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 // The GCLocker may not be active but the GCLocker initiated 956 // GC may not yet have been performed (GCLocker::needs_gc() 957 // returns true). In this case we do not try this GC and 958 // wait until the GCLocker initiated GC is performed, and 959 // then retry the allocation. 960 if (GC_locker::needs_gc()) { 961 should_try_gc = false; 962 } else { 963 // Read the GC count while still holding the Heap_lock. 964 gc_count_before = total_collections(); 965 should_try_gc = true; 966 } 967 } 968 } 969 970 if (should_try_gc) { 971 bool succeeded; 972 result = do_collection_pause(word_size, gc_count_before, &succeeded); 973 if (result != NULL) { 974 assert(succeeded, "only way to get back a non-NULL result"); 975 return result; 976 } 977 978 if (succeeded) { 979 // If we get here we successfully scheduled a collection which 980 // failed to allocate. No point in trying to allocate 981 // further. We'll just return NULL. 982 MutexLockerEx x(Heap_lock); 983 *gc_count_before_ret = total_collections(); 984 return NULL; 985 } 986 } else { 987 // The GCLocker is either active or the GCLocker initiated 988 // GC has not yet been performed. Stall until it is and 989 // then retry the allocation. 990 GC_locker::stall_until_clear(); 991 } 992 993 // We can reach here if we were unsuccessul in scheduling a 994 // collection (because another thread beat us to it) or if we were 995 // stalled due to the GC locker. In either can we should retry the 996 // allocation attempt in case another thread successfully 997 // performed a collection and reclaimed enough space. We do the 998 // first attempt (without holding the Heap_lock) here and the 999 // follow-on attempt will be at the start of the next loop 1000 // iteration (after taking the Heap_lock). 1001 result = _mutator_alloc_region.attempt_allocation(word_size, 1002 false /* bot_updates */); 1003 if (result != NULL) { 1004 return result; 1005 } 1006 1007 // Give a warning if we seem to be looping forever. 1008 if ((QueuedAllocationWarningCount > 0) && 1009 (try_count % QueuedAllocationWarningCount == 0)) { 1010 warning("G1CollectedHeap::attempt_allocation_slow() " 1011 "retries %d times", try_count); 1012 } 1013 } 1014 1015 ShouldNotReachHere(); 1016 return NULL; 1017 } 1018 1019 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1020 unsigned int * gc_count_before_ret) { 1021 // The structure of this method has a lot of similarities to 1022 // attempt_allocation_slow(). The reason these two were not merged 1023 // into a single one is that such a method would require several "if 1024 // allocation is not humongous do this, otherwise do that" 1025 // conditional paths which would obscure its flow. In fact, an early 1026 // version of this code did use a unified method which was harder to 1027 // follow and, as a result, it had subtle bugs that were hard to 1028 // track down. So keeping these two methods separate allows each to 1029 // be more readable. It will be good to keep these two in sync as 1030 // much as possible. 1031 1032 assert_heap_not_locked_and_not_at_safepoint(); 1033 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1034 "should only be called for humongous allocations"); 1035 1036 // Humongous objects can exhaust the heap quickly, so we should check if we 1037 // need to start a marking cycle at each humongous object allocation. We do 1038 // the check before we do the actual allocation. The reason for doing it 1039 // before the allocation is that we avoid having to keep track of the newly 1040 // allocated memory while we do a GC. 1041 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1042 word_size)) { 1043 collect(GCCause::_g1_humongous_allocation); 1044 } 1045 1046 // We will loop until a) we manage to successfully perform the 1047 // allocation or b) we successfully schedule a collection which 1048 // fails to perform the allocation. b) is the only case when we'll 1049 // return NULL. 1050 HeapWord* result = NULL; 1051 for (int try_count = 1; /* we'll return */; try_count += 1) { 1052 bool should_try_gc; 1053 unsigned int gc_count_before; 1054 1055 { 1056 MutexLockerEx x(Heap_lock); 1057 1058 // Given that humongous objects are not allocated in young 1059 // regions, we'll first try to do the allocation without doing a 1060 // collection hoping that there's enough space in the heap. 1061 result = humongous_obj_allocate(word_size); 1062 if (result != NULL) { 1063 return result; 1064 } 1065 1066 if (GC_locker::is_active_and_needs_gc()) { 1067 should_try_gc = false; 1068 } else { 1069 // The GCLocker may not be active but the GCLocker initiated 1070 // GC may not yet have been performed (GCLocker::needs_gc() 1071 // returns true). In this case we do not try this GC and 1072 // wait until the GCLocker initiated GC is performed, and 1073 // then retry the allocation. 1074 if (GC_locker::needs_gc()) { 1075 should_try_gc = false; 1076 } else { 1077 // Read the GC count while still holding the Heap_lock. 1078 gc_count_before = total_collections(); 1079 should_try_gc = true; 1080 } 1081 } 1082 } 1083 1084 if (should_try_gc) { 1085 // If we failed to allocate the humongous object, we should try to 1086 // do a collection pause (if we're allowed) in case it reclaims 1087 // enough space for the allocation to succeed after the pause. 1088 1089 bool succeeded; 1090 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1091 if (result != NULL) { 1092 assert(succeeded, "only way to get back a non-NULL result"); 1093 return result; 1094 } 1095 1096 if (succeeded) { 1097 // If we get here we successfully scheduled a collection which 1098 // failed to allocate. No point in trying to allocate 1099 // further. We'll just return NULL. 1100 MutexLockerEx x(Heap_lock); 1101 *gc_count_before_ret = total_collections(); 1102 return NULL; 1103 } 1104 } else { 1105 // The GCLocker is either active or the GCLocker initiated 1106 // GC has not yet been performed. Stall until it is and 1107 // then retry the allocation. 1108 GC_locker::stall_until_clear(); 1109 } 1110 1111 // We can reach here if we were unsuccessul in scheduling a 1112 // collection (because another thread beat us to it) or if we were 1113 // stalled due to the GC locker. In either can we should retry the 1114 // allocation attempt in case another thread successfully 1115 // performed a collection and reclaimed enough space. Give a 1116 // warning if we seem to be looping forever. 1117 1118 if ((QueuedAllocationWarningCount > 0) && 1119 (try_count % QueuedAllocationWarningCount == 0)) { 1120 warning("G1CollectedHeap::attempt_allocation_humongous() " 1121 "retries %d times", try_count); 1122 } 1123 } 1124 1125 ShouldNotReachHere(); 1126 return NULL; 1127 } 1128 1129 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1130 bool expect_null_mutator_alloc_region) { 1131 assert_at_safepoint(true /* should_be_vm_thread */); 1132 assert(_mutator_alloc_region.get() == NULL || 1133 !expect_null_mutator_alloc_region, 1134 "the current alloc region was unexpectedly found to be non-NULL"); 1135 1136 if (!isHumongous(word_size)) { 1137 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1138 false /* bot_updates */); 1139 } else { 1140 HeapWord* result = humongous_obj_allocate(word_size); 1141 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1142 g1_policy()->set_initiate_conc_mark_if_possible(); 1143 } 1144 return result; 1145 } 1146 1147 ShouldNotReachHere(); 1148 } 1149 1150 class PostMCRemSetClearClosure: public HeapRegionClosure { 1151 G1CollectedHeap* _g1h; 1152 ModRefBarrierSet* _mr_bs; 1153 public: 1154 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1155 _g1h(g1h), _mr_bs(mr_bs) { } 1156 bool doHeapRegion(HeapRegion* r) { 1157 if (r->continuesHumongous()) { 1158 return false; 1159 } 1160 _g1h->reset_gc_time_stamps(r); 1161 HeapRegionRemSet* hrrs = r->rem_set(); 1162 if (hrrs != NULL) hrrs->clear(); 1163 // You might think here that we could clear just the cards 1164 // corresponding to the used region. But no: if we leave a dirty card 1165 // in a region we might allocate into, then it would prevent that card 1166 // from being enqueued, and cause it to be missed. 1167 // Re: the performance cost: we shouldn't be doing full GC anyway! 1168 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1169 return false; 1170 } 1171 }; 1172 1173 void G1CollectedHeap::clear_rsets_post_compaction() { 1174 PostMCRemSetClearClosure rs_clear(this, mr_bs()); 1175 heap_region_iterate(&rs_clear); 1176 } 1177 1178 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1179 G1CollectedHeap* _g1h; 1180 UpdateRSOopClosure _cl; 1181 int _worker_i; 1182 public: 1183 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1184 _cl(g1->g1_rem_set(), worker_i), 1185 _worker_i(worker_i), 1186 _g1h(g1) 1187 { } 1188 1189 bool doHeapRegion(HeapRegion* r) { 1190 if (!r->continuesHumongous()) { 1191 _cl.set_from(r); 1192 r->oop_iterate(&_cl); 1193 } 1194 return false; 1195 } 1196 }; 1197 1198 class ParRebuildRSTask: public AbstractGangTask { 1199 G1CollectedHeap* _g1; 1200 public: 1201 ParRebuildRSTask(G1CollectedHeap* g1) 1202 : AbstractGangTask("ParRebuildRSTask"), 1203 _g1(g1) 1204 { } 1205 1206 void work(uint worker_id) { 1207 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1208 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1209 _g1->workers()->active_workers(), 1210 HeapRegion::RebuildRSClaimValue); 1211 } 1212 }; 1213 1214 class PostCompactionPrinterClosure: public HeapRegionClosure { 1215 private: 1216 G1HRPrinter* _hr_printer; 1217 public: 1218 bool doHeapRegion(HeapRegion* hr) { 1219 assert(!hr->is_young(), "not expecting to find young regions"); 1220 // We only generate output for non-empty regions. 1221 if (!hr->is_empty()) { 1222 if (!hr->isHumongous()) { 1223 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1224 } else if (hr->startsHumongous()) { 1225 if (hr->region_num() == 1) { 1226 // single humongous region 1227 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1228 } else { 1229 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1230 } 1231 } else { 1232 assert(hr->continuesHumongous(), "only way to get here"); 1233 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1234 } 1235 } 1236 return false; 1237 } 1238 1239 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1240 : _hr_printer(hr_printer) { } 1241 }; 1242 1243 void G1CollectedHeap::print_hrs_post_compaction() { 1244 PostCompactionPrinterClosure cl(hr_printer()); 1245 heap_region_iterate(&cl); 1246 } 1247 1248 bool G1CollectedHeap::do_collection(bool explicit_gc, 1249 bool clear_all_soft_refs, 1250 size_t word_size) { 1251 assert_at_safepoint(true /* should_be_vm_thread */); 1252 1253 if (GC_locker::check_active_before_gc()) { 1254 return false; 1255 } 1256 1257 SvcGCMarker sgcm(SvcGCMarker::FULL); 1258 ResourceMark rm; 1259 1260 print_heap_before_gc(); 1261 1262 HRSPhaseSetter x(HRSPhaseFullGC); 1263 verify_region_sets_optional(); 1264 1265 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1266 collector_policy()->should_clear_all_soft_refs(); 1267 1268 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1269 1270 { 1271 IsGCActiveMark x; 1272 1273 // Timing 1274 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1275 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1276 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1277 1278 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty); 1279 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1280 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1281 1282 double start = os::elapsedTime(); 1283 g1_policy()->record_full_collection_start(); 1284 1285 // Note: When we have a more flexible GC logging framework that 1286 // allows us to add optional attributes to a GC log record we 1287 // could consider timing and reporting how long we wait in the 1288 // following two methods. 1289 wait_while_free_regions_coming(); 1290 // If we start the compaction before the CM threads finish 1291 // scanning the root regions we might trip them over as we'll 1292 // be moving objects / updating references. So let's wait until 1293 // they are done. By telling them to abort, they should complete 1294 // early. 1295 _cm->root_regions()->abort(); 1296 _cm->root_regions()->wait_until_scan_finished(); 1297 append_secondary_free_list_if_not_empty_with_lock(); 1298 1299 gc_prologue(true); 1300 increment_total_collections(true /* full gc */); 1301 increment_old_marking_cycles_started(); 1302 1303 size_t g1h_prev_used = used(); 1304 assert(used() == recalculate_used(), "Should be equal"); 1305 1306 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 1307 HandleMark hm; // Discard invalid handles created during verification 1308 gclog_or_tty->print(" VerifyBeforeGC:"); 1309 prepare_for_verify(); 1310 Universe::verify(/* silent */ false, 1311 /* option */ VerifyOption_G1UsePrevMarking); 1312 1313 } 1314 pre_full_gc_dump(); 1315 1316 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1317 1318 // Disable discovery and empty the discovered lists 1319 // for the CM ref processor. 1320 ref_processor_cm()->disable_discovery(); 1321 ref_processor_cm()->abandon_partial_discovery(); 1322 ref_processor_cm()->verify_no_references_recorded(); 1323 1324 // Abandon current iterations of concurrent marking and concurrent 1325 // refinement, if any are in progress. We have to do this before 1326 // wait_until_scan_finished() below. 1327 concurrent_mark()->abort(); 1328 1329 // Make sure we'll choose a new allocation region afterwards. 1330 release_mutator_alloc_region(); 1331 abandon_gc_alloc_regions(); 1332 g1_rem_set()->cleanupHRRS(); 1333 1334 // We should call this after we retire any currently active alloc 1335 // regions so that all the ALLOC / RETIRE events are generated 1336 // before the start GC event. 1337 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1338 1339 // We may have added regions to the current incremental collection 1340 // set between the last GC or pause and now. We need to clear the 1341 // incremental collection set and then start rebuilding it afresh 1342 // after this full GC. 1343 abandon_collection_set(g1_policy()->inc_cset_head()); 1344 g1_policy()->clear_incremental_cset(); 1345 g1_policy()->stop_incremental_cset_building(); 1346 1347 tear_down_region_sets(false /* free_list_only */); 1348 g1_policy()->set_gcs_are_young(true); 1349 1350 // See the comments in g1CollectedHeap.hpp and 1351 // G1CollectedHeap::ref_processing_init() about 1352 // how reference processing currently works in G1. 1353 1354 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1355 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1356 1357 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1358 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1359 1360 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1361 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1362 1363 // Do collection work 1364 { 1365 HandleMark hm; // Discard invalid handles created during gc 1366 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1367 } 1368 1369 assert(free_regions() == 0, "we should not have added any free regions"); 1370 rebuild_region_sets(false /* free_list_only */); 1371 1372 // Enqueue any discovered reference objects that have 1373 // not been removed from the discovered lists. 1374 ref_processor_stw()->enqueue_discovered_references(); 1375 1376 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1377 1378 MemoryService::track_memory_usage(); 1379 1380 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 1381 HandleMark hm; // Discard invalid handles created during verification 1382 gclog_or_tty->print(" VerifyAfterGC:"); 1383 prepare_for_verify(); 1384 Universe::verify(/* silent */ false, 1385 /* option */ VerifyOption_G1UsePrevMarking); 1386 1387 } 1388 1389 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1390 ref_processor_stw()->verify_no_references_recorded(); 1391 1392 // Note: since we've just done a full GC, concurrent 1393 // marking is no longer active. Therefore we need not 1394 // re-enable reference discovery for the CM ref processor. 1395 // That will be done at the start of the next marking cycle. 1396 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1397 ref_processor_cm()->verify_no_references_recorded(); 1398 1399 reset_gc_time_stamp(); 1400 // Since everything potentially moved, we will clear all remembered 1401 // sets, and clear all cards. Later we will rebuild remebered 1402 // sets. We will also reset the GC time stamps of the regions. 1403 clear_rsets_post_compaction(); 1404 check_gc_time_stamps(); 1405 1406 // Resize the heap if necessary. 1407 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1408 1409 if (_hr_printer.is_active()) { 1410 // We should do this after we potentially resize the heap so 1411 // that all the COMMIT / UNCOMMIT events are generated before 1412 // the end GC event. 1413 1414 print_hrs_post_compaction(); 1415 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1416 } 1417 1418 if (_cg1r->use_cache()) { 1419 _cg1r->clear_and_record_card_counts(); 1420 _cg1r->clear_hot_cache(); 1421 } 1422 1423 // Rebuild remembered sets of all regions. 1424 if (G1CollectedHeap::use_parallel_gc_threads()) { 1425 uint n_workers = 1426 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1427 workers()->active_workers(), 1428 Threads::number_of_non_daemon_threads()); 1429 assert(UseDynamicNumberOfGCThreads || 1430 n_workers == workers()->total_workers(), 1431 "If not dynamic should be using all the workers"); 1432 workers()->set_active_workers(n_workers); 1433 // Set parallel threads in the heap (_n_par_threads) only 1434 // before a parallel phase and always reset it to 0 after 1435 // the phase so that the number of parallel threads does 1436 // no get carried forward to a serial phase where there 1437 // may be code that is "possibly_parallel". 