1 /* 2 * Copyright (c) 2001, 2012, 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/g1EvacFailure.hpp" 36 #include "gc_implementation/g1/g1Log.hpp" 37 #include "gc_implementation/g1/g1MarkSweep.hpp" 38 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 39 #include "gc_implementation/g1/g1RemSet.inline.hpp" 40 #include "gc_implementation/g1/heapRegion.inline.hpp" 41 #include "gc_implementation/g1/heapRegionRemSet.hpp" 42 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 43 #include "gc_implementation/g1/vm_operations_g1.hpp" 44 #include "gc_implementation/shared/isGCActiveMark.hpp" 45 #include "memory/gcLocker.inline.hpp" 46 #include "memory/genOopClosures.inline.hpp" 47 #include "memory/generationSpec.hpp" 48 #include "memory/referenceProcessor.hpp" 49 #include "oops/oop.inline.hpp" 50 #include "oops/oop.pcgc.inline.hpp" 51 #include "runtime/aprofiler.hpp" 52 #include "runtime/vmThread.hpp" 53 54 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 55 56 // turn it on so that the contents of the young list (scan-only / 57 // to-be-collected) are printed at "strategic" points before / during 58 // / after the collection --- this is useful for debugging 59 #define YOUNG_LIST_VERBOSE 0 60 // CURRENT STATUS 61 // This file is under construction. Search for "FIXME". 62 63 // INVARIANTS/NOTES 64 // 65 // All allocation activity covered by the G1CollectedHeap interface is 66 // serialized by acquiring the HeapLock. This happens in mem_allocate 67 // and allocate_new_tlab, which are the "entry" points to the 68 // allocation code from the rest of the JVM. (Note that this does not 69 // apply to TLAB allocation, which is not part of this interface: it 70 // is done by clients of this interface.) 71 72 // Notes on implementation of parallelism in different tasks. 73 // 74 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 75 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 76 // It does use run_task() which sets _n_workers in the task. 77 // G1ParTask executes g1_process_strong_roots() -> 78 // SharedHeap::process_strong_roots() which calls eventuall to 79 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 80 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 81 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 82 // 83 84 // Local to this file. 85 86 class RefineCardTableEntryClosure: public CardTableEntryClosure { 87 SuspendibleThreadSet* _sts; 88 G1RemSet* _g1rs; 89 ConcurrentG1Refine* _cg1r; 90 bool _concurrent; 91 public: 92 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 93 G1RemSet* g1rs, 94 ConcurrentG1Refine* cg1r) : 95 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 96 {} 97 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 98 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false); 99 // This path is executed by the concurrent refine or mutator threads, 100 // concurrently, and so we do not care if card_ptr contains references 101 // that point into the collection set. 102 assert(!oops_into_cset, "should be"); 103 104 if (_concurrent && _sts->should_yield()) { 105 // Caller will actually yield. 106 return false; 107 } 108 // Otherwise, we finished successfully; return true. 109 return true; 110 } 111 void set_concurrent(bool b) { _concurrent = b; } 112 }; 113 114 115 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 116 int _calls; 117 G1CollectedHeap* _g1h; 118 CardTableModRefBS* _ctbs; 119 int _histo[256]; 120 public: 121 ClearLoggedCardTableEntryClosure() : 122 _calls(0) 123 { 124 _g1h = G1CollectedHeap::heap(); 125 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 126 for (int i = 0; i < 256; i++) _histo[i] = 0; 127 } 128 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 129 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 130 _calls++; 131 unsigned char* ujb = (unsigned char*)card_ptr; 132 int ind = (int)(*ujb); 133 _histo[ind]++; 134 *card_ptr = -1; 135 } 136 return true; 137 } 138 int calls() { return _calls; } 139 void print_histo() { 140 gclog_or_tty->print_cr("Card table value histogram:"); 141 for (int i = 0; i < 256; i++) { 142 if (_histo[i] != 0) { 143 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 144 } 145 } 146 } 147 }; 148 149 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 150 int _calls; 151 G1CollectedHeap* _g1h; 152 CardTableModRefBS* _ctbs; 153 public: 154 RedirtyLoggedCardTableEntryClosure() : 155 _calls(0) 156 { 157 _g1h = G1CollectedHeap::heap(); 158 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 159 } 160 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 161 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 162 _calls++; 163 *card_ptr = 0; 164 } 165 return true; 166 } 167 int calls() { return _calls; } 168 }; 169 170 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 171 public: 172 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 173 *card_ptr = CardTableModRefBS::dirty_card_val(); 174 return true; 175 } 176 }; 177 178 YoungList::YoungList(G1CollectedHeap* g1h) : 179 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 180 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 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 uint 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 %u entries, _length is %u", 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 %u", 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((uint) young_index_in_cset == _survivor_length, "post-condition"); 341 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 342 343 _head = _survivor_head; 344 _length = _survivor_length; 345 if (_survivor_head != NULL) { 346 assert(_survivor_tail != NULL, "cause it shouldn't be"); 347 assert(_survivor_length > 0, "invariant"); 348 _survivor_tail->set_next_young_region(NULL); 349 } 350 351 // Don't clear the survivor list handles until the start of 352 // the next evacuation pause - we need it in order to re-tag 353 // the survivor regions from this evacuation pause as 'young' 354 // at the start of the next. 355 356 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 357 358 assert(check_list_well_formed(), "young list should be well formed"); 359 } 360 361 void YoungList::print() { 362 HeapRegion* lists[] = {_head, _survivor_head}; 363 const char* names[] = {"YOUNG", "SURVIVOR"}; 364 365 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 366 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 367 HeapRegion *curr = lists[list]; 368 if (curr == NULL) 369 gclog_or_tty->print_cr(" empty"); 370 while (curr != NULL) { 371 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 372 HR_FORMAT_PARAMS(curr), 373 curr->prev_top_at_mark_start(), 374 curr->next_top_at_mark_start(), 375 curr->age_in_surv_rate_group_cond()); 376 curr = curr->get_next_young_region(); 377 } 378 } 379 380 gclog_or_tty->print_cr(""); 381 } 382 383 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 384 { 385 // Claim the right to put the region on the dirty cards region list 386 // by installing a self pointer. 387 HeapRegion* next = hr->get_next_dirty_cards_region(); 388 if (next == NULL) { 389 HeapRegion* res = (HeapRegion*) 390 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 391 NULL); 392 if (res == NULL) { 393 HeapRegion* head; 394 do { 395 // Put the region to the dirty cards region list. 396 head = _dirty_cards_region_list; 397 next = (HeapRegion*) 398 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 399 if (next == head) { 400 assert(hr->get_next_dirty_cards_region() == hr, 401 "hr->get_next_dirty_cards_region() != hr"); 402 if (next == NULL) { 403 // The last region in the list points to itself. 404 hr->set_next_dirty_cards_region(hr); 405 } else { 406 hr->set_next_dirty_cards_region(next); 407 } 408 } 409 } while (next != head); 410 } 411 } 412 } 413 414 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 415 { 416 HeapRegion* head; 417 HeapRegion* hr; 418 do { 419 head = _dirty_cards_region_list; 420 if (head == NULL) { 421 return NULL; 422 } 423 HeapRegion* new_head = head->get_next_dirty_cards_region(); 424 if (head == new_head) { 425 // The last region. 426 new_head = NULL; 427 } 428 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 429 head); 430 } while (hr != head); 431 assert(hr != NULL, "invariant"); 432 hr->set_next_dirty_cards_region(NULL); 433 return hr; 434 } 435 436 void G1CollectedHeap::stop_conc_gc_threads() { 437 _cg1r->stop(); 438 _cmThread->stop(); 439 } 440 441 #ifdef ASSERT 442 // A region is added to the collection set as it is retired 443 // so an address p can point to a region which will be in the 444 // collection set but has not yet been retired. This method 445 // therefore is only accurate during a GC pause after all 446 // regions have been retired. It is used for debugging 447 // to check if an nmethod has references to objects that can 448 // be move during a partial collection. Though it can be 449 // inaccurate, it is sufficient for G1 because the conservative 450 // implementation of is_scavengable() for G1 will indicate that 451 // all nmethods must be scanned during a partial collection. 452 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 453 HeapRegion* hr = heap_region_containing(p); 454 return hr != NULL && hr->in_collection_set(); 455 } 456 #endif 457 458 // Returns true if the reference points to an object that 459 // can move in an incremental collecction. 460 bool G1CollectedHeap::is_scavengable(const void* p) { 461 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 462 G1CollectorPolicy* g1p = g1h->g1_policy(); 463 HeapRegion* hr = heap_region_containing(p); 464 if (hr == NULL) { 465 // perm gen (or null) 466 return false; 467 } else { 468 return !hr->isHumongous(); 469 } 470 } 471 472 void G1CollectedHeap::check_ct_logs_at_safepoint() { 473 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 474 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 475 476 // Count the dirty cards at the start. 477 CountNonCleanMemRegionClosure count1(this); 478 ct_bs->mod_card_iterate(&count1); 479 int orig_count = count1.n(); 480 481 // First clear the logged cards. 482 ClearLoggedCardTableEntryClosure clear; 483 dcqs.set_closure(&clear); 484 dcqs.apply_closure_to_all_completed_buffers(); 485 dcqs.iterate_closure_all_threads(false); 486 clear.print_histo(); 487 488 // Now ensure that there's no dirty cards. 489 CountNonCleanMemRegionClosure count2(this); 490 ct_bs->mod_card_iterate(&count2); 491 if (count2.n() != 0) { 492 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 493 count2.n(), orig_count); 494 } 495 guarantee(count2.n() == 0, "Card table should be clean."); 496 497 RedirtyLoggedCardTableEntryClosure redirty; 498 JavaThread::dirty_card_queue_set().set_closure(&redirty); 499 dcqs.apply_closure_to_all_completed_buffers(); 500 dcqs.iterate_closure_all_threads(false); 501 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 502 clear.calls(), orig_count); 503 guarantee(redirty.calls() == clear.calls(), 504 "Or else mechanism is broken."); 505 506 CountNonCleanMemRegionClosure count3(this); 507 ct_bs->mod_card_iterate(&count3); 508 if (count3.n() != orig_count) { 509 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 510 orig_count, count3.n()); 511 guarantee(count3.n() >= orig_count, "Should have restored them all."); 512 } 513 514 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 515 } 516 517 // Private class members. 518 519 G1CollectedHeap* G1CollectedHeap::_g1h; 520 521 // Private methods. 522 523 HeapRegion* 524 G1CollectedHeap::new_region_try_secondary_free_list() { 525 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 526 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 527 if (!_secondary_free_list.is_empty()) { 528 if (G1ConcRegionFreeingVerbose) { 529 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 530 "secondary_free_list has %u entries", 531 _secondary_free_list.length()); 532 } 533 // It looks as if there are free regions available on the 534 // secondary_free_list. Let's move them to the free_list and try 535 // again to allocate from it. 536 append_secondary_free_list(); 537 538 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 539 "empty we should have moved at least one entry to the free_list"); 540 HeapRegion* res = _free_list.remove_head(); 541 if (G1ConcRegionFreeingVerbose) { 542 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 543 "allocated "HR_FORMAT" from secondary_free_list", 544 HR_FORMAT_PARAMS(res)); 545 } 546 return res; 547 } 548 549 // Wait here until we get notifed either when (a) there are no 550 // more free regions coming or (b) some regions have been moved on 551 // the secondary_free_list. 552 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 553 } 554 555 if (G1ConcRegionFreeingVerbose) { 556 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 557 "could not allocate from secondary_free_list"); 558 } 559 return NULL; 560 } 561 562 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 563 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 564 "the only time we use this to allocate a humongous region is " 565 "when we are allocating a single humongous region"); 566 567 HeapRegion* res; 568 if (G1StressConcRegionFreeing) { 569 if (!_secondary_free_list.is_empty()) { 570 if (G1ConcRegionFreeingVerbose) { 571 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 572 "forced to look at the secondary_free_list"); 573 } 574 res = new_region_try_secondary_free_list(); 575 if (res != NULL) { 576 return res; 577 } 578 } 579 } 580 res = _free_list.remove_head_or_null(); 581 if (res == NULL) { 582 if (G1ConcRegionFreeingVerbose) { 583 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 584 "res == NULL, trying the secondary_free_list"); 585 } 586 res = new_region_try_secondary_free_list(); 587 } 588 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 589 // Currently, only attempts to allocate GC alloc regions set 590 // do_expand to true. So, we should only reach here during a 591 // safepoint. If this assumption changes we might have to 592 // reconsider the use of _expand_heap_after_alloc_failure. 593 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 594 595 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 // Given that expand() succeeded in expanding the heap, and we 602 // always expand the heap by an amount aligned to the heap 603 // region size, the free list should in theory not be empty. So 604 // it would probably be OK to use remove_head(). But the extra 605 // check for NULL is unlikely to be a performance issue here (we 606 // just expanded the heap!) so let's just be conservative and 607 // use remove_head_or_null(). 608 res = _free_list.remove_head_or_null(); 609 } else { 610 _expand_heap_after_alloc_failure = false; 611 } 612 } 613 return res; 614 } 615 616 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 617 size_t word_size) { 618 assert(isHumongous(word_size), "word_size should be humongous"); 619 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 620 621 uint first = G1_NULL_HRS_INDEX; 622 if (num_regions == 1) { 623 // Only one region to allocate, no need to go through the slower 624 // path. The caller will attempt the expasion if this fails, so 625 // let's not try to expand here too. 626 HeapRegion* hr = new_region(word_size, false /* do_expand */); 627 if (hr != NULL) { 628 first = hr->hrs_index(); 629 } else { 630 first = G1_NULL_HRS_INDEX; 631 } 632 } else { 633 // We can't allocate humongous regions while cleanupComplete() is 634 // running, since some of the regions we find to be empty might not 635 // yet be added to the free list and it is not straightforward to 636 // know which list they are on so that we can remove them. Note 637 // that we only need to do this if we need to allocate more than 638 // one region to satisfy the current humongous allocation 639 // request. If we are only allocating one region we use the common 640 // region allocation code (see above). 641 wait_while_free_regions_coming(); 642 append_secondary_free_list_if_not_empty_with_lock(); 643 644 if (free_regions() >= num_regions) { 645 first = _hrs.find_contiguous(num_regions); 646 if (first != G1_NULL_HRS_INDEX) { 647 for (uint i = first; i < first + num_regions; ++i) { 648 HeapRegion* hr = region_at(i); 649 assert(hr->is_empty(), "sanity"); 650 assert(is_on_master_free_list(hr), "sanity"); 651 hr->set_pending_removal(true); 652 } 653 _free_list.remove_all_pending(num_regions); 654 } 655 } 656 } 657 return first; 658 } 659 660 HeapWord* 661 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 662 uint num_regions, 663 size_t word_size) { 664 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 665 assert(isHumongous(word_size), "word_size should be humongous"); 666 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 667 668 // Index of last region in the series + 1. 669 uint last = first + num_regions; 670 671 // We need to initialize the region(s) we just discovered. This is 672 // a bit tricky given that it can happen concurrently with 673 // refinement threads refining cards on these regions and 674 // potentially wanting to refine the BOT as they are scanning 675 // those cards (this can happen shortly after a cleanup; see CR 676 // 6991377). So we have to set up the region(s) carefully and in 677 // a specific order. 678 679 // The word size sum of all the regions we will allocate. 680 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 681 assert(word_size <= word_size_sum, "sanity"); 682 683 // This will be the "starts humongous" region. 684 HeapRegion* first_hr = region_at(first); 685 // The header of the new object will be placed at the bottom of 686 // the first region. 687 HeapWord* new_obj = first_hr->bottom(); 688 // This will be the new end of the first region in the series that 689 // should also match the end of the last region in the seriers. 690 HeapWord* new_end = new_obj + word_size_sum; 691 // This will be the new top of the first region that will reflect 692 // this allocation. 693 HeapWord* new_top = new_obj + word_size; 694 695 // First, we need to zero the header of the space that we will be 696 // allocating. When we update top further down, some refinement 697 // threads might try to scan the region. By zeroing the header we 698 // ensure that any thread that will try to scan the region will 699 // come across the zero klass word and bail out. 700 // 701 // NOTE: It would not have been correct to have used 702 // CollectedHeap::fill_with_object() and make the space look like 703 // an int array. The thread that is doing the allocation will 704 // later update the object header to a potentially different array 705 // type and, for a very short period of time, the klass and length 706 // fields will be inconsistent. This could cause a refinement 707 // thread to calculate the object size incorrectly. 708 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 709 710 // We will set up the first region as "starts humongous". This 711 // will also update the BOT covering all the regions to reflect 712 // that there is a single object that starts at the bottom of the 713 // first region. 714 first_hr->set_startsHumongous(new_top, new_end); 715 716 // Then, if there are any, we will set up the "continues 717 // humongous" regions. 718 HeapRegion* hr = NULL; 719 for (uint i = first + 1; i < last; ++i) { 720 hr = region_at(i); 721 hr->set_continuesHumongous(first_hr); 722 } 723 // If we have "continues humongous" regions (hr != NULL), then the 724 // end of the last one should match new_end. 725 assert(hr == NULL || hr->end() == new_end, "sanity"); 726 727 // Up to this point no concurrent thread would have been able to 728 // do any scanning on any region in this series. All the top 729 // fields still point to bottom, so the intersection between 730 // [bottom,top] and [card_start,card_end] will be empty. Before we 731 // update the top fields, we'll do a storestore to make sure that 732 // no thread sees the update to top before the zeroing of the 733 // object header and the BOT initialization. 734 OrderAccess::storestore(); 735 736 // Now that the BOT and the object header have been initialized, 737 // we can update top of the "starts humongous" region. 738 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 739 "new_top should be in this region"); 740 first_hr->set_top(new_top); 741 if (_hr_printer.is_active()) { 742 HeapWord* bottom = first_hr->bottom(); 743 HeapWord* end = first_hr->orig_end(); 744 if ((first + 1) == last) { 745 // the series has a single humongous region 746 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 747 } else { 748 // the series has more than one humongous regions 749 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 750 } 751 } 752 753 // Now, we will update the top fields of the "continues humongous" 754 // regions. The reason we need to do this is that, otherwise, 755 // these regions would look empty and this will confuse parts of 756 // G1. For example, the code that looks for a consecutive number 757 // of empty regions will consider them empty and try to 758 // re-allocate them. We can extend is_empty() to also include 759 // !continuesHumongous(), but it is easier to just update the top 760 // fields here. The way we set top for all regions (i.e., top == 761 // end for all regions but the last one, top == new_top for the 762 // last one) is actually used when we will free up the humongous 763 // region in free_humongous_region(). 764 hr = NULL; 765 for (uint i = first + 1; i < last; ++i) { 766 hr = region_at(i); 767 if ((i + 1) == last) { 768 // last continues humongous region 769 assert(hr->bottom() < new_top && new_top <= hr->end(), 770 "new_top should fall on this region"); 771 hr->set_top(new_top); 772 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 773 } else { 774 // not last one 775 assert(new_top > hr->end(), "new_top should be above this region"); 776 hr->set_top(hr->end()); 777 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 778 } 779 } 780 // If we have continues humongous regions (hr != NULL), then the 781 // end of the last one should match new_end and its top should 782 // match new_top. 783 assert(hr == NULL || 784 (hr->end() == new_end && hr->top() == new_top), "sanity"); 785 786 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 787 _summary_bytes_used += first_hr->used(); 788 _humongous_set.add(first_hr); 789 790 return new_obj; 791 } 792 793 // If could fit into free regions w/o expansion, try. 794 // Otherwise, if can expand, do so. 795 // Otherwise, if using ex regions might help, try with ex given back. 796 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 797 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 798 799 verify_region_sets_optional(); 800 801 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 802 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 803 uint x_num = expansion_regions(); 804 uint fs = _hrs.free_suffix(); 805 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 806 if (first == G1_NULL_HRS_INDEX) { 807 // The only thing we can do now is attempt expansion. 808 if (fs + x_num >= num_regions) { 809 // If the number of regions we're trying to allocate for this 810 // object is at most the number of regions in the free suffix, 811 // then the call to humongous_obj_allocate_find_first() above 812 // should have succeeded and we wouldn't be here. 813 // 814 // We should only be trying to expand when the free suffix is 815 // not sufficient for the object _and_ we have some expansion 816 // room available. 817 assert(num_regions > fs, "earlier allocation should have succeeded"); 818 819 ergo_verbose1(ErgoHeapSizing, 820 "attempt heap expansion", 821 ergo_format_reason("humongous allocation request failed") 822 ergo_format_byte("allocation request"), 823 word_size * HeapWordSize); 824 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 825 // Even though the heap was expanded, it might not have 826 // reached the desired size. So, we cannot assume that the 827 // allocation will succeed. 828 first = humongous_obj_allocate_find_first(num_regions, word_size); 829 } 830 } 831 } 832 833 HeapWord* result = NULL; 834 if (first != G1_NULL_HRS_INDEX) { 835 result = 836 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 837 assert(result != NULL, "it should always return a valid result"); 838 839 // A successful humongous object allocation changes the used space 840 // information of the old generation so we need to recalculate the 841 // sizes and update the jstat counters here. 842 g1mm()->update_sizes(); 843 } 844 845 verify_region_sets_optional(); 846 847 return result; 848 } 849 850 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 851 assert_heap_not_locked_and_not_at_safepoint(); 852 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 853 854 unsigned int dummy_gc_count_before; 855 return attempt_allocation(word_size, &dummy_gc_count_before); 856 } 857 858 HeapWord* 859 G1CollectedHeap::mem_allocate(size_t word_size, 860 bool* gc_overhead_limit_was_exceeded) { 861 assert_heap_not_locked_and_not_at_safepoint(); 862 863 // Loop until the allocation is satisified, or unsatisfied after GC. 864 for (int try_count = 1; /* we'll return */; try_count += 1) { 865 unsigned int gc_count_before; 866 867 HeapWord* result = NULL; 868 if (!isHumongous(word_size)) { 869 result = attempt_allocation(word_size, &gc_count_before); 870 } else { 871 result = attempt_allocation_humongous(word_size, &gc_count_before); 872 } 873 if (result != NULL) { 874 return result; 875 } 876 877 // Create the garbage collection operation... 878 VM_G1CollectForAllocation op(gc_count_before, word_size); 879 // ...and get the VM thread to execute it. 880 VMThread::execute(&op); 881 882 if (op.prologue_succeeded() && op.pause_succeeded()) { 883 // If the operation was successful we'll return the result even 884 // if it is NULL. If the allocation attempt failed immediately 885 // after a Full GC, it's unlikely we'll be able to allocate now. 886 HeapWord* result = op.result(); 887 if (result != NULL && !isHumongous(word_size)) { 888 // Allocations that take place on VM operations do not do any 889 // card dirtying and we have to do it here. We only have to do 890 // this for non-humongous allocations, though. 891 dirty_young_block(result, word_size); 892 } 893 return result; 894 } else { 895 assert(op.result() == NULL, 896 "the result should be NULL if the VM op did not succeed"); 897 } 898 899 // Give a warning if we seem to be looping forever. 900 if ((QueuedAllocationWarningCount > 0) && 901 (try_count % QueuedAllocationWarningCount == 0)) { 902 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 903 } 904 } 905 906 ShouldNotReachHere(); 907 return NULL; 908 } 909 910 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 911 unsigned int *gc_count_before_ret) { 912 // Make sure you read the note in attempt_allocation_humongous(). 913 914 assert_heap_not_locked_and_not_at_safepoint(); 915 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 916 "be called for humongous allocation requests"); 917 918 // We should only get here after the first-level allocation attempt 919 // (attempt_allocation()) failed to allocate. 