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_time_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_time_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 1877 clear_cset_start_regions(); 1878 1879 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1880 } 1881 1882 jint G1CollectedHeap::initialize() { 1883 CollectedHeap::pre_initialize(); 1884 os::enable_vtime(); 1885 1886 // Necessary to satisfy locking discipline assertions. 1887 1888 MutexLocker x(Heap_lock); 1889 1890 // We have to initialize the printer before committing the heap, as 1891 // it will be used then. 1892 _hr_printer.set_active(G1PrintHeapRegions); 1893 1894 // While there are no constraints in the GC code that HeapWordSize 1895 // be any particular value, there are multiple other areas in the 1896 // system which believe this to be true (e.g. oop->object_size in some 1897 // cases incorrectly returns the size in wordSize units rather than 1898 // HeapWordSize). 1899 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1900 1901 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1902 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1903 1904 // Ensure that the sizes are properly aligned. 1905 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1906 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1907 1908 _cg1r = new ConcurrentG1Refine(); 1909 1910 // Reserve the maximum. 1911 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1912 // Includes the perm-gen. 1913 1914 // When compressed oops are enabled, the preferred heap base 1915 // is calculated by subtracting the requested size from the 1916 // 32Gb boundary and using the result as the base address for 1917 // heap reservation. If the requested size is not aligned to 1918 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1919 // into the ReservedHeapSpace constructor) then the actual 1920 // base of the reserved heap may end up differing from the 1921 // address that was requested (i.e. the preferred heap base). 1922 // If this happens then we could end up using a non-optimal 1923 // compressed oops mode. 1924 1925 // Since max_byte_size is aligned to the size of a heap region (checked 1926 // above), we also need to align the perm gen size as it might not be. 1927 const size_t total_reserved = max_byte_size + 1928 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1929 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1930 1931 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1932 1933 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1934 UseLargePages, addr); 1935 1936 if (UseCompressedOops) { 1937 if (addr != NULL && !heap_rs.is_reserved()) { 1938 // Failed to reserve at specified address - the requested memory 1939 // region is taken already, for example, by 'java' launcher. 1940 // Try again to reserver heap higher. 1941 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 1942 1943 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 1944 UseLargePages, addr); 1945 1946 if (addr != NULL && !heap_rs0.is_reserved()) { 1947 // Failed to reserve at specified address again - give up. 1948 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 1949 assert(addr == NULL, ""); 1950 1951 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 1952 UseLargePages, addr); 1953 heap_rs = heap_rs1; 1954 } else { 1955 heap_rs = heap_rs0; 1956 } 1957 } 1958 } 1959 1960 if (!heap_rs.is_reserved()) { 1961 vm_exit_during_initialization("Could not reserve enough space for object heap"); 1962 return JNI_ENOMEM; 1963 } 1964 1965 // It is important to do this in a way such that concurrent readers can't 1966 // temporarily think somethings in the heap. (I've actually seen this 1967 // happen in asserts: DLD.) 1968 _reserved.set_word_size(0); 1969 _reserved.set_start((HeapWord*)heap_rs.base()); 1970 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 1971 1972 _expansion_regions = max_byte_size/HeapRegion::GrainBytes; 1973 1974 // Create the gen rem set (and barrier set) for the entire reserved region. 1975 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 1976 set_barrier_set(rem_set()->bs()); 1977 if (barrier_set()->is_a(BarrierSet::ModRef)) { 1978 _mr_bs = (ModRefBarrierSet*)_barrier_set; 1979 } else { 1980 vm_exit_during_initialization("G1 requires a mod ref bs."); 1981 return JNI_ENOMEM; 1982 } 1983 1984 // Also create a G1 rem set. 1985 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 1986 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 1987 } else { 1988 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 1989 return JNI_ENOMEM; 1990 } 1991 1992 // Carve out the G1 part of the heap. 1993 1994 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 1995 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 1996 g1_rs.size()/HeapWordSize); 1997 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 1998 1999 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2000 2001 _g1_storage.initialize(g1_rs, 0); 2002 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2003 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2004 (HeapWord*) _g1_reserved.end(), 2005 _expansion_regions); 2006 2007 // 6843694 - ensure that the maximum region index can fit 2008 // in the remembered set structures. 2009 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2010 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2011 2012 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2013 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2014 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2015 "too many cards per region"); 2016 2017 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2018 2019 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2020 heap_word_size(init_byte_size)); 2021 2022 _g1h = this; 2023 2024 _in_cset_fast_test_length = max_regions(); 2025 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length); 2026 2027 // We're biasing _in_cset_fast_test to avoid subtracting the 2028 // beginning of the heap every time we want to index; basically 2029 // it's the same with what we do with the card table. 2030 _in_cset_fast_test = _in_cset_fast_test_base - 2031 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2032 2033 // Clear the _cset_fast_test bitmap in anticipation of adding 2034 // regions to the incremental collection set for the first 2035 // evacuation pause. 2036 clear_cset_fast_test(); 2037 2038 // Create the ConcurrentMark data structure and thread. 2039 // (Must do this late, so that "max_regions" is defined.) 2040 _cm = new ConcurrentMark(heap_rs, (int) max_regions()); 2041 _cmThread = _cm->cmThread(); 2042 2043 // Initialize the from_card cache structure of HeapRegionRemSet. 2044 HeapRegionRemSet::init_heap(max_regions()); 2045 2046 // Now expand into the initial heap size. 2047 if (!expand(init_byte_size)) { 2048 vm_exit_during_initialization("Failed to allocate initial heap."); 2049 return JNI_ENOMEM; 2050 } 2051 2052 // Perform any initialization actions delegated to the policy. 2053 g1_policy()->init(); 2054 2055 _refine_cte_cl = 2056 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2057 g1_rem_set(), 2058 concurrent_g1_refine()); 2059 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2060 2061 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2062 SATB_Q_FL_lock, 2063 G1SATBProcessCompletedThreshold, 2064 Shared_SATB_Q_lock); 2065 2066 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2067 DirtyCardQ_FL_lock, 2068 concurrent_g1_refine()->yellow_zone(), 2069 concurrent_g1_refine()->red_zone(), 2070 Shared_DirtyCardQ_lock); 2071 2072 if (G1DeferredRSUpdate) { 2073 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2074 DirtyCardQ_FL_lock, 2075 -1, // never trigger processing 2076 -1, // no limit on length 2077 Shared_DirtyCardQ_lock, 2078 &JavaThread::dirty_card_queue_set()); 2079 } 2080 2081 // Initialize the card queue set used to hold cards containing 2082 // references into the collection set. 2083 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2084 DirtyCardQ_FL_lock, 2085 -1, // never trigger processing 2086 -1, // no limit on length 2087 Shared_DirtyCardQ_lock, 2088 &JavaThread::dirty_card_queue_set()); 2089 2090 // In case we're keeping closure specialization stats, initialize those 2091 // counts and that mechanism. 2092 SpecializationStats::clear(); 2093 2094 // Do later initialization work for concurrent refinement. 2095 _cg1r->init(); 2096 2097 // Here we allocate the dummy full region that is required by the 2098 // G1AllocRegion class. If we don't pass an address in the reserved 2099 // space here, lots of asserts fire. 2100 2101 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2102 _g1_reserved.start()); 2103 // We'll re-use the same region whether the alloc region will 2104 // require BOT updates or not and, if it doesn't, then a non-young 2105 // region will complain that it cannot support allocations without 2106 // BOT updates. So we'll tag the dummy region as young to avoid that. 2107 dummy_region->set_young(); 2108 // Make sure it's full. 2109 dummy_region->set_top(dummy_region->end()); 2110 G1AllocRegion::setup(this, dummy_region); 2111 2112 init_mutator_alloc_region(); 2113 2114 // Do create of the monitoring and management support so that 2115 // values in the heap have been properly initialized. 2116 _g1mm = new G1MonitoringSupport(this); 2117 2118 return JNI_OK; 2119 } 2120 2121 void G1CollectedHeap::ref_processing_init() { 2122 // Reference processing in G1 currently works as follows: 2123 // 2124 // * There are two reference processor instances. One is 2125 // used to record and process discovered references 2126 // during concurrent marking; the other is used to 2127 // record and process references during STW pauses 2128 // (both full and incremental). 2129 // * Both ref processors need to 'span' the entire heap as 2130 // the regions in the collection set may be dotted around. 2131 // 2132 // * For the concurrent marking ref processor: 2133 // * Reference discovery is enabled at initial marking. 2134 // * Reference discovery is disabled and the discovered 2135 // references processed etc during remarking. 2136 // * Reference discovery is MT (see below). 2137 // * Reference discovery requires a barrier (see below). 2138 // * Reference processing may or may not be MT 2139 // (depending on the value of ParallelRefProcEnabled 2140 // and ParallelGCThreads). 2141 // * A full GC disables reference discovery by the CM 2142 // ref processor and abandons any entries on it's 2143 // discovered lists. 2144 // 2145 // * For the STW processor: 2146 // * Non MT discovery is enabled at the start of a full GC. 2147 // * Processing and enqueueing during a full GC is non-MT. 2148 // * During a full GC, references are processed after marking. 2149 // 2150 // * Discovery (may or may not be MT) is enabled at the start 2151 // of an incremental evacuation pause. 2152 // * References are processed near the end of a STW evacuation pause. 2153 // * For both types of GC: 2154 // * Discovery is atomic - i.e. not concurrent. 2155 // * Reference discovery will not need a barrier. 2156 2157 SharedHeap::ref_processing_init(); 2158 MemRegion mr = reserved_region(); 2159 2160 // Concurrent Mark ref processor 2161 _ref_processor_cm = 2162 new ReferenceProcessor(mr, // span 2163 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2164 // mt processing 2165 (int) ParallelGCThreads, 2166 // degree of mt processing 2167 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2168 // mt discovery 2169 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2170 // degree of mt discovery 2171 false, 2172 // Reference discovery is not atomic 2173 &_is_alive_closure_cm, 2174 // is alive closure 2175 // (for efficiency/performance) 2176 true); 2177 // Setting next fields of discovered 2178 // lists requires a barrier. 2179 2180 // STW ref processor 2181 _ref_processor_stw = 2182 new ReferenceProcessor(mr, // span 2183 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2184 // mt processing 2185 MAX2((int)ParallelGCThreads, 1), 2186 // degree of mt processing 2187 (ParallelGCThreads > 1), 2188 // mt discovery 2189 MAX2((int)ParallelGCThreads, 1), 2190 // degree of mt discovery 2191 true, 2192 // Reference discovery is atomic 2193 &_is_alive_closure_stw, 2194 // is alive closure 2195 // (for efficiency/performance) 2196 false); 2197 // Setting next fields of discovered 2198 // lists requires a barrier. 2199 } 2200 2201 size_t G1CollectedHeap::capacity() const { 2202 return _g1_committed.byte_size(); 2203 } 2204 2205 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2206 DirtyCardQueue* into_cset_dcq, 2207 bool concurrent, 2208 int worker_i) { 2209 // Clean cards in the hot card cache 2210 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2211 2212 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2213 int n_completed_buffers = 0; 2214 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2215 n_completed_buffers++; 2216 } 2217 g1_policy()->record_update_rs_processed_buffers(worker_i, 2218 (double) n_completed_buffers); 2219 dcqs.clear_n_completed_buffers(); 2220 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2221 } 2222 2223 2224 // Computes the sum of the storage used by the various regions. 2225 2226 size_t G1CollectedHeap::used() const { 2227 assert(Heap_lock->owner() != NULL, 2228 "Should be owned on this thread's behalf."); 2229 size_t result = _summary_bytes_used; 2230 // Read only once in case it is set to NULL concurrently 2231 HeapRegion* hr = _mutator_alloc_region.get(); 2232 if (hr != NULL) 2233 result += hr->used(); 2234 return result; 2235 } 2236 2237 size_t G1CollectedHeap::used_unlocked() const { 2238 size_t result = _summary_bytes_used; 2239 return result; 2240 } 2241 2242 class SumUsedClosure: public HeapRegionClosure { 2243 size_t _used; 2244 public: 2245 SumUsedClosure() : _used(0) {} 2246 bool doHeapRegion(HeapRegion* r) { 2247 if (!r->continuesHumongous()) { 2248 _used += r->used(); 2249 } 2250 return false; 2251 } 2252 size_t result() { return _used; } 2253 }; 2254 2255 size_t G1CollectedHeap::recalculate_used() const { 2256 SumUsedClosure blk; 2257 heap_region_iterate(&blk); 2258 return blk.result(); 2259 } 2260 2261 size_t G1CollectedHeap::unsafe_max_alloc() { 2262 if (free_regions() > 0) return HeapRegion::GrainBytes; 2263 // otherwise, is there space in the current allocation region? 2264 2265 // We need to store the current allocation region in a local variable 2266 // here. The problem is that this method doesn't take any locks and 2267 // there may be other threads which overwrite the current allocation 2268 // region field. attempt_allocation(), for example, sets it to NULL 2269 // and this can happen *after* the NULL check here but before the call 2270 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2271 // to be a problem in the optimized build, since the two loads of the 2272 // current allocation region field are optimized away. 2273 HeapRegion* hr = _mutator_alloc_region.get(); 2274 if (hr == NULL) { 2275 return 0; 2276 } 2277 return hr->free(); 2278 } 2279 2280 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2281 return 2282 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) || 2283 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent)); 2284 } 2285 2286 #ifndef PRODUCT 2287 void G1CollectedHeap::allocate_dummy_regions() { 2288 // Let's fill up most of the region 2289 size_t word_size = HeapRegion::GrainWords - 1024; 2290 // And as a result the region we'll allocate will be humongous. 2291 guarantee(isHumongous(word_size), "sanity"); 2292 2293 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2294 // Let's use the existing mechanism for the allocation 2295 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2296 if (dummy_obj != NULL) { 2297 MemRegion mr(dummy_obj, word_size); 2298 CollectedHeap::fill_with_object(mr); 2299 } else { 2300 // If we can't allocate once, we probably cannot allocate 2301 // again. Let's get out of the loop. 2302 break; 2303 } 2304 } 2305 } 2306 #endif // !PRODUCT 2307 2308 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) { 2309 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2310 2311 // We assume that if concurrent == true, then the caller is a 2312 // concurrent thread that was joined the Suspendible Thread 2313 // Set. If there's ever a cheap way to check this, we should add an 2314 // assert here. 2315 2316 // We have already incremented _total_full_collections at the start 2317 // of the GC, so total_full_collections() represents how many full 2318 // collections have been started. 2319 unsigned int full_collections_started = total_full_collections(); 2320 2321 // Given that this method is called at the end of a Full GC or of a 2322 // concurrent cycle, and those can be nested (i.e., a Full GC can 2323 // interrupt a concurrent cycle), the number of full collections 2324 // completed should be either one (in the case where there was no 2325 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2326 // behind the number of full collections started. 2327 2328 // This is the case for the inner caller, i.e. a Full GC. 2329 assert(concurrent || 2330 (full_collections_started == _full_collections_completed + 1) || 2331 (full_collections_started == _full_collections_completed + 2), 2332 err_msg("for inner caller (Full GC): full_collections_started = %u " 2333 "is inconsistent with _full_collections_completed = %u", 2334 full_collections_started, _full_collections_completed)); 2335 2336 // This is the case for the outer caller, i.e. the concurrent cycle. 2337 assert(!concurrent || 2338 (full_collections_started == _full_collections_completed + 1), 2339 err_msg("for outer caller (concurrent cycle): " 2340 "full_collections_started = %u " 2341 "is inconsistent with _full_collections_completed = %u", 2342 full_collections_started, _full_collections_completed)); 2343 2344 _full_collections_completed += 1; 2345 2346 // We need to clear the "in_progress" flag in the CM thread before 2347 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2348 // is set) so that if a waiter requests another System.gc() it doesn't 2349 // incorrectly see that a marking cyle is still in progress. 2350 if (concurrent) { 2351 _cmThread->clear_in_progress(); 2352 } 2353 2354 // This notify_all() will ensure that a thread that called 2355 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2356 // and it's waiting for a full GC to finish will be woken up. It is 2357 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2358 FullGCCount_lock->notify_all(); 2359 } 2360 2361 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2362 assert_at_safepoint(true /* should_be_vm_thread */); 2363 GCCauseSetter gcs(this, cause); 2364 switch (cause) { 2365 case GCCause::_heap_inspection: 2366 case GCCause::_heap_dump: { 2367 HandleMark hm; 2368 do_full_collection(false); // don't clear all soft refs 2369 break; 2370 } 2371 default: // XXX FIX ME 2372 ShouldNotReachHere(); // Unexpected use of this function 2373 } 2374 } 2375 2376 void G1CollectedHeap::collect(GCCause::Cause cause) { 2377 // The caller doesn't have the Heap_lock 2378 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock"); 2379 2380 unsigned int gc_count_before; 2381 unsigned int full_gc_count_before; 2382 { 2383 MutexLocker ml(Heap_lock); 2384 2385 // Read the GC count while holding the Heap_lock 2386 gc_count_before = SharedHeap::heap()->total_collections(); 2387 full_gc_count_before = SharedHeap::heap()->total_full_collections(); 2388 } 2389 2390 if (should_do_concurrent_full_gc(cause)) { 2391 // Schedule an initial-mark evacuation pause that will start a 2392 // concurrent cycle. We're setting word_size to 0 which means that 2393 // we are not requesting a post-GC allocation. 