1 /* 2 * Copyright (c) 2001, 2011, 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/icBuffer.hpp" 27 #include "gc_implementation/g1/bufferingOopClosure.hpp" 28 #include "gc_implementation/g1/concurrentG1Refine.hpp" 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp" 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp" 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 35 #include "gc_implementation/g1/g1MarkSweep.hpp" 36 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 37 #include "gc_implementation/g1/g1RemSet.inline.hpp" 38 #include "gc_implementation/g1/heapRegionRemSet.hpp" 39 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 40 #include "gc_implementation/g1/vm_operations_g1.hpp" 41 #include "gc_implementation/shared/isGCActiveMark.hpp" 42 #include "memory/gcLocker.inline.hpp" 43 #include "memory/genOopClosures.inline.hpp" 44 #include "memory/generationSpec.hpp" 45 #include "memory/referenceProcessor.hpp" 46 #include "oops/oop.inline.hpp" 47 #include "oops/oop.pcgc.inline.hpp" 48 #include "runtime/aprofiler.hpp" 49 #include "runtime/vmThread.hpp" 50 51 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 52 53 // turn it on so that the contents of the young list (scan-only / 54 // to-be-collected) are printed at "strategic" points before / during 55 // / after the collection --- this is useful for debugging 56 #define YOUNG_LIST_VERBOSE 0 57 // CURRENT STATUS 58 // This file is under construction. Search for "FIXME". 59 60 // INVARIANTS/NOTES 61 // 62 // All allocation activity covered by the G1CollectedHeap interface is 63 // serialized by acquiring the HeapLock. This happens in mem_allocate 64 // and allocate_new_tlab, which are the "entry" points to the 65 // allocation code from the rest of the JVM. (Note that this does not 66 // apply to TLAB allocation, which is not part of this interface: it 67 // is done by clients of this interface.) 68 69 // Notes on implementation of parallelism in different tasks. 70 // 71 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 72 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 73 // It does use run_task() which sets _n_workers in the task. 74 // G1ParTask executes g1_process_strong_roots() -> 75 // SharedHeap::process_strong_roots() which calls eventuall to 76 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 77 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 78 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 79 // 80 81 // Local to this file. 82 83 class RefineCardTableEntryClosure: public CardTableEntryClosure { 84 SuspendibleThreadSet* _sts; 85 G1RemSet* _g1rs; 86 ConcurrentG1Refine* _cg1r; 87 bool _concurrent; 88 public: 89 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 90 G1RemSet* g1rs, 91 ConcurrentG1Refine* cg1r) : 92 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 93 {} 94 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 95 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false); 96 // This path is executed by the concurrent refine or mutator threads, 97 // concurrently, and so we do not care if card_ptr contains references 98 // that point into the collection set. 99 assert(!oops_into_cset, "should be"); 100 101 if (_concurrent && _sts->should_yield()) { 102 // Caller will actually yield. 103 return false; 104 } 105 // Otherwise, we finished successfully; return true. 106 return true; 107 } 108 void set_concurrent(bool b) { _concurrent = b; } 109 }; 110 111 112 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 113 int _calls; 114 G1CollectedHeap* _g1h; 115 CardTableModRefBS* _ctbs; 116 int _histo[256]; 117 public: 118 ClearLoggedCardTableEntryClosure() : 119 _calls(0) 120 { 121 _g1h = G1CollectedHeap::heap(); 122 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 123 for (int i = 0; i < 256; i++) _histo[i] = 0; 124 } 125 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 126 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 127 _calls++; 128 unsigned char* ujb = (unsigned char*)card_ptr; 129 int ind = (int)(*ujb); 130 _histo[ind]++; 131 *card_ptr = -1; 132 } 133 return true; 134 } 135 int calls() { return _calls; } 136 void print_histo() { 137 gclog_or_tty->print_cr("Card table value histogram:"); 138 for (int i = 0; i < 256; i++) { 139 if (_histo[i] != 0) { 140 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 141 } 142 } 143 } 144 }; 145 146 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 147 int _calls; 148 G1CollectedHeap* _g1h; 149 CardTableModRefBS* _ctbs; 150 public: 151 RedirtyLoggedCardTableEntryClosure() : 152 _calls(0) 153 { 154 _g1h = G1CollectedHeap::heap(); 155 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 156 } 157 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 158 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 159 _calls++; 160 *card_ptr = 0; 161 } 162 return true; 163 } 164 int calls() { return _calls; } 165 }; 166 167 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 168 public: 169 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 170 *card_ptr = CardTableModRefBS::dirty_card_val(); 171 return true; 172 } 173 }; 174 175 YoungList::YoungList(G1CollectedHeap* g1h) 176 : _g1h(g1h), _head(NULL), 177 _length(0), 178 _last_sampled_rs_lengths(0), 179 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) 180 { 181 guarantee( check_list_empty(false), "just making sure..." ); 182 } 183 184 void YoungList::push_region(HeapRegion *hr) { 185 assert(!hr->is_young(), "should not already be young"); 186 assert(hr->get_next_young_region() == NULL, "cause it should!"); 187 188 hr->set_next_young_region(_head); 189 _head = hr; 190 191 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 192 ++_length; 193 } 194 195 void YoungList::add_survivor_region(HeapRegion* hr) { 196 assert(hr->is_survivor(), "should be flagged as survivor region"); 197 assert(hr->get_next_young_region() == NULL, "cause it should!"); 198 199 hr->set_next_young_region(_survivor_head); 200 if (_survivor_head == NULL) { 201 _survivor_tail = hr; 202 } 203 _survivor_head = hr; 204 ++_survivor_length; 205 } 206 207 void YoungList::empty_list(HeapRegion* list) { 208 while (list != NULL) { 209 HeapRegion* next = list->get_next_young_region(); 210 list->set_next_young_region(NULL); 211 list->uninstall_surv_rate_group(); 212 list->set_not_young(); 213 list = next; 214 } 215 } 216 217 void YoungList::empty_list() { 218 assert(check_list_well_formed(), "young list should be well formed"); 219 220 empty_list(_head); 221 _head = NULL; 222 _length = 0; 223 224 empty_list(_survivor_head); 225 _survivor_head = NULL; 226 _survivor_tail = NULL; 227 _survivor_length = 0; 228 229 _last_sampled_rs_lengths = 0; 230 231 assert(check_list_empty(false), "just making sure..."); 232 } 233 234 bool YoungList::check_list_well_formed() { 235 bool ret = true; 236 237 size_t length = 0; 238 HeapRegion* curr = _head; 239 HeapRegion* last = NULL; 240 while (curr != NULL) { 241 if (!curr->is_young()) { 242 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 243 "incorrectly tagged (y: %d, surv: %d)", 244 curr->bottom(), curr->end(), 245 curr->is_young(), curr->is_survivor()); 246 ret = false; 247 } 248 ++length; 249 last = curr; 250 curr = curr->get_next_young_region(); 251 } 252 ret = ret && (length == _length); 253 254 if (!ret) { 255 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 256 gclog_or_tty->print_cr("### list has %d entries, _length is %d", 257 length, _length); 258 } 259 260 return ret; 261 } 262 263 bool YoungList::check_list_empty(bool check_sample) { 264 bool ret = true; 265 266 if (_length != 0) { 267 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %d", 268 _length); 269 ret = false; 270 } 271 if (check_sample && _last_sampled_rs_lengths != 0) { 272 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 273 ret = false; 274 } 275 if (_head != NULL) { 276 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 277 ret = false; 278 } 279 if (!ret) { 280 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 281 } 282 283 return ret; 284 } 285 286 void 287 YoungList::rs_length_sampling_init() { 288 _sampled_rs_lengths = 0; 289 _curr = _head; 290 } 291 292 bool 293 YoungList::rs_length_sampling_more() { 294 return _curr != NULL; 295 } 296 297 void 298 YoungList::rs_length_sampling_next() { 299 assert( _curr != NULL, "invariant" ); 300 size_t rs_length = _curr->rem_set()->occupied(); 301 302 _sampled_rs_lengths += rs_length; 303 304 // The current region may not yet have been added to the 305 // incremental collection set (it gets added when it is 306 // retired as the current allocation region). 307 if (_curr->in_collection_set()) { 308 // Update the collection set policy information for this region 309 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 310 } 311 312 _curr = _curr->get_next_young_region(); 313 if (_curr == NULL) { 314 _last_sampled_rs_lengths = _sampled_rs_lengths; 315 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 316 } 317 } 318 319 void 320 YoungList::reset_auxilary_lists() { 321 guarantee( is_empty(), "young list should be empty" ); 322 assert(check_list_well_formed(), "young list should be well formed"); 323 324 // Add survivor regions to SurvRateGroup. 325 _g1h->g1_policy()->note_start_adding_survivor_regions(); 326 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 327 328 int young_index_in_cset = 0; 329 for (HeapRegion* curr = _survivor_head; 330 curr != NULL; 331 curr = curr->get_next_young_region()) { 332 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 333 334 // The region is a non-empty survivor so let's add it to 335 // the incremental collection set for the next evacuation 336 // pause. 337 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 338 young_index_in_cset += 1; 339 } 340 assert((size_t) young_index_in_cset == _survivor_length, 341 "post-condition"); 342 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 343 344 _head = _survivor_head; 345 _length = _survivor_length; 346 if (_survivor_head != NULL) { 347 assert(_survivor_tail != NULL, "cause it shouldn't be"); 348 assert(_survivor_length > 0, "invariant"); 349 _survivor_tail->set_next_young_region(NULL); 350 } 351 352 // Don't clear the survivor list handles until the start of 353 // the next evacuation pause - we need it in order to re-tag 354 // the survivor regions from this evacuation pause as 'young' 355 // at the start of the next. 356 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 358 359 assert(check_list_well_formed(), "young list should be well formed"); 360 } 361 362 void YoungList::print() { 363 HeapRegion* lists[] = {_head, _survivor_head}; 364 const char* names[] = {"YOUNG", "SURVIVOR"}; 365 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 368 HeapRegion *curr = lists[list]; 369 if (curr == NULL) 370 gclog_or_tty->print_cr(" empty"); 371 while (curr != NULL) { 372 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, " 373 "age: %4d, y: %d, surv: %d", 374 curr->bottom(), curr->end(), 375 curr->top(), 376 curr->prev_top_at_mark_start(), 377 curr->next_top_at_mark_start(), 378 curr->top_at_conc_mark_count(), 379 curr->age_in_surv_rate_group_cond(), 380 curr->is_young(), 381 curr->is_survivor()); 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 collecction. 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 // perm gen (or null) 472 return false; 473 } else { 474 return !hr->isHumongous(); 475 } 476 } 477 478 void G1CollectedHeap::check_ct_logs_at_safepoint() { 479 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 480 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 481 482 // Count the dirty cards at the start. 483 CountNonCleanMemRegionClosure count1(this); 484 ct_bs->mod_card_iterate(&count1); 485 int orig_count = count1.n(); 486 487 // First clear the logged cards. 488 ClearLoggedCardTableEntryClosure clear; 489 dcqs.set_closure(&clear); 490 dcqs.apply_closure_to_all_completed_buffers(); 491 dcqs.iterate_closure_all_threads(false); 492 clear.print_histo(); 493 494 // Now ensure that there's no dirty cards. 495 CountNonCleanMemRegionClosure count2(this); 496 ct_bs->mod_card_iterate(&count2); 497 if (count2.n() != 0) { 498 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 499 count2.n(), orig_count); 500 } 501 guarantee(count2.n() == 0, "Card table should be clean."); 502 503 RedirtyLoggedCardTableEntryClosure redirty; 504 JavaThread::dirty_card_queue_set().set_closure(&redirty); 505 dcqs.apply_closure_to_all_completed_buffers(); 506 dcqs.iterate_closure_all_threads(false); 507 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 508 clear.calls(), orig_count); 509 guarantee(redirty.calls() == clear.calls(), 510 "Or else mechanism is broken."); 511 512 CountNonCleanMemRegionClosure count3(this); 513 ct_bs->mod_card_iterate(&count3); 514 if (count3.n() != orig_count) { 515 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 516 orig_count, count3.n()); 517 guarantee(count3.n() >= orig_count, "Should have restored them all."); 518 } 519 520 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 521 } 522 523 // Private class members. 524 525 G1CollectedHeap* G1CollectedHeap::_g1h; 526 527 // Private methods. 528 529 HeapRegion* 530 G1CollectedHeap::new_region_try_secondary_free_list() { 531 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 532 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 533 if (!_secondary_free_list.is_empty()) { 534 if (G1ConcRegionFreeingVerbose) { 535 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 536 "secondary_free_list has "SIZE_FORMAT" entries", 537 _secondary_free_list.length()); 538 } 539 // It looks as if there are free regions available on the 540 // secondary_free_list. Let's move them to the free_list and try 541 // again to allocate from it. 542 append_secondary_free_list(); 543 544 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 545 "empty we should have moved at least one entry to the free_list"); 546 HeapRegion* res = _free_list.remove_head(); 547 if (G1ConcRegionFreeingVerbose) { 548 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 549 "allocated "HR_FORMAT" from secondary_free_list", 550 HR_FORMAT_PARAMS(res)); 551 } 552 return res; 553 } 554 555 // Wait here until we get notifed either when (a) there are no 556 // more free regions coming or (b) some regions have been moved on 557 // the secondary_free_list. 558 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 559 } 560 561 if (G1ConcRegionFreeingVerbose) { 562 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 563 "could not allocate from secondary_free_list"); 564 } 565 return NULL; 566 } 567 568 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 569 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 570 "the only time we use this to allocate a humongous region is " 571 "when we are allocating a single humongous region"); 572 573 HeapRegion* res; 574 if (G1StressConcRegionFreeing) { 575 if (!_secondary_free_list.is_empty()) { 576 if (G1ConcRegionFreeingVerbose) { 577 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 578 "forced to look at the secondary_free_list"); 579 } 580 res = new_region_try_secondary_free_list(); 581 if (res != NULL) { 582 return res; 583 } 584 } 585 } 586 res = _free_list.remove_head_or_null(); 587 if (res == NULL) { 588 if (G1ConcRegionFreeingVerbose) { 589 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 590 "res == NULL, trying the secondary_free_list"); 591 } 592 res = new_region_try_secondary_free_list(); 593 } 594 if (res == NULL && do_expand) { 595 ergo_verbose1(ErgoHeapSizing, 596 "attempt heap expansion", 597 ergo_format_reason("region allocation request failed") 598 ergo_format_byte("allocation request"), 599 word_size * HeapWordSize); 600 if (expand(word_size * HeapWordSize)) { 601 // Even though the heap was expanded, it might not have reached 602 // the desired size. So, we cannot assume that the allocation 603 // will succeed. 604 res = _free_list.remove_head_or_null(); 605 } 606 } 607 return res; 608 } 609 610 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions, 611 size_t word_size) { 612 assert(isHumongous(word_size), "word_size should be humongous"); 613 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 614 615 size_t first = G1_NULL_HRS_INDEX; 616 if (num_regions == 1) { 617 // Only one region to allocate, no need to go through the slower 618 // path. The caller will attempt the expasion if this fails, so 619 // let's not try to expand here too. 620 HeapRegion* hr = new_region(word_size, false /* do_expand */); 621 if (hr != NULL) { 622 first = hr->hrs_index(); 623 } else { 624 first = G1_NULL_HRS_INDEX; 625 } 626 } else { 627 // We can't allocate humongous regions while cleanupComplete() is 628 // running, since some of the regions we find to be empty might not 629 // yet be added to the free list and it is not straightforward to 630 // know which list they are on so that we can remove them. Note 631 // that we only need to do this if we need to allocate more than 632 // one region to satisfy the current humongous allocation 633 // request. If we are only allocating one region we use the common 634 // region allocation code (see above). 635 wait_while_free_regions_coming(); 636 append_secondary_free_list_if_not_empty_with_lock(); 637 638 if (free_regions() >= num_regions) { 639 first = _hrs.find_contiguous(num_regions); 640 if (first != G1_NULL_HRS_INDEX) { 641 for (size_t i = first; i < first + num_regions; ++i) { 642 HeapRegion* hr = region_at(i); 643 assert(hr->is_empty(), "sanity"); 644 assert(is_on_master_free_list(hr), "sanity"); 645 hr->set_pending_removal(true); 646 } 647 _free_list.remove_all_pending(num_regions); 648 } 649 } 650 } 651 return first; 652 } 653 654 HeapWord* 655 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first, 656 size_t num_regions, 657 size_t word_size) { 658 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 659 assert(isHumongous(word_size), "word_size should be humongous"); 660 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 661 662 // Index of last region in the series + 1. 663 size_t last = first + num_regions; 664 665 // We need to initialize the region(s) we just discovered. This is 666 // a bit tricky given that it can happen concurrently with 667 // refinement threads refining cards on these regions and 668 // potentially wanting to refine the BOT as they are scanning 669 // those cards (this can happen shortly after a cleanup; see CR 670 // 6991377). So we have to set up the region(s) carefully and in 671 // a specific order. 672 673 // The word size sum of all the regions we will allocate. 674 size_t word_size_sum = num_regions * HeapRegion::GrainWords; 675 assert(word_size <= word_size_sum, "sanity"); 676 677 // This will be the "starts humongous" region. 678 HeapRegion* first_hr = region_at(first); 679 // The header of the new object will be placed at the bottom of 680 // the first region. 681 HeapWord* new_obj = first_hr->bottom(); 682 // This will be the new end of the first region in the series that 683 // should also match the end of the last region in the seriers. 684 HeapWord* new_end = new_obj + word_size_sum; 685 // This will be the new top of the first region that will reflect 686 // this allocation. 687 HeapWord* new_top = new_obj + word_size; 688 689 // First, we need to zero the header of the space that we will be 690 // allocating. When we update top further down, some refinement 691 // threads might try to scan the region. By zeroing the header we 692 // ensure that any thread that will try to scan the region will 693 // come across the zero klass word and bail out. 694 // 695 // NOTE: It would not have been correct to have used 696 // CollectedHeap::fill_with_object() and make the space look like 697 // an int array. The thread that is doing the allocation will 698 // later update the object header to a potentially different array 699 // type and, for a very short period of time, the klass and length 700 // fields will be inconsistent. This could cause a refinement 701 // thread to calculate the object size incorrectly. 702 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 703 704 // We will set up the first region as "starts humongous". This 705 // will also update the BOT covering all the regions to reflect 706 // that there is a single object that starts at the bottom of the 707 // first region. 708 first_hr->set_startsHumongous(new_top, new_end); 709 710 // Then, if there are any, we will set up the "continues 711 // humongous" regions. 712 HeapRegion* hr = NULL; 713 for (size_t i = first + 1; i < last; ++i) { 714 hr = region_at(i); 715 hr->set_continuesHumongous(first_hr); 716 } 717 // If we have "continues humongous" regions (hr != NULL), then the 718 // end of the last one should match new_end. 719 assert(hr == NULL || hr->end() == new_end, "sanity"); 720 721 // Up to this point no concurrent thread would have been able to 722 // do any scanning on any region in this series. All the top 723 // fields still point to bottom, so the intersection between 724 // [bottom,top] and [card_start,card_end] will be empty. Before we 725 // update the top fields, we'll do a storestore to make sure that 726 // no thread sees the update to top before the zeroing of the 727 // object header and the BOT initialization. 728 OrderAccess::storestore(); 729 730 // Now that the BOT and the object header have been initialized, 731 // we can update top of the "starts humongous" region. 732 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 733 "new_top should be in this region"); 734 first_hr->set_top(new_top); 735 if (_hr_printer.is_active()) { 736 HeapWord* bottom = first_hr->bottom(); 737 HeapWord* end = first_hr->orig_end(); 738 if ((first + 1) == last) { 739 // the series has a single humongous region 740 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 741 } else { 742 // the series has more than one humongous regions 743 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 744 } 745 } 746 747 // Now, we will update the top fields of the "continues humongous" 748 // regions. The reason we need to do this is that, otherwise, 749 // these regions would look empty and this will confuse parts of 750 // G1. For example, the code that looks for a consecutive number 751 // of empty regions will consider them empty and try to 752 // re-allocate them. We can extend is_empty() to also include 753 // !continuesHumongous(), but it is easier to just update the top 754 // fields here. The way we set top for all regions (i.e., top == 755 // end for all regions but the last one, top == new_top for the 756 // last one) is actually used when we will free up the humongous 757 // region in free_humongous_region(). 758 hr = NULL; 759 for (size_t i = first + 1; i < last; ++i) { 760 hr = region_at(i); 761 if ((i + 1) == last) { 762 // last continues humongous region 763 assert(hr->bottom() < new_top && new_top <= hr->end(), 764 "new_top should fall on this region"); 765 hr->set_top(new_top); 766 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 767 } else { 768 // not last one 769 assert(new_top > hr->end(), "new_top should be above this region"); 770 hr->set_top(hr->end()); 771 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 772 } 773 } 774 // If we have continues humongous regions (hr != NULL), then the 775 // end of the last one should match new_end and its top should 776 // match new_top. 777 assert(hr == NULL || 778 (hr->end() == new_end && hr->top() == new_top), "sanity"); 779 780 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 781 _summary_bytes_used += first_hr->used(); 782 _humongous_set.add(first_hr); 783 784 return new_obj; 785 } 786 787 // If could fit into free regions w/o expansion, try. 788 // Otherwise, if can expand, do so. 789 // Otherwise, if using ex regions might help, try with ex given back. 790 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 791 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 792 793 verify_region_sets_optional(); 794 795 size_t num_regions = 796 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords; 797 size_t x_size = expansion_regions(); 798 size_t fs = _hrs.free_suffix(); 799 size_t first = humongous_obj_allocate_find_first(num_regions, word_size); 800 if (first == G1_NULL_HRS_INDEX) { 801 // The only thing we can do now is attempt expansion. 802 if (fs + x_size >= num_regions) { 803 // If the number of regions we're trying to allocate for this 804 // object is at most the number of regions in the free suffix, 805 // then the call to humongous_obj_allocate_find_first() above 806 // should have succeeded and we wouldn't be here. 807 // 808 // We should only be trying to expand when the free suffix is 809 // not sufficient for the object _and_ we have some expansion 810 // room available. 811 assert(num_regions > fs, "earlier allocation should have succeeded"); 812 813 ergo_verbose1(ErgoHeapSizing, 814 "attempt heap expansion", 815 ergo_format_reason("humongous allocation request failed") 816 ergo_format_byte("allocation request"), 817 word_size * HeapWordSize); 818 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 819 // Even though the heap was expanded, it might not have 820 // reached the desired size. So, we cannot assume that the 821 // allocation will succeed. 822 first = humongous_obj_allocate_find_first(num_regions, word_size); 823 } 824 } 825 } 826 827 HeapWord* result = NULL; 828 if (first != G1_NULL_HRS_INDEX) { 829 result = 830 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 831 assert(result != NULL, "it should always return a valid result"); 832 833 // A successful humongous object allocation changes the used space 834 // information of the old generation so we need to recalculate the 835 // sizes and update the jstat counters here. 836 g1mm()->update_sizes(); 837 } 838 839 verify_region_sets_optional(); 840 841 return result; 842 } 843 844 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 845 assert_heap_not_locked_and_not_at_safepoint(); 846 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 847 848 unsigned int dummy_gc_count_before; 849 return attempt_allocation(word_size, &dummy_gc_count_before); 850 } 851 852 HeapWord* 853 G1CollectedHeap::mem_allocate(size_t word_size, 854 bool* gc_overhead_limit_was_exceeded) { 855 assert_heap_not_locked_and_not_at_safepoint(); 856 857 // Loop until the allocation is satisified, or unsatisfied after GC. 858 for (int try_count = 1; /* we'll return */; try_count += 1) { 859 unsigned int gc_count_before; 860 861 HeapWord* result = NULL; 862 if (!isHumongous(word_size)) { 863 result = attempt_allocation(word_size, &gc_count_before); 864 } else { 865 result = attempt_allocation_humongous(word_size, &gc_count_before); 866 } 867 if (result != NULL) { 868 return result; 869 } 870 871 // Create the garbage collection operation... 872 VM_G1CollectForAllocation op(gc_count_before, word_size); 873 // ...and get the VM thread to execute it. 874 VMThread::execute(&op); 875 876 if (op.prologue_succeeded() && op.pause_succeeded()) { 877 // If the operation was successful we'll return the result even 878 // if it is NULL. If the allocation attempt failed immediately 879 // after a Full GC, it's unlikely we'll be able to allocate now. 880 HeapWord* result = op.result(); 881 if (result != NULL && !isHumongous(word_size)) { 882 // Allocations that take place on VM operations do not do any 883 // card dirtying and we have to do it here. We only have to do 884 // this for non-humongous allocations, though. 