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, uint 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, uint 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, uint 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, uint 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 GCCause::_g1_inc_collection_pause); 986 if (result != NULL) { 987 assert(succeeded, "only way to get back a non-NULL result"); 988 return result; 989 } 990 991 if (succeeded) { 992 // If we get here we successfully scheduled a collection which 993 // failed to allocate. No point in trying to allocate 994 // further. We'll just return NULL. 995 MutexLockerEx x(Heap_lock); 996 *gc_count_before_ret = total_collections(); 997 return NULL; 998 } 999 } else { 1000 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1001 MutexLockerEx x(Heap_lock); 1002 *gc_count_before_ret = total_collections(); 1003 return NULL; 1004 } 1005 // The GCLocker is either active or the GCLocker initiated 1006 // GC has not yet been performed. Stall until it is and 1007 // then retry the allocation. 1008 GC_locker::stall_until_clear(); 1009 (*gclocker_retry_count_ret) += 1; 1010 } 1011 1012 // We can reach here if we were unsuccessful in scheduling a 1013 // collection (because another thread beat us to it) or if we were 1014 // stalled due to the GC locker. In either can we should retry the 1015 // allocation attempt in case another thread successfully 1016 // performed a collection and reclaimed enough space. We do the 1017 // first attempt (without holding the Heap_lock) here and the 1018 // follow-on attempt will be at the start of the next loop 1019 // iteration (after taking the Heap_lock). 1020 result = _mutator_alloc_region.attempt_allocation(word_size, 1021 false /* bot_updates */); 1022 if (result != NULL) { 1023 return result; 1024 } 1025 1026 // Give a warning if we seem to be looping forever. 1027 if ((QueuedAllocationWarningCount > 0) && 1028 (try_count % QueuedAllocationWarningCount == 0)) { 1029 warning("G1CollectedHeap::attempt_allocation_slow() " 1030 "retries %d times", try_count); 1031 } 1032 } 1033 1034 ShouldNotReachHere(); 1035 return NULL; 1036 } 1037 1038 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1039 unsigned int * gc_count_before_ret, 1040 int* gclocker_retry_count_ret) { 1041 // The structure of this method has a lot of similarities to 1042 // attempt_allocation_slow(). The reason these two were not merged 1043 // into a single one is that such a method would require several "if 1044 // allocation is not humongous do this, otherwise do that" 1045 // conditional paths which would obscure its flow. In fact, an early 1046 // version of this code did use a unified method which was harder to 1047 // follow and, as a result, it had subtle bugs that were hard to 1048 // track down. So keeping these two methods separate allows each to 1049 // be more readable. It will be good to keep these two in sync as 1050 // much as possible. 1051 1052 assert_heap_not_locked_and_not_at_safepoint(); 1053 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1054 "should only be called for humongous allocations"); 1055 1056 // Humongous objects can exhaust the heap quickly, so we should check if we 1057 // need to start a marking cycle at each humongous object allocation. We do 1058 // the check before we do the actual allocation. The reason for doing it 1059 // before the allocation is that we avoid having to keep track of the newly 1060 // allocated memory while we do a GC. 1061 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1062 word_size)) { 1063 collect(GCCause::_g1_humongous_allocation); 1064 } 1065 1066 // We will loop until a) we manage to successfully perform the 1067 // allocation or b) we successfully schedule a collection which 1068 // fails to perform the allocation. b) is the only case when we'll 1069 // return NULL. 1070 HeapWord* result = NULL; 1071 for (int try_count = 1; /* we'll return */; try_count += 1) { 1072 bool should_try_gc; 1073 unsigned int gc_count_before; 1074 1075 { 1076 MutexLockerEx x(Heap_lock); 1077 1078 // Given that humongous objects are not allocated in young 1079 // regions, we'll first try to do the allocation without doing a 1080 // collection hoping that there's enough space in the heap. 1081 result = humongous_obj_allocate(word_size); 1082 if (result != NULL) { 1083 return result; 1084 } 1085 1086 if (GC_locker::is_active_and_needs_gc()) { 1087 should_try_gc = false; 1088 } else { 1089 // The GCLocker may not be active but the GCLocker initiated 1090 // GC may not yet have been performed (GCLocker::needs_gc() 1091 // returns true). In this case we do not try this GC and 1092 // wait until the GCLocker initiated GC is performed, and 1093 // then retry the allocation. 1094 if (GC_locker::needs_gc()) { 1095 should_try_gc = false; 1096 } else { 1097 // Read the GC count while still holding the Heap_lock. 1098 gc_count_before = total_collections(); 1099 should_try_gc = true; 1100 } 1101 } 1102 } 1103 1104 if (should_try_gc) { 1105 // If we failed to allocate the humongous object, we should try to 1106 // do a collection pause (if we're allowed) in case it reclaims 1107 // enough space for the allocation to succeed after the pause. 1108 1109 bool succeeded; 1110 result = do_collection_pause(word_size, gc_count_before, &succeeded, 1111 GCCause::_g1_humongous_allocation); 1112 if (result != NULL) { 1113 assert(succeeded, "only way to get back a non-NULL result"); 1114 return result; 1115 } 1116 1117 if (succeeded) { 1118 // If we get here we successfully scheduled a collection which 1119 // failed to allocate. No point in trying to allocate 1120 // further. We'll just return NULL. 1121 MutexLockerEx x(Heap_lock); 1122 *gc_count_before_ret = total_collections(); 1123 return NULL; 1124 } 1125 } else { 1126 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) { 1127 MutexLockerEx x(Heap_lock); 1128 *gc_count_before_ret = total_collections(); 1129 return NULL; 1130 } 1131 // The GCLocker is either active or the GCLocker initiated 1132 // GC has not yet been performed. Stall until it is and 1133 // then retry the allocation. 1134 GC_locker::stall_until_clear(); 1135 (*gclocker_retry_count_ret) += 1; 1136 } 1137 1138 // We can reach here if we were unsuccessful in scheduling a 1139 // collection (because another thread beat us to it) or if we were 1140 // stalled due to the GC locker. In either can we should retry the 1141 // allocation attempt in case another thread successfully 1142 // performed a collection and reclaimed enough space. Give a 1143 // warning if we seem to be looping forever. 1144 1145 if ((QueuedAllocationWarningCount > 0) && 1146 (try_count % QueuedAllocationWarningCount == 0)) { 1147 warning("G1CollectedHeap::attempt_allocation_humongous() " 1148 "retries %d times", try_count); 1149 } 1150 } 1151 1152 ShouldNotReachHere(); 1153 return NULL; 1154 } 1155 1156 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1157 bool expect_null_mutator_alloc_region) { 1158 assert_at_safepoint(true /* should_be_vm_thread */); 1159 assert(_mutator_alloc_region.get() == NULL || 1160 !expect_null_mutator_alloc_region, 1161 "the current alloc region was unexpectedly found to be non-NULL"); 1162 1163 if (!isHumongous(word_size)) { 1164 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1165 false /* bot_updates */); 1166 } else { 1167 HeapWord* result = humongous_obj_allocate(word_size); 1168 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1169 g1_policy()->set_initiate_conc_mark_if_possible(); 1170 } 1171 return result; 1172 } 1173 1174 ShouldNotReachHere(); 1175 } 1176 1177 class PostMCRemSetClearClosure: public HeapRegionClosure { 1178 G1CollectedHeap* _g1h; 1179 ModRefBarrierSet* _mr_bs; 1180 public: 1181 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1182 _g1h(g1h), _mr_bs(mr_bs) {} 1183 1184 bool doHeapRegion(HeapRegion* r) { 1185 HeapRegionRemSet* hrrs = r->rem_set(); 1186 1187 if (r->continuesHumongous()) { 1188 // We'll assert that the strong code root list and RSet is empty 1189 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 1190 assert(hrrs->occupied() == 0, "RSet should be empty"); 1191 return false; 1192 } 1193 1194 _g1h->reset_gc_time_stamps(r); 1195 hrrs->clear(); 1196 // You might think here that we could clear just the cards 1197 // corresponding to the used region. But no: if we leave a dirty card 1198 // in a region we might allocate into, then it would prevent that card 1199 // from being enqueued, and cause it to be missed. 1200 // Re: the performance cost: we shouldn't be doing full GC anyway! 1201 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1202 1203 return false; 1204 } 1205 }; 1206 1207 void G1CollectedHeap::clear_rsets_post_compaction() { 1208 PostMCRemSetClearClosure rs_clear(this, mr_bs()); 1209 heap_region_iterate(&rs_clear); 1210 } 1211 1212 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1213 G1CollectedHeap* _g1h; 1214 UpdateRSOopClosure _cl; 1215 uint _worker_i; 1216 public: 1217 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, uint worker_i = 0) : 1218 _cl(g1->g1_rem_set(), worker_i), 1219 _worker_i(worker_i), 1220 _g1h(g1) 1221 { } 1222 1223 bool doHeapRegion(HeapRegion* r) { 1224 if (!r->continuesHumongous()) { 1225 _cl.set_from(r); 1226 r->oop_iterate(&_cl); 1227 } 1228 return false; 1229 } 1230 }; 1231 1232 class ParRebuildRSTask: public AbstractGangTask { 1233 G1CollectedHeap* _g1; 1234 public: 1235 ParRebuildRSTask(G1CollectedHeap* g1) 1236 : AbstractGangTask("ParRebuildRSTask"), 1237 _g1(g1) 1238 { } 1239 1240 void work(uint worker_id) { 1241 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1242 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1243 _g1->workers()->active_workers(), 1244 HeapRegion::RebuildRSClaimValue); 1245 } 1246 }; 1247 1248 class PostCompactionPrinterClosure: public HeapRegionClosure { 1249 private: 1250 G1HRPrinter* _hr_printer; 1251 public: 1252 bool doHeapRegion(HeapRegion* hr) { 1253 assert(!hr->is_young(), "not expecting to find young regions"); 1254 // We only generate output for non-empty regions. 1255 if (!hr->is_empty()) { 1256 if (!hr->isHumongous()) { 1257 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1258 } else if (hr->startsHumongous()) { 1259 if (hr->region_num() == 1) { 1260 // single humongous region 1261 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1262 } else { 1263 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1264 } 1265 } else { 1266 assert(hr->continuesHumongous(), "only way to get here"); 1267 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1268 } 1269 } 1270 return false; 1271 } 1272 1273 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1274 : _hr_printer(hr_printer) { } 1275 }; 1276 1277 void G1CollectedHeap::print_hrs_post_compaction() { 1278 PostCompactionPrinterClosure cl(hr_printer()); 1279 heap_region_iterate(&cl); 1280 } 1281 1282 bool G1CollectedHeap::do_collection(bool explicit_gc, 1283 bool clear_all_soft_refs, 1284 size_t word_size) { 1285 assert_at_safepoint(true /* should_be_vm_thread */); 1286 1287 if (GC_locker::check_active_before_gc()) { 1288 return false; 1289 } 1290 1291 STWGCTimer* gc_timer = G1MarkSweep::gc_timer(); 1292 gc_timer->register_gc_start(os::elapsed_counter()); 1293 1294 SerialOldTracer* gc_tracer = G1MarkSweep::gc_tracer(); 1295 gc_tracer->report_gc_start(gc_cause(), gc_timer->gc_start()); 1296 1297 SvcGCMarker sgcm(SvcGCMarker::FULL); 1298 ResourceMark rm; 1299 1300 print_heap_before_gc(); 1301 trace_heap_before_gc(gc_tracer); 1302 1303 size_t metadata_prev_used = MetaspaceAux::allocated_used_bytes(); 1304 1305 HRSPhaseSetter x(HRSPhaseFullGC); 1306 verify_region_sets_optional(); 1307 1308 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1309 collector_policy()->should_clear_all_soft_refs(); 1310 1311 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1312 1313 { 1314 IsGCActiveMark x; 1315 1316 // Timing 1317 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1318 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1319 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1320 1321 { 1322 GCTraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, NULL); 1323 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1324 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1325 1326 double start = os::elapsedTime(); 1327 g1_policy()->record_full_collection_start(); 1328 1329 // Note: When we have a more flexible GC logging framework that 1330 // allows us to add optional attributes to a GC log record we 1331 // could consider timing and reporting how long we wait in the 1332 // following two methods. 1333 wait_while_free_regions_coming(); 1334 // If we start the compaction before the CM threads finish 1335 // scanning the root regions we might trip them over as we'll 1336 // be moving objects / updating references. So let's wait until 1337 // they are done. By telling them to abort, they should complete 1338 // early. 1339 _cm->root_regions()->abort(); 1340 _cm->root_regions()->wait_until_scan_finished(); 1341 append_secondary_free_list_if_not_empty_with_lock(); 1342 1343 gc_prologue(true); 1344 increment_total_collections(true /* full gc */); 1345 increment_old_marking_cycles_started(); 1346 1347 assert(used() == recalculate_used(), "Should be equal"); 1348 1349 verify_before_gc(); 1350 1351 pre_full_gc_dump(gc_timer); 1352 1353 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1354 1355 // Disable discovery and empty the discovered lists 1356 // for the CM ref processor. 1357 ref_processor_cm()->disable_discovery(); 1358 ref_processor_cm()->abandon_partial_discovery(); 1359 ref_processor_cm()->verify_no_references_recorded(); 1360 1361 // Abandon current iterations of concurrent marking and concurrent 1362 // refinement, if any are in progress. We have to do this before 1363 // wait_until_scan_finished() below. 1364 concurrent_mark()->abort(); 1365 1366 // Make sure we'll choose a new allocation region afterwards. 1367 release_mutator_alloc_region(); 1368 abandon_gc_alloc_regions(); 1369 g1_rem_set()->cleanupHRRS(); 1370 1371 // We should call this after we retire any currently active alloc 1372 // regions so that all the ALLOC / RETIRE events are generated 1373 // before the start GC event. 1374 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1375 1376 // We may have added regions to the current incremental collection 1377 // set between the last GC or pause and now. We need to clear the 1378 // incremental collection set and then start rebuilding it afresh 1379 // after this full GC. 1380 abandon_collection_set(g1_policy()->inc_cset_head()); 1381 g1_policy()->clear_incremental_cset(); 1382 g1_policy()->stop_incremental_cset_building(); 1383 1384 tear_down_region_sets(false /* free_list_only */); 1385 g1_policy()->set_gcs_are_young(true); 1386 1387 // See the comments in g1CollectedHeap.hpp and 1388 // G1CollectedHeap::ref_processing_init() about 1389 // how reference processing currently works in G1. 1390 1391 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1392 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1393 1394 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1395 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1396 1397 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1398 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1399 1400 // Do collection work 1401 { 1402 HandleMark hm; // Discard invalid handles created during gc 1403 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1404 } 1405 1406 assert(free_regions() == 0, "we should not have added any free regions"); 1407 rebuild_region_sets(false /* free_list_only */); 1408 1409 // Enqueue any discovered reference objects that have 1410 // not been removed from the discovered lists. 1411 ref_processor_stw()->enqueue_discovered_references(); 1412 1413 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1414 1415 MemoryService::track_memory_usage(); 1416 1417 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1418 ref_processor_stw()->verify_no_references_recorded(); 1419 1420 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1421 ClassLoaderDataGraph::purge(); 1422 MetaspaceAux::verify_metrics(); 1423 1424 // Note: since we've just done a full GC, concurrent 1425 // marking is no longer active. Therefore we need not 1426 // re-enable reference discovery for the CM ref processor. 1427 // That will be done at the start of the next marking cycle. 1428 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1429 ref_processor_cm()->verify_no_references_recorded(); 1430 1431 reset_gc_time_stamp(); 1432 // Since everything potentially moved, we will clear all remembered 1433 // sets, and clear all cards. Later we will rebuild remembered 1434 // sets. We will also reset the GC time stamps of the regions. 1435 clear_rsets_post_compaction(); 1436 check_gc_time_stamps(); 1437 1438 // Resize the heap if necessary. 1439 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1440 1441 if (_hr_printer.is_active()) { 1442 // We should do this after we potentially resize the heap so 1443 // that all the COMMIT / UNCOMMIT events are generated before 1444 // the end GC event. 1445 1446 print_hrs_post_compaction(); 1447 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1448 } 1449 1450 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 1451 if (hot_card_cache->use_cache()) { 1452 hot_card_cache->reset_card_counts(); 1453 hot_card_cache->reset_hot_cache(); 1454 } 1455 1456 // Rebuild remembered sets of all regions. 1457 if (G1CollectedHeap::use_parallel_gc_threads()) { 1458 uint n_workers = 1459 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1460 workers()->active_workers(), 1461 Threads::number_of_non_daemon_threads()); 1462 assert(UseDynamicNumberOfGCThreads || 1463 n_workers == workers()->total_workers(), 1464 "If not dynamic should be using all the workers"); 1465 workers()->set_active_workers(n_workers); 1466 // Set parallel threads in the heap (_n_par_threads) only 1467 // before a parallel phase and always reset it to 0 after 1468 // the phase so that the number of parallel threads does 1469 // no get carried forward to a serial phase where there 1470 // may be code that is "possibly_parallel". 1471 set_par_threads(n_workers); 1472 1473 ParRebuildRSTask rebuild_rs_task(this); 1474 assert(check_heap_region_claim_values( 1475 HeapRegion::InitialClaimValue), "sanity check"); 1476 assert(UseDynamicNumberOfGCThreads || 1477 workers()->active_workers() == workers()->total_workers(), 1478 "Unless dynamic should use total workers"); 1479 // Use the most recent number of active workers 1480 assert(workers()->active_workers() > 0, 1481 "Active workers not properly set"); 1482 set_par_threads(workers()->active_workers()); 1483 workers()->run_task(&rebuild_rs_task); 1484 set_par_threads(0); 1485 assert(check_heap_region_claim_values( 1486 HeapRegion::RebuildRSClaimValue), "sanity check"); 1487 reset_heap_region_claim_values(); 1488 } else { 1489 RebuildRSOutOfRegionClosure rebuild_rs(this); 1490 heap_region_iterate(&rebuild_rs); 1491 } 1492 1493 // Rebuild the strong code root lists for each region 1494 rebuild_strong_code_roots(); 1495 1496 if (true) { // FIXME 1497 MetaspaceGC::compute_new_size(); 1498 } 1499 1500 #ifdef TRACESPINNING 1501 ParallelTaskTerminator::print_termination_counts(); 1502 #endif 1503 1504 // Discard all rset updates 1505 JavaThread::dirty_card_queue_set().abandon_logs(); 1506 assert(!G1DeferredRSUpdate 1507 || (G1DeferredRSUpdate && 1508 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1509 1510 _young_list->reset_sampled_info(); 1511 // At this point there should be no regions in the 1512 // entire heap tagged as young. 1513 assert(check_young_list_empty(true /* check_heap */), 1514 "young list should be empty at this point"); 1515 1516 // Update the number of full collections that have been completed. 1517 increment_old_marking_cycles_completed(false /* concurrent */); 1518 1519 _hrs.verify_optional(); 1520 verify_region_sets_optional(); 1521 1522 verify_after_gc(); 1523 1524 // Start a new incremental collection set for the next pause 1525 assert(g1_policy()->collection_set() == NULL, "must be"); 1526 g1_policy()->start_incremental_cset_building(); 1527 1528 // Clear the _cset_fast_test bitmap in anticipation of adding 1529 // regions to the incremental collection set for the next 1530 // evacuation pause. 1531 clear_cset_fast_test(); 1532 1533 init_mutator_alloc_region(); 1534 1535 double end = os::elapsedTime(); 1536 g1_policy()->record_full_collection_end(); 1537 1538 if (G1Log::fine()) { 1539 g1_policy()->print_heap_transition(); 1540 } 1541 1542 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1543 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1544 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1545 // before any GC notifications are raised. 1546 g1mm()->update_sizes(); 1547 1548 gc_epilogue(true); 1549 } 1550 1551 if (G1Log::finer()) { 1552 g1_policy()->print_detailed_heap_transition(true /* full */); 1553 } 1554 1555 print_heap_after_gc(); 1556 trace_heap_after_gc(gc_tracer); 1557 1558 post_full_gc_dump(gc_timer); 1559 1560 gc_timer->register_gc_end(os::elapsed_counter()); 1561 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1562 } 1563 1564 return true; 1565 } 1566 1567 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1568 // do_collection() will return whether it succeeded in performing 1569 // the GC. Currently, there is no facility on the 1570 // do_full_collection() API to notify the caller than the collection 1571 // did not succeed (e.g., because it was locked out by the GC 1572 // locker). So, right now, we'll ignore the return value. 1573 bool dummy = do_collection(true, /* explicit_gc */ 1574 clear_all_soft_refs, 1575 0 /* word_size */); 1576 } 1577 1578 // This code is mostly copied from TenuredGeneration. 1579 void 1580 G1CollectedHeap:: 1581 resize_if_necessary_after_full_collection(size_t word_size) { 1582 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1583 1584 // Include the current allocation, if any, and bytes that will be 1585 // pre-allocated to support collections, as "used". 1586 const size_t used_after_gc = used(); 1587 const size_t capacity_after_gc = capacity(); 1588 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1589 1590 // This is enforced in arguments.cpp. 1591 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1592 "otherwise the code below doesn't make sense"); 1593 1594 // We don't have floating point command-line arguments 1595 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1596 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1597 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1598 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1599 1600 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1601 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1602 1603 // We have to be careful here as these two calculations can overflow 1604 // 32-bit size_t's. 1605 double used_after_gc_d = (double) used_after_gc; 1606 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1607 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1608 1609 // Let's make sure that they are both under the max heap size, which 1610 // by default will make them fit into a size_t. 1611 double desired_capacity_upper_bound = (double) max_heap_size; 1612 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1613 desired_capacity_upper_bound); 1614 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1615 desired_capacity_upper_bound); 1616 1617 // We can now safely turn them into size_t's. 1618 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1619 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1620 1621 // This assert only makes sense here, before we adjust them 1622 // with respect to the min and max heap size. 1623 assert(minimum_desired_capacity <= maximum_desired_capacity, 1624 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1625 "maximum_desired_capacity = "SIZE_FORMAT, 1626 minimum_desired_capacity, maximum_desired_capacity)); 1627 1628 // Should not be greater than the heap max size. No need to adjust 1629 // it with respect to the heap min size as it's a lower bound (i.e., 1630 // we'll try to make the capacity larger than it, not smaller). 1631 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1632 // Should not be less than the heap min size. No need to adjust it 1633 // with respect to the heap max size as it's an upper bound (i.e., 1634 // we'll try to make the capacity smaller than it, not greater). 1635 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1636 1637 if (capacity_after_gc < minimum_desired_capacity) { 1638 // Don't expand unless it's significant 1639 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1640 ergo_verbose4(ErgoHeapSizing, 1641 "attempt heap expansion", 1642 ergo_format_reason("capacity lower than " 1643 "min desired capacity after Full GC") 1644 ergo_format_byte("capacity") 1645 ergo_format_byte("occupancy") 1646 ergo_format_byte_perc("min desired capacity"), 1647 capacity_after_gc, used_after_gc, 1648 minimum_desired_capacity, (double) MinHeapFreeRatio); 1649 expand(expand_bytes); 1650 1651 // No expansion, now see if we want to shrink 1652 } else if (capacity_after_gc > maximum_desired_capacity) { 1653 // Capacity too large, compute shrinking size 1654 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1655 ergo_verbose4(ErgoHeapSizing, 1656 "attempt heap shrinking", 1657 ergo_format_reason("capacity higher than " 1658 "max desired capacity after Full GC") 1659 ergo_format_byte("capacity") 1660 ergo_format_byte("occupancy") 1661 ergo_format_byte_perc("max desired capacity"), 1662 capacity_after_gc, used_after_gc, 1663 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1664 shrink(shrink_bytes); 1665 } 1666 } 1667 1668 1669 HeapWord* 1670 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1671 bool* succeeded) { 1672 assert_at_safepoint(true /* should_be_vm_thread */); 1673 1674 *succeeded = true; 1675 // Let's attempt the allocation first. 