2394 VM_G1IncCollectionPause op(gc_count_before, 2395 0, /* word_size */ 2396 true, /* should_initiate_conc_mark */ 2397 g1_policy()->max_pause_time_ms(), 2398 cause); 2399 VMThread::execute(&op); 2400 } else { 2401 if (cause == GCCause::_gc_locker 2402 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2403 2404 // Schedule a standard evacuation pause. We're setting word_size 2405 // to 0 which means that we are not requesting a post-GC allocation. 2406 VM_G1IncCollectionPause op(gc_count_before, 2407 0, /* word_size */ 2408 false, /* should_initiate_conc_mark */ 2409 g1_policy()->max_pause_time_ms(), 2410 cause); 2411 VMThread::execute(&op); 2412 } else { 2413 // Schedule a Full GC. 2414 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause); 2415 VMThread::execute(&op); 2416 } 2417 } 2418 } 2419 2420 bool G1CollectedHeap::is_in(const void* p) const { 2421 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p); 2422 if (hr != NULL) { 2423 return hr->is_in(p); 2424 } else { 2425 return _perm_gen->as_gen()->is_in(p); 2426 } 2427 } 2428 2429 // Iteration functions. 2430 2431 // Iterates an OopClosure over all ref-containing fields of objects 2432 // within a HeapRegion. 2433 2434 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2435 MemRegion _mr; 2436 OopClosure* _cl; 2437 public: 2438 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2439 : _mr(mr), _cl(cl) {} 2440 bool doHeapRegion(HeapRegion* r) { 2441 if (! r->continuesHumongous()) { 2442 r->oop_iterate(_cl); 2443 } 2444 return false; 2445 } 2446 }; 2447 2448 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2449 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2450 heap_region_iterate(&blk); 2451 if (do_perm) { 2452 perm_gen()->oop_iterate(cl); 2453 } 2454 } 2455 2456 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2457 IterateOopClosureRegionClosure blk(mr, cl); 2458 heap_region_iterate(&blk); 2459 if (do_perm) { 2460 perm_gen()->oop_iterate(cl); 2461 } 2462 } 2463 2464 // Iterates an ObjectClosure over all objects within a HeapRegion. 2465 2466 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2467 ObjectClosure* _cl; 2468 public: 2469 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2470 bool doHeapRegion(HeapRegion* r) { 2471 if (! r->continuesHumongous()) { 2472 r->object_iterate(_cl); 2473 } 2474 return false; 2475 } 2476 }; 2477 2478 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2479 IterateObjectClosureRegionClosure blk(cl); 2480 heap_region_iterate(&blk); 2481 if (do_perm) { 2482 perm_gen()->object_iterate(cl); 2483 } 2484 } 2485 2486 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2487 // FIXME: is this right? 2488 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2489 } 2490 2491 // Calls a SpaceClosure on a HeapRegion. 2492 2493 class SpaceClosureRegionClosure: public HeapRegionClosure { 2494 SpaceClosure* _cl; 2495 public: 2496 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2497 bool doHeapRegion(HeapRegion* r) { 2498 _cl->do_space(r); 2499 return false; 2500 } 2501 }; 2502 2503 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2504 SpaceClosureRegionClosure blk(cl); 2505 heap_region_iterate(&blk); 2506 } 2507 2508 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2509 _hrs.iterate(cl); 2510 } 2511 2512 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r, 2513 HeapRegionClosure* cl) const { 2514 _hrs.iterate_from(r, cl); 2515 } 2516 2517 void 2518 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2519 int worker, 2520 int no_of_par_workers, 2521 jint claim_value) { 2522 const size_t regions = n_regions(); 2523 const size_t max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2524 no_of_par_workers : 2525 1); 2526 assert(UseDynamicNumberOfGCThreads || 2527 no_of_par_workers == workers()->total_workers(), 2528 "Non dynamic should use fixed number of workers"); 2529 // try to spread out the starting points of the workers 2530 const size_t start_index = regions / max_workers * (size_t) worker; 2531 2532 // each worker will actually look at all regions 2533 for (size_t count = 0; count < regions; ++count) { 2534 const size_t index = (start_index + count) % regions; 2535 assert(0 <= index && index < regions, "sanity"); 2536 HeapRegion* r = region_at(index); 2537 // we'll ignore "continues humongous" regions (we'll process them 2538 // when we come across their corresponding "start humongous" 2539 // region) and regions already claimed 2540 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2541 continue; 2542 } 2543 // OK, try to claim it 2544 if (r->claimHeapRegion(claim_value)) { 2545 // success! 2546 assert(!r->continuesHumongous(), "sanity"); 2547 if (r->startsHumongous()) { 2548 // If the region is "starts humongous" we'll iterate over its 2549 // "continues humongous" first; in fact we'll do them 2550 // first. The order is important. In on case, calling the 2551 // closure on the "starts humongous" region might de-allocate 2552 // and clear all its "continues humongous" regions and, as a 2553 // result, we might end up processing them twice. So, we'll do 2554 // them first (notice: most closures will ignore them anyway) and 2555 // then we'll do the "starts humongous" region. 2556 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) { 2557 HeapRegion* chr = region_at(ch_index); 2558 2559 // if the region has already been claimed or it's not 2560 // "continues humongous" we're done 2561 if (chr->claim_value() == claim_value || 2562 !chr->continuesHumongous()) { 2563 break; 2564 } 2565 2566 // Noone should have claimed it directly. We can given 2567 // that we claimed its "starts humongous" region. 2568 assert(chr->claim_value() != claim_value, "sanity"); 2569 assert(chr->humongous_start_region() == r, "sanity"); 2570 2571 if (chr->claimHeapRegion(claim_value)) { 2572 // we should always be able to claim it; noone else should 2573 // be trying to claim this region 2574 2575 bool res2 = cl->doHeapRegion(chr); 2576 assert(!res2, "Should not abort"); 2577 2578 // Right now, this holds (i.e., no closure that actually 2579 // does something with "continues humongous" regions 2580 // clears them). We might have to weaken it in the future, 2581 // but let's leave these two asserts here for extra safety. 2582 assert(chr->continuesHumongous(), "should still be the case"); 2583 assert(chr->humongous_start_region() == r, "sanity"); 2584 } else { 2585 guarantee(false, "we should not reach here"); 2586 } 2587 } 2588 } 2589 2590 assert(!r->continuesHumongous(), "sanity"); 2591 bool res = cl->doHeapRegion(r); 2592 assert(!res, "Should not abort"); 2593 } 2594 } 2595 } 2596 2597 class ResetClaimValuesClosure: public HeapRegionClosure { 2598 public: 2599 bool doHeapRegion(HeapRegion* r) { 2600 r->set_claim_value(HeapRegion::InitialClaimValue); 2601 return false; 2602 } 2603 }; 2604 2605 void 2606 G1CollectedHeap::reset_heap_region_claim_values() { 2607 ResetClaimValuesClosure blk; 2608 heap_region_iterate(&blk); 2609 } 2610 2611 #ifdef ASSERT 2612 // This checks whether all regions in the heap have the correct claim 2613 // value. I also piggy-backed on this a check to ensure that the 2614 // humongous_start_region() information on "continues humongous" 2615 // regions is correct. 2616 2617 class CheckClaimValuesClosure : public HeapRegionClosure { 2618 private: 2619 jint _claim_value; 2620 size_t _failures; 2621 HeapRegion* _sh_region; 2622 public: 2623 CheckClaimValuesClosure(jint claim_value) : 2624 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2625 bool doHeapRegion(HeapRegion* r) { 2626 if (r->claim_value() != _claim_value) { 2627 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2628 "claim value = %d, should be %d", 2629 HR_FORMAT_PARAMS(r), 2630 r->claim_value(), _claim_value); 2631 ++_failures; 2632 } 2633 if (!r->isHumongous()) { 2634 _sh_region = NULL; 2635 } else if (r->startsHumongous()) { 2636 _sh_region = r; 2637 } else if (r->continuesHumongous()) { 2638 if (r->humongous_start_region() != _sh_region) { 2639 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2640 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2641 HR_FORMAT_PARAMS(r), 2642 r->humongous_start_region(), 2643 _sh_region); 2644 ++_failures; 2645 } 2646 } 2647 return false; 2648 } 2649 size_t failures() { 2650 return _failures; 2651 } 2652 }; 2653 2654 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2655 CheckClaimValuesClosure cl(claim_value); 2656 heap_region_iterate(&cl); 2657 return cl.failures() == 0; 2658 } 2659 2660 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2661 jint _claim_value; 2662 size_t _failures; 2663 2664 public: 2665 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2666 _claim_value(claim_value), 2667 _failures(0) { } 2668 2669 size_t failures() { 2670 return _failures; 2671 } 2672 2673 bool doHeapRegion(HeapRegion* hr) { 2674 assert(hr->in_collection_set(), "how?"); 2675 assert(!hr->isHumongous(), "H-region in CSet"); 2676 if (hr->claim_value() != _claim_value) { 2677 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2678 "claim value = %d, should be %d", 2679 HR_FORMAT_PARAMS(hr), 2680 hr->claim_value(), _claim_value); 2681 _failures += 1; 2682 } 2683 return false; 2684 } 2685 }; 2686 2687 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2688 CheckClaimValuesInCSetHRClosure cl(claim_value); 2689 collection_set_iterate(&cl); 2690 return cl.failures() == 0; 2691 } 2692 #endif // ASSERT 2693 2694 // Clear the cached CSet starting regions and (more importantly) 2695 // the time stamps. Called when we reset the GC time stamp. 2696 void G1CollectedHeap::clear_cset_start_regions() { 2697 assert(_worker_cset_start_region != NULL, "sanity"); 2698 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2699 2700 int n_queues = MAX2((int)ParallelGCThreads, 1); 2701 for (int i = 0; i < n_queues; i++) { 2702 _worker_cset_start_region[i] = NULL; 2703 _worker_cset_start_region_time_stamp[i] = 0; 2704 } 2705 } 2706 2707 // Given the id of a worker, obtain or calculate a suitable 2708 // starting region for iterating over the current collection set. 2709 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2710 assert(_worker_cset_start_region != NULL, "sanity"); 2711 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2712 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2713 2714 HeapRegion* result = NULL; 2715 unsigned gc_time_stamp = get_gc_time_stamp(); 2716 2717 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2718 // Cached starting region for current worker was set 2719 // during the current pause - so it's valid. 2720 // Note: the cached starting heap region may be NULL 2721 // (when the collection set is empty). 2722 result = _worker_cset_start_region[worker_i]; 2723 assert(result == NULL || result->in_collection_set(), "sanity"); 2724 return result; 2725 } 2726 2727 // The cached entry was not valid so let's calculate 2728 // a suitable starting heap region for this worker. 2729 2730 // We want the parallel threads to start their collection 2731 // set iteration at different collection set regions to 2732 // avoid contention. 2733 // If we have: 2734 // n collection set regions 2735 // p threads 2736 // Then thread t will start at region floor ((t * n) / p) 2737 2738 result = g1_policy()->collection_set(); 2739 if (G1CollectedHeap::use_parallel_gc_threads()) { 2740 size_t cs_size = g1_policy()->cset_region_length(); 2741 int active_workers = workers()->active_workers(); 2742 assert(UseDynamicNumberOfGCThreads || 2743 active_workers == workers()->total_workers(), 2744 "Unless dynamic should use total workers"); 2745 2746 size_t end_ind = (cs_size * worker_i) / active_workers; 2747 size_t start_ind = 0; 2748 2749 if (worker_i > 0 && 2750 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2751 // Previous workers starting region is valid 2752 // so let's iterate from there 2753 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2754 result = _worker_cset_start_region[worker_i - 1]; 2755 } 2756 2757 for (size_t i = start_ind; i < end_ind; i++) { 2758 result = result->next_in_collection_set(); 2759 } 2760 } 2761 2762 // Note: the calculated starting heap region may be NULL 2763 // (when the collection set is empty). 2764 assert(result == NULL || result->in_collection_set(), "sanity"); 2765 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2766 "should be updated only once per pause"); 2767 _worker_cset_start_region[worker_i] = result; 2768 OrderAccess::storestore(); 2769 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2770 return result; 2771 } 2772 2773 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2774 HeapRegion* r = g1_policy()->collection_set(); 2775 while (r != NULL) { 2776 HeapRegion* next = r->next_in_collection_set(); 2777 if (cl->doHeapRegion(r)) { 2778 cl->incomplete(); 2779 return; 2780 } 2781 r = next; 2782 } 2783 } 2784 2785 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2786 HeapRegionClosure *cl) { 2787 if (r == NULL) { 2788 // The CSet is empty so there's nothing to do. 2789 return; 2790 } 2791 2792 assert(r->in_collection_set(), 2793 "Start region must be a member of the collection set."); 2794 HeapRegion* cur = r; 2795 while (cur != NULL) { 2796 HeapRegion* next = cur->next_in_collection_set(); 2797 if (cl->doHeapRegion(cur) && false) { 2798 cl->incomplete(); 2799 return; 2800 } 2801 cur = next; 2802 } 2803 cur = g1_policy()->collection_set(); 2804 while (cur != r) { 2805 HeapRegion* next = cur->next_in_collection_set(); 2806 if (cl->doHeapRegion(cur) && false) { 2807 cl->incomplete(); 2808 return; 2809 } 2810 cur = next; 2811 } 2812 } 2813 2814 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2815 return n_regions() > 0 ? region_at(0) : NULL; 2816 } 2817 2818 2819 Space* G1CollectedHeap::space_containing(const void* addr) const { 2820 Space* res = heap_region_containing(addr); 2821 if (res == NULL) 2822 res = perm_gen()->space_containing(addr); 2823 return res; 2824 } 2825 2826 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2827 Space* sp = space_containing(addr); 2828 if (sp != NULL) { 2829 return sp->block_start(addr); 2830 } 2831 return NULL; 2832 } 2833 2834 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2835 Space* sp = space_containing(addr); 2836 assert(sp != NULL, "block_size of address outside of heap"); 2837 return sp->block_size(addr); 2838 } 2839 2840 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2841 Space* sp = space_containing(addr); 2842 return sp->block_is_obj(addr); 2843 } 2844 2845 bool G1CollectedHeap::supports_tlab_allocation() const { 2846 return true; 2847 } 2848 2849 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2850 return HeapRegion::GrainBytes; 2851 } 2852 2853 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2854 // Return the remaining space in the cur alloc region, but not less than 2855 // the min TLAB size. 2856 2857 // Also, this value can be at most the humongous object threshold, 2858 // since we can't allow tlabs to grow big enough to accomodate 2859 // humongous objects. 2860 2861 HeapRegion* hr = _mutator_alloc_region.get(); 2862 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2863 if (hr == NULL) { 2864 return max_tlab_size; 2865 } else { 2866 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2867 } 2868 } 2869 2870 size_t G1CollectedHeap::max_capacity() const { 2871 return _g1_reserved.byte_size(); 2872 } 2873 2874 jlong G1CollectedHeap::millis_since_last_gc() { 2875 // assert(false, "NYI"); 2876 return 0; 2877 } 2878 2879 void G1CollectedHeap::prepare_for_verify() { 2880 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2881 ensure_parsability(false); 2882 } 2883 g1_rem_set()->prepare_for_verify(); 2884 } 2885 2886 class VerifyLivenessOopClosure: public OopClosure { 2887 G1CollectedHeap* _g1h; 2888 VerifyOption _vo; 2889 public: 2890 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 2891 _g1h(g1h), _vo(vo) 2892 { } 2893 void do_oop(narrowOop *p) { do_oop_work(p); } 2894 void do_oop( oop *p) { do_oop_work(p); } 2895 2896 template <class T> void do_oop_work(T *p) { 2897 oop obj = oopDesc::load_decode_heap_oop(p); 2898 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 2899 "Dead object referenced by a not dead object"); 2900 } 2901 }; 2902 2903 class VerifyObjsInRegionClosure: public ObjectClosure { 2904 private: 2905 G1CollectedHeap* _g1h; 2906 size_t _live_bytes; 2907 HeapRegion *_hr; 2908 VerifyOption _vo; 2909 public: 2910 // _vo == UsePrevMarking -> use "prev" marking information, 2911 // _vo == UseNextMarking -> use "next" marking information, 2912 // _vo == UseMarkWord -> use mark word from object header. 2913 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 2914 : _live_bytes(0), _hr(hr), _vo(vo) { 2915 _g1h = G1CollectedHeap::heap(); 2916 } 2917 void do_object(oop o) { 2918 VerifyLivenessOopClosure isLive(_g1h, _vo); 2919 assert(o != NULL, "Huh?"); 2920 if (!_g1h->is_obj_dead_cond(o, _vo)) { 2921 // If the object is alive according to the mark word, 2922 // then verify that the marking information agrees. 2923 // Note we can't verify the contra-positive of the 2924 // above: if the object is dead (according to the mark 2925 // word), it may not be marked, or may have been marked 2926 // but has since became dead, or may have been allocated 2927 // since the last marking. 2928 if (_vo == VerifyOption_G1UseMarkWord) { 2929 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 2930 } 2931 2932 o->oop_iterate(&isLive); 2933 if (!_hr->obj_allocated_since_prev_marking(o)) { 2934 size_t obj_size = o->size(); // Make sure we don't overflow 2935 _live_bytes += (obj_size * HeapWordSize); 2936 } 2937 } 2938 } 2939 size_t live_bytes() { return _live_bytes; } 2940 }; 2941 2942 class PrintObjsInRegionClosure : public ObjectClosure { 2943 HeapRegion *_hr; 2944 G1CollectedHeap *_g1; 2945 public: 2946 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 2947 _g1 = G1CollectedHeap::heap(); 2948 }; 2949 2950 void do_object(oop o) { 2951 if (o != NULL) { 2952 HeapWord *start = (HeapWord *) o; 2953 size_t word_sz = o->size(); 2954 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 2955 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 2956 (void*) o, word_sz, 2957 _g1->isMarkedPrev(o), 2958 _g1->isMarkedNext(o), 2959 _hr->obj_allocated_since_prev_marking(o)); 2960 HeapWord *end = start + word_sz; 2961 HeapWord *cur; 2962 int *val; 2963 for (cur = start; cur < end; cur++) { 2964 val = (int *) cur; 2965 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 2966 } 2967 } 2968 } 2969 }; 2970 2971 class VerifyRegionClosure: public HeapRegionClosure { 2972 private: 2973 bool _allow_dirty; 2974 bool _par; 2975 VerifyOption _vo; 2976 bool _failures; 2977 public: 2978 // _vo == UsePrevMarking -> use "prev" marking information, 2979 // _vo == UseNextMarking -> use "next" marking information, 2980 // _vo == UseMarkWord -> use mark word from object header. 