885 dirty_young_block(result, word_size); 886 } 887 return result; 888 } else { 889 assert(op.result() == NULL, 890 "the result should be NULL if the VM op did not succeed"); 891 } 892 893 // Give a warning if we seem to be looping forever. 894 if ((QueuedAllocationWarningCount > 0) && 895 (try_count % QueuedAllocationWarningCount == 0)) { 896 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 897 } 898 } 899 900 ShouldNotReachHere(); 901 return NULL; 902 } 903 904 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 905 unsigned int *gc_count_before_ret) { 906 // Make sure you read the note in attempt_allocation_humongous(). 907 908 assert_heap_not_locked_and_not_at_safepoint(); 909 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 910 "be called for humongous allocation requests"); 911 912 // We should only get here after the first-level allocation attempt 913 // (attempt_allocation()) failed to allocate. 914 915 // We will loop until a) we manage to successfully perform the 916 // allocation or b) we successfully schedule a collection which 917 // fails to perform the allocation. b) is the only case when we'll 918 // return NULL. 919 HeapWord* result = NULL; 920 for (int try_count = 1; /* we'll return */; try_count += 1) { 921 bool should_try_gc; 922 unsigned int gc_count_before; 923 924 { 925 MutexLockerEx x(Heap_lock); 926 927 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 928 false /* bot_updates */); 929 if (result != NULL) { 930 return result; 931 } 932 933 // If we reach here, attempt_allocation_locked() above failed to 934 // allocate a new region. So the mutator alloc region should be NULL. 935 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 936 937 if (GC_locker::is_active_and_needs_gc()) { 938 if (g1_policy()->can_expand_young_list()) { 939 // No need for an ergo verbose message here, 940 // can_expand_young_list() does this when it returns true. 941 result = _mutator_alloc_region.attempt_allocation_force(word_size, 942 false /* bot_updates */); 943 if (result != NULL) { 944 return result; 945 } 946 } 947 should_try_gc = false; 948 } else { 949 // Read the GC count while still holding the Heap_lock. 950 gc_count_before = SharedHeap::heap()->total_collections(); 951 should_try_gc = true; 952 } 953 } 954 955 if (should_try_gc) { 956 bool succeeded; 957 result = do_collection_pause(word_size, gc_count_before, &succeeded); 958 if (result != NULL) { 959 assert(succeeded, "only way to get back a non-NULL result"); 960 return result; 961 } 962 963 if (succeeded) { 964 // If we get here we successfully scheduled a collection which 965 // failed to allocate. No point in trying to allocate 966 // further. We'll just return NULL. 967 MutexLockerEx x(Heap_lock); 968 *gc_count_before_ret = SharedHeap::heap()->total_collections(); 969 return NULL; 970 } 971 } else { 972 GC_locker::stall_until_clear(); 973 } 974 975 // We can reach here if we were unsuccessul in scheduling a 976 // collection (because another thread beat us to it) or if we were 977 // stalled due to the GC locker. In either can we should retry the 978 // allocation attempt in case another thread successfully 979 // performed a collection and reclaimed enough space. We do the 980 // first attempt (without holding the Heap_lock) here and the 981 // follow-on attempt will be at the start of the next loop 982 // iteration (after taking the Heap_lock). 983 result = _mutator_alloc_region.attempt_allocation(word_size, 984 false /* bot_updates */); 985 if (result != NULL ){ 986 return result; 987 } 988 989 // Give a warning if we seem to be looping forever. 990 if ((QueuedAllocationWarningCount > 0) && 991 (try_count % QueuedAllocationWarningCount == 0)) { 992 warning("G1CollectedHeap::attempt_allocation_slow() " 993 "retries %d times", try_count); 994 } 995 } 996 997 ShouldNotReachHere(); 998 return NULL; 999 } 1000 1001 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1002 unsigned int * gc_count_before_ret) { 1003 // The structure of this method has a lot of similarities to 1004 // attempt_allocation_slow(). The reason these two were not merged 1005 // into a single one is that such a method would require several "if 1006 // allocation is not humongous do this, otherwise do that" 1007 // conditional paths which would obscure its flow. In fact, an early 1008 // version of this code did use a unified method which was harder to 1009 // follow and, as a result, it had subtle bugs that were hard to 1010 // track down. So keeping these two methods separate allows each to 1011 // be more readable. It will be good to keep these two in sync as 1012 // much as possible. 1013 1014 assert_heap_not_locked_and_not_at_safepoint(); 1015 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1016 "should only be called for humongous allocations"); 1017 1018 // We will loop until a) we manage to successfully perform the 1019 // allocation or b) we successfully schedule a collection which 1020 // fails to perform the allocation. b) is the only case when we'll 1021 // return NULL. 1022 HeapWord* result = NULL; 1023 for (int try_count = 1; /* we'll return */; try_count += 1) { 1024 bool should_try_gc; 1025 unsigned int gc_count_before; 1026 1027 { 1028 MutexLockerEx x(Heap_lock); 1029 1030 // Given that humongous objects are not allocated in young 1031 // regions, we'll first try to do the allocation without doing a 1032 // collection hoping that there's enough space in the heap. 1033 result = humongous_obj_allocate(word_size); 1034 if (result != NULL) { 1035 return result; 1036 } 1037 1038 if (GC_locker::is_active_and_needs_gc()) { 1039 should_try_gc = false; 1040 } else { 1041 // Read the GC count while still holding the Heap_lock. 1042 gc_count_before = SharedHeap::heap()->total_collections(); 1043 should_try_gc = true; 1044 } 1045 } 1046 1047 if (should_try_gc) { 1048 // If we failed to allocate the humongous object, we should try to 1049 // do a collection pause (if we're allowed) in case it reclaims 1050 // enough space for the allocation to succeed after the pause. 1051 1052 bool succeeded; 1053 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1054 if (result != NULL) { 1055 assert(succeeded, "only way to get back a non-NULL result"); 1056 return result; 1057 } 1058 1059 if (succeeded) { 1060 // If we get here we successfully scheduled a collection which 1061 // failed to allocate. No point in trying to allocate 1062 // further. We'll just return NULL. 1063 MutexLockerEx x(Heap_lock); 1064 *gc_count_before_ret = SharedHeap::heap()->total_collections(); 1065 return NULL; 1066 } 1067 } else { 1068 GC_locker::stall_until_clear(); 1069 } 1070 1071 // We can reach here if we were unsuccessul in scheduling a 1072 // collection (because another thread beat us to it) or if we were 1073 // stalled due to the GC locker. In either can we should retry the 1074 // allocation attempt in case another thread successfully 1075 // performed a collection and reclaimed enough space. Give a 1076 // warning if we seem to be looping forever. 1077 1078 if ((QueuedAllocationWarningCount > 0) && 1079 (try_count % QueuedAllocationWarningCount == 0)) { 1080 warning("G1CollectedHeap::attempt_allocation_humongous() " 1081 "retries %d times", try_count); 1082 } 1083 } 1084 1085 ShouldNotReachHere(); 1086 return NULL; 1087 } 1088 1089 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1090 bool expect_null_mutator_alloc_region) { 1091 assert_at_safepoint(true /* should_be_vm_thread */); 1092 assert(_mutator_alloc_region.get() == NULL || 1093 !expect_null_mutator_alloc_region, 1094 "the current alloc region was unexpectedly found to be non-NULL"); 1095 1096 if (!isHumongous(word_size)) { 1097 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1098 false /* bot_updates */); 1099 } else { 1100 return humongous_obj_allocate(word_size); 1101 } 1102 1103 ShouldNotReachHere(); 1104 } 1105 1106 class PostMCRemSetClearClosure: public HeapRegionClosure { 1107 ModRefBarrierSet* _mr_bs; 1108 public: 1109 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1110 bool doHeapRegion(HeapRegion* r) { 1111 r->reset_gc_time_stamp(); 1112 if (r->continuesHumongous()) 1113 return false; 1114 HeapRegionRemSet* hrrs = r->rem_set(); 1115 if (hrrs != NULL) hrrs->clear(); 1116 // You might think here that we could clear just the cards 1117 // corresponding to the used region. But no: if we leave a dirty card 1118 // in a region we might allocate into, then it would prevent that card 1119 // from being enqueued, and cause it to be missed. 1120 // Re: the performance cost: we shouldn't be doing full GC anyway! 1121 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1122 return false; 1123 } 1124 }; 1125 1126 1127 class PostMCRemSetInvalidateClosure: public HeapRegionClosure { 1128 ModRefBarrierSet* _mr_bs; 1129 public: 1130 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1131 bool doHeapRegion(HeapRegion* r) { 1132 if (r->continuesHumongous()) return false; 1133 if (r->used_region().word_size() != 0) { 1134 _mr_bs->invalidate(r->used_region(), true /*whole heap*/); 1135 } 1136 return false; 1137 } 1138 }; 1139 1140 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1141 G1CollectedHeap* _g1h; 1142 UpdateRSOopClosure _cl; 1143 int _worker_i; 1144 public: 1145 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1146 _cl(g1->g1_rem_set(), worker_i), 1147 _worker_i(worker_i), 1148 _g1h(g1) 1149 { } 1150 1151 bool doHeapRegion(HeapRegion* r) { 1152 if (!r->continuesHumongous()) { 1153 _cl.set_from(r); 1154 r->oop_iterate(&_cl); 1155 } 1156 return false; 1157 } 1158 }; 1159 1160 class ParRebuildRSTask: public AbstractGangTask { 1161 G1CollectedHeap* _g1; 1162 public: 1163 ParRebuildRSTask(G1CollectedHeap* g1) 1164 : AbstractGangTask("ParRebuildRSTask"), 1165 _g1(g1) 1166 { } 1167 1168 void work(int i) { 1169 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i); 1170 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i, 1171 _g1->workers()->active_workers(), 1172 HeapRegion::RebuildRSClaimValue); 1173 } 1174 }; 1175 1176 class PostCompactionPrinterClosure: public HeapRegionClosure { 1177 private: 1178 G1HRPrinter* _hr_printer; 1179 public: 1180 bool doHeapRegion(HeapRegion* hr) { 1181 assert(!hr->is_young(), "not expecting to find young regions"); 1182 // We only generate output for non-empty regions. 1183 if (!hr->is_empty()) { 1184 if (!hr->isHumongous()) { 1185 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1186 } else if (hr->startsHumongous()) { 1187 if (hr->capacity() == HeapRegion::GrainBytes) { 1188 // single humongous region 1189 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1190 } else { 1191 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1192 } 1193 } else { 1194 assert(hr->continuesHumongous(), "only way to get here"); 1195 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1196 } 1197 } 1198 return false; 1199 } 1200 1201 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1202 : _hr_printer(hr_printer) { } 1203 }; 1204 1205 bool G1CollectedHeap::do_collection(bool explicit_gc, 1206 bool clear_all_soft_refs, 1207 size_t word_size) { 1208 assert_at_safepoint(true /* should_be_vm_thread */); 1209 1210 if (GC_locker::check_active_before_gc()) { 1211 return false; 1212 } 1213 1214 SvcGCMarker sgcm(SvcGCMarker::FULL); 1215 ResourceMark rm; 1216 1217 if (PrintHeapAtGC) { 1218 Universe::print_heap_before_gc(); 1219 } 1220 1221 HRSPhaseSetter x(HRSPhaseFullGC); 1222 verify_region_sets_optional(); 1223 1224 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1225 collector_policy()->should_clear_all_soft_refs(); 1226 1227 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1228 1229 { 1230 IsGCActiveMark x; 1231 1232 // Timing 1233 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc); 1234 assert(!system_gc || explicit_gc, "invariant"); 1235 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 1236 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 1237 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC", 1238 PrintGC, true, gclog_or_tty); 1239 1240 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1241 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1242 1243 double start = os::elapsedTime(); 1244 g1_policy()->record_full_collection_start(); 1245 1246 wait_while_free_regions_coming(); 1247 append_secondary_free_list_if_not_empty_with_lock(); 1248 1249 gc_prologue(true); 1250 increment_total_collections(true /* full gc */); 1251 1252 size_t g1h_prev_used = used(); 1253 assert(used() == recalculate_used(), "Should be equal"); 1254 1255 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 1256 HandleMark hm; // Discard invalid handles created during verification 1257 gclog_or_tty->print(" VerifyBeforeGC:"); 1258 prepare_for_verify(); 1259 Universe::verify(/* allow dirty */ true, 1260 /* silent */ false, 1261 /* option */ VerifyOption_G1UsePrevMarking); 1262 1263 } 1264 pre_full_gc_dump(); 1265 1266 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1267 1268 // Disable discovery and empty the discovered lists 1269 // for the CM ref processor. 1270 ref_processor_cm()->disable_discovery(); 1271 ref_processor_cm()->abandon_partial_discovery(); 1272 ref_processor_cm()->verify_no_references_recorded(); 1273 1274 // Abandon current iterations of concurrent marking and concurrent 1275 // refinement, if any are in progress. 1276 concurrent_mark()->abort(); 1277 1278 // Make sure we'll choose a new allocation region afterwards. 1279 release_mutator_alloc_region(); 1280 abandon_gc_alloc_regions(); 1281 g1_rem_set()->cleanupHRRS(); 1282 1283 // We should call this after we retire any currently active alloc 1284 // regions so that all the ALLOC / RETIRE events are generated 1285 // before the start GC event. 1286 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1287 1288 // We may have added regions to the current incremental collection 1289 // set between the last GC or pause and now. We need to clear the 1290 // incremental collection set and then start rebuilding it afresh 1291 // after this full GC. 1292 abandon_collection_set(g1_policy()->inc_cset_head()); 1293 g1_policy()->clear_incremental_cset(); 1294 g1_policy()->stop_incremental_cset_building(); 1295 1296 tear_down_region_sets(false /* free_list_only */); 1297 g1_policy()->set_full_young_gcs(true); 1298 1299 // See the comments in g1CollectedHeap.hpp and 1300 // G1CollectedHeap::ref_processing_init() about 1301 // how reference processing currently works in G1. 1302 1303 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1304 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1305 1306 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1307 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1308 1309 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1310 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1311 1312 // Do collection work 1313 { 1314 HandleMark hm; // Discard invalid handles created during gc 1315 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1316 } 1317 1318 assert(free_regions() == 0, "we should not have added any free regions"); 1319 rebuild_region_sets(false /* free_list_only */); 1320 1321 // Enqueue any discovered reference objects that have 1322 // not been removed from the discovered lists. 1323 ref_processor_stw()->enqueue_discovered_references(); 1324 1325 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1326 1327 MemoryService::track_memory_usage(); 1328 1329 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 1330 HandleMark hm; // Discard invalid handles created during verification 1331 gclog_or_tty->print(" VerifyAfterGC:"); 1332 prepare_for_verify(); 1333 Universe::verify(/* allow dirty */ false, 1334 /* silent */ false, 1335 /* option */ VerifyOption_G1UsePrevMarking); 1336 1337 } 1338 1339 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1340 ref_processor_stw()->verify_no_references_recorded(); 1341 1342 // Note: since we've just done a full GC, concurrent 1343 // marking is no longer active. Therefore we need not 1344 // re-enable reference discovery for the CM ref processor. 1345 // That will be done at the start of the next marking cycle. 1346 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1347 ref_processor_cm()->verify_no_references_recorded(); 1348 1349 reset_gc_time_stamp(); 1350 // Since everything potentially moved, we will clear all remembered 1351 // sets, and clear all cards. Later we will rebuild remebered 1352 // sets. We will also reset the GC time stamps of the regions. 1353 PostMCRemSetClearClosure rs_clear(mr_bs()); 1354 heap_region_iterate(&rs_clear); 1355 1356 // Resize the heap if necessary. 1357 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1358 1359 if (_hr_printer.is_active()) { 1360 // We should do this after we potentially resize the heap so 1361 // that all the COMMIT / UNCOMMIT events are generated before 1362 // the end GC event. 1363 1364 PostCompactionPrinterClosure cl(hr_printer()); 1365 heap_region_iterate(&cl); 1366 1367 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1368 } 1369 1370 if (_cg1r->use_cache()) { 1371 _cg1r->clear_and_record_card_counts(); 1372 _cg1r->clear_hot_cache(); 1373 } 1374 1375 // Rebuild remembered sets of all regions. 1376 if (G1CollectedHeap::use_parallel_gc_threads()) { 1377 int n_workers = 1378 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1379 workers()->active_workers(), 1380 Threads::number_of_non_daemon_threads()); 1381 assert(UseDynamicNumberOfGCThreads || 1382 n_workers == workers()->total_workers(), 1383 "If not dynamic should be using all the workers"); 1384 workers()->set_active_workers(n_workers); 1385 // Set parallel threads in the heap (_n_par_threads) only 1386 // before a parallel phase and always reset it to 0 after 1387 // the phase so that the number of parallel threads does 1388 // no get carried forward to a serial phase where there 1389 // may be code that is "possibly_parallel". 1390 set_par_threads(n_workers); 1391 1392 ParRebuildRSTask rebuild_rs_task(this); 1393 assert(check_heap_region_claim_values( 1394 HeapRegion::InitialClaimValue), "sanity check"); 1395 assert(UseDynamicNumberOfGCThreads || 1396 workers()->active_workers() == workers()->total_workers(), 1397 "Unless dynamic should use total workers"); 1398 // Use the most recent number of active workers 1399 assert(workers()->active_workers() > 0, 1400 "Active workers not properly set"); 1401 set_par_threads(workers()->active_workers()); 1402 workers()->run_task(&rebuild_rs_task); 1403 set_par_threads(0); 1404 assert(check_heap_region_claim_values( 1405 HeapRegion::RebuildRSClaimValue), "sanity check"); 1406 reset_heap_region_claim_values(); 1407 } else { 1408 RebuildRSOutOfRegionClosure rebuild_rs(this); 1409 heap_region_iterate(&rebuild_rs); 1410 } 1411 1412 if (PrintGC) { 1413 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1414 } 1415 1416 if (true) { // FIXME 1417 // Ask the permanent generation to adjust size for full collections 1418 perm()->compute_new_size(); 1419 } 1420 1421 // Start a new incremental collection set for the next pause 1422 assert(g1_policy()->collection_set() == NULL, "must be"); 1423 g1_policy()->start_incremental_cset_building(); 1424 1425 // Clear the _cset_fast_test bitmap in anticipation of adding 1426 // regions to the incremental collection set for the next 1427 // evacuation pause. 1428 clear_cset_fast_test(); 1429 1430 init_mutator_alloc_region(); 1431 1432 double end = os::elapsedTime(); 1433 g1_policy()->record_full_collection_end(); 1434 1435 #ifdef TRACESPINNING 1436 ParallelTaskTerminator::print_termination_counts(); 1437 #endif 1438 1439 gc_epilogue(true); 1440 1441 // Discard all rset updates 1442 JavaThread::dirty_card_queue_set().abandon_logs(); 1443 assert(!G1DeferredRSUpdate 1444 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1445 } 1446 1447 _young_list->reset_sampled_info(); 1448 // At this point there should be no regions in the 1449 // entire heap tagged as young. 1450 assert( check_young_list_empty(true /* check_heap */), 1451 "young list should be empty at this point"); 1452 1453 // Update the number of full collections that have been completed. 1454 increment_full_collections_completed(false /* concurrent */); 1455 1456 _hrs.verify_optional(); 1457 verify_region_sets_optional(); 1458 1459 if (PrintHeapAtGC) { 1460 Universe::print_heap_after_gc(); 1461 } 1462 g1mm()->update_sizes(); 1463 post_full_gc_dump(); 1464 1465 return true; 1466 } 1467 1468 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1469 // do_collection() will return whether it succeeded in performing 1470 // the GC. Currently, there is no facility on the 1471 // do_full_collection() API to notify the caller than the collection 1472 // did not succeed (e.g., because it was locked out by the GC 1473 // locker). So, right now, we'll ignore the return value. 1474 bool dummy = do_collection(true, /* explicit_gc */ 1475 clear_all_soft_refs, 1476 0 /* word_size */); 1477 } 1478 1479 // This code is mostly copied from TenuredGeneration. 1480 void 1481 G1CollectedHeap:: 1482 resize_if_necessary_after_full_collection(size_t word_size) { 1483 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1484 1485 // Include the current allocation, if any, and bytes that will be 1486 // pre-allocated to support collections, as "used". 1487 const size_t used_after_gc = used(); 1488 const size_t capacity_after_gc = capacity(); 1489 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1490 1491 // This is enforced in arguments.cpp. 1492 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1493 "otherwise the code below doesn't make sense"); 1494 1495 // We don't have floating point command-line arguments 1496 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1497 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1498 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1499 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1500 1501 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1502 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1503 1504 // We have to be careful here as these two calculations can overflow 1505 // 32-bit size_t's. 1506 double used_after_gc_d = (double) used_after_gc; 1507 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1508 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1509 1510 // Let's make sure that they are both under the max heap size, which 1511 // by default will make them fit into a size_t. 1512 double desired_capacity_upper_bound = (double) max_heap_size; 1513 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1514 desired_capacity_upper_bound); 1515 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1516 desired_capacity_upper_bound); 1517 1518 // We can now safely turn them into size_t's. 1519 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1520 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1521 1522 // This assert only makes sense here, before we adjust them 1523 // with respect to the min and max heap size. 1524 assert(minimum_desired_capacity <= maximum_desired_capacity, 1525 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1526 "maximum_desired_capacity = "SIZE_FORMAT, 1527 minimum_desired_capacity, maximum_desired_capacity)); 1528 1529 // Should not be greater than the heap max size. No need to adjust 1530 // it with respect to the heap min size as it's a lower bound (i.e., 1531 // we'll try to make the capacity larger than it, not smaller). 1532 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1533 // Should not be less than the heap min size. No need to adjust it 1534 // with respect to the heap max size as it's an upper bound (i.e., 1535 // we'll try to make the capacity smaller than it, not greater). 1536 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1537 1538 if (capacity_after_gc < minimum_desired_capacity) { 1539 // Don't expand unless it's significant 1540 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1541 ergo_verbose4(ErgoHeapSizing, 1542 "attempt heap expansion", 1543 ergo_format_reason("capacity lower than " 1544 "min desired capacity after Full GC") 1545 ergo_format_byte("capacity") 1546 ergo_format_byte("occupancy") 1547 ergo_format_byte_perc("min desired capacity"), 1548 capacity_after_gc, used_after_gc, 1549 minimum_desired_capacity, (double) MinHeapFreeRatio); 1550 expand(expand_bytes); 1551 1552 // No expansion, now see if we want to shrink 1553 } else if (capacity_after_gc > maximum_desired_capacity) { 1554 // Capacity too large, compute shrinking size 1555 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1556 ergo_verbose4(ErgoHeapSizing, 1557 "attempt heap shrinking", 1558 ergo_format_reason("capacity higher than " 1559 "max desired capacity after Full GC") 1560 ergo_format_byte("capacity") 1561 ergo_format_byte("occupancy") 1562 ergo_format_byte_perc("max desired capacity"), 1563 capacity_after_gc, used_after_gc, 1564 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1565 shrink(shrink_bytes); 1566 } 1567 } 1568 1569 1570 HeapWord* 1571 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1572 bool* succeeded) { 1573 assert_at_safepoint(true /* should_be_vm_thread */); 1574 1575 *succeeded = true; 1576 // Let's attempt the allocation first. 1577 HeapWord* result = 1578 attempt_allocation_at_safepoint(word_size, 1579 false /* expect_null_mutator_alloc_region */); 1580 if (result != NULL) { 1581 assert(*succeeded, "sanity"); 1582 return result; 1583 } 1584 1585 // In a G1 heap, we're supposed to keep allocation from failing by 1586 // incremental pauses. Therefore, at least for now, we'll favor 1587 // expansion over collection. (This might change in the future if we can 1588 // do something smarter than full collection to satisfy a failed alloc.) 1589 result = expand_and_allocate(word_size); 1590 if (result != NULL) { 1591 assert(*succeeded, "sanity"); 1592 return result; 1593 } 1594 1595 // Expansion didn't work, we'll try to do a Full GC. 1596 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1597 false, /* clear_all_soft_refs */ 1598 word_size); 1599 if (!gc_succeeded) { 1600 *succeeded = false; 1601 return NULL; 1602 } 1603 1604 // Retry the allocation 1605 result = attempt_allocation_at_safepoint(word_size, 1606 true /* expect_null_mutator_alloc_region */); 1607 if (result != NULL) { 1608 assert(*succeeded, "sanity"); 1609 return result; 1610 } 1611 1612 // Then, try a Full GC that will collect all soft references. 1613 gc_succeeded = do_collection(false, /* explicit_gc */ 1614 true, /* clear_all_soft_refs */ 1615 word_size); 1616 if (!gc_succeeded) { 1617 *succeeded = false; 1618 return NULL; 1619 } 1620 1621 // Retry the allocation once more 1622 result = attempt_allocation_at_safepoint(word_size, 1623 true /* expect_null_mutator_alloc_region */); 1624 if (result != NULL) { 1625 assert(*succeeded, "sanity"); 1626 return result; 1627 } 1628 1629 assert(!collector_policy()->should_clear_all_soft_refs(), 1630 "Flag should have been handled and cleared prior to this point"); 1631 1632 // What else? We might try synchronous finalization later. If the total 1633 // space available is large enough for the allocation, then a more 1634 // complete compaction phase than we've tried so far might be 1635 // appropriate. 1636 assert(*succeeded, "sanity"); 1637 return NULL; 1638 } 1639 1640 // Attempting to expand the heap sufficiently 1641 // to support an allocation of the given "word_size". If 1642 // successful, perform the allocation and return the address of the 1643 // allocated block, or else "NULL". 