1676 HeapWord* result = 1677 attempt_allocation_at_safepoint(word_size, 1678 false /* expect_null_mutator_alloc_region */); 1679 if (result != NULL) { 1680 assert(*succeeded, "sanity"); 1681 return result; 1682 } 1683 1684 // In a G1 heap, we're supposed to keep allocation from failing by 1685 // incremental pauses. Therefore, at least for now, we'll favor 1686 // expansion over collection. (This might change in the future if we can 1687 // do something smarter than full collection to satisfy a failed alloc.) 1688 result = expand_and_allocate(word_size); 1689 if (result != NULL) { 1690 assert(*succeeded, "sanity"); 1691 return result; 1692 } 1693 1694 // Expansion didn't work, we'll try to do a Full GC. 1695 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1696 false, /* clear_all_soft_refs */ 1697 word_size); 1698 if (!gc_succeeded) { 1699 *succeeded = false; 1700 return NULL; 1701 } 1702 1703 // Retry the allocation 1704 result = attempt_allocation_at_safepoint(word_size, 1705 true /* expect_null_mutator_alloc_region */); 1706 if (result != NULL) { 1707 assert(*succeeded, "sanity"); 1708 return result; 1709 } 1710 1711 // Then, try a Full GC that will collect all soft references. 1712 gc_succeeded = do_collection(false, /* explicit_gc */ 1713 true, /* clear_all_soft_refs */ 1714 word_size); 1715 if (!gc_succeeded) { 1716 *succeeded = false; 1717 return NULL; 1718 } 1719 1720 // Retry the allocation once more 1721 result = attempt_allocation_at_safepoint(word_size, 1722 true /* expect_null_mutator_alloc_region */); 1723 if (result != NULL) { 1724 assert(*succeeded, "sanity"); 1725 return result; 1726 } 1727 1728 assert(!collector_policy()->should_clear_all_soft_refs(), 1729 "Flag should have been handled and cleared prior to this point"); 1730 1731 // What else? We might try synchronous finalization later. If the total 1732 // space available is large enough for the allocation, then a more 1733 // complete compaction phase than we've tried so far might be 1734 // appropriate. 1735 assert(*succeeded, "sanity"); 1736 return NULL; 1737 } 1738 1739 // Attempting to expand the heap sufficiently 1740 // to support an allocation of the given "word_size". If 1741 // successful, perform the allocation and return the address of the 1742 // allocated block, or else "NULL". 1743 1744 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1745 assert_at_safepoint(true /* should_be_vm_thread */); 1746 1747 verify_region_sets_optional(); 1748 1749 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1750 ergo_verbose1(ErgoHeapSizing, 1751 "attempt heap expansion", 1752 ergo_format_reason("allocation request failed") 1753 ergo_format_byte("allocation request"), 1754 word_size * HeapWordSize); 1755 if (expand(expand_bytes)) { 1756 _hrs.verify_optional(); 1757 verify_region_sets_optional(); 1758 return attempt_allocation_at_safepoint(word_size, 1759 false /* expect_null_mutator_alloc_region */); 1760 } 1761 return NULL; 1762 } 1763 1764 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1765 HeapWord* new_end) { 1766 assert(old_end != new_end, "don't call this otherwise"); 1767 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1768 1769 // Update the committed mem region. 1770 _g1_committed.set_end(new_end); 1771 // Tell the card table about the update. 1772 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1773 // Tell the BOT about the update. 1774 _bot_shared->resize(_g1_committed.word_size()); 1775 // Tell the hot card cache about the update 1776 _cg1r->hot_card_cache()->resize_card_counts(capacity()); 1777 } 1778 1779 bool G1CollectedHeap::expand(size_t expand_bytes) { 1780 size_t old_mem_size = _g1_storage.committed_size(); 1781 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1782 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1783 HeapRegion::GrainBytes); 1784 ergo_verbose2(ErgoHeapSizing, 1785 "expand the heap", 1786 ergo_format_byte("requested expansion amount") 1787 ergo_format_byte("attempted expansion amount"), 1788 expand_bytes, aligned_expand_bytes); 1789 1790 // First commit the memory. 1791 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1792 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1793 if (successful) { 1794 // Then propagate this update to the necessary data structures. 1795 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1796 update_committed_space(old_end, new_end); 1797 1798 FreeRegionList expansion_list("Local Expansion List"); 1799 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1800 assert(mr.start() == old_end, "post-condition"); 1801 // mr might be a smaller region than what was requested if 1802 // expand_by() was unable to allocate the HeapRegion instances 1803 assert(mr.end() <= new_end, "post-condition"); 1804 1805 size_t actual_expand_bytes = mr.byte_size(); 1806 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1807 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1808 "post-condition"); 1809 if (actual_expand_bytes < aligned_expand_bytes) { 1810 // We could not expand _hrs to the desired size. In this case we 1811 // need to shrink the committed space accordingly. 1812 assert(mr.end() < new_end, "invariant"); 1813 1814 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1815 // First uncommit the memory. 1816 _g1_storage.shrink_by(diff_bytes); 1817 // Then propagate this update to the necessary data structures. 1818 update_committed_space(new_end, mr.end()); 1819 } 1820 _free_list.add_as_tail(&expansion_list); 1821 1822 if (_hr_printer.is_active()) { 1823 HeapWord* curr = mr.start(); 1824 while (curr < mr.end()) { 1825 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1826 _hr_printer.commit(curr, curr_end); 1827 curr = curr_end; 1828 } 1829 assert(curr == mr.end(), "post-condition"); 1830 } 1831 g1_policy()->record_new_heap_size(n_regions()); 1832 } else { 1833 ergo_verbose0(ErgoHeapSizing, 1834 "did not expand the heap", 1835 ergo_format_reason("heap expansion operation failed")); 1836 // The expansion of the virtual storage space was unsuccessful. 1837 // Let's see if it was because we ran out of swap. 1838 if (G1ExitOnExpansionFailure && 1839 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1840 // We had head room... 1841 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion"); 1842 } 1843 } 1844 return successful; 1845 } 1846 1847 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1848 size_t old_mem_size = _g1_storage.committed_size(); 1849 size_t aligned_shrink_bytes = 1850 ReservedSpace::page_align_size_down(shrink_bytes); 1851 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1852 HeapRegion::GrainBytes); 1853 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes); 1854 1855 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove); 1856 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1857 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes; 1858 1859 ergo_verbose3(ErgoHeapSizing, 1860 "shrink the heap", 1861 ergo_format_byte("requested shrinking amount") 1862 ergo_format_byte("aligned shrinking amount") 1863 ergo_format_byte("attempted shrinking amount"), 1864 shrink_bytes, aligned_shrink_bytes, shrunk_bytes); 1865 if (num_regions_removed > 0) { 1866 _g1_storage.shrink_by(shrunk_bytes); 1867 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1868 1869 if (_hr_printer.is_active()) { 1870 HeapWord* curr = old_end; 1871 while (curr > new_end) { 1872 HeapWord* curr_end = curr; 1873 curr -= HeapRegion::GrainWords; 1874 _hr_printer.uncommit(curr, curr_end); 1875 } 1876 } 1877 1878 _expansion_regions += num_regions_removed; 1879 update_committed_space(old_end, new_end); 1880 HeapRegionRemSet::shrink_heap(n_regions()); 1881 g1_policy()->record_new_heap_size(n_regions()); 1882 } else { 1883 ergo_verbose0(ErgoHeapSizing, 1884 "did not shrink the heap", 1885 ergo_format_reason("heap shrinking operation failed")); 1886 } 1887 } 1888 1889 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1890 verify_region_sets_optional(); 1891 1892 // We should only reach here at the end of a Full GC which means we 1893 // should not not be holding to any GC alloc regions. The method 1894 // below will make sure of that and do any remaining clean up. 1895 abandon_gc_alloc_regions(); 1896 1897 // Instead of tearing down / rebuilding the free lists here, we 1898 // could instead use the remove_all_pending() method on free_list to 1899 // remove only the ones that we need to remove. 1900 tear_down_region_sets(true /* free_list_only */); 1901 shrink_helper(shrink_bytes); 1902 rebuild_region_sets(true /* free_list_only */); 1903 1904 _hrs.verify_optional(); 1905 verify_region_sets_optional(); 1906 } 1907 1908 // Public methods. 1909 1910 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1911 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1912 #endif // _MSC_VER 1913 1914 1915 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1916 SharedHeap(policy_), 1917 _g1_policy(policy_), 1918 _dirty_card_queue_set(false), 1919 _into_cset_dirty_card_queue_set(false), 1920 _is_alive_closure_cm(this), 1921 _is_alive_closure_stw(this), 1922 _ref_processor_cm(NULL), 1923 _ref_processor_stw(NULL), 1924 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1925 _bot_shared(NULL), 1926 _evac_failure_scan_stack(NULL), 1927 _mark_in_progress(false), 1928 _cg1r(NULL), _summary_bytes_used(0), 1929 _g1mm(NULL), 1930 _refine_cte_cl(NULL), 1931 _full_collection(false), 1932 _free_list("Master Free List"), 1933 _secondary_free_list("Secondary Free List"), 1934 _old_set("Old Set"), 1935 _humongous_set("Master Humongous Set"), 1936 _free_regions_coming(false), 1937 _young_list(new YoungList(this)), 1938 _gc_time_stamp(0), 1939 _retained_old_gc_alloc_region(NULL), 1940 _survivor_plab_stats(YoungPLABSize, PLABWeight), 1941 _old_plab_stats(OldPLABSize, PLABWeight), 1942 _expand_heap_after_alloc_failure(true), 1943 _surviving_young_words(NULL), 1944 _old_marking_cycles_started(0), 1945 _old_marking_cycles_completed(0), 1946 _concurrent_cycle_started(false), 1947 _in_cset_fast_test(NULL), 1948 _in_cset_fast_test_base(NULL), 1949 _dirty_cards_region_list(NULL), 1950 _worker_cset_start_region(NULL), 1951 _worker_cset_start_region_time_stamp(NULL), 1952 _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()), 1953 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 1954 _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()), 1955 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()) { 1956 1957 _g1h = this; 1958 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1959 vm_exit_during_initialization("Failed necessary allocation."); 1960 } 1961 1962 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1963 1964 int n_queues = MAX2((int)ParallelGCThreads, 1); 1965 _task_queues = new RefToScanQueueSet(n_queues); 1966 1967 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1968 assert(n_rem_sets > 0, "Invariant."); 1969 1970 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1971 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1972 _evacuation_failed_info_array = NEW_C_HEAP_ARRAY(EvacuationFailedInfo, n_queues, mtGC); 1973 1974 for (int i = 0; i < n_queues; i++) { 1975 RefToScanQueue* q = new RefToScanQueue(); 1976 q->initialize(); 1977 _task_queues->register_queue(i, q); 1978 ::new (&_evacuation_failed_info_array[i]) EvacuationFailedInfo(); 1979 } 1980 clear_cset_start_regions(); 1981 1982 // Initialize the G1EvacuationFailureALot counters and flags. 1983 NOT_PRODUCT(reset_evacuation_should_fail();) 1984 1985 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1986 } 1987 1988 jint G1CollectedHeap::initialize() { 1989 CollectedHeap::pre_initialize(); 1990 os::enable_vtime(); 1991 1992 G1Log::init(); 1993 1994 // Necessary to satisfy locking discipline assertions. 1995 1996 MutexLocker x(Heap_lock); 1997 1998 // We have to initialize the printer before committing the heap, as 1999 // it will be used then. 2000 _hr_printer.set_active(G1PrintHeapRegions); 2001 2002 // While there are no constraints in the GC code that HeapWordSize 2003 // be any particular value, there are multiple other areas in the 2004 // system which believe this to be true (e.g. oop->object_size in some 2005 // cases incorrectly returns the size in wordSize units rather than 2006 // HeapWordSize). 2007 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 2008 2009 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 2010 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 2011 size_t heap_alignment = collector_policy()->max_alignment(); 2012 2013 // Ensure that the sizes are properly aligned. 2014 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2015 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2016 Universe::check_alignment(max_byte_size, heap_alignment, "g1 heap"); 2017 2018 _cg1r = new ConcurrentG1Refine(this); 2019 2020 // Reserve the maximum. 2021 2022 // When compressed oops are enabled, the preferred heap base 2023 // is calculated by subtracting the requested size from the 2024 // 32Gb boundary and using the result as the base address for 2025 // heap reservation. If the requested size is not aligned to 2026 // HeapRegion::GrainBytes (i.e. the alignment that is passed 2027 // into the ReservedHeapSpace constructor) then the actual 2028 // base of the reserved heap may end up differing from the 2029 // address that was requested (i.e. the preferred heap base). 2030 // If this happens then we could end up using a non-optimal 2031 // compressed oops mode. 2032 2033 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2034 heap_alignment); 2035 2036 // It is important to do this in a way such that concurrent readers can't 2037 // temporarily think something is in the heap. (I've actually seen this 2038 // happen in asserts: DLD.) 2039 _reserved.set_word_size(0); 2040 _reserved.set_start((HeapWord*)heap_rs.base()); 2041 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2042 2043 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2044 2045 // Create the gen rem set (and barrier set) for the entire reserved region. 2046 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2047 set_barrier_set(rem_set()->bs()); 2048 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2049 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2050 } else { 2051 vm_exit_during_initialization("G1 requires a mod ref bs."); 2052 return JNI_ENOMEM; 2053 } 2054 2055 // Also create a G1 rem set. 2056 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2057 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2058 } else { 2059 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2060 return JNI_ENOMEM; 2061 } 2062 2063 // Carve out the G1 part of the heap. 2064 2065 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2066 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2067 g1_rs.size()/HeapWordSize); 2068 2069 _g1_storage.initialize(g1_rs, 0); 2070 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2071 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2072 (HeapWord*) _g1_reserved.end(), 2073 _expansion_regions); 2074 2075 // Do later initialization work for concurrent refinement. 2076 _cg1r->init(); 2077 2078 // 6843694 - ensure that the maximum region index can fit 2079 // in the remembered set structures. 2080 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2081 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2082 2083 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2084 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2085 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2086 "too many cards per region"); 2087 2088 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2089 2090 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2091 heap_word_size(init_byte_size)); 2092 2093 _g1h = this; 2094 2095 _in_cset_fast_test_length = max_regions(); 2096 _in_cset_fast_test_base = 2097 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2098 2099 // We're biasing _in_cset_fast_test to avoid subtracting the 2100 // beginning of the heap every time we want to index; basically 2101 // it's the same with what we do with the card table. 2102 _in_cset_fast_test = _in_cset_fast_test_base - 2103 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2104 2105 // Clear the _cset_fast_test bitmap in anticipation of adding 2106 // regions to the incremental collection set for the first 2107 // evacuation pause. 2108 clear_cset_fast_test(); 2109 2110 // Create the ConcurrentMark data structure and thread. 2111 // (Must do this late, so that "max_regions" is defined.) 2112 _cm = new ConcurrentMark(this, heap_rs); 2113 if (_cm == NULL || !_cm->completed_initialization()) { 2114 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2115 return JNI_ENOMEM; 2116 } 2117 _cmThread = _cm->cmThread(); 2118 2119 // Initialize the from_card cache structure of HeapRegionRemSet. 2120 HeapRegionRemSet::init_heap(max_regions()); 2121 2122 // Now expand into the initial heap size. 2123 if (!expand(init_byte_size)) { 2124 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2125 return JNI_ENOMEM; 2126 } 2127 2128 // Perform any initialization actions delegated to the policy. 2129 g1_policy()->init(); 2130 2131 _refine_cte_cl = 2132 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2133 g1_rem_set(), 2134 concurrent_g1_refine()); 2135 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2136 2137 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2138 SATB_Q_FL_lock, 2139 G1SATBProcessCompletedThreshold, 2140 Shared_SATB_Q_lock); 2141 2142 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2143 DirtyCardQ_FL_lock, 2144 concurrent_g1_refine()->yellow_zone(), 2145 concurrent_g1_refine()->red_zone(), 2146 Shared_DirtyCardQ_lock); 2147 2148 if (G1DeferredRSUpdate) { 2149 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2150 DirtyCardQ_FL_lock, 2151 -1, // never trigger processing 2152 -1, // no limit on length 2153 Shared_DirtyCardQ_lock, 2154 &JavaThread::dirty_card_queue_set()); 2155 } 2156 2157 // Initialize the card queue set used to hold cards containing 2158 // references into the collection set. 2159 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2160 DirtyCardQ_FL_lock, 2161 -1, // never trigger processing 2162 -1, // no limit on length 2163 Shared_DirtyCardQ_lock, 2164 &JavaThread::dirty_card_queue_set()); 2165 2166 // In case we're keeping closure specialization stats, initialize those 2167 // counts and that mechanism. 2168 SpecializationStats::clear(); 2169 2170 // Here we allocate the dummy full region that is required by the 2171 // G1AllocRegion class. If we don't pass an address in the reserved 2172 // space here, lots of asserts fire. 2173 2174 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2175 _g1_reserved.start()); 2176 // We'll re-use the same region whether the alloc region will 2177 // require BOT updates or not and, if it doesn't, then a non-young 2178 // region will complain that it cannot support allocations without 2179 // BOT updates. So we'll tag the dummy region as young to avoid that. 2180 dummy_region->set_young(); 2181 // Make sure it's full. 2182 dummy_region->set_top(dummy_region->end()); 2183 G1AllocRegion::setup(this, dummy_region); 2184 2185 init_mutator_alloc_region(); 2186 2187 // Do create of the monitoring and management support so that 2188 // values in the heap have been properly initialized. 2189 _g1mm = new G1MonitoringSupport(this); 2190 2191 return JNI_OK; 2192 } 2193 2194 void G1CollectedHeap::ref_processing_init() { 2195 // Reference processing in G1 currently works as follows: 2196 // 2197 // * There are two reference processor instances. One is 2198 // used to record and process discovered references 2199 // during concurrent marking; the other is used to 2200 // record and process references during STW pauses 2201 // (both full and incremental). 2202 // * Both ref processors need to 'span' the entire heap as 2203 // the regions in the collection set may be dotted around. 2204 // 2205 // * For the concurrent marking ref processor: 2206 // * Reference discovery is enabled at initial marking. 2207 // * Reference discovery is disabled and the discovered 2208 // references processed etc during remarking. 2209 // * Reference discovery is MT (see below). 2210 // * Reference discovery requires a barrier (see below). 2211 // * Reference processing may or may not be MT 2212 // (depending on the value of ParallelRefProcEnabled 2213 // and ParallelGCThreads). 2214 // * A full GC disables reference discovery by the CM 2215 // ref processor and abandons any entries on it's 2216 // discovered lists. 2217 // 2218 // * For the STW processor: 2219 // * Non MT discovery is enabled at the start of a full GC. 2220 // * Processing and enqueueing during a full GC is non-MT. 2221 // * During a full GC, references are processed after marking. 2222 // 2223 // * Discovery (may or may not be MT) is enabled at the start 2224 // of an incremental evacuation pause. 2225 // * References are processed near the end of a STW evacuation pause. 2226 // * For both types of GC: 2227 // * Discovery is atomic - i.e. not concurrent. 2228 // * Reference discovery will not need a barrier. 2229 2230 SharedHeap::ref_processing_init(); 2231 MemRegion mr = reserved_region(); 2232 2233 // Concurrent Mark ref processor 2234 _ref_processor_cm = 2235 new ReferenceProcessor(mr, // span 2236 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2237 // mt processing 2238 (int) ParallelGCThreads, 2239 // degree of mt processing 2240 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2241 // mt discovery 2242 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2243 // degree of mt discovery 2244 false, 2245 // Reference discovery is not atomic 2246 &_is_alive_closure_cm, 2247 // is alive closure 2248 // (for efficiency/performance) 2249 true); 2250 // Setting next fields of discovered 2251 // lists requires a barrier. 2252 2253 // STW ref processor 2254 _ref_processor_stw = 2255 new ReferenceProcessor(mr, // span 2256 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2257 // mt processing 2258 MAX2((int)ParallelGCThreads, 1), 2259 // degree of mt processing 2260 (ParallelGCThreads > 1), 2261 // mt discovery 2262 MAX2((int)ParallelGCThreads, 1), 2263 // degree of mt discovery 2264 true, 2265 // Reference discovery is atomic 2266 &_is_alive_closure_stw, 2267 // is alive closure 2268 // (for efficiency/performance) 2269 false); 2270 // Setting next fields of discovered 2271 // lists requires a barrier. 2272 } 2273 2274 size_t G1CollectedHeap::capacity() const { 2275 return _g1_committed.byte_size(); 2276 } 2277 2278 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2279 assert(!hr->continuesHumongous(), "pre-condition"); 2280 hr->reset_gc_time_stamp(); 2281 if (hr->startsHumongous()) { 2282 uint first_index = hr->hrs_index() + 1; 2283 uint last_index = hr->last_hc_index(); 2284 for (uint i = first_index; i < last_index; i += 1) { 2285 HeapRegion* chr = region_at(i); 2286 assert(chr->continuesHumongous(), "sanity"); 2287 chr->reset_gc_time_stamp(); 2288 } 2289 } 2290 } 2291 2292 #ifndef PRODUCT 2293 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2294 private: 2295 unsigned _gc_time_stamp; 2296 bool _failures; 2297 2298 public: 2299 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2300 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2301 2302 virtual bool doHeapRegion(HeapRegion* hr) { 2303 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2304 if (_gc_time_stamp != region_gc_time_stamp) { 2305 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2306 "expected %d", HR_FORMAT_PARAMS(hr), 2307 region_gc_time_stamp, _gc_time_stamp); 2308 _failures = true; 2309 } 2310 return false; 2311 } 2312 2313 bool failures() { return _failures; } 2314 }; 2315 2316 void G1CollectedHeap::check_gc_time_stamps() { 2317 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2318 heap_region_iterate(&cl); 2319 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2320 } 2321 #endif // PRODUCT 2322 2323 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2324 DirtyCardQueue* into_cset_dcq, 2325 bool concurrent, 2326 uint worker_i) { 2327 // Clean cards in the hot card cache 2328 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 2329 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq); 2330 2331 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2332 int n_completed_buffers = 0; 2333 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2334 n_completed_buffers++; 2335 } 2336 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2337 dcqs.clear_n_completed_buffers(); 2338 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2339 } 2340 2341 2342 // Computes the sum of the storage used by the various regions. 2343 2344 size_t G1CollectedHeap::used() const { 2345 assert(Heap_lock->owner() != NULL, 2346 "Should be owned on this thread's behalf."); 2347 size_t result = _summary_bytes_used; 2348 // Read only once in case it is set to NULL concurrently 2349 HeapRegion* hr = _mutator_alloc_region.get(); 2350 if (hr != NULL) 2351 result += hr->used(); 2352 return result; 2353 } 2354 2355 size_t G1CollectedHeap::used_unlocked() const { 2356 size_t result = _summary_bytes_used; 2357 return result; 2358 } 2359 2360 class SumUsedClosure: public HeapRegionClosure { 2361 size_t _used; 2362 public: 2363 SumUsedClosure() : _used(0) {} 2364 bool doHeapRegion(HeapRegion* r) { 2365 if (!r->continuesHumongous()) { 2366 _used += r->used(); 2367 } 2368 return false; 2369 } 2370 size_t result() { return _used; } 2371 }; 2372 2373 size_t G1CollectedHeap::recalculate_used() const { 2374 SumUsedClosure blk; 2375 heap_region_iterate(&blk); 2376 return blk.result(); 2377 } 2378 2379 size_t G1CollectedHeap::unsafe_max_alloc() { 2380 if (free_regions() > 0) return HeapRegion::GrainBytes; 2381 // otherwise, is there space in the current allocation region? 