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