1644 1645 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1646 assert_at_safepoint(true /* should_be_vm_thread */); 1647 1648 verify_region_sets_optional(); 1649 1650 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1651 ergo_verbose1(ErgoHeapSizing, 1652 "attempt heap expansion", 1653 ergo_format_reason("allocation request failed") 1654 ergo_format_byte("allocation request"), 1655 word_size * HeapWordSize); 1656 if (expand(expand_bytes)) { 1657 _hrs.verify_optional(); 1658 verify_region_sets_optional(); 1659 return attempt_allocation_at_safepoint(word_size, 1660 false /* expect_null_mutator_alloc_region */); 1661 } 1662 return NULL; 1663 } 1664 1665 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1666 HeapWord* new_end) { 1667 assert(old_end != new_end, "don't call this otherwise"); 1668 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1669 1670 // Update the committed mem region. 1671 _g1_committed.set_end(new_end); 1672 // Tell the card table about the update. 1673 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1674 // Tell the BOT about the update. 1675 _bot_shared->resize(_g1_committed.word_size()); 1676 } 1677 1678 bool G1CollectedHeap::expand(size_t expand_bytes) { 1679 size_t old_mem_size = _g1_storage.committed_size(); 1680 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1681 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1682 HeapRegion::GrainBytes); 1683 ergo_verbose2(ErgoHeapSizing, 1684 "expand the heap", 1685 ergo_format_byte("requested expansion amount") 1686 ergo_format_byte("attempted expansion amount"), 1687 expand_bytes, aligned_expand_bytes); 1688 1689 // First commit the memory. 1690 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1691 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1692 if (successful) { 1693 // Then propagate this update to the necessary data structures. 1694 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1695 update_committed_space(old_end, new_end); 1696 1697 FreeRegionList expansion_list("Local Expansion List"); 1698 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1699 assert(mr.start() == old_end, "post-condition"); 1700 // mr might be a smaller region than what was requested if 1701 // expand_by() was unable to allocate the HeapRegion instances 1702 assert(mr.end() <= new_end, "post-condition"); 1703 1704 size_t actual_expand_bytes = mr.byte_size(); 1705 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1706 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1707 "post-condition"); 1708 if (actual_expand_bytes < aligned_expand_bytes) { 1709 // We could not expand _hrs to the desired size. In this case we 1710 // need to shrink the committed space accordingly. 1711 assert(mr.end() < new_end, "invariant"); 1712 1713 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1714 // First uncommit the memory. 1715 _g1_storage.shrink_by(diff_bytes); 1716 // Then propagate this update to the necessary data structures. 1717 update_committed_space(new_end, mr.end()); 1718 } 1719 _free_list.add_as_tail(&expansion_list); 1720 1721 if (_hr_printer.is_active()) { 1722 HeapWord* curr = mr.start(); 1723 while (curr < mr.end()) { 1724 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1725 _hr_printer.commit(curr, curr_end); 1726 curr = curr_end; 1727 } 1728 assert(curr == mr.end(), "post-condition"); 1729 } 1730 g1_policy()->record_new_heap_size(n_regions()); 1731 } else { 1732 ergo_verbose0(ErgoHeapSizing, 1733 "did not expand the heap", 1734 ergo_format_reason("heap expansion operation failed")); 1735 // The expansion of the virtual storage space was unsuccessful. 1736 // Let's see if it was because we ran out of swap. 1737 if (G1ExitOnExpansionFailure && 1738 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1739 // We had head room... 1740 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1741 } 1742 } 1743 return successful; 1744 } 1745 1746 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1747 size_t old_mem_size = _g1_storage.committed_size(); 1748 size_t aligned_shrink_bytes = 1749 ReservedSpace::page_align_size_down(shrink_bytes); 1750 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1751 HeapRegion::GrainBytes); 1752 size_t num_regions_deleted = 0; 1753 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted); 1754 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1755 assert(mr.end() == old_end, "post-condition"); 1756 1757 ergo_verbose3(ErgoHeapSizing, 1758 "shrink the heap", 1759 ergo_format_byte("requested shrinking amount") 1760 ergo_format_byte("aligned shrinking amount") 1761 ergo_format_byte("attempted shrinking amount"), 1762 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1763 if (mr.byte_size() > 0) { 1764 if (_hr_printer.is_active()) { 1765 HeapWord* curr = mr.end(); 1766 while (curr > mr.start()) { 1767 HeapWord* curr_end = curr; 1768 curr -= HeapRegion::GrainWords; 1769 _hr_printer.uncommit(curr, curr_end); 1770 } 1771 assert(curr == mr.start(), "post-condition"); 1772 } 1773 1774 _g1_storage.shrink_by(mr.byte_size()); 1775 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1776 assert(mr.start() == new_end, "post-condition"); 1777 1778 _expansion_regions += num_regions_deleted; 1779 update_committed_space(old_end, new_end); 1780 HeapRegionRemSet::shrink_heap(n_regions()); 1781 g1_policy()->record_new_heap_size(n_regions()); 1782 } else { 1783 ergo_verbose0(ErgoHeapSizing, 1784 "did not shrink the heap", 1785 ergo_format_reason("heap shrinking operation failed")); 1786 } 1787 } 1788 1789 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1790 verify_region_sets_optional(); 1791 1792 // We should only reach here at the end of a Full GC which means we 1793 // should not not be holding to any GC alloc regions. The method 1794 // below will make sure of that and do any remaining clean up. 1795 abandon_gc_alloc_regions(); 1796 1797 // Instead of tearing down / rebuilding the free lists here, we 1798 // could instead use the remove_all_pending() method on free_list to 1799 // remove only the ones that we need to remove. 1800 tear_down_region_sets(true /* free_list_only */); 1801 shrink_helper(shrink_bytes); 1802 rebuild_region_sets(true /* free_list_only */); 1803 1804 _hrs.verify_optional(); 1805 verify_region_sets_optional(); 1806 } 1807 1808 // Public methods. 1809 1810 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1811 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1812 #endif // _MSC_VER 1813 1814 1815 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1816 SharedHeap(policy_), 1817 _g1_policy(policy_), 1818 _dirty_card_queue_set(false), 1819 _into_cset_dirty_card_queue_set(false), 1820 _is_alive_closure_cm(this), 1821 _is_alive_closure_stw(this), 1822 _ref_processor_cm(NULL), 1823 _ref_processor_stw(NULL), 1824 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1825 _bot_shared(NULL), 1826 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL), 1827 _evac_failure_scan_stack(NULL) , 1828 _mark_in_progress(false), 1829 _cg1r(NULL), _summary_bytes_used(0), 1830 _g1mm(NULL), 1831 _refine_cte_cl(NULL), 1832 _full_collection(false), 1833 _free_list("Master Free List"), 1834 _secondary_free_list("Secondary Free List"), 1835 _old_set("Old Set"), 1836 _humongous_set("Master Humongous Set"), 1837 _free_regions_coming(false), 1838 _young_list(new YoungList(this)), 1839 _gc_time_stamp(0), 1840 _retained_old_gc_alloc_region(NULL), 1841 _surviving_young_words(NULL), 1842 _full_collections_completed(0), 1843 _in_cset_fast_test(NULL), 1844 _in_cset_fast_test_base(NULL), 1845 _dirty_cards_region_list(NULL), 1846 _worker_cset_start_region(NULL), 1847 _worker_cset_start_region_stamp(NULL) { 1848 _g1h = this; // To catch bugs. 1849 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1850 vm_exit_during_initialization("Failed necessary allocation."); 1851 } 1852 1853 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1854 1855 int n_queues = MAX2((int)ParallelGCThreads, 1); 1856 _task_queues = new RefToScanQueueSet(n_queues); 1857 1858 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1859 assert(n_rem_sets > 0, "Invariant."); 1860 1861 HeapRegionRemSetIterator** iter_arr = 1862 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues); 1863 for (int i = 0; i < n_queues; i++) { 1864 iter_arr[i] = new HeapRegionRemSetIterator(); 1865 } 1866 _rem_set_iterator = iter_arr; 1867 1868 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues); 1869 _worker_cset_start_region_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues); 1870 1871 for (int i = 0; i < n_queues; i++) { 1872 RefToScanQueue* q = new RefToScanQueue(); 1873 q->initialize(); 1874 _task_queues->register_queue(i, q); 1875 1876 _worker_cset_start_region[i] = NULL; 1877 _worker_cset_start_region_stamp[i] = 0; 1878 } 1879 1880 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1881 } 1882 1883 jint G1CollectedHeap::initialize() { 1884 CollectedHeap::pre_initialize(); 1885 os::enable_vtime(); 1886 1887 // Necessary to satisfy locking discipline assertions. 1888 1889 MutexLocker x(Heap_lock); 1890 1891 // We have to initialize the printer before committing the heap, as 1892 // it will be used then. 1893 _hr_printer.set_active(G1PrintHeapRegions); 1894 1895 // While there are no constraints in the GC code that HeapWordSize 1896 // be any particular value, there are multiple other areas in the 1897 // system which believe this to be true (e.g. oop->object_size in some 1898 // cases incorrectly returns the size in wordSize units rather than 1899 // HeapWordSize). 1900 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1901 1902 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1903 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1904 1905 // Ensure that the sizes are properly aligned. 1906 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1907 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1908 1909 _cg1r = new ConcurrentG1Refine(); 1910 1911 // Reserve the maximum. 1912 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1913 // Includes the perm-gen. 1914 1915 // When compressed oops are enabled, the preferred heap base 1916 // is calculated by subtracting the requested size from the 1917 // 32Gb boundary and using the result as the base address for 1918 // heap reservation. If the requested size is not aligned to 1919 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1920 // into the ReservedHeapSpace constructor) then the actual 1921 // base of the reserved heap may end up differing from the 1922 // address that was requested (i.e. the preferred heap base). 1923 // If this happens then we could end up using a non-optimal 1924 // compressed oops mode. 1925 1926 // Since max_byte_size is aligned to the size of a heap region (checked 1927 // above), we also need to align the perm gen size as it might not be. 1928 const size_t total_reserved = max_byte_size + 1929 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1930 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1931 1932 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1933 1934 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1935 UseLargePages, addr); 1936 1937 if (UseCompressedOops) { 1938 if (addr != NULL && !heap_rs.is_reserved()) { 1939 // Failed to reserve at specified address - the requested memory 1940 // region is taken already, for example, by 'java' launcher. 1941 // Try again to reserver heap higher. 1942 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 1943 1944 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 1945 UseLargePages, addr); 1946 1947 if (addr != NULL && !heap_rs0.is_reserved()) { 1948 // Failed to reserve at specified address again - give up. 1949 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 1950 assert(addr == NULL, ""); 1951 1952 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 1953 UseLargePages, addr); 1954 heap_rs = heap_rs1; 1955 } else { 1956 heap_rs = heap_rs0; 1957 } 1958 } 1959 } 1960 1961 if (!heap_rs.is_reserved()) { 1962 vm_exit_during_initialization("Could not reserve enough space for object heap"); 1963 return JNI_ENOMEM; 1964 } 1965 1966 // It is important to do this in a way such that concurrent readers can't 1967 // temporarily think somethings in the heap. (I've actually seen this 1968 // happen in asserts: DLD.) 1969 _reserved.set_word_size(0); 1970 _reserved.set_start((HeapWord*)heap_rs.base()); 1971 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 1972 1973 _expansion_regions = max_byte_size/HeapRegion::GrainBytes; 1974 1975 // Create the gen rem set (and barrier set) for the entire reserved region. 1976 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 1977 set_barrier_set(rem_set()->bs()); 1978 if (barrier_set()->is_a(BarrierSet::ModRef)) { 1979 _mr_bs = (ModRefBarrierSet*)_barrier_set; 1980 } else { 1981 vm_exit_during_initialization("G1 requires a mod ref bs."); 1982 return JNI_ENOMEM; 1983 } 1984 1985 // Also create a G1 rem set. 1986 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 1987 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 1988 } else { 1989 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 1990 return JNI_ENOMEM; 1991 } 1992 1993 // Carve out the G1 part of the heap. 1994 1995 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 1996 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 1997 g1_rs.size()/HeapWordSize); 1998 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 1999 2000 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2001 2002 _g1_storage.initialize(g1_rs, 0); 2003 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2004 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2005 (HeapWord*) _g1_reserved.end(), 2006 _expansion_regions); 2007 2008 // 6843694 - ensure that the maximum region index can fit 2009 // in the remembered set structures. 2010 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2011 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2012 2013 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2014 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2015 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2016 "too many cards per region"); 2017 2018 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2019 2020 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2021 heap_word_size(init_byte_size)); 2022 2023 _g1h = this; 2024 2025 _in_cset_fast_test_length = max_regions(); 2026 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length); 2027 2028 // We're biasing _in_cset_fast_test to avoid subtracting the 2029 // beginning of the heap every time we want to index; basically 2030 // it's the same with what we do with the card table. 2031 _in_cset_fast_test = _in_cset_fast_test_base - 2032 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2033 2034 // Clear the _cset_fast_test bitmap in anticipation of adding 2035 // regions to the incremental collection set for the first 2036 // evacuation pause. 2037 clear_cset_fast_test(); 2038 2039 // Create the ConcurrentMark data structure and thread. 2040 // (Must do this late, so that "max_regions" is defined.) 2041 _cm = new ConcurrentMark(heap_rs, (int) max_regions()); 2042 _cmThread = _cm->cmThread(); 2043 2044 // Initialize the from_card cache structure of HeapRegionRemSet. 2045 HeapRegionRemSet::init_heap(max_regions()); 2046 2047 // Now expand into the initial heap size. 2048 if (!expand(init_byte_size)) { 2049 vm_exit_during_initialization("Failed to allocate initial heap."); 2050 return JNI_ENOMEM; 2051 } 2052 2053 // Perform any initialization actions delegated to the policy. 2054 g1_policy()->init(); 2055 2056 _refine_cte_cl = 2057 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2058 g1_rem_set(), 2059 concurrent_g1_refine()); 2060 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2061 2062 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2063 SATB_Q_FL_lock, 2064 G1SATBProcessCompletedThreshold, 2065 Shared_SATB_Q_lock); 2066 2067 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2068 DirtyCardQ_FL_lock, 2069 concurrent_g1_refine()->yellow_zone(), 2070 concurrent_g1_refine()->red_zone(), 2071 Shared_DirtyCardQ_lock); 2072 2073 if (G1DeferredRSUpdate) { 2074 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2075 DirtyCardQ_FL_lock, 2076 -1, // never trigger processing 2077 -1, // no limit on length 2078 Shared_DirtyCardQ_lock, 2079 &JavaThread::dirty_card_queue_set()); 2080 } 2081 2082 // Initialize the card queue set used to hold cards containing 2083 // references into the collection set. 2084 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2085 DirtyCardQ_FL_lock, 2086 -1, // never trigger processing 2087 -1, // no limit on length 2088 Shared_DirtyCardQ_lock, 2089 &JavaThread::dirty_card_queue_set()); 2090 2091 // In case we're keeping closure specialization stats, initialize those 2092 // counts and that mechanism. 2093 SpecializationStats::clear(); 2094 2095 // Do later initialization work for concurrent refinement. 2096 _cg1r->init(); 2097 2098 // Here we allocate the dummy full region that is required by the 2099 // G1AllocRegion class. If we don't pass an address in the reserved 2100 // space here, lots of asserts fire. 2101 2102 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2103 _g1_reserved.start()); 2104 // We'll re-use the same region whether the alloc region will 2105 // require BOT updates or not and, if it doesn't, then a non-young 2106 // region will complain that it cannot support allocations without 2107 // BOT updates. So we'll tag the dummy region as young to avoid that. 2108 dummy_region->set_young(); 2109 // Make sure it's full. 2110 dummy_region->set_top(dummy_region->end()); 2111 G1AllocRegion::setup(this, dummy_region); 2112 2113 init_mutator_alloc_region(); 2114 2115 // Do create of the monitoring and management support so that 2116 // values in the heap have been properly initialized. 2117 _g1mm = new G1MonitoringSupport(this); 2118 2119 return JNI_OK; 2120 } 2121 2122 void G1CollectedHeap::ref_processing_init() { 2123 // Reference processing in G1 currently works as follows: 2124 // 2125 // * There are two reference processor instances. One is 2126 // used to record and process discovered references 2127 // during concurrent marking; the other is used to 2128 // record and process references during STW pauses 2129 // (both full and incremental). 2130 // * Both ref processors need to 'span' the entire heap as 2131 // the regions in the collection set may be dotted around. 2132 // 2133 // * For the concurrent marking ref processor: 2134 // * Reference discovery is enabled at initial marking. 2135 // * Reference discovery is disabled and the discovered 2136 // references processed etc during remarking. 2137 // * Reference discovery is MT (see below). 2138 // * Reference discovery requires a barrier (see below). 2139 // * Reference processing may or may not be MT 2140 // (depending on the value of ParallelRefProcEnabled 2141 // and ParallelGCThreads). 2142 // * A full GC disables reference discovery by the CM 2143 // ref processor and abandons any entries on it's 2144 // discovered lists. 2145 // 2146 // * For the STW processor: 2147 // * Non MT discovery is enabled at the start of a full GC. 2148 // * Processing and enqueueing during a full GC is non-MT. 2149 // * During a full GC, references are processed after marking. 2150 // 2151 // * Discovery (may or may not be MT) is enabled at the start 2152 // of an incremental evacuation pause. 2153 // * References are processed near the end of a STW evacuation pause. 2154 // * For both types of GC: 2155 // * Discovery is atomic - i.e. not concurrent. 2156 // * Reference discovery will not need a barrier. 2157 2158 SharedHeap::ref_processing_init(); 2159 MemRegion mr = reserved_region(); 2160 2161 // Concurrent Mark ref processor 2162 _ref_processor_cm = 2163 new ReferenceProcessor(mr, // span 2164 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2165 // mt processing 2166 (int) ParallelGCThreads, 2167 // degree of mt processing 2168 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2169 // mt discovery 2170 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2171 // degree of mt discovery 2172 false, 2173 // Reference discovery is not atomic 2174 &_is_alive_closure_cm, 2175 // is alive closure 2176 // (for efficiency/performance) 2177 true); 2178 // Setting next fields of discovered 2179 // lists requires a barrier. 2180 2181 // STW ref processor 2182 _ref_processor_stw = 2183 new ReferenceProcessor(mr, // span 2184 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2185 // mt processing 2186 MAX2((int)ParallelGCThreads, 1), 2187 // degree of mt processing 2188 (ParallelGCThreads > 1), 2189 // mt discovery 2190 MAX2((int)ParallelGCThreads, 1), 2191 // degree of mt discovery 2192 true, 2193 // Reference discovery is atomic 2194 &_is_alive_closure_stw, 2195 // is alive closure 2196 // (for efficiency/performance) 2197 false); 2198 // Setting next fields of discovered 2199 // lists requires a barrier. 2200 } 2201 2202 size_t G1CollectedHeap::capacity() const { 2203 return _g1_committed.byte_size(); 2204 } 2205 2206 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2207 DirtyCardQueue* into_cset_dcq, 2208 bool concurrent, 2209 int worker_i) { 2210 // Clean cards in the hot card cache 2211 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2212 2213 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2214 int n_completed_buffers = 0; 2215 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2216 n_completed_buffers++; 2217 } 2218 g1_policy()->record_update_rs_processed_buffers(worker_i, 2219 (double) n_completed_buffers); 2220 dcqs.clear_n_completed_buffers(); 2221 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2222 } 2223 2224 2225 // Computes the sum of the storage used by the various regions. 2226 2227 size_t G1CollectedHeap::used() const { 2228 assert(Heap_lock->owner() != NULL, 2229 "Should be owned on this thread's behalf."); 2230 size_t result = _summary_bytes_used; 2231 // Read only once in case it is set to NULL concurrently 2232 HeapRegion* hr = _mutator_alloc_region.get(); 2233 if (hr != NULL) 2234 result += hr->used(); 2235 return result; 2236 } 2237 2238 size_t G1CollectedHeap::used_unlocked() const { 2239 size_t result = _summary_bytes_used; 2240 return result; 2241 } 2242 2243 class SumUsedClosure: public HeapRegionClosure { 2244 size_t _used; 2245 public: 2246 SumUsedClosure() : _used(0) {} 2247 bool doHeapRegion(HeapRegion* r) { 2248 if (!r->continuesHumongous()) { 2249 _used += r->used(); 2250 } 2251 return false; 2252 } 2253 size_t result() { return _used; } 2254 }; 2255 2256 size_t G1CollectedHeap::recalculate_used() const { 2257 SumUsedClosure blk; 2258 heap_region_iterate(&blk); 2259 return blk.result(); 2260 } 2261 2262 size_t G1CollectedHeap::unsafe_max_alloc() { 2263 if (free_regions() > 0) return HeapRegion::GrainBytes; 2264 // otherwise, is there space in the current allocation region? 2265 2266 // We need to store the current allocation region in a local variable 2267 // here. The problem is that this method doesn't take any locks and 2268 // there may be other threads which overwrite the current allocation 2269 // region field. attempt_allocation(), for example, sets it to NULL 2270 // and this can happen *after* the NULL check here but before the call 2271 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2272 // to be a problem in the optimized build, since the two loads of the 2273 // current allocation region field are optimized away. 2274 HeapRegion* hr = _mutator_alloc_region.get(); 2275 if (hr == NULL) { 2276 return 0; 2277 } 2278 return hr->free(); 2279 } 2280 2281 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2282 return 2283 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) || 2284 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent)); 2285 } 2286 2287 #ifndef PRODUCT 2288 void G1CollectedHeap::allocate_dummy_regions() { 2289 // Let's fill up most of the region 2290 size_t word_size = HeapRegion::GrainWords - 1024; 2291 // And as a result the region we'll allocate will be humongous. 2292 guarantee(isHumongous(word_size), "sanity"); 2293 2294 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2295 // Let's use the existing mechanism for the allocation 2296 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2297 if (dummy_obj != NULL) { 2298 MemRegion mr(dummy_obj, word_size); 2299 CollectedHeap::fill_with_object(mr); 2300 } else { 2301 // If we can't allocate once, we probably cannot allocate 2302 // again. Let's get out of the loop. 2303 break; 2304 } 2305 } 2306 } 2307 #endif // !PRODUCT 2308 2309 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) { 2310 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2311 2312 // We assume that if concurrent == true, then the caller is a 2313 // concurrent thread that was joined the Suspendible Thread 2314 // Set. If there's ever a cheap way to check this, we should add an 2315 // assert here. 2316 2317 // We have already incremented _total_full_collections at the start 2318 // of the GC, so total_full_collections() represents how many full 2319 // collections have been started. 2320 unsigned int full_collections_started = total_full_collections(); 2321 2322 // Given that this method is called at the end of a Full GC or of a 2323 // concurrent cycle, and those can be nested (i.e., a Full GC can 2324 // interrupt a concurrent cycle), the number of full collections 2325 // completed should be either one (in the case where there was no 2326 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2327 // behind the number of full collections started. 2328 2329 // This is the case for the inner caller, i.e. a Full GC. 2330 assert(concurrent || 2331 (full_collections_started == _full_collections_completed + 1) || 2332 (full_collections_started == _full_collections_completed + 2), 2333 err_msg("for inner caller (Full GC): full_collections_started = %u " 2334 "is inconsistent with _full_collections_completed = %u", 2335 full_collections_started, _full_collections_completed)); 2336 2337 // This is the case for the outer caller, i.e. the concurrent cycle. 2338 assert(!concurrent || 2339 (full_collections_started == _full_collections_completed + 1), 2340 err_msg("for outer caller (concurrent cycle): " 2341 "full_collections_started = %u " 2342 "is inconsistent with _full_collections_completed = %u", 2343 full_collections_started, _full_collections_completed)); 2344 2345 _full_collections_completed += 1; 2346 2347 // We need to clear the "in_progress" flag in the CM thread before 2348 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2349 // is set) so that if a waiter requests another System.gc() it doesn't 2350 // incorrectly see that a marking cyle is still in progress. 