920 921 // We will loop until a) we manage to successfully perform the 922 // allocation or b) we successfully schedule a collection which 923 // fails to perform the allocation. b) is the only case when we'll 924 // return NULL. 925 HeapWord* result = NULL; 926 for (int try_count = 1; /* we'll return */; try_count += 1) { 927 bool should_try_gc; 928 unsigned int gc_count_before; 929 930 { 931 MutexLockerEx x(Heap_lock); 932 933 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 934 false /* bot_updates */); 935 if (result != NULL) { 936 return result; 937 } 938 939 // If we reach here, attempt_allocation_locked() above failed to 940 // allocate a new region. So the mutator alloc region should be NULL. 941 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 942 943 if (GC_locker::is_active_and_needs_gc()) { 944 if (g1_policy()->can_expand_young_list()) { 945 // No need for an ergo verbose message here, 946 // can_expand_young_list() does this when it returns true. 947 result = _mutator_alloc_region.attempt_allocation_force(word_size, 948 false /* bot_updates */); 949 if (result != NULL) { 950 return result; 951 } 952 } 953 should_try_gc = false; 954 } else { 955 // Read the GC count while still holding the Heap_lock. 956 gc_count_before = total_collections(); 957 should_try_gc = true; 958 } 959 } 960 961 if (should_try_gc) { 962 bool succeeded; 963 result = do_collection_pause(word_size, gc_count_before, &succeeded); 964 if (result != NULL) { 965 assert(succeeded, "only way to get back a non-NULL result"); 966 return result; 967 } 968 969 if (succeeded) { 970 // If we get here we successfully scheduled a collection which 971 // failed to allocate. No point in trying to allocate 972 // further. We'll just return NULL. 973 MutexLockerEx x(Heap_lock); 974 *gc_count_before_ret = total_collections(); 975 return NULL; 976 } 977 } else { 978 GC_locker::stall_until_clear(); 979 } 980 981 // We can reach here if we were unsuccessul in scheduling a 982 // collection (because another thread beat us to it) or if we were 983 // stalled due to the GC locker. In either can we should retry the 984 // allocation attempt in case another thread successfully 985 // performed a collection and reclaimed enough space. We do the 986 // first attempt (without holding the Heap_lock) here and the 987 // follow-on attempt will be at the start of the next loop 988 // iteration (after taking the Heap_lock). 989 result = _mutator_alloc_region.attempt_allocation(word_size, 990 false /* bot_updates */); 991 if (result != NULL) { 992 return result; 993 } 994 995 // Give a warning if we seem to be looping forever. 996 if ((QueuedAllocationWarningCount > 0) && 997 (try_count % QueuedAllocationWarningCount == 0)) { 998 warning("G1CollectedHeap::attempt_allocation_slow() " 999 "retries %d times", try_count); 1000 } 1001 } 1002 1003 ShouldNotReachHere(); 1004 return NULL; 1005 } 1006 1007 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1008 unsigned int * gc_count_before_ret) { 1009 // The structure of this method has a lot of similarities to 1010 // attempt_allocation_slow(). The reason these two were not merged 1011 // into a single one is that such a method would require several "if 1012 // allocation is not humongous do this, otherwise do that" 1013 // conditional paths which would obscure its flow. In fact, an early 1014 // version of this code did use a unified method which was harder to 1015 // follow and, as a result, it had subtle bugs that were hard to 1016 // track down. So keeping these two methods separate allows each to 1017 // be more readable. It will be good to keep these two in sync as 1018 // much as possible. 1019 1020 assert_heap_not_locked_and_not_at_safepoint(); 1021 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1022 "should only be called for humongous allocations"); 1023 1024 // Humongous objects can exhaust the heap quickly, so we should check if we 1025 // need to start a marking cycle at each humongous object allocation. We do 1026 // the check before we do the actual allocation. The reason for doing it 1027 // before the allocation is that we avoid having to keep track of the newly 1028 // allocated memory while we do a GC. 1029 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1030 word_size)) { 1031 collect(GCCause::_g1_humongous_allocation); 1032 } 1033 1034 // We will loop until a) we manage to successfully perform the 1035 // allocation or b) we successfully schedule a collection which 1036 // fails to perform the allocation. b) is the only case when we'll 1037 // return NULL. 1038 HeapWord* result = NULL; 1039 for (int try_count = 1; /* we'll return */; try_count += 1) { 1040 bool should_try_gc; 1041 unsigned int gc_count_before; 1042 1043 { 1044 MutexLockerEx x(Heap_lock); 1045 1046 // Given that humongous objects are not allocated in young 1047 // regions, we'll first try to do the allocation without doing a 1048 // collection hoping that there's enough space in the heap. 1049 result = humongous_obj_allocate(word_size); 1050 if (result != NULL) { 1051 return result; 1052 } 1053 1054 if (GC_locker::is_active_and_needs_gc()) { 1055 should_try_gc = false; 1056 } else { 1057 // Read the GC count while still holding the Heap_lock. 1058 gc_count_before = total_collections(); 1059 should_try_gc = true; 1060 } 1061 } 1062 1063 if (should_try_gc) { 1064 // If we failed to allocate the humongous object, we should try to 1065 // do a collection pause (if we're allowed) in case it reclaims 1066 // enough space for the allocation to succeed after the pause. 1067 1068 bool succeeded; 1069 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1070 if (result != NULL) { 1071 assert(succeeded, "only way to get back a non-NULL result"); 1072 return result; 1073 } 1074 1075 if (succeeded) { 1076 // If we get here we successfully scheduled a collection which 1077 // failed to allocate. No point in trying to allocate 1078 // further. We'll just return NULL. 1079 MutexLockerEx x(Heap_lock); 1080 *gc_count_before_ret = total_collections(); 1081 return NULL; 1082 } 1083 } else { 1084 GC_locker::stall_until_clear(); 1085 } 1086 1087 // We can reach here if we were unsuccessul in scheduling a 1088 // collection (because another thread beat us to it) or if we were 1089 // stalled due to the GC locker. In either can we should retry the 1090 // allocation attempt in case another thread successfully 1091 // performed a collection and reclaimed enough space. Give a 1092 // warning if we seem to be looping forever. 1093 1094 if ((QueuedAllocationWarningCount > 0) && 1095 (try_count % QueuedAllocationWarningCount == 0)) { 1096 warning("G1CollectedHeap::attempt_allocation_humongous() " 1097 "retries %d times", try_count); 1098 } 1099 } 1100 1101 ShouldNotReachHere(); 1102 return NULL; 1103 } 1104 1105 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1106 bool expect_null_mutator_alloc_region) { 1107 assert_at_safepoint(true /* should_be_vm_thread */); 1108 assert(_mutator_alloc_region.get() == NULL || 1109 !expect_null_mutator_alloc_region, 1110 "the current alloc region was unexpectedly found to be non-NULL"); 1111 1112 if (!isHumongous(word_size)) { 1113 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1114 false /* bot_updates */); 1115 } else { 1116 HeapWord* result = humongous_obj_allocate(word_size); 1117 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1118 g1_policy()->set_initiate_conc_mark_if_possible(); 1119 } 1120 return result; 1121 } 1122 1123 ShouldNotReachHere(); 1124 } 1125 1126 class PostMCRemSetClearClosure: public HeapRegionClosure { 1127 ModRefBarrierSet* _mr_bs; 1128 public: 1129 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1130 bool doHeapRegion(HeapRegion* r) { 1131 r->reset_gc_time_stamp(); 1132 if (r->continuesHumongous()) 1133 return false; 1134 HeapRegionRemSet* hrrs = r->rem_set(); 1135 if (hrrs != NULL) hrrs->clear(); 1136 // You might think here that we could clear just the cards 1137 // corresponding to the used region. But no: if we leave a dirty card 1138 // in a region we might allocate into, then it would prevent that card 1139 // from being enqueued, and cause it to be missed. 1140 // Re: the performance cost: we shouldn't be doing full GC anyway! 1141 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1142 return false; 1143 } 1144 }; 1145 1146 1147 class PostMCRemSetInvalidateClosure: public HeapRegionClosure { 1148 ModRefBarrierSet* _mr_bs; 1149 public: 1150 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1151 bool doHeapRegion(HeapRegion* r) { 1152 if (r->continuesHumongous()) return false; 1153 if (r->used_region().word_size() != 0) { 1154 _mr_bs->invalidate(r->used_region(), true /*whole heap*/); 1155 } 1156 return false; 1157 } 1158 }; 1159 1160 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1161 G1CollectedHeap* _g1h; 1162 UpdateRSOopClosure _cl; 1163 int _worker_i; 1164 public: 1165 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1166 _cl(g1->g1_rem_set(), worker_i), 1167 _worker_i(worker_i), 1168 _g1h(g1) 1169 { } 1170 1171 bool doHeapRegion(HeapRegion* r) { 1172 if (!r->continuesHumongous()) { 1173 _cl.set_from(r); 1174 r->oop_iterate(&_cl); 1175 } 1176 return false; 1177 } 1178 }; 1179 1180 class ParRebuildRSTask: public AbstractGangTask { 1181 G1CollectedHeap* _g1; 1182 public: 1183 ParRebuildRSTask(G1CollectedHeap* g1) 1184 : AbstractGangTask("ParRebuildRSTask"), 1185 _g1(g1) 1186 { } 1187 1188 void work(uint worker_id) { 1189 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1190 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1191 _g1->workers()->active_workers(), 1192 HeapRegion::RebuildRSClaimValue); 1193 } 1194 }; 1195 1196 class PostCompactionPrinterClosure: public HeapRegionClosure { 1197 private: 1198 G1HRPrinter* _hr_printer; 1199 public: 1200 bool doHeapRegion(HeapRegion* hr) { 1201 assert(!hr->is_young(), "not expecting to find young regions"); 1202 // We only generate output for non-empty regions. 1203 if (!hr->is_empty()) { 1204 if (!hr->isHumongous()) { 1205 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1206 } else if (hr->startsHumongous()) { 1207 if (hr->capacity() == HeapRegion::GrainBytes) { 1208 // single humongous region 1209 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1210 } else { 1211 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1212 } 1213 } else { 1214 assert(hr->continuesHumongous(), "only way to get here"); 1215 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1216 } 1217 } 1218 return false; 1219 } 1220 1221 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1222 : _hr_printer(hr_printer) { } 1223 }; 1224 1225 bool G1CollectedHeap::do_collection(bool explicit_gc, 1226 bool clear_all_soft_refs, 1227 size_t word_size) { 1228 assert_at_safepoint(true /* should_be_vm_thread */); 1229 1230 if (GC_locker::check_active_before_gc()) { 1231 return false; 1232 } 1233 1234 SvcGCMarker sgcm(SvcGCMarker::FULL); 1235 ResourceMark rm; 1236 1237 print_heap_before_gc(); 1238 1239 HRSPhaseSetter x(HRSPhaseFullGC); 1240 verify_region_sets_optional(); 1241 1242 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1243 collector_policy()->should_clear_all_soft_refs(); 1244 1245 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1246 1247 { 1248 IsGCActiveMark x; 1249 1250 // Timing 1251 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1252 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1253 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1254 1255 char verbose_str[128]; 1256 sprintf(verbose_str, "Full GC (%s)", GCCause::to_string(gc_cause())); 1257 TraceTime t(verbose_str, G1Log::fine(), true, gclog_or_tty); 1258 1259 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1260 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1261 1262 double start = os::elapsedTime(); 1263 g1_policy()->record_full_collection_start(); 1264 1265 // Note: When we have a more flexible GC logging framework that 1266 // allows us to add optional attributes to a GC log record we 1267 // could consider timing and reporting how long we wait in the 1268 // following two methods. 1269 wait_while_free_regions_coming(); 1270 // If we start the compaction before the CM threads finish 1271 // scanning the root regions we might trip them over as we'll 1272 // be moving objects / updating references. So let's wait until 1273 // they are done. By telling them to abort, they should complete 1274 // early. 1275 _cm->root_regions()->abort(); 1276 _cm->root_regions()->wait_until_scan_finished(); 1277 append_secondary_free_list_if_not_empty_with_lock(); 1278 1279 gc_prologue(true); 1280 increment_total_collections(true /* full gc */); 1281 1282 size_t g1h_prev_used = used(); 1283 assert(used() == recalculate_used(), "Should be equal"); 1284 1285 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 1286 HandleMark hm; // Discard invalid handles created during verification 1287 gclog_or_tty->print(" VerifyBeforeGC:"); 1288 prepare_for_verify(); 1289 Universe::verify(/* silent */ false, 1290 /* option */ VerifyOption_G1UsePrevMarking); 1291 1292 } 1293 pre_full_gc_dump(); 1294 1295 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1296 1297 // Disable discovery and empty the discovered lists 1298 // for the CM ref processor. 1299 ref_processor_cm()->disable_discovery(); 1300 ref_processor_cm()->abandon_partial_discovery(); 1301 ref_processor_cm()->verify_no_references_recorded(); 1302 1303 // Abandon current iterations of concurrent marking and concurrent 1304 // refinement, if any are in progress. We have to do this before 1305 // wait_until_scan_finished() below. 1306 concurrent_mark()->abort(); 1307 1308 // Make sure we'll choose a new allocation region afterwards. 1309 release_mutator_alloc_region(); 1310 abandon_gc_alloc_regions(); 1311 g1_rem_set()->cleanupHRRS(); 1312 1313 // We should call this after we retire any currently active alloc 1314 // regions so that all the ALLOC / RETIRE events are generated 1315 // before the start GC event. 1316 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1317 1318 // We may have added regions to the current incremental collection 1319 // set between the last GC or pause and now. We need to clear the 1320 // incremental collection set and then start rebuilding it afresh 1321 // after this full GC. 1322 abandon_collection_set(g1_policy()->inc_cset_head()); 1323 g1_policy()->clear_incremental_cset(); 1324 g1_policy()->stop_incremental_cset_building(); 1325 1326 tear_down_region_sets(false /* free_list_only */); 1327 g1_policy()->set_gcs_are_young(true); 1328 1329 // See the comments in g1CollectedHeap.hpp and 1330 // G1CollectedHeap::ref_processing_init() about 1331 // how reference processing currently works in G1. 1332 1333 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1334 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1335 1336 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1337 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1338 1339 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1340 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1341 1342 // Do collection work 1343 { 1344 HandleMark hm; // Discard invalid handles created during gc 1345 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1346 } 1347 1348 assert(free_regions() == 0, "we should not have added any free regions"); 1349 rebuild_region_sets(false /* free_list_only */); 1350 1351 // Enqueue any discovered reference objects that have 1352 // not been removed from the discovered lists. 1353 ref_processor_stw()->enqueue_discovered_references(); 1354 1355 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1356 1357 MemoryService::track_memory_usage(); 1358 1359 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 1360 HandleMark hm; // Discard invalid handles created during verification 1361 gclog_or_tty->print(" VerifyAfterGC:"); 1362 prepare_for_verify(); 1363 Universe::verify(/* silent */ false, 1364 /* option */ VerifyOption_G1UsePrevMarking); 1365 1366 } 1367 1368 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1369 ref_processor_stw()->verify_no_references_recorded(); 1370 1371 // Note: since we've just done a full GC, concurrent 1372 // marking is no longer active. Therefore we need not 1373 // re-enable reference discovery for the CM ref processor. 1374 // That will be done at the start of the next marking cycle. 1375 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1376 ref_processor_cm()->verify_no_references_recorded(); 1377 1378 reset_gc_time_stamp(); 1379 // Since everything potentially moved, we will clear all remembered 1380 // sets, and clear all cards. Later we will rebuild remebered 1381 // sets. We will also reset the GC time stamps of the regions. 1382 PostMCRemSetClearClosure rs_clear(mr_bs()); 1383 heap_region_iterate(&rs_clear); 1384 1385 // Resize the heap if necessary. 1386 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1387 1388 if (_hr_printer.is_active()) { 1389 // We should do this after we potentially resize the heap so 1390 // that all the COMMIT / UNCOMMIT events are generated before 1391 // the end GC event. 1392 1393 PostCompactionPrinterClosure cl(hr_printer()); 1394 heap_region_iterate(&cl); 1395 1396 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1397 } 1398 1399 if (_cg1r->use_cache()) { 1400 _cg1r->clear_and_record_card_counts(); 1401 _cg1r->clear_hot_cache(); 1402 } 1403 1404 // Rebuild remembered sets of all regions. 1405 if (G1CollectedHeap::use_parallel_gc_threads()) { 1406 uint n_workers = 1407 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1408 workers()->active_workers(), 1409 Threads::number_of_non_daemon_threads()); 1410 assert(UseDynamicNumberOfGCThreads || 1411 n_workers == workers()->total_workers(), 1412 "If not dynamic should be using all the workers"); 1413 workers()->set_active_workers(n_workers); 1414 // Set parallel threads in the heap (_n_par_threads) only 1415 // before a parallel phase and always reset it to 0 after 1416 // the phase so that the number of parallel threads does 1417 // no get carried forward to a serial phase where there 1418 // may be code that is "possibly_parallel". 1419 set_par_threads(n_workers); 1420 1421 ParRebuildRSTask rebuild_rs_task(this); 1422 assert(check_heap_region_claim_values( 1423 HeapRegion::InitialClaimValue), "sanity check"); 1424 assert(UseDynamicNumberOfGCThreads || 1425 workers()->active_workers() == workers()->total_workers(), 1426 "Unless dynamic should use total workers"); 1427 // Use the most recent number of active workers 1428 assert(workers()->active_workers() > 0, 1429 "Active workers not properly set"); 1430 set_par_threads(workers()->active_workers()); 1431 workers()->run_task(&rebuild_rs_task); 1432 set_par_threads(0); 1433 assert(check_heap_region_claim_values( 1434 HeapRegion::RebuildRSClaimValue), "sanity check"); 1435 reset_heap_region_claim_values(); 1436 } else { 1437 RebuildRSOutOfRegionClosure rebuild_rs(this); 1438 heap_region_iterate(&rebuild_rs); 1439 } 1440 1441 if (G1Log::fine()) { 1442 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1443 } 1444 1445 if (true) { // FIXME 1446 // Ask the permanent generation to adjust size for full collections 1447 perm()->compute_new_size(); 1448 } 1449 1450 // Start a new incremental collection set for the next pause 1451 assert(g1_policy()->collection_set() == NULL, "must be"); 1452 g1_policy()->start_incremental_cset_building(); 1453 1454 // Clear the _cset_fast_test bitmap in anticipation of adding 1455 // regions to the incremental collection set for the next 1456 // evacuation pause. 1457 clear_cset_fast_test(); 1458 1459 init_mutator_alloc_region(); 1460 1461 double end = os::elapsedTime(); 1462 g1_policy()->record_full_collection_end(); 1463 1464 #ifdef TRACESPINNING 1465 ParallelTaskTerminator::print_termination_counts(); 1466 #endif 1467 1468 gc_epilogue(true); 1469 1470 // Discard all rset updates 1471 JavaThread::dirty_card_queue_set().abandon_logs(); 1472 assert(!G1DeferredRSUpdate 1473 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1474 } 1475 1476 _young_list->reset_sampled_info(); 1477 // At this point there should be no regions in the 1478 // entire heap tagged as young. 1479 assert( check_young_list_empty(true /* check_heap */), 1480 "young list should be empty at this point"); 1481 1482 // Update the number of full collections that have been completed. 1483 increment_full_collections_completed(false /* concurrent */); 1484 1485 _hrs.verify_optional(); 1486 verify_region_sets_optional(); 1487 1488 print_heap_after_gc(); 1489 g1mm()->update_sizes(); 1490 post_full_gc_dump(); 1491 1492 return true; 1493 } 1494 1495 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1496 // do_collection() will return whether it succeeded in performing 1497 // the GC. Currently, there is no facility on the 1498 // do_full_collection() API to notify the caller than the collection 1499 // did not succeed (e.g., because it was locked out by the GC 1500 // locker). So, right now, we'll ignore the return value. 1501 bool dummy = do_collection(true, /* explicit_gc */ 1502 clear_all_soft_refs, 1503 0 /* word_size */); 1504 } 1505 1506 // This code is mostly copied from TenuredGeneration. 1507 void 1508 G1CollectedHeap:: 1509 resize_if_necessary_after_full_collection(size_t word_size) { 1510 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1511 1512 // Include the current allocation, if any, and bytes that will be 1513 // pre-allocated to support collections, as "used". 1514 const size_t used_after_gc = used(); 1515 const size_t capacity_after_gc = capacity(); 1516 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1517 1518 // This is enforced in arguments.cpp. 1519 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1520 "otherwise the code below doesn't make sense"); 1521 1522 // We don't have floating point command-line arguments 1523 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1524 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1525 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1526 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1527 1528 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1529 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1530 1531 // We have to be careful here as these two calculations can overflow 1532 // 32-bit size_t's. 1533 double used_after_gc_d = (double) used_after_gc; 1534 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1535 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1536 1537 // Let's make sure that they are both under the max heap size, which 1538 // by default will make them fit into a size_t. 1539 double desired_capacity_upper_bound = (double) max_heap_size; 1540 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1541 desired_capacity_upper_bound); 1542 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1543 desired_capacity_upper_bound); 1544 1545 // We can now safely turn them into size_t's. 1546 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1547 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1548 1549 // This assert only makes sense here, before we adjust them 1550 // with respect to the min and max heap size. 1551 assert(minimum_desired_capacity <= maximum_desired_capacity, 1552 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1553 "maximum_desired_capacity = "SIZE_FORMAT, 1554 minimum_desired_capacity, maximum_desired_capacity)); 1555 1556 // Should not be greater than the heap max size. No need to adjust 1557 // it with respect to the heap min size as it's a lower bound (i.e., 1558 // we'll try to make the capacity larger than it, not smaller). 1559 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1560 // Should not be less than the heap min size. No need to adjust it 1561 // with respect to the heap max size as it's an upper bound (i.e., 1562 // we'll try to make the capacity smaller than it, not greater). 1563 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1564 1565 if (capacity_after_gc < minimum_desired_capacity) { 1566 // Don't expand unless it's significant 1567 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1568 ergo_verbose4(ErgoHeapSizing, 1569 "attempt heap expansion", 1570 ergo_format_reason("capacity lower than " 1571 "min desired capacity after Full GC") 1572 ergo_format_byte("capacity") 1573 ergo_format_byte("occupancy") 1574 ergo_format_byte_perc("min desired capacity"), 1575 capacity_after_gc, used_after_gc, 1576 minimum_desired_capacity, (double) MinHeapFreeRatio); 1577 expand(expand_bytes); 1578 1579 // No expansion, now see if we want to shrink 1580 } else if (capacity_after_gc > maximum_desired_capacity) { 1581 // Capacity too large, compute shrinking size 1582 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1583 ergo_verbose4(ErgoHeapSizing, 1584 "attempt heap shrinking", 1585 ergo_format_reason("capacity higher than " 1586 "max desired capacity after Full GC") 1587 ergo_format_byte("capacity") 1588 ergo_format_byte("occupancy") 1589 ergo_format_byte_perc("max desired capacity"), 1590 capacity_after_gc, used_after_gc, 1591 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1592 shrink(shrink_bytes); 1593 } 1594 } 1595 1596 1597 HeapWord* 1598 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1599 bool* succeeded) { 1600 assert_at_safepoint(true /* should_be_vm_thread */); 1601 1602 *succeeded = true; 1603 // Let's attempt the allocation first. 1604 HeapWord* result = 1605 attempt_allocation_at_safepoint(word_size, 1606 false /* expect_null_mutator_alloc_region */); 1607 if (result != NULL) { 1608 assert(*succeeded, "sanity"); 1609 return result; 1610 } 1611 1612 // In a G1 heap, we're supposed to keep allocation from failing by 1613 // incremental pauses. Therefore, at least for now, we'll favor 1614 // expansion over collection. (This might change in the future if we can 1615 // do something smarter than full collection to satisfy a failed alloc.) 1616 result = expand_and_allocate(word_size); 1617 if (result != NULL) { 1618 assert(*succeeded, "sanity"); 1619 return result; 1620 } 1621 1622 // Expansion didn't work, we'll try to do a Full GC. 1623 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1624 false, /* clear_all_soft_refs */ 1625 word_size); 1626 if (!gc_succeeded) { 1627 *succeeded = false; 1628 return NULL; 1629 } 1630 1631 // Retry the allocation 1632 result = attempt_allocation_at_safepoint(word_size, 1633 true /* expect_null_mutator_alloc_region */); 1634 if (result != NULL) { 1635 assert(*succeeded, "sanity"); 1636 return result; 1637 } 1638 1639 // Then, try a Full GC that will collect all soft references. 1640 gc_succeeded = do_collection(false, /* explicit_gc */ 1641 true, /* clear_all_soft_refs */ 1642 word_size); 1643 if (!gc_succeeded) { 1644 *succeeded = false; 1645 return NULL; 1646 } 1647 1648 // Retry the allocation once more 1649 result = attempt_allocation_at_safepoint(word_size, 1650 true /* expect_null_mutator_alloc_region */); 1651 if (result != NULL) { 1652 assert(*succeeded, "sanity"); 1653 return result; 1654 } 1655 1656 assert(!collector_policy()->should_clear_all_soft_refs(), 1657 "Flag should have been handled and cleared prior to this point"); 1658 1659 // What else? We might try synchronous finalization later. If the total 1660 // space available is large enough for the allocation, then a more 1661 // complete compaction phase than we've tried so far might be 1662 // appropriate. 