2981 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo) 2982 : _allow_dirty(allow_dirty), 2983 _par(par), 2984 _vo(vo), 2985 _failures(false) {} 2986 2987 bool failures() { 2988 return _failures; 2989 } 2990 2991 bool doHeapRegion(HeapRegion* r) { 2992 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue, 2993 "Should be unclaimed at verify points."); 2994 if (!r->continuesHumongous()) { 2995 bool failures = false; 2996 r->verify(_allow_dirty, _vo, &failures); 2997 if (failures) { 2998 _failures = true; 2999 } else { 3000 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3001 r->object_iterate(¬_dead_yet_cl); 3002 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3003 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3004 "max_live_bytes "SIZE_FORMAT" " 3005 "< calculated "SIZE_FORMAT, 3006 r->bottom(), r->end(), 3007 r->max_live_bytes(), 3008 not_dead_yet_cl.live_bytes()); 3009 _failures = true; 3010 } 3011 } 3012 } 3013 return false; // stop the region iteration if we hit a failure 3014 } 3015 }; 3016 3017 class VerifyRootsClosure: public OopsInGenClosure { 3018 private: 3019 G1CollectedHeap* _g1h; 3020 VerifyOption _vo; 3021 bool _failures; 3022 public: 3023 // _vo == UsePrevMarking -> use "prev" marking information, 3024 // _vo == UseNextMarking -> use "next" marking information, 3025 // _vo == UseMarkWord -> use mark word from object header. 3026 VerifyRootsClosure(VerifyOption vo) : 3027 _g1h(G1CollectedHeap::heap()), 3028 _vo(vo), 3029 _failures(false) { } 3030 3031 bool failures() { return _failures; } 3032 3033 template <class T> void do_oop_nv(T* p) { 3034 T heap_oop = oopDesc::load_heap_oop(p); 3035 if (!oopDesc::is_null(heap_oop)) { 3036 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3037 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3038 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3039 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3040 if (_vo == VerifyOption_G1UseMarkWord) { 3041 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3042 } 3043 obj->print_on(gclog_or_tty); 3044 _failures = true; 3045 } 3046 } 3047 } 3048 3049 void do_oop(oop* p) { do_oop_nv(p); } 3050 void do_oop(narrowOop* p) { do_oop_nv(p); } 3051 }; 3052 3053 // This is the task used for parallel heap verification. 3054 3055 class G1ParVerifyTask: public AbstractGangTask { 3056 private: 3057 G1CollectedHeap* _g1h; 3058 bool _allow_dirty; 3059 VerifyOption _vo; 3060 bool _failures; 3061 3062 public: 3063 // _vo == UsePrevMarking -> use "prev" marking information, 3064 // _vo == UseNextMarking -> use "next" marking information, 3065 // _vo == UseMarkWord -> use mark word from object header. 3066 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) : 3067 AbstractGangTask("Parallel verify task"), 3068 _g1h(g1h), 3069 _allow_dirty(allow_dirty), 3070 _vo(vo), 3071 _failures(false) { } 3072 3073 bool failures() { 3074 return _failures; 3075 } 3076 3077 void work(int worker_i) { 3078 HandleMark hm; 3079 VerifyRegionClosure blk(_allow_dirty, true, _vo); 3080 _g1h->heap_region_par_iterate_chunked(&blk, worker_i, 3081 _g1h->workers()->active_workers(), 3082 HeapRegion::ParVerifyClaimValue); 3083 if (blk.failures()) { 3084 _failures = true; 3085 } 3086 } 3087 }; 3088 3089 void G1CollectedHeap::verify(bool allow_dirty, bool silent) { 3090 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking); 3091 } 3092 3093 void G1CollectedHeap::verify(bool allow_dirty, 3094 bool silent, 3095 VerifyOption vo) { 3096 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3097 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3098 VerifyRootsClosure rootsCl(vo); 3099 3100 assert(Thread::current()->is_VM_thread(), 3101 "Expected to be executed serially by the VM thread at this point"); 3102 3103 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3104 3105 // We apply the relevant closures to all the oops in the 3106 // system dictionary, the string table and the code cache. 3107 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3108 3109 process_strong_roots(true, // activate StrongRootsScope 3110 true, // we set "collecting perm gen" to true, 3111 // so we don't reset the dirty cards in the perm gen. 3112 SharedHeap::ScanningOption(so), // roots scanning options 3113 &rootsCl, 3114 &blobsCl, 3115 &rootsCl); 3116 3117 // If we're verifying after the marking phase of a Full GC then we can't 3118 // treat the perm gen as roots into the G1 heap. Some of the objects in 3119 // the perm gen may be dead and hence not marked. If one of these dead 3120 // objects is considered to be a root then we may end up with a false 3121 // "Root location <x> points to dead ob <y>" failure. 3122 if (vo != VerifyOption_G1UseMarkWord) { 3123 // Since we used "collecting_perm_gen" == true above, we will not have 3124 // checked the refs from perm into the G1-collected heap. We check those 3125 // references explicitly below. Whether the relevant cards are dirty 3126 // is checked further below in the rem set verification. 3127 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3128 perm_gen()->oop_iterate(&rootsCl); 3129 } 3130 bool failures = rootsCl.failures(); 3131 3132 if (vo != VerifyOption_G1UseMarkWord) { 3133 // If we're verifying during a full GC then the region sets 3134 // will have been torn down at the start of the GC. Therefore 3135 // verifying the region sets will fail. So we only verify 3136 // the region sets when not in a full GC. 3137 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3138 verify_region_sets(); 3139 } 3140 3141 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3142 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3143 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3144 "sanity check"); 3145 3146 G1ParVerifyTask task(this, allow_dirty, vo); 3147 assert(UseDynamicNumberOfGCThreads || 3148 workers()->active_workers() == workers()->total_workers(), 3149 "If not dynamic should be using all the workers"); 3150 int n_workers = workers()->active_workers(); 3151 set_par_threads(n_workers); 3152 workers()->run_task(&task); 3153 set_par_threads(0); 3154 if (task.failures()) { 3155 failures = true; 3156 } 3157 3158 // Checks that the expected amount of parallel work was done. 3159 // The implication is that n_workers is > 0. 3160 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3161 "sanity check"); 3162 3163 reset_heap_region_claim_values(); 3164 3165 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3166 "sanity check"); 3167 } else { 3168 VerifyRegionClosure blk(allow_dirty, false, vo); 3169 heap_region_iterate(&blk); 3170 if (blk.failures()) { 3171 failures = true; 3172 } 3173 } 3174 if (!silent) gclog_or_tty->print("RemSet "); 3175 rem_set()->verify(); 3176 3177 if (failures) { 3178 gclog_or_tty->print_cr("Heap:"); 3179 // It helps to have the per-region information in the output to 3180 // help us track down what went wrong. This is why we call 3181 // print_extended_on() instead of print_on(). 3182 print_extended_on(gclog_or_tty); 3183 gclog_or_tty->print_cr(""); 3184 #ifndef PRODUCT 3185 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3186 concurrent_mark()->print_reachable("at-verification-failure", 3187 vo, false /* all */); 3188 } 3189 #endif 3190 gclog_or_tty->flush(); 3191 } 3192 guarantee(!failures, "there should not have been any failures"); 3193 } else { 3194 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3195 } 3196 } 3197 3198 class PrintRegionClosure: public HeapRegionClosure { 3199 outputStream* _st; 3200 public: 3201 PrintRegionClosure(outputStream* st) : _st(st) {} 3202 bool doHeapRegion(HeapRegion* r) { 3203 r->print_on(_st); 3204 return false; 3205 } 3206 }; 3207 3208 void G1CollectedHeap::print_on(outputStream* st) const { 3209 st->print(" %-20s", "garbage-first heap"); 3210 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3211 capacity()/K, used_unlocked()/K); 3212 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3213 _g1_storage.low_boundary(), 3214 _g1_storage.high(), 3215 _g1_storage.high_boundary()); 3216 st->cr(); 3217 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3218 size_t young_regions = _young_list->length(); 3219 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ", 3220 young_regions, young_regions * HeapRegion::GrainBytes / K); 3221 size_t survivor_regions = g1_policy()->recorded_survivor_regions(); 3222 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)", 3223 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K); 3224 st->cr(); 3225 perm()->as_gen()->print_on(st); 3226 } 3227 3228 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3229 print_on(st); 3230 3231 // Print the per-region information. 3232 st->cr(); 3233 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)"); 3234 PrintRegionClosure blk(st); 3235 heap_region_iterate(&blk); 3236 } 3237 3238 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3239 if (G1CollectedHeap::use_parallel_gc_threads()) { 3240 workers()->print_worker_threads_on(st); 3241 } 3242 _cmThread->print_on(st); 3243 st->cr(); 3244 _cm->print_worker_threads_on(st); 3245 _cg1r->print_worker_threads_on(st); 3246 st->cr(); 3247 } 3248 3249 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3250 if (G1CollectedHeap::use_parallel_gc_threads()) { 3251 workers()->threads_do(tc); 3252 } 3253 tc->do_thread(_cmThread); 3254 _cg1r->threads_do(tc); 3255 } 3256 3257 void G1CollectedHeap::print_tracing_info() const { 3258 // We'll overload this to mean "trace GC pause statistics." 3259 if (TraceGen0Time || TraceGen1Time) { 3260 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3261 // to that. 3262 g1_policy()->print_tracing_info(); 3263 } 3264 if (G1SummarizeRSetStats) { 3265 g1_rem_set()->print_summary_info(); 3266 } 3267 if (G1SummarizeConcMark) { 3268 concurrent_mark()->print_summary_info(); 3269 } 3270 g1_policy()->print_yg_surv_rate_info(); 3271 SpecializationStats::print(); 3272 } 3273 3274 #ifndef PRODUCT 3275 // Helpful for debugging RSet issues. 3276 3277 class PrintRSetsClosure : public HeapRegionClosure { 3278 private: 3279 const char* _msg; 3280 size_t _occupied_sum; 3281 3282 public: 3283 bool doHeapRegion(HeapRegion* r) { 3284 HeapRegionRemSet* hrrs = r->rem_set(); 3285 size_t occupied = hrrs->occupied(); 3286 _occupied_sum += occupied; 3287 3288 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3289 HR_FORMAT_PARAMS(r)); 3290 if (occupied == 0) { 3291 gclog_or_tty->print_cr(" RSet is empty"); 3292 } else { 3293 hrrs->print(); 3294 } 3295 gclog_or_tty->print_cr("----------"); 3296 return false; 3297 } 3298 3299 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3300 gclog_or_tty->cr(); 3301 gclog_or_tty->print_cr("========================================"); 3302 gclog_or_tty->print_cr(msg); 3303 gclog_or_tty->cr(); 3304 } 3305 3306 ~PrintRSetsClosure() { 3307 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3308 gclog_or_tty->print_cr("========================================"); 3309 gclog_or_tty->cr(); 3310 } 3311 }; 3312 3313 void G1CollectedHeap::print_cset_rsets() { 3314 PrintRSetsClosure cl("Printing CSet RSets"); 3315 collection_set_iterate(&cl); 3316 } 3317 3318 void G1CollectedHeap::print_all_rsets() { 3319 PrintRSetsClosure cl("Printing All RSets");; 3320 heap_region_iterate(&cl); 3321 } 3322 #endif // PRODUCT 3323 3324 G1CollectedHeap* G1CollectedHeap::heap() { 3325 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3326 "not a garbage-first heap"); 3327 return _g1h; 3328 } 3329 3330 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3331 // always_do_update_barrier = false; 3332 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3333 // Call allocation profiler 3334 AllocationProfiler::iterate_since_last_gc(); 3335 // Fill TLAB's and such 3336 ensure_parsability(true); 3337 } 3338 3339 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3340 // FIXME: what is this about? 3341 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3342 // is set. 3343 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3344 "derived pointer present")); 3345 // always_do_update_barrier = true; 3346 3347 // We have just completed a GC. Update the soft reference 3348 // policy with the new heap occupancy 3349 Universe::update_heap_info_at_gc(); 3350 } 3351 3352 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3353 unsigned int gc_count_before, 3354 bool* succeeded) { 3355 assert_heap_not_locked_and_not_at_safepoint(); 3356 g1_policy()->record_stop_world_start(); 3357 VM_G1IncCollectionPause op(gc_count_before, 3358 word_size, 3359 false, /* should_initiate_conc_mark */ 3360 g1_policy()->max_pause_time_ms(), 3361 GCCause::_g1_inc_collection_pause); 3362 VMThread::execute(&op); 3363 3364 HeapWord* result = op.result(); 3365 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3366 assert(result == NULL || ret_succeeded, 3367 "the result should be NULL if the VM did not succeed"); 3368 *succeeded = ret_succeeded; 3369 3370 assert_heap_not_locked(); 3371 return result; 3372 } 3373 3374 void 3375 G1CollectedHeap::doConcurrentMark() { 3376 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3377 if (!_cmThread->in_progress()) { 3378 _cmThread->set_started(); 3379 CGC_lock->notify(); 3380 } 3381 } 3382 3383 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr, 3384 bool young) { 3385 return _g1_policy->predict_region_elapsed_time_ms(hr, young); 3386 } 3387 3388 void G1CollectedHeap::check_if_region_is_too_expensive(double 3389 predicted_time_ms) { 3390 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms); 3391 } 3392 3393 size_t G1CollectedHeap::pending_card_num() { 3394 size_t extra_cards = 0; 3395 JavaThread *curr = Threads::first(); 3396 while (curr != NULL) { 3397 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3398 extra_cards += dcq.size(); 3399 curr = curr->next(); 3400 } 3401 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3402 size_t buffer_size = dcqs.buffer_size(); 3403 size_t buffer_num = dcqs.completed_buffers_num(); 3404 return buffer_size * buffer_num + extra_cards; 3405 } 3406 3407 size_t G1CollectedHeap::max_pending_card_num() { 3408 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3409 size_t buffer_size = dcqs.buffer_size(); 3410 size_t buffer_num = dcqs.completed_buffers_num(); 3411 int thread_num = Threads::number_of_threads(); 3412 return (buffer_num + thread_num) * buffer_size; 3413 } 3414 3415 size_t G1CollectedHeap::cards_scanned() { 3416 return g1_rem_set()->cardsScanned(); 3417 } 3418 3419 void 3420 G1CollectedHeap::setup_surviving_young_words() { 3421 guarantee( _surviving_young_words == NULL, "pre-condition" ); 3422 size_t array_length = g1_policy()->young_cset_region_length(); 3423 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length); 3424 if (_surviving_young_words == NULL) { 3425 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3426 "Not enough space for young surv words summary."); 3427 } 3428 memset(_surviving_young_words, 0, array_length * sizeof(size_t)); 3429 #ifdef ASSERT 3430 for (size_t i = 0; i < array_length; ++i) { 3431 assert( _surviving_young_words[i] == 0, "memset above" ); 3432 } 3433 #endif // !ASSERT 3434 } 3435 3436 void 3437 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3438 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3439 size_t array_length = g1_policy()->young_cset_region_length(); 3440 for (size_t i = 0; i < array_length; ++i) 3441 _surviving_young_words[i] += surv_young_words[i]; 3442 } 3443 3444 void 3445 G1CollectedHeap::cleanup_surviving_young_words() { 3446 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3447 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words); 3448 _surviving_young_words = NULL; 3449 } 3450 3451 #ifdef ASSERT 3452 class VerifyCSetClosure: public HeapRegionClosure { 3453 public: 3454 bool doHeapRegion(HeapRegion* hr) { 3455 // Here we check that the CSet region's RSet is ready for parallel 3456 // iteration. The fields that we'll verify are only manipulated 3457 // when the region is part of a CSet and is collected. Afterwards, 3458 // we reset these fields when we clear the region's RSet (when the 3459 // region is freed) so they are ready when the region is 3460 // re-allocated. The only exception to this is if there's an 3461 // evacuation failure and instead of freeing the region we leave 3462 // it in the heap. In that case, we reset these fields during 3463 // evacuation failure handling. 3464 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3465 3466 // Here's a good place to add any other checks we'd like to 3467 // perform on CSet regions. 3468 return false; 3469 } 3470 }; 3471 #endif // ASSERT 3472 3473 #if TASKQUEUE_STATS 3474 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3475 st->print_raw_cr("GC Task Stats"); 3476 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3477 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3478 } 3479 3480 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3481 print_taskqueue_stats_hdr(st); 3482 3483 TaskQueueStats totals; 3484 const int n = workers() != NULL ? workers()->total_workers() : 1; 3485 for (int i = 0; i < n; ++i) { 3486 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3487 totals += task_queue(i)->stats; 3488 } 3489 st->print_raw("tot "); totals.print(st); st->cr(); 3490 3491 DEBUG_ONLY(totals.verify()); 3492 } 3493 3494 void G1CollectedHeap::reset_taskqueue_stats() { 3495 const int n = workers() != NULL ? workers()->total_workers() : 1; 3496 for (int i = 0; i < n; ++i) { 3497 task_queue(i)->stats.reset(); 3498 } 3499 } 3500 #endif // TASKQUEUE_STATS 3501 3502 bool 3503 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3504 assert_at_safepoint(true /* should_be_vm_thread */); 3505 guarantee(!is_gc_active(), "collection is not reentrant"); 3506 3507 if (GC_locker::check_active_before_gc()) { 3508 return false; 3509 } 3510 3511 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3512 ResourceMark rm; 3513 3514 if (PrintHeapAtGC) { 3515 Universe::print_heap_before_gc(); 3516 } 3517 3518 HRSPhaseSetter x(HRSPhaseEvacuation); 3519 verify_region_sets_optional(); 3520 verify_dirty_young_regions(); 3521 3522 { 3523 // This call will decide whether this pause is an initial-mark 3524 // pause. If it is, during_initial_mark_pause() will return true 3525 // for the duration of this pause. 3526 g1_policy()->decide_on_conc_mark_initiation(); 3527 3528 // We do not allow initial-mark to be piggy-backed on a 3529 // partially-young GC. 3530 assert(!g1_policy()->during_initial_mark_pause() || 3531 g1_policy()->full_young_gcs(), "sanity"); 3532 3533 // We also do not allow partially-young GCs during marking. 3534 assert(!mark_in_progress() || g1_policy()->full_young_gcs(), "sanity"); 3535 3536 char verbose_str[128]; 3537 sprintf(verbose_str, "GC pause "); 3538 if (g1_policy()->full_young_gcs()) { 3539 strcat(verbose_str, "(young)"); 3540 } else { 3541 strcat(verbose_str, "(partial)"); 3542 } 3543 if (g1_policy()->during_initial_mark_pause()) { 3544 strcat(verbose_str, " (initial-mark)"); 3545 // We are about to start a marking cycle, so we increment the 3546 // full collection counter. 