2351 if (concurrent) { 2352 _cmThread->clear_in_progress(); 2353 } 2354 2355 // This notify_all() will ensure that a thread that called 2356 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2357 // and it's waiting for a full GC to finish will be woken up. It is 2358 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2359 FullGCCount_lock->notify_all(); 2360 } 2361 2362 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2363 assert_at_safepoint(true /* should_be_vm_thread */); 2364 GCCauseSetter gcs(this, cause); 2365 switch (cause) { 2366 case GCCause::_heap_inspection: 2367 case GCCause::_heap_dump: { 2368 HandleMark hm; 2369 do_full_collection(false); // don't clear all soft refs 2370 break; 2371 } 2372 default: // XXX FIX ME 2373 ShouldNotReachHere(); // Unexpected use of this function 2374 } 2375 } 2376 2377 void G1CollectedHeap::collect(GCCause::Cause cause) { 2378 // The caller doesn't have the Heap_lock 2379 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock"); 2380 2381 unsigned int gc_count_before; 2382 unsigned int full_gc_count_before; 2383 { 2384 MutexLocker ml(Heap_lock); 2385 2386 // Read the GC count while holding the Heap_lock 2387 gc_count_before = SharedHeap::heap()->total_collections(); 2388 full_gc_count_before = SharedHeap::heap()->total_full_collections(); 2389 } 2390 2391 if (should_do_concurrent_full_gc(cause)) { 2392 // Schedule an initial-mark evacuation pause that will start a 2393 // concurrent cycle. We're setting word_size to 0 which means that 2394 // we are not requesting a post-GC allocation. 2395 VM_G1IncCollectionPause op(gc_count_before, 2396 0, /* word_size */ 2397 true, /* should_initiate_conc_mark */ 2398 g1_policy()->max_pause_time_ms(), 2399 cause); 2400 VMThread::execute(&op); 2401 } else { 2402 if (cause == GCCause::_gc_locker 2403 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2404 2405 // Schedule a standard evacuation pause. We're setting word_size 2406 // to 0 which means that we are not requesting a post-GC allocation. 2407 VM_G1IncCollectionPause op(gc_count_before, 2408 0, /* word_size */ 2409 false, /* should_initiate_conc_mark */ 2410 g1_policy()->max_pause_time_ms(), 2411 cause); 2412 VMThread::execute(&op); 2413 } else { 2414 // Schedule a Full GC. 2415 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause); 2416 VMThread::execute(&op); 2417 } 2418 } 2419 } 2420 2421 bool G1CollectedHeap::is_in(const void* p) const { 2422 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p); 2423 if (hr != NULL) { 2424 return hr->is_in(p); 2425 } else { 2426 return _perm_gen->as_gen()->is_in(p); 2427 } 2428 } 2429 2430 // Iteration functions. 2431 2432 // Iterates an OopClosure over all ref-containing fields of objects 2433 // within a HeapRegion. 2434 2435 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2436 MemRegion _mr; 2437 OopClosure* _cl; 2438 public: 2439 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2440 : _mr(mr), _cl(cl) {} 2441 bool doHeapRegion(HeapRegion* r) { 2442 if (! r->continuesHumongous()) { 2443 r->oop_iterate(_cl); 2444 } 2445 return false; 2446 } 2447 }; 2448 2449 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2450 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2451 heap_region_iterate(&blk); 2452 if (do_perm) { 2453 perm_gen()->oop_iterate(cl); 2454 } 2455 } 2456 2457 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2458 IterateOopClosureRegionClosure blk(mr, cl); 2459 heap_region_iterate(&blk); 2460 if (do_perm) { 2461 perm_gen()->oop_iterate(cl); 2462 } 2463 } 2464 2465 // Iterates an ObjectClosure over all objects within a HeapRegion. 2466 2467 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2468 ObjectClosure* _cl; 2469 public: 2470 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2471 bool doHeapRegion(HeapRegion* r) { 2472 if (! r->continuesHumongous()) { 2473 r->object_iterate(_cl); 2474 } 2475 return false; 2476 } 2477 }; 2478 2479 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2480 IterateObjectClosureRegionClosure blk(cl); 2481 heap_region_iterate(&blk); 2482 if (do_perm) { 2483 perm_gen()->object_iterate(cl); 2484 } 2485 } 2486 2487 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2488 // FIXME: is this right? 2489 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2490 } 2491 2492 // Calls a SpaceClosure on a HeapRegion. 2493 2494 class SpaceClosureRegionClosure: public HeapRegionClosure { 2495 SpaceClosure* _cl; 2496 public: 2497 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2498 bool doHeapRegion(HeapRegion* r) { 2499 _cl->do_space(r); 2500 return false; 2501 } 2502 }; 2503 2504 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2505 SpaceClosureRegionClosure blk(cl); 2506 heap_region_iterate(&blk); 2507 } 2508 2509 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2510 _hrs.iterate(cl); 2511 } 2512 2513 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r, 2514 HeapRegionClosure* cl) const { 2515 _hrs.iterate_from(r, cl); 2516 } 2517 2518 void 2519 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2520 int worker, 2521 int no_of_par_workers, 2522 jint claim_value) { 2523 const size_t regions = n_regions(); 2524 const size_t max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2525 no_of_par_workers : 2526 1); 2527 assert(UseDynamicNumberOfGCThreads || 2528 no_of_par_workers == workers()->total_workers(), 2529 "Non dynamic should use fixed number of workers"); 2530 // try to spread out the starting points of the workers 2531 const size_t start_index = regions / max_workers * (size_t) worker; 2532 2533 // each worker will actually look at all regions 2534 for (size_t count = 0; count < regions; ++count) { 2535 const size_t index = (start_index + count) % regions; 2536 assert(0 <= index && index < regions, "sanity"); 2537 HeapRegion* r = region_at(index); 2538 // we'll ignore "continues humongous" regions (we'll process them 2539 // when we come across their corresponding "start humongous" 2540 // region) and regions already claimed 2541 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2542 continue; 2543 } 2544 // OK, try to claim it 2545 if (r->claimHeapRegion(claim_value)) { 2546 // success! 2547 assert(!r->continuesHumongous(), "sanity"); 2548 if (r->startsHumongous()) { 2549 // If the region is "starts humongous" we'll iterate over its 2550 // "continues humongous" first; in fact we'll do them 2551 // first. The order is important. In on case, calling the 2552 // closure on the "starts humongous" region might de-allocate 2553 // and clear all its "continues humongous" regions and, as a 2554 // result, we might end up processing them twice. So, we'll do 2555 // them first (notice: most closures will ignore them anyway) and 2556 // then we'll do the "starts humongous" region. 2557 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) { 2558 HeapRegion* chr = region_at(ch_index); 2559 2560 // if the region has already been claimed or it's not 2561 // "continues humongous" we're done 2562 if (chr->claim_value() == claim_value || 2563 !chr->continuesHumongous()) { 2564 break; 2565 } 2566 2567 // Noone should have claimed it directly. We can given 2568 // that we claimed its "starts humongous" region. 2569 assert(chr->claim_value() != claim_value, "sanity"); 2570 assert(chr->humongous_start_region() == r, "sanity"); 2571 2572 if (chr->claimHeapRegion(claim_value)) { 2573 // we should always be able to claim it; noone else should 2574 // be trying to claim this region 2575 2576 bool res2 = cl->doHeapRegion(chr); 2577 assert(!res2, "Should not abort"); 2578 2579 // Right now, this holds (i.e., no closure that actually 2580 // does something with "continues humongous" regions 2581 // clears them). We might have to weaken it in the future, 2582 // but let's leave these two asserts here for extra safety. 2583 assert(chr->continuesHumongous(), "should still be the case"); 2584 assert(chr->humongous_start_region() == r, "sanity"); 2585 } else { 2586 guarantee(false, "we should not reach here"); 2587 } 2588 } 2589 } 2590 2591 assert(!r->continuesHumongous(), "sanity"); 2592 bool res = cl->doHeapRegion(r); 2593 assert(!res, "Should not abort"); 2594 } 2595 } 2596 } 2597 2598 class ResetClaimValuesClosure: public HeapRegionClosure { 2599 public: 2600 bool doHeapRegion(HeapRegion* r) { 2601 r->set_claim_value(HeapRegion::InitialClaimValue); 2602 return false; 2603 } 2604 }; 2605 2606 void 2607 G1CollectedHeap::reset_heap_region_claim_values() { 2608 ResetClaimValuesClosure blk; 2609 heap_region_iterate(&blk); 2610 } 2611 2612 #ifdef ASSERT 2613 // This checks whether all regions in the heap have the correct claim 2614 // value. I also piggy-backed on this a check to ensure that the 2615 // humongous_start_region() information on "continues humongous" 2616 // regions is correct. 2617 2618 class CheckClaimValuesClosure : public HeapRegionClosure { 2619 private: 2620 jint _claim_value; 2621 size_t _failures; 2622 HeapRegion* _sh_region; 2623 public: 2624 CheckClaimValuesClosure(jint claim_value) : 2625 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2626 bool doHeapRegion(HeapRegion* r) { 2627 if (r->claim_value() != _claim_value) { 2628 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2629 "claim value = %d, should be %d", 2630 HR_FORMAT_PARAMS(r), 2631 r->claim_value(), _claim_value); 2632 ++_failures; 2633 } 2634 if (!r->isHumongous()) { 2635 _sh_region = NULL; 2636 } else if (r->startsHumongous()) { 2637 _sh_region = r; 2638 } else if (r->continuesHumongous()) { 2639 if (r->humongous_start_region() != _sh_region) { 2640 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2641 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2642 HR_FORMAT_PARAMS(r), 2643 r->humongous_start_region(), 2644 _sh_region); 2645 ++_failures; 2646 } 2647 } 2648 return false; 2649 } 2650 size_t failures() { 2651 return _failures; 2652 } 2653 }; 2654 2655 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2656 CheckClaimValuesClosure cl(claim_value); 2657 heap_region_iterate(&cl); 2658 return cl.failures() == 0; 2659 } 2660 2661 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2662 jint _claim_value; 2663 size_t _failures; 2664 2665 public: 2666 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2667 _claim_value(claim_value), 2668 _failures(0) { } 2669 2670 size_t failures() { 2671 return _failures; 2672 } 2673 2674 bool doHeapRegion(HeapRegion* hr) { 2675 assert(hr->in_collection_set(), "how?"); 2676 assert(!hr->isHumongous(), "H-region in CSet"); 2677 if (hr->claim_value() != _claim_value) { 2678 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2679 "claim value = %d, should be %d", 2680 HR_FORMAT_PARAMS(hr), 2681 hr->claim_value(), _claim_value); 2682 _failures += 1; 2683 } 2684 return false; 2685 } 2686 }; 2687 2688 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2689 CheckClaimValuesInCSetHRClosure cl(claim_value); 2690 collection_set_iterate(&cl); 2691 return cl.failures() == 0; 2692 } 2693 #endif // ASSERT 2694 2695 // Given the id of a worker, obtain or calculate a suitable 2696 // starting region for iterating over the current collection set. 2697 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2698 assert(_worker_cset_start_region != NULL, "sanity"); 2699 assert(_worker_cset_start_region_stamp != NULL, "sanity"); 2700 assert(total_collections() > 0, "should have been updated by now"); 2701 2702 HeapRegion* result = NULL; 2703 2704 if (_worker_cset_start_region_stamp[worker_i] == total_collections()) { 2705 // Cached starting region for current worker was set 2706 // during the current pause - so it's valid. 2707 // Note: the cached starting heap region may be NULL 2708 // (when the collection set is empty). 2709 result = _worker_cset_start_region[worker_i]; 2710 assert(result == NULL || result->in_collection_set(), "sanity"); 2711 return result; 2712 } 2713 2714 // The cached entry was not valid so let's calculate 2715 // a suitable starting heap region for this worker. 2716 2717 // We want the parallel threads to start their collection 2718 // set iteration at different collection set regions to 2719 // avoid contention. 2720 // If we have: 2721 // n collection set regions 2722 // p threads 2723 // Then thread t will start at region t * floor (n/p) 2724 2725 result = g1_policy()->collection_set(); 2726 if (G1CollectedHeap::use_parallel_gc_threads()) { 2727 size_t cs_size = g1_policy()->cset_region_length(); 2728 int n_workers = workers()->total_workers(); 2729 size_t cs_spans = cs_size / n_workers; 2730 size_t ind = cs_spans * worker_i; 2731 for (size_t i = 0; i < ind; i++) { 2732 result = result->next_in_collection_set(); 2733 } 2734 } 2735 2736 // Note: the calculated starting heap region may be NULL 2737 // (when the collection set is empty). 2738 assert(result == NULL || result->in_collection_set(), "sanity"); 2739 assert(_worker_cset_start_region_stamp[worker_i] != total_collections(), 2740 "should be updated only once per pause"); 2741 _worker_cset_start_region[worker_i] = result; 2742 _worker_cset_start_region_stamp[worker_i] = total_collections(); 2743 return result; 2744 } 2745 2746 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2747 HeapRegion* r = g1_policy()->collection_set(); 2748 while (r != NULL) { 2749 HeapRegion* next = r->next_in_collection_set(); 2750 if (cl->doHeapRegion(r)) { 2751 cl->incomplete(); 2752 return; 2753 } 2754 r = next; 2755 } 2756 } 2757 2758 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2759 HeapRegionClosure *cl) { 2760 if (r == NULL) { 2761 // The CSet is empty so there's nothing to do. 2762 return; 2763 } 2764 2765 assert(r->in_collection_set(), 2766 "Start region must be a member of the collection set."); 2767 HeapRegion* cur = r; 2768 while (cur != NULL) { 2769 HeapRegion* next = cur->next_in_collection_set(); 2770 if (cl->doHeapRegion(cur) && false) { 2771 cl->incomplete(); 2772 return; 2773 } 2774 cur = next; 2775 } 2776 cur = g1_policy()->collection_set(); 2777 while (cur != r) { 2778 HeapRegion* next = cur->next_in_collection_set(); 2779 if (cl->doHeapRegion(cur) && false) { 2780 cl->incomplete(); 2781 return; 2782 } 2783 cur = next; 2784 } 2785 } 2786 2787 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2788 return n_regions() > 0 ? region_at(0) : NULL; 2789 } 2790 2791 2792 Space* G1CollectedHeap::space_containing(const void* addr) const { 2793 Space* res = heap_region_containing(addr); 2794 if (res == NULL) 2795 res = perm_gen()->space_containing(addr); 2796 return res; 2797 } 2798 2799 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2800 Space* sp = space_containing(addr); 2801 if (sp != NULL) { 2802 return sp->block_start(addr); 2803 } 2804 return NULL; 2805 } 2806 2807 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2808 Space* sp = space_containing(addr); 2809 assert(sp != NULL, "block_size of address outside of heap"); 2810 return sp->block_size(addr); 2811 } 2812 2813 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2814 Space* sp = space_containing(addr); 2815 return sp->block_is_obj(addr); 2816 } 2817 2818 bool G1CollectedHeap::supports_tlab_allocation() const { 2819 return true; 2820 } 2821 2822 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2823 return HeapRegion::GrainBytes; 2824 } 2825 2826 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2827 // Return the remaining space in the cur alloc region, but not less than 2828 // the min TLAB size. 2829 2830 // Also, this value can be at most the humongous object threshold, 2831 // since we can't allow tlabs to grow big enough to accomodate 2832 // humongous objects. 2833 2834 HeapRegion* hr = _mutator_alloc_region.get(); 2835 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2836 if (hr == NULL) { 2837 return max_tlab_size; 2838 } else { 2839 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2840 } 2841 } 2842 2843 size_t G1CollectedHeap::max_capacity() const { 2844 return _g1_reserved.byte_size(); 2845 } 2846 2847 jlong G1CollectedHeap::millis_since_last_gc() { 2848 // assert(false, "NYI"); 2849 return 0; 2850 } 2851 2852 void G1CollectedHeap::prepare_for_verify() { 2853 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2854 ensure_parsability(false); 2855 } 2856 g1_rem_set()->prepare_for_verify(); 2857 } 2858 2859 class VerifyLivenessOopClosure: public OopClosure { 2860 G1CollectedHeap* _g1h; 2861 VerifyOption _vo; 2862 public: 2863 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 2864 _g1h(g1h), _vo(vo) 2865 { } 2866 void do_oop(narrowOop *p) { do_oop_work(p); } 2867 void do_oop( oop *p) { do_oop_work(p); } 2868 2869 template <class T> void do_oop_work(T *p) { 2870 oop obj = oopDesc::load_decode_heap_oop(p); 2871 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 2872 "Dead object referenced by a not dead object"); 2873 } 2874 }; 2875 2876 class VerifyObjsInRegionClosure: public ObjectClosure { 2877 private: 2878 G1CollectedHeap* _g1h; 2879 size_t _live_bytes; 2880 HeapRegion *_hr; 2881 VerifyOption _vo; 2882 public: 2883 // _vo == UsePrevMarking -> use "prev" marking information, 2884 // _vo == UseNextMarking -> use "next" marking information, 2885 // _vo == UseMarkWord -> use mark word from object header. 2886 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 2887 : _live_bytes(0), _hr(hr), _vo(vo) { 2888 _g1h = G1CollectedHeap::heap(); 2889 } 2890 void do_object(oop o) { 2891 VerifyLivenessOopClosure isLive(_g1h, _vo); 2892 assert(o != NULL, "Huh?"); 2893 if (!_g1h->is_obj_dead_cond(o, _vo)) { 2894 // If the object is alive according to the mark word, 2895 // then verify that the marking information agrees. 2896 // Note we can't verify the contra-positive of the 2897 // above: if the object is dead (according to the mark 2898 // word), it may not be marked, or may have been marked 2899 // but has since became dead, or may have been allocated 2900 // since the last marking. 2901 if (_vo == VerifyOption_G1UseMarkWord) { 2902 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 2903 } 2904 2905 o->oop_iterate(&isLive); 2906 if (!_hr->obj_allocated_since_prev_marking(o)) { 2907 size_t obj_size = o->size(); // Make sure we don't overflow 2908 _live_bytes += (obj_size * HeapWordSize); 2909 } 2910 } 2911 } 2912 size_t live_bytes() { return _live_bytes; } 2913 }; 2914 2915 class PrintObjsInRegionClosure : public ObjectClosure { 2916 HeapRegion *_hr; 2917 G1CollectedHeap *_g1; 2918 public: 2919 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 2920 _g1 = G1CollectedHeap::heap(); 2921 }; 2922 2923 void do_object(oop o) { 2924 if (o != NULL) { 2925 HeapWord *start = (HeapWord *) o; 2926 size_t word_sz = o->size(); 2927 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 2928 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 2929 (void*) o, word_sz, 2930 _g1->isMarkedPrev(o), 2931 _g1->isMarkedNext(o), 2932 _hr->obj_allocated_since_prev_marking(o)); 2933 HeapWord *end = start + word_sz; 2934 HeapWord *cur; 2935 int *val; 2936 for (cur = start; cur < end; cur++) { 2937 val = (int *) cur; 2938 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 2939 } 2940 } 2941 } 2942 }; 2943 2944 class VerifyRegionClosure: public HeapRegionClosure { 2945 private: 2946 bool _allow_dirty; 2947 bool _par; 2948 VerifyOption _vo; 2949 bool _failures; 2950 public: 2951 // _vo == UsePrevMarking -> use "prev" marking information, 2952 // _vo == UseNextMarking -> use "next" marking information, 2953 // _vo == UseMarkWord -> use mark word from object header. 2954 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo) 2955 : _allow_dirty(allow_dirty), 2956 _par(par), 2957 _vo(vo), 2958 _failures(false) {} 2959 2960 bool failures() { 2961 return _failures; 2962 } 2963 2964 bool doHeapRegion(HeapRegion* r) { 2965 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue, 2966 "Should be unclaimed at verify points."); 2967 if (!r->continuesHumongous()) { 2968 bool failures = false; 2969 r->verify(_allow_dirty, _vo, &failures); 2970 if (failures) { 2971 _failures = true; 2972 } else { 2973 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 2974 r->object_iterate(¬_dead_yet_cl); 2975 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 2976 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 2977 "max_live_bytes "SIZE_FORMAT" " 2978 "< calculated "SIZE_FORMAT, 2979 r->bottom(), r->end(), 2980 r->max_live_bytes(), 2981 not_dead_yet_cl.live_bytes()); 2982 _failures = true; 2983 } 2984 } 2985 } 2986 return false; // stop the region iteration if we hit a failure 2987 } 2988 }; 2989 2990 class VerifyRootsClosure: public OopsInGenClosure { 2991 private: 2992 G1CollectedHeap* _g1h; 2993 VerifyOption _vo; 2994 bool _failures; 2995 public: 2996 // _vo == UsePrevMarking -> use "prev" marking information, 2997 // _vo == UseNextMarking -> use "next" marking information, 2998 // _vo == UseMarkWord -> use mark word from object header. 2999 VerifyRootsClosure(VerifyOption vo) : 3000 _g1h(G1CollectedHeap::heap()), 3001 _vo(vo), 3002 _failures(false) { } 3003 3004 bool failures() { return _failures; } 3005 3006 template <class T> void do_oop_nv(T* p) { 3007 T heap_oop = oopDesc::load_heap_oop(p); 3008 if (!oopDesc::is_null(heap_oop)) { 3009 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3010 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3011 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3012 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3013 if (_vo == VerifyOption_G1UseMarkWord) { 3014 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3015 } 3016 obj->print_on(gclog_or_tty); 3017 _failures = true; 3018 } 3019 } 3020 } 3021 3022 void do_oop(oop* p) { do_oop_nv(p); } 3023 void do_oop(narrowOop* p) { do_oop_nv(p); } 3024 }; 3025 3026 // This is the task used for parallel heap verification. 3027 3028 class G1ParVerifyTask: public AbstractGangTask { 3029 private: 3030 G1CollectedHeap* _g1h; 3031 bool _allow_dirty; 3032 VerifyOption _vo; 3033 bool _failures; 3034 3035 public: 3036 // _vo == UsePrevMarking -> use "prev" marking information, 3037 // _vo == UseNextMarking -> use "next" marking information, 3038 // _vo == UseMarkWord -> use mark word from object header. 3039 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) : 3040 AbstractGangTask("Parallel verify task"), 3041 _g1h(g1h), 3042 _allow_dirty(allow_dirty), 3043 _vo(vo), 3044 _failures(false) { } 3045 3046 bool failures() { 3047 return _failures; 3048 } 3049 3050 void work(int worker_i) { 3051 HandleMark hm; 3052 VerifyRegionClosure blk(_allow_dirty, true, _vo); 3053 _g1h->heap_region_par_iterate_chunked(&blk, worker_i, 3054 _g1h->workers()->active_workers(), 3055 HeapRegion::ParVerifyClaimValue); 3056 if (blk.failures()) { 3057 _failures = true; 3058 } 3059 } 3060 }; 3061 3062 void G1CollectedHeap::verify(bool allow_dirty, bool silent) { 3063 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking); 3064 } 3065 3066 void G1CollectedHeap::verify(bool allow_dirty, 3067 bool silent, 3068 VerifyOption vo) { 3069 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3070 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3071 VerifyRootsClosure rootsCl(vo); 3072 3073 assert(Thread::current()->is_VM_thread(), 3074 "Expected to be executed serially by the VM thread at this point"); 3075 3076 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3077 3078 // We apply the relevant closures to all the oops in the 3079 // system dictionary, the string table and the code cache. 3080 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3081 3082 process_strong_roots(true, // activate StrongRootsScope 3083 true, // we set "collecting perm gen" to true, 3084 // so we don't reset the dirty cards in the perm gen. 3085 SharedHeap::ScanningOption(so), // roots scanning options 3086 &rootsCl, 3087 &blobsCl, 3088 &rootsCl); 3089 3090 // If we're verifying after the marking phase of a Full GC then we can't 3091 // treat the perm gen as roots into the G1 heap. Some of the objects in 3092 // the perm gen may be dead and hence not marked. If one of these dead 3093 // objects is considered to be a root then we may end up with a false 3094 // "Root location <x> points to dead ob <y>" failure. 3095 if (vo != VerifyOption_G1UseMarkWord) { 3096 // Since we used "collecting_perm_gen" == true above, we will not have 3097 // checked the refs from perm into the G1-collected heap. We check those 3098 // references explicitly below. Whether the relevant cards are dirty 3099 // is checked further below in the rem set verification. 3100 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3101 perm_gen()->oop_iterate(&rootsCl); 3102 } 3103 bool failures = rootsCl.failures(); 3104 3105 if (vo != VerifyOption_G1UseMarkWord) { 3106 // If we're verifying during a full GC then the region sets 3107 // will have been torn down at the start of the GC. Therefore 3108 // verifying the region sets will fail. So we only verify 3109 // the region sets when not in a full GC. 3110 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3111 verify_region_sets(); 3112 } 3113 3114 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3115 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3116 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3117 "sanity check"); 3118 3119 G1ParVerifyTask task(this, allow_dirty, vo); 3120 assert(UseDynamicNumberOfGCThreads || 3121 workers()->active_workers() == workers()->total_workers(), 3122 "If not dynamic should be using all the workers"); 3123 int n_workers = workers()->active_workers(); 3124 set_par_threads(n_workers); 3125 workers()->run_task(&task); 3126 set_par_threads(0); 3127 if (task.failures()) { 3128 failures = true; 3129 } 3130 3131 // Checks that the expected amount of parallel work was done. 3132 // The implication is that n_workers is > 0. 3133 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3134 "sanity check"); 3135 3136 reset_heap_region_claim_values(); 3137 3138 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3139 "sanity check"); 3140 } else { 3141 VerifyRegionClosure blk(allow_dirty, false, vo); 3142 heap_region_iterate(&blk); 3143 if (blk.failures()) { 3144 failures = true; 3145 } 3146 } 3147 if (!silent) gclog_or_tty->print("RemSet "); 3148 rem_set()->verify(); 3149 3150 if (failures) { 3151 gclog_or_tty->print_cr("Heap:"); 3152 // It helps to have the per-region information in the output to 3153 // help us track down what went wrong. This is why we call 3154 // print_extended_on() instead of print_on(). 3155 print_extended_on(gclog_or_tty); 3156 gclog_or_tty->print_cr(""); 3157 #ifndef PRODUCT 3158 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3159 concurrent_mark()->print_reachable("at-verification-failure", 3160 vo, false /* all */); 3161 } 3162 #endif 3163 gclog_or_tty->flush(); 3164 } 3165 guarantee(!failures, "there should not have been any failures"); 3166 } else { 3167 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3168 } 3169 } 3170 3171 class PrintRegionClosure: public HeapRegionClosure { 3172 outputStream* _st; 3173 public: 3174 PrintRegionClosure(outputStream* st) : _st(st) {} 3175 bool doHeapRegion(HeapRegion* r) { 3176 r->print_on(_st); 3177 return false; 3178 } 3179 }; 3180 3181 void G1CollectedHeap::print_on(outputStream* st) const { 3182 st->print(" %-20s", "garbage-first heap"); 3183 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3184 capacity()/K, used_unlocked()/K); 3185 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3186 _g1_storage.low_boundary(), 3187 _g1_storage.high(), 3188 _g1_storage.high_boundary()); 3189 st->cr(); 3190 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3191 size_t young_regions = _young_list->length(); 3192 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ", 3193 young_regions, young_regions * HeapRegion::GrainBytes / K); 3194 size_t survivor_regions = g1_policy()->recorded_survivor_regions(); 3195 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)", 3196 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K); 3197 st->cr(); 3198 perm()->as_gen()->print_on(st); 3199 } 3200 3201 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3202 print_on(st); 3203 3204 // Print the per-region information. 