1438 set_par_threads(n_workers); 1439 1440 ParRebuildRSTask rebuild_rs_task(this); 1441 assert(check_heap_region_claim_values( 1442 HeapRegion::InitialClaimValue), "sanity check"); 1443 assert(UseDynamicNumberOfGCThreads || 1444 workers()->active_workers() == workers()->total_workers(), 1445 "Unless dynamic should use total workers"); 1446 // Use the most recent number of active workers 1447 assert(workers()->active_workers() > 0, 1448 "Active workers not properly set"); 1449 set_par_threads(workers()->active_workers()); 1450 workers()->run_task(&rebuild_rs_task); 1451 set_par_threads(0); 1452 assert(check_heap_region_claim_values( 1453 HeapRegion::RebuildRSClaimValue), "sanity check"); 1454 reset_heap_region_claim_values(); 1455 } else { 1456 RebuildRSOutOfRegionClosure rebuild_rs(this); 1457 heap_region_iterate(&rebuild_rs); 1458 } 1459 1460 if (G1Log::fine()) { 1461 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1462 } 1463 1464 if (true) { // FIXME 1465 // Ask the permanent generation to adjust size for full collections 1466 perm()->compute_new_size(); 1467 } 1468 1469 // Start a new incremental collection set for the next pause 1470 assert(g1_policy()->collection_set() == NULL, "must be"); 1471 g1_policy()->start_incremental_cset_building(); 1472 1473 // Clear the _cset_fast_test bitmap in anticipation of adding 1474 // regions to the incremental collection set for the next 1475 // evacuation pause. 1476 clear_cset_fast_test(); 1477 1478 init_mutator_alloc_region(); 1479 1480 double end = os::elapsedTime(); 1481 g1_policy()->record_full_collection_end(); 1482 1483 #ifdef TRACESPINNING 1484 ParallelTaskTerminator::print_termination_counts(); 1485 #endif 1486 1487 gc_epilogue(true); 1488 1489 // Discard all rset updates 1490 JavaThread::dirty_card_queue_set().abandon_logs(); 1491 assert(!G1DeferredRSUpdate 1492 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1493 1494 _young_list->reset_sampled_info(); 1495 // At this point there should be no regions in the 1496 // entire heap tagged as young. 1497 assert( check_young_list_empty(true /* check_heap */), 1498 "young list should be empty at this point"); 1499 1500 // Update the number of full collections that have been completed. 1501 increment_old_marking_cycles_completed(false /* concurrent */); 1502 1503 _hrs.verify_optional(); 1504 verify_region_sets_optional(); 1505 1506 print_heap_after_gc(); 1507 1508 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1509 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1510 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1511 // before any GC notifications are raised. 1512 g1mm()->update_sizes(); 1513 } 1514 1515 post_full_gc_dump(); 1516 1517 return true; 1518 } 1519 1520 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1521 // do_collection() will return whether it succeeded in performing 1522 // the GC. Currently, there is no facility on the 1523 // do_full_collection() API to notify the caller than the collection 1524 // did not succeed (e.g., because it was locked out by the GC 1525 // locker). So, right now, we'll ignore the return value. 1526 bool dummy = do_collection(true, /* explicit_gc */ 1527 clear_all_soft_refs, 1528 0 /* word_size */); 1529 } 1530 1531 // This code is mostly copied from TenuredGeneration. 1532 void 1533 G1CollectedHeap:: 1534 resize_if_necessary_after_full_collection(size_t word_size) { 1535 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1536 1537 // Include the current allocation, if any, and bytes that will be 1538 // pre-allocated to support collections, as "used". 1539 const size_t used_after_gc = used(); 1540 const size_t capacity_after_gc = capacity(); 1541 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1542 1543 // This is enforced in arguments.cpp. 1544 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1545 "otherwise the code below doesn't make sense"); 1546 1547 // We don't have floating point command-line arguments 1548 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1549 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1550 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1551 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1552 1553 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1554 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1555 1556 // We have to be careful here as these two calculations can overflow 1557 // 32-bit size_t's. 1558 double used_after_gc_d = (double) used_after_gc; 1559 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1560 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1561 1562 // Let's make sure that they are both under the max heap size, which 1563 // by default will make them fit into a size_t. 1564 double desired_capacity_upper_bound = (double) max_heap_size; 1565 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1566 desired_capacity_upper_bound); 1567 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1568 desired_capacity_upper_bound); 1569 1570 // We can now safely turn them into size_t's. 1571 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1572 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1573 1574 // This assert only makes sense here, before we adjust them 1575 // with respect to the min and max heap size. 1576 assert(minimum_desired_capacity <= maximum_desired_capacity, 1577 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1578 "maximum_desired_capacity = "SIZE_FORMAT, 1579 minimum_desired_capacity, maximum_desired_capacity)); 1580 1581 // Should not be greater than the heap max size. No need to adjust 1582 // it with respect to the heap min size as it's a lower bound (i.e., 1583 // we'll try to make the capacity larger than it, not smaller). 1584 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1585 // Should not be less than the heap min size. No need to adjust it 1586 // with respect to the heap max size as it's an upper bound (i.e., 1587 // we'll try to make the capacity smaller than it, not greater). 1588 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1589 1590 if (capacity_after_gc < minimum_desired_capacity) { 1591 // Don't expand unless it's significant 1592 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1593 ergo_verbose4(ErgoHeapSizing, 1594 "attempt heap expansion", 1595 ergo_format_reason("capacity lower than " 1596 "min desired capacity after Full GC") 1597 ergo_format_byte("capacity") 1598 ergo_format_byte("occupancy") 1599 ergo_format_byte_perc("min desired capacity"), 1600 capacity_after_gc, used_after_gc, 1601 minimum_desired_capacity, (double) MinHeapFreeRatio); 1602 expand(expand_bytes); 1603 1604 // No expansion, now see if we want to shrink 1605 } else if (capacity_after_gc > maximum_desired_capacity) { 1606 // Capacity too large, compute shrinking size 1607 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1608 ergo_verbose4(ErgoHeapSizing, 1609 "attempt heap shrinking", 1610 ergo_format_reason("capacity higher than " 1611 "max desired capacity after Full GC") 1612 ergo_format_byte("capacity") 1613 ergo_format_byte("occupancy") 1614 ergo_format_byte_perc("max desired capacity"), 1615 capacity_after_gc, used_after_gc, 1616 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1617 shrink(shrink_bytes); 1618 } 1619 } 1620 1621 1622 HeapWord* 1623 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1624 bool* succeeded) { 1625 assert_at_safepoint(true /* should_be_vm_thread */); 1626 1627 *succeeded = true; 1628 // Let's attempt the allocation first. 1629 HeapWord* result = 1630 attempt_allocation_at_safepoint(word_size, 1631 false /* expect_null_mutator_alloc_region */); 1632 if (result != NULL) { 1633 assert(*succeeded, "sanity"); 1634 return result; 1635 } 1636 1637 // In a G1 heap, we're supposed to keep allocation from failing by 1638 // incremental pauses. Therefore, at least for now, we'll favor 1639 // expansion over collection. (This might change in the future if we can 1640 // do something smarter than full collection to satisfy a failed alloc.) 1641 result = expand_and_allocate(word_size); 1642 if (result != NULL) { 1643 assert(*succeeded, "sanity"); 1644 return result; 1645 } 1646 1647 // Expansion didn't work, we'll try to do a Full GC. 1648 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1649 false, /* clear_all_soft_refs */ 1650 word_size); 1651 if (!gc_succeeded) { 1652 *succeeded = false; 1653 return NULL; 1654 } 1655 1656 // Retry the allocation 1657 result = attempt_allocation_at_safepoint(word_size, 1658 true /* expect_null_mutator_alloc_region */); 1659 if (result != NULL) { 1660 assert(*succeeded, "sanity"); 1661 return result; 1662 } 1663 1664 // Then, try a Full GC that will collect all soft references. 1665 gc_succeeded = do_collection(false, /* explicit_gc */ 1666 true, /* clear_all_soft_refs */ 1667 word_size); 1668 if (!gc_succeeded) { 1669 *succeeded = false; 1670 return NULL; 1671 } 1672 1673 // Retry the allocation once more 1674 result = attempt_allocation_at_safepoint(word_size, 1675 true /* expect_null_mutator_alloc_region */); 1676 if (result != NULL) { 1677 assert(*succeeded, "sanity"); 1678 return result; 1679 } 1680 1681 assert(!collector_policy()->should_clear_all_soft_refs(), 1682 "Flag should have been handled and cleared prior to this point"); 1683 1684 // What else? We might try synchronous finalization later. If the total 1685 // space available is large enough for the allocation, then a more 1686 // complete compaction phase than we've tried so far might be 1687 // appropriate. 1688 assert(*succeeded, "sanity"); 1689 return NULL; 1690 } 1691 1692 // Attempting to expand the heap sufficiently 1693 // to support an allocation of the given "word_size". If 1694 // successful, perform the allocation and return the address of the 1695 // allocated block, or else "NULL". 1696 1697 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1698 assert_at_safepoint(true /* should_be_vm_thread */); 1699 1700 verify_region_sets_optional(); 1701 1702 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1703 ergo_verbose1(ErgoHeapSizing, 1704 "attempt heap expansion", 1705 ergo_format_reason("allocation request failed") 1706 ergo_format_byte("allocation request"), 1707 word_size * HeapWordSize); 1708 if (expand(expand_bytes)) { 1709 _hrs.verify_optional(); 1710 verify_region_sets_optional(); 1711 return attempt_allocation_at_safepoint(word_size, 1712 false /* expect_null_mutator_alloc_region */); 1713 } 1714 return NULL; 1715 } 1716 1717 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1718 HeapWord* new_end) { 1719 assert(old_end != new_end, "don't call this otherwise"); 1720 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1721 1722 // Update the committed mem region. 1723 _g1_committed.set_end(new_end); 1724 // Tell the card table about the update. 1725 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1726 // Tell the BOT about the update. 1727 _bot_shared->resize(_g1_committed.word_size()); 1728 } 1729 1730 bool G1CollectedHeap::expand(size_t expand_bytes) { 1731 size_t old_mem_size = _g1_storage.committed_size(); 1732 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1733 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1734 HeapRegion::GrainBytes); 1735 ergo_verbose2(ErgoHeapSizing, 1736 "expand the heap", 1737 ergo_format_byte("requested expansion amount") 1738 ergo_format_byte("attempted expansion amount"), 1739 expand_bytes, aligned_expand_bytes); 1740 1741 // First commit the memory. 1742 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1743 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1744 if (successful) { 1745 // Then propagate this update to the necessary data structures. 1746 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1747 update_committed_space(old_end, new_end); 1748 1749 FreeRegionList expansion_list("Local Expansion List"); 1750 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1751 assert(mr.start() == old_end, "post-condition"); 1752 // mr might be a smaller region than what was requested if 1753 // expand_by() was unable to allocate the HeapRegion instances 1754 assert(mr.end() <= new_end, "post-condition"); 1755 1756 size_t actual_expand_bytes = mr.byte_size(); 1757 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1758 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1759 "post-condition"); 1760 if (actual_expand_bytes < aligned_expand_bytes) { 1761 // We could not expand _hrs to the desired size. In this case we 1762 // need to shrink the committed space accordingly. 1763 assert(mr.end() < new_end, "invariant"); 1764 1765 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1766 // First uncommit the memory. 1767 _g1_storage.shrink_by(diff_bytes); 1768 // Then propagate this update to the necessary data structures. 1769 update_committed_space(new_end, mr.end()); 1770 } 1771 _free_list.add_as_tail(&expansion_list); 1772 1773 if (_hr_printer.is_active()) { 1774 HeapWord* curr = mr.start(); 1775 while (curr < mr.end()) { 1776 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1777 _hr_printer.commit(curr, curr_end); 1778 curr = curr_end; 1779 } 1780 assert(curr == mr.end(), "post-condition"); 1781 } 1782 g1_policy()->record_new_heap_size(n_regions()); 1783 } else { 1784 ergo_verbose0(ErgoHeapSizing, 1785 "did not expand the heap", 1786 ergo_format_reason("heap expansion operation failed")); 1787 // The expansion of the virtual storage space was unsuccessful. 1788 // Let's see if it was because we ran out of swap. 1789 if (G1ExitOnExpansionFailure && 1790 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1791 // We had head room... 1792 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1793 } 1794 } 1795 return successful; 1796 } 1797 1798 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1799 size_t old_mem_size = _g1_storage.committed_size(); 1800 size_t aligned_shrink_bytes = 1801 ReservedSpace::page_align_size_down(shrink_bytes); 1802 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1803 HeapRegion::GrainBytes); 1804 uint num_regions_deleted = 0; 1805 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted); 1806 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1807 assert(mr.end() == old_end, "post-condition"); 1808 1809 ergo_verbose3(ErgoHeapSizing, 1810 "shrink the heap", 1811 ergo_format_byte("requested shrinking amount") 1812 ergo_format_byte("aligned shrinking amount") 1813 ergo_format_byte("attempted shrinking amount"), 1814 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1815 if (mr.byte_size() > 0) { 1816 if (_hr_printer.is_active()) { 1817 HeapWord* curr = mr.end(); 1818 while (curr > mr.start()) { 1819 HeapWord* curr_end = curr; 1820 curr -= HeapRegion::GrainWords; 1821 _hr_printer.uncommit(curr, curr_end); 1822 } 1823 assert(curr == mr.start(), "post-condition"); 1824 } 1825 1826 _g1_storage.shrink_by(mr.byte_size()); 1827 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1828 assert(mr.start() == new_end, "post-condition"); 1829 1830 _expansion_regions += num_regions_deleted; 1831 update_committed_space(old_end, new_end); 1832 HeapRegionRemSet::shrink_heap(n_regions()); 1833 g1_policy()->record_new_heap_size(n_regions()); 1834 } else { 1835 ergo_verbose0(ErgoHeapSizing, 1836 "did not shrink the heap", 1837 ergo_format_reason("heap shrinking operation failed")); 1838 } 1839 } 1840 1841 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1842 verify_region_sets_optional(); 1843 1844 // We should only reach here at the end of a Full GC which means we 1845 // should not not be holding to any GC alloc regions. The method 1846 // below will make sure of that and do any remaining clean up. 1847 abandon_gc_alloc_regions(); 1848 1849 // Instead of tearing down / rebuilding the free lists here, we 1850 // could instead use the remove_all_pending() method on free_list to 1851 // remove only the ones that we need to remove. 1852 tear_down_region_sets(true /* free_list_only */); 1853 shrink_helper(shrink_bytes); 1854 rebuild_region_sets(true /* free_list_only */); 1855 1856 _hrs.verify_optional(); 1857 verify_region_sets_optional(); 1858 } 1859 1860 // Public methods. 1861 1862 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1863 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1864 #endif // _MSC_VER 1865 1866 1867 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1868 SharedHeap(policy_), 1869 _g1_policy(policy_), 1870 _dirty_card_queue_set(false), 1871 _into_cset_dirty_card_queue_set(false), 1872 _is_alive_closure_cm(this), 1873 _is_alive_closure_stw(this), 1874 _ref_processor_cm(NULL), 1875 _ref_processor_stw(NULL), 1876 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1877 _bot_shared(NULL), 1878 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL), 1879 _evac_failure_scan_stack(NULL) , 1880 _mark_in_progress(false), 1881 _cg1r(NULL), _summary_bytes_used(0), 1882 _g1mm(NULL), 1883 _refine_cte_cl(NULL), 1884 _full_collection(false), 1885 _free_list("Master Free List"), 1886 _secondary_free_list("Secondary Free List"), 1887 _old_set("Old Set"), 1888 _humongous_set("Master Humongous Set"), 1889 _free_regions_coming(false), 1890 _young_list(new YoungList(this)), 1891 _gc_time_stamp(0), 1892 _retained_old_gc_alloc_region(NULL), 1893 _expand_heap_after_alloc_failure(true), 1894 _surviving_young_words(NULL), 1895 _old_marking_cycles_started(0), 1896 _old_marking_cycles_completed(0), 1897 _in_cset_fast_test(NULL), 1898 _in_cset_fast_test_base(NULL), 1899 _dirty_cards_region_list(NULL), 1900 _worker_cset_start_region(NULL), 1901 _worker_cset_start_region_time_stamp(NULL) { 1902 _g1h = this; // To catch bugs. 1903 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1904 vm_exit_during_initialization("Failed necessary allocation."); 1905 } 1906 1907 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1908 1909 int n_queues = MAX2((int)ParallelGCThreads, 1); 1910 _task_queues = new RefToScanQueueSet(n_queues); 1911 1912 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1913 assert(n_rem_sets > 0, "Invariant."); 1914 1915 HeapRegionRemSetIterator** iter_arr = 1916 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues); 1917 for (int i = 0; i < n_queues; i++) { 1918 iter_arr[i] = new HeapRegionRemSetIterator(); 1919 } 1920 _rem_set_iterator = iter_arr; 1921 1922 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues); 1923 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues); 1924 1925 for (int i = 0; i < n_queues; i++) { 1926 RefToScanQueue* q = new RefToScanQueue(); 1927 q->initialize(); 1928 _task_queues->register_queue(i, q); 1929 } 1930 1931 clear_cset_start_regions(); 1932 1933 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1934 } 1935 1936 jint G1CollectedHeap::initialize() { 1937 CollectedHeap::pre_initialize(); 1938 os::enable_vtime(); 1939 1940 G1Log::init(); 1941 1942 // Necessary to satisfy locking discipline assertions. 