2382 2383 // We need to store the current allocation region in a local variable 2384 // here. The problem is that this method doesn't take any locks and 2385 // there may be other threads which overwrite the current allocation 2386 // region field. attempt_allocation(), for example, sets it to NULL 2387 // and this can happen *after* the NULL check here but before the call 2388 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2389 // to be a problem in the optimized build, since the two loads of the 2390 // current allocation region field are optimized away. 2391 HeapRegion* hr = _mutator_alloc_region.get(); 2392 if (hr == NULL) { 2393 return 0; 2394 } 2395 return hr->free(); 2396 } 2397 2398 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2399 switch (cause) { 2400 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2401 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2402 case GCCause::_g1_humongous_allocation: return true; 2403 default: return false; 2404 } 2405 } 2406 2407 #ifndef PRODUCT 2408 void G1CollectedHeap::allocate_dummy_regions() { 2409 // Let's fill up most of the region 2410 size_t word_size = HeapRegion::GrainWords - 1024; 2411 // And as a result the region we'll allocate will be humongous. 2412 guarantee(isHumongous(word_size), "sanity"); 2413 2414 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2415 // Let's use the existing mechanism for the allocation 2416 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2417 if (dummy_obj != NULL) { 2418 MemRegion mr(dummy_obj, word_size); 2419 CollectedHeap::fill_with_object(mr); 2420 } else { 2421 // If we can't allocate once, we probably cannot allocate 2422 // again. Let's get out of the loop. 2423 break; 2424 } 2425 } 2426 } 2427 #endif // !PRODUCT 2428 2429 void G1CollectedHeap::increment_old_marking_cycles_started() { 2430 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2431 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2432 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2433 _old_marking_cycles_started, _old_marking_cycles_completed)); 2434 2435 _old_marking_cycles_started++; 2436 } 2437 2438 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2439 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2440 2441 // We assume that if concurrent == true, then the caller is a 2442 // concurrent thread that was joined the Suspendible Thread 2443 // Set. If there's ever a cheap way to check this, we should add an 2444 // assert here. 2445 2446 // Given that this method is called at the end of a Full GC or of a 2447 // concurrent cycle, and those can be nested (i.e., a Full GC can 2448 // interrupt a concurrent cycle), the number of full collections 2449 // completed should be either one (in the case where there was no 2450 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2451 // behind the number of full collections started. 2452 2453 // This is the case for the inner caller, i.e. a Full GC. 2454 assert(concurrent || 2455 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2456 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2457 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2458 "is inconsistent with _old_marking_cycles_completed = %u", 2459 _old_marking_cycles_started, _old_marking_cycles_completed)); 2460 2461 // This is the case for the outer caller, i.e. the concurrent cycle. 2462 assert(!concurrent || 2463 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2464 err_msg("for outer caller (concurrent cycle): " 2465 "_old_marking_cycles_started = %u " 2466 "is inconsistent with _old_marking_cycles_completed = %u", 2467 _old_marking_cycles_started, _old_marking_cycles_completed)); 2468 2469 _old_marking_cycles_completed += 1; 2470 2471 // We need to clear the "in_progress" flag in the CM thread before 2472 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2473 // is set) so that if a waiter requests another System.gc() it doesn't 2474 // incorrectly see that a marking cycle is still in progress. 2475 if (concurrent) { 2476 _cmThread->clear_in_progress(); 2477 } 2478 2479 // This notify_all() will ensure that a thread that called 2480 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2481 // and it's waiting for a full GC to finish will be woken up. It is 2482 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2483 FullGCCount_lock->notify_all(); 2484 } 2485 2486 void G1CollectedHeap::register_concurrent_cycle_start(jlong start_time) { 2487 _concurrent_cycle_started = true; 2488 _gc_timer_cm->register_gc_start(start_time); 2489 2490 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start()); 2491 trace_heap_before_gc(_gc_tracer_cm); 2492 } 2493 2494 void G1CollectedHeap::register_concurrent_cycle_end() { 2495 if (_concurrent_cycle_started) { 2496 if (_cm->has_aborted()) { 2497 _gc_tracer_cm->report_concurrent_mode_failure(); 2498 } 2499 2500 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2501 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2502 2503 _concurrent_cycle_started = false; 2504 } 2505 } 2506 2507 void G1CollectedHeap::trace_heap_after_concurrent_cycle() { 2508 if (_concurrent_cycle_started) { 2509 trace_heap_after_gc(_gc_tracer_cm); 2510 } 2511 } 2512 2513 G1YCType G1CollectedHeap::yc_type() { 2514 bool is_young = g1_policy()->gcs_are_young(); 2515 bool is_initial_mark = g1_policy()->during_initial_mark_pause(); 2516 bool is_during_mark = mark_in_progress(); 2517 2518 if (is_initial_mark) { 2519 return InitialMark; 2520 } else if (is_during_mark) { 2521 return DuringMark; 2522 } else if (is_young) { 2523 return Normal; 2524 } else { 2525 return Mixed; 2526 } 2527 } 2528 2529 void G1CollectedHeap::collect(GCCause::Cause cause) { 2530 assert_heap_not_locked(); 2531 2532 unsigned int gc_count_before; 2533 unsigned int old_marking_count_before; 2534 bool retry_gc; 2535 2536 do { 2537 retry_gc = false; 2538 2539 { 2540 MutexLocker ml(Heap_lock); 2541 2542 // Read the GC count while holding the Heap_lock 2543 gc_count_before = total_collections(); 2544 old_marking_count_before = _old_marking_cycles_started; 2545 } 2546 2547 if (should_do_concurrent_full_gc(cause)) { 2548 // Schedule an initial-mark evacuation pause that will start a 2549 // concurrent cycle. We're setting word_size to 0 which means that 2550 // we are not requesting a post-GC allocation. 2551 VM_G1IncCollectionPause op(gc_count_before, 2552 0, /* word_size */ 2553 true, /* should_initiate_conc_mark */ 2554 g1_policy()->max_pause_time_ms(), 2555 cause); 2556 2557 VMThread::execute(&op); 2558 if (!op.pause_succeeded()) { 2559 if (old_marking_count_before == _old_marking_cycles_started) { 2560 retry_gc = op.should_retry_gc(); 2561 } else { 2562 // A Full GC happened while we were trying to schedule the 2563 // initial-mark GC. No point in starting a new cycle given 2564 // that the whole heap was collected anyway. 2565 } 2566 2567 if (retry_gc) { 2568 if (GC_locker::is_active_and_needs_gc()) { 2569 GC_locker::stall_until_clear(); 2570 } 2571 } 2572 } 2573 } else { 2574 if (cause == GCCause::_gc_locker 2575 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2576 2577 // Schedule a standard evacuation pause. We're setting word_size 2578 // to 0 which means that we are not requesting a post-GC allocation. 2579 VM_G1IncCollectionPause op(gc_count_before, 2580 0, /* word_size */ 2581 false, /* should_initiate_conc_mark */ 2582 g1_policy()->max_pause_time_ms(), 2583 cause); 2584 VMThread::execute(&op); 2585 } else { 2586 // Schedule a Full GC. 2587 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2588 VMThread::execute(&op); 2589 } 2590 } 2591 } while (retry_gc); 2592 } 2593 2594 bool G1CollectedHeap::is_in(const void* p) const { 2595 if (_g1_committed.contains(p)) { 2596 // Given that we know that p is in the committed space, 2597 // heap_region_containing_raw() should successfully 2598 // return the containing region. 2599 HeapRegion* hr = heap_region_containing_raw(p); 2600 return hr->is_in(p); 2601 } else { 2602 return false; 2603 } 2604 } 2605 2606 // Iteration functions. 2607 2608 // Iterates an OopClosure over all ref-containing fields of objects 2609 // within a HeapRegion. 2610 2611 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2612 MemRegion _mr; 2613 ExtendedOopClosure* _cl; 2614 public: 2615 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2616 : _mr(mr), _cl(cl) {} 2617 bool doHeapRegion(HeapRegion* r) { 2618 if (!r->continuesHumongous()) { 2619 r->oop_iterate(_cl); 2620 } 2621 return false; 2622 } 2623 }; 2624 2625 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2626 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2627 heap_region_iterate(&blk); 2628 } 2629 2630 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2631 IterateOopClosureRegionClosure blk(mr, cl); 2632 heap_region_iterate(&blk); 2633 } 2634 2635 // Iterates an ObjectClosure over all objects within a HeapRegion. 2636 2637 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2638 ObjectClosure* _cl; 2639 public: 2640 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2641 bool doHeapRegion(HeapRegion* r) { 2642 if (! r->continuesHumongous()) { 2643 r->object_iterate(_cl); 2644 } 2645 return false; 2646 } 2647 }; 2648 2649 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2650 IterateObjectClosureRegionClosure blk(cl); 2651 heap_region_iterate(&blk); 2652 } 2653 2654 // Calls a SpaceClosure on a HeapRegion. 2655 2656 class SpaceClosureRegionClosure: public HeapRegionClosure { 2657 SpaceClosure* _cl; 2658 public: 2659 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2660 bool doHeapRegion(HeapRegion* r) { 2661 _cl->do_space(r); 2662 return false; 2663 } 2664 }; 2665 2666 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2667 SpaceClosureRegionClosure blk(cl); 2668 heap_region_iterate(&blk); 2669 } 2670 2671 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2672 _hrs.iterate(cl); 2673 } 2674 2675 void 2676 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2677 uint worker_id, 2678 uint no_of_par_workers, 2679 jint claim_value) { 2680 const uint regions = n_regions(); 2681 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2682 no_of_par_workers : 2683 1); 2684 assert(UseDynamicNumberOfGCThreads || 2685 no_of_par_workers == workers()->total_workers(), 2686 "Non dynamic should use fixed number of workers"); 2687 // try to spread out the starting points of the workers 2688 const HeapRegion* start_hr = 2689 start_region_for_worker(worker_id, no_of_par_workers); 2690 const uint start_index = start_hr->hrs_index(); 2691 2692 // each worker will actually look at all regions 2693 for (uint count = 0; count < regions; ++count) { 2694 const uint index = (start_index + count) % regions; 2695 assert(0 <= index && index < regions, "sanity"); 2696 HeapRegion* r = region_at(index); 2697 // we'll ignore "continues humongous" regions (we'll process them 2698 // when we come across their corresponding "start humongous" 2699 // region) and regions already claimed 2700 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2701 continue; 2702 } 2703 // OK, try to claim it 2704 if (r->claimHeapRegion(claim_value)) { 2705 // success! 2706 assert(!r->continuesHumongous(), "sanity"); 2707 if (r->startsHumongous()) { 2708 // If the region is "starts humongous" we'll iterate over its 2709 // "continues humongous" first; in fact we'll do them 2710 // first. The order is important. In on case, calling the 2711 // closure on the "starts humongous" region might de-allocate 2712 // and clear all its "continues humongous" regions and, as a 2713 // result, we might end up processing them twice. So, we'll do 2714 // them first (notice: most closures will ignore them anyway) and 2715 // then we'll do the "starts humongous" region. 2716 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2717 HeapRegion* chr = region_at(ch_index); 2718 2719 // if the region has already been claimed or it's not 2720 // "continues humongous" we're done 2721 if (chr->claim_value() == claim_value || 2722 !chr->continuesHumongous()) { 2723 break; 2724 } 2725 2726 // No one should have claimed it directly. We can given 2727 // that we claimed its "starts humongous" region. 2728 assert(chr->claim_value() != claim_value, "sanity"); 2729 assert(chr->humongous_start_region() == r, "sanity"); 2730 2731 if (chr->claimHeapRegion(claim_value)) { 2732 // we should always be able to claim it; no one else should 2733 // be trying to claim this region 2734 2735 bool res2 = cl->doHeapRegion(chr); 2736 assert(!res2, "Should not abort"); 2737 2738 // Right now, this holds (i.e., no closure that actually 2739 // does something with "continues humongous" regions 2740 // clears them). We might have to weaken it in the future, 2741 // but let's leave these two asserts here for extra safety. 2742 assert(chr->continuesHumongous(), "should still be the case"); 2743 assert(chr->humongous_start_region() == r, "sanity"); 2744 } else { 2745 guarantee(false, "we should not reach here"); 2746 } 2747 } 2748 } 2749 2750 assert(!r->continuesHumongous(), "sanity"); 2751 bool res = cl->doHeapRegion(r); 2752 assert(!res, "Should not abort"); 2753 } 2754 } 2755 } 2756 2757 class ResetClaimValuesClosure: public HeapRegionClosure { 2758 public: 2759 bool doHeapRegion(HeapRegion* r) { 2760 r->set_claim_value(HeapRegion::InitialClaimValue); 2761 return false; 2762 } 2763 }; 2764 2765 void G1CollectedHeap::reset_heap_region_claim_values() { 2766 ResetClaimValuesClosure blk; 2767 heap_region_iterate(&blk); 2768 } 2769 2770 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2771 ResetClaimValuesClosure blk; 2772 collection_set_iterate(&blk); 2773 } 2774 2775 #ifdef ASSERT 2776 // This checks whether all regions in the heap have the correct claim 2777 // value. I also piggy-backed on this a check to ensure that the 2778 // humongous_start_region() information on "continues humongous" 2779 // regions is correct. 2780 2781 class CheckClaimValuesClosure : public HeapRegionClosure { 2782 private: 2783 jint _claim_value; 2784 uint _failures; 2785 HeapRegion* _sh_region; 2786 2787 public: 2788 CheckClaimValuesClosure(jint claim_value) : 2789 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2790 bool doHeapRegion(HeapRegion* r) { 2791 if (r->claim_value() != _claim_value) { 2792 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2793 "claim value = %d, should be %d", 2794 HR_FORMAT_PARAMS(r), 2795 r->claim_value(), _claim_value); 2796 ++_failures; 2797 } 2798 if (!r->isHumongous()) { 2799 _sh_region = NULL; 2800 } else if (r->startsHumongous()) { 2801 _sh_region = r; 2802 } else if (r->continuesHumongous()) { 2803 if (r->humongous_start_region() != _sh_region) { 2804 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2805 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2806 HR_FORMAT_PARAMS(r), 2807 r->humongous_start_region(), 2808 _sh_region); 2809 ++_failures; 2810 } 2811 } 2812 return false; 2813 } 2814 uint failures() { return _failures; } 2815 }; 2816 2817 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2818 CheckClaimValuesClosure cl(claim_value); 2819 heap_region_iterate(&cl); 2820 return cl.failures() == 0; 2821 } 2822 2823 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2824 private: 2825 jint _claim_value; 2826 uint _failures; 2827 2828 public: 2829 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2830 _claim_value(claim_value), _failures(0) { } 2831 2832 uint failures() { return _failures; } 2833 2834 bool doHeapRegion(HeapRegion* hr) { 2835 assert(hr->in_collection_set(), "how?"); 2836 assert(!hr->isHumongous(), "H-region in CSet"); 2837 if (hr->claim_value() != _claim_value) { 2838 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2839 "claim value = %d, should be %d", 2840 HR_FORMAT_PARAMS(hr), 2841 hr->claim_value(), _claim_value); 2842 _failures += 1; 2843 } 2844 return false; 2845 } 2846 }; 2847 2848 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2849 CheckClaimValuesInCSetHRClosure cl(claim_value); 2850 collection_set_iterate(&cl); 2851 return cl.failures() == 0; 2852 } 2853 #endif // ASSERT 2854 2855 // Clear the cached CSet starting regions and (more importantly) 2856 // the time stamps. Called when we reset the GC time stamp. 2857 void G1CollectedHeap::clear_cset_start_regions() { 2858 assert(_worker_cset_start_region != NULL, "sanity"); 2859 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2860 2861 int n_queues = MAX2((int)ParallelGCThreads, 1); 2862 for (int i = 0; i < n_queues; i++) { 2863 _worker_cset_start_region[i] = NULL; 2864 _worker_cset_start_region_time_stamp[i] = 0; 2865 } 2866 } 2867 2868 // Given the id of a worker, obtain or calculate a suitable 2869 // starting region for iterating over the current collection set. 2870 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(uint worker_i) { 2871 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2872 2873 HeapRegion* result = NULL; 2874 unsigned gc_time_stamp = get_gc_time_stamp(); 2875 2876 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2877 // Cached starting region for current worker was set 2878 // during the current pause - so it's valid. 2879 // Note: the cached starting heap region may be NULL 2880 // (when the collection set is empty). 2881 result = _worker_cset_start_region[worker_i]; 2882 assert(result == NULL || result->in_collection_set(), "sanity"); 2883 return result; 2884 } 2885 2886 // The cached entry was not valid so let's calculate 2887 // a suitable starting heap region for this worker. 2888 2889 // We want the parallel threads to start their collection 2890 // set iteration at different collection set regions to 2891 // avoid contention. 2892 // If we have: 2893 // n collection set regions 2894 // p threads 2895 // Then thread t will start at region floor ((t * n) / p) 2896 2897 result = g1_policy()->collection_set(); 2898 if (G1CollectedHeap::use_parallel_gc_threads()) { 2899 uint cs_size = g1_policy()->cset_region_length(); 2900 uint active_workers = workers()->active_workers(); 2901 assert(UseDynamicNumberOfGCThreads || 2902 active_workers == workers()->total_workers(), 2903 "Unless dynamic should use total workers"); 2904 2905 uint end_ind = (cs_size * worker_i) / active_workers; 2906 uint start_ind = 0; 2907 2908 if (worker_i > 0 && 2909 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2910 // Previous workers starting region is valid 2911 // so let's iterate from there 2912 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2913 result = _worker_cset_start_region[worker_i - 1]; 2914 } 2915 2916 for (uint i = start_ind; i < end_ind; i++) { 2917 result = result->next_in_collection_set(); 2918 } 2919 } 2920 2921 // Note: the calculated starting heap region may be NULL 2922 // (when the collection set is empty). 2923 assert(result == NULL || result->in_collection_set(), "sanity"); 2924 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2925 "should be updated only once per pause"); 2926 _worker_cset_start_region[worker_i] = result; 2927 OrderAccess::storestore(); 2928 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2929 return result; 2930 } 2931 2932 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2933 uint no_of_par_workers) { 2934 uint worker_num = 2935 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2936 assert(UseDynamicNumberOfGCThreads || 2937 no_of_par_workers == workers()->total_workers(), 2938 "Non dynamic should use fixed number of workers"); 2939 const uint start_index = n_regions() * worker_i / worker_num; 2940 return region_at(start_index); 2941 } 2942 2943 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2944 HeapRegion* r = g1_policy()->collection_set(); 2945 while (r != NULL) { 2946 HeapRegion* next = r->next_in_collection_set(); 2947 if (cl->doHeapRegion(r)) { 2948 cl->incomplete(); 2949 return; 2950 } 2951 r = next; 2952 } 2953 } 2954 2955 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2956 HeapRegionClosure *cl) { 2957 if (r == NULL) { 2958 // The CSet is empty so there's nothing to do. 2959 return; 2960 } 2961 2962 assert(r->in_collection_set(), 2963 "Start region must be a member of the collection set."); 2964 HeapRegion* cur = r; 2965 while (cur != NULL) { 2966 HeapRegion* next = cur->next_in_collection_set(); 2967 if (cl->doHeapRegion(cur) && false) { 2968 cl->incomplete(); 2969 return; 2970 } 2971 cur = next; 2972 } 2973 cur = g1_policy()->collection_set(); 2974 while (cur != r) { 2975 HeapRegion* next = cur->next_in_collection_set(); 2976 if (cl->doHeapRegion(cur) && false) { 2977 cl->incomplete(); 2978 return; 2979 } 2980 cur = next; 2981 } 2982 } 2983 2984 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2985 return n_regions() > 0 ? region_at(0) : NULL; 2986 } 2987 2988 2989 Space* G1CollectedHeap::space_containing(const void* addr) const { 2990 Space* res = heap_region_containing(addr); 2991 return res; 2992 } 2993 2994 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2995 Space* sp = space_containing(addr); 2996 if (sp != NULL) { 2997 return sp->block_start(addr); 2998 } 2999 return NULL; 3000 } 3001 3002 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 3003 Space* sp = space_containing(addr); 3004 assert(sp != NULL, "block_size of address outside of heap"); 3005 return sp->block_size(addr); 3006 } 3007 3008 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 3009 Space* sp = space_containing(addr); 3010 return sp->block_is_obj(addr); 3011 } 3012 3013 bool G1CollectedHeap::supports_tlab_allocation() const { 3014 return true; 3015 } 3016 3017 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 3018 return HeapRegion::GrainBytes; 3019 } 3020 3021 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 3022 // Return the remaining space in the cur alloc region, but not less than 3023 // the min TLAB size. 3024 3025 // Also, this value can be at most the humongous object threshold, 3026 // since we can't allow tlabs to grow big enough to accommodate 3027 // humongous objects. 3028 3029 HeapRegion* hr = _mutator_alloc_region.get(); 3030 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 3031 if (hr == NULL) { 3032 return max_tlab_size; 3033 } else { 3034 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 3035 } 3036 } 3037 3038 size_t G1CollectedHeap::max_capacity() const { 3039 return _g1_reserved.byte_size(); 3040 } 3041 3042 jlong G1CollectedHeap::millis_since_last_gc() { 3043 // assert(false, "NYI"); 3044 return 0; 3045 } 3046 3047 void G1CollectedHeap::prepare_for_verify() { 3048 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3049 ensure_parsability(false); 3050 } 3051 g1_rem_set()->prepare_for_verify(); 3052 } 3053 3054 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3055 VerifyOption vo) { 3056 switch (vo) { 3057 case VerifyOption_G1UsePrevMarking: 3058 return hr->obj_allocated_since_prev_marking(obj); 3059 case VerifyOption_G1UseNextMarking: 3060 return hr->obj_allocated_since_next_marking(obj); 3061 case VerifyOption_G1UseMarkWord: 3062 return false; 3063 default: 3064 ShouldNotReachHere(); 3065 } 3066 return false; // keep some compilers happy 3067 } 3068 3069 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3070 switch (vo) { 3071 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3072 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3073 case VerifyOption_G1UseMarkWord: return NULL; 3074 default: ShouldNotReachHere(); 3075 } 3076 return NULL; // keep some compilers happy 3077 } 3078 3079 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3080 switch (vo) { 3081 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3082 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3083 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3084 default: ShouldNotReachHere(); 3085 } 3086 return false; // keep some compilers happy 3087 } 3088 3089 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3090 switch (vo) { 3091 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3092 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3093 case VerifyOption_G1UseMarkWord: return "NONE"; 3094 default: ShouldNotReachHere(); 3095 } 3096 return NULL; // keep some compilers happy 3097 } 3098 3099 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can 3100 // pass it as the perm_blk to SharedHeap::process_strong_roots. 3101 // When process_strong_roots stop calling perm_blk->younger_refs_iterate 3102 // we can change this closure to extend the simpler OopClosure. 3103 class VerifyRootsClosure: public OopsInGenClosure { 3104 private: 3105 G1CollectedHeap* _g1h; 3106 VerifyOption _vo; 3107 bool _failures; 3108 public: 3109 // _vo == UsePrevMarking -> use "prev" marking information, 3110 // _vo == UseNextMarking -> use "next" marking information, 3111 // _vo == UseMarkWord -> use mark word from object header. 3112 VerifyRootsClosure(VerifyOption vo) : 3113 _g1h(G1CollectedHeap::heap()), 3114 _vo(vo), 3115 _failures(false) { } 3116 3117 bool failures() { return _failures; } 3118 3119 template <class T> void do_oop_nv(T* p) { 3120 T heap_oop = oopDesc::load_heap_oop(p); 3121 if (!oopDesc::is_null(heap_oop)) { 3122 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3123 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3124 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3125 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3126 if (_vo == VerifyOption_G1UseMarkWord) { 3127 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3128 } 3129 obj->print_on(gclog_or_tty); 3130 _failures = true; 3131 } 3132 } 3133 } 3134 3135 void do_oop(oop* p) { do_oop_nv(p); } 3136 void do_oop(narrowOop* p) { do_oop_nv(p); } 3137 }; 3138 3139 class G1VerifyCodeRootOopClosure: public OopsInGenClosure { 3140 G1CollectedHeap* _g1h; 3141 OopClosure* _root_cl; 3142 nmethod* _nm; 3143 VerifyOption _vo; 3144 bool _failures; 3145 3146 template <class T> void do_oop_work(T* p) { 3147 // First verify that this root is live 3148 _root_cl->do_oop(p); 3149 3150 if (!G1VerifyHeapRegionCodeRoots) { 3151 // We're not verifying the code roots attached to heap region. 