3547 increment_total_full_collections(); 3548 } 3549 3550 // if PrintGCDetails is on, we'll print long statistics information 3551 // in the collector policy code, so let's not print this as the output 3552 // is messy if we do. 3553 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 3554 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3555 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty); 3556 3557 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3558 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3559 3560 // If the secondary_free_list is not empty, append it to the 3561 // free_list. No need to wait for the cleanup operation to finish; 3562 // the region allocation code will check the secondary_free_list 3563 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3564 // set, skip this step so that the region allocation code has to 3565 // get entries from the secondary_free_list. 3566 if (!G1StressConcRegionFreeing) { 3567 append_secondary_free_list_if_not_empty_with_lock(); 3568 } 3569 3570 assert(check_young_list_well_formed(), 3571 "young list should be well formed"); 3572 3573 // Don't dynamically change the number of GC threads this early. A value of 3574 // 0 is used to indicate serial work. When parallel work is done, 3575 // it will be set. 3576 3577 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3578 IsGCActiveMark x; 3579 3580 gc_prologue(false); 3581 increment_total_collections(false /* full gc */); 3582 increment_gc_time_stamp(); 3583 3584 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3585 HandleMark hm; // Discard invalid handles created during verification 3586 gclog_or_tty->print(" VerifyBeforeGC:"); 3587 prepare_for_verify(); 3588 Universe::verify(/* allow dirty */ false, 3589 /* silent */ false, 3590 /* option */ VerifyOption_G1UsePrevMarking); 3591 3592 } 3593 3594 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3595 3596 // Please see comment in g1CollectedHeap.hpp and 3597 // G1CollectedHeap::ref_processing_init() to see how 3598 // reference processing currently works in G1. 3599 3600 // Enable discovery in the STW reference processor 3601 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3602 true /*verify_no_refs*/); 3603 3604 { 3605 // We want to temporarily turn off discovery by the 3606 // CM ref processor, if necessary, and turn it back on 3607 // on again later if we do. Using a scoped 3608 // NoRefDiscovery object will do this. 3609 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3610 3611 // Forget the current alloc region (we might even choose it to be part 3612 // of the collection set!). 3613 release_mutator_alloc_region(); 3614 3615 // We should call this after we retire the mutator alloc 3616 // region(s) so that all the ALLOC / RETIRE events are generated 3617 // before the start GC event. 3618 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3619 3620 // The elapsed time induced by the start time below deliberately elides 3621 // the possible verification above. 3622 double start_time_sec = os::elapsedTime(); 3623 size_t start_used_bytes = used(); 3624 3625 #if YOUNG_LIST_VERBOSE 3626 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3627 _young_list->print(); 3628 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3629 #endif // YOUNG_LIST_VERBOSE 3630 3631 g1_policy()->record_collection_pause_start(start_time_sec, 3632 start_used_bytes); 3633 3634 #if YOUNG_LIST_VERBOSE 3635 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3636 _young_list->print(); 3637 #endif // YOUNG_LIST_VERBOSE 3638 3639 if (g1_policy()->during_initial_mark_pause()) { 3640 concurrent_mark()->checkpointRootsInitialPre(); 3641 } 3642 perm_gen()->save_marks(); 3643 3644 // We must do this before any possible evacuation that should propagate 3645 // marks. 3646 if (mark_in_progress()) { 3647 double start_time_sec = os::elapsedTime(); 3648 3649 _cm->drainAllSATBBuffers(); 3650 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0; 3651 g1_policy()->record_satb_drain_time(finish_mark_ms); 3652 } 3653 // Record the number of elements currently on the mark stack, so we 3654 // only iterate over these. (Since evacuation may add to the mark 3655 // stack, doing more exposes race conditions.) If no mark is in 3656 // progress, this will be zero. 3657 _cm->set_oops_do_bound(); 3658 3659 if (mark_in_progress()) { 3660 concurrent_mark()->newCSet(); 3661 } 3662 3663 #if YOUNG_LIST_VERBOSE 3664 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3665 _young_list->print(); 3666 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3667 #endif // YOUNG_LIST_VERBOSE 3668 3669 g1_policy()->choose_collection_set(target_pause_time_ms); 3670 3671 if (_hr_printer.is_active()) { 3672 HeapRegion* hr = g1_policy()->collection_set(); 3673 while (hr != NULL) { 3674 G1HRPrinter::RegionType type; 3675 if (!hr->is_young()) { 3676 type = G1HRPrinter::Old; 3677 } else if (hr->is_survivor()) { 3678 type = G1HRPrinter::Survivor; 3679 } else { 3680 type = G1HRPrinter::Eden; 3681 } 3682 _hr_printer.cset(hr); 3683 hr = hr->next_in_collection_set(); 3684 } 3685 } 3686 3687 // We have chosen the complete collection set. If marking is 3688 // active then, we clear the region fields of any of the 3689 // concurrent marking tasks whose region fields point into 3690 // the collection set as these values will become stale. This 3691 // will cause the owning marking threads to claim a new region 3692 // when marking restarts. 3693 if (mark_in_progress()) { 3694 concurrent_mark()->reset_active_task_region_fields_in_cset(); 3695 } 3696 3697 #ifdef ASSERT 3698 VerifyCSetClosure cl; 3699 collection_set_iterate(&cl); 3700 #endif // ASSERT 3701 3702 setup_surviving_young_words(); 3703 3704 // Initialize the GC alloc regions. 3705 init_gc_alloc_regions(); 3706 3707 // Actually do the work... 3708 evacuate_collection_set(); 3709 3710 free_collection_set(g1_policy()->collection_set()); 3711 g1_policy()->clear_collection_set(); 3712 3713 cleanup_surviving_young_words(); 3714 3715 // Start a new incremental collection set for the next pause. 3716 g1_policy()->start_incremental_cset_building(); 3717 3718 // Clear the _cset_fast_test bitmap in anticipation of adding 3719 // regions to the incremental collection set for the next 3720 // evacuation pause. 3721 clear_cset_fast_test(); 3722 3723 _young_list->reset_sampled_info(); 3724 3725 // Don't check the whole heap at this point as the 3726 // GC alloc regions from this pause have been tagged 3727 // as survivors and moved on to the survivor list. 3728 // Survivor regions will fail the !is_young() check. 3729 assert(check_young_list_empty(false /* check_heap */), 3730 "young list should be empty"); 3731 3732 #if YOUNG_LIST_VERBOSE 3733 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3734 _young_list->print(); 3735 #endif // YOUNG_LIST_VERBOSE 3736 3737 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3738 _young_list->first_survivor_region(), 3739 _young_list->last_survivor_region()); 3740 3741 _young_list->reset_auxilary_lists(); 3742 3743 if (evacuation_failed()) { 3744 _summary_bytes_used = recalculate_used(); 3745 } else { 3746 // The "used" of the the collection set have already been subtracted 3747 // when they were freed. Add in the bytes evacuated. 3748 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3749 } 3750 3751 if (g1_policy()->during_initial_mark_pause()) { 3752 concurrent_mark()->checkpointRootsInitialPost(); 3753 set_marking_started(); 3754 // CAUTION: after the doConcurrentMark() call below, 3755 // the concurrent marking thread(s) could be running 3756 // concurrently with us. Make sure that anything after 3757 // this point does not assume that we are the only GC thread 3758 // running. Note: of course, the actual marking work will 3759 // not start until the safepoint itself is released in 3760 // ConcurrentGCThread::safepoint_desynchronize(). 3761 doConcurrentMark(); 3762 } 3763 3764 allocate_dummy_regions(); 3765 3766 #if YOUNG_LIST_VERBOSE 3767 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3768 _young_list->print(); 3769 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3770 #endif // YOUNG_LIST_VERBOSE 3771 3772 init_mutator_alloc_region(); 3773 3774 { 3775 size_t expand_bytes = g1_policy()->expansion_amount(); 3776 if (expand_bytes > 0) { 3777 size_t bytes_before = capacity(); 3778 if (!expand(expand_bytes)) { 3779 // We failed to expand the heap so let's verify that 3780 // committed/uncommitted amount match the backing store 3781 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3782 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3783 } 3784 } 3785 } 3786 3787 double end_time_sec = os::elapsedTime(); 3788 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS; 3789 g1_policy()->record_pause_time_ms(pause_time_ms); 3790 int active_gc_threads = workers()->active_workers(); 3791 g1_policy()->record_collection_pause_end(active_gc_threads); 3792 3793 MemoryService::track_memory_usage(); 3794 3795 // In prepare_for_verify() below we'll need to scan the deferred 3796 // update buffers to bring the RSets up-to-date if 3797 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3798 // the update buffers we'll probably need to scan cards on the 3799 // regions we just allocated to (i.e., the GC alloc 3800 // regions). However, during the last GC we called 3801 // set_saved_mark() on all the GC alloc regions, so card 3802 // scanning might skip the [saved_mark_word()...top()] area of 3803 // those regions (i.e., the area we allocated objects into 3804 // during the last GC). But it shouldn't. Given that 3805 // saved_mark_word() is conditional on whether the GC time stamp 3806 // on the region is current or not, by incrementing the GC time 3807 // stamp here we invalidate all the GC time stamps on all the 3808 // regions and saved_mark_word() will simply return top() for 3809 // all the regions. This is a nicer way of ensuring this rather 3810 // than iterating over the regions and fixing them. In fact, the 3811 // GC time stamp increment here also ensures that 3812 // saved_mark_word() will return top() between pauses, i.e., 3813 // during concurrent refinement. So we don't need the 3814 // is_gc_active() check to decided which top to use when 3815 // scanning cards (see CR 7039627). 3816 increment_gc_time_stamp(); 3817 3818 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 3819 HandleMark hm; // Discard invalid handles created during verification 3820 gclog_or_tty->print(" VerifyAfterGC:"); 3821 prepare_for_verify(); 3822 Universe::verify(/* allow dirty */ true, 3823 /* silent */ false, 3824 /* option */ VerifyOption_G1UsePrevMarking); 3825 } 3826 3827 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 3828 ref_processor_stw()->verify_no_references_recorded(); 3829 3830 // CM reference discovery will be re-enabled if necessary. 3831 } 3832 3833 { 3834 size_t expand_bytes = g1_policy()->expansion_amount(); 3835 if (expand_bytes > 0) { 3836 size_t bytes_before = capacity(); 3837 // No need for an ergo verbose message here, 3838 // expansion_amount() does this when it returns a value > 0. 3839 if (!expand(expand_bytes)) { 3840 // We failed to expand the heap so let's verify that 3841 // committed/uncommitted amount match the backing store 3842 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3843 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3844 } 3845 } 3846 } 3847 3848 // We should do this after we potentially expand the heap so 3849 // that all the COMMIT events are generated before the end GC 3850 // event, and after we retire the GC alloc regions so that all 3851 // RETIRE events are generated before the end GC event. 3852 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 3853 3854 // We have to do this after we decide whether to expand the heap or not. 3855 g1_policy()->print_heap_transition(); 3856 3857 if (mark_in_progress()) { 3858 concurrent_mark()->update_g1_committed(); 3859 } 3860 3861 #ifdef TRACESPINNING 3862 ParallelTaskTerminator::print_termination_counts(); 3863 #endif 3864 3865 gc_epilogue(false); 3866 } 3867 3868 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) { 3869 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum); 3870 print_tracing_info(); 3871 vm_exit(-1); 3872 } 3873 } 3874 3875 _hrs.verify_optional(); 3876 verify_region_sets_optional(); 3877 3878 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 3879 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 3880 3881 if (PrintHeapAtGC) { 3882 Universe::print_heap_after_gc(); 3883 } 3884 g1mm()->update_sizes(); 3885 3886 if (G1SummarizeRSetStats && 3887 (G1SummarizeRSetStatsPeriod > 0) && 3888 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3889 g1_rem_set()->print_summary_info(); 3890 } 3891 3892 return true; 3893 } 3894 3895 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 3896 { 3897 size_t gclab_word_size; 3898 switch (purpose) { 3899 case GCAllocForSurvived: 3900 gclab_word_size = YoungPLABSize; 3901 break; 3902 case GCAllocForTenured: 3903 gclab_word_size = OldPLABSize; 3904 break; 3905 default: 3906 assert(false, "unknown GCAllocPurpose"); 3907 gclab_word_size = OldPLABSize; 3908 break; 3909 } 3910 return gclab_word_size; 3911 } 3912 3913 void G1CollectedHeap::init_mutator_alloc_region() { 3914 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 3915 _mutator_alloc_region.init(); 3916 } 3917 3918 void G1CollectedHeap::release_mutator_alloc_region() { 3919 _mutator_alloc_region.release(); 3920 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 3921 } 3922 3923 void G1CollectedHeap::init_gc_alloc_regions() { 3924 assert_at_safepoint(true /* should_be_vm_thread */); 3925 3926 _survivor_gc_alloc_region.init(); 3927 _old_gc_alloc_region.init(); 3928 HeapRegion* retained_region = _retained_old_gc_alloc_region; 3929 _retained_old_gc_alloc_region = NULL; 3930 3931 // We will discard the current GC alloc region if: 3932 // a) it's in the collection set (it can happen!), 3933 // b) it's already full (no point in using it), 3934 // c) it's empty (this means that it was emptied during 3935 // a cleanup and it should be on the free list now), or 3936 // d) it's humongous (this means that it was emptied 3937 // during a cleanup and was added to the free list, but 3938 // has been subseqently used to allocate a humongous 3939 // object that may be less than the region size). 3940 if (retained_region != NULL && 3941 !retained_region->in_collection_set() && 3942 !(retained_region->top() == retained_region->end()) && 3943 !retained_region->is_empty() && 3944 !retained_region->isHumongous()) { 3945 retained_region->set_saved_mark(); 3946 // The retained region was added to the old region set when it was 3947 // retired. We have to remove it now, since we don't allow regions 3948 // we allocate to in the region sets. We'll re-add it later, when 3949 // it's retired again. 3950 _old_set.remove(retained_region); 3951 _old_gc_alloc_region.set(retained_region); 3952 _hr_printer.reuse(retained_region); 3953 } 3954 } 3955 3956 void G1CollectedHeap::release_gc_alloc_regions() { 3957 _survivor_gc_alloc_region.release(); 3958 // If we have an old GC alloc region to release, we'll save it in 3959 // _retained_old_gc_alloc_region. If we don't 3960 // _retained_old_gc_alloc_region will become NULL. This is what we 3961 // want either way so no reason to check explicitly for either 3962 // condition. 3963 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 3964 } 3965 3966 void G1CollectedHeap::abandon_gc_alloc_regions() { 3967 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 3968 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 3969 _retained_old_gc_alloc_region = NULL; 3970 } 3971 3972 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 3973 _drain_in_progress = false; 3974 set_evac_failure_closure(cl); 3975 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 3976 } 3977 3978 void G1CollectedHeap::finalize_for_evac_failure() { 3979 assert(_evac_failure_scan_stack != NULL && 3980 _evac_failure_scan_stack->length() == 0, 3981 "Postcondition"); 3982 assert(!_drain_in_progress, "Postcondition"); 3983 delete _evac_failure_scan_stack; 3984 _evac_failure_scan_stack = NULL; 3985 } 3986 3987 class UpdateRSetDeferred : public OopsInHeapRegionClosure { 3988 private: 3989 G1CollectedHeap* _g1; 3990 DirtyCardQueue *_dcq; 3991 CardTableModRefBS* _ct_bs; 3992 3993 public: 3994 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) : 3995 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {} 3996 3997 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 3998 virtual void do_oop( oop* p) { do_oop_work(p); } 3999 template <class T> void do_oop_work(T* p) { 4000 assert(_from->is_in_reserved(p), "paranoia"); 4001 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) && 4002 !_from->is_survivor()) { 4003 size_t card_index = _ct_bs->index_for(p); 4004 if (_ct_bs->mark_card_deferred(card_index)) { 4005 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index)); 4006 } 4007 } 4008 } 4009 }; 4010 4011 class RemoveSelfPointerClosure: public ObjectClosure { 4012 private: 4013 G1CollectedHeap* _g1; 4014 ConcurrentMark* _cm; 4015 HeapRegion* _hr; 4016 size_t _prev_marked_bytes; 4017 size_t _next_marked_bytes; 4018 OopsInHeapRegionClosure *_cl; 4019 public: 4020 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr, 4021 OopsInHeapRegionClosure* cl) : 4022 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0), 4023 _next_marked_bytes(0), _cl(cl) {} 4024 4025 size_t prev_marked_bytes() { return _prev_marked_bytes; } 4026 size_t next_marked_bytes() { return _next_marked_bytes; } 4027 4028 // <original comment> 4029 // The original idea here was to coalesce evacuated and dead objects. 4030 // However that caused complications with the block offset table (BOT). 4031 // In particular if there were two TLABs, one of them partially refined. 4032 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~| 4033 // The BOT entries of the unrefined part of TLAB_2 point to the start 4034 // of TLAB_2. If the last object of the TLAB_1 and the first object 4035 // of TLAB_2 are coalesced, then the cards of the unrefined part 4036 // would point into middle of the filler object. 4037 // The current approach is to not coalesce and leave the BOT contents intact. 4038 // </original comment> 4039 // 4040 // We now reset the BOT when we start the object iteration over the 4041 // region and refine its entries for every object we come across. So 4042 // the above comment is not really relevant and we should be able 4043 // to coalesce dead objects if we want to. 4044 void do_object(oop obj) { 4045 HeapWord* obj_addr = (HeapWord*) obj; 4046 assert(_hr->is_in(obj_addr), "sanity"); 4047 size_t obj_size = obj->size(); 4048 _hr->update_bot_for_object(obj_addr, obj_size); 4049 if (obj->is_forwarded() && obj->forwardee() == obj) { 4050 // The object failed to move. 4051 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs."); 4052 _cm->markPrev(obj); 4053 assert(_cm->isPrevMarked(obj), "Should be marked!"); 4054 _prev_marked_bytes += (obj_size * HeapWordSize); 4055 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) { 4056 _cm->markAndGrayObjectIfNecessary(obj); 4057 } 4058 obj->set_mark(markOopDesc::prototype()); 4059 // While we were processing RSet buffers during the 4060 // collection, we actually didn't scan any cards on the 4061 // collection set, since we didn't want to update remebered 4062 // sets with entries that point into the collection set, given 4063 // that live objects fromthe collection set are about to move 4064 // and such entries will be stale very soon. This change also 4065 // dealt with a reliability issue which involved scanning a 4066 // card in the collection set and coming across an array that 4067 // was being chunked and looking malformed. The problem is 4068 // that, if evacuation fails, we might have remembered set 4069 // entries missing given that we skipped cards on the 4070 // collection set. So, we'll recreate such entries now. 4071 obj->oop_iterate(_cl); 4072 assert(_cm->isPrevMarked(obj), "Should be marked!"); 4073 } else { 4074 // The object has been either evacuated or is dead. Fill it with a 4075 // dummy object. 