1663 assert(*succeeded, "sanity"); 1664 return NULL; 1665 } 1666 1667 // Attempting to expand the heap sufficiently 1668 // to support an allocation of the given "word_size". If 1669 // successful, perform the allocation and return the address of the 1670 // allocated block, or else "NULL". 1671 1672 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1673 assert_at_safepoint(true /* should_be_vm_thread */); 1674 1675 verify_region_sets_optional(); 1676 1677 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1678 ergo_verbose1(ErgoHeapSizing, 1679 "attempt heap expansion", 1680 ergo_format_reason("allocation request failed") 1681 ergo_format_byte("allocation request"), 1682 word_size * HeapWordSize); 1683 if (expand(expand_bytes)) { 1684 _hrs.verify_optional(); 1685 verify_region_sets_optional(); 1686 return attempt_allocation_at_safepoint(word_size, 1687 false /* expect_null_mutator_alloc_region */); 1688 } 1689 return NULL; 1690 } 1691 1692 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1693 HeapWord* new_end) { 1694 assert(old_end != new_end, "don't call this otherwise"); 1695 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1696 1697 // Update the committed mem region. 1698 _g1_committed.set_end(new_end); 1699 // Tell the card table about the update. 1700 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1701 // Tell the BOT about the update. 1702 _bot_shared->resize(_g1_committed.word_size()); 1703 } 1704 1705 bool G1CollectedHeap::expand(size_t expand_bytes) { 1706 size_t old_mem_size = _g1_storage.committed_size(); 1707 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1708 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1709 HeapRegion::GrainBytes); 1710 ergo_verbose2(ErgoHeapSizing, 1711 "expand the heap", 1712 ergo_format_byte("requested expansion amount") 1713 ergo_format_byte("attempted expansion amount"), 1714 expand_bytes, aligned_expand_bytes); 1715 1716 // First commit the memory. 1717 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1718 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1719 if (successful) { 1720 // Then propagate this update to the necessary data structures. 1721 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1722 update_committed_space(old_end, new_end); 1723 1724 FreeRegionList expansion_list("Local Expansion List"); 1725 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1726 assert(mr.start() == old_end, "post-condition"); 1727 // mr might be a smaller region than what was requested if 1728 // expand_by() was unable to allocate the HeapRegion instances 1729 assert(mr.end() <= new_end, "post-condition"); 1730 1731 size_t actual_expand_bytes = mr.byte_size(); 1732 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1733 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1734 "post-condition"); 1735 if (actual_expand_bytes < aligned_expand_bytes) { 1736 // We could not expand _hrs to the desired size. In this case we 1737 // need to shrink the committed space accordingly. 1738 assert(mr.end() < new_end, "invariant"); 1739 1740 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1741 // First uncommit the memory. 1742 _g1_storage.shrink_by(diff_bytes); 1743 // Then propagate this update to the necessary data structures. 1744 update_committed_space(new_end, mr.end()); 1745 } 1746 _free_list.add_as_tail(&expansion_list); 1747 1748 if (_hr_printer.is_active()) { 1749 HeapWord* curr = mr.start(); 1750 while (curr < mr.end()) { 1751 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1752 _hr_printer.commit(curr, curr_end); 1753 curr = curr_end; 1754 } 1755 assert(curr == mr.end(), "post-condition"); 1756 } 1757 g1_policy()->record_new_heap_size(n_regions()); 1758 } else { 1759 ergo_verbose0(ErgoHeapSizing, 1760 "did not expand the heap", 1761 ergo_format_reason("heap expansion operation failed")); 1762 // The expansion of the virtual storage space was unsuccessful. 1763 // Let's see if it was because we ran out of swap. 1764 if (G1ExitOnExpansionFailure && 1765 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1766 // We had head room... 1767 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1768 } 1769 } 1770 return successful; 1771 } 1772 1773 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1774 size_t old_mem_size = _g1_storage.committed_size(); 1775 size_t aligned_shrink_bytes = 1776 ReservedSpace::page_align_size_down(shrink_bytes); 1777 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1778 HeapRegion::GrainBytes); 1779 uint num_regions_deleted = 0; 1780 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted); 1781 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1782 assert(mr.end() == old_end, "post-condition"); 1783 1784 ergo_verbose3(ErgoHeapSizing, 1785 "shrink the heap", 1786 ergo_format_byte("requested shrinking amount") 1787 ergo_format_byte("aligned shrinking amount") 1788 ergo_format_byte("attempted shrinking amount"), 1789 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1790 if (mr.byte_size() > 0) { 1791 if (_hr_printer.is_active()) { 1792 HeapWord* curr = mr.end(); 1793 while (curr > mr.start()) { 1794 HeapWord* curr_end = curr; 1795 curr -= HeapRegion::GrainWords; 1796 _hr_printer.uncommit(curr, curr_end); 1797 } 1798 assert(curr == mr.start(), "post-condition"); 1799 } 1800 1801 _g1_storage.shrink_by(mr.byte_size()); 1802 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1803 assert(mr.start() == new_end, "post-condition"); 1804 1805 _expansion_regions += num_regions_deleted; 1806 update_committed_space(old_end, new_end); 1807 HeapRegionRemSet::shrink_heap(n_regions()); 1808 g1_policy()->record_new_heap_size(n_regions()); 1809 } else { 1810 ergo_verbose0(ErgoHeapSizing, 1811 "did not shrink the heap", 1812 ergo_format_reason("heap shrinking operation failed")); 1813 } 1814 } 1815 1816 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1817 verify_region_sets_optional(); 1818 1819 // We should only reach here at the end of a Full GC which means we 1820 // should not not be holding to any GC alloc regions. The method 1821 // below will make sure of that and do any remaining clean up. 1822 abandon_gc_alloc_regions(); 1823 1824 // Instead of tearing down / rebuilding the free lists here, we 1825 // could instead use the remove_all_pending() method on free_list to 1826 // remove only the ones that we need to remove. 1827 tear_down_region_sets(true /* free_list_only */); 1828 shrink_helper(shrink_bytes); 1829 rebuild_region_sets(true /* free_list_only */); 1830 1831 _hrs.verify_optional(); 1832 verify_region_sets_optional(); 1833 } 1834 1835 // Public methods. 1836 1837 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1838 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1839 #endif // _MSC_VER 1840 1841 1842 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1843 SharedHeap(policy_), 1844 _g1_policy(policy_), 1845 _dirty_card_queue_set(false), 1846 _into_cset_dirty_card_queue_set(false), 1847 _is_alive_closure_cm(this), 1848 _is_alive_closure_stw(this), 1849 _ref_processor_cm(NULL), 1850 _ref_processor_stw(NULL), 1851 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1852 _bot_shared(NULL), 1853 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL), 1854 _evac_failure_scan_stack(NULL) , 1855 _mark_in_progress(false), 1856 _cg1r(NULL), _summary_bytes_used(0), 1857 _g1mm(NULL), 1858 _refine_cte_cl(NULL), 1859 _full_collection(false), 1860 _free_list("Master Free List"), 1861 _secondary_free_list("Secondary Free List"), 1862 _old_set("Old Set"), 1863 _humongous_set("Master Humongous Set"), 1864 _free_regions_coming(false), 1865 _young_list(new YoungList(this)), 1866 _gc_time_stamp(0), 1867 _retained_old_gc_alloc_region(NULL), 1868 _expand_heap_after_alloc_failure(true), 1869 _surviving_young_words(NULL), 1870 _full_collections_completed(0), 1871 _in_cset_fast_test(NULL), 1872 _in_cset_fast_test_base(NULL), 1873 _dirty_cards_region_list(NULL), 1874 _worker_cset_start_region(NULL), 1875 _worker_cset_start_region_time_stamp(NULL) { 1876 _g1h = this; // To catch bugs. 1877 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1878 vm_exit_during_initialization("Failed necessary allocation."); 1879 } 1880 1881 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1882 1883 int n_queues = MAX2((int)ParallelGCThreads, 1); 1884 _task_queues = new RefToScanQueueSet(n_queues); 1885 1886 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1887 assert(n_rem_sets > 0, "Invariant."); 1888 1889 HeapRegionRemSetIterator** iter_arr = 1890 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues); 1891 for (int i = 0; i < n_queues; i++) { 1892 iter_arr[i] = new HeapRegionRemSetIterator(); 1893 } 1894 _rem_set_iterator = iter_arr; 1895 1896 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues); 1897 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues); 1898 1899 for (int i = 0; i < n_queues; i++) { 1900 RefToScanQueue* q = new RefToScanQueue(); 1901 q->initialize(); 1902 _task_queues->register_queue(i, q); 1903 } 1904 1905 clear_cset_start_regions(); 1906 1907 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1908 } 1909 1910 jint G1CollectedHeap::initialize() { 1911 CollectedHeap::pre_initialize(); 1912 os::enable_vtime(); 1913 1914 G1Log::init(); 1915 1916 // Necessary to satisfy locking discipline assertions. 1917 1918 MutexLocker x(Heap_lock); 1919 1920 // We have to initialize the printer before committing the heap, as 1921 // it will be used then. 1922 _hr_printer.set_active(G1PrintHeapRegions); 1923 1924 // While there are no constraints in the GC code that HeapWordSize 1925 // be any particular value, there are multiple other areas in the 1926 // system which believe this to be true (e.g. oop->object_size in some 1927 // cases incorrectly returns the size in wordSize units rather than 1928 // HeapWordSize). 1929 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1930 1931 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1932 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1933 1934 // Ensure that the sizes are properly aligned. 1935 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1936 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1937 1938 _cg1r = new ConcurrentG1Refine(); 1939 1940 // Reserve the maximum. 1941 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1942 // Includes the perm-gen. 1943 1944 // When compressed oops are enabled, the preferred heap base 1945 // is calculated by subtracting the requested size from the 1946 // 32Gb boundary and using the result as the base address for 1947 // heap reservation. If the requested size is not aligned to 1948 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1949 // into the ReservedHeapSpace constructor) then the actual 1950 // base of the reserved heap may end up differing from the 1951 // address that was requested (i.e. the preferred heap base). 1952 // If this happens then we could end up using a non-optimal 1953 // compressed oops mode. 1954 1955 // Since max_byte_size is aligned to the size of a heap region (checked 1956 // above), we also need to align the perm gen size as it might not be. 1957 const size_t total_reserved = max_byte_size + 1958 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1959 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1960 1961 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1962 1963 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1964 UseLargePages, addr); 1965 1966 if (UseCompressedOops) { 1967 if (addr != NULL && !heap_rs.is_reserved()) { 1968 // Failed to reserve at specified address - the requested memory 1969 // region is taken already, for example, by 'java' launcher. 1970 // Try again to reserver heap higher. 1971 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 1972 1973 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 1974 UseLargePages, addr); 1975 1976 if (addr != NULL && !heap_rs0.is_reserved()) { 1977 // Failed to reserve at specified address again - give up. 1978 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 1979 assert(addr == NULL, ""); 1980 1981 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 1982 UseLargePages, addr); 1983 heap_rs = heap_rs1; 1984 } else { 1985 heap_rs = heap_rs0; 1986 } 1987 } 1988 } 1989 1990 if (!heap_rs.is_reserved()) { 1991 vm_exit_during_initialization("Could not reserve enough space for object heap"); 1992 return JNI_ENOMEM; 1993 } 1994 1995 // It is important to do this in a way such that concurrent readers can't 1996 // temporarily think somethings in the heap. (I've actually seen this 1997 // happen in asserts: DLD.) 1998 _reserved.set_word_size(0); 1999 _reserved.set_start((HeapWord*)heap_rs.base()); 2000 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2001 2002 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2003 2004 // Create the gen rem set (and barrier set) for the entire reserved region. 2005 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2006 set_barrier_set(rem_set()->bs()); 2007 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2008 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2009 } else { 2010 vm_exit_during_initialization("G1 requires a mod ref bs."); 2011 return JNI_ENOMEM; 2012 } 2013 2014 // Also create a G1 rem set. 2015 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2016 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2017 } else { 2018 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2019 return JNI_ENOMEM; 2020 } 2021 2022 // Carve out the G1 part of the heap. 2023 2024 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2025 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2026 g1_rs.size()/HeapWordSize); 2027 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 2028 2029 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2030 2031 _g1_storage.initialize(g1_rs, 0); 2032 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2033 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2034 (HeapWord*) _g1_reserved.end(), 2035 _expansion_regions); 2036 2037 // 6843694 - ensure that the maximum region index can fit 2038 // in the remembered set structures. 2039 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2040 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2041 2042 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2043 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2044 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2045 "too many cards per region"); 2046 2047 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2048 2049 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2050 heap_word_size(init_byte_size)); 2051 2052 _g1h = this; 2053 2054 _in_cset_fast_test_length = max_regions(); 2055 _in_cset_fast_test_base = 2056 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length); 2057 2058 // We're biasing _in_cset_fast_test to avoid subtracting the 2059 // beginning of the heap every time we want to index; basically 2060 // it's the same with what we do with the card table. 2061 _in_cset_fast_test = _in_cset_fast_test_base - 2062 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2063 2064 // Clear the _cset_fast_test bitmap in anticipation of adding 2065 // regions to the incremental collection set for the first 2066 // evacuation pause. 2067 clear_cset_fast_test(); 2068 2069 // Create the ConcurrentMark data structure and thread. 2070 // (Must do this late, so that "max_regions" is defined.) 2071 _cm = new ConcurrentMark(heap_rs, max_regions()); 2072 _cmThread = _cm->cmThread(); 2073 2074 // Initialize the from_card cache structure of HeapRegionRemSet. 2075 HeapRegionRemSet::init_heap(max_regions()); 2076 2077 // Now expand into the initial heap size. 2078 if (!expand(init_byte_size)) { 2079 vm_exit_during_initialization("Failed to allocate initial heap."); 2080 return JNI_ENOMEM; 2081 } 2082 2083 // Perform any initialization actions delegated to the policy. 2084 g1_policy()->init(); 2085 2086 _refine_cte_cl = 2087 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2088 g1_rem_set(), 2089 concurrent_g1_refine()); 2090 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2091 2092 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2093 SATB_Q_FL_lock, 2094 G1SATBProcessCompletedThreshold, 2095 Shared_SATB_Q_lock); 2096 2097 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2098 DirtyCardQ_FL_lock, 2099 concurrent_g1_refine()->yellow_zone(), 2100 concurrent_g1_refine()->red_zone(), 2101 Shared_DirtyCardQ_lock); 2102 2103 if (G1DeferredRSUpdate) { 2104 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2105 DirtyCardQ_FL_lock, 2106 -1, // never trigger processing 2107 -1, // no limit on length 2108 Shared_DirtyCardQ_lock, 2109 &JavaThread::dirty_card_queue_set()); 2110 } 2111 2112 // Initialize the card queue set used to hold cards containing 2113 // references into the collection set. 2114 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2115 DirtyCardQ_FL_lock, 2116 -1, // never trigger processing 2117 -1, // no limit on length 2118 Shared_DirtyCardQ_lock, 2119 &JavaThread::dirty_card_queue_set()); 2120 2121 // In case we're keeping closure specialization stats, initialize those 2122 // counts and that mechanism. 2123 SpecializationStats::clear(); 2124 2125 // Do later initialization work for concurrent refinement. 2126 _cg1r->init(); 2127 2128 // Here we allocate the dummy full region that is required by the 2129 // G1AllocRegion class. If we don't pass an address in the reserved 2130 // space here, lots of asserts fire. 2131 2132 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2133 _g1_reserved.start()); 2134 // We'll re-use the same region whether the alloc region will 2135 // require BOT updates or not and, if it doesn't, then a non-young 2136 // region will complain that it cannot support allocations without 2137 // BOT updates. So we'll tag the dummy region as young to avoid that. 2138 dummy_region->set_young(); 2139 // Make sure it's full. 2140 dummy_region->set_top(dummy_region->end()); 2141 G1AllocRegion::setup(this, dummy_region); 2142 2143 init_mutator_alloc_region(); 2144 2145 // Do create of the monitoring and management support so that 2146 // values in the heap have been properly initialized. 2147 _g1mm = new G1MonitoringSupport(this); 2148 2149 return JNI_OK; 2150 } 2151 2152 void G1CollectedHeap::ref_processing_init() { 2153 // Reference processing in G1 currently works as follows: 2154 // 2155 // * There are two reference processor instances. One is 2156 // used to record and process discovered references 2157 // during concurrent marking; the other is used to 2158 // record and process references during STW pauses 2159 // (both full and incremental). 2160 // * Both ref processors need to 'span' the entire heap as 2161 // the regions in the collection set may be dotted around. 2162 // 2163 // * For the concurrent marking ref processor: 2164 // * Reference discovery is enabled at initial marking. 2165 // * Reference discovery is disabled and the discovered 2166 // references processed etc during remarking. 2167 // * Reference discovery is MT (see below). 2168 // * Reference discovery requires a barrier (see below). 2169 // * Reference processing may or may not be MT 2170 // (depending on the value of ParallelRefProcEnabled 2171 // and ParallelGCThreads). 2172 // * A full GC disables reference discovery by the CM 2173 // ref processor and abandons any entries on it's 2174 // discovered lists. 2175 // 2176 // * For the STW processor: 2177 // * Non MT discovery is enabled at the start of a full GC. 2178 // * Processing and enqueueing during a full GC is non-MT. 2179 // * During a full GC, references are processed after marking. 2180 // 2181 // * Discovery (may or may not be MT) is enabled at the start 2182 // of an incremental evacuation pause. 2183 // * References are processed near the end of a STW evacuation pause. 2184 // * For both types of GC: 2185 // * Discovery is atomic - i.e. not concurrent. 2186 // * Reference discovery will not need a barrier. 2187 2188 SharedHeap::ref_processing_init(); 2189 MemRegion mr = reserved_region(); 2190 2191 // Concurrent Mark ref processor 2192 _ref_processor_cm = 2193 new ReferenceProcessor(mr, // span 2194 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2195 // mt processing 2196 (int) ParallelGCThreads, 2197 // degree of mt processing 2198 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2199 // mt discovery 2200 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2201 // degree of mt discovery 2202 false, 2203 // Reference discovery is not atomic 2204 &_is_alive_closure_cm, 2205 // is alive closure 2206 // (for efficiency/performance) 2207 true); 2208 // Setting next fields of discovered 2209 // lists requires a barrier. 2210 2211 // STW ref processor 2212 _ref_processor_stw = 2213 new ReferenceProcessor(mr, // span 2214 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2215 // mt processing 2216 MAX2((int)ParallelGCThreads, 1), 2217 // degree of mt processing 2218 (ParallelGCThreads > 1), 2219 // mt discovery 2220 MAX2((int)ParallelGCThreads, 1), 2221 // degree of mt discovery 2222 true, 2223 // Reference discovery is atomic 2224 &_is_alive_closure_stw, 2225 // is alive closure 2226 // (for efficiency/performance) 2227 false); 2228 // Setting next fields of discovered 2229 // lists requires a barrier. 2230 } 2231 2232 size_t G1CollectedHeap::capacity() const { 2233 return _g1_committed.byte_size(); 2234 } 2235 2236 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2237 DirtyCardQueue* into_cset_dcq, 2238 bool concurrent, 2239 int worker_i) { 2240 // Clean cards in the hot card cache 2241 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2242 2243 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2244 int n_completed_buffers = 0; 2245 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2246 n_completed_buffers++; 2247 } 2248 g1_policy()->record_update_rs_processed_buffers(worker_i, 2249 (double) n_completed_buffers); 2250 dcqs.clear_n_completed_buffers(); 2251 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2252 } 2253 2254 2255 // Computes the sum of the storage used by the various regions. 2256 2257 size_t G1CollectedHeap::used() const { 2258 assert(Heap_lock->owner() != NULL, 2259 "Should be owned on this thread's behalf."); 2260 size_t result = _summary_bytes_used; 2261 // Read only once in case it is set to NULL concurrently 2262 HeapRegion* hr = _mutator_alloc_region.get(); 2263 if (hr != NULL) 2264 result += hr->used(); 2265 return result; 2266 } 2267 2268 size_t G1CollectedHeap::used_unlocked() const { 2269 size_t result = _summary_bytes_used; 2270 return result; 2271 } 2272 2273 class SumUsedClosure: public HeapRegionClosure { 2274 size_t _used; 2275 public: 2276 SumUsedClosure() : _used(0) {} 2277 bool doHeapRegion(HeapRegion* r) { 2278 if (!r->continuesHumongous()) { 2279 _used += r->used(); 2280 } 2281 return false; 2282 } 2283 size_t result() { return _used; } 2284 }; 2285 2286 size_t G1CollectedHeap::recalculate_used() const { 2287 SumUsedClosure blk; 2288 heap_region_iterate(&blk); 2289 return blk.result(); 2290 } 2291 2292 size_t G1CollectedHeap::unsafe_max_alloc() { 2293 if (free_regions() > 0) return HeapRegion::GrainBytes; 2294 // otherwise, is there space in the current allocation region? 2295 2296 // We need to store the current allocation region in a local variable 2297 // here. The problem is that this method doesn't take any locks and 2298 // there may be other threads which overwrite the current allocation 2299 // region field. attempt_allocation(), for example, sets it to NULL 2300 // and this can happen *after* the NULL check here but before the call 2301 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2302 // to be a problem in the optimized build, since the two loads of the 2303 // current allocation region field are optimized away. 2304 HeapRegion* hr = _mutator_alloc_region.get(); 2305 if (hr == NULL) { 2306 return 0; 2307 } 2308 return hr->free(); 2309 } 2310 2311 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2312 switch (cause) { 2313 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2314 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2315 case GCCause::_g1_humongous_allocation: return true; 2316 default: return false; 2317 } 2318 } 2319 2320 #ifndef PRODUCT 2321 void G1CollectedHeap::allocate_dummy_regions() { 2322 // Let's fill up most of the region 2323 size_t word_size = HeapRegion::GrainWords - 1024; 2324 // And as a result the region we'll allocate will be humongous. 2325 guarantee(isHumongous(word_size), "sanity"); 2326 2327 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2328 // Let's use the existing mechanism for the allocation 2329 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2330 if (dummy_obj != NULL) { 2331 MemRegion mr(dummy_obj, word_size); 2332 CollectedHeap::fill_with_object(mr); 2333 } else { 2334 // If we can't allocate once, we probably cannot allocate 2335 // again. Let's get out of the loop. 2336 break; 2337 } 2338 } 2339 } 2340 #endif // !PRODUCT 2341 2342 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) { 2343 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2344 2345 // We assume that if concurrent == true, then the caller is a 2346 // concurrent thread that was joined the Suspendible Thread 2347 // Set. If there's ever a cheap way to check this, we should add an 2348 // assert here. 2349 2350 // We have already incremented _total_full_collections at the start 2351 // of the GC, so total_full_collections() represents how many full 2352 // collections have been started. 2353 unsigned int full_collections_started = total_full_collections(); 2354 2355 // Given that this method is called at the end of a Full GC or of a 2356 // concurrent cycle, and those can be nested (i.e., a Full GC can 2357 // interrupt a concurrent cycle), the number of full collections 2358 // completed should be either one (in the case where there was no 2359 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2360 // behind the number of full collections started. 2361 2362 // This is the case for the inner caller, i.e. a Full GC. 2363 assert(concurrent || 2364 (full_collections_started == _full_collections_completed + 1) || 2365 (full_collections_started == _full_collections_completed + 2), 2366 err_msg("for inner caller (Full GC): full_collections_started = %u " 2367 "is inconsistent with _full_collections_completed = %u", 2368 full_collections_started, _full_collections_completed)); 2369 2370 // This is the case for the outer caller, i.e. the concurrent cycle. 