3152 return; 3153 } 3154 3155 // Don't check the code roots during marking verification in a full GC 3156 if (_vo == VerifyOption_G1UseMarkWord) { 3157 return; 3158 } 3159 3160 // Now verify that the current nmethod (which contains p) is 3161 // in the code root list of the heap region containing the 3162 // object referenced by p. 3163 3164 T heap_oop = oopDesc::load_heap_oop(p); 3165 if (!oopDesc::is_null(heap_oop)) { 3166 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3167 3168 // Now fetch the region containing the object 3169 HeapRegion* hr = _g1h->heap_region_containing(obj); 3170 HeapRegionRemSet* hrrs = hr->rem_set(); 3171 // Verify that the strong code root list for this region 3172 // contains the nmethod 3173 if (!hrrs->strong_code_roots_list_contains(_nm)) { 3174 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" " 3175 "from nmethod "PTR_FORMAT" not in strong " 3176 "code roots for region ["PTR_FORMAT","PTR_FORMAT")", 3177 p, _nm, hr->bottom(), hr->end()); 3178 _failures = true; 3179 } 3180 } 3181 } 3182 3183 public: 3184 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo): 3185 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {} 3186 3187 void do_oop(oop* p) { do_oop_work(p); } 3188 void do_oop(narrowOop* p) { do_oop_work(p); } 3189 3190 void set_nmethod(nmethod* nm) { _nm = nm; } 3191 bool failures() { return _failures; } 3192 }; 3193 3194 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure { 3195 G1VerifyCodeRootOopClosure* _oop_cl; 3196 3197 public: 3198 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl): 3199 _oop_cl(oop_cl) {} 3200 3201 void do_code_blob(CodeBlob* cb) { 3202 nmethod* nm = cb->as_nmethod_or_null(); 3203 if (nm != NULL) { 3204 _oop_cl->set_nmethod(nm); 3205 nm->oops_do(_oop_cl); 3206 } 3207 } 3208 }; 3209 3210 class YoungRefCounterClosure : public OopClosure { 3211 G1CollectedHeap* _g1h; 3212 int _count; 3213 public: 3214 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3215 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3216 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3217 3218 int count() { return _count; } 3219 void reset_count() { _count = 0; }; 3220 }; 3221 3222 class VerifyKlassClosure: public KlassClosure { 3223 YoungRefCounterClosure _young_ref_counter_closure; 3224 OopClosure *_oop_closure; 3225 public: 3226 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3227 void do_klass(Klass* k) { 3228 k->oops_do(_oop_closure); 3229 3230 _young_ref_counter_closure.reset_count(); 3231 k->oops_do(&_young_ref_counter_closure); 3232 if (_young_ref_counter_closure.count() > 0) { 3233 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3234 } 3235 } 3236 }; 3237 3238 class VerifyLivenessOopClosure: public OopClosure { 3239 G1CollectedHeap* _g1h; 3240 VerifyOption _vo; 3241 public: 3242 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3243 _g1h(g1h), _vo(vo) 3244 { } 3245 void do_oop(narrowOop *p) { do_oop_work(p); } 3246 void do_oop( oop *p) { do_oop_work(p); } 3247 3248 template <class T> void do_oop_work(T *p) { 3249 oop obj = oopDesc::load_decode_heap_oop(p); 3250 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3251 "Dead object referenced by a not dead object"); 3252 } 3253 }; 3254 3255 class VerifyObjsInRegionClosure: public ObjectClosure { 3256 private: 3257 G1CollectedHeap* _g1h; 3258 size_t _live_bytes; 3259 HeapRegion *_hr; 3260 VerifyOption _vo; 3261 public: 3262 // _vo == UsePrevMarking -> use "prev" marking information, 3263 // _vo == UseNextMarking -> use "next" marking information, 3264 // _vo == UseMarkWord -> use mark word from object header. 3265 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3266 : _live_bytes(0), _hr(hr), _vo(vo) { 3267 _g1h = G1CollectedHeap::heap(); 3268 } 3269 void do_object(oop o) { 3270 VerifyLivenessOopClosure isLive(_g1h, _vo); 3271 assert(o != NULL, "Huh?"); 3272 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3273 // If the object is alive according to the mark word, 3274 // then verify that the marking information agrees. 3275 // Note we can't verify the contra-positive of the 3276 // above: if the object is dead (according to the mark 3277 // word), it may not be marked, or may have been marked 3278 // but has since became dead, or may have been allocated 3279 // since the last marking. 3280 if (_vo == VerifyOption_G1UseMarkWord) { 3281 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3282 } 3283 3284 o->oop_iterate_no_header(&isLive); 3285 if (!_hr->obj_allocated_since_prev_marking(o)) { 3286 size_t obj_size = o->size(); // Make sure we don't overflow 3287 _live_bytes += (obj_size * HeapWordSize); 3288 } 3289 } 3290 } 3291 size_t live_bytes() { return _live_bytes; } 3292 }; 3293 3294 class PrintObjsInRegionClosure : public ObjectClosure { 3295 HeapRegion *_hr; 3296 G1CollectedHeap *_g1; 3297 public: 3298 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3299 _g1 = G1CollectedHeap::heap(); 3300 }; 3301 3302 void do_object(oop o) { 3303 if (o != NULL) { 3304 HeapWord *start = (HeapWord *) o; 3305 size_t word_sz = o->size(); 3306 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3307 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3308 (void*) o, word_sz, 3309 _g1->isMarkedPrev(o), 3310 _g1->isMarkedNext(o), 3311 _hr->obj_allocated_since_prev_marking(o)); 3312 HeapWord *end = start + word_sz; 3313 HeapWord *cur; 3314 int *val; 3315 for (cur = start; cur < end; cur++) { 3316 val = (int *) cur; 3317 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3318 } 3319 } 3320 } 3321 }; 3322 3323 class VerifyRegionClosure: public HeapRegionClosure { 3324 private: 3325 bool _par; 3326 VerifyOption _vo; 3327 bool _failures; 3328 public: 3329 // _vo == UsePrevMarking -> use "prev" marking information, 3330 // _vo == UseNextMarking -> use "next" marking information, 3331 // _vo == UseMarkWord -> use mark word from object header. 3332 VerifyRegionClosure(bool par, VerifyOption vo) 3333 : _par(par), 3334 _vo(vo), 3335 _failures(false) {} 3336 3337 bool failures() { 3338 return _failures; 3339 } 3340 3341 bool doHeapRegion(HeapRegion* r) { 3342 if (!r->continuesHumongous()) { 3343 bool failures = false; 3344 r->verify(_vo, &failures); 3345 if (failures) { 3346 _failures = true; 3347 } else { 3348 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3349 r->object_iterate(¬_dead_yet_cl); 3350 if (_vo != VerifyOption_G1UseNextMarking) { 3351 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3352 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3353 "max_live_bytes "SIZE_FORMAT" " 3354 "< calculated "SIZE_FORMAT, 3355 r->bottom(), r->end(), 3356 r->max_live_bytes(), 3357 not_dead_yet_cl.live_bytes()); 3358 _failures = true; 3359 } 3360 } else { 3361 // When vo == UseNextMarking we cannot currently do a sanity 3362 // check on the live bytes as the calculation has not been 3363 // finalized yet. 3364 } 3365 } 3366 } 3367 return false; // stop the region iteration if we hit a failure 3368 } 3369 }; 3370 3371 // This is the task used for parallel verification of the heap regions 3372 3373 class G1ParVerifyTask: public AbstractGangTask { 3374 private: 3375 G1CollectedHeap* _g1h; 3376 VerifyOption _vo; 3377 bool _failures; 3378 3379 public: 3380 // _vo == UsePrevMarking -> use "prev" marking information, 3381 // _vo == UseNextMarking -> use "next" marking information, 3382 // _vo == UseMarkWord -> use mark word from object header. 3383 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3384 AbstractGangTask("Parallel verify task"), 3385 _g1h(g1h), 3386 _vo(vo), 3387 _failures(false) { } 3388 3389 bool failures() { 3390 return _failures; 3391 } 3392 3393 void work(uint worker_id) { 3394 HandleMark hm; 3395 VerifyRegionClosure blk(true, _vo); 3396 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3397 _g1h->workers()->active_workers(), 3398 HeapRegion::ParVerifyClaimValue); 3399 if (blk.failures()) { 3400 _failures = true; 3401 } 3402 } 3403 }; 3404 3405 void G1CollectedHeap::verify(bool silent, VerifyOption vo) { 3406 if (SafepointSynchronize::is_at_safepoint()) { 3407 assert(Thread::current()->is_VM_thread(), 3408 "Expected to be executed serially by the VM thread at this point"); 3409 3410 if (!silent) { gclog_or_tty->print("Roots "); } 3411 VerifyRootsClosure rootsCl(vo); 3412 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo); 3413 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl); 3414 VerifyKlassClosure klassCl(this, &rootsCl); 3415 3416 // We apply the relevant closures to all the oops in the 3417 // system dictionary, the string table and the code cache. 3418 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3419 3420 // Need cleared claim bits for the strong roots processing 3421 ClassLoaderDataGraph::clear_claimed_marks(); 3422 3423 process_strong_roots(true, // activate StrongRootsScope 3424 false, // we set "is scavenging" to false, 3425 // so we don't reset the dirty cards. 3426 ScanningOption(so), // roots scanning options 3427 &rootsCl, 3428 &blobsCl, 3429 &klassCl 3430 ); 3431 3432 bool failures = rootsCl.failures() || codeRootsCl.failures(); 3433 3434 if (vo != VerifyOption_G1UseMarkWord) { 3435 // If we're verifying during a full GC then the region sets 3436 // will have been torn down at the start of the GC. Therefore 3437 // verifying the region sets will fail. So we only verify 3438 // the region sets when not in a full GC. 3439 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3440 verify_region_sets(); 3441 } 3442 3443 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3444 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3445 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3446 "sanity check"); 3447 3448 G1ParVerifyTask task(this, vo); 3449 assert(UseDynamicNumberOfGCThreads || 3450 workers()->active_workers() == workers()->total_workers(), 3451 "If not dynamic should be using all the workers"); 3452 int n_workers = workers()->active_workers(); 3453 set_par_threads(n_workers); 3454 workers()->run_task(&task); 3455 set_par_threads(0); 3456 if (task.failures()) { 3457 failures = true; 3458 } 3459 3460 // Checks that the expected amount of parallel work was done. 3461 // The implication is that n_workers is > 0. 3462 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3463 "sanity check"); 3464 3465 reset_heap_region_claim_values(); 3466 3467 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3468 "sanity check"); 3469 } else { 3470 VerifyRegionClosure blk(false, vo); 3471 heap_region_iterate(&blk); 3472 if (blk.failures()) { 3473 failures = true; 3474 } 3475 } 3476 if (!silent) gclog_or_tty->print("RemSet "); 3477 rem_set()->verify(); 3478 3479 if (failures) { 3480 gclog_or_tty->print_cr("Heap:"); 3481 // It helps to have the per-region information in the output to 3482 // help us track down what went wrong. This is why we call 3483 // print_extended_on() instead of print_on(). 3484 print_extended_on(gclog_or_tty); 3485 gclog_or_tty->print_cr(""); 3486 #ifndef PRODUCT 3487 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3488 concurrent_mark()->print_reachable("at-verification-failure", 3489 vo, false /* all */); 3490 } 3491 #endif 3492 gclog_or_tty->flush(); 3493 } 3494 guarantee(!failures, "there should not have been any failures"); 3495 } else { 3496 if (!silent) 3497 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) "); 3498 } 3499 } 3500 3501 void G1CollectedHeap::verify(bool silent) { 3502 verify(silent, VerifyOption_G1UsePrevMarking); 3503 } 3504 3505 double G1CollectedHeap::verify(bool guard, const char* msg) { 3506 double verify_time_ms = 0.0; 3507 3508 if (guard && total_collections() >= VerifyGCStartAt) { 3509 double verify_start = os::elapsedTime(); 3510 HandleMark hm; // Discard invalid handles created during verification 3511 prepare_for_verify(); 3512 Universe::verify(VerifyOption_G1UsePrevMarking, msg); 3513 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 3514 } 3515 3516 return verify_time_ms; 3517 } 3518 3519 void G1CollectedHeap::verify_before_gc() { 3520 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 3521 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 3522 } 3523 3524 void G1CollectedHeap::verify_after_gc() { 3525 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 3526 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 3527 } 3528 3529 class PrintRegionClosure: public HeapRegionClosure { 3530 outputStream* _st; 3531 public: 3532 PrintRegionClosure(outputStream* st) : _st(st) {} 3533 bool doHeapRegion(HeapRegion* r) { 3534 r->print_on(_st); 3535 return false; 3536 } 3537 }; 3538 3539 void G1CollectedHeap::print_on(outputStream* st) const { 3540 st->print(" %-20s", "garbage-first heap"); 3541 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3542 capacity()/K, used_unlocked()/K); 3543 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3544 _g1_storage.low_boundary(), 3545 _g1_storage.high(), 3546 _g1_storage.high_boundary()); 3547 st->cr(); 3548 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3549 uint young_regions = _young_list->length(); 3550 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3551 (size_t) young_regions * HeapRegion::GrainBytes / K); 3552 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3553 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3554 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3555 st->cr(); 3556 MetaspaceAux::print_on(st); 3557 } 3558 3559 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3560 print_on(st); 3561 3562 // Print the per-region information. 3563 st->cr(); 3564 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3565 "HS=humongous(starts), HC=humongous(continues), " 3566 "CS=collection set, F=free, TS=gc time stamp, " 3567 "PTAMS=previous top-at-mark-start, " 3568 "NTAMS=next top-at-mark-start)"); 3569 PrintRegionClosure blk(st); 3570 heap_region_iterate(&blk); 3571 } 3572 3573 void G1CollectedHeap::print_on_error(outputStream* st) const { 3574 this->CollectedHeap::print_on_error(st); 3575 3576 if (_cm != NULL) { 3577 st->cr(); 3578 _cm->print_on_error(st); 3579 } 3580 } 3581 3582 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3583 if (G1CollectedHeap::use_parallel_gc_threads()) { 3584 workers()->print_worker_threads_on(st); 3585 } 3586 _cmThread->print_on(st); 3587 st->cr(); 3588 _cm->print_worker_threads_on(st); 3589 _cg1r->print_worker_threads_on(st); 3590 } 3591 3592 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3593 if (G1CollectedHeap::use_parallel_gc_threads()) { 3594 workers()->threads_do(tc); 3595 } 3596 tc->do_thread(_cmThread); 3597 _cg1r->threads_do(tc); 3598 } 3599 3600 void G1CollectedHeap::print_tracing_info() const { 3601 // We'll overload this to mean "trace GC pause statistics." 3602 if (TraceGen0Time || TraceGen1Time) { 3603 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3604 // to that. 3605 g1_policy()->print_tracing_info(); 3606 } 3607 if (G1SummarizeRSetStats) { 3608 g1_rem_set()->print_summary_info(); 3609 } 3610 if (G1SummarizeConcMark) { 3611 concurrent_mark()->print_summary_info(); 3612 } 3613 g1_policy()->print_yg_surv_rate_info(); 3614 SpecializationStats::print(); 3615 } 3616 3617 #ifndef PRODUCT 3618 // Helpful for debugging RSet issues. 3619 3620 class PrintRSetsClosure : public HeapRegionClosure { 3621 private: 3622 const char* _msg; 3623 size_t _occupied_sum; 3624 3625 public: 3626 bool doHeapRegion(HeapRegion* r) { 3627 HeapRegionRemSet* hrrs = r->rem_set(); 3628 size_t occupied = hrrs->occupied(); 3629 _occupied_sum += occupied; 3630 3631 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3632 HR_FORMAT_PARAMS(r)); 3633 if (occupied == 0) { 3634 gclog_or_tty->print_cr(" RSet is empty"); 3635 } else { 3636 hrrs->print(); 3637 } 3638 gclog_or_tty->print_cr("----------"); 3639 return false; 3640 } 3641 3642 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3643 gclog_or_tty->cr(); 3644 gclog_or_tty->print_cr("========================================"); 3645 gclog_or_tty->print_cr(msg); 3646 gclog_or_tty->cr(); 3647 } 3648 3649 ~PrintRSetsClosure() { 3650 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3651 gclog_or_tty->print_cr("========================================"); 3652 gclog_or_tty->cr(); 3653 } 3654 }; 3655 3656 void G1CollectedHeap::print_cset_rsets() { 3657 PrintRSetsClosure cl("Printing CSet RSets"); 3658 collection_set_iterate(&cl); 3659 } 3660 3661 void G1CollectedHeap::print_all_rsets() { 3662 PrintRSetsClosure cl("Printing All RSets");; 3663 heap_region_iterate(&cl); 3664 } 3665 #endif // PRODUCT 3666 3667 G1CollectedHeap* G1CollectedHeap::heap() { 3668 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3669 "not a garbage-first heap"); 3670 return _g1h; 3671 } 3672 3673 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3674 // always_do_update_barrier = false; 3675 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3676 // Fill TLAB's and such 3677 ensure_parsability(true); 3678 } 3679 3680 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3681 3682 if (G1SummarizeRSetStats && 3683 (G1SummarizeRSetStatsPeriod > 0) && 3684 // we are at the end of the GC. Total collections has already been increased. 3685 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) { 3686 g1_rem_set()->print_periodic_summary_info(); 3687 } 3688 3689 // FIXME: what is this about? 3690 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3691 // is set. 3692 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3693 "derived pointer present")); 3694 // always_do_update_barrier = true; 3695 3696 // We have just completed a GC. Update the soft reference 3697 // policy with the new heap occupancy 3698 Universe::update_heap_info_at_gc(); 3699 } 3700 3701 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3702 unsigned int gc_count_before, 3703 bool* succeeded, 3704 GCCause::Cause gc_cause) { 3705 assert_heap_not_locked_and_not_at_safepoint(); 3706 g1_policy()->record_stop_world_start(); 3707 VM_G1IncCollectionPause op(gc_count_before, 3708 word_size, 3709 false, /* should_initiate_conc_mark */ 3710 g1_policy()->max_pause_time_ms(), 3711 gc_cause); 3712 VMThread::execute(&op); 3713 3714 HeapWord* result = op.result(); 3715 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3716 assert(result == NULL || ret_succeeded, 3717 "the result should be NULL if the VM did not succeed"); 3718 *succeeded = ret_succeeded; 3719 3720 assert_heap_not_locked(); 3721 return result; 3722 } 3723 3724 void 3725 G1CollectedHeap::doConcurrentMark() { 3726 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3727 if (!_cmThread->in_progress()) { 3728 _cmThread->set_started(); 3729 CGC_lock->notify(); 3730 } 3731 } 3732 3733 size_t G1CollectedHeap::pending_card_num() { 3734 size_t extra_cards = 0; 3735 JavaThread *curr = Threads::first(); 3736 while (curr != NULL) { 3737 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3738 extra_cards += dcq.size(); 3739 curr = curr->next(); 3740 } 3741 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3742 size_t buffer_size = dcqs.buffer_size(); 3743 size_t buffer_num = dcqs.completed_buffers_num(); 3744 3745 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3746 // in bytes - not the number of 'entries'. We need to convert 3747 // into a number of cards. 3748 return (buffer_size * buffer_num + extra_cards) / oopSize; 3749 } 3750 3751 size_t G1CollectedHeap::cards_scanned() { 3752 return g1_rem_set()->cardsScanned(); 3753 } 3754 3755 void 3756 G1CollectedHeap::setup_surviving_young_words() { 3757 assert(_surviving_young_words == NULL, "pre-condition"); 3758 uint array_length = g1_policy()->young_cset_region_length(); 3759 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3760 if (_surviving_young_words == NULL) { 3761 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR, 3762 "Not enough space for young surv words summary."); 3763 } 3764 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3765 #ifdef ASSERT 3766 for (uint i = 0; i < array_length; ++i) { 3767 assert( _surviving_young_words[i] == 0, "memset above" ); 3768 } 3769 #endif // !ASSERT 3770 } 3771 3772 void 3773 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3774 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3775 uint array_length = g1_policy()->young_cset_region_length(); 3776 for (uint i = 0; i < array_length; ++i) { 3777 _surviving_young_words[i] += surv_young_words[i]; 3778 } 3779 } 3780 3781 void 3782 G1CollectedHeap::cleanup_surviving_young_words() { 3783 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3784 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3785 _surviving_young_words = NULL; 3786 } 3787 3788 #ifdef ASSERT 3789 class VerifyCSetClosure: public HeapRegionClosure { 3790 public: 3791 bool doHeapRegion(HeapRegion* hr) { 3792 // Here we check that the CSet region's RSet is ready for parallel 3793 // iteration. The fields that we'll verify are only manipulated 3794 // when the region is part of a CSet and is collected. Afterwards, 3795 // we reset these fields when we clear the region's RSet (when the 3796 // region is freed) so they are ready when the region is 3797 // re-allocated. The only exception to this is if there's an 3798 // evacuation failure and instead of freeing the region we leave 3799 // it in the heap. In that case, we reset these fields during 3800 // evacuation failure handling. 3801 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3802 3803 // Here's a good place to add any other checks we'd like to 3804 // perform on CSet regions. 3805 return false; 3806 } 3807 }; 3808 #endif // ASSERT 3809 3810 #if TASKQUEUE_STATS 3811 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3812 st->print_raw_cr("GC Task Stats"); 3813 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3814 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3815 } 3816 3817 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3818 print_taskqueue_stats_hdr(st); 3819 3820 TaskQueueStats totals; 3821 const int n = workers() != NULL ? workers()->total_workers() : 1; 3822 for (int i = 0; i < n; ++i) { 3823 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3824 totals += task_queue(i)->stats; 3825 } 3826 st->print_raw("tot "); totals.print(st); st->cr(); 3827 3828 DEBUG_ONLY(totals.verify()); 3829 } 3830 3831 void G1CollectedHeap::reset_taskqueue_stats() { 3832 const int n = workers() != NULL ? workers()->total_workers() : 1; 3833 for (int i = 0; i < n; ++i) { 3834 task_queue(i)->stats.reset(); 3835 } 3836 } 3837 #endif // TASKQUEUE_STATS 3838 3839 void G1CollectedHeap::log_gc_header() { 3840 if (!G1Log::fine()) { 3841 return; 3842 } 3843 3844 gclog_or_tty->date_stamp(PrintGCDateStamps); 3845 gclog_or_tty->stamp(PrintGCTimeStamps); 3846 3847 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3848 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3849 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3850 3851 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3852 } 3853 3854 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3855 if (!G1Log::fine()) { 3856 return; 3857 } 3858 3859 if (G1Log::finer()) { 3860 if (evacuation_failed()) { 3861 gclog_or_tty->print(" (to-space exhausted)"); 3862 } 3863 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3864 g1_policy()->phase_times()->note_gc_end(); 3865 g1_policy()->phase_times()->print(pause_time_sec); 3866 g1_policy()->print_detailed_heap_transition(); 3867 } else { 3868 if (evacuation_failed()) { 3869 gclog_or_tty->print("--"); 3870 } 3871 g1_policy()->print_heap_transition(); 3872 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3873 } 3874 gclog_or_tty->flush(); 3875 } 3876 3877 bool 3878 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3879 assert_at_safepoint(true /* should_be_vm_thread */); 3880 guarantee(!is_gc_active(), "collection is not reentrant"); 3881 3882 if (GC_locker::check_active_before_gc()) { 3883 return false; 3884 } 3885 3886 _gc_timer_stw->register_gc_start(os::elapsed_counter()); 3887 3888 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start()); 3889 3890 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3891 ResourceMark rm; 3892 3893 print_heap_before_gc(); 3894 trace_heap_before_gc(_gc_tracer_stw); 3895 3896 HRSPhaseSetter x(HRSPhaseEvacuation); 3897 verify_region_sets_optional(); 3898 verify_dirty_young_regions(); 3899 3900 // This call will decide whether this pause is an initial-mark 3901 // pause. If it is, during_initial_mark_pause() will return true 3902 // for the duration of this pause. 3903 g1_policy()->decide_on_conc_mark_initiation(); 3904 3905 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3906 assert(!g1_policy()->during_initial_mark_pause() || 3907 g1_policy()->gcs_are_young(), "sanity"); 3908 3909 // We also do not allow mixed GCs during marking. 3910 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3911 3912 // Record whether this pause is an initial mark. When the current 3913 // thread has completed its logging output and it's safe to signal 3914 // the CM thread, the flag's value in the policy has been reset. 3915 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3916 3917 // Inner scope for scope based logging, timers, and stats collection 3918 { 3919 EvacuationInfo evacuation_info; 3920 3921 if (g1_policy()->during_initial_mark_pause()) { 3922 // We are about to start a marking cycle, so we increment the 3923 // full collection counter. 