4076 MemRegion mr((HeapWord*)obj, obj_size); 4077 CollectedHeap::fill_with_object(mr); 4078 _cm->clearRangeBothMaps(mr); 4079 } 4080 } 4081 }; 4082 4083 void G1CollectedHeap::remove_self_forwarding_pointers() { 4084 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set()); 4085 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set()); 4086 UpdateRSetDeferred deferred_update(_g1h, &dcq); 4087 OopsInHeapRegionClosure *cl; 4088 if (G1DeferredRSUpdate) { 4089 cl = &deferred_update; 4090 } else { 4091 cl = &immediate_update; 4092 } 4093 HeapRegion* cur = g1_policy()->collection_set(); 4094 while (cur != NULL) { 4095 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!"); 4096 assert(!cur->isHumongous(), "sanity"); 4097 4098 if (cur->evacuation_failed()) { 4099 assert(cur->in_collection_set(), "bad CS"); 4100 RemoveSelfPointerClosure rspc(_g1h, cur, cl); 4101 4102 // In the common case we make sure that this is done when the 4103 // region is freed so that it is "ready-to-go" when it's 4104 // re-allocated. However, when evacuation failure happens, a 4105 // region will remain in the heap and might ultimately be added 4106 // to a CSet in the future. So we have to be careful here and 4107 // make sure the region's RSet is ready for parallel iteration 4108 // whenever this might be required in the future. 4109 cur->rem_set()->reset_for_par_iteration(); 4110 cur->reset_bot(); 4111 cl->set_region(cur); 4112 cur->object_iterate(&rspc); 4113 4114 // A number of manipulations to make the TAMS be the current top, 4115 // and the marked bytes be the ones observed in the iteration. 4116 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) { 4117 // The comments below are the postconditions achieved by the 4118 // calls. Note especially the last such condition, which says that 4119 // the count of marked bytes has been properly restored. 4120 cur->note_start_of_marking(false); 4121 // _next_top_at_mark_start == top, _next_marked_bytes == 0 4122 cur->add_to_marked_bytes(rspc.prev_marked_bytes()); 4123 // _next_marked_bytes == prev_marked_bytes. 4124 cur->note_end_of_marking(); 4125 // _prev_top_at_mark_start == top(), 4126 // _prev_marked_bytes == prev_marked_bytes 4127 } 4128 // If there is no mark in progress, we modified the _next variables 4129 // above needlessly, but harmlessly. 4130 if (_g1h->mark_in_progress()) { 4131 cur->note_start_of_marking(false); 4132 // _next_top_at_mark_start == top, _next_marked_bytes == 0 4133 // _next_marked_bytes == next_marked_bytes. 4134 } 4135 } 4136 cur = cur->next_in_collection_set(); 4137 } 4138 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!"); 4139 4140 // Now restore saved marks, if any. 4141 if (_objs_with_preserved_marks != NULL) { 4142 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4143 guarantee(_objs_with_preserved_marks->length() == 4144 _preserved_marks_of_objs->length(), "Both or none."); 4145 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4146 oop obj = _objs_with_preserved_marks->at(i); 4147 markOop m = _preserved_marks_of_objs->at(i); 4148 obj->set_mark(m); 4149 } 4150 // Delete the preserved marks growable arrays (allocated on the C heap). 4151 delete _objs_with_preserved_marks; 4152 delete _preserved_marks_of_objs; 4153 _objs_with_preserved_marks = NULL; 4154 _preserved_marks_of_objs = NULL; 4155 } 4156 } 4157 4158 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4159 _evac_failure_scan_stack->push(obj); 4160 } 4161 4162 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4163 assert(_evac_failure_scan_stack != NULL, "precondition"); 4164 4165 while (_evac_failure_scan_stack->length() > 0) { 4166 oop obj = _evac_failure_scan_stack->pop(); 4167 _evac_failure_closure->set_region(heap_region_containing(obj)); 4168 obj->oop_iterate_backwards(_evac_failure_closure); 4169 } 4170 } 4171 4172 oop 4173 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4174 oop old, 4175 bool should_mark_root) { 4176 assert(obj_in_cs(old), 4177 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4178 (HeapWord*) old)); 4179 markOop m = old->mark(); 4180 oop forward_ptr = old->forward_to_atomic(old); 4181 if (forward_ptr == NULL) { 4182 // Forward-to-self succeeded. 4183 4184 // should_mark_root will be true when this routine is called 4185 // from a root scanning closure during an initial mark pause. 4186 // In this case the thread that succeeds in self-forwarding the 4187 // object is also responsible for marking the object. 4188 if (should_mark_root) { 4189 assert(!oopDesc::is_null(old), "shouldn't be"); 4190 _cm->grayRoot(old); 4191 } 4192 4193 if (_evac_failure_closure != cl) { 4194 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4195 assert(!_drain_in_progress, 4196 "Should only be true while someone holds the lock."); 4197 // Set the global evac-failure closure to the current thread's. 4198 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4199 set_evac_failure_closure(cl); 4200 // Now do the common part. 4201 handle_evacuation_failure_common(old, m); 4202 // Reset to NULL. 4203 set_evac_failure_closure(NULL); 4204 } else { 4205 // The lock is already held, and this is recursive. 4206 assert(_drain_in_progress, "This should only be the recursive case."); 4207 handle_evacuation_failure_common(old, m); 4208 } 4209 return old; 4210 } else { 4211 // Forward-to-self failed. Either someone else managed to allocate 4212 // space for this object (old != forward_ptr) or they beat us in 4213 // self-forwarding it (old == forward_ptr). 4214 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4215 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4216 "should not be in the CSet", 4217 (HeapWord*) old, (HeapWord*) forward_ptr)); 4218 return forward_ptr; 4219 } 4220 } 4221 4222 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4223 set_evacuation_failed(true); 4224 4225 preserve_mark_if_necessary(old, m); 4226 4227 HeapRegion* r = heap_region_containing(old); 4228 if (!r->evacuation_failed()) { 4229 r->set_evacuation_failed(true); 4230 _hr_printer.evac_failure(r); 4231 } 4232 4233 push_on_evac_failure_scan_stack(old); 4234 4235 if (!_drain_in_progress) { 4236 // prevent recursion in copy_to_survivor_space() 4237 _drain_in_progress = true; 4238 drain_evac_failure_scan_stack(); 4239 _drain_in_progress = false; 4240 } 4241 } 4242 4243 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4244 assert(evacuation_failed(), "Oversaving!"); 4245 // We want to call the "for_promotion_failure" version only in the 4246 // case of a promotion failure. 4247 if (m->must_be_preserved_for_promotion_failure(obj)) { 4248 if (_objs_with_preserved_marks == NULL) { 4249 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4250 _objs_with_preserved_marks = 4251 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4252 _preserved_marks_of_objs = 4253 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true); 4254 } 4255 _objs_with_preserved_marks->push(obj); 4256 _preserved_marks_of_objs->push(m); 4257 } 4258 } 4259 4260 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4261 size_t word_size) { 4262 if (purpose == GCAllocForSurvived) { 4263 HeapWord* result = survivor_attempt_allocation(word_size); 4264 if (result != NULL) { 4265 return result; 4266 } else { 4267 // Let's try to allocate in the old gen in case we can fit the 4268 // object there. 4269 return old_attempt_allocation(word_size); 4270 } 4271 } else { 4272 assert(purpose == GCAllocForTenured, "sanity"); 4273 HeapWord* result = old_attempt_allocation(word_size); 4274 if (result != NULL) { 4275 return result; 4276 } else { 4277 // Let's try to allocate in the survivors in case we can fit the 4278 // object there. 4279 return survivor_attempt_allocation(word_size); 4280 } 4281 } 4282 4283 ShouldNotReachHere(); 4284 // Trying to keep some compilers happy. 4285 return NULL; 4286 } 4287 4288 #ifndef PRODUCT 4289 bool GCLabBitMapClosure::do_bit(size_t offset) { 4290 HeapWord* addr = _bitmap->offsetToHeapWord(offset); 4291 guarantee(_cm->isMarked(oop(addr)), "it should be!"); 4292 return true; 4293 } 4294 #endif // PRODUCT 4295 4296 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4297 ParGCAllocBuffer(gclab_word_size), 4298 _should_mark_objects(false), 4299 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size), 4300 _retired(false) 4301 { 4302 //_should_mark_objects is set to true when G1ParCopyHelper needs to 4303 // mark the forwarded location of an evacuated object. 4304 // We set _should_mark_objects to true if marking is active, i.e. when we 4305 // need to propagate a mark, or during an initial mark pause, i.e. when we 4306 // need to mark objects immediately reachable by the roots. 4307 if (G1CollectedHeap::heap()->mark_in_progress() || 4308 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) { 4309 _should_mark_objects = true; 4310 } 4311 } 4312 4313 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num) 4314 : _g1h(g1h), 4315 _refs(g1h->task_queue(queue_num)), 4316 _dcq(&g1h->dirty_card_queue_set()), 4317 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4318 _g1_rem(g1h->g1_rem_set()), 4319 _hash_seed(17), _queue_num(queue_num), 4320 _term_attempts(0), 4321 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4322 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4323 _age_table(false), 4324 _strong_roots_time(0), _term_time(0), 4325 _alloc_buffer_waste(0), _undo_waste(0) 4326 { 4327 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4328 // we "sacrifice" entry 0 to keep track of surviving bytes for 4329 // non-young regions (where the age is -1) 4330 // We also add a few elements at the beginning and at the end in 4331 // an attempt to eliminate cache contention 4332 size_t real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4333 size_t array_length = PADDING_ELEM_NUM + 4334 real_length + 4335 PADDING_ELEM_NUM; 4336 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length); 4337 if (_surviving_young_words_base == NULL) 4338 vm_exit_out_of_memory(array_length * sizeof(size_t), 4339 "Not enough space for young surv histo."); 4340 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4341 memset(_surviving_young_words, 0, real_length * sizeof(size_t)); 4342 4343 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4344 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4345 4346 _start = os::elapsedTime(); 4347 } 4348 4349 void 4350 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4351 { 4352 st->print_raw_cr("GC Termination Stats"); 4353 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4354 " ------waste (KiB)------"); 4355 st->print_raw_cr("thr ms ms % ms % attempts" 4356 " total alloc undo"); 4357 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4358 " ------- ------- -------"); 4359 } 4360 4361 void 4362 G1ParScanThreadState::print_termination_stats(int i, 4363 outputStream* const st) const 4364 { 4365 const double elapsed_ms = elapsed_time() * 1000.0; 4366 const double s_roots_ms = strong_roots_time() * 1000.0; 4367 const double term_ms = term_time() * 1000.0; 4368 st->print_cr("%3d %9.2f %9.2f %6.2f " 4369 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4370 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4371 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4372 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4373 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4374 alloc_buffer_waste() * HeapWordSize / K, 4375 undo_waste() * HeapWordSize / K); 4376 } 4377 4378 #ifdef ASSERT 4379 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4380 assert(ref != NULL, "invariant"); 4381 assert(UseCompressedOops, "sanity"); 4382 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4383 oop p = oopDesc::load_decode_heap_oop(ref); 4384 assert(_g1h->is_in_g1_reserved(p), 4385 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4386 return true; 4387 } 4388 4389 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4390 assert(ref != NULL, "invariant"); 4391 if (has_partial_array_mask(ref)) { 4392 // Must be in the collection set--it's already been copied. 4393 oop p = clear_partial_array_mask(ref); 4394 assert(_g1h->obj_in_cs(p), 4395 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4396 } else { 4397 oop p = oopDesc::load_decode_heap_oop(ref); 4398 assert(_g1h->is_in_g1_reserved(p), 4399 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4400 } 4401 return true; 4402 } 4403 4404 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4405 if (ref.is_narrow()) { 4406 return verify_ref((narrowOop*) ref); 4407 } else { 4408 return verify_ref((oop*) ref); 4409 } 4410 } 4411 #endif // ASSERT 4412 4413 void G1ParScanThreadState::trim_queue() { 4414 assert(_evac_cl != NULL, "not set"); 4415 assert(_evac_failure_cl != NULL, "not set"); 4416 assert(_partial_scan_cl != NULL, "not set"); 4417 4418 StarTask ref; 4419 do { 4420 // Drain the overflow stack first, so other threads can steal. 4421 while (refs()->pop_overflow(ref)) { 4422 deal_with_reference(ref); 4423 } 4424 4425 while (refs()->pop_local(ref)) { 4426 deal_with_reference(ref); 4427 } 4428 } while (!refs()->is_empty()); 4429 } 4430 4431 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) : 4432 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4433 _par_scan_state(par_scan_state), 4434 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4435 _mark_in_progress(_g1->mark_in_progress()) { } 4436 4437 template <class T> void G1ParCopyHelper::mark_object(T* p) { 4438 // This is called from do_oop_work for objects that are not 4439 // in the collection set. Objects in the collection set 4440 // are marked after they have been evacuated. 4441 4442 T heap_oop = oopDesc::load_heap_oop(p); 4443 if (!oopDesc::is_null(heap_oop)) { 4444 oop obj = oopDesc::decode_heap_oop(heap_oop); 4445 HeapWord* addr = (HeapWord*)obj; 4446 if (_g1->is_in_g1_reserved(addr)) { 4447 _cm->grayRoot(oop(addr)); 4448 } 4449 } 4450 } 4451 4452 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root, 4453 bool should_mark_copy) { 4454 size_t word_sz = old->size(); 4455 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4456 // +1 to make the -1 indexes valid... 4457 int young_index = from_region->young_index_in_cset()+1; 4458 assert( (from_region->is_young() && young_index > 0) || 4459 (!from_region->is_young() && young_index == 0), "invariant" ); 4460 G1CollectorPolicy* g1p = _g1->g1_policy(); 4461 markOop m = old->mark(); 4462 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4463 : m->age(); 4464 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4465 word_sz); 4466 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4467 oop obj = oop(obj_ptr); 4468 4469 if (obj_ptr == NULL) { 4470 // This will either forward-to-self, or detect that someone else has 4471 // installed a forwarding pointer. 4472 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4473 return _g1->handle_evacuation_failure_par(cl, old, should_mark_root); 4474 } 4475 4476 // We're going to allocate linearly, so might as well prefetch ahead. 4477 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4478 4479 oop forward_ptr = old->forward_to_atomic(obj); 4480 if (forward_ptr == NULL) { 4481 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4482 if (g1p->track_object_age(alloc_purpose)) { 4483 // We could simply do obj->incr_age(). However, this causes a 4484 // performance issue. obj->incr_age() will first check whether 4485 // the object has a displaced mark by checking its mark word; 4486 // getting the mark word from the new location of the object 4487 // stalls. So, given that we already have the mark word and we 4488 // are about to install it anyway, it's better to increase the 4489 // age on the mark word, when the object does not have a 4490 // displaced mark word. We're not expecting many objects to have 4491 // a displaced marked word, so that case is not optimized 4492 // further (it could be...) and we simply call obj->incr_age(). 4493 4494 if (m->has_displaced_mark_helper()) { 4495 // in this case, we have to install the mark word first, 4496 // otherwise obj looks to be forwarded (the old mark word, 4497 // which contains the forward pointer, was copied) 4498 obj->set_mark(m); 4499 obj->incr_age(); 4500 } else { 4501 m = m->incr_age(); 4502 obj->set_mark(m); 4503 } 4504 _par_scan_state->age_table()->add(obj, word_sz); 4505 } else { 4506 obj->set_mark(m); 4507 } 4508 4509 // Mark the evacuated object or propagate "next" mark bit 4510 if (should_mark_copy) { 4511 if (!use_local_bitmaps || 4512 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) { 4513 // if we couldn't mark it on the local bitmap (this happens when 4514 // the object was not allocated in the GCLab), we have to bite 4515 // the bullet and do the standard parallel mark 4516 _cm->markAndGrayObjectIfNecessary(obj); 4517 } 4518 4519 if (_g1->isMarkedNext(old)) { 4520 // Unmark the object's old location so that marking 4521 // doesn't think the old object is alive. 4522 _cm->nextMarkBitMap()->parClear((HeapWord*)old); 4523 } 4524 } 4525 4526 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4527 surv_young_words[young_index] += word_sz; 4528 4529 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4530 arrayOop(old)->set_length(0); 4531 oop* old_p = set_partial_array_mask(old); 4532 _par_scan_state->push_on_queue(old_p); 4533 } else { 4534 // No point in using the slower heap_region_containing() method, 4535 // given that we know obj is in the heap. 4536 _scanner->set_region(_g1->heap_region_containing_raw(obj)); 4537 obj->oop_iterate_backwards(_scanner); 4538 } 4539 } else { 4540 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4541 obj = forward_ptr; 4542 } 4543 return obj; 4544 } 4545 4546 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4547 template <class T> 4548 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4549 ::do_oop_work(T* p) { 4550 oop obj = oopDesc::load_decode_heap_oop(p); 4551 assert(barrier != G1BarrierRS || obj != NULL, 4552 "Precondition: G1BarrierRS implies obj is nonNull"); 4553 4554 // Marking: 4555 // If the object is in the collection set, then the thread 4556 // that copies the object should mark, or propagate the 4557 // mark to, the evacuated object. 4558 // If the object is not in the collection set then we 4559 // should call the mark_object() method depending on the 4560 // value of the template parameter do_mark_object (which will 4561 // be true for root scanning closures during an initial mark 4562 // pause). 4563 // The mark_object() method first checks whether the object 4564 // is marked and, if not, attempts to mark the object. 4565 4566 // here the null check is implicit in the cset_fast_test() test 4567 if (_g1->in_cset_fast_test(obj)) { 4568 if (obj->is_forwarded()) { 4569 oopDesc::encode_store_heap_oop(p, obj->forwardee()); 4570 // If we are a root scanning closure during an initial 4571 // mark pause (i.e. do_mark_object will be true) then 4572 // we also need to handle marking of roots in the 4573 // event of an evacuation failure. In the event of an 4574 // evacuation failure, the object is forwarded to itself 4575 // and not copied. For root-scanning closures, the 4576 // object would be marked after a successful self-forward 4577 // but an object could be pointed to by both a root and non 4578 // root location and be self-forwarded by a non-root-scanning 4579 // closure. Therefore we also have to attempt to mark the 4580 // self-forwarded root object here. 4581 if (do_mark_object && obj->forwardee() == obj) { 4582 mark_object(p); 4583 } 4584 } else { 4585 // During an initial mark pause, objects that are pointed to 4586 // by the roots need to be marked - even in the event of an 4587 // evacuation failure. We pass the template parameter 4588 // do_mark_object (which is true for root scanning closures 4589 // during an initial mark pause) to copy_to_survivor_space 4590 // which will pass it on to the evacuation failure handling 4591 // code. The thread that successfully self-forwards a root 4592 // object to itself is responsible for marking the object. 