1943 1944 MutexLocker x(Heap_lock); 1945 1946 // We have to initialize the printer before committing the heap, as 1947 // it will be used then. 1948 _hr_printer.set_active(G1PrintHeapRegions); 1949 1950 // While there are no constraints in the GC code that HeapWordSize 1951 // be any particular value, there are multiple other areas in the 1952 // system which believe this to be true (e.g. oop->object_size in some 1953 // cases incorrectly returns the size in wordSize units rather than 1954 // HeapWordSize). 1955 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1956 1957 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1958 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1959 1960 // Ensure that the sizes are properly aligned. 1961 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1962 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1963 1964 _cg1r = new ConcurrentG1Refine(); 1965 1966 // Reserve the maximum. 1967 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1968 // Includes the perm-gen. 1969 1970 // When compressed oops are enabled, the preferred heap base 1971 // is calculated by subtracting the requested size from the 1972 // 32Gb boundary and using the result as the base address for 1973 // heap reservation. If the requested size is not aligned to 1974 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1975 // into the ReservedHeapSpace constructor) then the actual 1976 // base of the reserved heap may end up differing from the 1977 // address that was requested (i.e. the preferred heap base). 1978 // If this happens then we could end up using a non-optimal 1979 // compressed oops mode. 1980 1981 // Since max_byte_size is aligned to the size of a heap region (checked 1982 // above), we also need to align the perm gen size as it might not be. 1983 const size_t total_reserved = max_byte_size + 1984 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1985 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1986 1987 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1988 1989 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1990 UseLargePages, addr); 1991 1992 if (UseCompressedOops) { 1993 if (addr != NULL && !heap_rs.is_reserved()) { 1994 // Failed to reserve at specified address - the requested memory 1995 // region is taken already, for example, by 'java' launcher. 1996 // Try again to reserver heap higher. 1997 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 1998 1999 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 2000 UseLargePages, addr); 2001 2002 if (addr != NULL && !heap_rs0.is_reserved()) { 2003 // Failed to reserve at specified address again - give up. 2004 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 2005 assert(addr == NULL, ""); 2006 2007 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 2008 UseLargePages, addr); 2009 heap_rs = heap_rs1; 2010 } else { 2011 heap_rs = heap_rs0; 2012 } 2013 } 2014 } 2015 2016 if (!heap_rs.is_reserved()) { 2017 vm_exit_during_initialization("Could not reserve enough space for object heap"); 2018 return JNI_ENOMEM; 2019 } 2020 2021 // It is important to do this in a way such that concurrent readers can't 2022 // temporarily think somethings in the heap. (I've actually seen this 2023 // happen in asserts: DLD.) 2024 _reserved.set_word_size(0); 2025 _reserved.set_start((HeapWord*)heap_rs.base()); 2026 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2027 2028 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2029 2030 // Create the gen rem set (and barrier set) for the entire reserved region. 2031 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2032 set_barrier_set(rem_set()->bs()); 2033 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2034 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2035 } else { 2036 vm_exit_during_initialization("G1 requires a mod ref bs."); 2037 return JNI_ENOMEM; 2038 } 2039 2040 // Also create a G1 rem set. 2041 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2042 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2043 } else { 2044 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2045 return JNI_ENOMEM; 2046 } 2047 2048 // Carve out the G1 part of the heap. 2049 2050 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2051 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2052 g1_rs.size()/HeapWordSize); 2053 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 2054 2055 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2056 2057 _g1_storage.initialize(g1_rs, 0); 2058 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2059 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2060 (HeapWord*) _g1_reserved.end(), 2061 _expansion_regions); 2062 2063 // 6843694 - ensure that the maximum region index can fit 2064 // in the remembered set structures. 2065 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2066 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2067 2068 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2069 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2070 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2071 "too many cards per region"); 2072 2073 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2074 2075 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2076 heap_word_size(init_byte_size)); 2077 2078 _g1h = this; 2079 2080 _in_cset_fast_test_length = max_regions(); 2081 _in_cset_fast_test_base = 2082 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length); 2083 2084 // We're biasing _in_cset_fast_test to avoid subtracting the 2085 // beginning of the heap every time we want to index; basically 2086 // it's the same with what we do with the card table. 2087 _in_cset_fast_test = _in_cset_fast_test_base - 2088 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2089 2090 // Clear the _cset_fast_test bitmap in anticipation of adding 2091 // regions to the incremental collection set for the first 2092 // evacuation pause. 2093 clear_cset_fast_test(); 2094 2095 // Create the ConcurrentMark data structure and thread. 2096 // (Must do this late, so that "max_regions" is defined.) 2097 _cm = new ConcurrentMark(heap_rs, max_regions()); 2098 _cmThread = _cm->cmThread(); 2099 2100 // Initialize the from_card cache structure of HeapRegionRemSet. 2101 HeapRegionRemSet::init_heap(max_regions()); 2102 2103 // Now expand into the initial heap size. 2104 if (!expand(init_byte_size)) { 2105 vm_exit_during_initialization("Failed to allocate initial heap."); 2106 return JNI_ENOMEM; 2107 } 2108 2109 // Perform any initialization actions delegated to the policy. 2110 g1_policy()->init(); 2111 2112 _refine_cte_cl = 2113 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2114 g1_rem_set(), 2115 concurrent_g1_refine()); 2116 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2117 2118 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2119 SATB_Q_FL_lock, 2120 G1SATBProcessCompletedThreshold, 2121 Shared_SATB_Q_lock); 2122 2123 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2124 DirtyCardQ_FL_lock, 2125 concurrent_g1_refine()->yellow_zone(), 2126 concurrent_g1_refine()->red_zone(), 2127 Shared_DirtyCardQ_lock); 2128 2129 if (G1DeferredRSUpdate) { 2130 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2131 DirtyCardQ_FL_lock, 2132 -1, // never trigger processing 2133 -1, // no limit on length 2134 Shared_DirtyCardQ_lock, 2135 &JavaThread::dirty_card_queue_set()); 2136 } 2137 2138 // Initialize the card queue set used to hold cards containing 2139 // references into the collection set. 2140 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2141 DirtyCardQ_FL_lock, 2142 -1, // never trigger processing 2143 -1, // no limit on length 2144 Shared_DirtyCardQ_lock, 2145 &JavaThread::dirty_card_queue_set()); 2146 2147 // In case we're keeping closure specialization stats, initialize those 2148 // counts and that mechanism. 2149 SpecializationStats::clear(); 2150 2151 // Do later initialization work for concurrent refinement. 2152 _cg1r->init(); 2153 2154 // Here we allocate the dummy full region that is required by the 2155 // G1AllocRegion class. If we don't pass an address in the reserved 2156 // space here, lots of asserts fire. 2157 2158 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2159 _g1_reserved.start()); 2160 // We'll re-use the same region whether the alloc region will 2161 // require BOT updates or not and, if it doesn't, then a non-young 2162 // region will complain that it cannot support allocations without 2163 // BOT updates. So we'll tag the dummy region as young to avoid that. 2164 dummy_region->set_young(); 2165 // Make sure it's full. 2166 dummy_region->set_top(dummy_region->end()); 2167 G1AllocRegion::setup(this, dummy_region); 2168 2169 init_mutator_alloc_region(); 2170 2171 // Do create of the monitoring and management support so that 2172 // values in the heap have been properly initialized. 2173 _g1mm = new G1MonitoringSupport(this); 2174 2175 return JNI_OK; 2176 } 2177 2178 void G1CollectedHeap::ref_processing_init() { 2179 // Reference processing in G1 currently works as follows: 2180 // 2181 // * There are two reference processor instances. One is 2182 // used to record and process discovered references 2183 // during concurrent marking; the other is used to 2184 // record and process references during STW pauses 2185 // (both full and incremental). 2186 // * Both ref processors need to 'span' the entire heap as 2187 // the regions in the collection set may be dotted around. 2188 // 2189 // * For the concurrent marking ref processor: 2190 // * Reference discovery is enabled at initial marking. 2191 // * Reference discovery is disabled and the discovered 2192 // references processed etc during remarking. 2193 // * Reference discovery is MT (see below). 2194 // * Reference discovery requires a barrier (see below). 2195 // * Reference processing may or may not be MT 2196 // (depending on the value of ParallelRefProcEnabled 2197 // and ParallelGCThreads). 2198 // * A full GC disables reference discovery by the CM 2199 // ref processor and abandons any entries on it's 2200 // discovered lists. 2201 // 2202 // * For the STW processor: 2203 // * Non MT discovery is enabled at the start of a full GC. 2204 // * Processing and enqueueing during a full GC is non-MT. 2205 // * During a full GC, references are processed after marking. 2206 // 2207 // * Discovery (may or may not be MT) is enabled at the start 2208 // of an incremental evacuation pause. 2209 // * References are processed near the end of a STW evacuation pause. 2210 // * For both types of GC: 2211 // * Discovery is atomic - i.e. not concurrent. 2212 // * Reference discovery will not need a barrier. 2213 2214 SharedHeap::ref_processing_init(); 2215 MemRegion mr = reserved_region(); 2216 2217 // Concurrent Mark ref processor 2218 _ref_processor_cm = 2219 new ReferenceProcessor(mr, // span 2220 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2221 // mt processing 2222 (int) ParallelGCThreads, 2223 // degree of mt processing 2224 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2225 // mt discovery 2226 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2227 // degree of mt discovery 2228 false, 2229 // Reference discovery is not atomic 2230 &_is_alive_closure_cm, 2231 // is alive closure 2232 // (for efficiency/performance) 2233 true); 2234 // Setting next fields of discovered 2235 // lists requires a barrier. 2236 2237 // STW ref processor 2238 _ref_processor_stw = 2239 new ReferenceProcessor(mr, // span 2240 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2241 // mt processing 2242 MAX2((int)ParallelGCThreads, 1), 2243 // degree of mt processing 2244 (ParallelGCThreads > 1), 2245 // mt discovery 2246 MAX2((int)ParallelGCThreads, 1), 2247 // degree of mt discovery 2248 true, 2249 // Reference discovery is atomic 2250 &_is_alive_closure_stw, 2251 // is alive closure 2252 // (for efficiency/performance) 2253 false); 2254 // Setting next fields of discovered 2255 // lists requires a barrier. 2256 } 2257 2258 size_t G1CollectedHeap::capacity() const { 2259 return _g1_committed.byte_size(); 2260 } 2261 2262 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2263 assert(!hr->continuesHumongous(), "pre-condition"); 2264 hr->reset_gc_time_stamp(); 2265 if (hr->startsHumongous()) { 2266 uint first_index = hr->hrs_index() + 1; 2267 uint last_index = hr->last_hc_index(); 2268 for (uint i = first_index; i < last_index; i += 1) { 2269 HeapRegion* chr = region_at(i); 2270 assert(chr->continuesHumongous(), "sanity"); 2271 chr->reset_gc_time_stamp(); 2272 } 2273 } 2274 } 2275 2276 #ifndef PRODUCT 2277 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2278 private: 2279 unsigned _gc_time_stamp; 2280 bool _failures; 2281 2282 public: 2283 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2284 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2285 2286 virtual bool doHeapRegion(HeapRegion* hr) { 2287 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2288 if (_gc_time_stamp != region_gc_time_stamp) { 2289 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2290 "expected %d", HR_FORMAT_PARAMS(hr), 2291 region_gc_time_stamp, _gc_time_stamp); 2292 _failures = true; 2293 } 2294 return false; 2295 } 2296 2297 bool failures() { return _failures; } 2298 }; 2299 2300 void G1CollectedHeap::check_gc_time_stamps() { 2301 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2302 heap_region_iterate(&cl); 2303 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2304 } 2305 #endif // PRODUCT 2306 2307 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2308 DirtyCardQueue* into_cset_dcq, 2309 bool concurrent, 2310 int worker_i) { 2311 // Clean cards in the hot card cache 2312 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2313 2314 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2315 int n_completed_buffers = 0; 2316 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2317 n_completed_buffers++; 2318 } 2319 g1_policy()->record_update_rs_processed_buffers(worker_i, 2320 (double) n_completed_buffers); 2321 dcqs.clear_n_completed_buffers(); 2322 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2323 } 2324 2325 2326 // Computes the sum of the storage used by the various regions. 2327 2328 size_t G1CollectedHeap::used() const { 2329 assert(Heap_lock->owner() != NULL, 2330 "Should be owned on this thread's behalf."); 2331 size_t result = _summary_bytes_used; 2332 // Read only once in case it is set to NULL concurrently 2333 HeapRegion* hr = _mutator_alloc_region.get(); 2334 if (hr != NULL) 2335 result += hr->used(); 2336 return result; 2337 } 2338 2339 size_t G1CollectedHeap::used_unlocked() const { 2340 size_t result = _summary_bytes_used; 2341 return result; 2342 } 2343 2344 class SumUsedClosure: public HeapRegionClosure { 2345 size_t _used; 2346 public: 2347 SumUsedClosure() : _used(0) {} 2348 bool doHeapRegion(HeapRegion* r) { 2349 if (!r->continuesHumongous()) { 2350 _used += r->used(); 2351 } 2352 return false; 2353 } 2354 size_t result() { return _used; } 2355 }; 2356 2357 size_t G1CollectedHeap::recalculate_used() const { 2358 SumUsedClosure blk; 2359 heap_region_iterate(&blk); 2360 return blk.result(); 2361 } 2362 2363 size_t G1CollectedHeap::unsafe_max_alloc() { 2364 if (free_regions() > 0) return HeapRegion::GrainBytes; 2365 // otherwise, is there space in the current allocation region? 2366 2367 // We need to store the current allocation region in a local variable 2368 // here. The problem is that this method doesn't take any locks and 2369 // there may be other threads which overwrite the current allocation 2370 // region field. attempt_allocation(), for example, sets it to NULL 2371 // and this can happen *after* the NULL check here but before the call 2372 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2373 // to be a problem in the optimized build, since the two loads of the 2374 // current allocation region field are optimized away. 2375 HeapRegion* hr = _mutator_alloc_region.get(); 2376 if (hr == NULL) { 2377 return 0; 2378 } 2379 return hr->free(); 2380 } 2381 2382 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2383 switch (cause) { 2384 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2385 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2386 case GCCause::_g1_humongous_allocation: return true; 2387 default: return false; 2388 } 2389 } 2390 2391 #ifndef PRODUCT 2392 void G1CollectedHeap::allocate_dummy_regions() { 2393 // Let's fill up most of the region 2394 size_t word_size = HeapRegion::GrainWords - 1024; 2395 // And as a result the region we'll allocate will be humongous. 