3924 increment_old_marking_cycles_started(); 3925 register_concurrent_cycle_start(_gc_timer_stw->gc_start()); 3926 } 3927 3928 _gc_tracer_stw->report_yc_type(yc_type()); 3929 3930 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3931 3932 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3933 workers()->active_workers() : 1); 3934 double pause_start_sec = os::elapsedTime(); 3935 g1_policy()->phase_times()->note_gc_start(active_workers); 3936 log_gc_header(); 3937 3938 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3939 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3940 3941 // If the secondary_free_list is not empty, append it to the 3942 // free_list. No need to wait for the cleanup operation to finish; 3943 // the region allocation code will check the secondary_free_list 3944 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3945 // set, skip this step so that the region allocation code has to 3946 // get entries from the secondary_free_list. 3947 if (!G1StressConcRegionFreeing) { 3948 append_secondary_free_list_if_not_empty_with_lock(); 3949 } 3950 3951 assert(check_young_list_well_formed(), "young list should be well formed"); 3952 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3953 "sanity check"); 3954 3955 // Don't dynamically change the number of GC threads this early. A value of 3956 // 0 is used to indicate serial work. When parallel work is done, 3957 // it will be set. 3958 3959 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3960 IsGCActiveMark x; 3961 3962 gc_prologue(false); 3963 increment_total_collections(false /* full gc */); 3964 increment_gc_time_stamp(); 3965 3966 verify_before_gc(); 3967 3968 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3969 3970 // Please see comment in g1CollectedHeap.hpp and 3971 // G1CollectedHeap::ref_processing_init() to see how 3972 // reference processing currently works in G1. 3973 3974 // Enable discovery in the STW reference processor 3975 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3976 true /*verify_no_refs*/); 3977 3978 { 3979 // We want to temporarily turn off discovery by the 3980 // CM ref processor, if necessary, and turn it back on 3981 // on again later if we do. Using a scoped 3982 // NoRefDiscovery object will do this. 3983 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3984 3985 // Forget the current alloc region (we might even choose it to be part 3986 // of the collection set!). 3987 release_mutator_alloc_region(); 3988 3989 // We should call this after we retire the mutator alloc 3990 // region(s) so that all the ALLOC / RETIRE events are generated 3991 // before the start GC event. 3992 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3993 3994 // This timing is only used by the ergonomics to handle our pause target. 3995 // It is unclear why this should not include the full pause. We will 3996 // investigate this in CR 7178365. 3997 // 3998 // Preserving the old comment here if that helps the investigation: 3999 // 4000 // The elapsed time induced by the start time below deliberately elides 4001 // the possible verification above. 4002 double sample_start_time_sec = os::elapsedTime(); 4003 4004 #if YOUNG_LIST_VERBOSE 4005 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 4006 _young_list->print(); 4007 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4008 #endif // YOUNG_LIST_VERBOSE 4009 4010 g1_policy()->record_collection_pause_start(sample_start_time_sec); 4011 4012 double scan_wait_start = os::elapsedTime(); 4013 // We have to wait until the CM threads finish scanning the 4014 // root regions as it's the only way to ensure that all the 4015 // objects on them have been correctly scanned before we start 4016 // moving them during the GC. 4017 bool waited = _cm->root_regions()->wait_until_scan_finished(); 4018 double wait_time_ms = 0.0; 4019 if (waited) { 4020 double scan_wait_end = os::elapsedTime(); 4021 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 4022 } 4023 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 4024 4025 #if YOUNG_LIST_VERBOSE 4026 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 4027 _young_list->print(); 4028 #endif // YOUNG_LIST_VERBOSE 4029 4030 if (g1_policy()->during_initial_mark_pause()) { 4031 concurrent_mark()->checkpointRootsInitialPre(); 4032 } 4033 4034 #if YOUNG_LIST_VERBOSE 4035 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 4036 _young_list->print(); 4037 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4038 #endif // YOUNG_LIST_VERBOSE 4039 4040 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); 4041 4042 _cm->note_start_of_gc(); 4043 // We should not verify the per-thread SATB buffers given that 4044 // we have not filtered them yet (we'll do so during the 4045 // GC). We also call this after finalize_cset() to 4046 // ensure that the CSet has been finalized. 4047 _cm->verify_no_cset_oops(true /* verify_stacks */, 4048 true /* verify_enqueued_buffers */, 4049 false /* verify_thread_buffers */, 4050 true /* verify_fingers */); 4051 4052 if (_hr_printer.is_active()) { 4053 HeapRegion* hr = g1_policy()->collection_set(); 4054 while (hr != NULL) { 4055 G1HRPrinter::RegionType type; 4056 if (!hr->is_young()) { 4057 type = G1HRPrinter::Old; 4058 } else if (hr->is_survivor()) { 4059 type = G1HRPrinter::Survivor; 4060 } else { 4061 type = G1HRPrinter::Eden; 4062 } 4063 _hr_printer.cset(hr); 4064 hr = hr->next_in_collection_set(); 4065 } 4066 } 4067 4068 #ifdef ASSERT 4069 VerifyCSetClosure cl; 4070 collection_set_iterate(&cl); 4071 #endif // ASSERT 4072 4073 setup_surviving_young_words(); 4074 4075 // Initialize the GC alloc regions. 4076 init_gc_alloc_regions(evacuation_info); 4077 4078 // Actually do the work... 4079 evacuate_collection_set(evacuation_info); 4080 4081 // We do this to mainly verify the per-thread SATB buffers 4082 // (which have been filtered by now) since we didn't verify 4083 // them earlier. No point in re-checking the stacks / enqueued 4084 // buffers given that the CSet has not changed since last time 4085 // we checked. 4086 _cm->verify_no_cset_oops(false /* verify_stacks */, 4087 false /* verify_enqueued_buffers */, 4088 true /* verify_thread_buffers */, 4089 true /* verify_fingers */); 4090 4091 free_collection_set(g1_policy()->collection_set(), evacuation_info); 4092 g1_policy()->clear_collection_set(); 4093 4094 cleanup_surviving_young_words(); 4095 4096 // Start a new incremental collection set for the next pause. 4097 g1_policy()->start_incremental_cset_building(); 4098 4099 // Clear the _cset_fast_test bitmap in anticipation of adding 4100 // regions to the incremental collection set for the next 4101 // evacuation pause. 4102 clear_cset_fast_test(); 4103 4104 _young_list->reset_sampled_info(); 4105 4106 // Don't check the whole heap at this point as the 4107 // GC alloc regions from this pause have been tagged 4108 // as survivors and moved on to the survivor list. 4109 // Survivor regions will fail the !is_young() check. 4110 assert(check_young_list_empty(false /* check_heap */), 4111 "young list should be empty"); 4112 4113 #if YOUNG_LIST_VERBOSE 4114 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 4115 _young_list->print(); 4116 #endif // YOUNG_LIST_VERBOSE 4117 4118 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 4119 _young_list->first_survivor_region(), 4120 _young_list->last_survivor_region()); 4121 4122 _young_list->reset_auxilary_lists(); 4123 4124 if (evacuation_failed()) { 4125 _summary_bytes_used = recalculate_used(); 4126 uint n_queues = MAX2((int)ParallelGCThreads, 1); 4127 for (uint i = 0; i < n_queues; i++) { 4128 if (_evacuation_failed_info_array[i].has_failed()) { 4129 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]); 4130 } 4131 } 4132 } else { 4133 // The "used" of the the collection set have already been subtracted 4134 // when they were freed. Add in the bytes evacuated. 4135 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 4136 } 4137 4138 if (g1_policy()->during_initial_mark_pause()) { 4139 // We have to do this before we notify the CM threads that 4140 // they can start working to make sure that all the 4141 // appropriate initialization is done on the CM object. 4142 concurrent_mark()->checkpointRootsInitialPost(); 4143 set_marking_started(); 4144 // Note that we don't actually trigger the CM thread at 4145 // this point. We do that later when we're sure that 4146 // the current thread has completed its logging output. 4147 } 4148 4149 allocate_dummy_regions(); 4150 4151 #if YOUNG_LIST_VERBOSE 4152 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 4153 _young_list->print(); 4154 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 4155 #endif // YOUNG_LIST_VERBOSE 4156 4157 init_mutator_alloc_region(); 4158 4159 { 4160 size_t expand_bytes = g1_policy()->expansion_amount(); 4161 if (expand_bytes > 0) { 4162 size_t bytes_before = capacity(); 4163 // No need for an ergo verbose message here, 4164 // expansion_amount() does this when it returns a value > 0. 4165 if (!expand(expand_bytes)) { 4166 // We failed to expand the heap so let's verify that 4167 // committed/uncommitted amount match the backing store 4168 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 4169 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 4170 } 4171 } 4172 } 4173 4174 // We redo the verification but now wrt to the new CSet which 4175 // has just got initialized after the previous CSet was freed. 4176 _cm->verify_no_cset_oops(true /* verify_stacks */, 4177 true /* verify_enqueued_buffers */, 4178 true /* verify_thread_buffers */, 4179 true /* verify_fingers */); 4180 _cm->note_end_of_gc(); 4181 4182 // This timing is only used by the ergonomics to handle our pause target. 4183 // It is unclear why this should not include the full pause. We will 4184 // investigate this in CR 7178365. 4185 double sample_end_time_sec = os::elapsedTime(); 4186 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 4187 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info); 4188 4189 MemoryService::track_memory_usage(); 4190 4191 // In prepare_for_verify() below we'll need to scan the deferred 4192 // update buffers to bring the RSets up-to-date if 4193 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4194 // the update buffers we'll probably need to scan cards on the 4195 // regions we just allocated to (i.e., the GC alloc 4196 // regions). However, during the last GC we called 4197 // set_saved_mark() on all the GC alloc regions, so card 4198 // scanning might skip the [saved_mark_word()...top()] area of 4199 // those regions (i.e., the area we allocated objects into 4200 // during the last GC). But it shouldn't. Given that 4201 // saved_mark_word() is conditional on whether the GC time stamp 4202 // on the region is current or not, by incrementing the GC time 4203 // stamp here we invalidate all the GC time stamps on all the 4204 // regions and saved_mark_word() will simply return top() for 4205 // all the regions. This is a nicer way of ensuring this rather 4206 // than iterating over the regions and fixing them. In fact, the 4207 // GC time stamp increment here also ensures that 4208 // saved_mark_word() will return top() between pauses, i.e., 4209 // during concurrent refinement. So we don't need the 4210 // is_gc_active() check to decided which top to use when 4211 // scanning cards (see CR 7039627). 4212 increment_gc_time_stamp(); 4213 4214 verify_after_gc(); 4215 4216 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4217 ref_processor_stw()->verify_no_references_recorded(); 4218 4219 // CM reference discovery will be re-enabled if necessary. 4220 } 4221 4222 // We should do this after we potentially expand the heap so 4223 // that all the COMMIT events are generated before the end GC 4224 // event, and after we retire the GC alloc regions so that all 4225 // RETIRE events are generated before the end GC event. 4226 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4227 4228 if (mark_in_progress()) { 4229 concurrent_mark()->update_g1_committed(); 4230 } 4231 4232 #ifdef TRACESPINNING 4233 ParallelTaskTerminator::print_termination_counts(); 4234 #endif 4235 4236 gc_epilogue(false); 4237 } 4238 4239 // Print the remainder of the GC log output. 4240 log_gc_footer(os::elapsedTime() - pause_start_sec); 4241 4242 // It is not yet to safe to tell the concurrent mark to 4243 // start as we have some optional output below. We don't want the 4244 // output from the concurrent mark thread interfering with this 4245 // logging output either. 4246 4247 _hrs.verify_optional(); 4248 verify_region_sets_optional(); 4249 4250 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4251 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4252 4253 print_heap_after_gc(); 4254 trace_heap_after_gc(_gc_tracer_stw); 4255 4256 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4257 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4258 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4259 // before any GC notifications are raised. 4260 g1mm()->update_sizes(); 4261 4262 _gc_tracer_stw->report_evacuation_info(&evacuation_info); 4263 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold()); 4264 _gc_timer_stw->register_gc_end(os::elapsed_counter()); 4265 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions()); 4266 } 4267 // It should now be safe to tell the concurrent mark thread to start 4268 // without its logging output interfering with the logging output 4269 // that came from the pause. 4270 4271 if (should_start_conc_mark) { 4272 // CAUTION: after the doConcurrentMark() call below, 4273 // the concurrent marking thread(s) could be running 4274 // concurrently with us. Make sure that anything after 4275 // this point does not assume that we are the only GC thread 4276 // running. Note: of course, the actual marking work will 4277 // not start until the safepoint itself is released in 4278 // ConcurrentGCThread::safepoint_desynchronize(). 4279 doConcurrentMark(); 4280 } 4281 4282 return true; 4283 } 4284 4285 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4286 { 4287 size_t gclab_word_size; 4288 switch (purpose) { 4289 case GCAllocForSurvived: 4290 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4291 break; 4292 case GCAllocForTenured: 4293 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4294 break; 4295 default: 4296 assert(false, "unknown GCAllocPurpose"); 4297 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4298 break; 4299 } 4300 4301 // Prevent humongous PLAB sizes for two reasons: 4302 // * PLABs are allocated using a similar paths as oops, but should 4303 // never be in a humongous region 4304 // * Allowing humongous PLABs needlessly churns the region free lists 4305 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4306 } 4307 4308 void G1CollectedHeap::init_mutator_alloc_region() { 4309 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4310 _mutator_alloc_region.init(); 4311 } 4312 4313 void G1CollectedHeap::release_mutator_alloc_region() { 4314 _mutator_alloc_region.release(); 4315 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4316 } 4317 4318 void G1CollectedHeap::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { 4319 assert_at_safepoint(true /* should_be_vm_thread */); 4320 4321 _survivor_gc_alloc_region.init(); 4322 _old_gc_alloc_region.init(); 4323 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4324 _retained_old_gc_alloc_region = NULL; 4325 4326 // We will discard the current GC alloc region if: 4327 // a) it's in the collection set (it can happen!), 4328 // b) it's already full (no point in using it), 4329 // c) it's empty (this means that it was emptied during 4330 // a cleanup and it should be on the free list now), or 4331 // d) it's humongous (this means that it was emptied 4332 // during a cleanup and was added to the free list, but 4333 // has been subsequently used to allocate a humongous 4334 // object that may be less than the region size). 4335 if (retained_region != NULL && 4336 !retained_region->in_collection_set() && 4337 !(retained_region->top() == retained_region->end()) && 4338 !retained_region->is_empty() && 4339 !retained_region->isHumongous()) { 4340 retained_region->set_saved_mark(); 4341 // The retained region was added to the old region set when it was 4342 // retired. We have to remove it now, since we don't allow regions 4343 // we allocate to in the region sets. We'll re-add it later, when 4344 // it's retired again. 4345 _old_set.remove(retained_region); 4346 bool during_im = g1_policy()->during_initial_mark_pause(); 4347 retained_region->note_start_of_copying(during_im); 4348 _old_gc_alloc_region.set(retained_region); 4349 _hr_printer.reuse(retained_region); 4350 evacuation_info.set_alloc_regions_used_before(retained_region->used()); 4351 } 4352 } 4353 4354 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { 4355 evacuation_info.set_allocation_regions(_survivor_gc_alloc_region.count() + 4356 _old_gc_alloc_region.count()); 4357 _survivor_gc_alloc_region.release(); 4358 // If we have an old GC alloc region to release, we'll save it in 4359 // _retained_old_gc_alloc_region. If we don't 4360 // _retained_old_gc_alloc_region will become NULL. This is what we 4361 // want either way so no reason to check explicitly for either 4362 // condition. 4363 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4364 4365 if (ResizePLAB) { 4366 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4367 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4368 } 4369 } 4370 4371 void G1CollectedHeap::abandon_gc_alloc_regions() { 4372 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4373 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4374 _retained_old_gc_alloc_region = NULL; 4375 } 4376 4377 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4378 _drain_in_progress = false; 4379 set_evac_failure_closure(cl); 4380 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4381 } 4382 4383 void G1CollectedHeap::finalize_for_evac_failure() { 4384 assert(_evac_failure_scan_stack != NULL && 4385 _evac_failure_scan_stack->length() == 0, 4386 "Postcondition"); 4387 assert(!_drain_in_progress, "Postcondition"); 4388 delete _evac_failure_scan_stack; 4389 _evac_failure_scan_stack = NULL; 4390 } 4391 4392 void G1CollectedHeap::remove_self_forwarding_pointers() { 4393 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4394 4395 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4396 4397 if (G1CollectedHeap::use_parallel_gc_threads()) { 4398 set_par_threads(); 4399 workers()->run_task(&rsfp_task); 4400 set_par_threads(0); 4401 } else { 4402 rsfp_task.work(0); 4403 } 4404 4405 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4406 4407 // Reset the claim values in the regions in the collection set. 4408 reset_cset_heap_region_claim_values(); 4409 4410 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4411 4412 // Now restore saved marks, if any. 4413 assert(_objs_with_preserved_marks.size() == 4414 _preserved_marks_of_objs.size(), "Both or none."); 4415 while (!_objs_with_preserved_marks.is_empty()) { 4416 oop obj = _objs_with_preserved_marks.pop(); 4417 markOop m = _preserved_marks_of_objs.pop(); 4418 obj->set_mark(m); 4419 } 4420 _objs_with_preserved_marks.clear(true); 4421 _preserved_marks_of_objs.clear(true); 4422 } 4423 4424 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4425 _evac_failure_scan_stack->push(obj); 4426 } 4427 4428 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4429 assert(_evac_failure_scan_stack != NULL, "precondition"); 4430 4431 while (_evac_failure_scan_stack->length() > 0) { 4432 oop obj = _evac_failure_scan_stack->pop(); 4433 _evac_failure_closure->set_region(heap_region_containing(obj)); 4434 obj->oop_iterate_backwards(_evac_failure_closure); 4435 } 4436 } 4437 4438 oop 4439 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state, 4440 oop old) { 4441 assert(obj_in_cs(old), 4442 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4443 (HeapWord*) old)); 4444 markOop m = old->mark(); 4445 oop forward_ptr = old->forward_to_atomic(old); 4446 if (forward_ptr == NULL) { 4447 // Forward-to-self succeeded. 4448 assert(_par_scan_state != NULL, "par scan state"); 4449 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4450 uint queue_num = _par_scan_state->queue_num(); 4451 4452 _evacuation_failed = true; 4453 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size()); 4454 if (_evac_failure_closure != cl) { 4455 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4456 assert(!_drain_in_progress, 4457 "Should only be true while someone holds the lock."); 4458 // Set the global evac-failure closure to the current thread's. 4459 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4460 set_evac_failure_closure(cl); 4461 // Now do the common part. 4462 handle_evacuation_failure_common(old, m); 4463 // Reset to NULL. 4464 set_evac_failure_closure(NULL); 4465 } else { 4466 // The lock is already held, and this is recursive. 4467 assert(_drain_in_progress, "This should only be the recursive case."); 4468 handle_evacuation_failure_common(old, m); 4469 } 4470 return old; 4471 } else { 4472 // Forward-to-self failed. Either someone else managed to allocate 4473 // space for this object (old != forward_ptr) or they beat us in 4474 // self-forwarding it (old == forward_ptr). 4475 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4476 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4477 "should not be in the CSet", 4478 (HeapWord*) old, (HeapWord*) forward_ptr)); 4479 return forward_ptr; 4480 } 4481 } 4482 4483 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4484 preserve_mark_if_necessary(old, m); 4485 4486 HeapRegion* r = heap_region_containing(old); 4487 if (!r->evacuation_failed()) { 4488 r->set_evacuation_failed(true); 4489 _hr_printer.evac_failure(r); 4490 } 4491 4492 push_on_evac_failure_scan_stack(old); 4493 4494 if (!_drain_in_progress) { 4495 // prevent recursion in copy_to_survivor_space() 4496 _drain_in_progress = true; 4497 drain_evac_failure_scan_stack(); 4498 _drain_in_progress = false; 4499 } 4500 } 4501 4502 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4503 assert(evacuation_failed(), "Oversaving!"); 4504 // We want to call the "for_promotion_failure" version only in the 4505 // case of a promotion failure. 4506 if (m->must_be_preserved_for_promotion_failure(obj)) { 4507 _objs_with_preserved_marks.push(obj); 4508 _preserved_marks_of_objs.push(m); 4509 } 4510 } 4511 4512 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4513 size_t word_size) { 4514 if (purpose == GCAllocForSurvived) { 4515 HeapWord* result = survivor_attempt_allocation(word_size); 4516 if (result != NULL) { 4517 return result; 4518 } else { 4519 // Let's try to allocate in the old gen in case we can fit the 4520 // object there. 4521 return old_attempt_allocation(word_size); 4522 } 4523 } else { 4524 assert(purpose == GCAllocForTenured, "sanity"); 4525 HeapWord* result = old_attempt_allocation(word_size); 4526 if (result != NULL) { 4527 return result; 4528 } else { 4529 // Let's try to allocate in the survivors in case we can fit the 4530 // object there. 4531 return survivor_attempt_allocation(word_size); 4532 } 4533 } 4534 4535 ShouldNotReachHere(); 4536 // Trying to keep some compilers happy. 4537 return NULL; 4538 } 4539 4540 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4541 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4542 4543 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4544 : _g1h(g1h), 4545 _refs(g1h->task_queue(queue_num)), 4546 _dcq(&g1h->dirty_card_queue_set()), 4547 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4548 _g1_rem(g1h->g1_rem_set()), 4549 _hash_seed(17), _queue_num(queue_num), 4550 _term_attempts(0), 4551 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4552 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4553 _age_table(false), 4554 _strong_roots_time(0), _term_time(0), 4555 _alloc_buffer_waste(0), _undo_waste(0) { 4556 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4557 // we "sacrifice" entry 0 to keep track of surviving bytes for 4558 // non-young regions (where the age is -1) 4559 // We also add a few elements at the beginning and at the end in 4560 // an attempt to eliminate cache contention 4561 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4562 uint array_length = PADDING_ELEM_NUM + 4563 real_length + 4564 PADDING_ELEM_NUM; 4565 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4566 if (_surviving_young_words_base == NULL) 4567 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR, 4568 "Not enough space for young surv histo."); 4569 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4570 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4571 4572 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4573 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4574 4575 _start = os::elapsedTime(); 4576 } 4577 4578 void 4579 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4580 { 4581 st->print_raw_cr("GC Termination Stats"); 4582 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4583 " ------waste (KiB)------"); 4584 st->print_raw_cr("thr ms ms % ms % attempts" 4585 " total alloc undo"); 4586 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4587 " ------- ------- -------"); 4588 } 4589 4590 void 4591 G1ParScanThreadState::print_termination_stats(int i, 4592 outputStream* const st) const 4593 { 4594 const double elapsed_ms = elapsed_time() * 1000.0; 4595 const double s_roots_ms = strong_roots_time() * 1000.0; 4596 const double term_ms = term_time() * 1000.0; 4597 st->print_cr("%3d %9.2f %9.2f %6.2f " 4598 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4599 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4600 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4601 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4602 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4603 alloc_buffer_waste() * HeapWordSize / K, 4604 undo_waste() * HeapWordSize / K); 4605 } 4606 4607 #ifdef ASSERT 4608 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4609 assert(ref != NULL, "invariant"); 4610 assert(UseCompressedOops, "sanity"); 4611 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4612 oop p = oopDesc::load_decode_heap_oop(ref); 4613 assert(_g1h->is_in_g1_reserved(p), 4614 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4615 return true; 4616 } 4617 4618 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4619 assert(ref != NULL, "invariant"); 4620 if (has_partial_array_mask(ref)) { 4621 // Must be in the collection set--it's already been copied. 