2371 assert(!concurrent || 2372 (full_collections_started == _full_collections_completed + 1), 2373 err_msg("for outer caller (concurrent cycle): " 2374 "full_collections_started = %u " 2375 "is inconsistent with _full_collections_completed = %u", 2376 full_collections_started, _full_collections_completed)); 2377 2378 _full_collections_completed += 1; 2379 2380 // We need to clear the "in_progress" flag in the CM thread before 2381 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2382 // is set) so that if a waiter requests another System.gc() it doesn't 2383 // incorrectly see that a marking cyle is still in progress. 2384 if (concurrent) { 2385 _cmThread->clear_in_progress(); 2386 } 2387 2388 // This notify_all() will ensure that a thread that called 2389 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2390 // and it's waiting for a full GC to finish will be woken up. It is 2391 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2392 FullGCCount_lock->notify_all(); 2393 } 2394 2395 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2396 assert_at_safepoint(true /* should_be_vm_thread */); 2397 GCCauseSetter gcs(this, cause); 2398 switch (cause) { 2399 case GCCause::_heap_inspection: 2400 case GCCause::_heap_dump: { 2401 HandleMark hm; 2402 do_full_collection(false); // don't clear all soft refs 2403 break; 2404 } 2405 default: // XXX FIX ME 2406 ShouldNotReachHere(); // Unexpected use of this function 2407 } 2408 } 2409 2410 void G1CollectedHeap::collect(GCCause::Cause cause) { 2411 assert_heap_not_locked(); 2412 2413 unsigned int gc_count_before; 2414 unsigned int full_gc_count_before; 2415 bool retry_gc; 2416 2417 do { 2418 retry_gc = false; 2419 2420 { 2421 MutexLocker ml(Heap_lock); 2422 2423 // Read the GC count while holding the Heap_lock 2424 gc_count_before = total_collections(); 2425 full_gc_count_before = total_full_collections(); 2426 } 2427 2428 if (should_do_concurrent_full_gc(cause)) { 2429 // Schedule an initial-mark evacuation pause that will start a 2430 // concurrent cycle. We're setting word_size to 0 which means that 2431 // we are not requesting a post-GC allocation. 2432 VM_G1IncCollectionPause op(gc_count_before, 2433 0, /* word_size */ 2434 true, /* should_initiate_conc_mark */ 2435 g1_policy()->max_pause_time_ms(), 2436 cause); 2437 2438 VMThread::execute(&op); 2439 if (!op.pause_succeeded()) { 2440 if (full_gc_count_before == total_full_collections()) { 2441 retry_gc = op.should_retry_gc(); 2442 } else { 2443 // A Full GC happened while we were trying to schedule the 2444 // initial-mark GC. No point in starting a new cycle given 2445 // that the whole heap was collected anyway. 2446 } 2447 2448 if (retry_gc) { 2449 if (GC_locker::is_active_and_needs_gc()) { 2450 GC_locker::stall_until_clear(); 2451 } 2452 } 2453 } 2454 } else { 2455 if (cause == GCCause::_gc_locker 2456 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2457 2458 // Schedule a standard evacuation pause. We're setting word_size 2459 // to 0 which means that we are not requesting a post-GC allocation. 2460 VM_G1IncCollectionPause op(gc_count_before, 2461 0, /* word_size */ 2462 false, /* should_initiate_conc_mark */ 2463 g1_policy()->max_pause_time_ms(), 2464 cause); 2465 VMThread::execute(&op); 2466 } else { 2467 // Schedule a Full GC. 2468 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause); 2469 VMThread::execute(&op); 2470 } 2471 } 2472 } while (retry_gc); 2473 } 2474 2475 bool G1CollectedHeap::is_in(const void* p) const { 2476 if (_g1_committed.contains(p)) { 2477 // Given that we know that p is in the committed space, 2478 // heap_region_containing_raw() should successfully 2479 // return the containing region. 2480 HeapRegion* hr = heap_region_containing_raw(p); 2481 return hr->is_in(p); 2482 } else { 2483 return _perm_gen->as_gen()->is_in(p); 2484 } 2485 } 2486 2487 // Iteration functions. 2488 2489 // Iterates an OopClosure over all ref-containing fields of objects 2490 // within a HeapRegion. 2491 2492 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2493 MemRegion _mr; 2494 OopClosure* _cl; 2495 public: 2496 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2497 : _mr(mr), _cl(cl) {} 2498 bool doHeapRegion(HeapRegion* r) { 2499 if (! r->continuesHumongous()) { 2500 r->oop_iterate(_cl); 2501 } 2502 return false; 2503 } 2504 }; 2505 2506 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2507 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2508 heap_region_iterate(&blk); 2509 if (do_perm) { 2510 perm_gen()->oop_iterate(cl); 2511 } 2512 } 2513 2514 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2515 IterateOopClosureRegionClosure blk(mr, cl); 2516 heap_region_iterate(&blk); 2517 if (do_perm) { 2518 perm_gen()->oop_iterate(cl); 2519 } 2520 } 2521 2522 // Iterates an ObjectClosure over all objects within a HeapRegion. 2523 2524 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2525 ObjectClosure* _cl; 2526 public: 2527 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2528 bool doHeapRegion(HeapRegion* r) { 2529 if (! r->continuesHumongous()) { 2530 r->object_iterate(_cl); 2531 } 2532 return false; 2533 } 2534 }; 2535 2536 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2537 IterateObjectClosureRegionClosure blk(cl); 2538 heap_region_iterate(&blk); 2539 if (do_perm) { 2540 perm_gen()->object_iterate(cl); 2541 } 2542 } 2543 2544 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2545 // FIXME: is this right? 2546 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2547 } 2548 2549 // Calls a SpaceClosure on a HeapRegion. 2550 2551 class SpaceClosureRegionClosure: public HeapRegionClosure { 2552 SpaceClosure* _cl; 2553 public: 2554 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2555 bool doHeapRegion(HeapRegion* r) { 2556 _cl->do_space(r); 2557 return false; 2558 } 2559 }; 2560 2561 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2562 SpaceClosureRegionClosure blk(cl); 2563 heap_region_iterate(&blk); 2564 } 2565 2566 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2567 _hrs.iterate(cl); 2568 } 2569 2570 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r, 2571 HeapRegionClosure* cl) const { 2572 _hrs.iterate_from(r, cl); 2573 } 2574 2575 void 2576 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2577 uint worker, 2578 uint no_of_par_workers, 2579 jint claim_value) { 2580 const uint regions = n_regions(); 2581 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2582 no_of_par_workers : 2583 1); 2584 assert(UseDynamicNumberOfGCThreads || 2585 no_of_par_workers == workers()->total_workers(), 2586 "Non dynamic should use fixed number of workers"); 2587 // try to spread out the starting points of the workers 2588 const uint start_index = regions / max_workers * worker; 2589 2590 // each worker will actually look at all regions 2591 for (uint count = 0; count < regions; ++count) { 2592 const uint index = (start_index + count) % regions; 2593 assert(0 <= index && index < regions, "sanity"); 2594 HeapRegion* r = region_at(index); 2595 // we'll ignore "continues humongous" regions (we'll process them 2596 // when we come across their corresponding "start humongous" 2597 // region) and regions already claimed 2598 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2599 continue; 2600 } 2601 // OK, try to claim it 2602 if (r->claimHeapRegion(claim_value)) { 2603 // success! 2604 assert(!r->continuesHumongous(), "sanity"); 2605 if (r->startsHumongous()) { 2606 // If the region is "starts humongous" we'll iterate over its 2607 // "continues humongous" first; in fact we'll do them 2608 // first. The order is important. In on case, calling the 2609 // closure on the "starts humongous" region might de-allocate 2610 // and clear all its "continues humongous" regions and, as a 2611 // result, we might end up processing them twice. So, we'll do 2612 // them first (notice: most closures will ignore them anyway) and 2613 // then we'll do the "starts humongous" region. 2614 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2615 HeapRegion* chr = region_at(ch_index); 2616 2617 // if the region has already been claimed or it's not 2618 // "continues humongous" we're done 2619 if (chr->claim_value() == claim_value || 2620 !chr->continuesHumongous()) { 2621 break; 2622 } 2623 2624 // Noone should have claimed it directly. We can given 2625 // that we claimed its "starts humongous" region. 2626 assert(chr->claim_value() != claim_value, "sanity"); 2627 assert(chr->humongous_start_region() == r, "sanity"); 2628 2629 if (chr->claimHeapRegion(claim_value)) { 2630 // we should always be able to claim it; noone else should 2631 // be trying to claim this region 2632 2633 bool res2 = cl->doHeapRegion(chr); 2634 assert(!res2, "Should not abort"); 2635 2636 // Right now, this holds (i.e., no closure that actually 2637 // does something with "continues humongous" regions 2638 // clears them). We might have to weaken it in the future, 2639 // but let's leave these two asserts here for extra safety. 2640 assert(chr->continuesHumongous(), "should still be the case"); 2641 assert(chr->humongous_start_region() == r, "sanity"); 2642 } else { 2643 guarantee(false, "we should not reach here"); 2644 } 2645 } 2646 } 2647 2648 assert(!r->continuesHumongous(), "sanity"); 2649 bool res = cl->doHeapRegion(r); 2650 assert(!res, "Should not abort"); 2651 } 2652 } 2653 } 2654 2655 class ResetClaimValuesClosure: public HeapRegionClosure { 2656 public: 2657 bool doHeapRegion(HeapRegion* r) { 2658 r->set_claim_value(HeapRegion::InitialClaimValue); 2659 return false; 2660 } 2661 }; 2662 2663 void G1CollectedHeap::reset_heap_region_claim_values() { 2664 ResetClaimValuesClosure blk; 2665 heap_region_iterate(&blk); 2666 } 2667 2668 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2669 ResetClaimValuesClosure blk; 2670 collection_set_iterate(&blk); 2671 } 2672 2673 #ifdef ASSERT 2674 // This checks whether all regions in the heap have the correct claim 2675 // value. I also piggy-backed on this a check to ensure that the 2676 // humongous_start_region() information on "continues humongous" 2677 // regions is correct. 2678 2679 class CheckClaimValuesClosure : public HeapRegionClosure { 2680 private: 2681 jint _claim_value; 2682 uint _failures; 2683 HeapRegion* _sh_region; 2684 2685 public: 2686 CheckClaimValuesClosure(jint claim_value) : 2687 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2688 bool doHeapRegion(HeapRegion* r) { 2689 if (r->claim_value() != _claim_value) { 2690 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2691 "claim value = %d, should be %d", 2692 HR_FORMAT_PARAMS(r), 2693 r->claim_value(), _claim_value); 2694 ++_failures; 2695 } 2696 if (!r->isHumongous()) { 2697 _sh_region = NULL; 2698 } else if (r->startsHumongous()) { 2699 _sh_region = r; 2700 } else if (r->continuesHumongous()) { 2701 if (r->humongous_start_region() != _sh_region) { 2702 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2703 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2704 HR_FORMAT_PARAMS(r), 2705 r->humongous_start_region(), 2706 _sh_region); 2707 ++_failures; 2708 } 2709 } 2710 return false; 2711 } 2712 uint failures() { return _failures; } 2713 }; 2714 2715 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2716 CheckClaimValuesClosure cl(claim_value); 2717 heap_region_iterate(&cl); 2718 return cl.failures() == 0; 2719 } 2720 2721 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2722 private: 2723 jint _claim_value; 2724 uint _failures; 2725 2726 public: 2727 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2728 _claim_value(claim_value), _failures(0) { } 2729 2730 uint failures() { return _failures; } 2731 2732 bool doHeapRegion(HeapRegion* hr) { 2733 assert(hr->in_collection_set(), "how?"); 2734 assert(!hr->isHumongous(), "H-region in CSet"); 2735 if (hr->claim_value() != _claim_value) { 2736 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2737 "claim value = %d, should be %d", 2738 HR_FORMAT_PARAMS(hr), 2739 hr->claim_value(), _claim_value); 2740 _failures += 1; 2741 } 2742 return false; 2743 } 2744 }; 2745 2746 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2747 CheckClaimValuesInCSetHRClosure cl(claim_value); 2748 collection_set_iterate(&cl); 2749 return cl.failures() == 0; 2750 } 2751 #endif // ASSERT 2752 2753 // Clear the cached CSet starting regions and (more importantly) 2754 // the time stamps. Called when we reset the GC time stamp. 2755 void G1CollectedHeap::clear_cset_start_regions() { 2756 assert(_worker_cset_start_region != NULL, "sanity"); 2757 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2758 2759 int n_queues = MAX2((int)ParallelGCThreads, 1); 2760 for (int i = 0; i < n_queues; i++) { 2761 _worker_cset_start_region[i] = NULL; 2762 _worker_cset_start_region_time_stamp[i] = 0; 2763 } 2764 } 2765 2766 // Given the id of a worker, obtain or calculate a suitable 2767 // starting region for iterating over the current collection set. 2768 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2769 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2770 2771 HeapRegion* result = NULL; 2772 unsigned gc_time_stamp = get_gc_time_stamp(); 2773 2774 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2775 // Cached starting region for current worker was set 2776 // during the current pause - so it's valid. 2777 // Note: the cached starting heap region may be NULL 2778 // (when the collection set is empty). 2779 result = _worker_cset_start_region[worker_i]; 2780 assert(result == NULL || result->in_collection_set(), "sanity"); 2781 return result; 2782 } 2783 2784 // The cached entry was not valid so let's calculate 2785 // a suitable starting heap region for this worker. 2786 2787 // We want the parallel threads to start their collection 2788 // set iteration at different collection set regions to 2789 // avoid contention. 2790 // If we have: 2791 // n collection set regions 2792 // p threads 2793 // Then thread t will start at region floor ((t * n) / p) 2794 2795 result = g1_policy()->collection_set(); 2796 if (G1CollectedHeap::use_parallel_gc_threads()) { 2797 uint cs_size = g1_policy()->cset_region_length(); 2798 uint active_workers = workers()->active_workers(); 2799 assert(UseDynamicNumberOfGCThreads || 2800 active_workers == workers()->total_workers(), 2801 "Unless dynamic should use total workers"); 2802 2803 uint end_ind = (cs_size * worker_i) / active_workers; 2804 uint start_ind = 0; 2805 2806 if (worker_i > 0 && 2807 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2808 // Previous workers starting region is valid 2809 // so let's iterate from there 2810 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2811 result = _worker_cset_start_region[worker_i - 1]; 2812 } 2813 2814 for (uint i = start_ind; i < end_ind; i++) { 2815 result = result->next_in_collection_set(); 2816 } 2817 } 2818 2819 // Note: the calculated starting heap region may be NULL 2820 // (when the collection set is empty). 2821 assert(result == NULL || result->in_collection_set(), "sanity"); 2822 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2823 "should be updated only once per pause"); 2824 _worker_cset_start_region[worker_i] = result; 2825 OrderAccess::storestore(); 2826 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2827 return result; 2828 } 2829 2830 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2831 HeapRegion* r = g1_policy()->collection_set(); 2832 while (r != NULL) { 2833 HeapRegion* next = r->next_in_collection_set(); 2834 if (cl->doHeapRegion(r)) { 2835 cl->incomplete(); 2836 return; 2837 } 2838 r = next; 2839 } 2840 } 2841 2842 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2843 HeapRegionClosure *cl) { 2844 if (r == NULL) { 2845 // The CSet is empty so there's nothing to do. 2846 return; 2847 } 2848 2849 assert(r->in_collection_set(), 2850 "Start region must be a member of the collection set."); 2851 HeapRegion* cur = r; 2852 while (cur != NULL) { 2853 HeapRegion* next = cur->next_in_collection_set(); 2854 if (cl->doHeapRegion(cur) && false) { 2855 cl->incomplete(); 2856 return; 2857 } 2858 cur = next; 2859 } 2860 cur = g1_policy()->collection_set(); 2861 while (cur != r) { 2862 HeapRegion* next = cur->next_in_collection_set(); 2863 if (cl->doHeapRegion(cur) && false) { 2864 cl->incomplete(); 2865 return; 2866 } 2867 cur = next; 2868 } 2869 } 2870 2871 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2872 return n_regions() > 0 ? region_at(0) : NULL; 2873 } 2874 2875 2876 Space* G1CollectedHeap::space_containing(const void* addr) const { 2877 Space* res = heap_region_containing(addr); 2878 if (res == NULL) 2879 res = perm_gen()->space_containing(addr); 2880 return res; 2881 } 2882 2883 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2884 Space* sp = space_containing(addr); 2885 if (sp != NULL) { 2886 return sp->block_start(addr); 2887 } 2888 return NULL; 2889 } 2890 2891 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2892 Space* sp = space_containing(addr); 2893 assert(sp != NULL, "block_size of address outside of heap"); 2894 return sp->block_size(addr); 2895 } 2896 2897 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2898 Space* sp = space_containing(addr); 2899 return sp->block_is_obj(addr); 2900 } 2901 2902 bool G1CollectedHeap::supports_tlab_allocation() const { 2903 return true; 2904 } 2905 2906 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2907 return HeapRegion::GrainBytes; 2908 } 2909 2910 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2911 // Return the remaining space in the cur alloc region, but not less than 2912 // the min TLAB size. 2913 2914 // Also, this value can be at most the humongous object threshold, 2915 // since we can't allow tlabs to grow big enough to accomodate 2916 // humongous objects. 2917 2918 HeapRegion* hr = _mutator_alloc_region.get(); 2919 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2920 if (hr == NULL) { 2921 return max_tlab_size; 2922 } else { 2923 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2924 } 2925 } 2926 2927 size_t G1CollectedHeap::max_capacity() const { 2928 return _g1_reserved.byte_size(); 2929 } 2930 2931 jlong G1CollectedHeap::millis_since_last_gc() { 2932 // assert(false, "NYI"); 2933 return 0; 2934 } 2935 2936 void G1CollectedHeap::prepare_for_verify() { 2937 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2938 ensure_parsability(false); 2939 } 2940 g1_rem_set()->prepare_for_verify(); 2941 } 2942 2943 class VerifyLivenessOopClosure: public OopClosure { 2944 G1CollectedHeap* _g1h; 2945 VerifyOption _vo; 2946 public: 2947 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 2948 _g1h(g1h), _vo(vo) 2949 { } 2950 void do_oop(narrowOop *p) { do_oop_work(p); } 2951 void do_oop( oop *p) { do_oop_work(p); } 2952 2953 template <class T> void do_oop_work(T *p) { 2954 oop obj = oopDesc::load_decode_heap_oop(p); 2955 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 2956 "Dead object referenced by a not dead object"); 2957 } 2958 }; 2959 2960 class VerifyObjsInRegionClosure: public ObjectClosure { 2961 private: 2962 G1CollectedHeap* _g1h; 2963 size_t _live_bytes; 2964 HeapRegion *_hr; 2965 VerifyOption _vo; 2966 public: 2967 // _vo == UsePrevMarking -> use "prev" marking information, 2968 // _vo == UseNextMarking -> use "next" marking information, 2969 // _vo == UseMarkWord -> use mark word from object header. 2970 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 2971 : _live_bytes(0), _hr(hr), _vo(vo) { 2972 _g1h = G1CollectedHeap::heap(); 2973 } 2974 void do_object(oop o) { 2975 VerifyLivenessOopClosure isLive(_g1h, _vo); 2976 assert(o != NULL, "Huh?"); 2977 if (!_g1h->is_obj_dead_cond(o, _vo)) { 2978 // If the object is alive according to the mark word, 2979 // then verify that the marking information agrees. 2980 // Note we can't verify the contra-positive of the 2981 // above: if the object is dead (according to the mark 2982 // word), it may not be marked, or may have been marked 2983 // but has since became dead, or may have been allocated 2984 // since the last marking. 2985 if (_vo == VerifyOption_G1UseMarkWord) { 2986 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 2987 } 2988 2989 o->oop_iterate(&isLive); 2990 if (!_hr->obj_allocated_since_prev_marking(o)) { 2991 size_t obj_size = o->size(); // Make sure we don't overflow 2992 _live_bytes += (obj_size * HeapWordSize); 2993 } 2994 } 2995 } 2996 size_t live_bytes() { return _live_bytes; } 2997 }; 2998 2999 class PrintObjsInRegionClosure : public ObjectClosure { 3000 HeapRegion *_hr; 3001 G1CollectedHeap *_g1; 3002 public: 3003 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3004 _g1 = G1CollectedHeap::heap(); 3005 }; 3006 3007 void do_object(oop o) { 3008 if (o != NULL) { 3009 HeapWord *start = (HeapWord *) o; 3010 size_t word_sz = o->size(); 3011 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3012 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3013 (void*) o, word_sz, 3014 _g1->isMarkedPrev(o), 3015 _g1->isMarkedNext(o), 3016 _hr->obj_allocated_since_prev_marking(o)); 3017 HeapWord *end = start + word_sz; 3018 HeapWord *cur; 3019 int *val; 3020 for (cur = start; cur < end; cur++) { 3021 val = (int *) cur; 3022 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3023 } 3024 } 3025 } 3026 }; 3027 3028 class VerifyRegionClosure: public HeapRegionClosure { 3029 private: 3030 bool _par; 3031 VerifyOption _vo; 3032 bool _failures; 3033 public: 3034 // _vo == UsePrevMarking -> use "prev" marking information, 3035 // _vo == UseNextMarking -> use "next" marking information, 3036 // _vo == UseMarkWord -> use mark word from object header. 3037 VerifyRegionClosure(bool par, VerifyOption vo) 3038 : _par(par), 3039 _vo(vo), 3040 _failures(false) {} 3041 3042 bool failures() { 3043 return _failures; 3044 } 3045 3046 bool doHeapRegion(HeapRegion* r) { 3047 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue, 3048 "Should be unclaimed at verify points."); 3049 if (!r->continuesHumongous()) { 3050 bool failures = false; 3051 r->verify(_vo, &failures); 3052 if (failures) { 3053 _failures = true; 3054 } else { 3055 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3056 r->object_iterate(¬_dead_yet_cl); 3057 if (_vo != VerifyOption_G1UseNextMarking) { 3058 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3059 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3060 "max_live_bytes "SIZE_FORMAT" " 3061 "< calculated "SIZE_FORMAT, 3062 r->bottom(), r->end(), 3063 r->max_live_bytes(), 3064 not_dead_yet_cl.live_bytes()); 3065 _failures = true; 3066 } 3067 } else { 3068 // When vo == UseNextMarking we cannot currently do a sanity 3069 // check on the live bytes as the calculation has not been 3070 // finalized yet. 3071 } 3072 } 3073 } 3074 return false; // stop the region iteration if we hit a failure 3075 } 3076 }; 3077 3078 class VerifyRootsClosure: public OopsInGenClosure { 3079 private: 3080 G1CollectedHeap* _g1h; 3081 VerifyOption _vo; 3082 bool _failures; 3083 public: 3084 // _vo == UsePrevMarking -> use "prev" marking information, 3085 // _vo == UseNextMarking -> use "next" marking information, 3086 // _vo == UseMarkWord -> use mark word from object header. 3087 VerifyRootsClosure(VerifyOption vo) : 3088 _g1h(G1CollectedHeap::heap()), 3089 _vo(vo), 3090 _failures(false) { } 3091 3092 bool failures() { return _failures; } 3093 3094 template <class T> void do_oop_nv(T* p) { 3095 T heap_oop = oopDesc::load_heap_oop(p); 3096 if (!oopDesc::is_null(heap_oop)) { 3097 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3098 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3099 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3100 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3101 if (_vo == VerifyOption_G1UseMarkWord) { 3102 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3103 } 3104 obj->print_on(gclog_or_tty); 3105 _failures = true; 3106 } 3107 } 3108 } 3109 3110 void do_oop(oop* p) { do_oop_nv(p); } 3111 void do_oop(narrowOop* p) { do_oop_nv(p); } 3112 }; 3113 3114 // This is the task used for parallel heap verification. 3115 3116 class G1ParVerifyTask: public AbstractGangTask { 3117 private: 3118 G1CollectedHeap* _g1h; 3119 VerifyOption _vo; 3120 bool _failures; 3121 3122 public: 3123 // _vo == UsePrevMarking -> use "prev" marking information, 3124 // _vo == UseNextMarking -> use "next" marking information, 3125 // _vo == UseMarkWord -> use mark word from object header. 3126 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3127 AbstractGangTask("Parallel verify task"), 3128 _g1h(g1h), 3129 _vo(vo), 3130 _failures(false) { } 3131 3132 bool failures() { 3133 return _failures; 3134 } 3135 3136 void work(uint worker_id) { 3137 HandleMark hm; 3138 VerifyRegionClosure blk(true, _vo); 3139 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3140 _g1h->workers()->active_workers(), 3141 HeapRegion::ParVerifyClaimValue); 3142 if (blk.failures()) { 3143 _failures = true; 3144 } 3145 } 3146 }; 3147 3148 void G1CollectedHeap::verify(bool silent) { 3149 verify(silent, VerifyOption_G1UsePrevMarking); 3150 } 3151 3152 void G1CollectedHeap::verify(bool silent, 3153 VerifyOption vo) { 3154 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3155 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3156 VerifyRootsClosure rootsCl(vo); 3157 3158 assert(Thread::current()->is_VM_thread(), 3159 "Expected to be executed serially by the VM thread at this point"); 3160 3161 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3162 3163 // We apply the relevant closures to all the oops in the 3164 // system dictionary, the string table and the code cache. 3165 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3166 3167 process_strong_roots(true, // activate StrongRootsScope 3168 true, // we set "collecting perm gen" to true, 3169 // so we don't reset the dirty cards in the perm gen. 3170 ScanningOption(so), // roots scanning options 3171 &rootsCl, 3172 &blobsCl, 3173 &rootsCl); 3174 3175 // If we're verifying after the marking phase of a Full GC then we can't 3176 // treat the perm gen as roots into the G1 heap. Some of the objects in 3177 // the perm gen may be dead and hence not marked. If one of these dead 3178 // objects is considered to be a root then we may end up with a false 3179 // "Root location <x> points to dead ob <y>" failure. 3180 if (vo != VerifyOption_G1UseMarkWord) { 3181 // Since we used "collecting_perm_gen" == true above, we will not have 3182 // checked the refs from perm into the G1-collected heap. We check those 3183 // references explicitly below. Whether the relevant cards are dirty 3184 // is checked further below in the rem set verification. 3185 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3186 perm_gen()->oop_iterate(&rootsCl); 3187 } 3188 bool failures = rootsCl.failures(); 3189 3190 if (vo != VerifyOption_G1UseMarkWord) { 3191 // If we're verifying during a full GC then the region sets 3192 // will have been torn down at the start of the GC. Therefore 3193 // verifying the region sets will fail. So we only verify 3194 // the region sets when not in a full GC. 3195 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3196 verify_region_sets(); 3197 } 3198 3199 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3200 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3201 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3202 "sanity check"); 3203 3204 G1ParVerifyTask task(this, vo); 3205 assert(UseDynamicNumberOfGCThreads || 3206 workers()->active_workers() == workers()->total_workers(), 3207 "If not dynamic should be using all the workers"); 3208 int n_workers = workers()->active_workers(); 3209 set_par_threads(n_workers); 3210 workers()->run_task(&task); 3211 set_par_threads(0); 3212 if (task.failures()) { 3213 failures = true; 3214 } 3215 3216 // Checks that the expected amount of parallel work was done. 3217 // The implication is that n_workers is > 0. 