4593 bool should_mark_root = do_mark_object; 4594 4595 // We need to mark the copied object if we're a root scanning 4596 // closure during an initial mark pause (i.e. do_mark_object 4597 // will be true), or the object is already marked and we need 4598 // to propagate the mark to the evacuated copy. 4599 bool should_mark_copy = do_mark_object || 4600 _during_initial_mark || 4601 (_mark_in_progress && !_g1->is_obj_ill(obj)); 4602 4603 oop copy_oop = copy_to_survivor_space(obj, should_mark_root, 4604 should_mark_copy); 4605 oopDesc::encode_store_heap_oop(p, copy_oop); 4606 } 4607 // When scanning the RS, we only care about objs in CS. 4608 if (barrier == G1BarrierRS) { 4609 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num()); 4610 } 4611 } else { 4612 // The object is not in collection set. If we're a root scanning 4613 // closure during an initial mark pause (i.e. do_mark_object will 4614 // be true) then attempt to mark the object. 4615 if (do_mark_object) { 4616 mark_object(p); 4617 } 4618 } 4619 4620 if (barrier == G1BarrierEvac && obj != NULL) { 4621 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num()); 4622 } 4623 4624 if (do_gen_barrier && obj != NULL) { 4625 par_do_barrier(p); 4626 } 4627 } 4628 4629 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4630 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4631 4632 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4633 assert(has_partial_array_mask(p), "invariant"); 4634 oop old = clear_partial_array_mask(p); 4635 assert(old->is_objArray(), "must be obj array"); 4636 assert(old->is_forwarded(), "must be forwarded"); 4637 assert(Universe::heap()->is_in_reserved(old), "must be in heap."); 4638 4639 objArrayOop obj = objArrayOop(old->forwardee()); 4640 assert((void*)old != (void*)old->forwardee(), "self forwarding here?"); 4641 // Process ParGCArrayScanChunk elements now 4642 // and push the remainder back onto queue 4643 int start = arrayOop(old)->length(); 4644 int end = obj->length(); 4645 int remainder = end - start; 4646 assert(start <= end, "just checking"); 4647 if (remainder > 2 * ParGCArrayScanChunk) { 4648 // Test above combines last partial chunk with a full chunk 4649 end = start + ParGCArrayScanChunk; 4650 arrayOop(old)->set_length(end); 4651 // Push remainder. 4652 oop* old_p = set_partial_array_mask(old); 4653 assert(arrayOop(old)->length() < obj->length(), "Empty push?"); 4654 _par_scan_state->push_on_queue(old_p); 4655 } else { 4656 // Restore length so that the heap remains parsable in 4657 // case of evacuation failure. 4658 arrayOop(old)->set_length(end); 4659 } 4660 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4661 // process our set of indices (include header in first chunk) 4662 obj->oop_iterate_range(&_scanner, start, end); 4663 } 4664 4665 class G1ParEvacuateFollowersClosure : public VoidClosure { 4666 protected: 4667 G1CollectedHeap* _g1h; 4668 G1ParScanThreadState* _par_scan_state; 4669 RefToScanQueueSet* _queues; 4670 ParallelTaskTerminator* _terminator; 4671 4672 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4673 RefToScanQueueSet* queues() { return _queues; } 4674 ParallelTaskTerminator* terminator() { return _terminator; } 4675 4676 public: 4677 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4678 G1ParScanThreadState* par_scan_state, 4679 RefToScanQueueSet* queues, 4680 ParallelTaskTerminator* terminator) 4681 : _g1h(g1h), _par_scan_state(par_scan_state), 4682 _queues(queues), _terminator(terminator) {} 4683 4684 void do_void(); 4685 4686 private: 4687 inline bool offer_termination(); 4688 }; 4689 4690 bool G1ParEvacuateFollowersClosure::offer_termination() { 4691 G1ParScanThreadState* const pss = par_scan_state(); 4692 pss->start_term_time(); 4693 const bool res = terminator()->offer_termination(); 4694 pss->end_term_time(); 4695 return res; 4696 } 4697 4698 void G1ParEvacuateFollowersClosure::do_void() { 4699 StarTask stolen_task; 4700 G1ParScanThreadState* const pss = par_scan_state(); 4701 pss->trim_queue(); 4702 4703 do { 4704 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4705 assert(pss->verify_task(stolen_task), "sanity"); 4706 if (stolen_task.is_narrow()) { 4707 pss->deal_with_reference((narrowOop*) stolen_task); 4708 } else { 4709 pss->deal_with_reference((oop*) stolen_task); 4710 } 4711 4712 // We've just processed a reference and we might have made 4713 // available new entries on the queues. So we have to make sure 4714 // we drain the queues as necessary. 4715 pss->trim_queue(); 4716 } 4717 } while (!offer_termination()); 4718 4719 pss->retire_alloc_buffers(); 4720 } 4721 4722 class G1ParTask : public AbstractGangTask { 4723 protected: 4724 G1CollectedHeap* _g1h; 4725 RefToScanQueueSet *_queues; 4726 ParallelTaskTerminator _terminator; 4727 int _n_workers; 4728 4729 Mutex _stats_lock; 4730 Mutex* stats_lock() { return &_stats_lock; } 4731 4732 size_t getNCards() { 4733 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4734 / G1BlockOffsetSharedArray::N_bytes; 4735 } 4736 4737 public: 4738 G1ParTask(G1CollectedHeap* g1h, 4739 RefToScanQueueSet *task_queues) 4740 : AbstractGangTask("G1 collection"), 4741 _g1h(g1h), 4742 _queues(task_queues), 4743 _terminator(0, _queues), 4744 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4745 {} 4746 4747 RefToScanQueueSet* queues() { return _queues; } 4748 4749 RefToScanQueue *work_queue(int i) { 4750 return queues()->queue(i); 4751 } 4752 4753 ParallelTaskTerminator* terminator() { return &_terminator; } 4754 4755 virtual void set_for_termination(int active_workers) { 4756 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4757 // in the young space (_par_seq_tasks) in the G1 heap 4758 // for SequentialSubTasksDone. 4759 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4760 // both of which need setting by set_n_termination(). 4761 _g1h->SharedHeap::set_n_termination(active_workers); 4762 _g1h->set_n_termination(active_workers); 4763 terminator()->reset_for_reuse(active_workers); 4764 _n_workers = active_workers; 4765 } 4766 4767 void work(int i) { 4768 if (i >= _n_workers) return; // no work needed this round 4769 4770 double start_time_ms = os::elapsedTime() * 1000.0; 4771 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms); 4772 4773 ResourceMark rm; 4774 HandleMark hm; 4775 4776 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4777 4778 G1ParScanThreadState pss(_g1h, i); 4779 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4780 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4781 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4782 4783 pss.set_evac_closure(&scan_evac_cl); 4784 pss.set_evac_failure_closure(&evac_failure_cl); 4785 pss.set_partial_scan_closure(&partial_scan_cl); 4786 4787 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4788 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4789 4790 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4791 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4792 4793 OopClosure* scan_root_cl = &only_scan_root_cl; 4794 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4795 4796 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4797 // We also need to mark copied objects. 4798 scan_root_cl = &scan_mark_root_cl; 4799 scan_perm_cl = &scan_mark_perm_cl; 4800 } 4801 4802 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4803 4804 pss.start_strong_roots(); 4805 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4806 SharedHeap::SO_AllClasses, 4807 scan_root_cl, 4808 &push_heap_rs_cl, 4809 scan_perm_cl, 4810 i); 4811 pss.end_strong_roots(); 4812 4813 { 4814 double start = os::elapsedTime(); 4815 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4816 evac.do_void(); 4817 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4818 double term_ms = pss.term_time()*1000.0; 4819 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms); 4820 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts()); 4821 } 4822 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4823 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4824 4825 // Clean up any par-expanded rem sets. 4826 HeapRegionRemSet::par_cleanup(); 4827 4828 if (ParallelGCVerbose) { 4829 MutexLocker x(stats_lock()); 4830 pss.print_termination_stats(i); 4831 } 4832 4833 assert(pss.refs()->is_empty(), "should be empty"); 4834 double end_time_ms = os::elapsedTime() * 1000.0; 4835 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms); 4836 } 4837 }; 4838 4839 // *** Common G1 Evacuation Stuff 4840 4841 // This method is run in a GC worker. 4842 4843 void 4844 G1CollectedHeap:: 4845 g1_process_strong_roots(bool collecting_perm_gen, 4846 SharedHeap::ScanningOption so, 4847 OopClosure* scan_non_heap_roots, 4848 OopsInHeapRegionClosure* scan_rs, 4849 OopsInGenClosure* scan_perm, 4850 int worker_i) { 4851 4852 // First scan the strong roots, including the perm gen. 4853 double ext_roots_start = os::elapsedTime(); 4854 double closure_app_time_sec = 0.0; 4855 4856 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4857 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4858 buf_scan_perm.set_generation(perm_gen()); 4859 4860 // Walk the code cache w/o buffering, because StarTask cannot handle 4861 // unaligned oop locations. 4862 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true); 4863 4864 process_strong_roots(false, // no scoping; this is parallel code 4865 collecting_perm_gen, so, 4866 &buf_scan_non_heap_roots, 4867 &eager_scan_code_roots, 4868 &buf_scan_perm); 4869 4870 // Now the CM ref_processor roots. 4871 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4872 // We need to treat the discovered reference lists of the 4873 // concurrent mark ref processor as roots and keep entries 4874 // (which are added by the marking threads) on them live 4875 // until they can be processed at the end of marking. 4876 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4877 } 4878 4879 // Finish up any enqueued closure apps (attributed as object copy time). 4880 buf_scan_non_heap_roots.done(); 4881 buf_scan_perm.done(); 4882 4883 double ext_roots_end = os::elapsedTime(); 4884 4885 g1_policy()->reset_obj_copy_time(worker_i); 4886 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4887 buf_scan_non_heap_roots.closure_app_seconds(); 4888 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4889 4890 double ext_root_time_ms = 4891 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4892 4893 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4894 4895 // Scan strong roots in mark stack. 4896 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) { 4897 concurrent_mark()->oops_do(scan_non_heap_roots); 4898 } 4899 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0; 4900 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms); 4901 4902 // Now scan the complement of the collection set. 4903 if (scan_rs != NULL) { 4904 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4905 } 4906 4907 _process_strong_tasks->all_tasks_completed(); 4908 } 4909 4910 void 4911 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 4912 OopClosure* non_root_closure) { 4913 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 4914 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 4915 } 4916 4917 // Weak Reference Processing support 4918 4919 // An always "is_alive" closure that is used to preserve referents. 4920 // If the object is non-null then it's alive. Used in the preservation 4921 // of referent objects that are pointed to by reference objects 4922 // discovered by the CM ref processor. 4923 class G1AlwaysAliveClosure: public BoolObjectClosure { 4924 G1CollectedHeap* _g1; 4925 public: 4926 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4927 void do_object(oop p) { assert(false, "Do not call."); } 4928 bool do_object_b(oop p) { 4929 if (p != NULL) { 4930 return true; 4931 } 4932 return false; 4933 } 4934 }; 4935 4936 bool G1STWIsAliveClosure::do_object_b(oop p) { 4937 // An object is reachable if it is outside the collection set, 4938 // or is inside and copied. 4939 return !_g1->obj_in_cs(p) || p->is_forwarded(); 4940 } 4941 4942 // Non Copying Keep Alive closure 4943 class G1KeepAliveClosure: public OopClosure { 4944 G1CollectedHeap* _g1; 4945 public: 4946 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4947 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 4948 void do_oop( oop* p) { 4949 oop obj = *p; 4950 4951 if (_g1->obj_in_cs(obj)) { 4952 assert( obj->is_forwarded(), "invariant" ); 4953 *p = obj->forwardee(); 4954 } 4955 } 4956 }; 4957 4958 // Copying Keep Alive closure - can be called from both 4959 // serial and parallel code as long as different worker 4960 // threads utilize different G1ParScanThreadState instances 4961 // and different queues. 4962 4963 class G1CopyingKeepAliveClosure: public OopClosure { 4964 G1CollectedHeap* _g1h; 4965 OopClosure* _copy_non_heap_obj_cl; 4966 OopsInHeapRegionClosure* _copy_perm_obj_cl; 4967 G1ParScanThreadState* _par_scan_state; 4968 4969 public: 4970 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 4971 OopClosure* non_heap_obj_cl, 4972 OopsInHeapRegionClosure* perm_obj_cl, 4973 G1ParScanThreadState* pss): 4974 _g1h(g1h), 4975 _copy_non_heap_obj_cl(non_heap_obj_cl), 4976 _copy_perm_obj_cl(perm_obj_cl), 4977 _par_scan_state(pss) 4978 {} 4979 4980 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 4981 virtual void do_oop( oop* p) { do_oop_work(p); } 4982 4983 template <class T> void do_oop_work(T* p) { 4984 oop obj = oopDesc::load_decode_heap_oop(p); 4985 4986 if (_g1h->obj_in_cs(obj)) { 4987 // If the referent object has been forwarded (either copied 4988 // to a new location or to itself in the event of an 4989 // evacuation failure) then we need to update the reference 4990 // field and, if both reference and referent are in the G1 4991 // heap, update the RSet for the referent. 4992 // 4993 // If the referent has not been forwarded then we have to keep 4994 // it alive by policy. Therefore we have copy the referent. 4995 // 4996 // If the reference field is in the G1 heap then we can push 4997 // on the PSS queue. When the queue is drained (after each 4998 // phase of reference processing) the object and it's followers 4999 // will be copied, the reference field set to point to the 5000 // new location, and the RSet updated. Otherwise we need to 5001 // use the the non-heap or perm closures directly to copy 5002 // the refernt object and update the pointer, while avoiding 5003 // updating the RSet. 5004 5005 if (_g1h->is_in_g1_reserved(p)) { 5006 _par_scan_state->push_on_queue(p); 5007 } else { 5008 // The reference field is not in the G1 heap. 5009 if (_g1h->perm_gen()->is_in(p)) { 5010 _copy_perm_obj_cl->do_oop(p); 5011 } else { 5012 _copy_non_heap_obj_cl->do_oop(p); 5013 } 5014 } 5015 } 5016 } 5017 }; 5018 5019 // Serial drain queue closure. Called as the 'complete_gc' 5020 // closure for each discovered list in some of the 5021 // reference processing phases. 5022 5023 class G1STWDrainQueueClosure: public VoidClosure { 5024 protected: 5025 G1CollectedHeap* _g1h; 5026 G1ParScanThreadState* _par_scan_state; 5027 5028 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5029 5030 public: 5031 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5032 _g1h(g1h), 5033 _par_scan_state(pss) 5034 { } 5035 5036 void do_void() { 5037 G1ParScanThreadState* const pss = par_scan_state(); 5038 pss->trim_queue(); 5039 } 5040 }; 5041 5042 // Parallel Reference Processing closures 5043 5044 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5045 // processing during G1 evacuation pauses. 5046 5047 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5048 private: 5049 G1CollectedHeap* _g1h; 5050 RefToScanQueueSet* _queues; 5051 FlexibleWorkGang* _workers; 5052 int _active_workers; 5053 5054 public: 5055 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5056 FlexibleWorkGang* workers, 5057 RefToScanQueueSet *task_queues, 5058 int n_workers) : 5059 _g1h(g1h), 5060 _queues(task_queues), 5061 _workers(workers), 5062 _active_workers(n_workers) 5063 { 5064 assert(n_workers > 0, "shouldn't call this otherwise"); 5065 } 5066 5067 // Executes the given task using concurrent marking worker threads. 5068 virtual void execute(ProcessTask& task); 5069 virtual void execute(EnqueueTask& task); 5070 }; 5071 5072 // Gang task for possibly parallel reference processing 5073 5074 class G1STWRefProcTaskProxy: public AbstractGangTask { 5075 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5076 ProcessTask& _proc_task; 5077 G1CollectedHeap* _g1h; 5078 RefToScanQueueSet *_task_queues; 5079 ParallelTaskTerminator* _terminator; 5080 5081 public: 5082 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5083 G1CollectedHeap* g1h, 5084 RefToScanQueueSet *task_queues, 5085 ParallelTaskTerminator* terminator) : 5086 AbstractGangTask("Process reference objects in parallel"), 5087 _proc_task(proc_task), 5088 _g1h(g1h), 5089 _task_queues(task_queues), 5090 _terminator(terminator) 5091 {} 5092 5093 virtual void work(int i) { 5094 // The reference processing task executed by a single worker. 5095 ResourceMark rm; 5096 HandleMark hm; 5097 5098 G1STWIsAliveClosure is_alive(_g1h); 5099 5100 G1ParScanThreadState pss(_g1h, i); 5101 5102 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5103 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5104 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5105 5106 pss.set_evac_closure(&scan_evac_cl); 5107 pss.set_evac_failure_closure(&evac_failure_cl); 5108 pss.set_partial_scan_closure(&partial_scan_cl); 5109 5110 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5111 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5112 5113 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5114 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5115 5116 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5117 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5118 5119 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5120 // We also need to mark copied objects. 5121 copy_non_heap_cl = ©_mark_non_heap_cl; 5122 copy_perm_cl = ©_mark_perm_cl; 5123 } 5124 5125 // Keep alive closure. 5126 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5127 5128 // Complete GC closure 5129 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5130 5131 // Call the reference processing task's work routine. 5132 _proc_task.work(i, is_alive, keep_alive, drain_queue); 5133 5134 // Note we cannot assert that the refs array is empty here as not all 5135 // of the processing tasks (specifically phase2 - pp2_work) execute 5136 // the complete_gc closure (which ordinarily would drain the queue) so 5137 // the queue may not be empty. 5138 } 5139 }; 5140 5141 // Driver routine for parallel reference processing. 5142 // Creates an instance of the ref processing gang 5143 // task and has the worker threads execute it. 5144 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5145 assert(_workers != NULL, "Need parallel worker threads."); 5146 5147 ParallelTaskTerminator terminator(_active_workers, _queues); 5148 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5149 5150 _g1h->set_par_threads(_active_workers); 5151 _workers->run_task(&proc_task_proxy); 5152 _g1h->set_par_threads(0); 5153 } 5154 5155 // Gang task for parallel reference enqueueing. 5156 5157 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5158 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5159 EnqueueTask& _enq_task; 5160 5161 public: 5162 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5163 AbstractGangTask("Enqueue reference objects in parallel"), 5164 _enq_task(enq_task) 5165 { } 5166 5167 virtual void work(int i) { 5168 _enq_task.work(i); 5169 } 5170 }; 5171 5172 // Driver routine for parallel reference enqueing. 