4622 oop p = clear_partial_array_mask(ref); 4623 assert(_g1h->obj_in_cs(p), 4624 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4625 } else { 4626 oop p = oopDesc::load_decode_heap_oop(ref); 4627 assert(_g1h->is_in_g1_reserved(p), 4628 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4629 } 4630 return true; 4631 } 4632 4633 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4634 if (ref.is_narrow()) { 4635 return verify_ref((narrowOop*) ref); 4636 } else { 4637 return verify_ref((oop*) ref); 4638 } 4639 } 4640 #endif // ASSERT 4641 4642 void G1ParScanThreadState::trim_queue() { 4643 assert(_evac_cl != NULL, "not set"); 4644 assert(_evac_failure_cl != NULL, "not set"); 4645 assert(_partial_scan_cl != NULL, "not set"); 4646 4647 StarTask ref; 4648 do { 4649 // Drain the overflow stack first, so other threads can steal. 4650 while (refs()->pop_overflow(ref)) { 4651 deal_with_reference(ref); 4652 } 4653 4654 while (refs()->pop_local(ref)) { 4655 deal_with_reference(ref); 4656 } 4657 } while (!refs()->is_empty()); 4658 } 4659 4660 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4661 G1ParScanThreadState* par_scan_state) : 4662 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4663 _par_scan_state(par_scan_state), 4664 _worker_id(par_scan_state->queue_num()), 4665 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4666 _mark_in_progress(_g1->mark_in_progress()) { } 4667 4668 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4669 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4670 #ifdef ASSERT 4671 HeapRegion* hr = _g1->heap_region_containing(obj); 4672 assert(hr != NULL, "sanity"); 4673 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4674 #endif // ASSERT 4675 4676 // We know that the object is not moving so it's safe to read its size. 4677 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4678 } 4679 4680 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4681 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4682 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4683 #ifdef ASSERT 4684 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4685 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4686 assert(from_obj != to_obj, "should not be self-forwarded"); 4687 4688 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4689 assert(from_hr != NULL, "sanity"); 4690 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4691 4692 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4693 assert(to_hr != NULL, "sanity"); 4694 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4695 #endif // ASSERT 4696 4697 // The object might be in the process of being copied by another 4698 // worker so we cannot trust that its to-space image is 4699 // well-formed. So we have to read its size from its from-space 4700 // image which we know should not be changing. 4701 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4702 } 4703 4704 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4705 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4706 ::copy_to_survivor_space(oop old) { 4707 size_t word_sz = old->size(); 4708 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4709 // +1 to make the -1 indexes valid... 4710 int young_index = from_region->young_index_in_cset()+1; 4711 assert( (from_region->is_young() && young_index > 0) || 4712 (!from_region->is_young() && young_index == 0), "invariant" ); 4713 G1CollectorPolicy* g1p = _g1->g1_policy(); 4714 markOop m = old->mark(); 4715 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4716 : m->age(); 4717 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4718 word_sz); 4719 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4720 #ifndef PRODUCT 4721 // Should this evacuation fail? 4722 if (_g1->evacuation_should_fail()) { 4723 if (obj_ptr != NULL) { 4724 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4725 obj_ptr = NULL; 4726 } 4727 } 4728 #endif // !PRODUCT 4729 4730 if (obj_ptr == NULL) { 4731 // This will either forward-to-self, or detect that someone else has 4732 // installed a forwarding pointer. 4733 return _g1->handle_evacuation_failure_par(_par_scan_state, old); 4734 } 4735 4736 oop obj = oop(obj_ptr); 4737 4738 // We're going to allocate linearly, so might as well prefetch ahead. 4739 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4740 4741 oop forward_ptr = old->forward_to_atomic(obj); 4742 if (forward_ptr == NULL) { 4743 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4744 if (g1p->track_object_age(alloc_purpose)) { 4745 // We could simply do obj->incr_age(). However, this causes a 4746 // performance issue. obj->incr_age() will first check whether 4747 // the object has a displaced mark by checking its mark word; 4748 // getting the mark word from the new location of the object 4749 // stalls. So, given that we already have the mark word and we 4750 // are about to install it anyway, it's better to increase the 4751 // age on the mark word, when the object does not have a 4752 // displaced mark word. We're not expecting many objects to have 4753 // a displaced marked word, so that case is not optimized 4754 // further (it could be...) and we simply call obj->incr_age(). 4755 4756 if (m->has_displaced_mark_helper()) { 4757 // in this case, we have to install the mark word first, 4758 // otherwise obj looks to be forwarded (the old mark word, 4759 // which contains the forward pointer, was copied) 4760 obj->set_mark(m); 4761 obj->incr_age(); 4762 } else { 4763 m = m->incr_age(); 4764 obj->set_mark(m); 4765 } 4766 _par_scan_state->age_table()->add(obj, word_sz); 4767 } else { 4768 obj->set_mark(m); 4769 } 4770 4771 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4772 surv_young_words[young_index] += word_sz; 4773 4774 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4775 // We keep track of the next start index in the length field of 4776 // the to-space object. The actual length can be found in the 4777 // length field of the from-space object. 4778 arrayOop(obj)->set_length(0); 4779 oop* old_p = set_partial_array_mask(old); 4780 _par_scan_state->push_on_queue(old_p); 4781 } else { 4782 // No point in using the slower heap_region_containing() method, 4783 // given that we know obj is in the heap. 4784 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4785 obj->oop_iterate_backwards(&_scanner); 4786 } 4787 } else { 4788 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4789 obj = forward_ptr; 4790 } 4791 return obj; 4792 } 4793 4794 template <class T> 4795 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4796 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4797 _scanned_klass->record_modified_oops(); 4798 } 4799 } 4800 4801 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4802 template <class T> 4803 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4804 ::do_oop_work(T* p) { 4805 oop obj = oopDesc::load_decode_heap_oop(p); 4806 assert(barrier != G1BarrierRS || obj != NULL, 4807 "Precondition: G1BarrierRS implies obj is non-NULL"); 4808 4809 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4810 4811 // here the null check is implicit in the cset_fast_test() test 4812 if (_g1->in_cset_fast_test(obj)) { 4813 oop forwardee; 4814 if (obj->is_forwarded()) { 4815 forwardee = obj->forwardee(); 4816 } else { 4817 forwardee = copy_to_survivor_space(obj); 4818 } 4819 assert(forwardee != NULL, "forwardee should not be NULL"); 4820 oopDesc::encode_store_heap_oop(p, forwardee); 4821 if (do_mark_object && forwardee != obj) { 4822 // If the object is self-forwarded we don't need to explicitly 4823 // mark it, the evacuation failure protocol will do so. 4824 mark_forwarded_object(obj, forwardee); 4825 } 4826 4827 // When scanning the RS, we only care about objs in CS. 4828 if (barrier == G1BarrierRS) { 4829 _par_scan_state->update_rs(_from, p, _worker_id); 4830 } else if (barrier == G1BarrierKlass) { 4831 do_klass_barrier(p, forwardee); 4832 } 4833 } else { 4834 // The object is not in collection set. If we're a root scanning 4835 // closure during an initial mark pause (i.e. do_mark_object will 4836 // be true) then attempt to mark the object. 4837 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4838 mark_object(obj); 4839 } 4840 } 4841 4842 if (barrier == G1BarrierEvac && obj != NULL) { 4843 _par_scan_state->update_rs(_from, p, _worker_id); 4844 } 4845 4846 if (do_gen_barrier && obj != NULL) { 4847 par_do_barrier(p); 4848 } 4849 } 4850 4851 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4852 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4853 4854 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4855 assert(has_partial_array_mask(p), "invariant"); 4856 oop from_obj = clear_partial_array_mask(p); 4857 4858 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4859 assert(from_obj->is_objArray(), "must be obj array"); 4860 objArrayOop from_obj_array = objArrayOop(from_obj); 4861 // The from-space object contains the real length. 4862 int length = from_obj_array->length(); 4863 4864 assert(from_obj->is_forwarded(), "must be forwarded"); 4865 oop to_obj = from_obj->forwardee(); 4866 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4867 objArrayOop to_obj_array = objArrayOop(to_obj); 4868 // We keep track of the next start index in the length field of the 4869 // to-space object. 4870 int next_index = to_obj_array->length(); 4871 assert(0 <= next_index && next_index < length, 4872 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4873 4874 int start = next_index; 4875 int end = length; 4876 int remainder = end - start; 4877 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4878 if (remainder > 2 * ParGCArrayScanChunk) { 4879 end = start + ParGCArrayScanChunk; 4880 to_obj_array->set_length(end); 4881 // Push the remainder before we process the range in case another 4882 // worker has run out of things to do and can steal it. 4883 oop* from_obj_p = set_partial_array_mask(from_obj); 4884 _par_scan_state->push_on_queue(from_obj_p); 4885 } else { 4886 assert(length == end, "sanity"); 4887 // We'll process the final range for this object. Restore the length 4888 // so that the heap remains parsable in case of evacuation failure. 4889 to_obj_array->set_length(end); 4890 } 4891 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4892 // Process indexes [start,end). It will also process the header 4893 // along with the first chunk (i.e., the chunk with start == 0). 4894 // Note that at this point the length field of to_obj_array is not 4895 // correct given that we are using it to keep track of the next 4896 // start index. oop_iterate_range() (thankfully!) ignores the length 4897 // field and only relies on the start / end parameters. It does 4898 // however return the size of the object which will be incorrect. So 4899 // we have to ignore it even if we wanted to use it. 4900 to_obj_array->oop_iterate_range(&_scanner, start, end); 4901 } 4902 4903 class G1ParEvacuateFollowersClosure : public VoidClosure { 4904 protected: 4905 G1CollectedHeap* _g1h; 4906 G1ParScanThreadState* _par_scan_state; 4907 RefToScanQueueSet* _queues; 4908 ParallelTaskTerminator* _terminator; 4909 4910 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4911 RefToScanQueueSet* queues() { return _queues; } 4912 ParallelTaskTerminator* terminator() { return _terminator; } 4913 4914 public: 4915 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4916 G1ParScanThreadState* par_scan_state, 4917 RefToScanQueueSet* queues, 4918 ParallelTaskTerminator* terminator) 4919 : _g1h(g1h), _par_scan_state(par_scan_state), 4920 _queues(queues), _terminator(terminator) {} 4921 4922 void do_void(); 4923 4924 private: 4925 inline bool offer_termination(); 4926 }; 4927 4928 bool G1ParEvacuateFollowersClosure::offer_termination() { 4929 G1ParScanThreadState* const pss = par_scan_state(); 4930 pss->start_term_time(); 4931 const bool res = terminator()->offer_termination(); 4932 pss->end_term_time(); 4933 return res; 4934 } 4935 4936 void G1ParEvacuateFollowersClosure::do_void() { 4937 StarTask stolen_task; 4938 G1ParScanThreadState* const pss = par_scan_state(); 4939 pss->trim_queue(); 4940 4941 do { 4942 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4943 assert(pss->verify_task(stolen_task), "sanity"); 4944 if (stolen_task.is_narrow()) { 4945 pss->deal_with_reference((narrowOop*) stolen_task); 4946 } else { 4947 pss->deal_with_reference((oop*) stolen_task); 4948 } 4949 4950 // We've just processed a reference and we might have made 4951 // available new entries on the queues. So we have to make sure 4952 // we drain the queues as necessary. 4953 pss->trim_queue(); 4954 } 4955 } while (!offer_termination()); 4956 4957 pss->retire_alloc_buffers(); 4958 } 4959 4960 class G1KlassScanClosure : public KlassClosure { 4961 G1ParCopyHelper* _closure; 4962 bool _process_only_dirty; 4963 int _count; 4964 public: 4965 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4966 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4967 void do_klass(Klass* klass) { 4968 // If the klass has not been dirtied we know that there's 4969 // no references into the young gen and we can skip it. 4970 if (!_process_only_dirty || klass->has_modified_oops()) { 4971 // Clean the klass since we're going to scavenge all the metadata. 4972 klass->clear_modified_oops(); 4973 4974 // Tell the closure that this klass is the Klass to scavenge 4975 // and is the one to dirty if oops are left pointing into the young gen. 4976 _closure->set_scanned_klass(klass); 4977 4978 klass->oops_do(_closure); 4979 4980 _closure->set_scanned_klass(NULL); 4981 } 4982 _count++; 4983 } 4984 }; 4985 4986 class G1ParTask : public AbstractGangTask { 4987 protected: 4988 G1CollectedHeap* _g1h; 4989 RefToScanQueueSet *_queues; 4990 ParallelTaskTerminator _terminator; 4991 uint _n_workers; 4992 4993 Mutex _stats_lock; 4994 Mutex* stats_lock() { return &_stats_lock; } 4995 4996 size_t getNCards() { 4997 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4998 / G1BlockOffsetSharedArray::N_bytes; 4999 } 5000 5001 public: 5002 G1ParTask(G1CollectedHeap* g1h, 5003 RefToScanQueueSet *task_queues) 5004 : AbstractGangTask("G1 collection"), 5005 _g1h(g1h), 5006 _queues(task_queues), 5007 _terminator(0, _queues), 5008 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 5009 {} 5010 5011 RefToScanQueueSet* queues() { return _queues; } 5012 5013 RefToScanQueue *work_queue(int i) { 5014 return queues()->queue(i); 5015 } 5016 5017 ParallelTaskTerminator* terminator() { return &_terminator; } 5018 5019 virtual void set_for_termination(int active_workers) { 5020 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 5021 // in the young space (_par_seq_tasks) in the G1 heap 5022 // for SequentialSubTasksDone. 5023 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 5024 // both of which need setting by set_n_termination(). 5025 _g1h->SharedHeap::set_n_termination(active_workers); 5026 _g1h->set_n_termination(active_workers); 5027 terminator()->reset_for_reuse(active_workers); 5028 _n_workers = active_workers; 5029 } 5030 5031 void work(uint worker_id) { 5032 if (worker_id >= _n_workers) return; // no work needed this round 5033 5034 double start_time_ms = os::elapsedTime() * 1000.0; 5035 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 5036 5037 { 5038 ResourceMark rm; 5039 HandleMark hm; 5040 5041 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 5042 5043 G1ParScanThreadState pss(_g1h, worker_id); 5044 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 5045 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 5046 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 5047 5048 pss.set_evac_closure(&scan_evac_cl); 5049 pss.set_evac_failure_closure(&evac_failure_cl); 5050 pss.set_partial_scan_closure(&partial_scan_cl); 5051 5052 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 5053 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 5054 5055 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 5056 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 5057 5058 bool only_young = _g1h->g1_policy()->gcs_are_young(); 5059 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 5060 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 5061 5062 OopClosure* scan_root_cl = &only_scan_root_cl; 5063 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 5064 5065 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5066 // We also need to mark copied objects. 5067 scan_root_cl = &scan_mark_root_cl; 5068 scan_klasses_cl = &scan_mark_klasses_cl_s; 5069 } 5070 5071 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 5072 5073 // Don't scan the scavengable methods in the code cache as part 5074 // of strong root scanning. The code roots that point into a 5075 // region in the collection set are scanned when we scan the 5076 // region's RSet. 5077 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings; 5078 5079 pss.start_strong_roots(); 5080 _g1h->g1_process_strong_roots(/* is scavenging */ true, 5081 SharedHeap::ScanningOption(so), 5082 scan_root_cl, 5083 &push_heap_rs_cl, 5084 scan_klasses_cl, 5085 worker_id); 5086 pss.end_strong_roots(); 5087 5088 { 5089 double start = os::elapsedTime(); 5090 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 5091 evac.do_void(); 5092 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 5093 double term_ms = pss.term_time()*1000.0; 5094 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 5095 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 5096 } 5097 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 5098 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 5099 5100 if (ParallelGCVerbose) { 5101 MutexLocker x(stats_lock()); 5102 pss.print_termination_stats(worker_id); 5103 } 5104 5105 assert(pss.refs()->is_empty(), "should be empty"); 5106 5107 // Close the inner scope so that the ResourceMark and HandleMark 5108 // destructors are executed here and are included as part of the 5109 // "GC Worker Time". 5110 } 5111 5112 double end_time_ms = os::elapsedTime() * 1000.0; 5113 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 5114 } 5115 }; 5116 5117 // *** Common G1 Evacuation Stuff 5118 5119 // This method is run in a GC worker. 5120 5121 void 5122 G1CollectedHeap:: 5123 g1_process_strong_roots(bool is_scavenging, 5124 ScanningOption so, 5125 OopClosure* scan_non_heap_roots, 5126 OopsInHeapRegionClosure* scan_rs, 5127 G1KlassScanClosure* scan_klasses, 5128 uint worker_i) { 5129 5130 // First scan the strong roots 5131 double ext_roots_start = os::elapsedTime(); 5132 double closure_app_time_sec = 0.0; 5133 5134 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 5135 5136 assert(so & SO_CodeCache || scan_rs != NULL, "must scan code roots somehow"); 5137 // Walk the code cache/strong code roots w/o buffering, because StarTask 5138 // cannot handle unaligned oop locations. 5139 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, true /* do_marking */); 5140 5141 process_strong_roots(false, // no scoping; this is parallel code 5142 is_scavenging, so, 5143 &buf_scan_non_heap_roots, 5144 &eager_scan_code_roots, 5145 scan_klasses 5146 ); 5147 5148 // Now the CM ref_processor roots. 5149 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5150 // We need to treat the discovered reference lists of the 5151 // concurrent mark ref processor as roots and keep entries 5152 // (which are added by the marking threads) on them live 5153 // until they can be processed at the end of marking. 5154 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5155 } 5156 5157 // Finish up any enqueued closure apps (attributed as object copy time). 5158 buf_scan_non_heap_roots.done(); 5159 5160 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5161 5162 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5163 5164 double ext_root_time_ms = 5165 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5166 5167 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5168 5169 // During conc marking we have to filter the per-thread SATB buffers 5170 // to make sure we remove any oops into the CSet (which will show up 5171 // as implicitly live). 5172 double satb_filtering_ms = 0.0; 5173 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5174 if (mark_in_progress()) { 5175 double satb_filter_start = os::elapsedTime(); 5176 5177 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5178 5179 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5180 } 5181 } 5182 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5183 5184 // If this is an initial mark pause, and we're not scanning 5185 // the entire code cache, we need to mark the oops in the 5186 // strong code root lists for the regions that are not in 5187 // the collection set. 5188 // Note all threads participate in this set of root tasks. 5189 double mark_strong_code_roots_ms = 0.0; 5190 if (g1_policy()->during_initial_mark_pause() && !(so & SO_CodeCache)) { 5191 double mark_strong_roots_start = os::elapsedTime(); 5192 mark_strong_code_roots(worker_i); 5193 mark_strong_code_roots_ms = (os::elapsedTime() - mark_strong_roots_start) * 1000.0; 5194 } 5195 g1_policy()->phase_times()->record_strong_code_root_mark_time(worker_i, mark_strong_code_roots_ms); 5196 5197 // Now scan the complement of the collection set. 5198 if (scan_rs != NULL) { 5199 g1_rem_set()->oops_into_collection_set_do(scan_rs, &eager_scan_code_roots, worker_i); 5200 } 5201 _process_strong_tasks->all_tasks_completed(); 5202 } 5203 5204 void 5205 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) { 5206 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5207 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs); 5208 } 5209 5210 // Weak Reference Processing support 5211 5212 // An always "is_alive" closure that is used to preserve referents. 5213 // If the object is non-null then it's alive. Used in the preservation 5214 // of referent objects that are pointed to by reference objects 5215 // discovered by the CM ref processor. 5216 class G1AlwaysAliveClosure: public BoolObjectClosure { 5217 G1CollectedHeap* _g1; 5218 public: 5219 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5220 bool do_object_b(oop p) { 5221 if (p != NULL) { 5222 return true; 5223 } 5224 return false; 5225 } 5226 }; 5227 5228 bool G1STWIsAliveClosure::do_object_b(oop p) { 5229 // An object is reachable if it is outside the collection set, 5230 // or is inside and copied. 5231 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5232 } 5233 5234 // Non Copying Keep Alive closure 5235 class G1KeepAliveClosure: public OopClosure { 5236 G1CollectedHeap* _g1; 5237 public: 5238 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5239 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5240 void do_oop( oop* p) { 5241 oop obj = *p; 5242 5243 if (_g1->obj_in_cs(obj)) { 5244 assert( obj->is_forwarded(), "invariant" ); 5245 *p = obj->forwardee(); 5246 } 5247 } 5248 }; 5249 5250 // Copying Keep Alive closure - can be called from both 5251 // serial and parallel code as long as different worker 5252 // threads utilize different G1ParScanThreadState instances 5253 // and different queues. 5254 5255 class G1CopyingKeepAliveClosure: public OopClosure { 5256 G1CollectedHeap* _g1h; 5257 OopClosure* _copy_non_heap_obj_cl; 5258 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5259 G1ParScanThreadState* _par_scan_state; 5260 5261 public: 5262 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5263 OopClosure* non_heap_obj_cl, 5264 OopsInHeapRegionClosure* metadata_obj_cl, 5265 G1ParScanThreadState* pss): 5266 _g1h(g1h), 5267 _copy_non_heap_obj_cl(non_heap_obj_cl), 5268 _copy_metadata_obj_cl(metadata_obj_cl), 5269 _par_scan_state(pss) 5270 {} 5271 5272 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5273 virtual void do_oop( oop* p) { do_oop_work(p); } 5274 5275 template <class T> void do_oop_work(T* p) { 5276 oop obj = oopDesc::load_decode_heap_oop(p); 5277 5278 if (_g1h->obj_in_cs(obj)) { 5279 // If the referent object has been forwarded (either copied 5280 // to a new location or to itself in the event of an 5281 // evacuation failure) then we need to update the reference 5282 // field and, if both reference and referent are in the G1 5283 // heap, update the RSet for the referent. 5284 // 5285 // If the referent has not been forwarded then we have to keep 5286 // it alive by policy. Therefore we have copy the referent. 5287 // 5288 // If the reference field is in the G1 heap then we can push 5289 // on the PSS queue. When the queue is drained (after each 5290 // phase of reference processing) the object and it's followers 5291 // will be copied, the reference field set to point to the 5292 // new location, and the RSet updated. Otherwise we need to 5293 // use the the non-heap or metadata closures directly to copy 5294 // the referent object and update the pointer, while avoiding 5295 // updating the RSet. 5296 5297 if (_g1h->is_in_g1_reserved(p)) { 5298 _par_scan_state->push_on_queue(p); 5299 } else { 5300 assert(!ClassLoaderDataGraph::contains((address)p), 5301 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5302 PTR_FORMAT, p)); 5303 _copy_non_heap_obj_cl->do_oop(p); 5304 } 5305 } 5306 } 5307 }; 5308 5309 // Serial drain queue closure. Called as the 'complete_gc' 5310 // closure for each discovered list in some of the 5311 // reference processing phases. 5312 5313 class G1STWDrainQueueClosure: public VoidClosure { 5314 protected: 5315 G1CollectedHeap* _g1h; 5316 G1ParScanThreadState* _par_scan_state; 5317 5318 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5319 5320 public: 5321 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5322 _g1h(g1h), 5323 _par_scan_state(pss) 5324 { } 5325 5326 void do_void() { 5327 G1ParScanThreadState* const pss = par_scan_state(); 5328 pss->trim_queue(); 5329 } 5330 }; 5331 5332 // Parallel Reference Processing closures 5333 5334 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5335 // processing during G1 evacuation pauses. 