3218 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3219 "sanity check"); 3220 3221 reset_heap_region_claim_values(); 3222 3223 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3224 "sanity check"); 3225 } else { 3226 VerifyRegionClosure blk(false, vo); 3227 heap_region_iterate(&blk); 3228 if (blk.failures()) { 3229 failures = true; 3230 } 3231 } 3232 if (!silent) gclog_or_tty->print("RemSet "); 3233 rem_set()->verify(); 3234 3235 if (failures) { 3236 gclog_or_tty->print_cr("Heap:"); 3237 // It helps to have the per-region information in the output to 3238 // help us track down what went wrong. This is why we call 3239 // print_extended_on() instead of print_on(). 3240 print_extended_on(gclog_or_tty); 3241 gclog_or_tty->print_cr(""); 3242 #ifndef PRODUCT 3243 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3244 concurrent_mark()->print_reachable("at-verification-failure", 3245 vo, false /* all */); 3246 } 3247 #endif 3248 gclog_or_tty->flush(); 3249 } 3250 guarantee(!failures, "there should not have been any failures"); 3251 } else { 3252 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3253 } 3254 } 3255 3256 class PrintRegionClosure: public HeapRegionClosure { 3257 outputStream* _st; 3258 public: 3259 PrintRegionClosure(outputStream* st) : _st(st) {} 3260 bool doHeapRegion(HeapRegion* r) { 3261 r->print_on(_st); 3262 return false; 3263 } 3264 }; 3265 3266 void G1CollectedHeap::print_on(outputStream* st) const { 3267 st->print(" %-20s", "garbage-first heap"); 3268 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3269 capacity()/K, used_unlocked()/K); 3270 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3271 _g1_storage.low_boundary(), 3272 _g1_storage.high(), 3273 _g1_storage.high_boundary()); 3274 st->cr(); 3275 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3276 uint young_regions = _young_list->length(); 3277 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3278 (size_t) young_regions * HeapRegion::GrainBytes / K); 3279 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3280 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3281 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3282 st->cr(); 3283 perm()->as_gen()->print_on(st); 3284 } 3285 3286 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3287 print_on(st); 3288 3289 // Print the per-region information. 3290 st->cr(); 3291 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3292 "HS=humongous(starts), HC=humongous(continues), " 3293 "CS=collection set, F=free, TS=gc time stamp, " 3294 "PTAMS=previous top-at-mark-start, " 3295 "NTAMS=next top-at-mark-start)"); 3296 PrintRegionClosure blk(st); 3297 heap_region_iterate(&blk); 3298 } 3299 3300 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3301 if (G1CollectedHeap::use_parallel_gc_threads()) { 3302 workers()->print_worker_threads_on(st); 3303 } 3304 _cmThread->print_on(st); 3305 st->cr(); 3306 _cm->print_worker_threads_on(st); 3307 _cg1r->print_worker_threads_on(st); 3308 st->cr(); 3309 } 3310 3311 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3312 if (G1CollectedHeap::use_parallel_gc_threads()) { 3313 workers()->threads_do(tc); 3314 } 3315 tc->do_thread(_cmThread); 3316 _cg1r->threads_do(tc); 3317 } 3318 3319 void G1CollectedHeap::print_tracing_info() const { 3320 // We'll overload this to mean "trace GC pause statistics." 3321 if (TraceGen0Time || TraceGen1Time) { 3322 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3323 // to that. 3324 g1_policy()->print_tracing_info(); 3325 } 3326 if (G1SummarizeRSetStats) { 3327 g1_rem_set()->print_summary_info(); 3328 } 3329 if (G1SummarizeConcMark) { 3330 concurrent_mark()->print_summary_info(); 3331 } 3332 g1_policy()->print_yg_surv_rate_info(); 3333 SpecializationStats::print(); 3334 } 3335 3336 #ifndef PRODUCT 3337 // Helpful for debugging RSet issues. 3338 3339 class PrintRSetsClosure : public HeapRegionClosure { 3340 private: 3341 const char* _msg; 3342 size_t _occupied_sum; 3343 3344 public: 3345 bool doHeapRegion(HeapRegion* r) { 3346 HeapRegionRemSet* hrrs = r->rem_set(); 3347 size_t occupied = hrrs->occupied(); 3348 _occupied_sum += occupied; 3349 3350 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3351 HR_FORMAT_PARAMS(r)); 3352 if (occupied == 0) { 3353 gclog_or_tty->print_cr(" RSet is empty"); 3354 } else { 3355 hrrs->print(); 3356 } 3357 gclog_or_tty->print_cr("----------"); 3358 return false; 3359 } 3360 3361 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3362 gclog_or_tty->cr(); 3363 gclog_or_tty->print_cr("========================================"); 3364 gclog_or_tty->print_cr(msg); 3365 gclog_or_tty->cr(); 3366 } 3367 3368 ~PrintRSetsClosure() { 3369 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3370 gclog_or_tty->print_cr("========================================"); 3371 gclog_or_tty->cr(); 3372 } 3373 }; 3374 3375 void G1CollectedHeap::print_cset_rsets() { 3376 PrintRSetsClosure cl("Printing CSet RSets"); 3377 collection_set_iterate(&cl); 3378 } 3379 3380 void G1CollectedHeap::print_all_rsets() { 3381 PrintRSetsClosure cl("Printing All RSets");; 3382 heap_region_iterate(&cl); 3383 } 3384 #endif // PRODUCT 3385 3386 G1CollectedHeap* G1CollectedHeap::heap() { 3387 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3388 "not a garbage-first heap"); 3389 return _g1h; 3390 } 3391 3392 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3393 // always_do_update_barrier = false; 3394 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3395 // Call allocation profiler 3396 AllocationProfiler::iterate_since_last_gc(); 3397 // Fill TLAB's and such 3398 ensure_parsability(true); 3399 } 3400 3401 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3402 // FIXME: what is this about? 3403 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3404 // is set. 3405 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3406 "derived pointer present")); 3407 // always_do_update_barrier = true; 3408 3409 // We have just completed a GC. Update the soft reference 3410 // policy with the new heap occupancy 3411 Universe::update_heap_info_at_gc(); 3412 } 3413 3414 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3415 unsigned int gc_count_before, 3416 bool* succeeded) { 3417 assert_heap_not_locked_and_not_at_safepoint(); 3418 g1_policy()->record_stop_world_start(); 3419 VM_G1IncCollectionPause op(gc_count_before, 3420 word_size, 3421 false, /* should_initiate_conc_mark */ 3422 g1_policy()->max_pause_time_ms(), 3423 GCCause::_g1_inc_collection_pause); 3424 VMThread::execute(&op); 3425 3426 HeapWord* result = op.result(); 3427 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3428 assert(result == NULL || ret_succeeded, 3429 "the result should be NULL if the VM did not succeed"); 3430 *succeeded = ret_succeeded; 3431 3432 assert_heap_not_locked(); 3433 return result; 3434 } 3435 3436 void 3437 G1CollectedHeap::doConcurrentMark() { 3438 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3439 if (!_cmThread->in_progress()) { 3440 _cmThread->set_started(); 3441 CGC_lock->notify(); 3442 } 3443 } 3444 3445 size_t G1CollectedHeap::pending_card_num() { 3446 size_t extra_cards = 0; 3447 JavaThread *curr = Threads::first(); 3448 while (curr != NULL) { 3449 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3450 extra_cards += dcq.size(); 3451 curr = curr->next(); 3452 } 3453 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3454 size_t buffer_size = dcqs.buffer_size(); 3455 size_t buffer_num = dcqs.completed_buffers_num(); 3456 return buffer_size * buffer_num + extra_cards; 3457 } 3458 3459 size_t G1CollectedHeap::max_pending_card_num() { 3460 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3461 size_t buffer_size = dcqs.buffer_size(); 3462 size_t buffer_num = dcqs.completed_buffers_num(); 3463 int thread_num = Threads::number_of_threads(); 3464 return (buffer_num + thread_num) * buffer_size; 3465 } 3466 3467 size_t G1CollectedHeap::cards_scanned() { 3468 return g1_rem_set()->cardsScanned(); 3469 } 3470 3471 void 3472 G1CollectedHeap::setup_surviving_young_words() { 3473 assert(_surviving_young_words == NULL, "pre-condition"); 3474 uint array_length = g1_policy()->young_cset_region_length(); 3475 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length); 3476 if (_surviving_young_words == NULL) { 3477 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3478 "Not enough space for young surv words summary."); 3479 } 3480 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3481 #ifdef ASSERT 3482 for (uint i = 0; i < array_length; ++i) { 3483 assert( _surviving_young_words[i] == 0, "memset above" ); 3484 } 3485 #endif // !ASSERT 3486 } 3487 3488 void 3489 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3490 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3491 uint array_length = g1_policy()->young_cset_region_length(); 3492 for (uint i = 0; i < array_length; ++i) { 3493 _surviving_young_words[i] += surv_young_words[i]; 3494 } 3495 } 3496 3497 void 3498 G1CollectedHeap::cleanup_surviving_young_words() { 3499 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3500 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words); 3501 _surviving_young_words = NULL; 3502 } 3503 3504 #ifdef ASSERT 3505 class VerifyCSetClosure: public HeapRegionClosure { 3506 public: 3507 bool doHeapRegion(HeapRegion* hr) { 3508 // Here we check that the CSet region's RSet is ready for parallel 3509 // iteration. The fields that we'll verify are only manipulated 3510 // when the region is part of a CSet and is collected. Afterwards, 3511 // we reset these fields when we clear the region's RSet (when the 3512 // region is freed) so they are ready when the region is 3513 // re-allocated. The only exception to this is if there's an 3514 // evacuation failure and instead of freeing the region we leave 3515 // it in the heap. In that case, we reset these fields during 3516 // evacuation failure handling. 3517 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3518 3519 // Here's a good place to add any other checks we'd like to 3520 // perform on CSet regions. 3521 return false; 3522 } 3523 }; 3524 #endif // ASSERT 3525 3526 #if TASKQUEUE_STATS 3527 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3528 st->print_raw_cr("GC Task Stats"); 3529 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3530 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3531 } 3532 3533 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3534 print_taskqueue_stats_hdr(st); 3535 3536 TaskQueueStats totals; 3537 const int n = workers() != NULL ? workers()->total_workers() : 1; 3538 for (int i = 0; i < n; ++i) { 3539 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3540 totals += task_queue(i)->stats; 3541 } 3542 st->print_raw("tot "); totals.print(st); st->cr(); 3543 3544 DEBUG_ONLY(totals.verify()); 3545 } 3546 3547 void G1CollectedHeap::reset_taskqueue_stats() { 3548 const int n = workers() != NULL ? workers()->total_workers() : 1; 3549 for (int i = 0; i < n; ++i) { 3550 task_queue(i)->stats.reset(); 3551 } 3552 } 3553 #endif // TASKQUEUE_STATS 3554 3555 bool 3556 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3557 assert_at_safepoint(true /* should_be_vm_thread */); 3558 guarantee(!is_gc_active(), "collection is not reentrant"); 3559 3560 if (GC_locker::check_active_before_gc()) { 3561 return false; 3562 } 3563 3564 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3565 ResourceMark rm; 3566 3567 print_heap_before_gc(); 3568 3569 HRSPhaseSetter x(HRSPhaseEvacuation); 3570 verify_region_sets_optional(); 3571 verify_dirty_young_regions(); 3572 3573 // This call will decide whether this pause is an initial-mark 3574 // pause. If it is, during_initial_mark_pause() will return true 3575 // for the duration of this pause. 3576 g1_policy()->decide_on_conc_mark_initiation(); 3577 3578 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3579 assert(!g1_policy()->during_initial_mark_pause() || 3580 g1_policy()->gcs_are_young(), "sanity"); 3581 3582 // We also do not allow mixed GCs during marking. 3583 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3584 3585 // Record whether this pause is an initial mark. When the current 3586 // thread has completed its logging output and it's safe to signal 3587 // the CM thread, the flag's value in the policy has been reset. 3588 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3589 3590 // Inner scope for scope based logging, timers, and stats collection 3591 { 3592 if (g1_policy()->during_initial_mark_pause()) { 3593 // We are about to start a marking cycle, so we increment the 3594 // full collection counter. 3595 increment_total_full_collections(); 3596 } 3597 // if the log level is "finer" is on, we'll print long statistics information 3598 // in the collector policy code, so let's not print this as the output 3599 // is messy if we do. 3600 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 3601 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3602 3603 char verbose_str[128]; 3604 sprintf(verbose_str, "GC pause (%s) (%s)%s", 3605 GCCause::to_string(gc_cause()), 3606 g1_policy()->gcs_are_young() ? "young" : "mixed", 3607 g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3608 TraceTime t(verbose_str, G1Log::fine() && !G1Log::finer(), true, gclog_or_tty); 3609 3610 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3611 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3612 3613 // If the secondary_free_list is not empty, append it to the 3614 // free_list. No need to wait for the cleanup operation to finish; 3615 // the region allocation code will check the secondary_free_list 3616 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3617 // set, skip this step so that the region allocation code has to 3618 // get entries from the secondary_free_list. 3619 if (!G1StressConcRegionFreeing) { 3620 append_secondary_free_list_if_not_empty_with_lock(); 3621 } 3622 3623 assert(check_young_list_well_formed(), 3624 "young list should be well formed"); 3625 3626 // Don't dynamically change the number of GC threads this early. A value of 3627 // 0 is used to indicate serial work. When parallel work is done, 3628 // it will be set. 3629 3630 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3631 IsGCActiveMark x; 3632 3633 gc_prologue(false); 3634 increment_total_collections(false /* full gc */); 3635 increment_gc_time_stamp(); 3636 3637 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3638 HandleMark hm; // Discard invalid handles created during verification 3639 gclog_or_tty->print(" VerifyBeforeGC:"); 3640 prepare_for_verify(); 3641 Universe::verify(/* silent */ false, 3642 /* option */ VerifyOption_G1UsePrevMarking); 3643 } 3644 3645 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3646 3647 // Please see comment in g1CollectedHeap.hpp and 3648 // G1CollectedHeap::ref_processing_init() to see how 3649 // reference processing currently works in G1. 3650 3651 // Enable discovery in the STW reference processor 3652 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3653 true /*verify_no_refs*/); 3654 3655 { 3656 // We want to temporarily turn off discovery by the 3657 // CM ref processor, if necessary, and turn it back on 3658 // on again later if we do. Using a scoped 3659 // NoRefDiscovery object will do this. 3660 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3661 3662 // Forget the current alloc region (we might even choose it to be part 3663 // of the collection set!). 3664 release_mutator_alloc_region(); 3665 3666 // We should call this after we retire the mutator alloc 3667 // region(s) so that all the ALLOC / RETIRE events are generated 3668 // before the start GC event. 3669 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3670 3671 // The elapsed time induced by the start time below deliberately elides 3672 // the possible verification above. 3673 double start_time_sec = os::elapsedTime(); 3674 size_t start_used_bytes = used(); 3675 3676 #if YOUNG_LIST_VERBOSE 3677 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3678 _young_list->print(); 3679 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3680 #endif // YOUNG_LIST_VERBOSE 3681 3682 g1_policy()->record_collection_pause_start(start_time_sec, 3683 start_used_bytes); 3684 3685 double scan_wait_start = os::elapsedTime(); 3686 // We have to wait until the CM threads finish scanning the 3687 // root regions as it's the only way to ensure that all the 3688 // objects on them have been correctly scanned before we start 3689 // moving them during the GC. 3690 bool waited = _cm->root_regions()->wait_until_scan_finished(); 3691 if (waited) { 3692 double scan_wait_end = os::elapsedTime(); 3693 double wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 3694 g1_policy()->record_root_region_scan_wait_time(wait_time_ms); 3695 } 3696 3697 #if YOUNG_LIST_VERBOSE 3698 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3699 _young_list->print(); 3700 #endif // YOUNG_LIST_VERBOSE 3701 3702 if (g1_policy()->during_initial_mark_pause()) { 3703 concurrent_mark()->checkpointRootsInitialPre(); 3704 } 3705 perm_gen()->save_marks(); 3706 3707 #if YOUNG_LIST_VERBOSE 3708 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3709 _young_list->print(); 3710 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3711 #endif // YOUNG_LIST_VERBOSE 3712 3713 g1_policy()->finalize_cset(target_pause_time_ms); 3714 3715 _cm->note_start_of_gc(); 3716 // We should not verify the per-thread SATB buffers given that 3717 // we have not filtered them yet (we'll do so during the 3718 // GC). We also call this after finalize_cset() to 3719 // ensure that the CSet has been finalized. 3720 _cm->verify_no_cset_oops(true /* verify_stacks */, 3721 true /* verify_enqueued_buffers */, 3722 false /* verify_thread_buffers */, 3723 true /* verify_fingers */); 3724 3725 if (_hr_printer.is_active()) { 3726 HeapRegion* hr = g1_policy()->collection_set(); 3727 while (hr != NULL) { 3728 G1HRPrinter::RegionType type; 3729 if (!hr->is_young()) { 3730 type = G1HRPrinter::Old; 3731 } else if (hr->is_survivor()) { 3732 type = G1HRPrinter::Survivor; 3733 } else { 3734 type = G1HRPrinter::Eden; 3735 } 3736 _hr_printer.cset(hr); 3737 hr = hr->next_in_collection_set(); 3738 } 3739 } 3740 3741 #ifdef ASSERT 3742 VerifyCSetClosure cl; 3743 collection_set_iterate(&cl); 3744 #endif // ASSERT 3745 3746 setup_surviving_young_words(); 3747 3748 // Initialize the GC alloc regions. 3749 init_gc_alloc_regions(); 3750 3751 // Actually do the work... 3752 evacuate_collection_set(); 3753 3754 // We do this to mainly verify the per-thread SATB buffers 3755 // (which have been filtered by now) since we didn't verify 3756 // them earlier. No point in re-checking the stacks / enqueued 3757 // buffers given that the CSet has not changed since last time 3758 // we checked. 3759 _cm->verify_no_cset_oops(false /* verify_stacks */, 3760 false /* verify_enqueued_buffers */, 3761 true /* verify_thread_buffers */, 3762 true /* verify_fingers */); 3763 3764 free_collection_set(g1_policy()->collection_set()); 3765 g1_policy()->clear_collection_set(); 3766 3767 cleanup_surviving_young_words(); 3768 3769 // Start a new incremental collection set for the next pause. 3770 g1_policy()->start_incremental_cset_building(); 3771 3772 // Clear the _cset_fast_test bitmap in anticipation of adding 3773 // regions to the incremental collection set for the next 3774 // evacuation pause. 3775 clear_cset_fast_test(); 3776 3777 _young_list->reset_sampled_info(); 3778 3779 // Don't check the whole heap at this point as the 3780 // GC alloc regions from this pause have been tagged 3781 // as survivors and moved on to the survivor list. 3782 // Survivor regions will fail the !is_young() check. 3783 assert(check_young_list_empty(false /* check_heap */), 3784 "young list should be empty"); 3785 3786 #if YOUNG_LIST_VERBOSE 3787 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3788 _young_list->print(); 3789 #endif // YOUNG_LIST_VERBOSE 3790 3791 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3792 _young_list->first_survivor_region(), 3793 _young_list->last_survivor_region()); 3794 3795 _young_list->reset_auxilary_lists(); 3796 3797 if (evacuation_failed()) { 3798 _summary_bytes_used = recalculate_used(); 3799 } else { 3800 // The "used" of the the collection set have already been subtracted 3801 // when they were freed. Add in the bytes evacuated. 3802 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3803 } 3804 3805 if (g1_policy()->during_initial_mark_pause()) { 3806 // We have to do this before we notify the CM threads that 3807 // they can start working to make sure that all the 3808 // appropriate initialization is done on the CM object. 3809 concurrent_mark()->checkpointRootsInitialPost(); 3810 set_marking_started(); 3811 // Note that we don't actually trigger the CM thread at 3812 // this point. We do that later when we're sure that 3813 // the current thread has completed its logging output. 3814 } 3815 3816 allocate_dummy_regions(); 3817 3818 #if YOUNG_LIST_VERBOSE 3819 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3820 _young_list->print(); 3821 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3822 #endif // YOUNG_LIST_VERBOSE 3823 3824 init_mutator_alloc_region(); 3825 3826 { 3827 size_t expand_bytes = g1_policy()->expansion_amount(); 3828 if (expand_bytes > 0) { 3829 size_t bytes_before = capacity(); 3830 // No need for an ergo verbose message here, 3831 // expansion_amount() does this when it returns a value > 0. 3832 if (!expand(expand_bytes)) { 3833 // We failed to expand the heap so let's verify that 3834 // committed/uncommitted amount match the backing store 3835 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3836 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3837 } 3838 } 3839 } 3840 3841 // We redo the verificaiton but now wrt to the new CSet which 3842 // has just got initialized after the previous CSet was freed. 3843 _cm->verify_no_cset_oops(true /* verify_stacks */, 3844 true /* verify_enqueued_buffers */, 3845 true /* verify_thread_buffers */, 3846 true /* verify_fingers */); 3847 _cm->note_end_of_gc(); 3848 3849 double end_time_sec = os::elapsedTime(); 3850 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS; 3851 g1_policy()->record_pause_time_ms(pause_time_ms); 3852 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3853 workers()->active_workers() : 1); 3854 g1_policy()->record_collection_pause_end(active_workers); 3855 3856 MemoryService::track_memory_usage(); 3857 3858 // In prepare_for_verify() below we'll need to scan the deferred 3859 // update buffers to bring the RSets up-to-date if 3860 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3861 // the update buffers we'll probably need to scan cards on the 3862 // regions we just allocated to (i.e., the GC alloc 3863 // regions). However, during the last GC we called 3864 // set_saved_mark() on all the GC alloc regions, so card 3865 // scanning might skip the [saved_mark_word()...top()] area of 3866 // those regions (i.e., the area we allocated objects into 3867 // during the last GC). But it shouldn't. Given that 3868 // saved_mark_word() is conditional on whether the GC time stamp 3869 // on the region is current or not, by incrementing the GC time 3870 // stamp here we invalidate all the GC time stamps on all the 3871 // regions and saved_mark_word() will simply return top() for 3872 // all the regions. This is a nicer way of ensuring this rather 3873 // than iterating over the regions and fixing them. In fact, the 3874 // GC time stamp increment here also ensures that 3875 // saved_mark_word() will return top() between pauses, i.e., 3876 // during concurrent refinement. So we don't need the 3877 // is_gc_active() check to decided which top to use when 3878 // scanning cards (see CR 7039627). 3879 increment_gc_time_stamp(); 3880 3881 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 3882 HandleMark hm; // Discard invalid handles created during verification 3883 gclog_or_tty->print(" VerifyAfterGC:"); 3884 prepare_for_verify(); 3885 Universe::verify(/* silent */ false, 3886 /* option */ VerifyOption_G1UsePrevMarking); 3887 } 3888 3889 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 3890 ref_processor_stw()->verify_no_references_recorded(); 3891 3892 // CM reference discovery will be re-enabled if necessary. 3893 } 3894 3895 // We should do this after we potentially expand the heap so 3896 // that all the COMMIT events are generated before the end GC 3897 // event, and after we retire the GC alloc regions so that all 3898 // RETIRE events are generated before the end GC event. 3899 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 3900 3901 // We have to do this after we decide whether to expand the heap or not. 3902 g1_policy()->print_heap_transition(); 3903 3904 if (mark_in_progress()) { 3905 concurrent_mark()->update_g1_committed(); 3906 } 3907 3908 #ifdef TRACESPINNING 3909 ParallelTaskTerminator::print_termination_counts(); 3910 #endif 3911 3912 gc_epilogue(false); 3913 } 3914 3915 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) { 3916 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum); 3917 print_tracing_info(); 3918 vm_exit(-1); 3919 } 3920 } 3921 3922 // The closing of the inner scope, immediately above, will complete 3923 // logging at the "fine" level. The record_collection_pause_end() call 3924 // above will complete logging at the "finer" level. 3925 // 3926 // It is not yet to safe, however, to tell the concurrent mark to 3927 // start as we have some optional output below. We don't want the 3928 // output from the concurrent mark thread interfering with this 3929 // logging output either. 3930 3931 _hrs.verify_optional(); 3932 verify_region_sets_optional(); 3933 3934 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 3935 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 3936 3937 print_heap_after_gc(); 3938 g1mm()->update_sizes(); 3939 3940 if (G1SummarizeRSetStats && 3941 (G1SummarizeRSetStatsPeriod > 0) && 3942 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3943 g1_rem_set()->print_summary_info(); 3944 } 3945 3946 // It should now be safe to tell the concurrent mark thread to start 3947 // without its logging output interfering with the logging output 3948 // that came from the pause. 3949 3950 if (should_start_conc_mark) { 3951 // CAUTION: after the doConcurrentMark() call below, 3952 // the concurrent marking thread(s) could be running 3953 // concurrently with us. Make sure that anything after 3954 // this point does not assume that we are the only GC thread 3955 // running. Note: of course, the actual marking work will 3956 // not start until the safepoint itself is released in 3957 // ConcurrentGCThread::safepoint_desynchronize(). 