3205 st->cr(); 3206 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), HS=humongous(starts), HC=humongous(continues), CS=collection set, F=free, TS=gc time stamp, PTAMS=previous top-at-mark-start, NTAMS=next top-at-mark-start)"); 3207 PrintRegionClosure blk(st); 3208 heap_region_iterate(&blk); 3209 } 3210 3211 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3212 if (G1CollectedHeap::use_parallel_gc_threads()) { 3213 workers()->print_worker_threads_on(st); 3214 } 3215 _cmThread->print_on(st); 3216 st->cr(); 3217 _cm->print_worker_threads_on(st); 3218 _cg1r->print_worker_threads_on(st); 3219 st->cr(); 3220 } 3221 3222 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3223 if (G1CollectedHeap::use_parallel_gc_threads()) { 3224 workers()->threads_do(tc); 3225 } 3226 tc->do_thread(_cmThread); 3227 _cg1r->threads_do(tc); 3228 } 3229 3230 void G1CollectedHeap::print_tracing_info() const { 3231 // We'll overload this to mean "trace GC pause statistics." 3232 if (TraceGen0Time || TraceGen1Time) { 3233 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3234 // to that. 3235 g1_policy()->print_tracing_info(); 3236 } 3237 if (G1SummarizeRSetStats) { 3238 g1_rem_set()->print_summary_info(); 3239 } 3240 if (G1SummarizeConcMark) { 3241 concurrent_mark()->print_summary_info(); 3242 } 3243 g1_policy()->print_yg_surv_rate_info(); 3244 SpecializationStats::print(); 3245 } 3246 3247 #ifndef PRODUCT 3248 // Helpful for debugging RSet issues. 3249 3250 class PrintRSetsClosure : public HeapRegionClosure { 3251 private: 3252 const char* _msg; 3253 size_t _occupied_sum; 3254 3255 public: 3256 bool doHeapRegion(HeapRegion* r) { 3257 HeapRegionRemSet* hrrs = r->rem_set(); 3258 size_t occupied = hrrs->occupied(); 3259 _occupied_sum += occupied; 3260 3261 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3262 HR_FORMAT_PARAMS(r)); 3263 if (occupied == 0) { 3264 gclog_or_tty->print_cr(" RSet is empty"); 3265 } else { 3266 hrrs->print(); 3267 } 3268 gclog_or_tty->print_cr("----------"); 3269 return false; 3270 } 3271 3272 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3273 gclog_or_tty->cr(); 3274 gclog_or_tty->print_cr("========================================"); 3275 gclog_or_tty->print_cr(msg); 3276 gclog_or_tty->cr(); 3277 } 3278 3279 ~PrintRSetsClosure() { 3280 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3281 gclog_or_tty->print_cr("========================================"); 3282 gclog_or_tty->cr(); 3283 } 3284 }; 3285 3286 void G1CollectedHeap::print_cset_rsets() { 3287 PrintRSetsClosure cl("Printing CSet RSets"); 3288 collection_set_iterate(&cl); 3289 } 3290 3291 void G1CollectedHeap::print_all_rsets() { 3292 PrintRSetsClosure cl("Printing All RSets");; 3293 heap_region_iterate(&cl); 3294 } 3295 #endif // PRODUCT 3296 3297 G1CollectedHeap* G1CollectedHeap::heap() { 3298 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3299 "not a garbage-first heap"); 3300 return _g1h; 3301 } 3302 3303 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3304 // always_do_update_barrier = false; 3305 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3306 // Call allocation profiler 3307 AllocationProfiler::iterate_since_last_gc(); 3308 // Fill TLAB's and such 3309 ensure_parsability(true); 3310 } 3311 3312 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3313 // FIXME: what is this about? 3314 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3315 // is set. 3316 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3317 "derived pointer present")); 3318 // always_do_update_barrier = true; 3319 3320 // We have just completed a GC. Update the soft reference 3321 // policy with the new heap occupancy 3322 Universe::update_heap_info_at_gc(); 3323 } 3324 3325 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3326 unsigned int gc_count_before, 3327 bool* succeeded) { 3328 assert_heap_not_locked_and_not_at_safepoint(); 3329 g1_policy()->record_stop_world_start(); 3330 VM_G1IncCollectionPause op(gc_count_before, 3331 word_size, 3332 false, /* should_initiate_conc_mark */ 3333 g1_policy()->max_pause_time_ms(), 3334 GCCause::_g1_inc_collection_pause); 3335 VMThread::execute(&op); 3336 3337 HeapWord* result = op.result(); 3338 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3339 assert(result == NULL || ret_succeeded, 3340 "the result should be NULL if the VM did not succeed"); 3341 *succeeded = ret_succeeded; 3342 3343 assert_heap_not_locked(); 3344 return result; 3345 } 3346 3347 void 3348 G1CollectedHeap::doConcurrentMark() { 3349 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3350 if (!_cmThread->in_progress()) { 3351 _cmThread->set_started(); 3352 CGC_lock->notify(); 3353 } 3354 } 3355 3356 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr, 3357 bool young) { 3358 return _g1_policy->predict_region_elapsed_time_ms(hr, young); 3359 } 3360 3361 void G1CollectedHeap::check_if_region_is_too_expensive(double 3362 predicted_time_ms) { 3363 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms); 3364 } 3365 3366 size_t G1CollectedHeap::pending_card_num() { 3367 size_t extra_cards = 0; 3368 JavaThread *curr = Threads::first(); 3369 while (curr != NULL) { 3370 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3371 extra_cards += dcq.size(); 3372 curr = curr->next(); 3373 } 3374 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3375 size_t buffer_size = dcqs.buffer_size(); 3376 size_t buffer_num = dcqs.completed_buffers_num(); 3377 return buffer_size * buffer_num + extra_cards; 3378 } 3379 3380 size_t G1CollectedHeap::max_pending_card_num() { 3381 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3382 size_t buffer_size = dcqs.buffer_size(); 3383 size_t buffer_num = dcqs.completed_buffers_num(); 3384 int thread_num = Threads::number_of_threads(); 3385 return (buffer_num + thread_num) * buffer_size; 3386 } 3387 3388 size_t G1CollectedHeap::cards_scanned() { 3389 return g1_rem_set()->cardsScanned(); 3390 } 3391 3392 void 3393 G1CollectedHeap::setup_surviving_young_words() { 3394 guarantee( _surviving_young_words == NULL, "pre-condition" ); 3395 size_t array_length = g1_policy()->young_cset_region_length(); 3396 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length); 3397 if (_surviving_young_words == NULL) { 3398 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3399 "Not enough space for young surv words summary."); 3400 } 3401 memset(_surviving_young_words, 0, array_length * sizeof(size_t)); 3402 #ifdef ASSERT 3403 for (size_t i = 0; i < array_length; ++i) { 3404 assert( _surviving_young_words[i] == 0, "memset above" ); 3405 } 3406 #endif // !ASSERT 3407 } 3408 3409 void 3410 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3411 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3412 size_t array_length = g1_policy()->young_cset_region_length(); 3413 for (size_t i = 0; i < array_length; ++i) 3414 _surviving_young_words[i] += surv_young_words[i]; 3415 } 3416 3417 void 3418 G1CollectedHeap::cleanup_surviving_young_words() { 3419 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3420 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words); 3421 _surviving_young_words = NULL; 3422 } 3423 3424 #ifdef ASSERT 3425 class VerifyCSetClosure: public HeapRegionClosure { 3426 public: 3427 bool doHeapRegion(HeapRegion* hr) { 3428 // Here we check that the CSet region's RSet is ready for parallel 3429 // iteration. The fields that we'll verify are only manipulated 3430 // when the region is part of a CSet and is collected. Afterwards, 3431 // we reset these fields when we clear the region's RSet (when the 3432 // region is freed) so they are ready when the region is 3433 // re-allocated. The only exception to this is if there's an 3434 // evacuation failure and instead of freeing the region we leave 3435 // it in the heap. In that case, we reset these fields during 3436 // evacuation failure handling. 3437 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3438 3439 // Here's a good place to add any other checks we'd like to 3440 // perform on CSet regions. 3441 return false; 3442 } 3443 }; 3444 #endif // ASSERT 3445 3446 #if TASKQUEUE_STATS 3447 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3448 st->print_raw_cr("GC Task Stats"); 3449 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3450 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3451 } 3452 3453 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3454 print_taskqueue_stats_hdr(st); 3455 3456 TaskQueueStats totals; 3457 const int n = workers() != NULL ? workers()->total_workers() : 1; 3458 for (int i = 0; i < n; ++i) { 3459 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3460 totals += task_queue(i)->stats; 3461 } 3462 st->print_raw("tot "); totals.print(st); st->cr(); 3463 3464 DEBUG_ONLY(totals.verify()); 3465 } 3466 3467 void G1CollectedHeap::reset_taskqueue_stats() { 3468 const int n = workers() != NULL ? workers()->total_workers() : 1; 3469 for (int i = 0; i < n; ++i) { 3470 task_queue(i)->stats.reset(); 3471 } 3472 } 3473 #endif // TASKQUEUE_STATS 3474 3475 bool 3476 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3477 assert_at_safepoint(true /* should_be_vm_thread */); 3478 guarantee(!is_gc_active(), "collection is not reentrant"); 3479 3480 if (GC_locker::check_active_before_gc()) { 3481 return false; 3482 } 3483 3484 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3485 ResourceMark rm; 3486 3487 if (PrintHeapAtGC) { 3488 Universe::print_heap_before_gc(); 3489 } 3490 3491 HRSPhaseSetter x(HRSPhaseEvacuation); 3492 verify_region_sets_optional(); 3493 verify_dirty_young_regions(); 3494 3495 { 3496 // This call will decide whether this pause is an initial-mark 3497 // pause. If it is, during_initial_mark_pause() will return true 3498 // for the duration of this pause. 3499 g1_policy()->decide_on_conc_mark_initiation(); 3500 3501 // We do not allow initial-mark to be piggy-backed on a 3502 // partially-young GC. 3503 assert(!g1_policy()->during_initial_mark_pause() || 3504 g1_policy()->full_young_gcs(), "sanity"); 3505 3506 // We also do not allow partially-young GCs during marking. 3507 assert(!mark_in_progress() || g1_policy()->full_young_gcs(), "sanity"); 3508 3509 char verbose_str[128]; 3510 sprintf(verbose_str, "GC pause "); 3511 if (g1_policy()->full_young_gcs()) { 3512 strcat(verbose_str, "(young)"); 3513 } else { 3514 strcat(verbose_str, "(partial)"); 3515 } 3516 if (g1_policy()->during_initial_mark_pause()) { 3517 strcat(verbose_str, " (initial-mark)"); 3518 // We are about to start a marking cycle, so we increment the 3519 // full collection counter. 3520 increment_total_full_collections(); 3521 } 3522 3523 // if PrintGCDetails is on, we'll print long statistics information 3524 // in the collector policy code, so let's not print this as the output 3525 // is messy if we do. 3526 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 3527 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3528 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty); 3529 3530 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3531 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3532 3533 // If the secondary_free_list is not empty, append it to the 3534 // free_list. No need to wait for the cleanup operation to finish; 3535 // the region allocation code will check the secondary_free_list 3536 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3537 // set, skip this step so that the region allocation code has to 3538 // get entries from the secondary_free_list. 3539 if (!G1StressConcRegionFreeing) { 3540 append_secondary_free_list_if_not_empty_with_lock(); 3541 } 3542 3543 assert(check_young_list_well_formed(), 3544 "young list should be well formed"); 3545 3546 // Don't dynamically change the number of GC threads this early. A value of 3547 // 0 is used to indicate serial work. When parallel work is done, 3548 // it will be set. 3549 3550 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3551 IsGCActiveMark x; 3552 3553 gc_prologue(false); 3554 increment_total_collections(false /* full gc */); 3555 increment_gc_time_stamp(); 3556 3557 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3558 HandleMark hm; // Discard invalid handles created during verification 3559 gclog_or_tty->print(" VerifyBeforeGC:"); 3560 prepare_for_verify(); 3561 Universe::verify(/* allow dirty */ false, 3562 /* silent */ false, 3563 /* option */ VerifyOption_G1UsePrevMarking); 3564 3565 } 3566 3567 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3568 3569 // Please see comment in g1CollectedHeap.hpp and 3570 // G1CollectedHeap::ref_processing_init() to see how 3571 // reference processing currently works in G1. 3572 3573 // Enable discovery in the STW reference processor 3574 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3575 true /*verify_no_refs*/); 3576 3577 { 3578 // We want to temporarily turn off discovery by the 3579 // CM ref processor, if necessary, and turn it back on 3580 // on again later if we do. Using a scoped 3581 // NoRefDiscovery object will do this. 3582 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3583 3584 // Forget the current alloc region (we might even choose it to be part 3585 // of the collection set!). 3586 release_mutator_alloc_region(); 3587 3588 // We should call this after we retire the mutator alloc 3589 // region(s) so that all the ALLOC / RETIRE events are generated 3590 // before the start GC event. 3591 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3592 3593 // The elapsed time induced by the start time below deliberately elides 3594 // the possible verification above. 3595 double start_time_sec = os::elapsedTime(); 3596 size_t start_used_bytes = used(); 3597 3598 #if YOUNG_LIST_VERBOSE 3599 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3600 _young_list->print(); 3601 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3602 #endif // YOUNG_LIST_VERBOSE 3603 3604 g1_policy()->record_collection_pause_start(start_time_sec, 3605 start_used_bytes); 3606 3607 #if YOUNG_LIST_VERBOSE 3608 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3609 _young_list->print(); 3610 #endif // YOUNG_LIST_VERBOSE 3611 3612 if (g1_policy()->during_initial_mark_pause()) { 3613 concurrent_mark()->checkpointRootsInitialPre(); 3614 } 3615 perm_gen()->save_marks(); 3616 3617 // We must do this before any possible evacuation that should propagate 3618 // marks. 3619 if (mark_in_progress()) { 3620 double start_time_sec = os::elapsedTime(); 3621 3622 _cm->drainAllSATBBuffers(); 3623 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0; 3624 g1_policy()->record_satb_drain_time(finish_mark_ms); 3625 } 3626 // Record the number of elements currently on the mark stack, so we 3627 // only iterate over these. (Since evacuation may add to the mark 3628 // stack, doing more exposes race conditions.) If no mark is in 3629 // progress, this will be zero. 3630 _cm->set_oops_do_bound(); 3631 3632 if (mark_in_progress()) { 3633 concurrent_mark()->newCSet(); 3634 } 3635 3636 #if YOUNG_LIST_VERBOSE 3637 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3638 _young_list->print(); 3639 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3640 #endif // YOUNG_LIST_VERBOSE 3641 3642 g1_policy()->choose_collection_set(target_pause_time_ms); 3643 3644 if (_hr_printer.is_active()) { 3645 HeapRegion* hr = g1_policy()->collection_set(); 3646 while (hr != NULL) { 3647 G1HRPrinter::RegionType type; 3648 if (!hr->is_young()) { 3649 type = G1HRPrinter::Old; 3650 } else if (hr->is_survivor()) { 3651 type = G1HRPrinter::Survivor; 3652 } else { 3653 type = G1HRPrinter::Eden; 3654 } 3655 _hr_printer.cset(hr); 3656 hr = hr->next_in_collection_set(); 3657 } 3658 } 3659 3660 // We have chosen the complete collection set. If marking is 3661 // active then, we clear the region fields of any of the 3662 // concurrent marking tasks whose region fields point into 3663 // the collection set as these values will become stale. This 3664 // will cause the owning marking threads to claim a new region 3665 // when marking restarts. 3666 if (mark_in_progress()) { 3667 concurrent_mark()->reset_active_task_region_fields_in_cset(); 3668 } 3669 3670 #ifdef ASSERT 3671 VerifyCSetClosure cl; 3672 collection_set_iterate(&cl); 3673 #endif // ASSERT 3674 3675 setup_surviving_young_words(); 3676 3677 // Initialize the GC alloc regions. 3678 init_gc_alloc_regions(); 3679 3680 // Actually do the work... 3681 evacuate_collection_set(); 3682 3683 free_collection_set(g1_policy()->collection_set()); 3684 g1_policy()->clear_collection_set(); 3685 3686 cleanup_surviving_young_words(); 3687 3688 // Start a new incremental collection set for the next pause. 3689 g1_policy()->start_incremental_cset_building(); 3690 3691 // Clear the _cset_fast_test bitmap in anticipation of adding 3692 // regions to the incremental collection set for the next 3693 // evacuation pause. 3694 clear_cset_fast_test(); 3695 3696 _young_list->reset_sampled_info(); 3697 3698 // Don't check the whole heap at this point as the 3699 // GC alloc regions from this pause have been tagged 3700 // as survivors and moved on to the survivor list. 3701 // Survivor regions will fail the !is_young() check. 3702 assert(check_young_list_empty(false /* check_heap */), 3703 "young list should be empty"); 3704 3705 #if YOUNG_LIST_VERBOSE 3706 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3707 _young_list->print(); 3708 #endif // YOUNG_LIST_VERBOSE 3709 3710 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3711 _young_list->first_survivor_region(), 3712 _young_list->last_survivor_region()); 3713 3714 _young_list->reset_auxilary_lists(); 3715 3716 if (evacuation_failed()) { 3717 _summary_bytes_used = recalculate_used(); 3718 } else { 3719 // The "used" of the the collection set have already been subtracted 3720 // when they were freed. Add in the bytes evacuated. 3721 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3722 } 3723 3724 if (g1_policy()->during_initial_mark_pause()) { 3725 concurrent_mark()->checkpointRootsInitialPost(); 3726 set_marking_started(); 3727 // CAUTION: after the doConcurrentMark() call below, 3728 // the concurrent marking thread(s) could be running 3729 // concurrently with us. Make sure that anything after 3730 // this point does not assume that we are the only GC thread 3731 // running. Note: of course, the actual marking work will 3732 // not start until the safepoint itself is released in 3733 // ConcurrentGCThread::safepoint_desynchronize(). 3734 doConcurrentMark(); 3735 } 3736 3737 allocate_dummy_regions(); 3738 3739 #if YOUNG_LIST_VERBOSE 3740 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3741 _young_list->print(); 3742 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3743 #endif // YOUNG_LIST_VERBOSE 3744 3745 init_mutator_alloc_region(); 3746 3747 { 3748 size_t expand_bytes = g1_policy()->expansion_amount(); 3749 if (expand_bytes > 0) { 3750 size_t bytes_before = capacity(); 3751 if (!expand(expand_bytes)) { 3752 // We failed to expand the heap so let's verify that 3753 // committed/uncommitted amount match the backing store 3754 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3755 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3756 } 3757 } 3758 } 3759 3760 double end_time_sec = os::elapsedTime(); 3761 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS; 3762 g1_policy()->record_pause_time_ms(pause_time_ms); 3763 int active_gc_threads = workers()->active_workers(); 3764 g1_policy()->record_collection_pause_end(active_gc_threads); 3765 3766 MemoryService::track_memory_usage(); 3767 3768 // In prepare_for_verify() below we'll need to scan the deferred 3769 // update buffers to bring the RSets up-to-date if 3770 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3771 // the update buffers we'll probably need to scan cards on the 3772 // regions we just allocated to (i.e., the GC alloc 3773 // regions). However, during the last GC we called 3774 // set_saved_mark() on all the GC alloc regions, so card 3775 // scanning might skip the [saved_mark_word()...top()] area of 3776 // those regions (i.e., the area we allocated objects into 3777 // during the last GC). But it shouldn't. Given that 3778 // saved_mark_word() is conditional on whether the GC time stamp 3779 // on the region is current or not, by incrementing the GC time 3780 // stamp here we invalidate all the GC time stamps on all the 3781 // regions and saved_mark_word() will simply return top() for 3782 // all the regions. This is a nicer way of ensuring this rather 3783 // than iterating over the regions and fixing them. In fact, the 3784 // GC time stamp increment here also ensures that 3785 // saved_mark_word() will return top() between pauses, i.e., 3786 // during concurrent refinement. So we don't need the 3787 // is_gc_active() check to decided which top to use when 3788 // scanning cards (see CR 7039627). 3789 increment_gc_time_stamp(); 3790 3791 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 3792 HandleMark hm; // Discard invalid handles created during verification 3793 gclog_or_tty->print(" VerifyAfterGC:"); 3794 prepare_for_verify(); 3795 Universe::verify(/* allow dirty */ true, 3796 /* silent */ false, 3797 /* option */ VerifyOption_G1UsePrevMarking); 3798 } 3799 3800 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 3801 ref_processor_stw()->verify_no_references_recorded(); 3802 3803 // CM reference discovery will be re-enabled if necessary. 3804 } 3805 3806 { 3807 size_t expand_bytes = g1_policy()->expansion_amount(); 3808 if (expand_bytes > 0) { 3809 size_t bytes_before = capacity(); 3810 // No need for an ergo verbose message here, 3811 // expansion_amount() does this when it returns a value > 0. 3812 if (!expand(expand_bytes)) { 3813 // We failed to expand the heap so let's verify that 3814 // committed/uncommitted amount match the backing store 3815 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3816 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3817 } 3818 } 3819 } 3820 3821 // We should do this after we potentially expand the heap so 3822 // that all the COMMIT events are generated before the end GC 3823 // event, and after we retire the GC alloc regions so that all 3824 // RETIRE events are generated before the end GC event. 3825 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 3826 3827 // We have to do this after we decide whether to expand the heap or not. 3828 g1_policy()->print_heap_transition(); 3829 3830 if (mark_in_progress()) { 3831 concurrent_mark()->update_g1_committed(); 3832 } 3833 3834 #ifdef TRACESPINNING 3835 ParallelTaskTerminator::print_termination_counts(); 3836 #endif 3837 3838 gc_epilogue(false); 3839 } 3840 3841 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) { 3842 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum); 3843 print_tracing_info(); 3844 vm_exit(-1); 3845 } 3846 } 3847 3848 _hrs.verify_optional(); 3849 verify_region_sets_optional(); 3850 3851 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 3852 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 3853 3854 if (PrintHeapAtGC) { 3855 Universe::print_heap_after_gc(); 3856 } 3857 g1mm()->update_sizes(); 3858 3859 if (G1SummarizeRSetStats && 3860 (G1SummarizeRSetStatsPeriod > 0) && 3861 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3862 g1_rem_set()->print_summary_info(); 3863 } 3864 3865 return true; 3866 } 3867 3868 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 3869 { 3870 size_t gclab_word_size; 3871 switch (purpose) { 3872 case GCAllocForSurvived: 3873 gclab_word_size = YoungPLABSize; 3874 break; 3875 case GCAllocForTenured: 3876 gclab_word_size = OldPLABSize; 3877 break; 3878 default: 3879 assert(false, "unknown GCAllocPurpose"); 3880 gclab_word_size = OldPLABSize; 3881 break; 3882 } 3883 return gclab_word_size; 3884 } 3885 3886 void G1CollectedHeap::init_mutator_alloc_region() { 3887 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 3888 _mutator_alloc_region.init(); 3889 } 3890 3891 void G1CollectedHeap::release_mutator_alloc_region() { 3892 _mutator_alloc_region.release(); 3893 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 3894 } 3895 3896 void G1CollectedHeap::init_gc_alloc_regions() { 3897 assert_at_safepoint(true /* should_be_vm_thread */); 3898 3899 _survivor_gc_alloc_region.init(); 3900 _old_gc_alloc_region.init(); 3901 HeapRegion* retained_region = _retained_old_gc_alloc_region; 3902 _retained_old_gc_alloc_region = NULL; 3903 3904 // We will discard the current GC alloc region if: 3905 // a) it's in the collection set (it can happen!), 3906 // b) it's already full (no point in using it), 3907 // c) it's empty (this means that it was emptied during 3908 // a cleanup and it should be on the free list now), or 3909 // d) it's humongous (this means that it was emptied 3910 // during a cleanup and was added to the free list, but 3911 // has been subseqently used to allocate a humongous 3912 // object that may be less than the region size). 3913 if (retained_region != NULL && 3914 !retained_region->in_collection_set() && 3915 !(retained_region->top() == retained_region->end()) && 3916 !retained_region->is_empty() && 3917 !retained_region->isHumongous()) { 3918 retained_region->set_saved_mark(); 3919 // The retained region was added to the old region set when it was 3920 // retired. We have to remove it now, since we don't allow regions 3921 // we allocate to in the region sets. We'll re-add it later, when 3922 // it's retired again. 3923 _old_set.remove(retained_region); 3924 _old_gc_alloc_region.set(retained_region); 3925 _hr_printer.reuse(retained_region); 3926 } 3927 } 3928 3929 void G1CollectedHeap::release_gc_alloc_regions() { 3930 _survivor_gc_alloc_region.release(); 3931 // If we have an old GC alloc region to release, we'll save it in 3932 // _retained_old_gc_alloc_region. If we don't 3933 // _retained_old_gc_alloc_region will become NULL. This is what we 3934 // want either way so no reason to check explicitly for either 3935 // condition. 3936 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 3937 } 3938 3939 void G1CollectedHeap::abandon_gc_alloc_regions() { 3940 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 3941 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 3942 _retained_old_gc_alloc_region = NULL; 3943 } 3944 3945 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 3946 _drain_in_progress = false; 3947 set_evac_failure_closure(cl); 3948 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 3949 } 3950 3951 void G1CollectedHeap::finalize_for_evac_failure() { 3952 assert(_evac_failure_scan_stack != NULL && 3953 _evac_failure_scan_stack->length() == 0, 3954 "Postcondition"); 3955 assert(!_drain_in_progress, "Postcondition"); 3956 delete _evac_failure_scan_stack; 3957 _evac_failure_scan_stack = NULL; 3958 } 3959 3960 class UpdateRSetDeferred : public OopsInHeapRegionClosure { 3961 private: 3962 G1CollectedHeap* _g1; 3963 DirtyCardQueue *_dcq; 3964 CardTableModRefBS* _ct_bs; 3965 3966 public: 3967 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) : 3968 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {} 3969 3970 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 3971 virtual void do_oop( oop* p) { do_oop_work(p); } 3972 template <class T> void do_oop_work(T* p) { 3973 assert(_from->is_in_reserved(p), "paranoia"); 3974 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) && 3975 !