5336 5337 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5338 private: 5339 G1CollectedHeap* _g1h; 5340 RefToScanQueueSet* _queues; 5341 FlexibleWorkGang* _workers; 5342 int _active_workers; 5343 5344 public: 5345 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5346 FlexibleWorkGang* workers, 5347 RefToScanQueueSet *task_queues, 5348 int n_workers) : 5349 _g1h(g1h), 5350 _queues(task_queues), 5351 _workers(workers), 5352 _active_workers(n_workers) 5353 { 5354 assert(n_workers > 0, "shouldn't call this otherwise"); 5355 } 5356 5357 // Executes the given task using concurrent marking worker threads. 5358 virtual void execute(ProcessTask& task); 5359 virtual void execute(EnqueueTask& task); 5360 }; 5361 5362 // Gang task for possibly parallel reference processing 5363 5364 class G1STWRefProcTaskProxy: public AbstractGangTask { 5365 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5366 ProcessTask& _proc_task; 5367 G1CollectedHeap* _g1h; 5368 RefToScanQueueSet *_task_queues; 5369 ParallelTaskTerminator* _terminator; 5370 5371 public: 5372 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5373 G1CollectedHeap* g1h, 5374 RefToScanQueueSet *task_queues, 5375 ParallelTaskTerminator* terminator) : 5376 AbstractGangTask("Process reference objects in parallel"), 5377 _proc_task(proc_task), 5378 _g1h(g1h), 5379 _task_queues(task_queues), 5380 _terminator(terminator) 5381 {} 5382 5383 virtual void work(uint worker_id) { 5384 // The reference processing task executed by a single worker. 5385 ResourceMark rm; 5386 HandleMark hm; 5387 5388 G1STWIsAliveClosure is_alive(_g1h); 5389 5390 G1ParScanThreadState pss(_g1h, worker_id); 5391 5392 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5393 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5394 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5395 5396 pss.set_evac_closure(&scan_evac_cl); 5397 pss.set_evac_failure_closure(&evac_failure_cl); 5398 pss.set_partial_scan_closure(&partial_scan_cl); 5399 5400 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5401 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5402 5403 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5404 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5405 5406 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5407 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5408 5409 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5410 // We also need to mark copied objects. 5411 copy_non_heap_cl = ©_mark_non_heap_cl; 5412 copy_metadata_cl = ©_mark_metadata_cl; 5413 } 5414 5415 // Keep alive closure. 5416 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5417 5418 // Complete GC closure 5419 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5420 5421 // Call the reference processing task's work routine. 5422 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5423 5424 // Note we cannot assert that the refs array is empty here as not all 5425 // of the processing tasks (specifically phase2 - pp2_work) execute 5426 // the complete_gc closure (which ordinarily would drain the queue) so 5427 // the queue may not be empty. 5428 } 5429 }; 5430 5431 // Driver routine for parallel reference processing. 5432 // Creates an instance of the ref processing gang 5433 // task and has the worker threads execute it. 5434 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5435 assert(_workers != NULL, "Need parallel worker threads."); 5436 5437 ParallelTaskTerminator terminator(_active_workers, _queues); 5438 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5439 5440 _g1h->set_par_threads(_active_workers); 5441 _workers->run_task(&proc_task_proxy); 5442 _g1h->set_par_threads(0); 5443 } 5444 5445 // Gang task for parallel reference enqueueing. 5446 5447 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5448 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5449 EnqueueTask& _enq_task; 5450 5451 public: 5452 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5453 AbstractGangTask("Enqueue reference objects in parallel"), 5454 _enq_task(enq_task) 5455 { } 5456 5457 virtual void work(uint worker_id) { 5458 _enq_task.work(worker_id); 5459 } 5460 }; 5461 5462 // Driver routine for parallel reference enqueueing. 5463 // Creates an instance of the ref enqueueing gang 5464 // task and has the worker threads execute it. 5465 5466 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5467 assert(_workers != NULL, "Need parallel worker threads."); 5468 5469 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5470 5471 _g1h->set_par_threads(_active_workers); 5472 _workers->run_task(&enq_task_proxy); 5473 _g1h->set_par_threads(0); 5474 } 5475 5476 // End of weak reference support closures 5477 5478 // Abstract task used to preserve (i.e. copy) any referent objects 5479 // that are in the collection set and are pointed to by reference 5480 // objects discovered by the CM ref processor. 5481 5482 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5483 protected: 5484 G1CollectedHeap* _g1h; 5485 RefToScanQueueSet *_queues; 5486 ParallelTaskTerminator _terminator; 5487 uint _n_workers; 5488 5489 public: 5490 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5491 AbstractGangTask("ParPreserveCMReferents"), 5492 _g1h(g1h), 5493 _queues(task_queues), 5494 _terminator(workers, _queues), 5495 _n_workers(workers) 5496 { } 5497 5498 void work(uint worker_id) { 5499 ResourceMark rm; 5500 HandleMark hm; 5501 5502 G1ParScanThreadState pss(_g1h, worker_id); 5503 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5504 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5505 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5506 5507 pss.set_evac_closure(&scan_evac_cl); 5508 pss.set_evac_failure_closure(&evac_failure_cl); 5509 pss.set_partial_scan_closure(&partial_scan_cl); 5510 5511 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5512 5513 5514 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5515 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5516 5517 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5518 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5519 5520 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5521 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5522 5523 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5524 // We also need to mark copied objects. 5525 copy_non_heap_cl = ©_mark_non_heap_cl; 5526 copy_metadata_cl = ©_mark_metadata_cl; 5527 } 5528 5529 // Is alive closure 5530 G1AlwaysAliveClosure always_alive(_g1h); 5531 5532 // Copying keep alive closure. Applied to referent objects that need 5533 // to be copied. 5534 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5535 5536 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5537 5538 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5539 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5540 5541 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5542 // So this must be true - but assert just in case someone decides to 5543 // change the worker ids. 5544 assert(0 <= worker_id && worker_id < limit, "sanity"); 5545 assert(!rp->discovery_is_atomic(), "check this code"); 5546 5547 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5548 for (uint idx = worker_id; idx < limit; idx += stride) { 5549 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5550 5551 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5552 while (iter.has_next()) { 5553 // Since discovery is not atomic for the CM ref processor, we 5554 // can see some null referent objects. 5555 iter.load_ptrs(DEBUG_ONLY(true)); 5556 oop ref = iter.obj(); 5557 5558 // This will filter nulls. 5559 if (iter.is_referent_alive()) { 5560 iter.make_referent_alive(); 5561 } 5562 iter.move_to_next(); 5563 } 5564 } 5565 5566 // Drain the queue - which may cause stealing 5567 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5568 drain_queue.do_void(); 5569 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5570 assert(pss.refs()->is_empty(), "should be"); 5571 } 5572 }; 5573 5574 // Weak Reference processing during an evacuation pause (part 1). 5575 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5576 double ref_proc_start = os::elapsedTime(); 5577 5578 ReferenceProcessor* rp = _ref_processor_stw; 5579 assert(rp->discovery_enabled(), "should have been enabled"); 5580 5581 // Any reference objects, in the collection set, that were 'discovered' 5582 // by the CM ref processor should have already been copied (either by 5583 // applying the external root copy closure to the discovered lists, or 5584 // by following an RSet entry). 5585 // 5586 // But some of the referents, that are in the collection set, that these 5587 // reference objects point to may not have been copied: the STW ref 5588 // processor would have seen that the reference object had already 5589 // been 'discovered' and would have skipped discovering the reference, 5590 // but would not have treated the reference object as a regular oop. 5591 // As a result the copy closure would not have been applied to the 5592 // referent object. 5593 // 5594 // We need to explicitly copy these referent objects - the references 5595 // will be processed at the end of remarking. 5596 // 5597 // We also need to do this copying before we process the reference 5598 // objects discovered by the STW ref processor in case one of these 5599 // referents points to another object which is also referenced by an 5600 // object discovered by the STW ref processor. 5601 5602 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5603 no_of_gc_workers == workers()->active_workers(), 5604 "Need to reset active GC workers"); 5605 5606 set_par_threads(no_of_gc_workers); 5607 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5608 no_of_gc_workers, 5609 _task_queues); 5610 5611 if (G1CollectedHeap::use_parallel_gc_threads()) { 5612 workers()->run_task(&keep_cm_referents); 5613 } else { 5614 keep_cm_referents.work(0); 5615 } 5616 5617 set_par_threads(0); 5618 5619 // Closure to test whether a referent is alive. 5620 G1STWIsAliveClosure is_alive(this); 5621 5622 // Even when parallel reference processing is enabled, the processing 5623 // of JNI refs is serial and performed serially by the current thread 5624 // rather than by a worker. The following PSS will be used for processing 5625 // JNI refs. 5626 5627 // Use only a single queue for this PSS. 5628 G1ParScanThreadState pss(this, 0); 5629 5630 // We do not embed a reference processor in the copying/scanning 5631 // closures while we're actually processing the discovered 5632 // reference objects. 5633 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5634 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5635 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5636 5637 pss.set_evac_closure(&scan_evac_cl); 5638 pss.set_evac_failure_closure(&evac_failure_cl); 5639 pss.set_partial_scan_closure(&partial_scan_cl); 5640 5641 assert(pss.refs()->is_empty(), "pre-condition"); 5642 5643 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5644 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5645 5646 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5647 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5648 5649 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5650 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5651 5652 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5653 // We also need to mark copied objects. 5654 copy_non_heap_cl = ©_mark_non_heap_cl; 5655 copy_metadata_cl = ©_mark_metadata_cl; 5656 } 5657 5658 // Keep alive closure. 5659 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5660 5661 // Serial Complete GC closure 5662 G1STWDrainQueueClosure drain_queue(this, &pss); 5663 5664 // Setup the soft refs policy... 5665 rp->setup_policy(false); 5666 5667 ReferenceProcessorStats stats; 5668 if (!rp->processing_is_mt()) { 5669 // Serial reference processing... 5670 stats = rp->process_discovered_references(&is_alive, 5671 &keep_alive, 5672 &drain_queue, 5673 NULL, 5674 _gc_timer_stw); 5675 } else { 5676 // Parallel reference processing 5677 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5678 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5679 5680 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5681 stats = rp->process_discovered_references(&is_alive, 5682 &keep_alive, 5683 &drain_queue, 5684 &par_task_executor, 5685 _gc_timer_stw); 5686 } 5687 5688 _gc_tracer_stw->report_gc_reference_stats(stats); 5689 // We have completed copying any necessary live referent objects 5690 // (that were not copied during the actual pause) so we can 5691 // retire any active alloc buffers 5692 pss.retire_alloc_buffers(); 5693 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5694 5695 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5696 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5697 } 5698 5699 // Weak Reference processing during an evacuation pause (part 2). 5700 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5701 double ref_enq_start = os::elapsedTime(); 5702 5703 ReferenceProcessor* rp = _ref_processor_stw; 5704 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5705 5706 // Now enqueue any remaining on the discovered lists on to 5707 // the pending list. 5708 if (!rp->processing_is_mt()) { 5709 // Serial reference processing... 5710 rp->enqueue_discovered_references(); 5711 } else { 5712 // Parallel reference enqueueing 5713 5714 assert(no_of_gc_workers == workers()->active_workers(), 5715 "Need to reset active workers"); 5716 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5717 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5718 5719 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5720 rp->enqueue_discovered_references(&par_task_executor); 5721 } 5722 5723 rp->verify_no_references_recorded(); 5724 assert(!rp->discovery_enabled(), "should have been disabled"); 5725 5726 // FIXME 5727 // CM's reference processing also cleans up the string and symbol tables. 5728 // Should we do that here also? We could, but it is a serial operation 5729 // and could significantly increase the pause time. 5730 5731 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5732 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5733 } 5734 5735 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { 5736 _expand_heap_after_alloc_failure = true; 5737 _evacuation_failed = false; 5738 5739 // Should G1EvacuationFailureALot be in effect for this GC? 5740 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5741 5742 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5743 5744 // Disable the hot card cache. 5745 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache(); 5746 hot_card_cache->reset_hot_cache_claimed_index(); 5747 hot_card_cache->set_use_cache(false); 5748 5749 uint n_workers; 5750 if (G1CollectedHeap::use_parallel_gc_threads()) { 5751 n_workers = 5752 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5753 workers()->active_workers(), 5754 Threads::number_of_non_daemon_threads()); 5755 assert(UseDynamicNumberOfGCThreads || 5756 n_workers == workers()->total_workers(), 5757 "If not dynamic should be using all the workers"); 5758 workers()->set_active_workers(n_workers); 5759 set_par_threads(n_workers); 5760 } else { 5761 assert(n_par_threads() == 0, 5762 "Should be the original non-parallel value"); 5763 n_workers = 1; 5764 } 5765 5766 G1ParTask g1_par_task(this, _task_queues); 5767 5768 init_for_evac_failure(NULL); 5769 5770 rem_set()->prepare_for_younger_refs_iterate(true); 5771 5772 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5773 double start_par_time_sec = os::elapsedTime(); 5774 double end_par_time_sec; 5775 5776 { 5777 StrongRootsScope srs(this); 5778 5779 if (G1CollectedHeap::use_parallel_gc_threads()) { 5780 // The individual threads will set their evac-failure closures. 5781 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5782 // These tasks use ShareHeap::_process_strong_tasks 5783 assert(UseDynamicNumberOfGCThreads || 5784 workers()->active_workers() == workers()->total_workers(), 5785 "If not dynamic should be using all the workers"); 5786 workers()->run_task(&g1_par_task); 5787 } else { 5788 g1_par_task.set_for_termination(n_workers); 5789 g1_par_task.work(0); 5790 } 5791 end_par_time_sec = os::elapsedTime(); 5792 5793 // Closing the inner scope will execute the destructor 5794 // for the StrongRootsScope object. We record the current 5795 // elapsed time before closing the scope so that time 5796 // taken for the SRS destructor is NOT included in the 5797 // reported parallel time. 5798 } 5799 5800 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5801 g1_policy()->phase_times()->record_par_time(par_time_ms); 5802 5803 double code_root_fixup_time_ms = 5804 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5805 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5806 5807 set_par_threads(0); 5808 5809 // Process any discovered reference objects - we have 5810 // to do this _before_ we retire the GC alloc regions 5811 // as we may have to copy some 'reachable' referent 5812 // objects (and their reachable sub-graphs) that were 5813 // not copied during the pause. 5814 process_discovered_references(n_workers); 5815 5816 // Weak root processing. 5817 { 5818 G1STWIsAliveClosure is_alive(this); 5819 G1KeepAliveClosure keep_alive(this); 5820 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5821 } 5822 5823 release_gc_alloc_regions(n_workers, evacuation_info); 5824 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5825 5826 // Reset and re-enable the hot card cache. 5827 // Note the counts for the cards in the regions in the 5828 // collection set are reset when the collection set is freed. 5829 hot_card_cache->reset_hot_cache(); 5830 hot_card_cache->set_use_cache(true); 5831 5832 // Migrate the strong code roots attached to each region in 5833 // the collection set. Ideally we would like to do this 5834 // after we have finished the scanning/evacuation of the 5835 // strong code roots for a particular heap region. 5836 migrate_strong_code_roots(); 5837 5838 if (g1_policy()->during_initial_mark_pause()) { 5839 // Reset the claim values set during marking the strong code roots 5840 reset_heap_region_claim_values(); 5841 } 5842 5843 finalize_for_evac_failure(); 5844 5845 if (evacuation_failed()) { 5846 remove_self_forwarding_pointers(); 5847 5848 // Reset the G1EvacuationFailureALot counters and flags 5849 // Note: the values are reset only when an actual 5850 // evacuation failure occurs. 5851 NOT_PRODUCT(reset_evacuation_should_fail();) 5852 } 5853 5854 // Enqueue any remaining references remaining on the STW 5855 // reference processor's discovered lists. We need to do 5856 // this after the card table is cleaned (and verified) as 5857 // the act of enqueueing entries on to the pending list 5858 // will log these updates (and dirty their associated 5859 // cards). We need these updates logged to update any 5860 // RSets. 5861 enqueue_discovered_references(n_workers); 5862 5863 if (G1DeferredRSUpdate) { 5864 RedirtyLoggedCardTableEntryFastClosure redirty; 5865 dirty_card_queue_set().set_closure(&redirty); 5866 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5867 5868 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5869 dcq.merge_bufferlists(&dirty_card_queue_set()); 5870 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5871 } 5872 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5873 } 5874 5875 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5876 size_t* pre_used, 5877 FreeRegionList* free_list, 5878 OldRegionSet* old_proxy_set, 5879 HumongousRegionSet* humongous_proxy_set, 5880 HRRSCleanupTask* hrrs_cleanup_task, 5881 bool par) { 5882 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5883 if (hr->isHumongous()) { 5884 assert(hr->startsHumongous(), "we should only see starts humongous"); 5885 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5886 } else { 5887 _old_set.remove_with_proxy(hr, old_proxy_set); 5888 free_region(hr, pre_used, free_list, par); 5889 } 5890 } else { 5891 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5892 } 5893 } 5894 5895 void G1CollectedHeap::free_region(HeapRegion* hr, 5896 size_t* pre_used, 5897 FreeRegionList* free_list, 5898 bool par) { 5899 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5900 assert(!hr->is_empty(), "the region should not be empty"); 5901 assert(free_list != NULL, "pre-condition"); 5902 5903 // Clear the card counts for this region. 5904 // Note: we only need to do this if the region is not young 5905 // (since we don't refine cards in young regions). 5906 if (!hr->is_young()) { 5907 _cg1r->hot_card_cache()->reset_card_counts(hr); 5908 } 5909 *pre_used += hr->used(); 5910 hr->hr_clear(par, true /* clear_space */); 5911 free_list->add_as_head(hr); 5912 } 5913 5914 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5915 size_t* pre_used, 5916 FreeRegionList* free_list, 5917 HumongousRegionSet* humongous_proxy_set, 5918 bool par) { 5919 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5920 assert(free_list != NULL, "pre-condition"); 5921 assert(humongous_proxy_set != NULL, "pre-condition"); 5922 5923 size_t hr_used = hr->used(); 5924 size_t hr_capacity = hr->capacity(); 5925 size_t hr_pre_used = 0; 5926 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5927 // We need to read this before we make the region non-humongous, 5928 // otherwise the information will be gone. 5929 uint last_index = hr->last_hc_index(); 5930 hr->set_notHumongous(); 5931 free_region(hr, &hr_pre_used, free_list, par); 5932 5933 uint i = hr->hrs_index() + 1; 5934 while (i < last_index) { 5935 HeapRegion* curr_hr = region_at(i); 5936 assert(curr_hr->continuesHumongous(), "invariant"); 5937 curr_hr->set_notHumongous(); 5938 free_region(curr_hr, &hr_pre_used, free_list, par); 5939 i += 1; 5940 } 5941 assert(hr_pre_used == hr_used, 5942 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5943 "should be the same", hr_pre_used, hr_used)); 5944 *pre_used += hr_pre_used; 5945 } 5946 5947 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5948 FreeRegionList* free_list, 5949 OldRegionSet* old_proxy_set, 5950 HumongousRegionSet* humongous_proxy_set, 5951 bool par) { 5952 if (pre_used > 0) { 5953 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5954 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5955 assert(_summary_bytes_used >= pre_used, 5956 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5957 "should be >= pre_used: "SIZE_FORMAT, 5958 _summary_bytes_used, pre_used)); 5959 _summary_bytes_used -= pre_used; 5960 } 5961 if (free_list != NULL && !free_list->is_empty()) { 5962 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5963 _free_list.add_as_head(free_list); 5964 } 5965 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5966 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5967 _old_set.update_from_proxy(old_proxy_set); 5968 } 5969 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5970 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5971 _humongous_set.update_from_proxy(humongous_proxy_set); 5972 } 5973 } 5974 5975 class G1ParCleanupCTTask : public AbstractGangTask { 5976 CardTableModRefBS* _ct_bs; 5977 G1CollectedHeap* _g1h; 5978 HeapRegion* volatile _su_head; 5979 public: 5980 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5981 G1CollectedHeap* g1h) : 5982 AbstractGangTask("G1 Par Cleanup CT Task"), 5983 _ct_bs(ct_bs), _g1h(g1h) { } 5984 5985 void work(uint worker_id) { 5986 HeapRegion* r; 5987 while (r = _g1h->pop_dirty_cards_region()) { 5988 clear_cards(r); 5989 } 5990 } 5991 5992 void clear_cards(HeapRegion* r) { 5993 // Cards of the survivors should have already been dirtied. 5994 if (!r->is_survivor()) { 5995 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5996 } 5997 } 5998 }; 5999 6000 #ifndef PRODUCT 6001 class G1VerifyCardTableCleanup: public HeapRegionClosure { 6002 G1CollectedHeap* _g1h; 6003 CardTableModRefBS* _ct_bs; 6004 public: 6005 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 6006 : _g1h(g1h), _ct_bs(ct_bs) { } 6007 virtual bool doHeapRegion(HeapRegion* r) { 6008 if (r->is_survivor()) { 6009 _g1h->verify_dirty_region(r); 6010 } else { 6011 _g1h->verify_not_dirty_region(r); 6012 } 6013 return false; 6014 } 6015 }; 6016 6017 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 6018 // All of the region should be clean. 6019 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 6020 MemRegion mr(hr->bottom(), hr->end()); 6021 ct_bs->verify_not_dirty_region(mr); 6022 } 6023 6024 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 6025 // We cannot guarantee that [bottom(),end()] is dirty. Threads 6026 // dirty allocated blocks as they allocate them. The thread that 6027 // retires each region and replaces it with a new one will do a 6028 // maximal allocation to fill in [pre_dummy_top(),end()] but will 6029 // not dirty that area (one less thing to have to do while holding 6030 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 6031 // is dirty. 