3958 doConcurrentMark(); 3959 } 3960 3961 return true; 3962 } 3963 3964 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 3965 { 3966 size_t gclab_word_size; 3967 switch (purpose) { 3968 case GCAllocForSurvived: 3969 gclab_word_size = YoungPLABSize; 3970 break; 3971 case GCAllocForTenured: 3972 gclab_word_size = OldPLABSize; 3973 break; 3974 default: 3975 assert(false, "unknown GCAllocPurpose"); 3976 gclab_word_size = OldPLABSize; 3977 break; 3978 } 3979 return gclab_word_size; 3980 } 3981 3982 void G1CollectedHeap::init_mutator_alloc_region() { 3983 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 3984 _mutator_alloc_region.init(); 3985 } 3986 3987 void G1CollectedHeap::release_mutator_alloc_region() { 3988 _mutator_alloc_region.release(); 3989 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 3990 } 3991 3992 void G1CollectedHeap::init_gc_alloc_regions() { 3993 assert_at_safepoint(true /* should_be_vm_thread */); 3994 3995 _survivor_gc_alloc_region.init(); 3996 _old_gc_alloc_region.init(); 3997 HeapRegion* retained_region = _retained_old_gc_alloc_region; 3998 _retained_old_gc_alloc_region = NULL; 3999 4000 // We will discard the current GC alloc region if: 4001 // a) it's in the collection set (it can happen!), 4002 // b) it's already full (no point in using it), 4003 // c) it's empty (this means that it was emptied during 4004 // a cleanup and it should be on the free list now), or 4005 // d) it's humongous (this means that it was emptied 4006 // during a cleanup and was added to the free list, but 4007 // has been subseqently used to allocate a humongous 4008 // object that may be less than the region size). 4009 if (retained_region != NULL && 4010 !retained_region->in_collection_set() && 4011 !(retained_region->top() == retained_region->end()) && 4012 !retained_region->is_empty() && 4013 !retained_region->isHumongous()) { 4014 retained_region->set_saved_mark(); 4015 // The retained region was added to the old region set when it was 4016 // retired. We have to remove it now, since we don't allow regions 4017 // we allocate to in the region sets. We'll re-add it later, when 4018 // it's retired again. 4019 _old_set.remove(retained_region); 4020 bool during_im = g1_policy()->during_initial_mark_pause(); 4021 retained_region->note_start_of_copying(during_im); 4022 _old_gc_alloc_region.set(retained_region); 4023 _hr_printer.reuse(retained_region); 4024 } 4025 } 4026 4027 void G1CollectedHeap::release_gc_alloc_regions() { 4028 _survivor_gc_alloc_region.release(); 4029 // If we have an old GC alloc region to release, we'll save it in 4030 // _retained_old_gc_alloc_region. If we don't 4031 // _retained_old_gc_alloc_region will become NULL. This is what we 4032 // want either way so no reason to check explicitly for either 4033 // condition. 4034 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4035 } 4036 4037 void G1CollectedHeap::abandon_gc_alloc_regions() { 4038 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4039 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4040 _retained_old_gc_alloc_region = NULL; 4041 } 4042 4043 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4044 _drain_in_progress = false; 4045 set_evac_failure_closure(cl); 4046 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4047 } 4048 4049 void G1CollectedHeap::finalize_for_evac_failure() { 4050 assert(_evac_failure_scan_stack != NULL && 4051 _evac_failure_scan_stack->length() == 0, 4052 "Postcondition"); 4053 assert(!_drain_in_progress, "Postcondition"); 4054 delete _evac_failure_scan_stack; 4055 _evac_failure_scan_stack = NULL; 4056 } 4057 4058 void G1CollectedHeap::remove_self_forwarding_pointers() { 4059 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4060 4061 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4062 4063 if (G1CollectedHeap::use_parallel_gc_threads()) { 4064 set_par_threads(); 4065 workers()->run_task(&rsfp_task); 4066 set_par_threads(0); 4067 } else { 4068 rsfp_task.work(0); 4069 } 4070 4071 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4072 4073 // Reset the claim values in the regions in the collection set. 4074 reset_cset_heap_region_claim_values(); 4075 4076 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4077 4078 // Now restore saved marks, if any. 4079 if (_objs_with_preserved_marks != NULL) { 4080 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4081 guarantee(_objs_with_preserved_marks->length() == 4082 _preserved_marks_of_objs->length(), "Both or none."); 4083 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4084 oop obj = _objs_with_preserved_marks->at(i); 4085 markOop m = _preserved_marks_of_objs->at(i); 4086 obj->set_mark(m); 4087 } 4088 4089 // Delete the preserved marks growable arrays (allocated on the C heap). 4090 delete _objs_with_preserved_marks; 4091 delete _preserved_marks_of_objs; 4092 _objs_with_preserved_marks = NULL; 4093 _preserved_marks_of_objs = NULL; 4094 } 4095 } 4096 4097 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4098 _evac_failure_scan_stack->push(obj); 4099 } 4100 4101 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4102 assert(_evac_failure_scan_stack != NULL, "precondition"); 4103 4104 while (_evac_failure_scan_stack->length() > 0) { 4105 oop obj = _evac_failure_scan_stack->pop(); 4106 _evac_failure_closure->set_region(heap_region_containing(obj)); 4107 obj->oop_iterate_backwards(_evac_failure_closure); 4108 } 4109 } 4110 4111 oop 4112 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4113 oop old) { 4114 assert(obj_in_cs(old), 4115 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4116 (HeapWord*) old)); 4117 markOop m = old->mark(); 4118 oop forward_ptr = old->forward_to_atomic(old); 4119 if (forward_ptr == NULL) { 4120 // Forward-to-self succeeded. 4121 4122 if (_evac_failure_closure != cl) { 4123 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4124 assert(!_drain_in_progress, 4125 "Should only be true while someone holds the lock."); 4126 // Set the global evac-failure closure to the current thread's. 4127 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4128 set_evac_failure_closure(cl); 4129 // Now do the common part. 4130 handle_evacuation_failure_common(old, m); 4131 // Reset to NULL. 4132 set_evac_failure_closure(NULL); 4133 } else { 4134 // The lock is already held, and this is recursive. 4135 assert(_drain_in_progress, "This should only be the recursive case."); 4136 handle_evacuation_failure_common(old, m); 4137 } 4138 return old; 4139 } else { 4140 // Forward-to-self failed. Either someone else managed to allocate 4141 // space for this object (old != forward_ptr) or they beat us in 4142 // self-forwarding it (old == forward_ptr). 4143 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4144 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4145 "should not be in the CSet", 4146 (HeapWord*) old, (HeapWord*) forward_ptr)); 4147 return forward_ptr; 4148 } 4149 } 4150 4151 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4152 set_evacuation_failed(true); 4153 4154 preserve_mark_if_necessary(old, m); 4155 4156 HeapRegion* r = heap_region_containing(old); 4157 if (!r->evacuation_failed()) { 4158 r->set_evacuation_failed(true); 4159 _hr_printer.evac_failure(r); 4160 } 4161 4162 push_on_evac_failure_scan_stack(old); 4163 4164 if (!_drain_in_progress) { 4165 // prevent recursion in copy_to_survivor_space() 4166 _drain_in_progress = true; 4167 drain_evac_failure_scan_stack(); 4168 _drain_in_progress = false; 4169 } 4170 } 4171 4172 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4173 assert(evacuation_failed(), "Oversaving!"); 4174 // We want to call the "for_promotion_failure" version only in the 4175 // case of a promotion failure. 4176 if (m->must_be_preserved_for_promotion_failure(obj)) { 4177 if (_objs_with_preserved_marks == NULL) { 4178 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4179 _objs_with_preserved_marks = 4180 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4181 _preserved_marks_of_objs = 4182 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true); 4183 } 4184 _objs_with_preserved_marks->push(obj); 4185 _preserved_marks_of_objs->push(m); 4186 } 4187 } 4188 4189 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4190 size_t word_size) { 4191 if (purpose == GCAllocForSurvived) { 4192 HeapWord* result = survivor_attempt_allocation(word_size); 4193 if (result != NULL) { 4194 return result; 4195 } else { 4196 // Let's try to allocate in the old gen in case we can fit the 4197 // object there. 4198 return old_attempt_allocation(word_size); 4199 } 4200 } else { 4201 assert(purpose == GCAllocForTenured, "sanity"); 4202 HeapWord* result = old_attempt_allocation(word_size); 4203 if (result != NULL) { 4204 return result; 4205 } else { 4206 // Let's try to allocate in the survivors in case we can fit the 4207 // object there. 4208 return survivor_attempt_allocation(word_size); 4209 } 4210 } 4211 4212 ShouldNotReachHere(); 4213 // Trying to keep some compilers happy. 4214 return NULL; 4215 } 4216 4217 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4218 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4219 4220 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4221 : _g1h(g1h), 4222 _refs(g1h->task_queue(queue_num)), 4223 _dcq(&g1h->dirty_card_queue_set()), 4224 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4225 _g1_rem(g1h->g1_rem_set()), 4226 _hash_seed(17), _queue_num(queue_num), 4227 _term_attempts(0), 4228 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4229 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4230 _age_table(false), 4231 _strong_roots_time(0), _term_time(0), 4232 _alloc_buffer_waste(0), _undo_waste(0) { 4233 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4234 // we "sacrifice" entry 0 to keep track of surviving bytes for 4235 // non-young regions (where the age is -1) 4236 // We also add a few elements at the beginning and at the end in 4237 // an attempt to eliminate cache contention 4238 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4239 uint array_length = PADDING_ELEM_NUM + 4240 real_length + 4241 PADDING_ELEM_NUM; 4242 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length); 4243 if (_surviving_young_words_base == NULL) 4244 vm_exit_out_of_memory(array_length * sizeof(size_t), 4245 "Not enough space for young surv histo."); 4246 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4247 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4248 4249 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4250 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4251 4252 _start = os::elapsedTime(); 4253 } 4254 4255 void 4256 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4257 { 4258 st->print_raw_cr("GC Termination Stats"); 4259 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4260 " ------waste (KiB)------"); 4261 st->print_raw_cr("thr ms ms % ms % attempts" 4262 " total alloc undo"); 4263 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4264 " ------- ------- -------"); 4265 } 4266 4267 void 4268 G1ParScanThreadState::print_termination_stats(int i, 4269 outputStream* const st) const 4270 { 4271 const double elapsed_ms = elapsed_time() * 1000.0; 4272 const double s_roots_ms = strong_roots_time() * 1000.0; 4273 const double term_ms = term_time() * 1000.0; 4274 st->print_cr("%3d %9.2f %9.2f %6.2f " 4275 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4276 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4277 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4278 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4279 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4280 alloc_buffer_waste() * HeapWordSize / K, 4281 undo_waste() * HeapWordSize / K); 4282 } 4283 4284 #ifdef ASSERT 4285 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4286 assert(ref != NULL, "invariant"); 4287 assert(UseCompressedOops, "sanity"); 4288 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4289 oop p = oopDesc::load_decode_heap_oop(ref); 4290 assert(_g1h->is_in_g1_reserved(p), 4291 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4292 return true; 4293 } 4294 4295 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4296 assert(ref != NULL, "invariant"); 4297 if (has_partial_array_mask(ref)) { 4298 // Must be in the collection set--it's already been copied. 4299 oop p = clear_partial_array_mask(ref); 4300 assert(_g1h->obj_in_cs(p), 4301 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4302 } else { 4303 oop p = oopDesc::load_decode_heap_oop(ref); 4304 assert(_g1h->is_in_g1_reserved(p), 4305 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4306 } 4307 return true; 4308 } 4309 4310 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4311 if (ref.is_narrow()) { 4312 return verify_ref((narrowOop*) ref); 4313 } else { 4314 return verify_ref((oop*) ref); 4315 } 4316 } 4317 #endif // ASSERT 4318 4319 void G1ParScanThreadState::trim_queue() { 4320 assert(_evac_cl != NULL, "not set"); 4321 assert(_evac_failure_cl != NULL, "not set"); 4322 assert(_partial_scan_cl != NULL, "not set"); 4323 4324 StarTask ref; 4325 do { 4326 // Drain the overflow stack first, so other threads can steal. 4327 while (refs()->pop_overflow(ref)) { 4328 deal_with_reference(ref); 4329 } 4330 4331 while (refs()->pop_local(ref)) { 4332 deal_with_reference(ref); 4333 } 4334 } while (!refs()->is_empty()); 4335 } 4336 4337 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4338 G1ParScanThreadState* par_scan_state) : 4339 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4340 _par_scan_state(par_scan_state), 4341 _worker_id(par_scan_state->queue_num()), 4342 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4343 _mark_in_progress(_g1->mark_in_progress()) { } 4344 4345 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4346 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4347 #ifdef ASSERT 4348 HeapRegion* hr = _g1->heap_region_containing(obj); 4349 assert(hr != NULL, "sanity"); 4350 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4351 #endif // ASSERT 4352 4353 // We know that the object is not moving so it's safe to read its size. 4354 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4355 } 4356 4357 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4358 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4359 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4360 #ifdef ASSERT 4361 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4362 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4363 assert(from_obj != to_obj, "should not be self-forwarded"); 4364 4365 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4366 assert(from_hr != NULL, "sanity"); 4367 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4368 4369 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4370 assert(to_hr != NULL, "sanity"); 4371 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4372 #endif // ASSERT 4373 4374 // The object might be in the process of being copied by another 4375 // worker so we cannot trust that its to-space image is 4376 // well-formed. So we have to read its size from its from-space 4377 // image which we know should not be changing. 4378 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4379 } 4380 4381 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4382 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4383 ::copy_to_survivor_space(oop old) { 4384 size_t word_sz = old->size(); 4385 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4386 // +1 to make the -1 indexes valid... 4387 int young_index = from_region->young_index_in_cset()+1; 4388 assert( (from_region->is_young() && young_index > 0) || 4389 (!from_region->is_young() && young_index == 0), "invariant" ); 4390 G1CollectorPolicy* g1p = _g1->g1_policy(); 4391 markOop m = old->mark(); 4392 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4393 : m->age(); 4394 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4395 word_sz); 4396 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4397 oop obj = oop(obj_ptr); 4398 4399 if (obj_ptr == NULL) { 4400 // This will either forward-to-self, or detect that someone else has 4401 // installed a forwarding pointer. 4402 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4403 return _g1->handle_evacuation_failure_par(cl, old); 4404 } 4405 4406 // We're going to allocate linearly, so might as well prefetch ahead. 4407 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4408 4409 oop forward_ptr = old->forward_to_atomic(obj); 4410 if (forward_ptr == NULL) { 4411 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4412 if (g1p->track_object_age(alloc_purpose)) { 4413 // We could simply do obj->incr_age(). However, this causes a 4414 // performance issue. obj->incr_age() will first check whether 4415 // the object has a displaced mark by checking its mark word; 4416 // getting the mark word from the new location of the object 4417 // stalls. So, given that we already have the mark word and we 4418 // are about to install it anyway, it's better to increase the 4419 // age on the mark word, when the object does not have a 4420 // displaced mark word. We're not expecting many objects to have 4421 // a displaced marked word, so that case is not optimized 4422 // further (it could be...) and we simply call obj->incr_age(). 4423 4424 if (m->has_displaced_mark_helper()) { 4425 // in this case, we have to install the mark word first, 4426 // otherwise obj looks to be forwarded (the old mark word, 4427 // which contains the forward pointer, was copied) 4428 obj->set_mark(m); 4429 obj->incr_age(); 4430 } else { 4431 m = m->incr_age(); 4432 obj->set_mark(m); 4433 } 4434 _par_scan_state->age_table()->add(obj, word_sz); 4435 } else { 4436 obj->set_mark(m); 4437 } 4438 4439 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4440 surv_young_words[young_index] += word_sz; 4441 4442 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4443 // We keep track of the next start index in the length field of 4444 // the to-space object. The actual length can be found in the 4445 // length field of the from-space object. 4446 arrayOop(obj)->set_length(0); 4447 oop* old_p = set_partial_array_mask(old); 4448 _par_scan_state->push_on_queue(old_p); 4449 } else { 4450 // No point in using the slower heap_region_containing() method, 4451 // given that we know obj is in the heap. 4452 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4453 obj->oop_iterate_backwards(&_scanner); 4454 } 4455 } else { 4456 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4457 obj = forward_ptr; 4458 } 4459 return obj; 4460 } 4461 4462 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4463 template <class T> 4464 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4465 ::do_oop_work(T* p) { 4466 oop obj = oopDesc::load_decode_heap_oop(p); 4467 assert(barrier != G1BarrierRS || obj != NULL, 4468 "Precondition: G1BarrierRS implies obj is non-NULL"); 4469 4470 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4471 4472 // here the null check is implicit in the cset_fast_test() test 4473 if (_g1->in_cset_fast_test(obj)) { 4474 oop forwardee; 4475 if (obj->is_forwarded()) { 4476 forwardee = obj->forwardee(); 4477 } else { 4478 forwardee = copy_to_survivor_space(obj); 4479 } 4480 assert(forwardee != NULL, "forwardee should not be NULL"); 4481 oopDesc::encode_store_heap_oop(p, forwardee); 4482 if (do_mark_object && forwardee != obj) { 4483 // If the object is self-forwarded we don't need to explicitly 4484 // mark it, the evacuation failure protocol will do so. 4485 mark_forwarded_object(obj, forwardee); 4486 } 4487 4488 // When scanning the RS, we only care about objs in CS. 4489 if (barrier == G1BarrierRS) { 4490 _par_scan_state->update_rs(_from, p, _worker_id); 4491 } 4492 } else { 4493 // The object is not in collection set. If we're a root scanning 4494 // closure during an initial mark pause (i.e. do_mark_object will 4495 // be true) then attempt to mark the object. 4496 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4497 mark_object(obj); 4498 } 4499 } 4500 4501 if (barrier == G1BarrierEvac && obj != NULL) { 4502 _par_scan_state->update_rs(_from, p, _worker_id); 4503 } 4504 4505 if (do_gen_barrier && obj != NULL) { 4506 par_do_barrier(p); 4507 } 4508 } 4509 4510 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4511 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4512 4513 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4514 assert(has_partial_array_mask(p), "invariant"); 4515 oop from_obj = clear_partial_array_mask(p); 4516 4517 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4518 assert(from_obj->is_objArray(), "must be obj array"); 4519 objArrayOop from_obj_array = objArrayOop(from_obj); 4520 // The from-space object contains the real length. 4521 int length = from_obj_array->length(); 4522 4523 assert(from_obj->is_forwarded(), "must be forwarded"); 4524 oop to_obj = from_obj->forwardee(); 4525 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4526 objArrayOop to_obj_array = objArrayOop(to_obj); 4527 // We keep track of the next start index in the length field of the 4528 // to-space object. 4529 int next_index = to_obj_array->length(); 4530 assert(0 <= next_index && next_index < length, 4531 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4532 4533 int start = next_index; 4534 int end = length; 4535 int remainder = end - start; 4536 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4537 if (remainder > 2 * ParGCArrayScanChunk) { 4538 end = start + ParGCArrayScanChunk; 4539 to_obj_array->set_length(end); 4540 // Push the remainder before we process the range in case another 4541 // worker has run out of things to do and can steal it. 4542 oop* from_obj_p = set_partial_array_mask(from_obj); 4543 _par_scan_state->push_on_queue(from_obj_p); 4544 } else { 4545 assert(length == end, "sanity"); 4546 // We'll process the final range for this object. Restore the length 4547 // so that the heap remains parsable in case of evacuation failure. 4548 to_obj_array->set_length(end); 4549 } 4550 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4551 // Process indexes [start,end). It will also process the header 4552 // along with the first chunk (i.e., the chunk with start == 0). 4553 // Note that at this point the length field of to_obj_array is not 4554 // correct given that we are using it to keep track of the next 4555 // start index. oop_iterate_range() (thankfully!) ignores the length 4556 // field and only relies on the start / end parameters. It does 4557 // however return the size of the object which will be incorrect. So 4558 // we have to ignore it even if we wanted to use it. 4559 to_obj_array->oop_iterate_range(&_scanner, start, end); 4560 } 4561 4562 class G1ParEvacuateFollowersClosure : public VoidClosure { 4563 protected: 4564 G1CollectedHeap* _g1h; 4565 G1ParScanThreadState* _par_scan_state; 4566 RefToScanQueueSet* _queues; 4567 ParallelTaskTerminator* _terminator; 4568 4569 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4570 RefToScanQueueSet* queues() { return _queues; } 4571 ParallelTaskTerminator* terminator() { return _terminator; } 4572 4573 public: 4574 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4575 G1ParScanThreadState* par_scan_state, 4576 RefToScanQueueSet* queues, 4577 ParallelTaskTerminator* terminator) 4578 : _g1h(g1h), _par_scan_state(par_scan_state), 4579 _queues(queues), _terminator(terminator) {} 4580 4581 void do_void(); 4582 4583 private: 4584 inline bool offer_termination(); 4585 }; 4586 4587 bool G1ParEvacuateFollowersClosure::offer_termination() { 4588 G1ParScanThreadState* const pss = par_scan_state(); 4589 pss->start_term_time(); 4590 const bool res = terminator()->offer_termination(); 4591 pss->end_term_time(); 4592 return res; 4593 } 4594 4595 void G1ParEvacuateFollowersClosure::do_void() { 4596 StarTask stolen_task; 4597 G1ParScanThreadState* const pss = par_scan_state(); 4598 pss->trim_queue(); 4599 4600 do { 4601 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4602 assert(pss->verify_task(stolen_task), "sanity"); 4603 if (stolen_task.is_narrow()) { 4604 pss->deal_with_reference((narrowOop*) stolen_task); 4605 } else { 4606 pss->deal_with_reference((oop*) stolen_task); 4607 } 4608 4609 // We've just processed a reference and we might have made 4610 // available new entries on the queues. So we have to make sure 4611 // we drain the queues as necessary. 4612 pss->trim_queue(); 4613 } 4614 } while (!offer_termination()); 4615 4616 pss->retire_alloc_buffers(); 4617 } 4618 4619 class G1ParTask : public AbstractGangTask { 4620 protected: 4621 G1CollectedHeap* _g1h; 4622 RefToScanQueueSet *_queues; 4623 ParallelTaskTerminator _terminator; 4624 uint _n_workers; 4625 4626 Mutex _stats_lock; 4627 Mutex* stats_lock() { return &_stats_lock; } 4628 4629 size_t getNCards() { 4630 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4631 / G1BlockOffsetSharedArray::N_bytes; 4632 } 4633 4634 public: 4635 G1ParTask(G1CollectedHeap* g1h, 4636 RefToScanQueueSet *task_queues) 4637 : AbstractGangTask("G1 collection"), 4638 _g1h(g1h), 4639 _queues(task_queues), 4640 _terminator(0, _queues), 4641 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4642 {} 4643 4644 RefToScanQueueSet* queues() { return _queues; } 4645 4646 RefToScanQueue *work_queue(int i) { 4647 return queues()->queue(i); 4648 } 4649 4650 ParallelTaskTerminator* terminator() { return &_terminator; } 4651 4652 virtual void set_for_termination(int active_workers) { 4653 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4654 // in the young space (_par_seq_tasks) in the G1 heap 4655 // for SequentialSubTasksDone. 4656 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4657 // both of which need setting by set_n_termination(). 4658 _g1h->SharedHeap::set_n_termination(active_workers); 4659 _g1h->set_n_termination(active_workers); 4660 terminator()->reset_for_reuse(active_workers); 4661 _n_workers = active_workers; 4662 } 4663 4664 void work(uint worker_id) { 4665 if (worker_id >= _n_workers) return; // no work needed this round 4666 4667 double start_time_ms = os::elapsedTime() * 1000.0; 4668 _g1h->g1_policy()->record_gc_worker_start_time(worker_id, start_time_ms); 4669 4670 { 4671 ResourceMark rm; 4672 HandleMark hm; 4673 4674 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4675 4676 G1ParScanThreadState pss(_g1h, worker_id); 4677 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4678 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4679 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4680 4681 pss.set_evac_closure(&scan_evac_cl); 4682 pss.set_evac_failure_closure(&evac_failure_cl); 4683 pss.set_partial_scan_closure(&partial_scan_cl); 4684 4685 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4686 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4687 4688 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4689 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4690 4691 OopClosure* scan_root_cl = &only_scan_root_cl; 4692 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4693 4694 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4695 // We also need to mark copied objects. 