_from->is_survivor()) { 3976 size_t card_index = _ct_bs->index_for(p); 3977 if (_ct_bs->mark_card_deferred(card_index)) { 3978 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index)); 3979 } 3980 } 3981 } 3982 }; 3983 3984 class RemoveSelfPointerClosure: public ObjectClosure { 3985 private: 3986 G1CollectedHeap* _g1; 3987 ConcurrentMark* _cm; 3988 HeapRegion* _hr; 3989 size_t _prev_marked_bytes; 3990 size_t _next_marked_bytes; 3991 OopsInHeapRegionClosure *_cl; 3992 public: 3993 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr, 3994 OopsInHeapRegionClosure* cl) : 3995 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0), 3996 _next_marked_bytes(0), _cl(cl) {} 3997 3998 size_t prev_marked_bytes() { return _prev_marked_bytes; } 3999 size_t next_marked_bytes() { return _next_marked_bytes; } 4000 4001 // <original comment> 4002 // The original idea here was to coalesce evacuated and dead objects. 4003 // However that caused complications with the block offset table (BOT). 4004 // In particular if there were two TLABs, one of them partially refined. 4005 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~| 4006 // The BOT entries of the unrefined part of TLAB_2 point to the start 4007 // of TLAB_2. If the last object of the TLAB_1 and the first object 4008 // of TLAB_2 are coalesced, then the cards of the unrefined part 4009 // would point into middle of the filler object. 4010 // The current approach is to not coalesce and leave the BOT contents intact. 4011 // </original comment> 4012 // 4013 // We now reset the BOT when we start the object iteration over the 4014 // region and refine its entries for every object we come across. So 4015 // the above comment is not really relevant and we should be able 4016 // to coalesce dead objects if we want to. 4017 void do_object(oop obj) { 4018 HeapWord* obj_addr = (HeapWord*) obj; 4019 assert(_hr->is_in(obj_addr), "sanity"); 4020 size_t obj_size = obj->size(); 4021 _hr->update_bot_for_object(obj_addr, obj_size); 4022 if (obj->is_forwarded() && obj->forwardee() == obj) { 4023 // The object failed to move. 4024 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs."); 4025 _cm->markPrev(obj); 4026 assert(_cm->isPrevMarked(obj), "Should be marked!"); 4027 _prev_marked_bytes += (obj_size * HeapWordSize); 4028 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) { 4029 _cm->markAndGrayObjectIfNecessary(obj); 4030 } 4031 obj->set_mark(markOopDesc::prototype()); 4032 // While we were processing RSet buffers during the 4033 // collection, we actually didn't scan any cards on the 4034 // collection set, since we didn't want to update remebered 4035 // sets with entries that point into the collection set, given 4036 // that live objects fromthe collection set are about to move 4037 // and such entries will be stale very soon. This change also 4038 // dealt with a reliability issue which involved scanning a 4039 // card in the collection set and coming across an array that 4040 // was being chunked and looking malformed. The problem is 4041 // that, if evacuation fails, we might have remembered set 4042 // entries missing given that we skipped cards on the 4043 // collection set. So, we'll recreate such entries now. 4044 obj->oop_iterate(_cl); 4045 assert(_cm->isPrevMarked(obj), "Should be marked!"); 4046 } else { 4047 // The object has been either evacuated or is dead. Fill it with a 4048 // dummy object. 4049 MemRegion mr((HeapWord*)obj, obj_size); 4050 CollectedHeap::fill_with_object(mr); 4051 _cm->clearRangeBothMaps(mr); 4052 } 4053 } 4054 }; 4055 4056 void G1CollectedHeap::remove_self_forwarding_pointers() { 4057 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set()); 4058 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set()); 4059 UpdateRSetDeferred deferred_update(_g1h, &dcq); 4060 OopsInHeapRegionClosure *cl; 4061 if (G1DeferredRSUpdate) { 4062 cl = &deferred_update; 4063 } else { 4064 cl = &immediate_update; 4065 } 4066 HeapRegion* cur = g1_policy()->collection_set(); 4067 while (cur != NULL) { 4068 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!"); 4069 assert(!cur->isHumongous(), "sanity"); 4070 4071 if (cur->evacuation_failed()) { 4072 assert(cur->in_collection_set(), "bad CS"); 4073 RemoveSelfPointerClosure rspc(_g1h, cur, cl); 4074 4075 // In the common case we make sure that this is done when the 4076 // region is freed so that it is "ready-to-go" when it's 4077 // re-allocated. However, when evacuation failure happens, a 4078 // region will remain in the heap and might ultimately be added 4079 // to a CSet in the future. So we have to be careful here and 4080 // make sure the region's RSet is ready for parallel iteration 4081 // whenever this might be required in the future. 4082 cur->rem_set()->reset_for_par_iteration(); 4083 cur->reset_bot(); 4084 cl->set_region(cur); 4085 cur->object_iterate(&rspc); 4086 4087 // A number of manipulations to make the TAMS be the current top, 4088 // and the marked bytes be the ones observed in the iteration. 4089 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) { 4090 // The comments below are the postconditions achieved by the 4091 // calls. Note especially the last such condition, which says that 4092 // the count of marked bytes has been properly restored. 4093 cur->note_start_of_marking(false); 4094 // _next_top_at_mark_start == top, _next_marked_bytes == 0 4095 cur->add_to_marked_bytes(rspc.prev_marked_bytes()); 4096 // _next_marked_bytes == prev_marked_bytes. 4097 cur->note_end_of_marking(); 4098 // _prev_top_at_mark_start == top(), 4099 // _prev_marked_bytes == prev_marked_bytes 4100 } 4101 // If there is no mark in progress, we modified the _next variables 4102 // above needlessly, but harmlessly. 4103 if (_g1h->mark_in_progress()) { 4104 cur->note_start_of_marking(false); 4105 // _next_top_at_mark_start == top, _next_marked_bytes == 0 4106 // _next_marked_bytes == next_marked_bytes. 4107 } 4108 } 4109 cur = cur->next_in_collection_set(); 4110 } 4111 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!"); 4112 4113 // Now restore saved marks, if any. 4114 if (_objs_with_preserved_marks != NULL) { 4115 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4116 guarantee(_objs_with_preserved_marks->length() == 4117 _preserved_marks_of_objs->length(), "Both or none."); 4118 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4119 oop obj = _objs_with_preserved_marks->at(i); 4120 markOop m = _preserved_marks_of_objs->at(i); 4121 obj->set_mark(m); 4122 } 4123 // Delete the preserved marks growable arrays (allocated on the C heap). 4124 delete _objs_with_preserved_marks; 4125 delete _preserved_marks_of_objs; 4126 _objs_with_preserved_marks = NULL; 4127 _preserved_marks_of_objs = NULL; 4128 } 4129 } 4130 4131 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4132 _evac_failure_scan_stack->push(obj); 4133 } 4134 4135 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4136 assert(_evac_failure_scan_stack != NULL, "precondition"); 4137 4138 while (_evac_failure_scan_stack->length() > 0) { 4139 oop obj = _evac_failure_scan_stack->pop(); 4140 _evac_failure_closure->set_region(heap_region_containing(obj)); 4141 obj->oop_iterate_backwards(_evac_failure_closure); 4142 } 4143 } 4144 4145 oop 4146 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4147 oop old, 4148 bool should_mark_root) { 4149 assert(obj_in_cs(old), 4150 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4151 (HeapWord*) old)); 4152 markOop m = old->mark(); 4153 oop forward_ptr = old->forward_to_atomic(old); 4154 if (forward_ptr == NULL) { 4155 // Forward-to-self succeeded. 4156 4157 // should_mark_root will be true when this routine is called 4158 // from a root scanning closure during an initial mark pause. 4159 // In this case the thread that succeeds in self-forwarding the 4160 // object is also responsible for marking the object. 4161 if (should_mark_root) { 4162 assert(!oopDesc::is_null(old), "shouldn't be"); 4163 _cm->grayRoot(old); 4164 } 4165 4166 if (_evac_failure_closure != cl) { 4167 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4168 assert(!_drain_in_progress, 4169 "Should only be true while someone holds the lock."); 4170 // Set the global evac-failure closure to the current thread's. 4171 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4172 set_evac_failure_closure(cl); 4173 // Now do the common part. 4174 handle_evacuation_failure_common(old, m); 4175 // Reset to NULL. 4176 set_evac_failure_closure(NULL); 4177 } else { 4178 // The lock is already held, and this is recursive. 4179 assert(_drain_in_progress, "This should only be the recursive case."); 4180 handle_evacuation_failure_common(old, m); 4181 } 4182 return old; 4183 } else { 4184 // Forward-to-self failed. Either someone else managed to allocate 4185 // space for this object (old != forward_ptr) or they beat us in 4186 // self-forwarding it (old == forward_ptr). 4187 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4188 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4189 "should not be in the CSet", 4190 (HeapWord*) old, (HeapWord*) forward_ptr)); 4191 return forward_ptr; 4192 } 4193 } 4194 4195 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4196 set_evacuation_failed(true); 4197 4198 preserve_mark_if_necessary(old, m); 4199 4200 HeapRegion* r = heap_region_containing(old); 4201 if (!r->evacuation_failed()) { 4202 r->set_evacuation_failed(true); 4203 _hr_printer.evac_failure(r); 4204 } 4205 4206 push_on_evac_failure_scan_stack(old); 4207 4208 if (!_drain_in_progress) { 4209 // prevent recursion in copy_to_survivor_space() 4210 _drain_in_progress = true; 4211 drain_evac_failure_scan_stack(); 4212 _drain_in_progress = false; 4213 } 4214 } 4215 4216 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4217 assert(evacuation_failed(), "Oversaving!"); 4218 // We want to call the "for_promotion_failure" version only in the 4219 // case of a promotion failure. 4220 if (m->must_be_preserved_for_promotion_failure(obj)) { 4221 if (_objs_with_preserved_marks == NULL) { 4222 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4223 _objs_with_preserved_marks = 4224 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4225 _preserved_marks_of_objs = 4226 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true); 4227 } 4228 _objs_with_preserved_marks->push(obj); 4229 _preserved_marks_of_objs->push(m); 4230 } 4231 } 4232 4233 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4234 size_t word_size) { 4235 if (purpose == GCAllocForSurvived) { 4236 HeapWord* result = survivor_attempt_allocation(word_size); 4237 if (result != NULL) { 4238 return result; 4239 } else { 4240 // Let's try to allocate in the old gen in case we can fit the 4241 // object there. 4242 return old_attempt_allocation(word_size); 4243 } 4244 } else { 4245 assert(purpose == GCAllocForTenured, "sanity"); 4246 HeapWord* result = old_attempt_allocation(word_size); 4247 if (result != NULL) { 4248 return result; 4249 } else { 4250 // Let's try to allocate in the survivors in case we can fit the 4251 // object there. 4252 return survivor_attempt_allocation(word_size); 4253 } 4254 } 4255 4256 ShouldNotReachHere(); 4257 // Trying to keep some compilers happy. 4258 return NULL; 4259 } 4260 4261 #ifndef PRODUCT 4262 bool GCLabBitMapClosure::do_bit(size_t offset) { 4263 HeapWord* addr = _bitmap->offsetToHeapWord(offset); 4264 guarantee(_cm->isMarked(oop(addr)), "it should be!"); 4265 return true; 4266 } 4267 #endif // PRODUCT 4268 4269 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4270 ParGCAllocBuffer(gclab_word_size), 4271 _should_mark_objects(false), 4272 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size), 4273 _retired(false) 4274 { 4275 //_should_mark_objects is set to true when G1ParCopyHelper needs to 4276 // mark the forwarded location of an evacuated object. 4277 // We set _should_mark_objects to true if marking is active, i.e. when we 4278 // need to propagate a mark, or during an initial mark pause, i.e. when we 4279 // need to mark objects immediately reachable by the roots. 4280 if (G1CollectedHeap::heap()->mark_in_progress() || 4281 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) { 4282 _should_mark_objects = true; 4283 } 4284 } 4285 4286 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num) 4287 : _g1h(g1h), 4288 _refs(g1h->task_queue(queue_num)), 4289 _dcq(&g1h->dirty_card_queue_set()), 4290 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4291 _g1_rem(g1h->g1_rem_set()), 4292 _hash_seed(17), _queue_num(queue_num), 4293 _term_attempts(0), 4294 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4295 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4296 _age_table(false), 4297 _strong_roots_time(0), _term_time(0), 4298 _alloc_buffer_waste(0), _undo_waste(0) 4299 { 4300 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4301 // we "sacrifice" entry 0 to keep track of surviving bytes for 4302 // non-young regions (where the age is -1) 4303 // We also add a few elements at the beginning and at the end in 4304 // an attempt to eliminate cache contention 4305 size_t real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4306 size_t array_length = PADDING_ELEM_NUM + 4307 real_length + 4308 PADDING_ELEM_NUM; 4309 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length); 4310 if (_surviving_young_words_base == NULL) 4311 vm_exit_out_of_memory(array_length * sizeof(size_t), 4312 "Not enough space for young surv histo."); 4313 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4314 memset(_surviving_young_words, 0, real_length * sizeof(size_t)); 4315 4316 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4317 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4318 4319 _start = os::elapsedTime(); 4320 } 4321 4322 void 4323 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4324 { 4325 st->print_raw_cr("GC Termination Stats"); 4326 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4327 " ------waste (KiB)------"); 4328 st->print_raw_cr("thr ms ms % ms % attempts" 4329 " total alloc undo"); 4330 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4331 " ------- ------- -------"); 4332 } 4333 4334 void 4335 G1ParScanThreadState::print_termination_stats(int i, 4336 outputStream* const st) const 4337 { 4338 const double elapsed_ms = elapsed_time() * 1000.0; 4339 const double s_roots_ms = strong_roots_time() * 1000.0; 4340 const double term_ms = term_time() * 1000.0; 4341 st->print_cr("%3d %9.2f %9.2f %6.2f " 4342 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4343 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4344 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4345 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4346 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4347 alloc_buffer_waste() * HeapWordSize / K, 4348 undo_waste() * HeapWordSize / K); 4349 } 4350 4351 #ifdef ASSERT 4352 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4353 assert(ref != NULL, "invariant"); 4354 assert(UseCompressedOops, "sanity"); 4355 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4356 oop p = oopDesc::load_decode_heap_oop(ref); 4357 assert(_g1h->is_in_g1_reserved(p), 4358 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4359 return true; 4360 } 4361 4362 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4363 assert(ref != NULL, "invariant"); 4364 if (has_partial_array_mask(ref)) { 4365 // Must be in the collection set--it's already been copied. 4366 oop p = clear_partial_array_mask(ref); 4367 assert(_g1h->obj_in_cs(p), 4368 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4369 } else { 4370 oop p = oopDesc::load_decode_heap_oop(ref); 4371 assert(_g1h->is_in_g1_reserved(p), 4372 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4373 } 4374 return true; 4375 } 4376 4377 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4378 if (ref.is_narrow()) { 4379 return verify_ref((narrowOop*) ref); 4380 } else { 4381 return verify_ref((oop*) ref); 4382 } 4383 } 4384 #endif // ASSERT 4385 4386 void G1ParScanThreadState::trim_queue() { 4387 assert(_evac_cl != NULL, "not set"); 4388 assert(_evac_failure_cl != NULL, "not set"); 4389 assert(_partial_scan_cl != NULL, "not set"); 4390 4391 StarTask ref; 4392 do { 4393 // Drain the overflow stack first, so other threads can steal. 4394 while (refs()->pop_overflow(ref)) { 4395 deal_with_reference(ref); 4396 } 4397 4398 while (refs()->pop_local(ref)) { 4399 deal_with_reference(ref); 4400 } 4401 } while (!refs()->is_empty()); 4402 } 4403 4404 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) : 4405 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4406 _par_scan_state(par_scan_state), 4407 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4408 _mark_in_progress(_g1->mark_in_progress()) { } 4409 4410 template <class T> void G1ParCopyHelper::mark_object(T* p) { 4411 // This is called from do_oop_work for objects that are not 4412 // in the collection set. Objects in the collection set 4413 // are marked after they have been evacuated. 4414 4415 T heap_oop = oopDesc::load_heap_oop(p); 4416 if (!oopDesc::is_null(heap_oop)) { 4417 oop obj = oopDesc::decode_heap_oop(heap_oop); 4418 HeapWord* addr = (HeapWord*)obj; 4419 if (_g1->is_in_g1_reserved(addr)) { 4420 _cm->grayRoot(oop(addr)); 4421 } 4422 } 4423 } 4424 4425 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root, 4426 bool should_mark_copy) { 4427 size_t word_sz = old->size(); 4428 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4429 // +1 to make the -1 indexes valid... 4430 int young_index = from_region->young_index_in_cset()+1; 4431 assert( (from_region->is_young() && young_index > 0) || 4432 (!from_region->is_young() && young_index == 0), "invariant" ); 4433 G1CollectorPolicy* g1p = _g1->g1_policy(); 4434 markOop m = old->mark(); 4435 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4436 : m->age(); 4437 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4438 word_sz); 4439 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4440 oop obj = oop(obj_ptr); 4441 4442 if (obj_ptr == NULL) { 4443 // This will either forward-to-self, or detect that someone else has 4444 // installed a forwarding pointer. 4445 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4446 return _g1->handle_evacuation_failure_par(cl, old, should_mark_root); 4447 } 4448 4449 // We're going to allocate linearly, so might as well prefetch ahead. 4450 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4451 4452 oop forward_ptr = old->forward_to_atomic(obj); 4453 if (forward_ptr == NULL) { 4454 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4455 if (g1p->track_object_age(alloc_purpose)) { 4456 // We could simply do obj->incr_age(). However, this causes a 4457 // performance issue. obj->incr_age() will first check whether 4458 // the object has a displaced mark by checking its mark word; 4459 // getting the mark word from the new location of the object 4460 // stalls. So, given that we already have the mark word and we 4461 // are about to install it anyway, it's better to increase the 4462 // age on the mark word, when the object does not have a 4463 // displaced mark word. We're not expecting many objects to have 4464 // a displaced marked word, so that case is not optimized 4465 // further (it could be...) and we simply call obj->incr_age(). 4466 4467 if (m->has_displaced_mark_helper()) { 4468 // in this case, we have to install the mark word first, 4469 // otherwise obj looks to be forwarded (the old mark word, 4470 // which contains the forward pointer, was copied) 4471 obj->set_mark(m); 4472 obj->incr_age(); 4473 } else { 4474 m = m->incr_age(); 4475 obj->set_mark(m); 4476 } 4477 _par_scan_state->age_table()->add(obj, word_sz); 4478 } else { 4479 obj->set_mark(m); 4480 } 4481 4482 // Mark the evacuated object or propagate "next" mark bit 4483 if (should_mark_copy) { 4484 if (!use_local_bitmaps || 4485 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) { 4486 // if we couldn't mark it on the local bitmap (this happens when 4487 // the object was not allocated in the GCLab), we have to bite 4488 // the bullet and do the standard parallel mark 4489 _cm->markAndGrayObjectIfNecessary(obj); 4490 } 4491 4492 if (_g1->isMarkedNext(old)) { 4493 // Unmark the object's old location so that marking 4494 // doesn't think the old object is alive. 4495 _cm->nextMarkBitMap()->parClear((HeapWord*)old); 4496 } 4497 } 4498 4499 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4500 surv_young_words[young_index] += word_sz; 4501 4502 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4503 arrayOop(old)->set_length(0); 4504 oop* old_p = set_partial_array_mask(old); 4505 _par_scan_state->push_on_queue(old_p); 4506 } else { 4507 // No point in using the slower heap_region_containing() method, 4508 // given that we know obj is in the heap. 4509 _scanner->set_region(_g1->heap_region_containing_raw(obj)); 4510 obj->oop_iterate_backwards(_scanner); 4511 } 4512 } else { 4513 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4514 obj = forward_ptr; 4515 } 4516 return obj; 4517 } 4518 4519 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4520 template <class T> 4521 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4522 ::do_oop_work(T* p) { 4523 oop obj = oopDesc::load_decode_heap_oop(p); 4524 assert(barrier != G1BarrierRS || obj != NULL, 4525 "Precondition: G1BarrierRS implies obj is nonNull"); 4526 4527 // Marking: 4528 // If the object is in the collection set, then the thread 4529 // that copies the object should mark, or propagate the 4530 // mark to, the evacuated object. 4531 // If the object is not in the collection set then we 4532 // should call the mark_object() method depending on the 4533 // value of the template parameter do_mark_object (which will 4534 // be true for root scanning closures during an initial mark 4535 // pause). 4536 // The mark_object() method first checks whether the object 4537 // is marked and, if not, attempts to mark the object. 4538 4539 // here the null check is implicit in the cset_fast_test() test 4540 if (_g1->in_cset_fast_test(obj)) { 4541 if (obj->is_forwarded()) { 4542 oopDesc::encode_store_heap_oop(p, obj->forwardee()); 4543 // If we are a root scanning closure during an initial 4544 // mark pause (i.e. do_mark_object will be true) then 4545 // we also need to handle marking of roots in the 4546 // event of an evacuation failure. In the event of an 4547 // evacuation failure, the object is forwarded to itself 4548 // and not copied. For root-scanning closures, the 4549 // object would be marked after a successful self-forward 4550 // but an object could be pointed to by both a root and non 4551 // root location and be self-forwarded by a non-root-scanning 4552 // closure. Therefore we also have to attempt to mark the 4553 // self-forwarded root object here. 4554 if (do_mark_object && obj->forwardee() == obj) { 4555 mark_object(p); 4556 } 4557 } else { 4558 // During an initial mark pause, objects that are pointed to 4559 // by the roots need to be marked - even in the event of an 4560 // evacuation failure. We pass the template parameter 4561 // do_mark_object (which is true for root scanning closures 4562 // during an initial mark pause) to copy_to_survivor_space 4563 // which will pass it on to the evacuation failure handling 4564 // code. The thread that successfully self-forwards a root 4565 // object to itself is responsible for marking the object. 4566 bool should_mark_root = do_mark_object; 4567 4568 // We need to mark the copied object if we're a root scanning 4569 // closure during an initial mark pause (i.e. do_mark_object 4570 // will be true), or the object is already marked and we need 4571 // to propagate the mark to the evacuated copy. 4572 bool should_mark_copy = do_mark_object || 4573 _during_initial_mark || 4574 (_mark_in_progress && !_g1->is_obj_ill(obj)); 4575 4576 oop copy_oop = copy_to_survivor_space(obj, should_mark_root, 4577 should_mark_copy); 4578 oopDesc::encode_store_heap_oop(p, copy_oop); 4579 } 4580 // When scanning the RS, we only care about objs in CS. 4581 if (barrier == G1BarrierRS) { 4582 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num()); 4583 } 4584 } else { 4585 // The object is not in collection set. If we're a root scanning 4586 // closure during an initial mark pause (i.e. do_mark_object will 4587 // be true) then attempt to mark the object. 4588 if (do_mark_object) { 4589 mark_object(p); 4590 } 4591 } 4592 4593 if (barrier == G1BarrierEvac && obj != NULL) { 4594 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num()); 4595 } 4596 4597 if (do_gen_barrier && obj != NULL) { 4598 par_do_barrier(p); 4599 } 4600 } 4601 4602 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4603 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4604 4605 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4606 assert(has_partial_array_mask(p), "invariant"); 4607 oop old = clear_partial_array_mask(p); 4608 assert(old->is_objArray(), "must be obj array"); 4609 assert(old->is_forwarded(), "must be forwarded"); 4610 assert(Universe::heap()->is_in_reserved(old), "must be in heap."); 4611 4612 objArrayOop obj = objArrayOop(old->forwardee()); 4613 assert((void*)old != (void*)old->forwardee(), "self forwarding here?"); 4614 // Process ParGCArrayScanChunk elements now 4615 // and push the remainder back onto queue 4616 int start = arrayOop(old)->length(); 4617 int end = obj->length(); 4618 int remainder = end - start; 4619 assert(start <= end, "just checking"); 4620 if (remainder > 2 * ParGCArrayScanChunk) { 4621 // Test above combines last partial chunk with a full chunk 4622 end = start + ParGCArrayScanChunk; 4623 arrayOop(old)->set_length(end); 4624 // Push remainder. 4625 oop* old_p = set_partial_array_mask(old); 4626 assert(arrayOop(old)->length() < obj->length(), "Empty push?"); 4627 _par_scan_state->push_on_queue(old_p); 4628 } else { 4629 // Restore length so that the heap remains parsable in 4630 // case of evacuation failure. 4631 arrayOop(old)->set_length(end); 4632 } 4633 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4634 // process our set of indices (include header in first chunk) 4635 obj->oop_iterate_range(&_scanner, start, end); 4636 } 4637 4638 class G1ParEvacuateFollowersClosure : public VoidClosure { 4639 protected: 4640 G1CollectedHeap* _g1h; 4641 G1ParScanThreadState* _par_scan_state; 4642 RefToScanQueueSet* _queues; 4643 ParallelTaskTerminator* _terminator; 4644 4645 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4646 RefToScanQueueSet* queues() { return _queues; } 4647 ParallelTaskTerminator* terminator() { return _terminator; } 4648 4649 public: 4650 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4651 G1ParScanThreadState* par_scan_state, 4652 RefToScanQueueSet* queues, 4653 ParallelTaskTerminator* terminator) 4654 : _g1h(g1h), _par_scan_state(par_scan_state), 4655 _queues(queues), _terminator(terminator) {} 4656 4657 void do_void(); 4658 4659 private: 4660 inline bool offer_termination(); 4661 }; 4662 4663 bool G1ParEvacuateFollowersClosure::offer_termination() { 4664 G1ParScanThreadState* const pss = par_scan_state(); 4665 pss->start_term_time(); 4666 const bool res = terminator()->offer_termination(); 4667 pss->end_term_time(); 4668 return res; 4669 } 4670 4671 void G1ParEvacuateFollowersClosure::do_void() { 4672 StarTask stolen_task; 4673 G1ParScanThreadState* const pss = par_scan_state(); 4674 pss->trim_queue(); 4675 4676 do { 4677 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4678 assert(pss->verify_task(stolen_task), "sanity"); 4679 if (stolen_task.is_narrow()) { 4680 pss->deal_with_reference((narrowOop*) stolen_task); 4681 } else { 4682 pss->deal_with_reference((oop*) stolen_task); 4683 } 4684 4685 // We've just processed a reference and we might have made 4686 // available new entries on the queues. So we have to make sure 4687 // we drain the queues as necessary. 