g1h(g1h), _worker_id(worker_id) {} 6789 6790 bool doHeapRegion(HeapRegion *hr) { 6791 HeapRegionRemSet* hrrs = hr->rem_set(); 6792 if (hr->isHumongous()) { 6793 // Code roots should never be attached to a humongous region 6794 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 6795 return false; 6796 } 6797 6798 if (hr->in_collection_set()) { 6799 // Don't mark code roots into regions in the collection set here. 6800 // They will be marked when we scan them. 6801 return false; 6802 } 6803 6804 MarkStrongCodeRootCodeBlobClosure cb_cl(_g1h->concurrent_mark(), hr, _worker_id); 6805 hr->strong_code_roots_do(&cb_cl); 6806 return false; 6807 } 6808 }; 6809 6810 void G1CollectedHeap::mark_strong_code_roots(uint worker_id) { 6811 MarkStrongCodeRootsHRClosure cl(this, worker_id); 6812 if (G1CollectedHeap::use_parallel_gc_threads()) { 6813 heap_region_par_iterate_chunked(&cl, 6814 worker_id, 6815 workers()->active_workers(), 6816 HeapRegion::ParMarkRootClaimValue); 6817 } else { 6818 heap_region_iterate(&cl); 6819 } 6820 } 6821 6822 class RebuildStrongCodeRootClosure: public CodeBlobClosure { 6823 G1CollectedHeap* _g1h; 6824 6825 public: 6826 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) : 6827 _g1h(g1h) {} 6828 6829 void do_code_blob(CodeBlob* cb) { 6830 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL; 6831 if (nm == NULL) { 6832 return; 6833 } 6834 6835 if (ScavengeRootsInCode && nm->detect_scavenge_root_oops()) { 6836 _g1h->register_nmethod(nm); 6837 } 6838 } 6839 }; 6840 6841 void G1CollectedHeap::rebuild_strong_code_roots() { 6842 RebuildStrongCodeRootClosure blob_cl(this); 6843 CodeCache::blobs_do(&blob_cl); 6844 }