2396 guarantee(isHumongous(word_size), "sanity"); 2397 2398 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2399 // Let's use the existing mechanism for the allocation 2400 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2401 if (dummy_obj != NULL) { 2402 MemRegion mr(dummy_obj, word_size); 2403 CollectedHeap::fill_with_object(mr); 2404 } else { 2405 // If we can't allocate once, we probably cannot allocate 2406 // again. Let's get out of the loop. 2407 break; 2408 } 2409 } 2410 } 2411 #endif // !PRODUCT 2412 2413 void G1CollectedHeap::increment_old_marking_cycles_started() { 2414 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2415 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2416 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2417 _old_marking_cycles_started, _old_marking_cycles_completed)); 2418 2419 _old_marking_cycles_started++; 2420 } 2421 2422 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2423 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2424 2425 // We assume that if concurrent == true, then the caller is a 2426 // concurrent thread that was joined the Suspendible Thread 2427 // Set. If there's ever a cheap way to check this, we should add an 2428 // assert here. 2429 2430 // Given that this method is called at the end of a Full GC or of a 2431 // concurrent cycle, and those can be nested (i.e., a Full GC can 2432 // interrupt a concurrent cycle), the number of full collections 2433 // completed should be either one (in the case where there was no 2434 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2435 // behind the number of full collections started. 2436 2437 // This is the case for the inner caller, i.e. a Full GC. 2438 assert(concurrent || 2439 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2440 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2441 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2442 "is inconsistent with _old_marking_cycles_completed = %u", 2443 _old_marking_cycles_started, _old_marking_cycles_completed)); 2444 2445 // This is the case for the outer caller, i.e. the concurrent cycle. 2446 assert(!concurrent || 2447 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2448 err_msg("for outer caller (concurrent cycle): " 2449 "_old_marking_cycles_started = %u " 2450 "is inconsistent with _old_marking_cycles_completed = %u", 2451 _old_marking_cycles_started, _old_marking_cycles_completed)); 2452 2453 _old_marking_cycles_completed += 1; 2454 2455 // We need to clear the "in_progress" flag in the CM thread before 2456 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2457 // is set) so that if a waiter requests another System.gc() it doesn't 2458 // incorrectly see that a marking cyle is still in progress. 2459 if (concurrent) { 2460 _cmThread->clear_in_progress(); 2461 } 2462 2463 // This notify_all() will ensure that a thread that called 2464 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2465 // and it's waiting for a full GC to finish will be woken up. It is 2466 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2467 FullGCCount_lock->notify_all(); 2468 } 2469 2470 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2471 assert_at_safepoint(true /* should_be_vm_thread */); 2472 GCCauseSetter gcs(this, cause); 2473 switch (cause) { 2474 case GCCause::_heap_inspection: 2475 case GCCause::_heap_dump: { 2476 HandleMark hm; 2477 do_full_collection(false); // don't clear all soft refs 2478 break; 2479 } 2480 default: // XXX FIX ME 2481 ShouldNotReachHere(); // Unexpected use of this function 2482 } 2483 } 2484 2485 void G1CollectedHeap::collect(GCCause::Cause cause) { 2486 assert_heap_not_locked(); 2487 2488 unsigned int gc_count_before; 2489 unsigned int old_marking_count_before; 2490 bool retry_gc; 2491 2492 do { 2493 retry_gc = false; 2494 2495 { 2496 MutexLocker ml(Heap_lock); 2497 2498 // Read the GC count while holding the Heap_lock 2499 gc_count_before = total_collections(); 2500 old_marking_count_before = _old_marking_cycles_started; 2501 } 2502 2503 if (should_do_concurrent_full_gc(cause)) { 2504 // Schedule an initial-mark evacuation pause that will start a 2505 // concurrent cycle. We're setting word_size to 0 which means that 2506 // we are not requesting a post-GC allocation. 2507 VM_G1IncCollectionPause op(gc_count_before, 2508 0, /* word_size */ 2509 true, /* should_initiate_conc_mark */ 2510 g1_policy()->max_pause_time_ms(), 2511 cause); 2512 2513 VMThread::execute(&op); 2514 if (!op.pause_succeeded()) { 2515 if (old_marking_count_before == _old_marking_cycles_started) { 2516 retry_gc = op.should_retry_gc(); 2517 } else { 2518 // A Full GC happened while we were trying to schedule the 2519 // initial-mark GC. No point in starting a new cycle given 2520 // that the whole heap was collected anyway. 2521 } 2522 2523 if (retry_gc) { 2524 if (GC_locker::is_active_and_needs_gc()) { 2525 GC_locker::stall_until_clear(); 2526 } 2527 } 2528 } 2529 } else { 2530 if (cause == GCCause::_gc_locker 2531 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2532 2533 // Schedule a standard evacuation pause. We're setting word_size 2534 // to 0 which means that we are not requesting a post-GC allocation. 2535 VM_G1IncCollectionPause op(gc_count_before, 2536 0, /* word_size */ 2537 false, /* should_initiate_conc_mark */ 2538 g1_policy()->max_pause_time_ms(), 2539 cause); 2540 VMThread::execute(&op); 2541 } else { 2542 // Schedule a Full GC. 2543 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2544 VMThread::execute(&op); 2545 } 2546 } 2547 } while (retry_gc); 2548 } 2549 2550 bool G1CollectedHeap::is_in(const void* p) const { 2551 if (_g1_committed.contains(p)) { 2552 // Given that we know that p is in the committed space, 2553 // heap_region_containing_raw() should successfully 2554 // return the containing region. 2555 HeapRegion* hr = heap_region_containing_raw(p); 2556 return hr->is_in(p); 2557 } else { 2558 return _perm_gen->as_gen()->is_in(p); 2559 } 2560 } 2561 2562 // Iteration functions. 2563 2564 // Iterates an OopClosure over all ref-containing fields of objects 2565 // within a HeapRegion. 2566 2567 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2568 MemRegion _mr; 2569 OopClosure* _cl; 2570 public: 2571 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2572 : _mr(mr), _cl(cl) {} 2573 bool doHeapRegion(HeapRegion* r) { 2574 if (!r->continuesHumongous()) { 2575 r->oop_iterate(_cl); 2576 } 2577 return false; 2578 } 2579 }; 2580 2581 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2582 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2583 heap_region_iterate(&blk); 2584 if (do_perm) { 2585 perm_gen()->oop_iterate(cl); 2586 } 2587 } 2588 2589 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2590 IterateOopClosureRegionClosure blk(mr, cl); 2591 heap_region_iterate(&blk); 2592 if (do_perm) { 2593 perm_gen()->oop_iterate(cl); 2594 } 2595 } 2596 2597 // Iterates an ObjectClosure over all objects within a HeapRegion. 2598 2599 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2600 ObjectClosure* _cl; 2601 public: 2602 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2603 bool doHeapRegion(HeapRegion* r) { 2604 if (! r->continuesHumongous()) { 2605 r->object_iterate(_cl); 2606 } 2607 return false; 2608 } 2609 }; 2610 2611 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2612 IterateObjectClosureRegionClosure blk(cl); 2613 heap_region_iterate(&blk); 2614 if (do_perm) { 2615 perm_gen()->object_iterate(cl); 2616 } 2617 } 2618 2619 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2620 // FIXME: is this right? 2621 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2622 } 2623 2624 // Calls a SpaceClosure on a HeapRegion. 2625 2626 class SpaceClosureRegionClosure: public HeapRegionClosure { 2627 SpaceClosure* _cl; 2628 public: 2629 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2630 bool doHeapRegion(HeapRegion* r) { 2631 _cl->do_space(r); 2632 return false; 2633 } 2634 }; 2635 2636 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2637 SpaceClosureRegionClosure blk(cl); 2638 heap_region_iterate(&blk); 2639 } 2640 2641 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2642 _hrs.iterate(cl); 2643 } 2644 2645 void 2646 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2647 uint worker_id, 2648 uint no_of_par_workers, 2649 jint claim_value) { 2650 const uint regions = n_regions(); 2651 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2652 no_of_par_workers : 2653 1); 2654 assert(UseDynamicNumberOfGCThreads || 2655 no_of_par_workers == workers()->total_workers(), 2656 "Non dynamic should use fixed number of workers"); 2657 // try to spread out the starting points of the workers 2658 const HeapRegion* start_hr = 2659 start_region_for_worker(worker_id, no_of_par_workers); 2660 const uint start_index = start_hr->hrs_index(); 2661 2662 // each worker will actually look at all regions 2663 for (uint count = 0; count < regions; ++count) { 2664 const uint index = (start_index + count) % regions; 2665 assert(0 <= index && index < regions, "sanity"); 2666 HeapRegion* r = region_at(index); 2667 // we'll ignore "continues humongous" regions (we'll process them 2668 // when we come across their corresponding "start humongous" 2669 // region) and regions already claimed 2670 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2671 continue; 2672 } 2673 // OK, try to claim it 2674 if (r->claimHeapRegion(claim_value)) { 2675 // success! 2676 assert(!r->continuesHumongous(), "sanity"); 2677 if (r->startsHumongous()) { 2678 // If the region is "starts humongous" we'll iterate over its 2679 // "continues humongous" first; in fact we'll do them 2680 // first. The order is important. In on case, calling the 2681 // closure on the "starts humongous" region might de-allocate 2682 // and clear all its "continues humongous" regions and, as a 2683 // result, we might end up processing them twice. So, we'll do 2684 // them first (notice: most closures will ignore them anyway) and 2685 // then we'll do the "starts humongous" region. 2686 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2687 HeapRegion* chr = region_at(ch_index); 2688 2689 // if the region has already been claimed or it's not 2690 // "continues humongous" we're done 2691 if (chr->claim_value() == claim_value || 2692 !chr->continuesHumongous()) { 2693 break; 2694 } 2695 2696 // Noone should have claimed it directly. We can given 2697 // that we claimed its "starts humongous" region. 2698 assert(chr->claim_value() != claim_value, "sanity"); 2699 assert(chr->humongous_start_region() == r, "sanity"); 2700 2701 if (chr->claimHeapRegion(claim_value)) { 2702 // we should always be able to claim it; noone else should 2703 // be trying to claim this region 2704 2705 bool res2 = cl->doHeapRegion(chr); 2706 assert(!res2, "Should not abort"); 2707 2708 // Right now, this holds (i.e., no closure that actually 2709 // does something with "continues humongous" regions 2710 // clears them). We might have to weaken it in the future, 2711 // but let's leave these two asserts here for extra safety. 2712 assert(chr->continuesHumongous(), "should still be the case"); 2713 assert(chr->humongous_start_region() == r, "sanity"); 2714 } else { 2715 guarantee(false, "we should not reach here"); 2716 } 2717 } 2718 } 2719 2720 assert(!r->continuesHumongous(), "sanity"); 2721 bool res = cl->doHeapRegion(r); 2722 assert(!res, "Should not abort"); 2723 } 2724 } 2725 } 2726 2727 class ResetClaimValuesClosure: public HeapRegionClosure { 2728 public: 2729 bool doHeapRegion(HeapRegion* r) { 2730 r->set_claim_value(HeapRegion::InitialClaimValue); 2731 return false; 2732 } 2733 }; 2734 2735 void G1CollectedHeap::reset_heap_region_claim_values() { 2736 ResetClaimValuesClosure blk; 2737 heap_region_iterate(&blk); 2738 } 2739 2740 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2741 ResetClaimValuesClosure blk; 2742 collection_set_iterate(&blk); 2743 } 2744 2745 #ifdef ASSERT 2746 // This checks whether all regions in the heap have the correct claim 2747 // value. I also piggy-backed on this a check to ensure that the 2748 // humongous_start_region() information on "continues humongous" 2749 // regions is correct. 2750 2751 class CheckClaimValuesClosure : public HeapRegionClosure { 2752 private: 2753 jint _claim_value; 2754 uint _failures; 2755 HeapRegion* _sh_region; 2756 2757 public: 2758 CheckClaimValuesClosure(jint claim_value) : 2759 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2760 bool doHeapRegion(HeapRegion* r) { 2761 if (r->claim_value() != _claim_value) { 2762 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2763 "claim value = %d, should be %d", 2764 HR_FORMAT_PARAMS(r), 2765 r->claim_value(), _claim_value); 2766 ++_failures; 2767 } 2768 if (!r->isHumongous()) { 2769 _sh_region = NULL; 2770 } else if (r->startsHumongous()) { 2771 _sh_region = r; 2772 } else if (r->continuesHumongous()) { 2773 if (r->humongous_start_region() != _sh_region) { 2774 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2775 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2776 HR_FORMAT_PARAMS(r), 2777 r->humongous_start_region(), 2778 _sh_region); 2779 ++_failures; 2780 } 2781 } 2782 return false; 2783 } 2784 uint failures() { return _failures; } 2785 }; 2786 2787 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2788 CheckClaimValuesClosure cl(claim_value); 2789 heap_region_iterate(&cl); 2790 return cl.failures() == 0; 2791 } 2792 2793 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2794 private: 2795 jint _claim_value; 2796 uint _failures; 2797 2798 public: 2799 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2800 _claim_value(claim_value), _failures(0) { } 2801 2802 uint failures() { return _failures; } 2803 2804 bool doHeapRegion(HeapRegion* hr) { 2805 assert(hr->in_collection_set(), "how?"); 2806 assert(!hr->isHumongous(), "H-region in CSet"); 2807 if (hr->claim_value() != _claim_value) { 2808 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2809 "claim value = %d, should be %d", 2810 HR_FORMAT_PARAMS(hr), 2811 hr->claim_value(), _claim_value); 2812 _failures += 1; 2813 } 2814 return false; 2815 } 2816 }; 2817 2818 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2819 CheckClaimValuesInCSetHRClosure cl(claim_value); 2820 collection_set_iterate(&cl); 2821 return cl.failures() == 0; 2822 } 2823 #endif // ASSERT 2824 2825 // Clear the cached CSet starting regions and (more importantly) 2826 // the time stamps. Called when we reset the GC time stamp. 2827 void G1CollectedHeap::clear_cset_start_regions() { 2828 assert(_worker_cset_start_region != NULL, "sanity"); 2829 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2830 2831 int n_queues = MAX2((int)ParallelGCThreads, 1); 2832 for (int i = 0; i < n_queues; i++) { 2833 _worker_cset_start_region[i] = NULL; 2834 _worker_cset_start_region_time_stamp[i] = 0; 2835 } 2836 } 2837 2838 // Given the id of a worker, obtain or calculate a suitable 2839 // starting region for iterating over the current collection set. 2840 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2841 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2842 2843 HeapRegion* result = NULL; 2844 unsigned gc_time_stamp = get_gc_time_stamp(); 2845 2846 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2847 // Cached starting region for current worker was set 2848 // during the current pause - so it's valid. 2849 // Note: the cached starting heap region may be NULL 2850 // (when the collection set is empty). 2851 result = _worker_cset_start_region[worker_i]; 2852 assert(result == NULL || result->in_collection_set(), "sanity"); 2853 return result; 2854 } 2855 2856 // The cached entry was not valid so let's calculate 2857 // a suitable starting heap region for this worker. 2858 2859 // We want the parallel threads to start their collection 2860 // set iteration at different collection set regions to 2861 // avoid contention. 2862 // If we have: 2863 // n collection set regions 2864 // p threads 2865 // Then thread t will start at region floor ((t * n) / p) 2866 2867 result = g1_policy()->collection_set(); 2868 if (G1CollectedHeap::use_parallel_gc_threads()) { 2869 uint cs_size = g1_policy()->cset_region_length(); 2870 uint active_workers = workers()->active_workers(); 2871 assert(UseDynamicNumberOfGCThreads || 2872 active_workers == workers()->total_workers(), 2873 "Unless dynamic should use total workers"); 2874 2875 uint end_ind = (cs_size * worker_i) / active_workers; 2876 uint start_ind = 0; 2877 2878 if (worker_i > 0 && 2879 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2880 // Previous workers starting region is valid 2881 // so let's iterate from there 2882 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2883 result = _worker_cset_start_region[worker_i - 1]; 2884 } 2885 2886 for (uint i = start_ind; i < end_ind; i++) { 2887 result = result->next_in_collection_set(); 2888 } 2889 } 2890 2891 // Note: the calculated starting heap region may be NULL 2892 // (when the collection set is empty). 