4696 scan_root_cl = &scan_mark_root_cl; 4697 scan_perm_cl = &scan_mark_perm_cl; 4698 } 4699 4700 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4701 4702 pss.start_strong_roots(); 4703 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4704 SharedHeap::SO_AllClasses, 4705 scan_root_cl, 4706 &push_heap_rs_cl, 4707 scan_perm_cl, 4708 worker_id); 4709 pss.end_strong_roots(); 4710 4711 { 4712 double start = os::elapsedTime(); 4713 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4714 evac.do_void(); 4715 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4716 double term_ms = pss.term_time()*1000.0; 4717 _g1h->g1_policy()->record_obj_copy_time(worker_id, elapsed_ms-term_ms); 4718 _g1h->g1_policy()->record_termination(worker_id, term_ms, pss.term_attempts()); 4719 } 4720 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4721 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4722 4723 // Clean up any par-expanded rem sets. 4724 HeapRegionRemSet::par_cleanup(); 4725 4726 if (ParallelGCVerbose) { 4727 MutexLocker x(stats_lock()); 4728 pss.print_termination_stats(worker_id); 4729 } 4730 4731 assert(pss.refs()->is_empty(), "should be empty"); 4732 4733 // Close the inner scope so that the ResourceMark and HandleMark 4734 // destructors are executed here and are included as part of the 4735 // "GC Worker Time". 4736 } 4737 4738 double end_time_ms = os::elapsedTime() * 1000.0; 4739 _g1h->g1_policy()->record_gc_worker_end_time(worker_id, end_time_ms); 4740 } 4741 }; 4742 4743 // *** Common G1 Evacuation Stuff 4744 4745 // Closures that support the filtering of CodeBlobs scanned during 4746 // external root scanning. 4747 4748 // Closure applied to reference fields in code blobs (specifically nmethods) 4749 // to determine whether an nmethod contains references that point into 4750 // the collection set. Used as a predicate when walking code roots so 4751 // that only nmethods that point into the collection set are added to the 4752 // 'marked' list. 4753 4754 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure { 4755 4756 class G1PointsIntoCSOopClosure : public OopClosure { 4757 G1CollectedHeap* _g1; 4758 bool _points_into_cs; 4759 public: 4760 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) : 4761 _g1(g1), _points_into_cs(false) { } 4762 4763 bool points_into_cs() const { return _points_into_cs; } 4764 4765 template <class T> 4766 void do_oop_nv(T* p) { 4767 if (!_points_into_cs) { 4768 T heap_oop = oopDesc::load_heap_oop(p); 4769 if (!oopDesc::is_null(heap_oop) && 4770 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) { 4771 _points_into_cs = true; 4772 } 4773 } 4774 } 4775 4776 virtual void do_oop(oop* p) { do_oop_nv(p); } 4777 virtual void do_oop(narrowOop* p) { do_oop_nv(p); } 4778 }; 4779 4780 G1CollectedHeap* _g1; 4781 4782 public: 4783 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) : 4784 CodeBlobToOopClosure(cl, true), _g1(g1) { } 4785 4786 virtual void do_code_blob(CodeBlob* cb) { 4787 nmethod* nm = cb->as_nmethod_or_null(); 4788 if (nm != NULL && !(nm->test_oops_do_mark())) { 4789 G1PointsIntoCSOopClosure predicate_cl(_g1); 4790 nm->oops_do(&predicate_cl); 4791 4792 if (predicate_cl.points_into_cs()) { 4793 // At least one of the reference fields or the oop relocations 4794 // in the nmethod points into the collection set. We have to 4795 // 'mark' this nmethod. 4796 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob() 4797 // or MarkingCodeBlobClosure::do_code_blob() change. 4798 if (!nm->test_set_oops_do_mark()) { 4799 do_newly_marked_nmethod(nm); 4800 } 4801 } 4802 } 4803 } 4804 }; 4805 4806 // This method is run in a GC worker. 4807 4808 void 4809 G1CollectedHeap:: 4810 g1_process_strong_roots(bool collecting_perm_gen, 4811 ScanningOption so, 4812 OopClosure* scan_non_heap_roots, 4813 OopsInHeapRegionClosure* scan_rs, 4814 OopsInGenClosure* scan_perm, 4815 int worker_i) { 4816 4817 // First scan the strong roots, including the perm gen. 4818 double ext_roots_start = os::elapsedTime(); 4819 double closure_app_time_sec = 0.0; 4820 4821 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4822 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4823 buf_scan_perm.set_generation(perm_gen()); 4824 4825 // Walk the code cache w/o buffering, because StarTask cannot handle 4826 // unaligned oop locations. 4827 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots); 4828 4829 process_strong_roots(false, // no scoping; this is parallel code 4830 collecting_perm_gen, so, 4831 &buf_scan_non_heap_roots, 4832 &eager_scan_code_roots, 4833 &buf_scan_perm); 4834 4835 // Now the CM ref_processor roots. 4836 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4837 // We need to treat the discovered reference lists of the 4838 // concurrent mark ref processor as roots and keep entries 4839 // (which are added by the marking threads) on them live 4840 // until they can be processed at the end of marking. 4841 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4842 } 4843 4844 // Finish up any enqueued closure apps (attributed as object copy time). 4845 buf_scan_non_heap_roots.done(); 4846 buf_scan_perm.done(); 4847 4848 double ext_roots_end = os::elapsedTime(); 4849 4850 g1_policy()->reset_obj_copy_time(worker_i); 4851 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4852 buf_scan_non_heap_roots.closure_app_seconds(); 4853 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4854 4855 double ext_root_time_ms = 4856 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4857 4858 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4859 4860 // During conc marking we have to filter the per-thread SATB buffers 4861 // to make sure we remove any oops into the CSet (which will show up 4862 // as implicitly live). 4863 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 4864 if (mark_in_progress()) { 4865 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 4866 } 4867 } 4868 double satb_filtering_ms = (os::elapsedTime() - ext_roots_end) * 1000.0; 4869 g1_policy()->record_satb_filtering_time(worker_i, satb_filtering_ms); 4870 4871 // Now scan the complement of the collection set. 4872 if (scan_rs != NULL) { 4873 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4874 } 4875 4876 _process_strong_tasks->all_tasks_completed(); 4877 } 4878 4879 void 4880 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 4881 OopClosure* non_root_closure) { 4882 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 4883 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 4884 } 4885 4886 // Weak Reference Processing support 4887 4888 // An always "is_alive" closure that is used to preserve referents. 4889 // If the object is non-null then it's alive. Used in the preservation 4890 // of referent objects that are pointed to by reference objects 4891 // discovered by the CM ref processor. 4892 class G1AlwaysAliveClosure: public BoolObjectClosure { 4893 G1CollectedHeap* _g1; 4894 public: 4895 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4896 void do_object(oop p) { assert(false, "Do not call."); } 4897 bool do_object_b(oop p) { 4898 if (p != NULL) { 4899 return true; 4900 } 4901 return false; 4902 } 4903 }; 4904 4905 bool G1STWIsAliveClosure::do_object_b(oop p) { 4906 // An object is reachable if it is outside the collection set, 4907 // or is inside and copied. 4908 return !_g1->obj_in_cs(p) || p->is_forwarded(); 4909 } 4910 4911 // Non Copying Keep Alive closure 4912 class G1KeepAliveClosure: public OopClosure { 4913 G1CollectedHeap* _g1; 4914 public: 4915 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4916 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 4917 void do_oop( oop* p) { 4918 oop obj = *p; 4919 4920 if (_g1->obj_in_cs(obj)) { 4921 assert( obj->is_forwarded(), "invariant" ); 4922 *p = obj->forwardee(); 4923 } 4924 } 4925 }; 4926 4927 // Copying Keep Alive closure - can be called from both 4928 // serial and parallel code as long as different worker 4929 // threads utilize different G1ParScanThreadState instances 4930 // and different queues. 4931 4932 class G1CopyingKeepAliveClosure: public OopClosure { 4933 G1CollectedHeap* _g1h; 4934 OopClosure* _copy_non_heap_obj_cl; 4935 OopsInHeapRegionClosure* _copy_perm_obj_cl; 4936 G1ParScanThreadState* _par_scan_state; 4937 4938 public: 4939 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 4940 OopClosure* non_heap_obj_cl, 4941 OopsInHeapRegionClosure* perm_obj_cl, 4942 G1ParScanThreadState* pss): 4943 _g1h(g1h), 4944 _copy_non_heap_obj_cl(non_heap_obj_cl), 4945 _copy_perm_obj_cl(perm_obj_cl), 4946 _par_scan_state(pss) 4947 {} 4948 4949 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 4950 virtual void do_oop( oop* p) { do_oop_work(p); } 4951 4952 template <class T> void do_oop_work(T* p) { 4953 oop obj = oopDesc::load_decode_heap_oop(p); 4954 4955 if (_g1h->obj_in_cs(obj)) { 4956 // If the referent object has been forwarded (either copied 4957 // to a new location or to itself in the event of an 4958 // evacuation failure) then we need to update the reference 4959 // field and, if both reference and referent are in the G1 4960 // heap, update the RSet for the referent. 4961 // 4962 // If the referent has not been forwarded then we have to keep 4963 // it alive by policy. Therefore we have copy the referent. 4964 // 4965 // If the reference field is in the G1 heap then we can push 4966 // on the PSS queue. When the queue is drained (after each 4967 // phase of reference processing) the object and it's followers 4968 // will be copied, the reference field set to point to the 4969 // new location, and the RSet updated. Otherwise we need to 4970 // use the the non-heap or perm closures directly to copy 4971 // the refernt object and update the pointer, while avoiding 4972 // updating the RSet. 4973 4974 if (_g1h->is_in_g1_reserved(p)) { 4975 _par_scan_state->push_on_queue(p); 4976 } else { 4977 // The reference field is not in the G1 heap. 4978 if (_g1h->perm_gen()->is_in(p)) { 4979 _copy_perm_obj_cl->do_oop(p); 4980 } else { 4981 _copy_non_heap_obj_cl->do_oop(p); 4982 } 4983 } 4984 } 4985 } 4986 }; 4987 4988 // Serial drain queue closure. Called as the 'complete_gc' 4989 // closure for each discovered list in some of the 4990 // reference processing phases. 4991 4992 class G1STWDrainQueueClosure: public VoidClosure { 4993 protected: 4994 G1CollectedHeap* _g1h; 4995 G1ParScanThreadState* _par_scan_state; 4996 4997 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4998 4999 public: 5000 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5001 _g1h(g1h), 5002 _par_scan_state(pss) 5003 { } 5004 5005 void do_void() { 5006 G1ParScanThreadState* const pss = par_scan_state(); 5007 pss->trim_queue(); 5008 } 5009 }; 5010 5011 // Parallel Reference Processing closures 5012 5013 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5014 // processing during G1 evacuation pauses. 5015 5016 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5017 private: 5018 G1CollectedHeap* _g1h; 5019 RefToScanQueueSet* _queues; 5020 FlexibleWorkGang* _workers; 5021 int _active_workers; 5022 5023 public: 5024 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5025 FlexibleWorkGang* workers, 5026 RefToScanQueueSet *task_queues, 5027 int n_workers) : 5028 _g1h(g1h), 5029 _queues(task_queues), 5030 _workers(workers), 5031 _active_workers(n_workers) 5032 { 5033 assert(n_workers > 0, "shouldn't call this otherwise"); 5034 } 5035 5036 // Executes the given task using concurrent marking worker threads. 5037 virtual void execute(ProcessTask& task); 5038 virtual void execute(EnqueueTask& task); 5039 }; 5040 5041 // Gang task for possibly parallel reference processing 5042 5043 class G1STWRefProcTaskProxy: public AbstractGangTask { 5044 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5045 ProcessTask& _proc_task; 5046 G1CollectedHeap* _g1h; 5047 RefToScanQueueSet *_task_queues; 5048 ParallelTaskTerminator* _terminator; 5049 5050 public: 5051 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5052 G1CollectedHeap* g1h, 5053 RefToScanQueueSet *task_queues, 5054 ParallelTaskTerminator* terminator) : 5055 AbstractGangTask("Process reference objects in parallel"), 5056 _proc_task(proc_task), 5057 _g1h(g1h), 5058 _task_queues(task_queues), 5059 _terminator(terminator) 5060 {} 5061 5062 virtual void work(uint worker_id) { 5063 // The reference processing task executed by a single worker. 5064 ResourceMark rm; 5065 HandleMark hm; 5066 5067 G1STWIsAliveClosure is_alive(_g1h); 5068 5069 G1ParScanThreadState pss(_g1h, worker_id); 5070 5071 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5072 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5073 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5074 5075 pss.set_evac_closure(&scan_evac_cl); 5076 pss.set_evac_failure_closure(&evac_failure_cl); 5077 pss.set_partial_scan_closure(&partial_scan_cl); 5078 5079 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5080 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5081 5082 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5083 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5084 5085 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5086 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5087 5088 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5089 // We also need to mark copied objects. 5090 copy_non_heap_cl = ©_mark_non_heap_cl; 5091 copy_perm_cl = ©_mark_perm_cl; 5092 } 5093 5094 // Keep alive closure. 5095 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5096 5097 // Complete GC closure 5098 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5099 5100 // Call the reference processing task's work routine. 5101 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5102 5103 // Note we cannot assert that the refs array is empty here as not all 5104 // of the processing tasks (specifically phase2 - pp2_work) execute 5105 // the complete_gc closure (which ordinarily would drain the queue) so 5106 // the queue may not be empty. 5107 } 5108 }; 5109 5110 // Driver routine for parallel reference processing. 5111 // Creates an instance of the ref processing gang 5112 // task and has the worker threads execute it. 5113 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5114 assert(_workers != NULL, "Need parallel worker threads."); 5115 5116 ParallelTaskTerminator terminator(_active_workers, _queues); 5117 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5118 5119 _g1h->set_par_threads(_active_workers); 5120 _workers->run_task(&proc_task_proxy); 5121 _g1h->set_par_threads(0); 5122 } 5123 5124 // Gang task for parallel reference enqueueing. 5125 5126 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5127 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5128 EnqueueTask& _enq_task; 5129 5130 public: 5131 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5132 AbstractGangTask("Enqueue reference objects in parallel"), 5133 _enq_task(enq_task) 5134 { } 5135 5136 virtual void work(uint worker_id) { 5137 _enq_task.work(worker_id); 5138 } 5139 }; 5140 5141 // Driver routine for parallel reference enqueing. 5142 // Creates an instance of the ref enqueueing gang 5143 // task and has the worker threads execute it. 5144 5145 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5146 assert(_workers != NULL, "Need parallel worker threads."); 5147 5148 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5149 5150 _g1h->set_par_threads(_active_workers); 5151 _workers->run_task(&enq_task_proxy); 5152 _g1h->set_par_threads(0); 5153 } 5154 5155 // End of weak reference support closures 5156 5157 // Abstract task used to preserve (i.e. copy) any referent objects 5158 // that are in the collection set and are pointed to by reference 5159 // objects discovered by the CM ref processor. 5160 5161 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5162 protected: 5163 G1CollectedHeap* _g1h; 5164 RefToScanQueueSet *_queues; 5165 ParallelTaskTerminator _terminator; 5166 uint _n_workers; 5167 5168 public: 5169 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5170 AbstractGangTask("ParPreserveCMReferents"), 5171 _g1h(g1h), 5172 _queues(task_queues), 5173 _terminator(workers, _queues), 5174 _n_workers(workers) 5175 { } 5176 5177 void work(uint worker_id) { 5178 ResourceMark rm; 5179 HandleMark hm; 5180 5181 G1ParScanThreadState pss(_g1h, worker_id); 5182 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5183 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5184 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5185 5186 pss.set_evac_closure(&scan_evac_cl); 5187 pss.set_evac_failure_closure(&evac_failure_cl); 5188 pss.set_partial_scan_closure(&partial_scan_cl); 5189 5190 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5191 5192 5193 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5194 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5195 5196 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5197 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5198 5199 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5200 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5201 5202 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5203 // We also need to mark copied objects. 5204 copy_non_heap_cl = ©_mark_non_heap_cl; 5205 copy_perm_cl = ©_mark_perm_cl; 5206 } 5207 5208 // Is alive closure 5209 G1AlwaysAliveClosure always_alive(_g1h); 5210 5211 // Copying keep alive closure. Applied to referent objects that need 5212 // to be copied. 5213 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5214 5215 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5216 5217 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5218 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5219 5220 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5221 // So this must be true - but assert just in case someone decides to 5222 // change the worker ids. 5223 assert(0 <= worker_id && worker_id < limit, "sanity"); 5224 assert(!rp->discovery_is_atomic(), "check this code"); 5225 5226 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5227 for (uint idx = worker_id; idx < limit; idx += stride) { 5228 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5229 5230 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5231 while (iter.has_next()) { 5232 // Since discovery is not atomic for the CM ref processor, we 5233 // can see some null referent objects. 5234 iter.load_ptrs(DEBUG_ONLY(true)); 5235 oop ref = iter.obj(); 5236 5237 // This will filter nulls. 5238 if (iter.is_referent_alive()) { 5239 iter.make_referent_alive(); 5240 } 5241 iter.move_to_next(); 5242 } 5243 } 5244 5245 // Drain the queue - which may cause stealing 5246 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5247 drain_queue.do_void(); 5248 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5249 assert(pss.refs()->is_empty(), "should be"); 5250 } 5251 }; 5252 5253 // Weak Reference processing during an evacuation pause (part 1). 5254 void G1CollectedHeap::process_discovered_references() { 5255 double ref_proc_start = os::elapsedTime(); 5256 5257 ReferenceProcessor* rp = _ref_processor_stw; 5258 assert(rp->discovery_enabled(), "should have been enabled"); 5259 5260 // Any reference objects, in the collection set, that were 'discovered' 5261 // by the CM ref processor should have already been copied (either by 5262 // applying the external root copy closure to the discovered lists, or 5263 // by following an RSet entry). 5264 // 5265 // But some of the referents, that are in the collection set, that these 5266 // reference objects point to may not have been copied: the STW ref 5267 // processor would have seen that the reference object had already 5268 // been 'discovered' and would have skipped discovering the reference, 5269 // but would not have treated the reference object as a regular oop. 5270 // As a reult the copy closure would not have been applied to the 5271 // referent object. 5272 // 5273 // We need to explicitly copy these referent objects - the references 5274 // will be processed at the end of remarking. 5275 // 5276 // We also need to do this copying before we process the reference 5277 // objects discovered by the STW ref processor in case one of these 5278 // referents points to another object which is also referenced by an 5279 // object discovered by the STW ref processor. 5280 5281 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5282 workers()->active_workers() : 1); 5283 5284 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5285 active_workers == workers()->active_workers(), 5286 "Need to reset active_workers"); 5287 5288 set_par_threads(active_workers); 5289 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5290 5291 if (G1CollectedHeap::use_parallel_gc_threads()) { 5292 workers()->run_task(&keep_cm_referents); 5293 } else { 5294 keep_cm_referents.work(0); 5295 } 5296 5297 set_par_threads(0); 5298 5299 // Closure to test whether a referent is alive. 5300 G1STWIsAliveClosure is_alive(this); 5301 5302 // Even when parallel reference processing is enabled, the processing 5303 // of JNI refs is serial and performed serially by the current thread 5304 // rather than by a worker. The following PSS will be used for processing 5305 // JNI refs. 5306 5307 // Use only a single queue for this PSS. 5308 G1ParScanThreadState pss(this, 0); 5309 5310 // We do not embed a reference processor in the copying/scanning 5311 // closures while we're actually processing the discovered 5312 // reference objects. 5313 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5314 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5315 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5316 5317 pss.set_evac_closure(&scan_evac_cl); 5318 pss.set_evac_failure_closure(&evac_failure_cl); 5319 pss.set_partial_scan_closure(&partial_scan_cl); 5320 5321 assert(pss.refs()->is_empty(), "pre-condition"); 5322 5323 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5324 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5325 5326 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5327 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5328 5329 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5330 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5331 5332 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5333 // We also need to mark copied objects. 5334 copy_non_heap_cl = ©_mark_non_heap_cl; 5335 copy_perm_cl = ©_mark_perm_cl; 5336 } 5337 5338 // Keep alive closure. 5339 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5340 5341 // Serial Complete GC closure 5342 G1STWDrainQueueClosure drain_queue(this, &pss); 5343 5344 // Setup the soft refs policy... 5345 rp->setup_policy(false); 5346 5347 if (!rp->processing_is_mt()) { 5348 // Serial reference processing... 5349 rp->process_discovered_references(&is_alive, 5350 &keep_alive, 5351 &drain_queue, 5352 NULL); 5353 } else { 5354 // Parallel reference processing 5355 assert(rp->num_q() == active_workers, "sanity"); 5356 assert(active_workers <= rp->max_num_q(), "sanity"); 5357 5358 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5359 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5360 } 5361 5362 // We have completed copying any necessary live referent objects 5363 // (that were not copied during the actual pause) so we can 5364 // retire any active alloc buffers 5365 pss.retire_alloc_buffers(); 5366 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5367 5368 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5369 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0); 5370 } 5371 5372 // Weak Reference processing during an evacuation pause (part 2). 5373 void G1CollectedHeap::enqueue_discovered_references() { 5374 double ref_enq_start = os::elapsedTime(); 5375 5376 ReferenceProcessor* rp = _ref_processor_stw; 5377 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5378 5379 // Now enqueue any remaining on the discovered lists on to 5380 // the pending list. 5381 if (!rp->processing_is_mt()) { 5382 // Serial reference processing... 5383 rp->enqueue_discovered_references(); 5384 } else { 5385 // Parallel reference enqueuing 5386 5387 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5388 assert(active_workers == workers()->active_workers(), 5389 "Need to reset active_workers"); 5390 assert(rp->num_q() == active_workers, "sanity"); 5391 assert(active_workers <= rp->max_num_q(), "sanity"); 5392 5393 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5394 rp->enqueue_discovered_references(&par_task_executor); 5395 } 5396 5397 rp->verify_no_references_recorded(); 5398 assert(!rp->discovery_enabled(), "should have been disabled"); 5399 5400 // FIXME 5401 // CM's reference processing also cleans up the string and symbol tables. 5402 // Should we do that here also? We could, but it is a serial operation 5403 // and could signicantly increase the pause time. 5404 5405 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5406 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0); 5407 } 5408 5409 void G1CollectedHeap::evacuate_collection_set() { 5410 _expand_heap_after_alloc_failure = true; 5411 set_evacuation_failed(false); 5412 5413 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5414 concurrent_g1_refine()->set_use_cache(false); 5415 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5416 5417 uint n_workers; 5418 if (G1CollectedHeap::use_parallel_gc_threads()) { 5419 n_workers = 5420 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5421 workers()->active_workers(), 5422 Threads::number_of_non_daemon_threads()); 5423 assert(UseDynamicNumberOfGCThreads || 5424 n_workers == workers()->total_workers(), 5425 "If not dynamic should be using all the workers"); 5426 workers()->set_active_workers(n_workers); 5427 set_par_threads(n_workers); 5428 } else { 5429 assert(n_par_threads() == 0, 5430 "Should be the original non-parallel value"); 5431 n_workers = 1; 5432 } 5433 5434 G1ParTask g1_par_task(this, _task_queues); 5435 5436 init_for_evac_failure(NULL); 5437 5438 rem_set()->prepare_for_younger_refs_iterate(true); 5439 5440 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5441 double start_par_time_sec = os::elapsedTime(); 5442 double end_par_time_sec; 5443 5444 { 5445 StrongRootsScope srs(this); 5446 5447 if (G1CollectedHeap::use_parallel_gc_threads()) { 5448 // The individual threads will set their evac-failure closures. 5449 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5450 // These tasks use ShareHeap::_process_strong_tasks 5451 assert(UseDynamicNumberOfGCThreads || 5452 workers()->active_workers() == workers()->total_workers(), 5453 "If not dynamic should be using all the workers"); 5454 workers()->run_task(&g1_par_task); 5455 } else { 5456 g1_par_task.set_for_termination(n_workers); 5457 g1_par_task.work(0); 5458 } 5459 end_par_time_sec = os::elapsedTime(); 5460 5461 // Closing the inner scope will execute the destructor 5462 // for the StrongRootsScope object. We record the current 5463 // elapsed time before closing the scope so that time 5464 // taken for the SRS destructor is NOT included in the 5465 // reported parallel time. 5466 } 5467 5468 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5469 g1_policy()->record_par_time(par_time_ms); 5470 5471 double code_root_fixup_time_ms = 5472 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5473 g1_policy()->record_code_root_fixup_time(code_root_fixup_time_ms); 5474 5475 set_par_threads(0); 5476 5477 // Process any discovered reference objects - we have 5478 // to do this _before_ we retire the GC alloc regions 5479 // as we may have to copy some 'reachable' referent 5480 // objects (and their reachable sub-graphs) that were 5481 // not copied during the pause. 