4688 pss->trim_queue(); 4689 } 4690 } while (!offer_termination()); 4691 4692 pss->retire_alloc_buffers(); 4693 } 4694 4695 class G1ParTask : public AbstractGangTask { 4696 protected: 4697 G1CollectedHeap* _g1h; 4698 RefToScanQueueSet *_queues; 4699 ParallelTaskTerminator _terminator; 4700 int _n_workers; 4701 4702 Mutex _stats_lock; 4703 Mutex* stats_lock() { return &_stats_lock; } 4704 4705 size_t getNCards() { 4706 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4707 / G1BlockOffsetSharedArray::N_bytes; 4708 } 4709 4710 public: 4711 G1ParTask(G1CollectedHeap* g1h, 4712 RefToScanQueueSet *task_queues) 4713 : AbstractGangTask("G1 collection"), 4714 _g1h(g1h), 4715 _queues(task_queues), 4716 _terminator(0, _queues), 4717 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4718 {} 4719 4720 RefToScanQueueSet* queues() { return _queues; } 4721 4722 RefToScanQueue *work_queue(int i) { 4723 return queues()->queue(i); 4724 } 4725 4726 ParallelTaskTerminator* terminator() { return &_terminator; } 4727 4728 virtual void set_for_termination(int active_workers) { 4729 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4730 // in the young space (_par_seq_tasks) in the G1 heap 4731 // for SequentialSubTasksDone. 4732 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4733 // both of which need setting by set_n_termination(). 4734 _g1h->SharedHeap::set_n_termination(active_workers); 4735 _g1h->set_n_termination(active_workers); 4736 terminator()->reset_for_reuse(active_workers); 4737 _n_workers = active_workers; 4738 } 4739 4740 void work(int i) { 4741 if (i >= _n_workers) return; // no work needed this round 4742 4743 double start_time_ms = os::elapsedTime() * 1000.0; 4744 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms); 4745 4746 ResourceMark rm; 4747 HandleMark hm; 4748 4749 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4750 4751 G1ParScanThreadState pss(_g1h, i); 4752 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4753 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4754 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4755 4756 pss.set_evac_closure(&scan_evac_cl); 4757 pss.set_evac_failure_closure(&evac_failure_cl); 4758 pss.set_partial_scan_closure(&partial_scan_cl); 4759 4760 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4761 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4762 4763 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4764 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4765 4766 OopClosure* scan_root_cl = &only_scan_root_cl; 4767 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4768 4769 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4770 // We also need to mark copied objects. 4771 scan_root_cl = &scan_mark_root_cl; 4772 scan_perm_cl = &scan_mark_perm_cl; 4773 } 4774 4775 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4776 4777 pss.start_strong_roots(); 4778 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4779 SharedHeap::SO_AllClasses, 4780 scan_root_cl, 4781 &push_heap_rs_cl, 4782 scan_perm_cl, 4783 i); 4784 pss.end_strong_roots(); 4785 4786 { 4787 double start = os::elapsedTime(); 4788 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4789 evac.do_void(); 4790 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4791 double term_ms = pss.term_time()*1000.0; 4792 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms); 4793 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts()); 4794 } 4795 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4796 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4797 4798 // Clean up any par-expanded rem sets. 4799 HeapRegionRemSet::par_cleanup(); 4800 4801 if (ParallelGCVerbose) { 4802 MutexLocker x(stats_lock()); 4803 pss.print_termination_stats(i); 4804 } 4805 4806 assert(pss.refs()->is_empty(), "should be empty"); 4807 double end_time_ms = os::elapsedTime() * 1000.0; 4808 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms); 4809 } 4810 }; 4811 4812 // *** Common G1 Evacuation Stuff 4813 4814 // This method is run in a GC worker. 4815 4816 void 4817 G1CollectedHeap:: 4818 g1_process_strong_roots(bool collecting_perm_gen, 4819 SharedHeap::ScanningOption so, 4820 OopClosure* scan_non_heap_roots, 4821 OopsInHeapRegionClosure* scan_rs, 4822 OopsInGenClosure* scan_perm, 4823 int worker_i) { 4824 4825 // First scan the strong roots, including the perm gen. 4826 double ext_roots_start = os::elapsedTime(); 4827 double closure_app_time_sec = 0.0; 4828 4829 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4830 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4831 buf_scan_perm.set_generation(perm_gen()); 4832 4833 // Walk the code cache w/o buffering, because StarTask cannot handle 4834 // unaligned oop locations. 4835 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true); 4836 4837 process_strong_roots(false, // no scoping; this is parallel code 4838 collecting_perm_gen, so, 4839 &buf_scan_non_heap_roots, 4840 &eager_scan_code_roots, 4841 &buf_scan_perm); 4842 4843 // Now the CM ref_processor roots. 4844 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4845 // We need to treat the discovered reference lists of the 4846 // concurrent mark ref processor as roots and keep entries 4847 // (which are added by the marking threads) on them live 4848 // until they can be processed at the end of marking. 4849 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4850 } 4851 4852 // Finish up any enqueued closure apps (attributed as object copy time). 4853 buf_scan_non_heap_roots.done(); 4854 buf_scan_perm.done(); 4855 4856 double ext_roots_end = os::elapsedTime(); 4857 4858 g1_policy()->reset_obj_copy_time(worker_i); 4859 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4860 buf_scan_non_heap_roots.closure_app_seconds(); 4861 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4862 4863 double ext_root_time_ms = 4864 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4865 4866 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4867 4868 // Scan strong roots in mark stack. 4869 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) { 4870 concurrent_mark()->oops_do(scan_non_heap_roots); 4871 } 4872 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0; 4873 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms); 4874 4875 // Now scan the complement of the collection set. 4876 if (scan_rs != NULL) { 4877 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4878 } 4879 4880 _process_strong_tasks->all_tasks_completed(); 4881 } 4882 4883 void 4884 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 4885 OopClosure* non_root_closure) { 4886 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 4887 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 4888 } 4889 4890 // Weak Reference Processing support 4891 4892 // An always "is_alive" closure that is used to preserve referents. 4893 // If the object is non-null then it's alive. Used in the preservation 4894 // of referent objects that are pointed to by reference objects 4895 // discovered by the CM ref processor. 4896 class G1AlwaysAliveClosure: public BoolObjectClosure { 4897 G1CollectedHeap* _g1; 4898 public: 4899 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4900 void do_object(oop p) { assert(false, "Do not call."); } 4901 bool do_object_b(oop p) { 4902 if (p != NULL) { 4903 return true; 4904 } 4905 return false; 4906 } 4907 }; 4908 4909 bool G1STWIsAliveClosure::do_object_b(oop p) { 4910 // An object is reachable if it is outside the collection set, 4911 // or is inside and copied. 4912 return !_g1->obj_in_cs(p) || p->is_forwarded(); 4913 } 4914 4915 // Non Copying Keep Alive closure 4916 class G1KeepAliveClosure: public OopClosure { 4917 G1CollectedHeap* _g1; 4918 public: 4919 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4920 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 4921 void do_oop( oop* p) { 4922 oop obj = *p; 4923 4924 if (_g1->obj_in_cs(obj)) { 4925 assert( obj->is_forwarded(), "invariant" ); 4926 *p = obj->forwardee(); 4927 } 4928 } 4929 }; 4930 4931 // Copying Keep Alive closure - can be called from both 4932 // serial and parallel code as long as different worker 4933 // threads utilize different G1ParScanThreadState instances 4934 // and different queues. 4935 4936 class G1CopyingKeepAliveClosure: public OopClosure { 4937 G1CollectedHeap* _g1h; 4938 OopClosure* _copy_non_heap_obj_cl; 4939 OopsInHeapRegionClosure* _copy_perm_obj_cl; 4940 G1ParScanThreadState* _par_scan_state; 4941 4942 public: 4943 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 4944 OopClosure* non_heap_obj_cl, 4945 OopsInHeapRegionClosure* perm_obj_cl, 4946 G1ParScanThreadState* pss): 4947 _g1h(g1h), 4948 _copy_non_heap_obj_cl(non_heap_obj_cl), 4949 _copy_perm_obj_cl(perm_obj_cl), 4950 _par_scan_state(pss) 4951 {} 4952 4953 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 4954 virtual void do_oop( oop* p) { do_oop_work(p); } 4955 4956 template <class T> void do_oop_work(T* p) { 4957 oop obj = oopDesc::load_decode_heap_oop(p); 4958 4959 if (_g1h->obj_in_cs(obj)) { 4960 // If the referent object has been forwarded (either copied 4961 // to a new location or to itself in the event of an 4962 // evacuation failure) then we need to update the reference 4963 // field and, if both reference and referent are in the G1 4964 // heap, update the RSet for the referent. 4965 // 4966 // If the referent has not been forwarded then we have to keep 4967 // it alive by policy. Therefore we have copy the referent. 4968 // 4969 // If the reference field is in the G1 heap then we can push 4970 // on the PSS queue. When the queue is drained (after each 4971 // phase of reference processing) the object and it's followers 4972 // will be copied, the reference field set to point to the 4973 // new location, and the RSet updated. Otherwise we need to 4974 // use the the non-heap or perm closures directly to copy 4975 // the refernt object and update the pointer, while avoiding 4976 // updating the RSet. 4977 4978 if (_g1h->is_in_g1_reserved(p)) { 4979 _par_scan_state->push_on_queue(p); 4980 } else { 4981 // The reference field is not in the G1 heap. 4982 if (_g1h->perm_gen()->is_in(p)) { 4983 _copy_perm_obj_cl->do_oop(p); 4984 } else { 4985 _copy_non_heap_obj_cl->do_oop(p); 4986 } 4987 } 4988 } 4989 } 4990 }; 4991 4992 // Serial drain queue closure. Called as the 'complete_gc' 4993 // closure for each discovered list in some of the 4994 // reference processing phases. 4995 4996 class G1STWDrainQueueClosure: public VoidClosure { 4997 protected: 4998 G1CollectedHeap* _g1h; 4999 G1ParScanThreadState* _par_scan_state; 5000 5001 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5002 5003 public: 5004 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5005 _g1h(g1h), 5006 _par_scan_state(pss) 5007 { } 5008 5009 void do_void() { 5010 G1ParScanThreadState* const pss = par_scan_state(); 5011 pss->trim_queue(); 5012 } 5013 }; 5014 5015 // Parallel Reference Processing closures 5016 5017 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5018 // processing during G1 evacuation pauses. 5019 5020 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5021 private: 5022 G1CollectedHeap* _g1h; 5023 RefToScanQueueSet* _queues; 5024 FlexibleWorkGang* _workers; 5025 int _active_workers; 5026 5027 public: 5028 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5029 FlexibleWorkGang* workers, 5030 RefToScanQueueSet *task_queues, 5031 int n_workers) : 5032 _g1h(g1h), 5033 _queues(task_queues), 5034 _workers(workers), 5035 _active_workers(n_workers) 5036 { 5037 assert(n_workers > 0, "shouldn't call this otherwise"); 5038 } 5039 5040 // Executes the given task using concurrent marking worker threads. 5041 virtual void execute(ProcessTask& task); 5042 virtual void execute(EnqueueTask& task); 5043 }; 5044 5045 // Gang task for possibly parallel reference processing 5046 5047 class G1STWRefProcTaskProxy: public AbstractGangTask { 5048 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5049 ProcessTask& _proc_task; 5050 G1CollectedHeap* _g1h; 5051 RefToScanQueueSet *_task_queues; 5052 ParallelTaskTerminator* _terminator; 5053 5054 public: 5055 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5056 G1CollectedHeap* g1h, 5057 RefToScanQueueSet *task_queues, 5058 ParallelTaskTerminator* terminator) : 5059 AbstractGangTask("Process reference objects in parallel"), 5060 _proc_task(proc_task), 5061 _g1h(g1h), 5062 _task_queues(task_queues), 5063 _terminator(terminator) 5064 {} 5065 5066 virtual void work(int i) { 5067 // The reference processing task executed by a single worker. 5068 ResourceMark rm; 5069 HandleMark hm; 5070 5071 G1STWIsAliveClosure is_alive(_g1h); 5072 5073 G1ParScanThreadState pss(_g1h, i); 5074 5075 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5076 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5077 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5078 5079 pss.set_evac_closure(&scan_evac_cl); 5080 pss.set_evac_failure_closure(&evac_failure_cl); 5081 pss.set_partial_scan_closure(&partial_scan_cl); 5082 5083 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5084 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5085 5086 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5087 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5088 5089 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5090 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5091 5092 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5093 // We also need to mark copied objects. 5094 copy_non_heap_cl = ©_mark_non_heap_cl; 5095 copy_perm_cl = ©_mark_perm_cl; 5096 } 5097 5098 // Keep alive closure. 5099 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5100 5101 // Complete GC closure 5102 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5103 5104 // Call the reference processing task's work routine. 5105 _proc_task.work(i, is_alive, keep_alive, drain_queue); 5106 5107 // Note we cannot assert that the refs array is empty here as not all 5108 // of the processing tasks (specifically phase2 - pp2_work) execute 5109 // the complete_gc closure (which ordinarily would drain the queue) so 5110 // the queue may not be empty. 5111 } 5112 }; 5113 5114 // Driver routine for parallel reference processing. 5115 // Creates an instance of the ref processing gang 5116 // task and has the worker threads execute it. 5117 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5118 assert(_workers != NULL, "Need parallel worker threads."); 5119 5120 ParallelTaskTerminator terminator(_active_workers, _queues); 5121 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5122 5123 _g1h->set_par_threads(_active_workers); 5124 _workers->run_task(&proc_task_proxy); 5125 _g1h->set_par_threads(0); 5126 } 5127 5128 // Gang task for parallel reference enqueueing. 5129 5130 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5131 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5132 EnqueueTask& _enq_task; 5133 5134 public: 5135 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5136 AbstractGangTask("Enqueue reference objects in parallel"), 5137 _enq_task(enq_task) 5138 { } 5139 5140 virtual void work(int i) { 5141 _enq_task.work(i); 5142 } 5143 }; 5144 5145 // Driver routine for parallel reference enqueing. 5146 // Creates an instance of the ref enqueueing gang 5147 // task and has the worker threads execute it. 5148 5149 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5150 assert(_workers != NULL, "Need parallel worker threads."); 5151 5152 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5153 5154 _g1h->set_par_threads(_active_workers); 5155 _workers->run_task(&enq_task_proxy); 5156 _g1h->set_par_threads(0); 5157 } 5158 5159 // End of weak reference support closures 5160 5161 // Abstract task used to preserve (i.e. copy) any referent objects 5162 // that are in the collection set and are pointed to by reference 5163 // objects discovered by the CM ref processor. 5164 5165 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5166 protected: 5167 G1CollectedHeap* _g1h; 5168 RefToScanQueueSet *_queues; 5169 ParallelTaskTerminator _terminator; 5170 int _n_workers; 5171 5172 public: 5173 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5174 AbstractGangTask("ParPreserveCMReferents"), 5175 _g1h(g1h), 5176 _queues(task_queues), 5177 _terminator(workers, _queues), 5178 _n_workers(workers) 5179 { } 5180 5181 void work(int i) { 5182 ResourceMark rm; 5183 HandleMark hm; 5184 5185 G1ParScanThreadState pss(_g1h, i); 5186 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5187 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5188 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5189 5190 pss.set_evac_closure(&scan_evac_cl); 5191 pss.set_evac_failure_closure(&evac_failure_cl); 5192 pss.set_partial_scan_closure(&partial_scan_cl); 5193 5194 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5195 5196 5197 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5198 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5199 5200 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5201 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5202 5203 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5204 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5205 5206 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5207 // We also need to mark copied objects. 5208 copy_non_heap_cl = ©_mark_non_heap_cl; 5209 copy_perm_cl = ©_mark_perm_cl; 5210 } 5211 5212 // Is alive closure 5213 G1AlwaysAliveClosure always_alive(_g1h); 5214 5215 // Copying keep alive closure. Applied to referent objects that need 5216 // to be copied. 5217 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5218 5219 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5220 5221 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5222 int stride = MIN2(MAX2(_n_workers, 1), limit); 5223 5224 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5225 // So this must be true - but assert just in case someone decides to 5226 // change the worker ids. 5227 assert(0 <= i && i < limit, "sanity"); 5228 assert(!rp->discovery_is_atomic(), "check this code"); 5229 5230 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5231 for (int idx = i; idx < limit; idx += stride) { 5232 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5233 5234 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5235 while (iter.has_next()) { 5236 // Since discovery is not atomic for the CM ref processor, we 5237 // can see some null referent objects. 5238 iter.load_ptrs(DEBUG_ONLY(true)); 5239 oop ref = iter.obj(); 5240 5241 // This will filter nulls. 5242 if (iter.is_referent_alive()) { 5243 iter.make_referent_alive(); 5244 } 5245 iter.move_to_next(); 5246 } 5247 } 5248 5249 // Drain the queue - which may cause stealing 5250 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5251 drain_queue.do_void(); 5252 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5253 assert(pss.refs()->is_empty(), "should be"); 5254 } 5255 }; 5256 5257 // Weak Reference processing during an evacuation pause (part 1). 5258 void G1CollectedHeap::process_discovered_references() { 5259 double ref_proc_start = os::elapsedTime(); 5260 5261 ReferenceProcessor* rp = _ref_processor_stw; 5262 assert(rp->discovery_enabled(), "should have been enabled"); 5263 5264 // Any reference objects, in the collection set, that were 'discovered' 5265 // by the CM ref processor should have already been copied (either by 5266 // applying the external root copy closure to the discovered lists, or 5267 // by following an RSet entry). 5268 // 5269 // But some of the referents, that are in the collection set, that these 5270 // reference objects point to may not have been copied: the STW ref 5271 // processor would have seen that the reference object had already 5272 // been 'discovered' and would have skipped discovering the reference, 5273 // but would not have treated the reference object as a regular oop. 5274 // As a reult the copy closure would not have been applied to the 5275 // referent object. 5276 // 5277 // We need to explicitly copy these referent objects - the references 5278 // will be processed at the end of remarking. 5279 // 5280 // We also need to do this copying before we process the reference 5281 // objects discovered by the STW ref processor in case one of these 5282 // referents points to another object which is also referenced by an 5283 // object discovered by the STW ref processor. 5284 5285 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5286 workers()->active_workers() : 1); 5287 5288 assert(active_workers == workers()->active_workers(), 5289 "Need to reset active_workers"); 5290 set_par_threads(active_workers); 5291 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5292 5293 if (G1CollectedHeap::use_parallel_gc_threads()) { 5294 workers()->run_task(&keep_cm_referents); 5295 } else { 5296 keep_cm_referents.work(0); 5297 } 5298 5299 set_par_threads(0); 5300 5301 // Closure to test whether a referent is alive. 5302 G1STWIsAliveClosure is_alive(this); 5303 5304 // Even when parallel reference processing is enabled, the processing 5305 // of JNI refs is serial and performed serially by the current thread 5306 // rather than by a worker. The following PSS will be used for processing 5307 // JNI refs. 5308 5309 // Use only a single queue for this PSS. 5310 G1ParScanThreadState pss(this, 0); 5311 5312 // We do not embed a reference processor in the copying/scanning 5313 // closures while we're actually processing the discovered 5314 // reference objects. 5315 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5316 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5317 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5318 5319 pss.set_evac_closure(&scan_evac_cl); 5320 pss.set_evac_failure_closure(&evac_failure_cl); 5321 pss.set_partial_scan_closure(&partial_scan_cl); 5322 5323 assert(pss.refs()->is_empty(), "pre-condition"); 5324 5325 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5326 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5327 5328 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5329 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5330 5331 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5332 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5333 5334 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5335 // We also need to mark copied objects. 5336 copy_non_heap_cl = ©_mark_non_heap_cl; 5337 copy_perm_cl = ©_mark_perm_cl; 5338 } 5339 5340 // Keep alive closure. 5341 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5342 5343 // Serial Complete GC closure 5344 G1STWDrainQueueClosure drain_queue(this, &pss); 5345 5346 // Setup the soft refs policy... 5347 rp->setup_policy(false); 5348 5349 if (!rp->processing_is_mt()) { 5350 // Serial reference processing... 5351 rp->process_discovered_references(&is_alive, 5352 &keep_alive, 5353 &drain_queue, 5354 NULL); 5355 } else { 5356 // Parallel reference processing 5357 assert(rp->num_q() == active_workers, "sanity"); 5358 assert(active_workers <= rp->max_num_q(), "sanity"); 5359 5360 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5361 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5362 } 5363 5364 // We have completed copying any necessary live referent objects 5365 // (that were not copied during the actual pause) so we can 5366 // retire any active alloc buffers 5367 pss.retire_alloc_buffers(); 5368 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5369 5370 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5371 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0); 5372 } 5373 5374 // Weak Reference processing during an evacuation pause (part 2). 5375 void G1CollectedHeap::enqueue_discovered_references() { 5376 double ref_enq_start = os::elapsedTime(); 5377 5378 ReferenceProcessor* rp = _ref_processor_stw; 5379 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5380 5381 // Now enqueue any remaining on the discovered lists on to 5382 // the pending list. 5383 if (!rp->processing_is_mt()) { 5384 // Serial reference processing... 5385 rp->enqueue_discovered_references(); 5386 } else { 5387 // Parallel reference enqueuing 5388 5389 int active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5390 assert(active_workers == workers()->active_workers(), 5391 "Need to reset active_workers"); 5392 assert(rp->num_q() == active_workers, "sanity"); 5393 assert(active_workers <= rp->max_num_q(), "sanity"); 5394 5395 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5396 rp->enqueue_discovered_references(&par_task_executor); 5397 } 5398 5399 rp->verify_no_references_recorded(); 5400 assert(!rp->discovery_enabled(), "should have been disabled"); 5401 5402 // FIXME 5403 // CM's reference processing also cleans up the string and symbol tables. 5404 // Should we do that here also? We could, but it is a serial operation 5405 // and could signicantly increase the pause time. 5406 5407 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5408 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0); 5409 } 5410 5411 void G1CollectedHeap::evacuate_collection_set() { 5412 set_evacuation_failed(false); 5413 5414 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5415 concurrent_g1_refine()->set_use_cache(false); 5416 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5417 5418 int n_workers; 5419 if (G1CollectedHeap::use_parallel_gc_threads()) { 5420 n_workers = 5421 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5422 workers()->active_workers(), 5423 Threads::number_of_non_daemon_threads()); 5424 assert(UseDynamicNumberOfGCThreads || 5425 n_workers == workers()->total_workers(), 5426 "If not dynamic should be using all the workers"); 5427 set_par_threads(n_workers); 5428 } else { 5429 assert(n_par_threads() == 0, 5430 "Should be the original non-parallel value"); 5431 n_workers = 1; 5432 } 5433 workers()->set_active_workers(n_workers); 5434 5435 G1ParTask g1_par_task(this, _task_queues); 5436 5437 init_for_evac_failure(NULL); 5438 5439 rem_set()->prepare_for_younger_refs_iterate(true); 5440 5441 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5442 double start_par = os::elapsedTime(); 5443 5444 if (G1CollectedHeap::use_parallel_gc_threads()) { 5445 // The individual threads will set their evac-failure closures. 5446 StrongRootsScope srs(this); 5447 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5448 // These tasks use ShareHeap::_process_strong_tasks 5449 assert(UseDynamicNumberOfGCThreads || 5450 workers()->active_workers() == workers()->total_workers(), 5451 "If not dynamic should be using all the workers"); 5452 workers()->run_task(&g1_par_task); 5453 } else { 5454 StrongRootsScope srs(this); 5455 g1_par_task.work(0); 5456 } 5457 5458 double par_time = (os::elapsedTime() - start_par) * 1000.0; 5459 g1_policy()->record_par_time(par_time); 5460 5461 set_par_threads(0); 5462 5463 // Process any discovered reference objects - we have 5464 // to do this _before_ we retire the GC alloc regions 5465 // as we may have to copy some 'reachable' referent 5466 // objects (and their reachable sub-graphs) that were 5467 // not copied during the pause. 