5173 // Creates an instance of the ref enqueueing gang 5174 // task and has the worker threads execute it. 5175 5176 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5177 assert(_workers != NULL, "Need parallel worker threads."); 5178 5179 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5180 5181 _g1h->set_par_threads(_active_workers); 5182 _workers->run_task(&enq_task_proxy); 5183 _g1h->set_par_threads(0); 5184 } 5185 5186 // End of weak reference support closures 5187 5188 // Abstract task used to preserve (i.e. copy) any referent objects 5189 // that are in the collection set and are pointed to by reference 5190 // objects discovered by the CM ref processor. 5191 5192 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5193 protected: 5194 G1CollectedHeap* _g1h; 5195 RefToScanQueueSet *_queues; 5196 ParallelTaskTerminator _terminator; 5197 int _n_workers; 5198 5199 public: 5200 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5201 AbstractGangTask("ParPreserveCMReferents"), 5202 _g1h(g1h), 5203 _queues(task_queues), 5204 _terminator(workers, _queues), 5205 _n_workers(workers) 5206 { } 5207 5208 void work(int i) { 5209 ResourceMark rm; 5210 HandleMark hm; 5211 5212 G1ParScanThreadState pss(_g1h, i); 5213 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5214 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5215 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5216 5217 pss.set_evac_closure(&scan_evac_cl); 5218 pss.set_evac_failure_closure(&evac_failure_cl); 5219 pss.set_partial_scan_closure(&partial_scan_cl); 5220 5221 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5222 5223 5224 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5225 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5226 5227 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5228 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5229 5230 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5231 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5232 5233 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5234 // We also need to mark copied objects. 5235 copy_non_heap_cl = ©_mark_non_heap_cl; 5236 copy_perm_cl = ©_mark_perm_cl; 5237 } 5238 5239 // Is alive closure 5240 G1AlwaysAliveClosure always_alive(_g1h); 5241 5242 // Copying keep alive closure. Applied to referent objects that need 5243 // to be copied. 5244 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5245 5246 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5247 5248 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5249 int stride = MIN2(MAX2(_n_workers, 1), limit); 5250 5251 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5252 // So this must be true - but assert just in case someone decides to 5253 // change the worker ids. 5254 assert(0 <= i && i < limit, "sanity"); 5255 assert(!rp->discovery_is_atomic(), "check this code"); 5256 5257 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5258 for (int idx = i; idx < limit; idx += stride) { 5259 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5260 5261 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5262 while (iter.has_next()) { 5263 // Since discovery is not atomic for the CM ref processor, we 5264 // can see some null referent objects. 5265 iter.load_ptrs(DEBUG_ONLY(true)); 5266 oop ref = iter.obj(); 5267 5268 // This will filter nulls. 5269 if (iter.is_referent_alive()) { 5270 iter.make_referent_alive(); 5271 } 5272 iter.move_to_next(); 5273 } 5274 } 5275 5276 // Drain the queue - which may cause stealing 5277 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5278 drain_queue.do_void(); 5279 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5280 assert(pss.refs()->is_empty(), "should be"); 5281 } 5282 }; 5283 5284 // Weak Reference processing during an evacuation pause (part 1). 5285 void G1CollectedHeap::process_discovered_references() { 5286 double ref_proc_start = os::elapsedTime(); 5287 5288 ReferenceProcessor* rp = _ref_processor_stw; 5289 assert(rp->discovery_enabled(), "should have been enabled"); 5290 5291 // Any reference objects, in the collection set, that were 'discovered' 5292 // by the CM ref processor should have already been copied (either by 5293 // applying the external root copy closure to the discovered lists, or 5294 // by following an RSet entry). 5295 // 5296 // But some of the referents, that are in the collection set, that these 5297 // reference objects point to may not have been copied: the STW ref 5298 // processor would have seen that the reference object had already 5299 // been 'discovered' and would have skipped discovering the reference, 5300 // but would not have treated the reference object as a regular oop. 5301 // As a reult the copy closure would not have been applied to the 5302 // referent object. 5303 // 5304 // We need to explicitly copy these referent objects - the references 5305 // will be processed at the end of remarking. 5306 // 5307 // We also need to do this copying before we process the reference 5308 // objects discovered by the STW ref processor in case one of these 5309 // referents points to another object which is also referenced by an 5310 // object discovered by the STW ref processor. 5311 5312 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5313 workers()->active_workers() : 1); 5314 5315 assert(active_workers == workers()->active_workers(), 5316 "Need to reset active_workers"); 5317 set_par_threads(active_workers); 5318 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5319 5320 if (G1CollectedHeap::use_parallel_gc_threads()) { 5321 workers()->run_task(&keep_cm_referents); 5322 } else { 5323 keep_cm_referents.work(0); 5324 } 5325 5326 set_par_threads(0); 5327 5328 // Closure to test whether a referent is alive. 5329 G1STWIsAliveClosure is_alive(this); 5330 5331 // Even when parallel reference processing is enabled, the processing 5332 // of JNI refs is serial and performed serially by the current thread 5333 // rather than by a worker. The following PSS will be used for processing 5334 // JNI refs. 5335 5336 // Use only a single queue for this PSS. 5337 G1ParScanThreadState pss(this, 0); 5338 5339 // We do not embed a reference processor in the copying/scanning 5340 // closures while we're actually processing the discovered 5341 // reference objects. 5342 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5343 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5344 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5345 5346 pss.set_evac_closure(&scan_evac_cl); 5347 pss.set_evac_failure_closure(&evac_failure_cl); 5348 pss.set_partial_scan_closure(&partial_scan_cl); 5349 5350 assert(pss.refs()->is_empty(), "pre-condition"); 5351 5352 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5353 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5354 5355 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5356 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5357 5358 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5359 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5360 5361 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5362 // We also need to mark copied objects. 5363 copy_non_heap_cl = ©_mark_non_heap_cl; 5364 copy_perm_cl = ©_mark_perm_cl; 5365 } 5366 5367 // Keep alive closure. 5368 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5369 5370 // Serial Complete GC closure 5371 G1STWDrainQueueClosure drain_queue(this, &pss); 5372 5373 // Setup the soft refs policy... 5374 rp->setup_policy(false); 5375 5376 if (!rp->processing_is_mt()) { 5377 // Serial reference processing... 5378 rp->process_discovered_references(&is_alive, 5379 &keep_alive, 5380 &drain_queue, 5381 NULL); 5382 } else { 5383 // Parallel reference processing 5384 assert(rp->num_q() == active_workers, "sanity"); 5385 assert(active_workers <= rp->max_num_q(), "sanity"); 5386 5387 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5388 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5389 } 5390 5391 // We have completed copying any necessary live referent objects 5392 // (that were not copied during the actual pause) so we can 5393 // retire any active alloc buffers 5394 pss.retire_alloc_buffers(); 5395 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5396 5397 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5398 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0); 5399 } 5400 5401 // Weak Reference processing during an evacuation pause (part 2). 5402 void G1CollectedHeap::enqueue_discovered_references() { 5403 double ref_enq_start = os::elapsedTime(); 5404 5405 ReferenceProcessor* rp = _ref_processor_stw; 5406 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5407 5408 // Now enqueue any remaining on the discovered lists on to 5409 // the pending list. 5410 if (!rp->processing_is_mt()) { 5411 // Serial reference processing... 5412 rp->enqueue_discovered_references(); 5413 } else { 5414 // Parallel reference enqueuing 5415 5416 int active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5417 assert(active_workers == workers()->active_workers(), 5418 "Need to reset active_workers"); 5419 assert(rp->num_q() == active_workers, "sanity"); 5420 assert(active_workers <= rp->max_num_q(), "sanity"); 5421 5422 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5423 rp->enqueue_discovered_references(&par_task_executor); 5424 } 5425 5426 rp->verify_no_references_recorded(); 5427 assert(!rp->discovery_enabled(), "should have been disabled"); 5428 5429 // FIXME 5430 // CM's reference processing also cleans up the string and symbol tables. 5431 // Should we do that here also? We could, but it is a serial operation 5432 // and could signicantly increase the pause time. 5433 5434 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5435 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0); 5436 } 5437 5438 void G1CollectedHeap::evacuate_collection_set() { 5439 set_evacuation_failed(false); 5440 5441 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5442 concurrent_g1_refine()->set_use_cache(false); 5443 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5444 5445 int n_workers; 5446 if (G1CollectedHeap::use_parallel_gc_threads()) { 5447 n_workers = 5448 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5449 workers()->active_workers(), 5450 Threads::number_of_non_daemon_threads()); 5451 assert(UseDynamicNumberOfGCThreads || 5452 n_workers == workers()->total_workers(), 5453 "If not dynamic should be using all the workers"); 5454 set_par_threads(n_workers); 5455 } else { 5456 assert(n_par_threads() == 0, 5457 "Should be the original non-parallel value"); 5458 n_workers = 1; 5459 } 5460 workers()->set_active_workers(n_workers); 5461 5462 G1ParTask g1_par_task(this, _task_queues); 5463 5464 init_for_evac_failure(NULL); 5465 5466 rem_set()->prepare_for_younger_refs_iterate(true); 5467 5468 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5469 double start_par = os::elapsedTime(); 5470 5471 if (G1CollectedHeap::use_parallel_gc_threads()) { 5472 // The individual threads will set their evac-failure closures. 5473 StrongRootsScope srs(this); 5474 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5475 // These tasks use ShareHeap::_process_strong_tasks 5476 assert(UseDynamicNumberOfGCThreads || 5477 workers()->active_workers() == workers()->total_workers(), 5478 "If not dynamic should be using all the workers"); 5479 workers()->run_task(&g1_par_task); 5480 } else { 5481 StrongRootsScope srs(this); 5482 g1_par_task.work(0); 5483 } 5484 5485 double par_time = (os::elapsedTime() - start_par) * 1000.0; 5486 g1_policy()->record_par_time(par_time); 5487 5488 set_par_threads(0); 5489 5490 // Process any discovered reference objects - we have 5491 // to do this _before_ we retire the GC alloc regions 5492 // as we may have to copy some 'reachable' referent 5493 // objects (and their reachable sub-graphs) that were 5494 // not copied during the pause. 5495 process_discovered_references(); 5496 5497 // Weak root processing. 5498 // Note: when JSR 292 is enabled and code blobs can contain 5499 // non-perm oops then we will need to process the code blobs 5500 // here too. 5501 { 5502 G1STWIsAliveClosure is_alive(this); 5503 G1KeepAliveClosure keep_alive(this); 5504 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5505 } 5506 5507 release_gc_alloc_regions(); 5508 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5509 5510 concurrent_g1_refine()->clear_hot_cache(); 5511 concurrent_g1_refine()->set_use_cache(true); 5512 5513 finalize_for_evac_failure(); 5514 5515 // Must do this before clearing the per-region evac-failure flags 5516 // (which is currently done when we free the collection set). 5517 // We also only do this if marking is actually in progress and so 5518 // have to do this before we set the mark_in_progress flag at the 5519 // end of an initial mark pause. 5520 concurrent_mark()->complete_marking_in_collection_set(); 5521 5522 if (evacuation_failed()) { 5523 remove_self_forwarding_pointers(); 5524 if (PrintGCDetails) { 5525 gclog_or_tty->print(" (to-space overflow)"); 5526 } else if (PrintGC) { 5527 gclog_or_tty->print("--"); 5528 } 5529 } 5530 5531 // Enqueue any remaining references remaining on the STW 5532 // reference processor's discovered lists. We need to do 5533 // this after the card table is cleaned (and verified) as 5534 // the act of enqueuing entries on to the pending list 5535 // will log these updates (and dirty their associated 5536 // cards). We need these updates logged to update any 5537 // RSets. 5538 enqueue_discovered_references(); 5539 5540 if (G1DeferredRSUpdate) { 5541 RedirtyLoggedCardTableEntryFastClosure redirty; 5542 dirty_card_queue_set().set_closure(&redirty); 5543 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5544 5545 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5546 dcq.merge_bufferlists(&dirty_card_queue_set()); 5547 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5548 } 5549 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5550 } 5551 5552 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5553 size_t* pre_used, 5554 FreeRegionList* free_list, 5555 OldRegionSet* old_proxy_set, 5556 HumongousRegionSet* humongous_proxy_set, 5557 HRRSCleanupTask* hrrs_cleanup_task, 5558 bool par) { 5559 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5560 if (hr->isHumongous()) { 5561 assert(hr->startsHumongous(), "we should only see starts humongous"); 5562 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5563 } else { 5564 _old_set.remove_with_proxy(hr, old_proxy_set); 5565 free_region(hr, pre_used, free_list, par); 5566 } 5567 } else { 5568 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5569 } 5570 } 5571 5572 void G1CollectedHeap::free_region(HeapRegion* hr, 5573 size_t* pre_used, 5574 FreeRegionList* free_list, 5575 bool par) { 5576 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5577 assert(!hr->is_empty(), "the region should not be empty"); 5578 assert(free_list != NULL, "pre-condition"); 5579 5580 *pre_used += hr->used(); 5581 hr->hr_clear(par, true /* clear_space */); 5582 free_list->add_as_head(hr); 5583 } 5584 5585 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5586 size_t* pre_used, 5587 FreeRegionList* free_list, 5588 HumongousRegionSet* humongous_proxy_set, 5589 bool par) { 5590 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5591 assert(free_list != NULL, "pre-condition"); 5592 assert(humongous_proxy_set != NULL, "pre-condition"); 5593 5594 size_t hr_used = hr->used(); 5595 size_t hr_capacity = hr->capacity(); 5596 size_t hr_pre_used = 0; 5597 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5598 hr->set_notHumongous(); 5599 free_region(hr, &hr_pre_used, free_list, par); 5600 5601 size_t i = hr->hrs_index() + 1; 5602 size_t num = 1; 5603 while (i < n_regions()) { 5604 HeapRegion* curr_hr = region_at(i); 5605 if (!curr_hr->continuesHumongous()) { 5606 break; 5607 } 5608 curr_hr->set_notHumongous(); 5609 free_region(curr_hr, &hr_pre_used, free_list, par); 5610 num += 1; 5611 i += 1; 5612 } 5613 assert(hr_pre_used == hr_used, 5614 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5615 "should be the same", hr_pre_used, hr_used)); 5616 *pre_used += hr_pre_used; 5617 } 5618 5619 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5620 FreeRegionList* free_list, 5621 OldRegionSet* old_proxy_set, 5622 HumongousRegionSet* humongous_proxy_set, 5623 bool par) { 5624 if (pre_used > 0) { 5625 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5626 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5627 assert(_summary_bytes_used >= pre_used, 5628 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5629 "should be >= pre_used: "SIZE_FORMAT, 5630 _summary_bytes_used, pre_used)); 5631 _summary_bytes_used -= pre_used; 5632 } 5633 if (free_list != NULL && !free_list->is_empty()) { 5634 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5635 _free_list.add_as_head(free_list); 5636 } 5637 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5638 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5639 _old_set.update_from_proxy(old_proxy_set); 5640 } 5641 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5642 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5643 _humongous_set.update_from_proxy(humongous_proxy_set); 5644 } 5645 } 5646 5647 class G1ParCleanupCTTask : public AbstractGangTask { 5648 CardTableModRefBS* _ct_bs; 5649 G1CollectedHeap* _g1h; 5650 HeapRegion* volatile _su_head; 5651 public: 5652 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5653 G1CollectedHeap* g1h) : 5654 AbstractGangTask("G1 Par Cleanup CT Task"), 5655 _ct_bs(ct_bs), _g1h(g1h) { } 5656 5657 void work(int i) { 5658 HeapRegion* r; 5659 while (r = _g1h->pop_dirty_cards_region()) { 5660 clear_cards(r); 5661 } 5662 } 5663 5664 void clear_cards(HeapRegion* r) { 5665 // Cards of the survivors should have already been dirtied. 5666 if (!r->is_survivor()) { 5667 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5668 } 5669 } 5670 }; 5671 5672 #ifndef PRODUCT 5673 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5674 G1CollectedHeap* _g1h; 5675 CardTableModRefBS* _ct_bs; 5676 public: 5677 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5678 : _g1h(g1h), _ct_bs(ct_bs) { } 5679 virtual bool doHeapRegion(HeapRegion* r) { 5680 if (r->is_survivor()) { 5681 _g1h->verify_dirty_region(r); 5682 } else { 5683 _g1h->verify_not_dirty_region(r); 5684 } 5685 return false; 5686 } 5687 }; 5688 5689 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5690 // All of the region should be clean. 5691 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5692 MemRegion mr(hr->bottom(), hr->end()); 5693 ct_bs->verify_not_dirty_region(mr); 5694 } 5695 5696 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5697 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5698 // dirty allocated blocks as they allocate them. The thread that 5699 // retires each region and replaces it with a new one will do a 5700 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5701 // not dirty that area (one less thing to have to do while holding 5702 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5703 // is dirty. 