5482 process_discovered_references(); 5483 5484 // Weak root processing. 5485 // Note: when JSR 292 is enabled and code blobs can contain 5486 // non-perm oops then we will need to process the code blobs 5487 // here too. 5488 { 5489 G1STWIsAliveClosure is_alive(this); 5490 G1KeepAliveClosure keep_alive(this); 5491 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5492 } 5493 5494 release_gc_alloc_regions(); 5495 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5496 5497 concurrent_g1_refine()->clear_hot_cache(); 5498 concurrent_g1_refine()->set_use_cache(true); 5499 5500 finalize_for_evac_failure(); 5501 5502 if (evacuation_failed()) { 5503 remove_self_forwarding_pointers(); 5504 if (G1Log::finer()) { 5505 gclog_or_tty->print(" (to-space overflow)"); 5506 } else if (G1Log::fine()) { 5507 gclog_or_tty->print("--"); 5508 } 5509 } 5510 5511 // Enqueue any remaining references remaining on the STW 5512 // reference processor's discovered lists. We need to do 5513 // this after the card table is cleaned (and verified) as 5514 // the act of enqueuing entries on to the pending list 5515 // will log these updates (and dirty their associated 5516 // cards). We need these updates logged to update any 5517 // RSets. 5518 enqueue_discovered_references(); 5519 5520 if (G1DeferredRSUpdate) { 5521 RedirtyLoggedCardTableEntryFastClosure redirty; 5522 dirty_card_queue_set().set_closure(&redirty); 5523 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5524 5525 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5526 dcq.merge_bufferlists(&dirty_card_queue_set()); 5527 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5528 } 5529 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5530 } 5531 5532 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5533 size_t* pre_used, 5534 FreeRegionList* free_list, 5535 OldRegionSet* old_proxy_set, 5536 HumongousRegionSet* humongous_proxy_set, 5537 HRRSCleanupTask* hrrs_cleanup_task, 5538 bool par) { 5539 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5540 if (hr->isHumongous()) { 5541 assert(hr->startsHumongous(), "we should only see starts humongous"); 5542 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5543 } else { 5544 _old_set.remove_with_proxy(hr, old_proxy_set); 5545 free_region(hr, pre_used, free_list, par); 5546 } 5547 } else { 5548 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5549 } 5550 } 5551 5552 void G1CollectedHeap::free_region(HeapRegion* hr, 5553 size_t* pre_used, 5554 FreeRegionList* free_list, 5555 bool par) { 5556 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5557 assert(!hr->is_empty(), "the region should not be empty"); 5558 assert(free_list != NULL, "pre-condition"); 5559 5560 *pre_used += hr->used(); 5561 hr->hr_clear(par, true /* clear_space */); 5562 free_list->add_as_head(hr); 5563 } 5564 5565 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5566 size_t* pre_used, 5567 FreeRegionList* free_list, 5568 HumongousRegionSet* humongous_proxy_set, 5569 bool par) { 5570 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5571 assert(free_list != NULL, "pre-condition"); 5572 assert(humongous_proxy_set != NULL, "pre-condition"); 5573 5574 size_t hr_used = hr->used(); 5575 size_t hr_capacity = hr->capacity(); 5576 size_t hr_pre_used = 0; 5577 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5578 hr->set_notHumongous(); 5579 free_region(hr, &hr_pre_used, free_list, par); 5580 5581 uint i = hr->hrs_index() + 1; 5582 uint num = 1; 5583 while (i < n_regions()) { 5584 HeapRegion* curr_hr = region_at(i); 5585 if (!curr_hr->continuesHumongous()) { 5586 break; 5587 } 5588 curr_hr->set_notHumongous(); 5589 free_region(curr_hr, &hr_pre_used, free_list, par); 5590 num += 1; 5591 i += 1; 5592 } 5593 assert(hr_pre_used == hr_used, 5594 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5595 "should be the same", hr_pre_used, hr_used)); 5596 *pre_used += hr_pre_used; 5597 } 5598 5599 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5600 FreeRegionList* free_list, 5601 OldRegionSet* old_proxy_set, 5602 HumongousRegionSet* humongous_proxy_set, 5603 bool par) { 5604 if (pre_used > 0) { 5605 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5606 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5607 assert(_summary_bytes_used >= pre_used, 5608 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5609 "should be >= pre_used: "SIZE_FORMAT, 5610 _summary_bytes_used, pre_used)); 5611 _summary_bytes_used -= pre_used; 5612 } 5613 if (free_list != NULL && !free_list->is_empty()) { 5614 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5615 _free_list.add_as_head(free_list); 5616 } 5617 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5618 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5619 _old_set.update_from_proxy(old_proxy_set); 5620 } 5621 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5622 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5623 _humongous_set.update_from_proxy(humongous_proxy_set); 5624 } 5625 } 5626 5627 class G1ParCleanupCTTask : public AbstractGangTask { 5628 CardTableModRefBS* _ct_bs; 5629 G1CollectedHeap* _g1h; 5630 HeapRegion* volatile _su_head; 5631 public: 5632 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5633 G1CollectedHeap* g1h) : 5634 AbstractGangTask("G1 Par Cleanup CT Task"), 5635 _ct_bs(ct_bs), _g1h(g1h) { } 5636 5637 void work(uint worker_id) { 5638 HeapRegion* r; 5639 while (r = _g1h->pop_dirty_cards_region()) { 5640 clear_cards(r); 5641 } 5642 } 5643 5644 void clear_cards(HeapRegion* r) { 5645 // Cards of the survivors should have already been dirtied. 5646 if (!r->is_survivor()) { 5647 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5648 } 5649 } 5650 }; 5651 5652 #ifndef PRODUCT 5653 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5654 G1CollectedHeap* _g1h; 5655 CardTableModRefBS* _ct_bs; 5656 public: 5657 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5658 : _g1h(g1h), _ct_bs(ct_bs) { } 5659 virtual bool doHeapRegion(HeapRegion* r) { 5660 if (r->is_survivor()) { 5661 _g1h->verify_dirty_region(r); 5662 } else { 5663 _g1h->verify_not_dirty_region(r); 5664 } 5665 return false; 5666 } 5667 }; 5668 5669 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5670 // All of the region should be clean. 5671 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5672 MemRegion mr(hr->bottom(), hr->end()); 5673 ct_bs->verify_not_dirty_region(mr); 5674 } 5675 5676 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5677 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5678 // dirty allocated blocks as they allocate them. The thread that 5679 // retires each region and replaces it with a new one will do a 5680 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5681 // not dirty that area (one less thing to have to do while holding 5682 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5683 // is dirty. 5684 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5685 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5686 ct_bs->verify_dirty_region(mr); 5687 } 5688 5689 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5690 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5691 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5692 verify_dirty_region(hr); 5693 } 5694 } 5695 5696 void G1CollectedHeap::verify_dirty_young_regions() { 5697 verify_dirty_young_list(_young_list->first_region()); 5698 verify_dirty_young_list(_young_list->first_survivor_region()); 5699 } 5700 #endif 5701 5702 void G1CollectedHeap::cleanUpCardTable() { 5703 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5704 double start = os::elapsedTime(); 5705 5706 { 5707 // Iterate over the dirty cards region list. 5708 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5709 5710 if (G1CollectedHeap::use_parallel_gc_threads()) { 5711 set_par_threads(); 5712 workers()->run_task(&cleanup_task); 5713 set_par_threads(0); 5714 } else { 5715 while (_dirty_cards_region_list) { 5716 HeapRegion* r = _dirty_cards_region_list; 5717 cleanup_task.clear_cards(r); 5718 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5719 if (_dirty_cards_region_list == r) { 5720 // The last region. 5721 _dirty_cards_region_list = NULL; 5722 } 5723 r->set_next_dirty_cards_region(NULL); 5724 } 5725 } 5726 #ifndef PRODUCT 5727 if (G1VerifyCTCleanup || VerifyAfterGC) { 5728 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5729 heap_region_iterate(&cleanup_verifier); 5730 } 5731 #endif 5732 } 5733 5734 double elapsed = os::elapsedTime() - start; 5735 g1_policy()->record_clear_ct_time(elapsed * 1000.0); 5736 } 5737 5738 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5739 size_t pre_used = 0; 5740 FreeRegionList local_free_list("Local List for CSet Freeing"); 5741 5742 double young_time_ms = 0.0; 5743 double non_young_time_ms = 0.0; 5744 5745 // Since the collection set is a superset of the the young list, 5746 // all we need to do to clear the young list is clear its 5747 // head and length, and unlink any young regions in the code below 5748 _young_list->clear(); 5749 5750 G1CollectorPolicy* policy = g1_policy(); 5751 5752 double start_sec = os::elapsedTime(); 5753 bool non_young = true; 5754 5755 HeapRegion* cur = cs_head; 5756 int age_bound = -1; 5757 size_t rs_lengths = 0; 5758 5759 while (cur != NULL) { 5760 assert(!is_on_master_free_list(cur), "sanity"); 5761 if (non_young) { 5762 if (cur->is_young()) { 5763 double end_sec = os::elapsedTime(); 5764 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5765 non_young_time_ms += elapsed_ms; 5766 5767 start_sec = os::elapsedTime(); 5768 non_young = false; 5769 } 5770 } else { 5771 if (!cur->is_young()) { 5772 double end_sec = os::elapsedTime(); 5773 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5774 young_time_ms += elapsed_ms; 5775 5776 start_sec = os::elapsedTime(); 5777 non_young = true; 5778 } 5779 } 5780 5781 rs_lengths += cur->rem_set()->occupied(); 5782 5783 HeapRegion* next = cur->next_in_collection_set(); 5784 assert(cur->in_collection_set(), "bad CS"); 5785 cur->set_next_in_collection_set(NULL); 5786 cur->set_in_collection_set(false); 5787 5788 if (cur->is_young()) { 5789 int index = cur->young_index_in_cset(); 5790 assert(index != -1, "invariant"); 5791 assert((uint) index < policy->young_cset_region_length(), "invariant"); 5792 size_t words_survived = _surviving_young_words[index]; 5793 cur->record_surv_words_in_group(words_survived); 5794 5795 // At this point the we have 'popped' cur from the collection set 5796 // (linked via next_in_collection_set()) but it is still in the 5797 // young list (linked via next_young_region()). Clear the 5798 // _next_young_region field. 5799 cur->set_next_young_region(NULL); 5800 } else { 5801 int index = cur->young_index_in_cset(); 5802 assert(index == -1, "invariant"); 5803 } 5804 5805 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5806 (!cur->is_young() && cur->young_index_in_cset() == -1), 5807 "invariant" ); 5808 5809 if (!cur->evacuation_failed()) { 5810 MemRegion used_mr = cur->used_region(); 5811 5812 // And the region is empty. 5813 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5814 free_region(cur, &pre_used, &local_free_list, false /* par */); 5815 } else { 5816 cur->uninstall_surv_rate_group(); 5817 if (cur->is_young()) { 5818 cur->set_young_index_in_cset(-1); 5819 } 5820 cur->set_not_young(); 5821 cur->set_evacuation_failed(false); 5822 // The region is now considered to be old. 5823 _old_set.add(cur); 5824 } 5825 cur = next; 5826 } 5827 5828 policy->record_max_rs_lengths(rs_lengths); 5829 policy->cset_regions_freed(); 5830 5831 double end_sec = os::elapsedTime(); 5832 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5833 5834 if (non_young) { 5835 non_young_time_ms += elapsed_ms; 5836 } else { 5837 young_time_ms += elapsed_ms; 5838 } 5839 5840 update_sets_after_freeing_regions(pre_used, &local_free_list, 5841 NULL /* old_proxy_set */, 5842 NULL /* humongous_proxy_set */, 5843 false /* par */); 5844 policy->record_young_free_cset_time_ms(young_time_ms); 5845 policy->record_non_young_free_cset_time_ms(non_young_time_ms); 5846 } 5847 5848 // This routine is similar to the above but does not record 5849 // any policy statistics or update free lists; we are abandoning 5850 // the current incremental collection set in preparation of a 5851 // full collection. After the full GC we will start to build up 5852 // the incremental collection set again. 5853 // This is only called when we're doing a full collection 5854 // and is immediately followed by the tearing down of the young list. 5855 5856 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5857 HeapRegion* cur = cs_head; 5858 5859 while (cur != NULL) { 5860 HeapRegion* next = cur->next_in_collection_set(); 5861 assert(cur->in_collection_set(), "bad CS"); 5862 cur->set_next_in_collection_set(NULL); 5863 cur->set_in_collection_set(false); 5864 cur->set_young_index_in_cset(-1); 5865 cur = next; 5866 } 5867 } 5868 5869 void G1CollectedHeap::set_free_regions_coming() { 5870 if (G1ConcRegionFreeingVerbose) { 5871 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5872 "setting free regions coming"); 5873 } 5874 5875 assert(!free_regions_coming(), "pre-condition"); 5876 _free_regions_coming = true; 5877 } 5878 5879 void G1CollectedHeap::reset_free_regions_coming() { 5880 assert(free_regions_coming(), "pre-condition"); 5881 5882 { 5883 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5884 _free_regions_coming = false; 5885 SecondaryFreeList_lock->notify_all(); 5886 } 5887 5888 if (G1ConcRegionFreeingVerbose) { 5889 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5890 "reset free regions coming"); 5891 } 5892 } 5893 5894 void G1CollectedHeap::wait_while_free_regions_coming() { 5895 // Most of the time we won't have to wait, so let's do a quick test 5896 // first before we take the lock. 5897 if (!free_regions_coming()) { 5898 return; 5899 } 5900 5901 if (G1ConcRegionFreeingVerbose) { 5902 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5903 "waiting for free regions"); 5904 } 5905 5906 { 5907 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5908 while (free_regions_coming()) { 5909 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 5910 } 5911 } 5912 5913 if (G1ConcRegionFreeingVerbose) { 5914 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5915 "done waiting for free regions"); 5916 } 5917 } 5918 5919 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 5920 assert(heap_lock_held_for_gc(), 5921 "the heap lock should already be held by or for this thread"); 5922 _young_list->push_region(hr); 5923 } 5924 5925 class NoYoungRegionsClosure: public HeapRegionClosure { 5926 private: 5927 bool _success; 5928 public: 5929 NoYoungRegionsClosure() : _success(true) { } 5930 bool doHeapRegion(HeapRegion* r) { 5931 if (r->is_young()) { 5932 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 5933 r->bottom(), r->end()); 5934 _success = false; 5935 } 5936 return false; 5937 } 5938 bool success() { return _success; } 5939 }; 5940 5941 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 5942 bool ret = _young_list->check_list_empty(check_sample); 5943 5944 if (check_heap) { 5945 NoYoungRegionsClosure closure; 5946 heap_region_iterate(&closure); 5947 ret = ret && closure.success(); 5948 } 5949 5950 return ret; 5951 } 5952 5953 class TearDownRegionSetsClosure : public HeapRegionClosure { 5954 private: 5955 OldRegionSet *_old_set; 5956 5957 public: 5958 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 5959 5960 bool doHeapRegion(HeapRegion* r) { 5961 if (r->is_empty()) { 5962 // We ignore empty regions, we'll empty the free list afterwards 5963 } else if (r->is_young()) { 5964 // We ignore young regions, we'll empty the young list afterwards 5965 } else if (r->isHumongous()) { 5966 // We ignore humongous regions, we're not tearing down the 5967 // humongous region set 5968 } else { 5969 // The rest should be old 5970 _old_set->remove(r); 5971 } 5972 return false; 5973 } 5974 5975 ~TearDownRegionSetsClosure() { 5976 assert(_old_set->is_empty(), "post-condition"); 5977 } 5978 }; 5979 5980 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 5981 assert_at_safepoint(true /* should_be_vm_thread */); 5982 5983 if (!free_list_only) { 5984 TearDownRegionSetsClosure cl(&_old_set); 5985 heap_region_iterate(&cl); 5986 5987 // Need to do this after the heap iteration to be able to 5988 // recognize the young regions and ignore them during the iteration. 5989 _young_list->empty_list(); 5990 } 5991 _free_list.remove_all(); 5992 } 5993 5994 class RebuildRegionSetsClosure : public HeapRegionClosure { 5995 private: 5996 bool _free_list_only; 5997 OldRegionSet* _old_set; 5998 FreeRegionList* _free_list; 5999 size_t _total_used; 6000 6001 public: 6002 RebuildRegionSetsClosure(bool free_list_only, 6003 OldRegionSet* old_set, FreeRegionList* free_list) : 6004 _free_list_only(free_list_only), 6005 _old_set(old_set), _free_list(free_list), _total_used(0) { 6006 assert(_free_list->is_empty(), "pre-condition"); 6007 if (!free_list_only) { 6008 assert(_old_set->is_empty(), "pre-condition"); 6009 } 6010 } 6011 6012 bool doHeapRegion(HeapRegion* r) { 6013 if (r->continuesHumongous()) { 6014 return false; 6015 } 6016 6017 if (r->is_empty()) { 6018 // Add free regions to the free list 6019 _free_list->add_as_tail(r); 6020 } else if (!_free_list_only) { 6021 assert(!r->is_young(), "we should not come across young regions"); 6022 6023 if (r->isHumongous()) { 6024 // We ignore humongous regions, we left the humongous set unchanged 6025 } else { 6026 // The rest should be old, add them to the old set 6027 _old_set->add(r); 6028 } 6029 _total_used += r->used(); 6030 } 6031 6032 return false; 6033 } 6034 6035 size_t total_used() { 6036 return _total_used; 6037 } 6038 }; 6039 6040 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6041 assert_at_safepoint(true /* should_be_vm_thread */); 6042 6043 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6044 heap_region_iterate(&cl); 6045 6046 if (!free_list_only) { 6047 _summary_bytes_used = cl.total_used(); 6048 } 6049 assert(_summary_bytes_used == recalculate_used(), 6050 err_msg("inconsistent _summary_bytes_used, " 6051 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6052 _summary_bytes_used, recalculate_used())); 6053 } 6054 6055 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6056 _refine_cte_cl->set_concurrent(concurrent); 6057 } 6058 6059 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6060 HeapRegion* hr = heap_region_containing(p); 6061 if (hr == NULL) { 6062 return is_in_permanent(p); 6063 } else { 6064 return hr->is_in(p); 6065 } 6066 } 6067 6068 // Methods for the mutator alloc region 6069 6070 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6071 bool force) { 6072 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6073 assert(!force || g1_policy()->can_expand_young_list(), 6074 "if force is true we should be able to expand the young list"); 6075 bool young_list_full = g1_policy()->is_young_list_full(); 6076 if (force || !young_list_full) { 6077 HeapRegion* new_alloc_region = new_region(word_size, 6078 false /* do_expand */); 6079 if (new_alloc_region != NULL) { 6080 set_region_short_lived_locked(new_alloc_region); 6081 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6082 return new_alloc_region; 6083 } 6084 } 6085 return NULL; 6086 } 6087 6088 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6089 size_t allocated_bytes) { 6090 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6091 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6092 6093 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6094 _summary_bytes_used += allocated_bytes; 6095 _hr_printer.retire(alloc_region); 6096 // We update the eden sizes here, when the region is retired, 6097 // instead of when it's allocated, since this is the point that its 6098 // used space has been recored in _summary_bytes_used. 6099 g1mm()->update_eden_size(); 6100 } 6101 6102 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6103 bool force) { 6104 return _g1h->new_mutator_alloc_region(word_size, force); 6105 } 6106 6107 void G1CollectedHeap::set_par_threads() { 6108 // Don't change the number of workers. Use the value previously set 6109 // in the workgroup. 6110 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6111 uint n_workers = workers()->active_workers(); 6112 assert(UseDynamicNumberOfGCThreads || 6113 n_workers == workers()->total_workers(), 6114 "Otherwise should be using the total number of workers"); 6115 if (n_workers == 0) { 6116 assert(false, "Should have been set in prior evacuation pause."); 6117 n_workers = ParallelGCThreads; 6118 workers()->set_active_workers(n_workers); 6119 } 6120 set_par_threads(n_workers); 6121 } 6122 6123 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6124 size_t allocated_bytes) { 6125 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6126 } 6127 6128 // Methods for the GC alloc regions 6129 6130 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6131 uint count, 6132 GCAllocPurpose ap) { 6133 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6134 6135 if (count < g1_policy()->max_regions(ap)) { 6136 HeapRegion* new_alloc_region = new_region(word_size, 6137 true /* do_expand */); 6138 if (new_alloc_region != NULL) { 6139 // We really only need to do this for old regions given that we 6140 // should never scan survivors. But it doesn't hurt to do it 6141 // for survivors too. 6142 new_alloc_region->set_saved_mark(); 6143 if (ap == GCAllocForSurvived) { 6144 new_alloc_region->set_survivor(); 6145 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6146 } else { 6147 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6148 } 6149 bool during_im = g1_policy()->during_initial_mark_pause(); 6150 new_alloc_region->note_start_of_copying(during_im); 6151 return new_alloc_region; 6152 } else { 6153 g1_policy()->note_alloc_region_limit_reached(ap); 6154 } 6155 } 6156 return NULL; 6157 } 6158 6159 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6160 size_t allocated_bytes, 6161 GCAllocPurpose ap) { 6162 bool during_im = g1_policy()->during_initial_mark_pause(); 6163 alloc_region->note_end_of_copying(during_im); 6164 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6165 if (ap == GCAllocForSurvived) { 6166 young_list()->add_survivor_region(alloc_region); 6167 } else { 6168 _old_set.add(alloc_region); 6169 } 6170 _hr_printer.retire(alloc_region); 6171 } 6172 6173 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6174 bool force) { 6175 assert(!force, "not supported for GC alloc regions"); 6176 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6177 } 6178 6179 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6180 size_t allocated_bytes) { 6181 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6182 GCAllocForSurvived); 6183 } 6184 6185 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6186 bool force) { 6187 assert(!force, "not supported for GC alloc regions"); 6188 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6189 } 6190 6191 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6192 size_t allocated_bytes) { 6193 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6194 GCAllocForTenured); 6195 } 6196 // Heap region set verification 6197 6198 class VerifyRegionListsClosure : public HeapRegionClosure { 6199 private: 6200 FreeRegionList* _free_list; 6201 OldRegionSet* _old_set; 6202 HumongousRegionSet* _humongous_set; 6203 uint _region_count; 6204 6205 public: 6206 VerifyRegionListsClosure(OldRegionSet* old_set, 6207 HumongousRegionSet* humongous_set, 6208 FreeRegionList* free_list) : 6209 _old_set(old_set), _humongous_set(humongous_set), 6210 _free_list(free_list), _region_count(0) { } 6211 6212 uint region_count() { return _region_count; } 6213 6214 bool doHeapRegion(HeapRegion* hr) { 6215 _region_count += 1; 6216 6217 if (hr->continuesHumongous()) { 6218 return false; 6219 } 6220 6221 if (hr->is_young()) { 6222 // TODO 6223 } else if (hr->startsHumongous()) { 6224 _humongous_set->verify_next_region(hr); 6225 } else if (hr->is_empty()) { 6226 _free_list->verify_next_region(hr); 6227 } else { 6228 _old_set->verify_next_region(hr); 6229 } 6230 return false; 6231 } 6232 }; 6233 6234 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6235 HeapWord* bottom) { 6236 HeapWord* end = bottom + HeapRegion::GrainWords; 6237 MemRegion mr(bottom, end); 6238 assert(_g1_reserved.contains(mr), "invariant"); 6239 // This might return NULL if the allocation fails 6240 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6241 } 6242 6243 void G1CollectedHeap::verify_region_sets() { 6244 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6245 6246 // First, check the explicit lists. 6247 _free_list.verify(); 6248 { 6249 // Given that a concurrent operation might be adding regions to 6250 // the secondary free list we have to take the lock before 6251 // verifying it. 6252 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6253 _secondary_free_list.verify(); 6254 } 6255 _old_set.verify(); 6256 _humongous_set.verify(); 6257 6258 // If a concurrent region freeing operation is in progress it will 6259 // be difficult to correctly attributed any free regions we come 6260 // across to the correct free list given that they might belong to 6261 // one of several (free_list, secondary_free_list, any local lists, 6262 // etc.). So, if that's the case we will skip the rest of the 6263 // verification operation. Alternatively, waiting for the concurrent 6264 // operation to complete will have a non-trivial effect on the GC's 6265 // operation (no concurrent operation will last longer than the 6266 // interval between two calls to verification) and it might hide 6267 // any issues that we would like to catch during testing. 6268 if (free_regions_coming()) { 6269 return; 6270 } 6271 6272 // Make sure we append the secondary_free_list on the free_list so 6273 // that all free regions we will come across can be safely 6274 // attributed to the free_list. 6275 append_secondary_free_list_if_not_empty_with_lock(); 6276 6277 // Finally, make sure that the region accounting in the lists is 6278 // consistent with what we see in the heap. 6279 _old_set.verify_start(); 6280 _humongous_set.verify_start(); 6281 _free_list.verify_start(); 6282 6283 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6284 heap_region_iterate(&cl); 6285 6286 _old_set.verify_end(); 6287 _humongous_set.verify_end(); 6288 _free_list.verify_end(); 6289 }