5468 process_discovered_references(); 5469 5470 // Weak root processing. 5471 // Note: when JSR 292 is enabled and code blobs can contain 5472 // non-perm oops then we will need to process the code blobs 5473 // here too. 5474 { 5475 G1STWIsAliveClosure is_alive(this); 5476 G1KeepAliveClosure keep_alive(this); 5477 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5478 } 5479 5480 release_gc_alloc_regions(); 5481 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5482 5483 concurrent_g1_refine()->clear_hot_cache(); 5484 concurrent_g1_refine()->set_use_cache(true); 5485 5486 finalize_for_evac_failure(); 5487 5488 // Must do this before clearing the per-region evac-failure flags 5489 // (which is currently done when we free the collection set). 5490 // We also only do this if marking is actually in progress and so 5491 // have to do this before we set the mark_in_progress flag at the 5492 // end of an initial mark pause. 5493 concurrent_mark()->complete_marking_in_collection_set(); 5494 5495 if (evacuation_failed()) { 5496 remove_self_forwarding_pointers(); 5497 if (PrintGCDetails) { 5498 gclog_or_tty->print(" (to-space overflow)"); 5499 } else if (PrintGC) { 5500 gclog_or_tty->print("--"); 5501 } 5502 } 5503 5504 // Enqueue any remaining references remaining on the STW 5505 // reference processor's discovered lists. We need to do 5506 // this after the card table is cleaned (and verified) as 5507 // the act of enqueuing entries on to the pending list 5508 // will log these updates (and dirty their associated 5509 // cards). We need these updates logged to update any 5510 // RSets. 5511 enqueue_discovered_references(); 5512 5513 if (G1DeferredRSUpdate) { 5514 RedirtyLoggedCardTableEntryFastClosure redirty; 5515 dirty_card_queue_set().set_closure(&redirty); 5516 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5517 5518 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5519 dcq.merge_bufferlists(&dirty_card_queue_set()); 5520 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5521 } 5522 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5523 } 5524 5525 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5526 size_t* pre_used, 5527 FreeRegionList* free_list, 5528 OldRegionSet* old_proxy_set, 5529 HumongousRegionSet* humongous_proxy_set, 5530 HRRSCleanupTask* hrrs_cleanup_task, 5531 bool par) { 5532 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5533 if (hr->isHumongous()) { 5534 assert(hr->startsHumongous(), "we should only see starts humongous"); 5535 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5536 } else { 5537 _old_set.remove_with_proxy(hr, old_proxy_set); 5538 free_region(hr, pre_used, free_list, par); 5539 } 5540 } else { 5541 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5542 } 5543 } 5544 5545 void G1CollectedHeap::free_region(HeapRegion* hr, 5546 size_t* pre_used, 5547 FreeRegionList* free_list, 5548 bool par) { 5549 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5550 assert(!hr->is_empty(), "the region should not be empty"); 5551 assert(free_list != NULL, "pre-condition"); 5552 5553 *pre_used += hr->used(); 5554 hr->hr_clear(par, true /* clear_space */); 5555 free_list->add_as_head(hr); 5556 } 5557 5558 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5559 size_t* pre_used, 5560 FreeRegionList* free_list, 5561 HumongousRegionSet* humongous_proxy_set, 5562 bool par) { 5563 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5564 assert(free_list != NULL, "pre-condition"); 5565 assert(humongous_proxy_set != NULL, "pre-condition"); 5566 5567 size_t hr_used = hr->used(); 5568 size_t hr_capacity = hr->capacity(); 5569 size_t hr_pre_used = 0; 5570 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5571 hr->set_notHumongous(); 5572 free_region(hr, &hr_pre_used, free_list, par); 5573 5574 size_t i = hr->hrs_index() + 1; 5575 size_t num = 1; 5576 while (i < n_regions()) { 5577 HeapRegion* curr_hr = region_at(i); 5578 if (!curr_hr->continuesHumongous()) { 5579 break; 5580 } 5581 curr_hr->set_notHumongous(); 5582 free_region(curr_hr, &hr_pre_used, free_list, par); 5583 num += 1; 5584 i += 1; 5585 } 5586 assert(hr_pre_used == hr_used, 5587 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5588 "should be the same", hr_pre_used, hr_used)); 5589 *pre_used += hr_pre_used; 5590 } 5591 5592 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5593 FreeRegionList* free_list, 5594 OldRegionSet* old_proxy_set, 5595 HumongousRegionSet* humongous_proxy_set, 5596 bool par) { 5597 if (pre_used > 0) { 5598 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5599 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5600 assert(_summary_bytes_used >= pre_used, 5601 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5602 "should be >= pre_used: "SIZE_FORMAT, 5603 _summary_bytes_used, pre_used)); 5604 _summary_bytes_used -= pre_used; 5605 } 5606 if (free_list != NULL && !free_list->is_empty()) { 5607 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5608 _free_list.add_as_head(free_list); 5609 } 5610 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5611 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5612 _old_set.update_from_proxy(old_proxy_set); 5613 } 5614 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5615 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5616 _humongous_set.update_from_proxy(humongous_proxy_set); 5617 } 5618 } 5619 5620 class G1ParCleanupCTTask : public AbstractGangTask { 5621 CardTableModRefBS* _ct_bs; 5622 G1CollectedHeap* _g1h; 5623 HeapRegion* volatile _su_head; 5624 public: 5625 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5626 G1CollectedHeap* g1h) : 5627 AbstractGangTask("G1 Par Cleanup CT Task"), 5628 _ct_bs(ct_bs), _g1h(g1h) { } 5629 5630 void work(int i) { 5631 HeapRegion* r; 5632 while (r = _g1h->pop_dirty_cards_region()) { 5633 clear_cards(r); 5634 } 5635 } 5636 5637 void clear_cards(HeapRegion* r) { 5638 // Cards of the survivors should have already been dirtied. 5639 if (!r->is_survivor()) { 5640 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5641 } 5642 } 5643 }; 5644 5645 #ifndef PRODUCT 5646 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5647 G1CollectedHeap* _g1h; 5648 CardTableModRefBS* _ct_bs; 5649 public: 5650 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5651 : _g1h(g1h), _ct_bs(ct_bs) { } 5652 virtual bool doHeapRegion(HeapRegion* r) { 5653 if (r->is_survivor()) { 5654 _g1h->verify_dirty_region(r); 5655 } else { 5656 _g1h->verify_not_dirty_region(r); 5657 } 5658 return false; 5659 } 5660 }; 5661 5662 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5663 // All of the region should be clean. 5664 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5665 MemRegion mr(hr->bottom(), hr->end()); 5666 ct_bs->verify_not_dirty_region(mr); 5667 } 5668 5669 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5670 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5671 // dirty allocated blocks as they allocate them. The thread that 5672 // retires each region and replaces it with a new one will do a 5673 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5674 // not dirty that area (one less thing to have to do while holding 5675 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5676 // is dirty. 5677 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5678 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5679 ct_bs->verify_dirty_region(mr); 5680 } 5681 5682 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5683 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5684 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5685 verify_dirty_region(hr); 5686 } 5687 } 5688 5689 void G1CollectedHeap::verify_dirty_young_regions() { 5690 verify_dirty_young_list(_young_list->first_region()); 5691 verify_dirty_young_list(_young_list->first_survivor_region()); 5692 } 5693 #endif 5694 5695 void G1CollectedHeap::cleanUpCardTable() { 5696 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5697 double start = os::elapsedTime(); 5698 5699 { 5700 // Iterate over the dirty cards region list. 5701 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5702 5703 if (ParallelGCThreads > 0) { 5704 set_par_threads(workers()->total_workers()); 5705 workers()->run_task(&cleanup_task); 5706 set_par_threads(0); 5707 } else { 5708 while (_dirty_cards_region_list) { 5709 HeapRegion* r = _dirty_cards_region_list; 5710 cleanup_task.clear_cards(r); 5711 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5712 if (_dirty_cards_region_list == r) { 5713 // The last region. 5714 _dirty_cards_region_list = NULL; 5715 } 5716 r->set_next_dirty_cards_region(NULL); 5717 } 5718 } 5719 #ifndef PRODUCT 5720 if (G1VerifyCTCleanup || VerifyAfterGC) { 5721 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5722 heap_region_iterate(&cleanup_verifier); 5723 } 5724 #endif 5725 } 5726 5727 double elapsed = os::elapsedTime() - start; 5728 g1_policy()->record_clear_ct_time(elapsed * 1000.0); 5729 } 5730 5731 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5732 size_t pre_used = 0; 5733 FreeRegionList local_free_list("Local List for CSet Freeing"); 5734 5735 double young_time_ms = 0.0; 5736 double non_young_time_ms = 0.0; 5737 5738 // Since the collection set is a superset of the the young list, 5739 // all we need to do to clear the young list is clear its 5740 // head and length, and unlink any young regions in the code below 5741 _young_list->clear(); 5742 5743 G1CollectorPolicy* policy = g1_policy(); 5744 5745 double start_sec = os::elapsedTime(); 5746 bool non_young = true; 5747 5748 HeapRegion* cur = cs_head; 5749 int age_bound = -1; 5750 size_t rs_lengths = 0; 5751 5752 while (cur != NULL) { 5753 assert(!is_on_master_free_list(cur), "sanity"); 5754 if (non_young) { 5755 if (cur->is_young()) { 5756 double end_sec = os::elapsedTime(); 5757 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5758 non_young_time_ms += elapsed_ms; 5759 5760 start_sec = os::elapsedTime(); 5761 non_young = false; 5762 } 5763 } else { 5764 if (!cur->is_young()) { 5765 double end_sec = os::elapsedTime(); 5766 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5767 young_time_ms += elapsed_ms; 5768 5769 start_sec = os::elapsedTime(); 5770 non_young = true; 5771 } 5772 } 5773 5774 rs_lengths += cur->rem_set()->occupied(); 5775 5776 HeapRegion* next = cur->next_in_collection_set(); 5777 assert(cur->in_collection_set(), "bad CS"); 5778 cur->set_next_in_collection_set(NULL); 5779 cur->set_in_collection_set(false); 5780 5781 if (cur->is_young()) { 5782 int index = cur->young_index_in_cset(); 5783 assert(index != -1, "invariant"); 5784 assert((size_t) index < policy->young_cset_region_length(), "invariant"); 5785 size_t words_survived = _surviving_young_words[index]; 5786 cur->record_surv_words_in_group(words_survived); 5787 5788 // At this point the we have 'popped' cur from the collection set 5789 // (linked via next_in_collection_set()) but it is still in the 5790 // young list (linked via next_young_region()). Clear the 5791 // _next_young_region field. 5792 cur->set_next_young_region(NULL); 5793 } else { 5794 int index = cur->young_index_in_cset(); 5795 assert(index == -1, "invariant"); 5796 } 5797 5798 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5799 (!cur->is_young() && cur->young_index_in_cset() == -1), 5800 "invariant" ); 5801 5802 if (!cur->evacuation_failed()) { 5803 MemRegion used_mr = cur->used_region(); 5804 5805 // And the region is empty. 5806 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5807 5808 // If marking is in progress then clear any objects marked in 5809 // the current region. Note mark_in_progress() returns false, 5810 // even during an initial mark pause, until the set_marking_started() 5811 // call which takes place later in the pause. 5812 if (mark_in_progress()) { 5813 assert(!g1_policy()->during_initial_mark_pause(), "sanity"); 5814 _cm->nextMarkBitMap()->clearRange(used_mr); 5815 } 5816 5817 free_region(cur, &pre_used, &local_free_list, false /* par */); 5818 } else { 5819 cur->uninstall_surv_rate_group(); 5820 if (cur->is_young()) { 5821 cur->set_young_index_in_cset(-1); 5822 } 5823 cur->set_not_young(); 5824 cur->set_evacuation_failed(false); 5825 // The region is now considered to be old. 5826 _old_set.add(cur); 5827 } 5828 cur = next; 5829 } 5830 5831 policy->record_max_rs_lengths(rs_lengths); 5832 policy->cset_regions_freed(); 5833 5834 double end_sec = os::elapsedTime(); 5835 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5836 5837 if (non_young) { 5838 non_young_time_ms += elapsed_ms; 5839 } else { 5840 young_time_ms += elapsed_ms; 5841 } 5842 5843 update_sets_after_freeing_regions(pre_used, &local_free_list, 5844 NULL /* old_proxy_set */, 5845 NULL /* humongous_proxy_set */, 5846 false /* par */); 5847 policy->record_young_free_cset_time_ms(young_time_ms); 5848 policy->record_non_young_free_cset_time_ms(non_young_time_ms); 5849 } 5850 5851 // This routine is similar to the above but does not record 5852 // any policy statistics or update free lists; we are abandoning 5853 // the current incremental collection set in preparation of a 5854 // full collection. After the full GC we will start to build up 5855 // the incremental collection set again. 5856 // This is only called when we're doing a full collection 5857 // and is immediately followed by the tearing down of the young list. 5858 5859 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5860 HeapRegion* cur = cs_head; 5861 5862 while (cur != NULL) { 5863 HeapRegion* next = cur->next_in_collection_set(); 5864 assert(cur->in_collection_set(), "bad CS"); 5865 cur->set_next_in_collection_set(NULL); 5866 cur->set_in_collection_set(false); 5867 cur->set_young_index_in_cset(-1); 5868 cur = next; 5869 } 5870 } 5871 5872 void G1CollectedHeap::set_free_regions_coming() { 5873 if (G1ConcRegionFreeingVerbose) { 5874 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5875 "setting free regions coming"); 5876 } 5877 5878 assert(!free_regions_coming(), "pre-condition"); 5879 _free_regions_coming = true; 5880 } 5881 5882 void G1CollectedHeap::reset_free_regions_coming() { 5883 { 5884 assert(free_regions_coming(), "pre-condition"); 5885 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5886 _free_regions_coming = false; 5887 SecondaryFreeList_lock->notify_all(); 5888 } 5889 5890 if (G1ConcRegionFreeingVerbose) { 5891 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5892 "reset free regions coming"); 5893 } 5894 } 5895 5896 void G1CollectedHeap::wait_while_free_regions_coming() { 5897 // Most of the time we won't have to wait, so let's do a quick test 5898 // first before we take the lock. 5899 if (!free_regions_coming()) { 5900 return; 5901 } 5902 5903 if (G1ConcRegionFreeingVerbose) { 5904 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5905 "waiting for free regions"); 5906 } 5907 5908 { 5909 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5910 while (free_regions_coming()) { 5911 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 5912 } 5913 } 5914 5915 if (G1ConcRegionFreeingVerbose) { 5916 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5917 "done waiting for free regions"); 5918 } 5919 } 5920 5921 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 5922 assert(heap_lock_held_for_gc(), 5923 "the heap lock should already be held by or for this thread"); 5924 _young_list->push_region(hr); 5925 } 5926 5927 class NoYoungRegionsClosure: public HeapRegionClosure { 5928 private: 5929 bool _success; 5930 public: 5931 NoYoungRegionsClosure() : _success(true) { } 5932 bool doHeapRegion(HeapRegion* r) { 5933 if (r->is_young()) { 5934 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 5935 r->bottom(), r->end()); 5936 _success = false; 5937 } 5938 return false; 5939 } 5940 bool success() { return _success; } 5941 }; 5942 5943 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 5944 bool ret = _young_list->check_list_empty(check_sample); 5945 5946 if (check_heap) { 5947 NoYoungRegionsClosure closure; 5948 heap_region_iterate(&closure); 5949 ret = ret && closure.success(); 5950 } 5951 5952 return ret; 5953 } 5954 5955 class TearDownRegionSetsClosure : public HeapRegionClosure { 5956 private: 5957 OldRegionSet *_old_set; 5958 5959 public: 5960 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 5961 5962 bool doHeapRegion(HeapRegion* r) { 5963 if (r->is_empty()) { 5964 // We ignore empty regions, we'll empty the free list afterwards 5965 } else if (r->is_young()) { 5966 // We ignore young regions, we'll empty the young list afterwards 5967 } else if (r->isHumongous()) { 5968 // We ignore humongous regions, we're not tearing down the 5969 // humongous region set 5970 } else { 5971 // The rest should be old 5972 _old_set->remove(r); 5973 } 5974 return false; 5975 } 5976 5977 ~TearDownRegionSetsClosure() { 5978 assert(_old_set->is_empty(), "post-condition"); 5979 } 5980 }; 5981 5982 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 5983 assert_at_safepoint(true /* should_be_vm_thread */); 5984 5985 if (!free_list_only) { 5986 TearDownRegionSetsClosure cl(&_old_set); 5987 heap_region_iterate(&cl); 5988 5989 // Need to do this after the heap iteration to be able to 5990 // recognize the young regions and ignore them during the iteration. 5991 _young_list->empty_list(); 5992 } 5993 _free_list.remove_all(); 5994 } 5995 5996 class RebuildRegionSetsClosure : public HeapRegionClosure { 5997 private: 5998 bool _free_list_only; 5999 OldRegionSet* _old_set; 6000 FreeRegionList* _free_list; 6001 size_t _total_used; 6002 6003 public: 6004 RebuildRegionSetsClosure(bool free_list_only, 6005 OldRegionSet* old_set, FreeRegionList* free_list) : 6006 _free_list_only(free_list_only), 6007 _old_set(old_set), _free_list(free_list), _total_used(0) { 6008 assert(_free_list->is_empty(), "pre-condition"); 6009 if (!free_list_only) { 6010 assert(_old_set->is_empty(), "pre-condition"); 6011 } 6012 } 6013 6014 bool doHeapRegion(HeapRegion* r) { 6015 if (r->continuesHumongous()) { 6016 return false; 6017 } 6018 6019 if (r->is_empty()) { 6020 // Add free regions to the free list 6021 _free_list->add_as_tail(r); 6022 } else if (!_free_list_only) { 6023 assert(!r->is_young(), "we should not come across young regions"); 6024 6025 if (r->isHumongous()) { 6026 // We ignore humongous regions, we left the humongous set unchanged 6027 } else { 6028 // The rest should be old, add them to the old set 6029 _old_set->add(r); 6030 } 6031 _total_used += r->used(); 6032 } 6033 6034 return false; 6035 } 6036 6037 size_t total_used() { 6038 return _total_used; 6039 } 6040 }; 6041 6042 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6043 assert_at_safepoint(true /* should_be_vm_thread */); 6044 6045 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6046 heap_region_iterate(&cl); 6047 6048 if (!free_list_only) { 6049 _summary_bytes_used = cl.total_used(); 6050 } 6051 assert(_summary_bytes_used == recalculate_used(), 6052 err_msg("inconsistent _summary_bytes_used, " 6053 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6054 _summary_bytes_used, recalculate_used())); 6055 } 6056 6057 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6058 _refine_cte_cl->set_concurrent(concurrent); 6059 } 6060 6061 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6062 HeapRegion* hr = heap_region_containing(p); 6063 if (hr == NULL) { 6064 return is_in_permanent(p); 6065 } else { 6066 return hr->is_in(p); 6067 } 6068 } 6069 6070 // Methods for the mutator alloc region 6071 6072 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6073 bool force) { 6074 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6075 assert(!force || g1_policy()->can_expand_young_list(), 6076 "if force is true we should be able to expand the young list"); 6077 bool young_list_full = g1_policy()->is_young_list_full(); 6078 if (force || !young_list_full) { 6079 HeapRegion* new_alloc_region = new_region(word_size, 6080 false /* do_expand */); 6081 if (new_alloc_region != NULL) { 6082 set_region_short_lived_locked(new_alloc_region); 6083 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6084 return new_alloc_region; 6085 } 6086 } 6087 return NULL; 6088 } 6089 6090 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6091 size_t allocated_bytes) { 6092 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6093 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6094 6095 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6096 _summary_bytes_used += allocated_bytes; 6097 _hr_printer.retire(alloc_region); 6098 // We update the eden sizes here, when the region is retired, 6099 // instead of when it's allocated, since this is the point that its 6100 // used space has been recored in _summary_bytes_used. 6101 g1mm()->update_eden_size(); 6102 } 6103 6104 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6105 bool force) { 6106 return _g1h->new_mutator_alloc_region(word_size, force); 6107 } 6108 6109 void G1CollectedHeap::set_par_threads() { 6110 // Don't change the number of workers. Use the value previously set 6111 // in the workgroup. 6112 int n_workers = workers()->active_workers(); 6113 assert(UseDynamicNumberOfGCThreads || 6114 n_workers == workers()->total_workers(), 6115 "Otherwise should be using the total number of workers"); 6116 if (n_workers == 0) { 6117 assert(false, "Should have been set in prior evacuation pause."); 6118 n_workers = ParallelGCThreads; 6119 workers()->set_active_workers(n_workers); 6120 } 6121 set_par_threads(n_workers); 6122 } 6123 6124 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6125 size_t allocated_bytes) { 6126 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6127 } 6128 6129 // Methods for the GC alloc regions 6130 6131 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6132 size_t count, 6133 GCAllocPurpose ap) { 6134 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6135 6136 if (count < g1_policy()->max_regions(ap)) { 6137 HeapRegion* new_alloc_region = new_region(word_size, 6138 true /* do_expand */); 6139 if (new_alloc_region != NULL) { 6140 // We really only need to do this for old regions given that we 6141 // should never scan survivors. But it doesn't hurt to do it 6142 // for survivors too. 6143 new_alloc_region->set_saved_mark(); 6144 if (ap == GCAllocForSurvived) { 6145 new_alloc_region->set_survivor(); 6146 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6147 } else { 6148 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6149 } 6150 return new_alloc_region; 6151 } else { 6152 g1_policy()->note_alloc_region_limit_reached(ap); 6153 } 6154 } 6155 return NULL; 6156 } 6157 6158 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6159 size_t allocated_bytes, 6160 GCAllocPurpose ap) { 6161 alloc_region->note_end_of_copying(); 6162 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6163 if (ap == GCAllocForSurvived) { 6164 young_list()->add_survivor_region(alloc_region); 6165 } else { 6166 _old_set.add(alloc_region); 6167 } 6168 _hr_printer.retire(alloc_region); 6169 } 6170 6171 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6172 bool force) { 6173 assert(!force, "not supported for GC alloc regions"); 6174 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6175 } 6176 6177 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6178 size_t allocated_bytes) { 6179 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6180 GCAllocForSurvived); 6181 } 6182 6183 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6184 bool force) { 6185 assert(!force, "not supported for GC alloc regions"); 6186 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6187 } 6188 6189 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6190 size_t allocated_bytes) { 6191 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6192 GCAllocForTenured); 6193 } 6194 // Heap region set verification 6195 6196 class VerifyRegionListsClosure : public HeapRegionClosure { 6197 private: 6198 FreeRegionList* _free_list; 6199 OldRegionSet* _old_set; 6200 HumongousRegionSet* _humongous_set; 6201 size_t _region_count; 6202 6203 public: 6204 VerifyRegionListsClosure(OldRegionSet* old_set, 6205 HumongousRegionSet* humongous_set, 6206 FreeRegionList* free_list) : 6207 _old_set(old_set), _humongous_set(humongous_set), 6208 _free_list(free_list), _region_count(0) { } 6209 6210 size_t region_count() { return _region_count; } 6211 6212 bool doHeapRegion(HeapRegion* hr) { 6213 _region_count += 1; 6214 6215 if (hr->continuesHumongous()) { 6216 return false; 6217 } 6218 6219 if (hr->is_young()) { 6220 // TODO 6221 } else if (hr->startsHumongous()) { 6222 _humongous_set->verify_next_region(hr); 6223 } else if (hr->is_empty()) { 6224 _free_list->verify_next_region(hr); 6225 } else { 6226 _old_set->verify_next_region(hr); 6227 } 6228 return false; 6229 } 6230 }; 6231 6232 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index, 6233 HeapWord* bottom) { 6234 HeapWord* end = bottom + HeapRegion::GrainWords; 6235 MemRegion mr(bottom, end); 6236 assert(_g1_reserved.contains(mr), "invariant"); 6237 // This might return NULL if the allocation fails 6238 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6239 } 6240 6241 void G1CollectedHeap::verify_region_sets() { 6242 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6243 6244 // First, check the explicit lists. 6245 _free_list.verify(); 6246 { 6247 // Given that a concurrent operation might be adding regions to 6248 // the secondary free list we have to take the lock before 6249 // verifying it. 6250 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6251 _secondary_free_list.verify(); 6252 } 6253 _old_set.verify(); 6254 _humongous_set.verify(); 6255 6256 // If a concurrent region freeing operation is in progress it will 6257 // be difficult to correctly attributed any free regions we come 6258 // across to the correct free list given that they might belong to 6259 // one of several (free_list, secondary_free_list, any local lists, 6260 // etc.). So, if that's the case we will skip the rest of the 6261 // verification operation. Alternatively, waiting for the concurrent 6262 // operation to complete will have a non-trivial effect on the GC's 6263 // operation (no concurrent operation will last longer than the 6264 // interval between two calls to verification) and it might hide 6265 // any issues that we would like to catch during testing. 6266 if (free_regions_coming()) { 6267 return; 6268 } 6269 6270 // Make sure we append the secondary_free_list on the free_list so 6271 // that all free regions we will come across can be safely 6272 // attributed to the free_list. 6273 append_secondary_free_list_if_not_empty_with_lock(); 6274 6275 // Finally, make sure that the region accounting in the lists is 6276 // consistent with what we see in the heap. 6277 _old_set.verify_start(); 6278 _humongous_set.verify_start(); 6279 _free_list.verify_start(); 6280 6281 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6282 heap_region_iterate(&cl); 6283 6284 _old_set.verify_end(); 6285 _humongous_set.verify_end(); 6286 _free_list.verify_end(); 6287 }