2893 assert(result == NULL || result->in_collection_set(), "sanity"); 2894 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2895 "should be updated only once per pause"); 2896 _worker_cset_start_region[worker_i] = result; 2897 OrderAccess::storestore(); 2898 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2899 return result; 2900 } 2901 2902 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2903 uint no_of_par_workers) { 2904 uint worker_num = 2905 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2906 assert(UseDynamicNumberOfGCThreads || 2907 no_of_par_workers == workers()->total_workers(), 2908 "Non dynamic should use fixed number of workers"); 2909 const uint start_index = n_regions() * worker_i / worker_num; 2910 return region_at(start_index); 2911 } 2912 2913 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2914 HeapRegion* r = g1_policy()->collection_set(); 2915 while (r != NULL) { 2916 HeapRegion* next = r->next_in_collection_set(); 2917 if (cl->doHeapRegion(r)) { 2918 cl->incomplete(); 2919 return; 2920 } 2921 r = next; 2922 } 2923 } 2924 2925 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2926 HeapRegionClosure *cl) { 2927 if (r == NULL) { 2928 // The CSet is empty so there's nothing to do. 2929 return; 2930 } 2931 2932 assert(r->in_collection_set(), 2933 "Start region must be a member of the collection set."); 2934 HeapRegion* cur = r; 2935 while (cur != NULL) { 2936 HeapRegion* next = cur->next_in_collection_set(); 2937 if (cl->doHeapRegion(cur) && false) { 2938 cl->incomplete(); 2939 return; 2940 } 2941 cur = next; 2942 } 2943 cur = g1_policy()->collection_set(); 2944 while (cur != r) { 2945 HeapRegion* next = cur->next_in_collection_set(); 2946 if (cl->doHeapRegion(cur) && false) { 2947 cl->incomplete(); 2948 return; 2949 } 2950 cur = next; 2951 } 2952 } 2953 2954 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2955 return n_regions() > 0 ? region_at(0) : NULL; 2956 } 2957 2958 2959 Space* G1CollectedHeap::space_containing(const void* addr) const { 2960 Space* res = heap_region_containing(addr); 2961 if (res == NULL) 2962 res = perm_gen()->space_containing(addr); 2963 return res; 2964 } 2965 2966 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2967 Space* sp = space_containing(addr); 2968 if (sp != NULL) { 2969 return sp->block_start(addr); 2970 } 2971 return NULL; 2972 } 2973 2974 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2975 Space* sp = space_containing(addr); 2976 assert(sp != NULL, "block_size of address outside of heap"); 2977 return sp->block_size(addr); 2978 } 2979 2980 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2981 Space* sp = space_containing(addr); 2982 return sp->block_is_obj(addr); 2983 } 2984 2985 bool G1CollectedHeap::supports_tlab_allocation() const { 2986 return true; 2987 } 2988 2989 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2990 return HeapRegion::GrainBytes; 2991 } 2992 2993 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2994 // Return the remaining space in the cur alloc region, but not less than 2995 // the min TLAB size. 2996 2997 // Also, this value can be at most the humongous object threshold, 2998 // since we can't allow tlabs to grow big enough to accomodate 2999 // humongous objects. 3000 3001 HeapRegion* hr = _mutator_alloc_region.get(); 3002 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 3003 if (hr == NULL) { 3004 return max_tlab_size; 3005 } else { 3006 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 3007 } 3008 } 3009 3010 size_t G1CollectedHeap::max_capacity() const { 3011 return _g1_reserved.byte_size(); 3012 } 3013 3014 jlong G1CollectedHeap::millis_since_last_gc() { 3015 // assert(false, "NYI"); 3016 return 0; 3017 } 3018 3019 void G1CollectedHeap::prepare_for_verify() { 3020 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3021 ensure_parsability(false); 3022 } 3023 g1_rem_set()->prepare_for_verify(); 3024 } 3025 3026 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 3027 VerifyOption vo) { 3028 switch (vo) { 3029 case VerifyOption_G1UsePrevMarking: 3030 return hr->obj_allocated_since_prev_marking(obj); 3031 case VerifyOption_G1UseNextMarking: 3032 return hr->obj_allocated_since_next_marking(obj); 3033 case VerifyOption_G1UseMarkWord: 3034 return false; 3035 default: 3036 ShouldNotReachHere(); 3037 } 3038 return false; // keep some compilers happy 3039 } 3040 3041 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3042 switch (vo) { 3043 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3044 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3045 case VerifyOption_G1UseMarkWord: return NULL; 3046 default: ShouldNotReachHere(); 3047 } 3048 return NULL; // keep some compilers happy 3049 } 3050 3051 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3052 switch (vo) { 3053 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3054 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3055 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3056 default: ShouldNotReachHere(); 3057 } 3058 return false; // keep some compilers happy 3059 } 3060 3061 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3062 switch (vo) { 3063 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3064 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3065 case VerifyOption_G1UseMarkWord: return "NONE"; 3066 default: ShouldNotReachHere(); 3067 } 3068 return NULL; // keep some compilers happy 3069 } 3070 3071 class VerifyLivenessOopClosure: public OopClosure { 3072 G1CollectedHeap* _g1h; 3073 VerifyOption _vo; 3074 public: 3075 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3076 _g1h(g1h), _vo(vo) 3077 { } 3078 void do_oop(narrowOop *p) { do_oop_work(p); } 3079 void do_oop( oop *p) { do_oop_work(p); } 3080 3081 template <class T> void do_oop_work(T *p) { 3082 oop obj = oopDesc::load_decode_heap_oop(p); 3083 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3084 "Dead object referenced by a not dead object"); 3085 } 3086 }; 3087 3088 class VerifyObjsInRegionClosure: public ObjectClosure { 3089 private: 3090 G1CollectedHeap* _g1h; 3091 size_t _live_bytes; 3092 HeapRegion *_hr; 3093 VerifyOption _vo; 3094 public: 3095 // _vo == UsePrevMarking -> use "prev" marking information, 3096 // _vo == UseNextMarking -> use "next" marking information, 3097 // _vo == UseMarkWord -> use mark word from object header. 3098 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3099 : _live_bytes(0), _hr(hr), _vo(vo) { 3100 _g1h = G1CollectedHeap::heap(); 3101 } 3102 void do_object(oop o) { 3103 VerifyLivenessOopClosure isLive(_g1h, _vo); 3104 assert(o != NULL, "Huh?"); 3105 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3106 // If the object is alive according to the mark word, 3107 // then verify that the marking information agrees. 3108 // Note we can't verify the contra-positive of the 3109 // above: if the object is dead (according to the mark 3110 // word), it may not be marked, or may have been marked 3111 // but has since became dead, or may have been allocated 3112 // since the last marking. 3113 if (_vo == VerifyOption_G1UseMarkWord) { 3114 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3115 } 3116 3117 o->oop_iterate(&isLive); 3118 if (!_hr->obj_allocated_since_prev_marking(o)) { 3119 size_t obj_size = o->size(); // Make sure we don't overflow 3120 _live_bytes += (obj_size * HeapWordSize); 3121 } 3122 } 3123 } 3124 size_t live_bytes() { return _live_bytes; } 3125 }; 3126 3127 class PrintObjsInRegionClosure : public ObjectClosure { 3128 HeapRegion *_hr; 3129 G1CollectedHeap *_g1; 3130 public: 3131 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3132 _g1 = G1CollectedHeap::heap(); 3133 }; 3134 3135 void do_object(oop o) { 3136 if (o != NULL) { 3137 HeapWord *start = (HeapWord *) o; 3138 size_t word_sz = o->size(); 3139 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3140 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3141 (void*) o, word_sz, 3142 _g1->isMarkedPrev(o), 3143 _g1->isMarkedNext(o), 3144 _hr->obj_allocated_since_prev_marking(o)); 3145 HeapWord *end = start + word_sz; 3146 HeapWord *cur; 3147 int *val; 3148 for (cur = start; cur < end; cur++) { 3149 val = (int *) cur; 3150 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3151 } 3152 } 3153 } 3154 }; 3155 3156 class VerifyRegionClosure: public HeapRegionClosure { 3157 private: 3158 bool _par; 3159 VerifyOption _vo; 3160 bool _failures; 3161 public: 3162 // _vo == UsePrevMarking -> use "prev" marking information, 3163 // _vo == UseNextMarking -> use "next" marking information, 3164 // _vo == UseMarkWord -> use mark word from object header. 3165 VerifyRegionClosure(bool par, VerifyOption vo) 3166 : _par(par), 3167 _vo(vo), 3168 _failures(false) {} 3169 3170 bool failures() { 3171 return _failures; 3172 } 3173 3174 bool doHeapRegion(HeapRegion* r) { 3175 if (!r->continuesHumongous()) { 3176 bool failures = false; 3177 r->verify(_vo, &failures); 3178 if (failures) { 3179 _failures = true; 3180 } else { 3181 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3182 r->object_iterate(¬_dead_yet_cl); 3183 if (_vo != VerifyOption_G1UseNextMarking) { 3184 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3185 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3186 "max_live_bytes "SIZE_FORMAT" " 3187 "< calculated "SIZE_FORMAT, 3188 r->bottom(), r->end(), 3189 r->max_live_bytes(), 3190 not_dead_yet_cl.live_bytes()); 3191 _failures = true; 3192 } 3193 } else { 3194 // When vo == UseNextMarking we cannot currently do a sanity 3195 // check on the live bytes as the calculation has not been 3196 // finalized yet. 3197 } 3198 } 3199 } 3200 return false; // stop the region iteration if we hit a failure 3201 } 3202 }; 3203 3204 class VerifyRootsClosure: public OopsInGenClosure { 3205 private: 3206 G1CollectedHeap* _g1h; 3207 VerifyOption _vo; 3208 bool _failures; 3209 public: 3210 // _vo == UsePrevMarking -> use "prev" marking information, 3211 // _vo == UseNextMarking -> use "next" marking information, 3212 // _vo == UseMarkWord -> use mark word from object header. 3213 VerifyRootsClosure(VerifyOption vo) : 3214 _g1h(G1CollectedHeap::heap()), 3215 _vo(vo), 3216 _failures(false) { } 3217 3218 bool failures() { return _failures; } 3219 3220 template <class T> void do_oop_nv(T* p) { 3221 T heap_oop = oopDesc::load_heap_oop(p); 3222 if (!oopDesc::is_null(heap_oop)) { 3223 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3224 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3225 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3226 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3227 if (_vo == VerifyOption_G1UseMarkWord) { 3228 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3229 } 3230 obj->print_on(gclog_or_tty); 3231 _failures = true; 3232 } 3233 } 3234 } 3235 3236 void do_oop(oop* p) { do_oop_nv(p); } 3237 void do_oop(narrowOop* p) { do_oop_nv(p); } 3238 }; 3239 3240 // This is the task used for parallel heap verification. 3241 3242 class G1ParVerifyTask: public AbstractGangTask { 3243 private: 3244 G1CollectedHeap* _g1h; 3245 VerifyOption _vo; 3246 bool _failures; 3247 3248 public: 3249 // _vo == UsePrevMarking -> use "prev" marking information, 3250 // _vo == UseNextMarking -> use "next" marking information, 3251 // _vo == UseMarkWord -> use mark word from object header. 3252 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3253 AbstractGangTask("Parallel verify task"), 3254 _g1h(g1h), 3255 _vo(vo), 3256 _failures(false) { } 3257 3258 bool failures() { 3259 return _failures; 3260 } 3261 3262 void work(uint worker_id) { 3263 HandleMark hm; 3264 VerifyRegionClosure blk(true, _vo); 3265 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3266 _g1h->workers()->active_workers(), 3267 HeapRegion::ParVerifyClaimValue); 3268 if (blk.failures()) { 3269 _failures = true; 3270 } 3271 } 3272 }; 3273 3274 void G1CollectedHeap::verify(bool silent) { 3275 verify(silent, VerifyOption_G1UsePrevMarking); 3276 } 3277 3278 void G1CollectedHeap::verify(bool silent, 3279 VerifyOption vo) { 3280 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3281 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3282 VerifyRootsClosure rootsCl(vo); 3283 3284 assert(Thread::current()->is_VM_thread(), 3285 "Expected to be executed serially by the VM thread at this point"); 3286 3287 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3288 3289 // We apply the relevant closures to all the oops in the 3290 // system dictionary, the string table and the code cache. 3291 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3292 3293 process_strong_roots(true, // activate StrongRootsScope 3294 true, // we set "collecting perm gen" to true, 3295 // so we don't reset the dirty cards in the perm gen. 3296 ScanningOption(so), // roots scanning options 3297 &rootsCl, 3298 &blobsCl, 3299 &rootsCl); 3300 3301 // If we're verifying after the marking phase of a Full GC then we can't 3302 // treat the perm gen as roots into the G1 heap. Some of the objects in 3303 // the perm gen may be dead and hence not marked. If one of these dead 3304 // objects is considered to be a root then we may end up with a false 3305 // "Root location <x> points to dead ob <y>" failure. 3306 if (vo != VerifyOption_G1UseMarkWord) { 3307 // Since we used "collecting_perm_gen" == true above, we will not have 3308 // checked the refs from perm into the G1-collected heap. We check those 3309 // references explicitly below. Whether the relevant cards are dirty 3310 // is checked further below in the rem set verification. 3311 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3312 perm_gen()->oop_iterate(&rootsCl); 3313 } 3314 bool failures = rootsCl.failures(); 3315 3316 if (vo != VerifyOption_G1UseMarkWord) { 3317 // If we're verifying during a full GC then the region sets 3318 // will have been torn down at the start of the GC. Therefore 3319 // verifying the region sets will fail. So we only verify 3320 // the region sets when not in a full GC. 3321 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3322 verify_region_sets(); 3323 } 3324 3325 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3326 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3327 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3328 "sanity check"); 3329 3330 G1ParVerifyTask task(this, vo); 3331 assert(UseDynamicNumberOfGCThreads || 3332 workers()->active_workers() == workers()->total_workers(), 3333 "If not dynamic should be using all the workers"); 3334 int n_workers = workers()->active_workers(); 3335 set_par_threads(n_workers); 3336 workers()->run_task(&task); 3337 set_par_threads(0); 3338 if (task.failures()) { 3339 failures = true; 3340 } 3341 3342 // Checks that the expected amount of parallel work was done. 3343 // The implication is that n_workers is > 0. 3344 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3345 "sanity check"); 3346 3347 reset_heap_region_claim_values(); 3348 3349 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3350 "sanity check"); 3351 } else { 3352 VerifyRegionClosure blk(false, vo); 3353 heap_region_iterate(&blk); 3354 if (blk.failures()) { 3355 failures = true; 3356 } 3357 } 3358 if (!silent) gclog_or_tty->print("RemSet "); 3359 rem_set()->verify(); 3360 3361 if (failures) { 3362 gclog_or_tty->print_cr("Heap:"); 3363 // It helps to have the per-region information in the output to 3364 // help us track down what went wrong. This is why we call 3365 // print_extended_on() instead of print_on(). 3366 print_extended_on(gclog_or_tty); 3367 gclog_or_tty->print_cr(""); 3368 #ifndef PRODUCT 3369 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3370 concurrent_mark()->print_reachable("at-verification-failure", 3371 vo, false /* all */); 3372 } 3373 #endif 3374 gclog_or_tty->flush(); 3375 } 3376 guarantee(!failures, "there should not have been any failures"); 3377 } else { 3378 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3379 } 3380 } 3381 3382 class PrintRegionClosure: public HeapRegionClosure { 3383 outputStream* _st; 3384 public: 3385 PrintRegionClosure(outputStream* st) : _st(st) {} 3386 bool doHeapRegion(HeapRegion* r) { 3387 r->print_on(_st); 3388 return false; 3389 } 3390 }; 3391 3392 void G1CollectedHeap::print_on(outputStream* st) const { 3393 st->print(" %-20s", "garbage-first heap"); 3394 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3395 capacity()/K, used_unlocked()/K); 3396 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3397 _g1_storage.low_boundary(), 3398 _g1_storage.high(), 3399 _g1_storage.high_boundary()); 3400 st->cr(); 3401 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3402 uint young_regions = _young_list->length(); 3403 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3404 (size_t) young_regions * HeapRegion::GrainBytes / K); 3405 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3406 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3407 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3408 st->cr(); 3409 perm()->as_gen()->print_on(st); 3410 } 3411 3412 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3413 print_on(st); 3414 3415 // Print the per-region information. 3416 st->cr(); 3417 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3418 "HS=humongous(starts), HC=humongous(continues), " 3419 "CS=collection set, F=free, TS=gc time stamp, " 3420 "PTAMS=previous top-at-mark-start, " 3421 "NTAMS=next top-at-mark-start)"); 3422 PrintRegionClosure blk(st); 3423 heap_region_iterate(&blk); 3424 } 3425 3426 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3427 if (G1CollectedHeap::use_parallel_gc_threads()) { 3428 workers()->print_worker_threads_on(st); 3429 } 3430 _cmThread->print_on(st); 3431 st->cr(); 3432 _cm->print_worker_threads_on(st); 3433 _cg1r->print_worker_threads_on(st); 3434 st->cr(); 3435 } 3436 3437 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3438 if (G1CollectedHeap::use_parallel_gc_threads()) { 3439 workers()->threads_do(tc); 3440 } 3441 tc->do_thread(_cmThread); 3442 _cg1r->threads_do(tc); 3443 } 3444 3445 void G1CollectedHeap::print_tracing_info() const { 3446 // We'll overload this to mean "trace GC pause statistics." 3447 if (TraceGen0Time || TraceGen1Time) { 3448 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3449 // to that. 3450 g1_policy()->print_tracing_info(); 3451 } 3452 if (G1SummarizeRSetStats) { 3453 g1_rem_set()->print_summary_info(); 3454 } 3455 if (G1SummarizeConcMark) { 3456 concurrent_mark()->print_summary_info(); 3457 } 3458 g1_policy()->print_yg_surv_rate_info(); 3459 SpecializationStats::print(); 3460 } 3461 3462 #ifndef PRODUCT 3463 // Helpful for debugging RSet issues. 