6032 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6033 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 6034 ct_bs->verify_dirty_region(mr); 6035 } 6036 6037 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 6038 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 6039 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 6040 verify_dirty_region(hr); 6041 } 6042 } 6043 6044 void G1CollectedHeap::verify_dirty_young_regions() { 6045 verify_dirty_young_list(_young_list->first_region()); 6046 } 6047 #endif 6048 6049 void G1CollectedHeap::cleanUpCardTable() { 6050 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 6051 double start = os::elapsedTime(); 6052 6053 { 6054 // Iterate over the dirty cards region list. 6055 G1ParCleanupCTTask cleanup_task(ct_bs, this); 6056 6057 if (G1CollectedHeap::use_parallel_gc_threads()) { 6058 set_par_threads(); 6059 workers()->run_task(&cleanup_task); 6060 set_par_threads(0); 6061 } else { 6062 while (_dirty_cards_region_list) { 6063 HeapRegion* r = _dirty_cards_region_list; 6064 cleanup_task.clear_cards(r); 6065 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 6066 if (_dirty_cards_region_list == r) { 6067 // The last region. 6068 _dirty_cards_region_list = NULL; 6069 } 6070 r->set_next_dirty_cards_region(NULL); 6071 } 6072 } 6073 #ifndef PRODUCT 6074 if (G1VerifyCTCleanup || VerifyAfterGC) { 6075 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 6076 heap_region_iterate(&cleanup_verifier); 6077 } 6078 #endif 6079 } 6080 6081 double elapsed = os::elapsedTime() - start; 6082 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 6083 } 6084 6085 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) { 6086 size_t pre_used = 0; 6087 FreeRegionList local_free_list("Local List for CSet Freeing"); 6088 6089 double young_time_ms = 0.0; 6090 double non_young_time_ms = 0.0; 6091 6092 // Since the collection set is a superset of the the young list, 6093 // all we need to do to clear the young list is clear its 6094 // head and length, and unlink any young regions in the code below 6095 _young_list->clear(); 6096 6097 G1CollectorPolicy* policy = g1_policy(); 6098 6099 double start_sec = os::elapsedTime(); 6100 bool non_young = true; 6101 6102 HeapRegion* cur = cs_head; 6103 int age_bound = -1; 6104 size_t rs_lengths = 0; 6105 6106 while (cur != NULL) { 6107 assert(!is_on_master_free_list(cur), "sanity"); 6108 if (non_young) { 6109 if (cur->is_young()) { 6110 double end_sec = os::elapsedTime(); 6111 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6112 non_young_time_ms += elapsed_ms; 6113 6114 start_sec = os::elapsedTime(); 6115 non_young = false; 6116 } 6117 } else { 6118 if (!cur->is_young()) { 6119 double end_sec = os::elapsedTime(); 6120 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6121 young_time_ms += elapsed_ms; 6122 6123 start_sec = os::elapsedTime(); 6124 non_young = true; 6125 } 6126 } 6127 6128 rs_lengths += cur->rem_set()->occupied(); 6129 6130 HeapRegion* next = cur->next_in_collection_set(); 6131 assert(cur->in_collection_set(), "bad CS"); 6132 cur->set_next_in_collection_set(NULL); 6133 cur->set_in_collection_set(false); 6134 6135 if (cur->is_young()) { 6136 int index = cur->young_index_in_cset(); 6137 assert(index != -1, "invariant"); 6138 assert((uint) index < policy->young_cset_region_length(), "invariant"); 6139 size_t words_survived = _surviving_young_words[index]; 6140 cur->record_surv_words_in_group(words_survived); 6141 6142 // At this point the we have 'popped' cur from the collection set 6143 // (linked via next_in_collection_set()) but it is still in the 6144 // young list (linked via next_young_region()). Clear the 6145 // _next_young_region field. 6146 cur->set_next_young_region(NULL); 6147 } else { 6148 int index = cur->young_index_in_cset(); 6149 assert(index == -1, "invariant"); 6150 } 6151 6152 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 6153 (!cur->is_young() && cur->young_index_in_cset() == -1), 6154 "invariant" ); 6155 6156 if (!cur->evacuation_failed()) { 6157 MemRegion used_mr = cur->used_region(); 6158 6159 // And the region is empty. 6160 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 6161 free_region(cur, &pre_used, &local_free_list, false /* par */); 6162 } else { 6163 cur->uninstall_surv_rate_group(); 6164 if (cur->is_young()) { 6165 cur->set_young_index_in_cset(-1); 6166 } 6167 cur->set_not_young(); 6168 cur->set_evacuation_failed(false); 6169 // The region is now considered to be old. 6170 _old_set.add(cur); 6171 evacuation_info.increment_collectionset_used_after(cur->used()); 6172 } 6173 cur = next; 6174 } 6175 6176 evacuation_info.set_regions_freed(local_free_list.length()); 6177 policy->record_max_rs_lengths(rs_lengths); 6178 policy->cset_regions_freed(); 6179 6180 double end_sec = os::elapsedTime(); 6181 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6182 6183 if (non_young) { 6184 non_young_time_ms += elapsed_ms; 6185 } else { 6186 young_time_ms += elapsed_ms; 6187 } 6188 6189 update_sets_after_freeing_regions(pre_used, &local_free_list, 6190 NULL /* old_proxy_set */, 6191 NULL /* humongous_proxy_set */, 6192 false /* par */); 6193 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6194 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6195 } 6196 6197 // This routine is similar to the above but does not record 6198 // any policy statistics or update free lists; we are abandoning 6199 // the current incremental collection set in preparation of a 6200 // full collection. After the full GC we will start to build up 6201 // the incremental collection set again. 6202 // This is only called when we're doing a full collection 6203 // and is immediately followed by the tearing down of the young list. 6204 6205 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6206 HeapRegion* cur = cs_head; 6207 6208 while (cur != NULL) { 6209 HeapRegion* next = cur->next_in_collection_set(); 6210 assert(cur->in_collection_set(), "bad CS"); 6211 cur->set_next_in_collection_set(NULL); 6212 cur->set_in_collection_set(false); 6213 cur->set_young_index_in_cset(-1); 6214 cur = next; 6215 } 6216 } 6217 6218 void G1CollectedHeap::set_free_regions_coming() { 6219 if (G1ConcRegionFreeingVerbose) { 6220 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6221 "setting free regions coming"); 6222 } 6223 6224 assert(!free_regions_coming(), "pre-condition"); 6225 _free_regions_coming = true; 6226 } 6227 6228 void G1CollectedHeap::reset_free_regions_coming() { 6229 assert(free_regions_coming(), "pre-condition"); 6230 6231 { 6232 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6233 _free_regions_coming = false; 6234 SecondaryFreeList_lock->notify_all(); 6235 } 6236 6237 if (G1ConcRegionFreeingVerbose) { 6238 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6239 "reset free regions coming"); 6240 } 6241 } 6242 6243 void G1CollectedHeap::wait_while_free_regions_coming() { 6244 // Most of the time we won't have to wait, so let's do a quick test 6245 // first before we take the lock. 6246 if (!free_regions_coming()) { 6247 return; 6248 } 6249 6250 if (G1ConcRegionFreeingVerbose) { 6251 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6252 "waiting for free regions"); 6253 } 6254 6255 { 6256 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6257 while (free_regions_coming()) { 6258 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6259 } 6260 } 6261 6262 if (G1ConcRegionFreeingVerbose) { 6263 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6264 "done waiting for free regions"); 6265 } 6266 } 6267 6268 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6269 assert(heap_lock_held_for_gc(), 6270 "the heap lock should already be held by or for this thread"); 6271 _young_list->push_region(hr); 6272 } 6273 6274 class NoYoungRegionsClosure: public HeapRegionClosure { 6275 private: 6276 bool _success; 6277 public: 6278 NoYoungRegionsClosure() : _success(true) { } 6279 bool doHeapRegion(HeapRegion* r) { 6280 if (r->is_young()) { 6281 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6282 r->bottom(), r->end()); 6283 _success = false; 6284 } 6285 return false; 6286 } 6287 bool success() { return _success; } 6288 }; 6289 6290 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6291 bool ret = _young_list->check_list_empty(check_sample); 6292 6293 if (check_heap) { 6294 NoYoungRegionsClosure closure; 6295 heap_region_iterate(&closure); 6296 ret = ret && closure.success(); 6297 } 6298 6299 return ret; 6300 } 6301 6302 class TearDownRegionSetsClosure : public HeapRegionClosure { 6303 private: 6304 OldRegionSet *_old_set; 6305 6306 public: 6307 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6308 6309 bool doHeapRegion(HeapRegion* r) { 6310 if (r->is_empty()) { 6311 // We ignore empty regions, we'll empty the free list afterwards 6312 } else if (r->is_young()) { 6313 // We ignore young regions, we'll empty the young list afterwards 6314 } else if (r->isHumongous()) { 6315 // We ignore humongous regions, we're not tearing down the 6316 // humongous region set 6317 } else { 6318 // The rest should be old 6319 _old_set->remove(r); 6320 } 6321 return false; 6322 } 6323 6324 ~TearDownRegionSetsClosure() { 6325 assert(_old_set->is_empty(), "post-condition"); 6326 } 6327 }; 6328 6329 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6330 assert_at_safepoint(true /* should_be_vm_thread */); 6331 6332 if (!free_list_only) { 6333 TearDownRegionSetsClosure cl(&_old_set); 6334 heap_region_iterate(&cl); 6335 6336 // Need to do this after the heap iteration to be able to 6337 // recognize the young regions and ignore them during the iteration. 6338 _young_list->empty_list(); 6339 } 6340 _free_list.remove_all(); 6341 } 6342 6343 class RebuildRegionSetsClosure : public HeapRegionClosure { 6344 private: 6345 bool _free_list_only; 6346 OldRegionSet* _old_set; 6347 FreeRegionList* _free_list; 6348 size_t _total_used; 6349 6350 public: 6351 RebuildRegionSetsClosure(bool free_list_only, 6352 OldRegionSet* old_set, FreeRegionList* free_list) : 6353 _free_list_only(free_list_only), 6354 _old_set(old_set), _free_list(free_list), _total_used(0) { 6355 assert(_free_list->is_empty(), "pre-condition"); 6356 if (!free_list_only) { 6357 assert(_old_set->is_empty(), "pre-condition"); 6358 } 6359 } 6360 6361 bool doHeapRegion(HeapRegion* r) { 6362 if (r->continuesHumongous()) { 6363 return false; 6364 } 6365 6366 if (r->is_empty()) { 6367 // Add free regions to the free list 6368 _free_list->add_as_tail(r); 6369 } else if (!_free_list_only) { 6370 assert(!r->is_young(), "we should not come across young regions"); 6371 6372 if (r->isHumongous()) { 6373 // We ignore humongous regions, we left the humongous set unchanged 6374 } else { 6375 // The rest should be old, add them to the old set 6376 _old_set->add(r); 6377 } 6378 _total_used += r->used(); 6379 } 6380 6381 return false; 6382 } 6383 6384 size_t total_used() { 6385 return _total_used; 6386 } 6387 }; 6388 6389 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6390 assert_at_safepoint(true /* should_be_vm_thread */); 6391 6392 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6393 heap_region_iterate(&cl); 6394 6395 if (!free_list_only) { 6396 _summary_bytes_used = cl.total_used(); 6397 } 6398 assert(_summary_bytes_used == recalculate_used(), 6399 err_msg("inconsistent _summary_bytes_used, " 6400 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6401 _summary_bytes_used, recalculate_used())); 6402 } 6403 6404 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6405 _refine_cte_cl->set_concurrent(concurrent); 6406 } 6407 6408 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6409 HeapRegion* hr = heap_region_containing(p); 6410 if (hr == NULL) { 6411 return false; 6412 } else { 6413 return hr->is_in(p); 6414 } 6415 } 6416 6417 // Methods for the mutator alloc region 6418 6419 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6420 bool force) { 6421 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6422 assert(!force || g1_policy()->can_expand_young_list(), 6423 "if force is true we should be able to expand the young list"); 6424 bool young_list_full = g1_policy()->is_young_list_full(); 6425 if (force || !young_list_full) { 6426 HeapRegion* new_alloc_region = new_region(word_size, 6427 false /* do_expand */); 6428 if (new_alloc_region != NULL) { 6429 set_region_short_lived_locked(new_alloc_region); 6430 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6431 return new_alloc_region; 6432 } 6433 } 6434 return NULL; 6435 } 6436 6437 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6438 size_t allocated_bytes) { 6439 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6440 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6441 6442 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6443 _summary_bytes_used += allocated_bytes; 6444 _hr_printer.retire(alloc_region); 6445 // We update the eden sizes here, when the region is retired, 6446 // instead of when it's allocated, since this is the point that its 6447 // used space has been recored in _summary_bytes_used. 6448 g1mm()->update_eden_size(); 6449 } 6450 6451 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6452 bool force) { 6453 return _g1h->new_mutator_alloc_region(word_size, force); 6454 } 6455 6456 void G1CollectedHeap::set_par_threads() { 6457 // Don't change the number of workers. Use the value previously set 6458 // in the workgroup. 6459 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6460 uint n_workers = workers()->active_workers(); 6461 assert(UseDynamicNumberOfGCThreads || 6462 n_workers == workers()->total_workers(), 6463 "Otherwise should be using the total number of workers"); 6464 if (n_workers == 0) { 6465 assert(false, "Should have been set in prior evacuation pause."); 6466 n_workers = ParallelGCThreads; 6467 workers()->set_active_workers(n_workers); 6468 } 6469 set_par_threads(n_workers); 6470 } 6471 6472 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6473 size_t allocated_bytes) { 6474 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6475 } 6476 6477 // Methods for the GC alloc regions 6478 6479 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6480 uint count, 6481 GCAllocPurpose ap) { 6482 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6483 6484 if (count < g1_policy()->max_regions(ap)) { 6485 HeapRegion* new_alloc_region = new_region(word_size, 6486 true /* do_expand */); 6487 if (new_alloc_region != NULL) { 6488 // We really only need to do this for old regions given that we 6489 // should never scan survivors. But it doesn't hurt to do it 6490 // for survivors too. 6491 new_alloc_region->set_saved_mark(); 6492 if (ap == GCAllocForSurvived) { 6493 new_alloc_region->set_survivor(); 6494 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6495 } else { 6496 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6497 } 6498 bool during_im = g1_policy()->during_initial_mark_pause(); 6499 new_alloc_region->note_start_of_copying(during_im); 6500 return new_alloc_region; 6501 } else { 6502 g1_policy()->note_alloc_region_limit_reached(ap); 6503 } 6504 } 6505 return NULL; 6506 } 6507 6508 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6509 size_t allocated_bytes, 6510 GCAllocPurpose ap) { 6511 bool during_im = g1_policy()->during_initial_mark_pause(); 6512 alloc_region->note_end_of_copying(during_im); 6513 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6514 if (ap == GCAllocForSurvived) { 6515 young_list()->add_survivor_region(alloc_region); 6516 } else { 6517 _old_set.add(alloc_region); 6518 } 6519 _hr_printer.retire(alloc_region); 6520 } 6521 6522 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6523 bool force) { 6524 assert(!force, "not supported for GC alloc regions"); 6525 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6526 } 6527 6528 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6529 size_t allocated_bytes) { 6530 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6531 GCAllocForSurvived); 6532 } 6533 6534 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6535 bool force) { 6536 assert(!force, "not supported for GC alloc regions"); 6537 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6538 } 6539 6540 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6541 size_t allocated_bytes) { 6542 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6543 GCAllocForTenured); 6544 } 6545 // Heap region set verification 6546 6547 class VerifyRegionListsClosure : public HeapRegionClosure { 6548 private: 6549 FreeRegionList* _free_list; 6550 OldRegionSet* _old_set; 6551 HumongousRegionSet* _humongous_set; 6552 uint _region_count; 6553 6554 public: 6555 VerifyRegionListsClosure(OldRegionSet* old_set, 6556 HumongousRegionSet* humongous_set, 6557 FreeRegionList* free_list) : 6558 _old_set(old_set), _humongous_set(humongous_set), 6559 _free_list(free_list), _region_count(0) { } 6560 6561 uint region_count() { return _region_count; } 6562 6563 bool doHeapRegion(HeapRegion* hr) { 6564 _region_count += 1; 6565 6566 if (hr->continuesHumongous()) { 6567 return false; 6568 } 6569 6570 if (hr->is_young()) { 6571 // TODO 6572 } else if (hr->startsHumongous()) { 6573 _humongous_set->verify_next_region(hr); 6574 } else if (hr->is_empty()) { 6575 _free_list->verify_next_region(hr); 6576 } else { 6577 _old_set->verify_next_region(hr); 6578 } 6579 return false; 6580 } 6581 }; 6582 6583 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6584 HeapWord* bottom) { 6585 HeapWord* end = bottom + HeapRegion::GrainWords; 6586 MemRegion mr(bottom, end); 6587 assert(_g1_reserved.contains(mr), "invariant"); 6588 // This might return NULL if the allocation fails 6589 return new HeapRegion(hrs_index, _bot_shared, mr); 6590 } 6591 6592 void G1CollectedHeap::verify_region_sets() { 6593 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6594 6595 // First, check the explicit lists. 6596 _free_list.verify(); 6597 { 6598 // Given that a concurrent operation might be adding regions to 6599 // the secondary free list we have to take the lock before 6600 // verifying it. 6601 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6602 _secondary_free_list.verify(); 6603 } 6604 _old_set.verify(); 6605 _humongous_set.verify(); 6606 6607 // If a concurrent region freeing operation is in progress it will 6608 // be difficult to correctly attributed any free regions we come 6609 // across to the correct free list given that they might belong to 6610 // one of several (free_list, secondary_free_list, any local lists, 6611 // etc.). So, if that's the case we will skip the rest of the 6612 // verification operation. Alternatively, waiting for the concurrent 6613 // operation to complete will have a non-trivial effect on the GC's 6614 // operation (no concurrent operation will last longer than the 6615 // interval between two calls to verification) and it might hide 6616 // any issues that we would like to catch during testing. 6617 if (free_regions_coming()) { 6618 return; 6619 } 6620 6621 // Make sure we append the secondary_free_list on the free_list so 6622 // that all free regions we will come across can be safely 6623 // attributed to the free_list. 6624 append_secondary_free_list_if_not_empty_with_lock(); 6625 6626 // Finally, make sure that the region accounting in the lists is 6627 // consistent with what we see in the heap. 6628 _old_set.verify_start(); 6629 _humongous_set.verify_start(); 6630 _free_list.verify_start(); 6631 6632 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6633 heap_region_iterate(&cl); 6634 6635 _old_set.verify_end(); 6636 _humongous_set.verify_end(); 6637 _free_list.verify_end(); 6638 } 6639 6640 // Optimized nmethod scanning 6641 6642 class RegisterNMethodOopClosure: public OopClosure { 6643 G1CollectedHeap* _g1h; 6644 nmethod* _nm; 6645 6646 template <class T> void do_oop_work(T* p) { 6647 T heap_oop = oopDesc::load_heap_oop(p); 6648 if (!oopDesc::is_null(heap_oop)) { 6649 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6650 HeapRegion* hr = _g1h->heap_region_containing(obj); 6651 assert(!hr->isHumongous(), "code root in humongous region?"); 6652 6653 // HeapRegion::add_strong_code_root() avoids adding duplicate 6654 // entries but having duplicates is OK since we "mark" nmethods 6655 // as visited when we scan the strong code root lists during the GC. 6656 hr->add_strong_code_root(_nm); 6657 assert(hr->rem_set()->strong_code_roots_list_contains(_nm), "add failed?"); 6658 } 6659 } 6660 6661 public: 6662 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6663 _g1h(g1h), _nm(nm) {} 6664 6665 void do_oop(oop* p) { do_oop_work(p); } 6666 void do_oop(narrowOop* p) { do_oop_work(p); } 6667 }; 6668 6669 class UnregisterNMethodOopClosure: public OopClosure { 6670 G1CollectedHeap* _g1h; 6671 nmethod* _nm; 6672 6673 template <class T> void do_oop_work(T* p) { 6674 T heap_oop = oopDesc::load_heap_oop(p); 6675 if (!oopDesc::is_null(heap_oop)) { 6676 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6677 HeapRegion* hr = _g1h->heap_region_containing(obj); 6678 assert(!hr->isHumongous(), "code root in humongous region?"); 6679 hr->remove_strong_code_root(_nm); 6680 assert(!hr->rem_set()->strong_code_roots_list_contains(_nm), "remove failed?"); 6681 } 6682 } 6683 6684 public: 6685 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) : 6686 _g1h(g1h), _nm(nm) {} 6687 6688 void do_oop(oop* p) { do_oop_work(p); } 6689 void do_oop(narrowOop* p) { do_oop_work(p); } 6690 }; 6691 6692 void G1CollectedHeap::register_nmethod(nmethod* nm) { 6693 CollectedHeap::register_nmethod(nm); 6694 6695 guarantee(nm != NULL, "sanity"); 6696 RegisterNMethodOopClosure reg_cl(this, nm); 6697 nm->oops_do(®_cl); 6698 } 6699 6700 void G1CollectedHeap::unregister_nmethod(nmethod* nm) { 6701 CollectedHeap::unregister_nmethod(nm); 6702 6703 guarantee(nm != NULL, "sanity"); 6704 UnregisterNMethodOopClosure reg_cl(this, nm); 6705 nm->oops_do(®_cl, true); 6706 } 6707 6708 class MigrateCodeRootsHeapRegionClosure: public HeapRegionClosure { 6709 public: 6710 bool doHeapRegion(HeapRegion *hr) { 6711 assert(!hr->isHumongous(), "humongous region in collection set?"); 6712 hr->migrate_strong_code_roots(); 6713 return false; 6714 } 6715 }; 6716 6717 void G1CollectedHeap::migrate_strong_code_roots() { 6718 MigrateCodeRootsHeapRegionClosure cl; 6719 double migrate_start = os::elapsedTime(); 6720 collection_set_iterate(&cl); 6721 double migration_time_ms = (os::elapsedTime() - migrate_start) * 1000.0; 6722 g1_policy()->phase_times()->record_strong_code_root_migration_time(migration_time_ms); 6723 } 6724 6725 // Mark all the code roots that point into regions *not* in the 6726 // collection set. 6727 // 6728 // Note we do not want to use a "marking" CodeBlobToOopClosure while 6729 // walking the the code roots lists of regions not in the collection 6730 // set. Suppose we have an nmethod (M) that points to objects in two 6731 // separate regions - one in the collection set (R1) and one not (R2). 6732 // Using a "marking" CodeBlobToOopClosure here would result in "marking" 6733 // nmethod M when walking the code roots for R1. When we come to scan 6734 // the code roots for R2, we would see that M is already marked and it 6735 // would be skipped and the objects in R2 that are referenced from M 6736 // would not be evacuated. 6737 6738 class MarkStrongCodeRootCodeBlobClosure: public CodeBlobClosure { 6739 6740 class MarkStrongCodeRootOopClosure: public OopClosure { 6741 ConcurrentMark* _cm; 6742 HeapRegion* _hr; 6743 uint _worker_id; 6744 6745 template <class T> void do_oop_work(T* p) { 6746 T heap_oop = oopDesc::load_heap_oop(p); 6747 if (!oopDesc::is_null(heap_oop)) { 6748 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 6749 // Only mark objects in the region (which is assumed 6750 // to be not in the collection set). 6751 if (_hr->is_in(obj)) { 6752 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 6753 } 6754 } 6755 } 6756 6757 public: 6758 MarkStrongCodeRootOopClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id) : 6759 _cm(cm), _hr(hr), _worker_id(worker_id) { 6760 assert(!_hr->in_collection_set(), "sanity"); 6761 } 6762 6763 void do_oop(narrowOop* p) { do_oop_work(p); } 6764 void do_oop(oop* p) { do_oop_work(p); } 6765 }; 6766 6767 MarkStrongCodeRootOopClosure _oop_cl; 6768 6769 public: 6770 MarkStrongCodeRootCodeBlobClosure(ConcurrentMark* cm, HeapRegion* hr, uint worker_id): 6771 _oop_cl(cm, hr, worker_id) {} 6772 6773 void do_code_blob(CodeBlob* cb) { 6774 nmethod* nm = (cb == NULL) ? NULL : cb->as_nmethod_or_null(); 6775 if (nm != NULL) { 6776 nm->oops_do(&_oop_cl); 6777 } 6778 } 6779 }; 6780 6781 class MarkStrongCodeRootsHRClosure: public HeapRegionClosure { 6782 G1CollectedHeap* _g1h; 6783 uint _worker_id; 6784 6785 public: 6786 MarkStrongCodeRootsHRClosure(G1CollectedHeap* g1h, uint worker_id) : 6787 _g1h(g1h), _worker_id(worker_id) {} 6788 6789 bool doHeapRegion(HeapRegion *hr) { 6790 HeapRegionRemSet* hrrs = hr->rem_set(); 6791 if (hr->isHumongous()) { 6792 // Code roots should never be attached to a humongous region 6793 assert(hrrs->strong_code_roots_list_length() == 0, "sanity"); 6794 return false; 6795 } 6796 6797 if (hr->in_collection_set()) { 6798 // Don't mark code roots into regions in the collection set here. 6799 // They will be marked when we scan them. 6800 return false; 6801 } 6802 6803 MarkStrongCodeRootCodeBlobClosure cb_cl(_g1h->concurrent_mark(), hr, _worker_id); 6804 hr->strong_code_roots_do(&cb_cl); 6805 return false; 6806 } 6807 }; 6808 6809 void G1CollectedHeap::mark_strong_code_roots(uint worker_id) { 6810 MarkStrongCodeRootsHRClosure cl(this, worker_id); 6811 if (G1CollectedHeap::use_parallel_gc_threads()) { 6812 heap_region_par_iterate_chunked(&cl, 6813 worker_id, 6814 workers()->active_workers(), 6815 HeapRegion::ParMarkRootClaimValue); 6816 } else { 6817 heap_region_iterate(&cl); 6818 } 6819 } 6820 6821 class RebuildStrongCodeRootClosure: public CodeBlobClosure { 6822 G1CollectedHeap* _g1h; 6823 6824 public: 6825 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) : 6826 _g1h(g1h) {} 6827 6828 void do_code_blob(CodeBlob* cb) { 6829 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL; 6830 if (nm == NULL) { 6831 return; 6832 } 6833 6834 if (ScavengeRootsInCode && nm->detect_scavenge_root_oops()) { 6835 _g1h->register_nmethod(nm); 6836 } 6837 } 6838 }; 6839 6840 void G1CollectedHeap::rebuild_strong_code_roots() { 6841 RebuildStrongCodeRootClosure blob_cl(this); 6842 CodeCache::blobs_do(&blob_cl); 6843 }