5704 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5705 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5706 ct_bs->verify_dirty_region(mr); 5707 } 5708 5709 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5710 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5711 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5712 verify_dirty_region(hr); 5713 } 5714 } 5715 5716 void G1CollectedHeap::verify_dirty_young_regions() { 5717 verify_dirty_young_list(_young_list->first_region()); 5718 verify_dirty_young_list(_young_list->first_survivor_region()); 5719 } 5720 #endif 5721 5722 void G1CollectedHeap::cleanUpCardTable() { 5723 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5724 double start = os::elapsedTime(); 5725 5726 { 5727 // Iterate over the dirty cards region list. 5728 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5729 5730 if (ParallelGCThreads > 0) { 5731 set_par_threads(workers()->total_workers()); 5732 workers()->run_task(&cleanup_task); 5733 set_par_threads(0); 5734 } else { 5735 while (_dirty_cards_region_list) { 5736 HeapRegion* r = _dirty_cards_region_list; 5737 cleanup_task.clear_cards(r); 5738 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5739 if (_dirty_cards_region_list == r) { 5740 // The last region. 5741 _dirty_cards_region_list = NULL; 5742 } 5743 r->set_next_dirty_cards_region(NULL); 5744 } 5745 } 5746 #ifndef PRODUCT 5747 if (G1VerifyCTCleanup || VerifyAfterGC) { 5748 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5749 heap_region_iterate(&cleanup_verifier); 5750 } 5751 #endif 5752 } 5753 5754 double elapsed = os::elapsedTime() - start; 5755 g1_policy()->record_clear_ct_time(elapsed * 1000.0); 5756 } 5757 5758 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5759 size_t pre_used = 0; 5760 FreeRegionList local_free_list("Local List for CSet Freeing"); 5761 5762 double young_time_ms = 0.0; 5763 double non_young_time_ms = 0.0; 5764 5765 // Since the collection set is a superset of the the young list, 5766 // all we need to do to clear the young list is clear its 5767 // head and length, and unlink any young regions in the code below 5768 _young_list->clear(); 5769 5770 G1CollectorPolicy* policy = g1_policy(); 5771 5772 double start_sec = os::elapsedTime(); 5773 bool non_young = true; 5774 5775 HeapRegion* cur = cs_head; 5776 int age_bound = -1; 5777 size_t rs_lengths = 0; 5778 5779 while (cur != NULL) { 5780 assert(!is_on_master_free_list(cur), "sanity"); 5781 if (non_young) { 5782 if (cur->is_young()) { 5783 double end_sec = os::elapsedTime(); 5784 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5785 non_young_time_ms += elapsed_ms; 5786 5787 start_sec = os::elapsedTime(); 5788 non_young = false; 5789 } 5790 } else { 5791 if (!cur->is_young()) { 5792 double end_sec = os::elapsedTime(); 5793 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5794 young_time_ms += elapsed_ms; 5795 5796 start_sec = os::elapsedTime(); 5797 non_young = true; 5798 } 5799 } 5800 5801 rs_lengths += cur->rem_set()->occupied(); 5802 5803 HeapRegion* next = cur->next_in_collection_set(); 5804 assert(cur->in_collection_set(), "bad CS"); 5805 cur->set_next_in_collection_set(NULL); 5806 cur->set_in_collection_set(false); 5807 5808 if (cur->is_young()) { 5809 int index = cur->young_index_in_cset(); 5810 assert(index != -1, "invariant"); 5811 assert((size_t) index < policy->young_cset_region_length(), "invariant"); 5812 size_t words_survived = _surviving_young_words[index]; 5813 cur->record_surv_words_in_group(words_survived); 5814 5815 // At this point the we have 'popped' cur from the collection set 5816 // (linked via next_in_collection_set()) but it is still in the 5817 // young list (linked via next_young_region()). Clear the 5818 // _next_young_region field. 5819 cur->set_next_young_region(NULL); 5820 } else { 5821 int index = cur->young_index_in_cset(); 5822 assert(index == -1, "invariant"); 5823 } 5824 5825 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5826 (!cur->is_young() && cur->young_index_in_cset() == -1), 5827 "invariant" ); 5828 5829 if (!cur->evacuation_failed()) { 5830 MemRegion used_mr = cur->used_region(); 5831 5832 // And the region is empty. 5833 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5834 5835 // If marking is in progress then clear any objects marked in 5836 // the current region. Note mark_in_progress() returns false, 5837 // even during an initial mark pause, until the set_marking_started() 5838 // call which takes place later in the pause. 5839 if (mark_in_progress()) { 5840 assert(!g1_policy()->during_initial_mark_pause(), "sanity"); 5841 _cm->nextMarkBitMap()->clearRange(used_mr); 5842 } 5843 5844 free_region(cur, &pre_used, &local_free_list, false /* par */); 5845 } else { 5846 cur->uninstall_surv_rate_group(); 5847 if (cur->is_young()) { 5848 cur->set_young_index_in_cset(-1); 5849 } 5850 cur->set_not_young(); 5851 cur->set_evacuation_failed(false); 5852 // The region is now considered to be old. 5853 _old_set.add(cur); 5854 } 5855 cur = next; 5856 } 5857 5858 policy->record_max_rs_lengths(rs_lengths); 5859 policy->cset_regions_freed(); 5860 5861 double end_sec = os::elapsedTime(); 5862 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5863 5864 if (non_young) { 5865 non_young_time_ms += elapsed_ms; 5866 } else { 5867 young_time_ms += elapsed_ms; 5868 } 5869 5870 update_sets_after_freeing_regions(pre_used, &local_free_list, 5871 NULL /* old_proxy_set */, 5872 NULL /* humongous_proxy_set */, 5873 false /* par */); 5874 policy->record_young_free_cset_time_ms(young_time_ms); 5875 policy->record_non_young_free_cset_time_ms(non_young_time_ms); 5876 } 5877 5878 // This routine is similar to the above but does not record 5879 // any policy statistics or update free lists; we are abandoning 5880 // the current incremental collection set in preparation of a 5881 // full collection. After the full GC we will start to build up 5882 // the incremental collection set again. 5883 // This is only called when we're doing a full collection 5884 // and is immediately followed by the tearing down of the young list. 5885 5886 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5887 HeapRegion* cur = cs_head; 5888 5889 while (cur != NULL) { 5890 HeapRegion* next = cur->next_in_collection_set(); 5891 assert(cur->in_collection_set(), "bad CS"); 5892 cur->set_next_in_collection_set(NULL); 5893 cur->set_in_collection_set(false); 5894 cur->set_young_index_in_cset(-1); 5895 cur = next; 5896 } 5897 } 5898 5899 void G1CollectedHeap::set_free_regions_coming() { 5900 if (G1ConcRegionFreeingVerbose) { 5901 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5902 "setting free regions coming"); 5903 } 5904 5905 assert(!free_regions_coming(), "pre-condition"); 5906 _free_regions_coming = true; 5907 } 5908 5909 void G1CollectedHeap::reset_free_regions_coming() { 5910 { 5911 assert(free_regions_coming(), "pre-condition"); 5912 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5913 _free_regions_coming = false; 5914 SecondaryFreeList_lock->notify_all(); 5915 } 5916 5917 if (G1ConcRegionFreeingVerbose) { 5918 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5919 "reset free regions coming"); 5920 } 5921 } 5922 5923 void G1CollectedHeap::wait_while_free_regions_coming() { 5924 // Most of the time we won't have to wait, so let's do a quick test 5925 // first before we take the lock. 5926 if (!free_regions_coming()) { 5927 return; 5928 } 5929 5930 if (G1ConcRegionFreeingVerbose) { 5931 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5932 "waiting for free regions"); 5933 } 5934 5935 { 5936 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5937 while (free_regions_coming()) { 5938 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 5939 } 5940 } 5941 5942 if (G1ConcRegionFreeingVerbose) { 5943 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5944 "done waiting for free regions"); 5945 } 5946 } 5947 5948 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 5949 assert(heap_lock_held_for_gc(), 5950 "the heap lock should already be held by or for this thread"); 5951 _young_list->push_region(hr); 5952 } 5953 5954 class NoYoungRegionsClosure: public HeapRegionClosure { 5955 private: 5956 bool _success; 5957 public: 5958 NoYoungRegionsClosure() : _success(true) { } 5959 bool doHeapRegion(HeapRegion* r) { 5960 if (r->is_young()) { 5961 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 5962 r->bottom(), r->end()); 5963 _success = false; 5964 } 5965 return false; 5966 } 5967 bool success() { return _success; } 5968 }; 5969 5970 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 5971 bool ret = _young_list->check_list_empty(check_sample); 5972 5973 if (check_heap) { 5974 NoYoungRegionsClosure closure; 5975 heap_region_iterate(&closure); 5976 ret = ret && closure.success(); 5977 } 5978 5979 return ret; 5980 } 5981 5982 class TearDownRegionSetsClosure : public HeapRegionClosure { 5983 private: 5984 OldRegionSet *_old_set; 5985 5986 public: 5987 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 5988 5989 bool doHeapRegion(HeapRegion* r) { 5990 if (r->is_empty()) { 5991 // We ignore empty regions, we'll empty the free list afterwards 5992 } else if (r->is_young()) { 5993 // We ignore young regions, we'll empty the young list afterwards 5994 } else if (r->isHumongous()) { 5995 // We ignore humongous regions, we're not tearing down the 5996 // humongous region set 5997 } else { 5998 // The rest should be old 5999 _old_set->remove(r); 6000 } 6001 return false; 6002 } 6003 6004 ~TearDownRegionSetsClosure() { 6005 assert(_old_set->is_empty(), "post-condition"); 6006 } 6007 }; 6008 6009 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6010 assert_at_safepoint(true /* should_be_vm_thread */); 6011 6012 if (!free_list_only) { 6013 TearDownRegionSetsClosure cl(&_old_set); 6014 heap_region_iterate(&cl); 6015 6016 // Need to do this after the heap iteration to be able to 6017 // recognize the young regions and ignore them during the iteration. 6018 _young_list->empty_list(); 6019 } 6020 _free_list.remove_all(); 6021 } 6022 6023 class RebuildRegionSetsClosure : public HeapRegionClosure { 6024 private: 6025 bool _free_list_only; 6026 OldRegionSet* _old_set; 6027 FreeRegionList* _free_list; 6028 size_t _total_used; 6029 6030 public: 6031 RebuildRegionSetsClosure(bool free_list_only, 6032 OldRegionSet* old_set, FreeRegionList* free_list) : 6033 _free_list_only(free_list_only), 6034 _old_set(old_set), _free_list(free_list), _total_used(0) { 6035 assert(_free_list->is_empty(), "pre-condition"); 6036 if (!free_list_only) { 6037 assert(_old_set->is_empty(), "pre-condition"); 6038 } 6039 } 6040 6041 bool doHeapRegion(HeapRegion* r) { 6042 if (r->continuesHumongous()) { 6043 return false; 6044 } 6045 6046 if (r->is_empty()) { 6047 // Add free regions to the free list 6048 _free_list->add_as_tail(r); 6049 } else if (!_free_list_only) { 6050 assert(!r->is_young(), "we should not come across young regions"); 6051 6052 if (r->isHumongous()) { 6053 // We ignore humongous regions, we left the humongous set unchanged 6054 } else { 6055 // The rest should be old, add them to the old set 6056 _old_set->add(r); 6057 } 6058 _total_used += r->used(); 6059 } 6060 6061 return false; 6062 } 6063 6064 size_t total_used() { 6065 return _total_used; 6066 } 6067 }; 6068 6069 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6070 assert_at_safepoint(true /* should_be_vm_thread */); 6071 6072 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6073 heap_region_iterate(&cl); 6074 6075 if (!free_list_only) { 6076 _summary_bytes_used = cl.total_used(); 6077 } 6078 assert(_summary_bytes_used == recalculate_used(), 6079 err_msg("inconsistent _summary_bytes_used, " 6080 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6081 _summary_bytes_used, recalculate_used())); 6082 } 6083 6084 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6085 _refine_cte_cl->set_concurrent(concurrent); 6086 } 6087 6088 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6089 HeapRegion* hr = heap_region_containing(p); 6090 if (hr == NULL) { 6091 return is_in_permanent(p); 6092 } else { 6093 return hr->is_in(p); 6094 } 6095 } 6096 6097 // Methods for the mutator alloc region 6098 6099 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6100 bool force) { 6101 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6102 assert(!force || g1_policy()->can_expand_young_list(), 6103 "if force is true we should be able to expand the young list"); 6104 bool young_list_full = g1_policy()->is_young_list_full(); 6105 if (force || !young_list_full) { 6106 HeapRegion* new_alloc_region = new_region(word_size, 6107 false /* do_expand */); 6108 if (new_alloc_region != NULL) { 6109 set_region_short_lived_locked(new_alloc_region); 6110 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6111 return new_alloc_region; 6112 } 6113 } 6114 return NULL; 6115 } 6116 6117 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6118 size_t allocated_bytes) { 6119 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6120 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6121 6122 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6123 _summary_bytes_used += allocated_bytes; 6124 _hr_printer.retire(alloc_region); 6125 // We update the eden sizes here, when the region is retired, 6126 // instead of when it's allocated, since this is the point that its 6127 // used space has been recored in _summary_bytes_used. 6128 g1mm()->update_eden_size(); 6129 } 6130 6131 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6132 bool force) { 6133 return _g1h->new_mutator_alloc_region(word_size, force); 6134 } 6135 6136 void G1CollectedHeap::set_par_threads() { 6137 // Don't change the number of workers. Use the value previously set 6138 // in the workgroup. 6139 int n_workers = workers()->active_workers(); 6140 assert(UseDynamicNumberOfGCThreads || 6141 n_workers == workers()->total_workers(), 6142 "Otherwise should be using the total number of workers"); 6143 if (n_workers == 0) { 6144 assert(false, "Should have been set in prior evacuation pause."); 6145 n_workers = ParallelGCThreads; 6146 workers()->set_active_workers(n_workers); 6147 } 6148 set_par_threads(n_workers); 6149 } 6150 6151 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6152 size_t allocated_bytes) { 6153 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6154 } 6155 6156 // Methods for the GC alloc regions 6157 6158 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6159 size_t count, 6160 GCAllocPurpose ap) { 6161 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6162 6163 if (count < g1_policy()->max_regions(ap)) { 6164 HeapRegion* new_alloc_region = new_region(word_size, 6165 true /* do_expand */); 6166 if (new_alloc_region != NULL) { 6167 // We really only need to do this for old regions given that we 6168 // should never scan survivors. But it doesn't hurt to do it 6169 // for survivors too. 6170 new_alloc_region->set_saved_mark(); 6171 if (ap == GCAllocForSurvived) { 6172 new_alloc_region->set_survivor(); 6173 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6174 } else { 6175 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6176 } 6177 return new_alloc_region; 6178 } else { 6179 g1_policy()->note_alloc_region_limit_reached(ap); 6180 } 6181 } 6182 return NULL; 6183 } 6184 6185 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6186 size_t allocated_bytes, 6187 GCAllocPurpose ap) { 6188 alloc_region->note_end_of_copying(); 6189 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6190 if (ap == GCAllocForSurvived) { 6191 young_list()->add_survivor_region(alloc_region); 6192 } else { 6193 _old_set.add(alloc_region); 6194 } 6195 _hr_printer.retire(alloc_region); 6196 } 6197 6198 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6199 bool force) { 6200 assert(!force, "not supported for GC alloc regions"); 6201 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6202 } 6203 6204 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6205 size_t allocated_bytes) { 6206 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6207 GCAllocForSurvived); 6208 } 6209 6210 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6211 bool force) { 6212 assert(!force, "not supported for GC alloc regions"); 6213 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6214 } 6215 6216 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6217 size_t allocated_bytes) { 6218 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6219 GCAllocForTenured); 6220 } 6221 // Heap region set verification 6222 6223 class VerifyRegionListsClosure : public HeapRegionClosure { 6224 private: 6225 FreeRegionList* _free_list; 6226 OldRegionSet* _old_set; 6227 HumongousRegionSet* _humongous_set; 6228 size_t _region_count; 6229 6230 public: 6231 VerifyRegionListsClosure(OldRegionSet* old_set, 6232 HumongousRegionSet* humongous_set, 6233 FreeRegionList* free_list) : 6234 _old_set(old_set), _humongous_set(humongous_set), 6235 _free_list(free_list), _region_count(0) { } 6236 6237 size_t region_count() { return _region_count; } 6238 6239 bool doHeapRegion(HeapRegion* hr) { 6240 _region_count += 1; 6241 6242 if (hr->continuesHumongous()) { 6243 return false; 6244 } 6245 6246 if (hr->is_young()) { 6247 // TODO 6248 } else if (hr->startsHumongous()) { 6249 _humongous_set->verify_next_region(hr); 6250 } else if (hr->is_empty()) { 6251 _free_list->verify_next_region(hr); 6252 } else { 6253 _old_set->verify_next_region(hr); 6254 } 6255 return false; 6256 } 6257 }; 6258 6259 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index, 6260 HeapWord* bottom) { 6261 HeapWord* end = bottom + HeapRegion::GrainWords; 6262 MemRegion mr(bottom, end); 6263 assert(_g1_reserved.contains(mr), "invariant"); 6264 // This might return NULL if the allocation fails 6265 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6266 } 6267 6268 void G1CollectedHeap::verify_region_sets() { 6269 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6270 6271 // First, check the explicit lists. 6272 _free_list.verify(); 6273 { 6274 // Given that a concurrent operation might be adding regions to 6275 // the secondary free list we have to take the lock before 6276 // verifying it. 6277 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6278 _secondary_free_list.verify(); 6279 } 6280 _old_set.verify(); 6281 _humongous_set.verify(); 6282 6283 // If a concurrent region freeing operation is in progress it will 6284 // be difficult to correctly attributed any free regions we come 6285 // across to the correct free list given that they might belong to 6286 // one of several (free_list, secondary_free_list, any local lists, 6287 // etc.). So, if that's the case we will skip the rest of the 6288 // verification operation. Alternatively, waiting for the concurrent 6289 // operation to complete will have a non-trivial effect on the GC's 6290 // operation (no concurrent operation will last longer than the 6291 // interval between two calls to verification) and it might hide 6292 // any issues that we would like to catch during testing. 6293 if (free_regions_coming()) { 6294 return; 6295 } 6296 6297 // Make sure we append the secondary_free_list on the free_list so 6298 // that all free regions we will come across can be safely 6299 // attributed to the free_list. 6300 append_secondary_free_list_if_not_empty_with_lock(); 6301 6302 // Finally, make sure that the region accounting in the lists is 6303 // consistent with what we see in the heap. 6304 _old_set.verify_start(); 6305 _humongous_set.verify_start(); 6306 _free_list.verify_start(); 6307 6308 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6309 heap_region_iterate(&cl); 6310 6311 _old_set.verify_end(); 6312 _humongous_set.verify_end(); 6313 _free_list.verify_end(); 6314 }