3464 3465 class PrintRSetsClosure : public HeapRegionClosure { 3466 private: 3467 const char* _msg; 3468 size_t _occupied_sum; 3469 3470 public: 3471 bool doHeapRegion(HeapRegion* r) { 3472 HeapRegionRemSet* hrrs = r->rem_set(); 3473 size_t occupied = hrrs->occupied(); 3474 _occupied_sum += occupied; 3475 3476 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3477 HR_FORMAT_PARAMS(r)); 3478 if (occupied == 0) { 3479 gclog_or_tty->print_cr(" RSet is empty"); 3480 } else { 3481 hrrs->print(); 3482 } 3483 gclog_or_tty->print_cr("----------"); 3484 return false; 3485 } 3486 3487 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3488 gclog_or_tty->cr(); 3489 gclog_or_tty->print_cr("========================================"); 3490 gclog_or_tty->print_cr(msg); 3491 gclog_or_tty->cr(); 3492 } 3493 3494 ~PrintRSetsClosure() { 3495 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3496 gclog_or_tty->print_cr("========================================"); 3497 gclog_or_tty->cr(); 3498 } 3499 }; 3500 3501 void G1CollectedHeap::print_cset_rsets() { 3502 PrintRSetsClosure cl("Printing CSet RSets"); 3503 collection_set_iterate(&cl); 3504 } 3505 3506 void G1CollectedHeap::print_all_rsets() { 3507 PrintRSetsClosure cl("Printing All RSets");; 3508 heap_region_iterate(&cl); 3509 } 3510 #endif // PRODUCT 3511 3512 G1CollectedHeap* G1CollectedHeap::heap() { 3513 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3514 "not a garbage-first heap"); 3515 return _g1h; 3516 } 3517 3518 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3519 // always_do_update_barrier = false; 3520 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3521 // Call allocation profiler 3522 AllocationProfiler::iterate_since_last_gc(); 3523 // Fill TLAB's and such 3524 ensure_parsability(true); 3525 } 3526 3527 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3528 // FIXME: what is this about? 3529 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3530 // is set. 3531 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3532 "derived pointer present")); 3533 // always_do_update_barrier = true; 3534 3535 // We have just completed a GC. Update the soft reference 3536 // policy with the new heap occupancy 3537 Universe::update_heap_info_at_gc(); 3538 } 3539 3540 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3541 unsigned int gc_count_before, 3542 bool* succeeded) { 3543 assert_heap_not_locked_and_not_at_safepoint(); 3544 g1_policy()->record_stop_world_start(); 3545 VM_G1IncCollectionPause op(gc_count_before, 3546 word_size, 3547 false, /* should_initiate_conc_mark */ 3548 g1_policy()->max_pause_time_ms(), 3549 GCCause::_g1_inc_collection_pause); 3550 VMThread::execute(&op); 3551 3552 HeapWord* result = op.result(); 3553 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3554 assert(result == NULL || ret_succeeded, 3555 "the result should be NULL if the VM did not succeed"); 3556 *succeeded = ret_succeeded; 3557 3558 assert_heap_not_locked(); 3559 return result; 3560 } 3561 3562 void 3563 G1CollectedHeap::doConcurrentMark() { 3564 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3565 if (!_cmThread->in_progress()) { 3566 _cmThread->set_started(); 3567 CGC_lock->notify(); 3568 } 3569 } 3570 3571 size_t G1CollectedHeap::pending_card_num() { 3572 size_t extra_cards = 0; 3573 JavaThread *curr = Threads::first(); 3574 while (curr != NULL) { 3575 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3576 extra_cards += dcq.size(); 3577 curr = curr->next(); 3578 } 3579 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3580 size_t buffer_size = dcqs.buffer_size(); 3581 size_t buffer_num = dcqs.completed_buffers_num(); 3582 return buffer_size * buffer_num + extra_cards; 3583 } 3584 3585 size_t G1CollectedHeap::max_pending_card_num() { 3586 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3587 size_t buffer_size = dcqs.buffer_size(); 3588 size_t buffer_num = dcqs.completed_buffers_num(); 3589 int thread_num = Threads::number_of_threads(); 3590 return (buffer_num + thread_num) * buffer_size; 3591 } 3592 3593 size_t G1CollectedHeap::cards_scanned() { 3594 return g1_rem_set()->cardsScanned(); 3595 } 3596 3597 void 3598 G1CollectedHeap::setup_surviving_young_words() { 3599 assert(_surviving_young_words == NULL, "pre-condition"); 3600 uint array_length = g1_policy()->young_cset_region_length(); 3601 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length); 3602 if (_surviving_young_words == NULL) { 3603 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3604 "Not enough space for young surv words summary."); 3605 } 3606 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3607 #ifdef ASSERT 3608 for (uint i = 0; i < array_length; ++i) { 3609 assert( _surviving_young_words[i] == 0, "memset above" ); 3610 } 3611 #endif // !ASSERT 3612 } 3613 3614 void 3615 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3616 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3617 uint array_length = g1_policy()->young_cset_region_length(); 3618 for (uint i = 0; i < array_length; ++i) { 3619 _surviving_young_words[i] += surv_young_words[i]; 3620 } 3621 } 3622 3623 void 3624 G1CollectedHeap::cleanup_surviving_young_words() { 3625 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3626 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words); 3627 _surviving_young_words = NULL; 3628 } 3629 3630 #ifdef ASSERT 3631 class VerifyCSetClosure: public HeapRegionClosure { 3632 public: 3633 bool doHeapRegion(HeapRegion* hr) { 3634 // Here we check that the CSet region's RSet is ready for parallel 3635 // iteration. The fields that we'll verify are only manipulated 3636 // when the region is part of a CSet and is collected. Afterwards, 3637 // we reset these fields when we clear the region's RSet (when the 3638 // region is freed) so they are ready when the region is 3639 // re-allocated. The only exception to this is if there's an 3640 // evacuation failure and instead of freeing the region we leave 3641 // it in the heap. In that case, we reset these fields during 3642 // evacuation failure handling. 3643 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3644 3645 // Here's a good place to add any other checks we'd like to 3646 // perform on CSet regions. 3647 return false; 3648 } 3649 }; 3650 #endif // ASSERT 3651 3652 #if TASKQUEUE_STATS 3653 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3654 st->print_raw_cr("GC Task Stats"); 3655 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3656 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3657 } 3658 3659 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3660 print_taskqueue_stats_hdr(st); 3661 3662 TaskQueueStats totals; 3663 const int n = workers() != NULL ? workers()->total_workers() : 1; 3664 for (int i = 0; i < n; ++i) { 3665 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3666 totals += task_queue(i)->stats; 3667 } 3668 st->print_raw("tot "); totals.print(st); st->cr(); 3669 3670 DEBUG_ONLY(totals.verify()); 3671 } 3672 3673 void G1CollectedHeap::reset_taskqueue_stats() { 3674 const int n = workers() != NULL ? workers()->total_workers() : 1; 3675 for (int i = 0; i < n; ++i) { 3676 task_queue(i)->stats.reset(); 3677 } 3678 } 3679 #endif // TASKQUEUE_STATS 3680 3681 bool 3682 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3683 assert_at_safepoint(true /* should_be_vm_thread */); 3684 guarantee(!is_gc_active(), "collection is not reentrant"); 3685 3686 if (GC_locker::check_active_before_gc()) { 3687 return false; 3688 } 3689 3690 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3691 ResourceMark rm; 3692 3693 print_heap_before_gc(); 3694 3695 HRSPhaseSetter x(HRSPhaseEvacuation); 3696 verify_region_sets_optional(); 3697 verify_dirty_young_regions(); 3698 3699 // This call will decide whether this pause is an initial-mark 3700 // pause. If it is, during_initial_mark_pause() will return true 3701 // for the duration of this pause. 3702 g1_policy()->decide_on_conc_mark_initiation(); 3703 3704 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3705 assert(!g1_policy()->during_initial_mark_pause() || 3706 g1_policy()->gcs_are_young(), "sanity"); 3707 3708 // We also do not allow mixed GCs during marking. 3709 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3710 3711 // Record whether this pause is an initial mark. When the current 3712 // thread has completed its logging output and it's safe to signal 3713 // the CM thread, the flag's value in the policy has been reset. 3714 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3715 3716 // Inner scope for scope based logging, timers, and stats collection 3717 { 3718 if (g1_policy()->during_initial_mark_pause()) { 3719 // We are about to start a marking cycle, so we increment the 3720 // full collection counter. 3721 increment_old_marking_cycles_started(); 3722 } 3723 // if the log level is "finer" is on, we'll print long statistics information 3724 // in the collector policy code, so let's not print this as the output 3725 // is messy if we do. 3726 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 3727 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3728 3729 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3730 .append(g1_policy()->gcs_are_young() ? " (young)" : " (mixed)") 3731 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3732 TraceTime t(gc_cause_str, G1Log::fine() && !G1Log::finer(), true, gclog_or_tty); 3733 3734 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3735 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3736 3737 // If the secondary_free_list is not empty, append it to the 3738 // free_list. No need to wait for the cleanup operation to finish; 3739 // the region allocation code will check the secondary_free_list 3740 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3741 // set, skip this step so that the region allocation code has to 3742 // get entries from the secondary_free_list. 3743 if (!G1StressConcRegionFreeing) { 3744 append_secondary_free_list_if_not_empty_with_lock(); 3745 } 3746 3747 assert(check_young_list_well_formed(), 3748 "young list should be well formed"); 3749 3750 // Don't dynamically change the number of GC threads this early. A value of 3751 // 0 is used to indicate serial work. When parallel work is done, 3752 // it will be set. 3753 3754 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3755 IsGCActiveMark x; 3756 3757 gc_prologue(false); 3758 increment_total_collections(false /* full gc */); 3759 increment_gc_time_stamp(); 3760 3761 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3762 HandleMark hm; // Discard invalid handles created during verification 3763 gclog_or_tty->print(" VerifyBeforeGC:"); 3764 prepare_for_verify(); 3765 Universe::verify(/* silent */ false, 3766 /* option */ VerifyOption_G1UsePrevMarking); 3767 } 3768 3769 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3770 3771 // Please see comment in g1CollectedHeap.hpp and 3772 // G1CollectedHeap::ref_processing_init() to see how 3773 // reference processing currently works in G1. 3774 3775 // Enable discovery in the STW reference processor 3776 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3777 true /*verify_no_refs*/); 3778 3779 { 3780 // We want to temporarily turn off discovery by the 3781 // CM ref processor, if necessary, and turn it back on 3782 // on again later if we do. Using a scoped 3783 // NoRefDiscovery object will do this. 3784 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3785 3786 // Forget the current alloc region (we might even choose it to be part 3787 // of the collection set!). 3788 release_mutator_alloc_region(); 3789 3790 // We should call this after we retire the mutator alloc 3791 // region(s) so that all the ALLOC / RETIRE events are generated 3792 // before the start GC event. 3793 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3794 3795 // The elapsed time induced by the start time below deliberately elides 3796 // the possible verification above. 3797 double start_time_sec = os::elapsedTime(); 3798 size_t start_used_bytes = used(); 3799 3800 #if YOUNG_LIST_VERBOSE 3801 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3802 _young_list->print(); 3803 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3804 #endif // YOUNG_LIST_VERBOSE 3805 3806 g1_policy()->record_collection_pause_start(start_time_sec, 3807 start_used_bytes); 3808 3809 double scan_wait_start = os::elapsedTime(); 3810 // We have to wait until the CM threads finish scanning the 3811 // root regions as it's the only way to ensure that all the 3812 // objects on them have been correctly scanned before we start 3813 // moving them during the GC. 3814 bool waited = _cm->root_regions()->wait_until_scan_finished(); 3815 if (waited) { 3816 double scan_wait_end = os::elapsedTime(); 3817 double wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 3818 g1_policy()->record_root_region_scan_wait_time(wait_time_ms); 3819 } 3820 3821 #if YOUNG_LIST_VERBOSE 3822 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3823 _young_list->print(); 3824 #endif // YOUNG_LIST_VERBOSE 3825 3826 if (g1_policy()->during_initial_mark_pause()) { 3827 concurrent_mark()->checkpointRootsInitialPre(); 3828 } 3829 perm_gen()->save_marks(); 3830 3831 #if YOUNG_LIST_VERBOSE 3832 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3833 _young_list->print(); 3834 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3835 #endif // YOUNG_LIST_VERBOSE 3836 3837 g1_policy()->finalize_cset(target_pause_time_ms); 3838 3839 _cm->note_start_of_gc(); 3840 // We should not verify the per-thread SATB buffers given that 3841 // we have not filtered them yet (we'll do so during the 3842 // GC). We also call this after finalize_cset() to 3843 // ensure that the CSet has been finalized. 3844 _cm->verify_no_cset_oops(true /* verify_stacks */, 3845 true /* verify_enqueued_buffers */, 3846 false /* verify_thread_buffers */, 3847 true /* verify_fingers */); 3848 3849 if (_hr_printer.is_active()) { 3850 HeapRegion* hr = g1_policy()->collection_set(); 3851 while (hr != NULL) { 3852 G1HRPrinter::RegionType type; 3853 if (!hr->is_young()) { 3854 type = G1HRPrinter::Old; 3855 } else if (hr->is_survivor()) { 3856 type = G1HRPrinter::Survivor; 3857 } else { 3858 type = G1HRPrinter::Eden; 3859 } 3860 _hr_printer.cset(hr); 3861 hr = hr->next_in_collection_set(); 3862 } 3863 } 3864 3865 #ifdef ASSERT 3866 VerifyCSetClosure cl; 3867 collection_set_iterate(&cl); 3868 #endif // ASSERT 3869 3870 setup_surviving_young_words(); 3871 3872 // Initialize the GC alloc regions. 3873 init_gc_alloc_regions(); 3874 3875 // Actually do the work... 3876 evacuate_collection_set(); 3877 3878 // We do this to mainly verify the per-thread SATB buffers 3879 // (which have been filtered by now) since we didn't verify 3880 // them earlier. No point in re-checking the stacks / enqueued 3881 // buffers given that the CSet has not changed since last time 3882 // we checked. 3883 _cm->verify_no_cset_oops(false /* verify_stacks */, 3884 false /* verify_enqueued_buffers */, 3885 true /* verify_thread_buffers */, 3886 true /* verify_fingers */); 3887 3888 free_collection_set(g1_policy()->collection_set()); 3889 g1_policy()->clear_collection_set(); 3890 3891 cleanup_surviving_young_words(); 3892 3893 // Start a new incremental collection set for the next pause. 3894 g1_policy()->start_incremental_cset_building(); 3895 3896 // Clear the _cset_fast_test bitmap in anticipation of adding 3897 // regions to the incremental collection set for the next 3898 // evacuation pause. 3899 clear_cset_fast_test(); 3900 3901 _young_list->reset_sampled_info(); 3902 3903 // Don't check the whole heap at this point as the 3904 // GC alloc regions from this pause have been tagged 3905 // as survivors and moved on to the survivor list. 3906 // Survivor regions will fail the !is_young() check. 3907 assert(check_young_list_empty(false /* check_heap */), 3908 "young list should be empty"); 3909 3910 #if YOUNG_LIST_VERBOSE 3911 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3912 _young_list->print(); 3913 #endif // YOUNG_LIST_VERBOSE 3914 3915 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3916 _young_list->first_survivor_region(), 3917 _young_list->last_survivor_region()); 3918 3919 _young_list->reset_auxilary_lists(); 3920 3921 if (evacuation_failed()) { 3922 _summary_bytes_used = recalculate_used(); 3923 } else { 3924 // The "used" of the the collection set have already been subtracted 3925 // when they were freed. Add in the bytes evacuated. 3926 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3927 } 3928 3929 if (g1_policy()->during_initial_mark_pause()) { 3930 // We have to do this before we notify the CM threads that 3931 // they can start working to make sure that all the 3932 // appropriate initialization is done on the CM object. 3933 concurrent_mark()->checkpointRootsInitialPost(); 3934 set_marking_started(); 3935 // Note that we don't actually trigger the CM thread at 3936 // this point. We do that later when we're sure that 3937 // the current thread has completed its logging output. 3938 } 3939 3940 allocate_dummy_regions(); 3941 3942 #if YOUNG_LIST_VERBOSE 3943 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3944 _young_list->print(); 3945 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3946 #endif // YOUNG_LIST_VERBOSE 3947 3948 init_mutator_alloc_region(); 3949 3950 { 3951 size_t expand_bytes = g1_policy()->expansion_amount(); 3952 if (expand_bytes > 0) { 3953 size_t bytes_before = capacity(); 3954 // No need for an ergo verbose message here, 3955 // expansion_amount() does this when it returns a value > 0. 3956 if (!expand(expand_bytes)) { 3957 // We failed to expand the heap so let's verify that 3958 // committed/uncommitted amount match the backing store 3959 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3960 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3961 } 3962 } 3963 } 3964 3965 // We redo the verificaiton but now wrt to the new CSet which 3966 // has just got initialized after the previous CSet was freed. 3967 _cm->verify_no_cset_oops(true /* verify_stacks */, 3968 true /* verify_enqueued_buffers */, 3969 true /* verify_thread_buffers */, 3970 true /* verify_fingers */); 3971 _cm->note_end_of_gc(); 3972 3973 double end_time_sec = os::elapsedTime(); 3974 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS; 3975 g1_policy()->record_pause_time_ms(pause_time_ms); 3976 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3977 workers()->active_workers() : 1); 3978 g1_policy()->record_collection_pause_end(active_workers); 3979 3980 MemoryService::track_memory_usage(); 3981 3982 // In prepare_for_verify() below we'll need to scan the deferred 3983 // update buffers to bring the RSets up-to-date if 3984 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3985 // the update buffers we'll probably need to scan cards on the 3986 // regions we just allocated to (i.e., the GC alloc 3987 // regions). However, during the last GC we called 3988 // set_saved_mark() on all the GC alloc regions, so card 3989 // scanning might skip the [saved_mark_word()...top()] area of 3990 // those regions (i.e., the area we allocated objects into 3991 // during the last GC). But it shouldn't. Given that 3992 // saved_mark_word() is conditional on whether the GC time stamp 3993 // on the region is current or not, by incrementing the GC time 3994 // stamp here we invalidate all the GC time stamps on all the 3995 // regions and saved_mark_word() will simply return top() for 3996 // all the regions. This is a nicer way of ensuring this rather 3997 // than iterating over the regions and fixing them. In fact, the 3998 // GC time stamp increment here also ensures that 3999 // saved_mark_word() will return top() between pauses, i.e., 4000 // during concurrent refinement. So we don't need the 4001 // is_gc_active() check to decided which top to use when 4002 // scanning cards (see CR 7039627). 4003 increment_gc_time_stamp(); 4004 4005 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 4006 HandleMark hm; // Discard invalid handles created during verification 4007 gclog_or_tty->print(" VerifyAfterGC:"); 4008 prepare_for_verify(); 4009 Universe::verify(/* silent */ false, 4010 /* option */ VerifyOption_G1UsePrevMarking); 4011 } 4012 4013 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4014 ref_processor_stw()->verify_no_references_recorded(); 4015 4016 // CM reference discovery will be re-enabled if necessary. 4017 } 4018 4019 // We should do this after we potentially expand the heap so 4020 // that all the COMMIT events are generated before the end GC 4021 // event, and after we retire the GC alloc regions so that all 4022 // RETIRE events are generated before the end GC event. 4023 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4024 4025 // We have to do this after we decide whether to expand the heap or not. 4026 g1_policy()->print_heap_transition(); 4027 4028 if (mark_in_progress()) { 4029 concurrent_mark()->update_g1_committed(); 4030 } 4031 4032 #ifdef TRACESPINNING 4033 ParallelTaskTerminator::print_termination_counts(); 4034 #endif 4035 4036 gc_epilogue(false); 4037 } 4038 4039 // The closing of the inner scope, immediately above, will complete 4040 // logging at the "fine" level. The record_collection_pause_end() call 4041 // above will complete logging at the "finer" level. 4042 // 4043 // It is not yet to safe, however, to tell the concurrent mark to 4044 // start as we have some optional output below. We don't want the 4045 // output from the concurrent mark thread interfering with this 4046 // logging output either. 4047 4048 _hrs.verify_optional(); 4049 verify_region_sets_optional(); 4050 4051 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4052 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4053 4054 print_heap_after_gc(); 4055 4056 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4057 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4058 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4059 // before any GC notifications are raised. 4060 g1mm()->update_sizes(); 4061 } 4062 4063 if (G1SummarizeRSetStats && 4064 (G1SummarizeRSetStatsPeriod > 0) && 4065 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 4066 g1_rem_set()->print_summary_info(); 4067 } 4068 4069 // It should now be safe to tell the concurrent mark thread to start 4070 // without its logging output interfering with the logging output 4071 // that came from the pause. 4072 4073 if (should_start_conc_mark) { 4074 // CAUTION: after the doConcurrentMark() call below, 4075 // the concurrent marking thread(s) could be running 4076 // concurrently with us. Make sure that anything after 4077 // this point does not assume that we are the only GC thread 4078 // running. Note: of course, the actual marking work will 4079 // not start until the safepoint itself is released in 4080 // ConcurrentGCThread::safepoint_desynchronize(). 4081 doConcurrentMark(); 4082 } 4083 4084 return true; 4085 } 4086 4087 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4088 { 4089 size_t gclab_word_size; 4090 switch (purpose) { 4091 case GCAllocForSurvived: 4092 gclab_word_size = YoungPLABSize; 4093 break; 4094 case GCAllocForTenured: 4095 gclab_word_size = OldPLABSize; 4096 break; 4097 default: 4098 assert(false, "unknown GCAllocPurpose"); 4099 gclab_word_size = OldPLABSize; 4100 break; 4101 } 4102 return gclab_word_size; 4103 } 4104 4105 void G1CollectedHeap::init_mutator_alloc_region() { 4106 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4107 _mutator_alloc_region.init(); 4108 } 4109 4110 void G1CollectedHeap::release_mutator_alloc_region() { 4111 _mutator_alloc_region.release(); 4112 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4113 } 4114 4115 void G1CollectedHeap::init_gc_alloc_regions() { 4116 assert_at_safepoint(true /* should_be_vm_thread */); 4117 4118 _survivor_gc_alloc_region.init(); 4119 _old_gc_alloc_region.init(); 4120 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4121 _retained_old_gc_alloc_region = NULL; 4122 4123 // We will discard the current GC alloc region if: 4124 // a) it's in the collection set (it can happen!), 4125 // b) it's already full (no point in using it), 4126 // c) it's empty (this means that it was emptied during 4127 // a cleanup and it should be on the free list now), or 4128 // d) it's humongous (this means that it was emptied 4129 // during a cleanup and was added to the free list, but 4130 // has been subseqently used to allocate a humongous 4131 // object that may be less than the region size). 4132 if (retained_region != NULL && 4133 !retained_region->in_collection_set() && 4134 !(retained_region->top() == retained_region->end()) && 4135 !retained_region->is_empty() && 4136 !retained_region->isHumongous()) { 4137 retained_region->set_saved_mark(); 4138 // The retained region was added to the old region set when it was 4139 // retired. We have to remove it now, since we don't allow regions 4140 // we allocate to in the region sets. We'll re-add it later, when 4141 // it's retired again. 4142 _old_set.remove(retained_region); 4143 bool during_im = g1_policy()->during_initial_mark_pause(); 4144 retained_region->note_start_of_copying(during_im); 4145 _old_gc_alloc_region.set(retained_region); 4146 _hr_printer.reuse(retained_region); 4147 } 4148 } 4149 4150 void G1CollectedHeap::release_gc_alloc_regions() { 4151 _survivor_gc_alloc_region.release(); 4152 // If we have an old GC alloc region to release, we'll save it in 4153 // _retained_old_gc_alloc_region. If we don't 4154 // _retained_old_gc_alloc_region will become NULL. This is what we 4155 // want either way so no reason to check explicitly for either 4156 // condition. 4157 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4158 } 4159 4160 void G1CollectedHeap::abandon_gc_alloc_regions() { 4161 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4162 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4163 _retained_old_gc_alloc_region = NULL; 4164 } 4165 4166 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4167 _drain_in_progress = false; 4168 set_evac_failure_closure(cl); 4169 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4170 } 4171 4172 void G1CollectedHeap::finalize_for_evac_failure() { 4173 assert(_evac_failure_scan_stack != NULL && 4174 _evac_failure_scan_stack->length() == 0, 4175 "Postcondition"); 4176 assert(!_drain_in_progress, "Postcondition"); 4177 delete _evac_failure_scan_stack; 4178 _evac_failure_scan_stack = NULL; 4179 } 4180 4181 void G1CollectedHeap::remove_self_forwarding_pointers() { 4182 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4183 4184 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4185 4186 if (G1CollectedHeap::use_parallel_gc_threads()) { 4187 set_par_threads(); 4188 workers()->run_task(&rsfp_task); 4189 set_par_threads(0); 4190 } else { 4191 rsfp_task.work(0); 4192 } 4193 4194 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4195 4196 // Reset the claim values in the regions in the collection set. 4197 reset_cset_heap_region_claim_values(); 4198 4199 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4200 4201 // Now restore saved marks, if any. 4202 if (_objs_with_preserved_marks != NULL) { 4203 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4204 guarantee(_objs_with_preserved_marks->length() == 4205 _preserved_marks_of_objs->length(), "Both or none."); 4206 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4207 oop obj = _objs_with_preserved_marks->at(i); 4208 markOop m = _preserved_marks_of_objs->at(i); 4209 obj->set_mark(m); 4210 } 4211 4212 // Delete the preserved marks growable arrays (allocated on the C heap). 4213 delete _objs_with_preserved_marks; 4214 delete _preserved_marks_of_objs; 4215 _objs_with_preserved_marks = NULL; 4216 _preserved_marks_of_objs = NULL; 4217 } 4218 } 4219 4220 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4221 _evac_failure_scan_stack->push(obj); 4222 } 4223 4224 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4225 assert(_evac_failure_scan_stack != NULL, "precondition"); 4226 4227 while (_evac_failure_scan_stack->length() > 0) { 4228 oop obj = _evac_failure_scan_stack->pop(); 4229 _evac_failure_closure->set_region(heap_region_containing(obj)); 4230 obj->oop_iterate_backwards(_evac_failure_closure); 4231 } 4232 } 4233 4234 oop 4235 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4236 oop old) { 4237 assert(obj_in_cs(old), 4238 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4239 (HeapWord*) old)); 4240 markOop m = old->mark(); 4241 oop forward_ptr = old->forward_to_atomic(old); 4242 if (forward_ptr == NULL) { 4243 // Forward-to-self succeeded. 4244 4245 if (_evac_failure_closure != cl) { 4246 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4247 assert(!_drain_in_progress, 4248 "Should only be true while someone holds the lock."); 4249 // Set the global evac-failure closure to the current thread's. 4250 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4251 set_evac_failure_closure(cl); 4252 // Now do the common part. 4253 handle_evacuation_failure_common(old, m); 4254 // Reset to NULL. 4255 set_evac_failure_closure(NULL); 4256 } else { 4257 // The lock is already held, and this is recursive. 4258 assert(_drain_in_progress, "This should only be the recursive case."); 4259 handle_evacuation_failure_common(old, m); 4260 } 4261 return old; 4262 } else { 4263 // Forward-to-self failed. Either someone else managed to allocate 4264 // space for this object (old != forward_ptr) or they beat us in 4265 // self-forwarding it (old == forward_ptr). 4266 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4267 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4268 "should not be in the CSet", 4269 (HeapWord*) old, (HeapWord*) forward_ptr)); 4270 return forward_ptr; 4271 } 4272 } 4273 4274 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4275 set_evacuation_failed(true); 4276 4277 preserve_mark_if_necessary(old, m); 4278 4279 HeapRegion* r = heap_region_containing(old); 4280 if (!r->evacuation_failed()) { 4281 r->set_evacuation_failed(true); 4282 _hr_printer.evac_failure(r); 4283 } 4284 4285 push_on_evac_failure_scan_stack(old); 4286 4287 if (!_drain_in_progress) { 4288 // prevent recursion in copy_to_survivor_space() 4289 _drain_in_progress = true; 4290 drain_evac_failure_scan_stack(); 4291 _drain_in_progress = false; 4292 } 4293 } 4294 4295 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4296 assert(evacuation_failed(), "Oversaving!"); 4297 // We want to call the "for_promotion_failure" version only in the 4298 // case of a promotion failure. 4299 if (m->must_be_preserved_for_promotion_failure(obj)) { 4300 if (_objs_with_preserved_marks == NULL) { 4301 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4302 _objs_with_preserved_marks = 4303 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4304 _preserved_marks_of_objs = 4305 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true); 4306 } 4307 _objs_with_preserved_marks->push(obj); 4308 _preserved_marks_of_objs->push(m); 4309 } 4310 } 4311 4312 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4313 size_t word_size) { 4314 if (purpose == GCAllocForSurvived) { 4315 HeapWord* result = survivor_attempt_allocation(word_size); 4316 if (result != NULL) { 4317 return result; 4318 } else { 4319 // Let's try to allocate in the old gen in case we can fit the 4320 // object there. 4321 return old_attempt_allocation(word_size); 4322 } 4323 } else { 4324 assert(purpose == GCAllocForTenured, "sanity"); 4325 HeapWord* result = old_attempt_allocation(word_size); 4326 if (result != NULL) { 4327 return result; 4328 } else { 4329 // Let's try to allocate in the survivors in case we can fit the 4330 // object there. 4331 return survivor_attempt_allocation(word_size); 4332 } 4333 } 4334 4335 ShouldNotReachHere(); 4336 // Trying to keep some compilers happy. 4337 return NULL; 4338 } 4339 4340 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4341 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4342 4343 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4344 : _g1h(g1h), 4345 _refs(g1h->task_queue(queue_num)), 4346 _dcq(&g1h->dirty_card_queue_set()), 4347 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4348 _g1_rem(g1h->g1_rem_set()), 4349 _hash_seed(17), _queue_num(queue_num), 4350 _term_attempts(0), 4351 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4352 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4353 _age_table(false), 4354 _strong_roots_time(0), _term_time(0), 4355 _alloc_buffer_waste(0), _undo_waste(0) { 4356 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4357 // we "sacrifice" entry 0 to keep track of surviving bytes for 4358 // non-young regions (where the age is -1) 4359 // We also add a few elements at the beginning and at the end in 4360 // an attempt to eliminate cache contention 4361 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4362 uint array_length = PADDING_ELEM_NUM + 4363 real_length + 4364 PADDING_ELEM_NUM; 4365 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length); 4366 if (_surviving_young_words_base == NULL) 4367 vm_exit_out_of_memory(array_length * sizeof(size_t), 4368 "Not enough space for young surv histo."); 4369 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4370 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4371 4372 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4373 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4374 4375 _start = os::elapsedTime(); 4376 } 4377 4378 void 4379 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4380 { 4381 st->print_raw_cr("GC Termination Stats"); 4382 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4383 " ------waste (KiB)------"); 4384 st->print_raw_cr("thr ms ms % ms % attempts" 4385 " total alloc undo"); 4386 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4387 " ------- ------- -------"); 4388 } 4389 4390 void 4391 G1ParScanThreadState::print_termination_stats(int i, 4392 outputStream* const st) const 4393 { 4394 const double elapsed_ms = elapsed_time() * 1000.0; 4395 const double s_roots_ms = strong_roots_time() * 1000.0; 4396 const double term_ms = term_time() * 1000.0; 4397 st->print_cr("%3d %9.2f %9.2f %6.2f " 4398 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4399 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4400 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4401 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4402 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4403 alloc_buffer_waste() * HeapWordSize / K, 4404 undo_waste() * HeapWordSize / K); 4405 } 4406 4407 #ifdef ASSERT 4408 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4409 assert(ref != NULL, "invariant"); 4410 assert(UseCompressedOops, "sanity"); 4411 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4412 oop p = oopDesc::load_decode_heap_oop(ref); 4413 assert(_g1h->is_in_g1_reserved(p), 4414 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4415 return true; 4416 } 4417 4418 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4419 assert(ref != NULL, "invariant"); 4420 if (has_partial_array_mask(ref)) { 4421 // Must be in the collection set--it's already been copied. 4422 oop p = clear_partial_array_mask(ref); 4423 assert(_g1h->obj_in_cs(p), 4424 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4425 } else { 4426 oop p = oopDesc::load_decode_heap_oop(ref); 4427 assert(_g1h->is_in_g1_reserved(p), 4428 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4429 } 4430 return true; 4431 } 4432 4433 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4434 if (ref.is_narrow()) { 4435 return verify_ref((narrowOop*) ref); 4436 } else { 4437 return verify_ref((oop*) ref); 4438 } 4439 } 4440 #endif // ASSERT 4441 4442 void G1ParScanThreadState::trim_queue() { 4443 assert(_evac_cl != NULL, "not set"); 4444 assert(_evac_failure_cl != NULL, "not set"); 4445 assert(_partial_scan_cl != NULL, "not set"); 4446 4447 StarTask ref; 4448 do { 4449 // Drain the overflow stack first, so other threads can steal. 4450 while (refs()->pop_overflow(ref)) { 4451 deal_with_reference(ref); 4452 } 4453 4454 while (refs()->pop_local(ref)) { 4455 deal_with_reference(ref); 4456 } 4457 } while (!refs()->is_empty()); 4458 } 4459 4460 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4461 G1ParScanThreadState* par_scan_state) : 4462 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4463 _par_scan_state(par_scan_state), 4464 _worker_id(par_scan_state->queue_num()), 4465 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4466 _mark_in_progress(_g1->mark_in_progress()) { } 4467 4468 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4469 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4470 #ifdef ASSERT 4471 HeapRegion* hr = _g1->heap_region_containing(obj); 4472 assert(hr != NULL, "sanity"); 4473 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4474 #endif // ASSERT 4475 4476 // We know that the object is not moving so it's safe to read its size. 4477 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4478 } 4479 4480 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4481 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4482 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4483 #ifdef ASSERT 4484 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4485 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4486 assert(from_obj != to_obj, "should not be self-forwarded"); 4487 4488 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4489 assert(from_hr != NULL, "sanity"); 4490 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4491 4492 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4493 assert(to_hr != NULL, "sanity"); 4494 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4495 #endif // ASSERT 4496 4497 // The object might be in the process of being copied by another 4498 // worker so we cannot trust that its to-space image is 4499 // well-formed. So we have to read its size from its from-space 4500 // image which we know should not be changing. 4501 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4502 } 4503 4504 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4505 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4506 ::copy_to_survivor_space(oop old) { 4507 size_t word_sz = old->size(); 4508 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4509 // +1 to make the -1 indexes valid... 4510 int young_index = from_region->young_index_in_cset()+1; 4511 assert( (from_region->is_young() && young_index > 0) || 4512 (!from_region->is_young() && young_index == 0), "invariant" ); 4513 G1CollectorPolicy* g1p = _g1->g1_policy(); 4514 markOop m = old->mark(); 4515 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4516 : m->age(); 4517 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4518 word_sz); 4519 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4520 oop obj = oop(obj_ptr); 4521 4522 if (obj_ptr == NULL) { 4523 // This will either forward-to-self, or detect that someone else has 4524 // installed a forwarding pointer. 4525 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4526 return _g1->handle_evacuation_failure_par(cl, old); 4527 } 4528 4529 // We're going to allocate linearly, so might as well prefetch ahead. 4530 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4531 4532 oop forward_ptr = old->forward_to_atomic(obj); 4533 if (forward_ptr == NULL) { 4534 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4535 if (g1p->track_object_age(alloc_purpose)) { 4536 // We could simply do obj->incr_age(). However, this causes a 4537 // performance issue. obj->incr_age() will first check whether 4538 // the object has a displaced mark by checking its mark word; 4539 // getting the mark word from the new location of the object 4540 // stalls. So, given that we already have the mark word and we 4541 // are about to install it anyway, it's better to increase the 4542 // age on the mark word, when the object does not have a 4543 // displaced mark word. We're not expecting many objects to have 4544 // a displaced marked word, so that case is not optimized 4545 // further (it could be...) and we simply call obj->incr_age(). 4546 4547 if (m->has_displaced_mark_helper()) { 4548 // in this case, we have to install the mark word first, 4549 // otherwise obj looks to be forwarded (the old mark word, 4550 // which contains the forward pointer, was copied) 4551 obj->set_mark(m); 4552 obj->incr_age(); 4553 } else { 4554 m = m->incr_age(); 4555 obj->set_mark(m); 4556 } 4557 _par_scan_state->age_table()->add(obj, word_sz); 4558 } else { 4559 obj->set_mark(m); 4560 } 4561 4562 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4563 surv_young_words[young_index] += word_sz; 4564 4565 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4566 // We keep track of the next start index in the length field of 4567 // the to-space object. The actual length can be found in the 4568 // length field of the from-space object. 4569 arrayOop(obj)->set_length(0); 4570 oop* old_p = set_partial_array_mask(old); 4571 _par_scan_state->push_on_queue(old_p); 4572 } else { 4573 // No point in using the slower heap_region_containing() method, 4574 // given that we know obj is in the heap. 4575 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4576 obj->oop_iterate_backwards(&_scanner); 4577 } 4578 } else { 4579 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4580 obj = forward_ptr; 4581 } 4582 return obj; 4583 } 4584 4585 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4586 template <class T> 4587 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4588 ::do_oop_work(T* p) { 4589 oop obj = oopDesc::load_decode_heap_oop(p); 4590 assert(barrier != G1BarrierRS || obj != NULL, 4591 "Precondition: G1BarrierRS implies obj is non-NULL"); 4592 4593 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4594 4595 // here the null check is implicit in the cset_fast_test() test 4596 if (_g1->in_cset_fast_test(obj)) { 4597 oop forwardee; 4598 if (obj->is_forwarded()) { 4599 forwardee = obj->forwardee(); 4600 } else { 4601 forwardee = copy_to_survivor_space(obj); 4602 } 4603 assert(forwardee != NULL, "forwardee should not be NULL"); 4604 oopDesc::encode_store_heap_oop(p, forwardee); 4605 if (do_mark_object && forwardee != obj) { 4606 // If the object is self-forwarded we don't need to explicitly 4607 // mark it, the evacuation failure protocol will do so. 4608 mark_forwarded_object(obj, forwardee); 4609 } 4610 4611 // When scanning the RS, we only care about objs in CS. 4612 if (barrier == G1BarrierRS) { 4613 _par_scan_state->update_rs(_from, p, _worker_id); 4614 } 4615 } else { 4616 // The object is not in collection set. If we're a root scanning 4617 // closure during an initial mark pause (i.e. do_mark_object will 4618 // be true) then attempt to mark the object. 4619 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4620 mark_object(obj); 4621 } 4622 } 4623 4624 if (barrier == G1BarrierEvac && obj != NULL) { 4625 _par_scan_state->update_rs(_from, p, _worker_id); 4626 } 4627 4628 if (do_gen_barrier && obj != NULL) { 4629 par_do_barrier(p); 4630 } 4631 } 4632 4633 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4634 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4635 4636 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4637 assert(has_partial_array_mask(p), "invariant"); 4638 oop from_obj = clear_partial_array_mask(p); 4639 4640 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4641 assert(from_obj->is_objArray(), "must be obj array"); 4642 objArrayOop from_obj_array = objArrayOop(from_obj); 4643 // The from-space object contains the real length. 4644 int length = from_obj_array->length(); 4645 4646 assert(from_obj->is_forwarded(), "must be forwarded"); 4647 oop to_obj = from_obj->forwardee(); 4648 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4649 objArrayOop to_obj_array = objArrayOop(to_obj); 4650 // We keep track of the next start index in the length field of the 4651 // to-space object. 4652 int next_index = to_obj_array->length(); 4653 assert(0 <= next_index && next_index < length, 4654 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4655 4656 int start = next_index; 4657 int end = length; 4658 int remainder = end - start; 4659 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4660 if (remainder > 2 * ParGCArrayScanChunk) { 4661 end = start + ParGCArrayScanChunk; 4662 to_obj_array->set_length(end); 4663 // Push the remainder before we process the range in case another 4664 // worker has run out of things to do and can steal it. 4665 oop* from_obj_p = set_partial_array_mask(from_obj); 4666 _par_scan_state->push_on_queue(from_obj_p); 4667 } else { 4668 assert(length == end, "sanity"); 4669 // We'll process the final range for this object. Restore the length 4670 // so that the heap remains parsable in case of evacuation failure. 4671 to_obj_array->set_length(end); 4672 } 4673 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4674 // Process indexes [start,end). It will also process the header 4675 // along with the first chunk (i.e., the chunk with start == 0). 4676 // Note that at this point the length field of to_obj_array is not 4677 // correct given that we are using it to keep track of the next 4678 // start index. oop_iterate_range() (thankfully!) ignores the length 4679 // field and only relies on the start / end parameters. It does 4680 // however return the size of the object which will be incorrect. So 4681 // we have to ignore it even if we wanted to use it. 4682 to_obj_array->oop_iterate_range(&_scanner, start, end); 4683 } 4684 4685 class G1ParEvacuateFollowersClosure : public VoidClosure { 4686 protected: 4687 G1CollectedHeap* _g1h; 4688 G1ParScanThreadState* _par_scan_state; 4689 RefToScanQueueSet* _queues; 4690 ParallelTaskTerminator* _terminator; 4691 4692 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4693 RefToScanQueueSet* queues() { return _queues; } 4694 ParallelTaskTerminator* terminator() { return _terminator; } 4695 4696 public: 4697 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4698 G1ParScanThreadState* par_scan_state, 4699 RefToScanQueueSet* queues, 4700 ParallelTaskTerminator* terminator) 4701 : _g1h(g1h), _par_scan_state(par_scan_state), 4702 _queues(queues), _terminator(terminator) {} 4703 4704 void do_void(); 4705 4706 private: 4707 inline bool offer_termination(); 4708 }; 4709 4710 bool G1ParEvacuateFollowersClosure::offer_termination() { 4711 G1ParScanThreadState* const pss = par_scan_state(); 4712 pss->start_term_time(); 4713 const bool res = terminator()->offer_termination(); 4714 pss->end_term_time(); 4715 return res; 4716 } 4717 4718 void G1ParEvacuateFollowersClosure::do_void() { 4719 StarTask stolen_task; 4720 G1ParScanThreadState* const pss = par_scan_state(); 4721 pss->trim_queue(); 4722 4723 do { 4724 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4725 assert(pss->verify_task(stolen_task), "sanity"); 4726 if (stolen_task.is_narrow()) { 4727 pss->deal_with_reference((narrowOop*) stolen_task); 4728 } else { 4729 pss->deal_with_reference((oop*) stolen_task); 4730 } 4731 4732 // We've just processed a reference and we might have made 4733 // available new entries on the queues. So we have to make sure 4734 // we drain the queues as necessary. 4735 pss->trim_queue(); 4736 } 4737 } while (!offer_termination()); 4738 4739 pss->retire_alloc_buffers(); 4740 } 4741 4742 class G1ParTask : public AbstractGangTask { 4743 protected: 4744 G1CollectedHeap* _g1h; 4745 RefToScanQueueSet *_queues; 4746 ParallelTaskTerminator _terminator; 4747 uint _n_workers; 4748 4749 Mutex _stats_lock; 4750 Mutex* stats_lock() { return &_stats_lock; } 4751 4752 size_t getNCards() { 4753 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4754 / G1BlockOffsetSharedArray::N_bytes; 4755 } 4756 4757 public: 4758 G1ParTask(G1CollectedHeap* g1h, 4759 RefToScanQueueSet *task_queues) 4760 : AbstractGangTask("G1 collection"), 4761 _g1h(g1h), 4762 _queues(task_queues), 4763 _terminator(0, _queues), 4764 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4765 {} 4766 4767 RefToScanQueueSet* queues() { return _queues; } 4768 4769 RefToScanQueue *work_queue(int i) { 4770 return queues()->queue(i); 4771 } 4772 4773 ParallelTaskTerminator* terminator() { return &_terminator; } 4774 4775 virtual void set_for_termination(int active_workers) { 4776 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4777 // in the young space (_par_seq_tasks) in the G1 heap 4778 // for SequentialSubTasksDone. 4779 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4780 // both of which need setting by set_n_termination(). 4781 _g1h->SharedHeap::set_n_termination(active_workers); 4782 _g1h->set_n_termination(active_workers); 4783 terminator()->reset_for_reuse(active_workers); 4784 _n_workers = active_workers; 4785 } 4786 4787 void work(uint worker_id) { 4788 if (worker_id >= _n_workers) return; // no work needed this round 4789 4790 double start_time_ms = os::elapsedTime() * 1000.0; 4791 _g1h->g1_policy()->record_gc_worker_start_time(worker_id, start_time_ms); 4792 4793 { 4794 ResourceMark rm; 4795 HandleMark hm; 4796 4797 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4798 4799 G1ParScanThreadState pss(_g1h, worker_id); 4800 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4801 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4802 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4803 4804 pss.set_evac_closure(&scan_evac_cl); 4805 pss.set_evac_failure_closure(&evac_failure_cl); 4806 pss.set_partial_scan_closure(&partial_scan_cl); 4807 4808 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4809 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4810 4811 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4812 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4813 4814 OopClosure* scan_root_cl = &only_scan_root_cl; 4815 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4816 4817 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4818 // We also need to mark copied objects. 4819 scan_root_cl = &scan_mark_root_cl; 4820 scan_perm_cl = &scan_mark_perm_cl; 4821 } 4822 4823 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4824 4825 pss.start_strong_roots(); 4826 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4827 SharedHeap::SO_AllClasses, 4828 scan_root_cl, 4829 &push_heap_rs_cl, 4830 scan_perm_cl, 4831 worker_id); 4832 pss.end_strong_roots(); 4833 4834 { 4835 double start = os::elapsedTime(); 4836 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4837 evac.do_void(); 4838 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4839 double term_ms = pss.term_time()*1000.0; 4840 _g1h->g1_policy()->record_obj_copy_time(worker_id, elapsed_ms-term_ms); 4841 _g1h->g1_policy()->record_termination(worker_id, term_ms, pss.term_attempts()); 4842 } 4843 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4844 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4845 4846 if (ParallelGCVerbose) { 4847 MutexLocker x(stats_lock()); 4848 pss.print_termination_stats(worker_id); 4849 } 4850 4851 assert(pss.refs()->is_empty(), "should be empty"); 4852 4853 // Close the inner scope so that the ResourceMark and HandleMark 4854 // destructors are executed here and are included as part of the 4855 // "GC Worker Time". 4856 } 4857 4858 double end_time_ms = os::elapsedTime() * 1000.0; 4859 _g1h->g1_policy()->record_gc_worker_end_time(worker_id, end_time_ms); 4860 } 4861 }; 4862 4863 // *** Common G1 Evacuation Stuff 4864 4865 // Closures that support the filtering of CodeBlobs scanned during 4866 // external root scanning. 4867 4868 // Closure applied to reference fields in code blobs (specifically nmethods) 4869 // to determine whether an nmethod contains references that point into 4870 // the collection set. Used as a predicate when walking code roots so 4871 // that only nmethods that point into the collection set are added to the 4872 // 'marked' list. 4873 4874 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure { 4875 4876 class G1PointsIntoCSOopClosure : public OopClosure { 4877 G1CollectedHeap* _g1; 4878 bool _points_into_cs; 4879 public: 4880 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) : 4881 _g1(g1), _points_into_cs(false) { } 4882 4883 bool points_into_cs() const { return _points_into_cs; } 4884 4885 template <class T> 4886 void do_oop_nv(T* p) { 4887 if (!_points_into_cs) { 4888 T heap_oop = oopDesc::load_heap_oop(p); 4889 if (!oopDesc::is_null(heap_oop) && 4890 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) { 4891 _points_into_cs = true; 4892 } 4893 } 4894 } 4895 4896 virtual void do_oop(oop* p) { do_oop_nv(p); } 4897 virtual void do_oop(narrowOop* p) { do_oop_nv(p); } 4898 }; 4899 4900 G1CollectedHeap* _g1; 4901 4902 public: 4903 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) : 4904 CodeBlobToOopClosure(cl, true), _g1(g1) { } 4905 4906 virtual void do_code_blob(CodeBlob* cb) { 4907 nmethod* nm = cb->as_nmethod_or_null(); 4908 if (nm != NULL && !(nm->test_oops_do_mark())) { 4909 G1PointsIntoCSOopClosure predicate_cl(_g1); 4910 nm->oops_do(&predicate_cl); 4911 4912 if (predicate_cl.points_into_cs()) { 4913 // At least one of the reference fields or the oop relocations 4914 // in the nmethod points into the collection set. We have to 4915 // 'mark' this nmethod. 4916 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob() 4917 // or MarkingCodeBlobClosure::do_code_blob() change. 4918 if (!nm->test_set_oops_do_mark()) { 4919 do_newly_marked_nmethod(nm); 4920 } 4921 } 4922 } 4923 } 4924 }; 4925 4926 // This method is run in a GC worker. 4927 4928 void 4929 G1CollectedHeap:: 4930 g1_process_strong_roots(bool collecting_perm_gen, 4931 ScanningOption so, 4932 OopClosure* scan_non_heap_roots, 4933 OopsInHeapRegionClosure* scan_rs, 4934 OopsInGenClosure* scan_perm, 4935 int worker_i) { 4936 4937 // First scan the strong roots, including the perm gen. 4938 double ext_roots_start = os::elapsedTime(); 4939 double closure_app_time_sec = 0.0; 4940 4941 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4942 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4943 buf_scan_perm.set_generation(perm_gen()); 4944 4945 // Walk the code cache w/o buffering, because StarTask cannot handle 4946 // unaligned oop locations. 4947 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots); 4948 4949 process_strong_roots(false, // no scoping; this is parallel code 4950 collecting_perm_gen, so, 4951 &buf_scan_non_heap_roots, 4952 &eager_scan_code_roots, 4953 &buf_scan_perm); 4954 4955 // Now the CM ref_processor roots. 4956 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4957 // We need to treat the discovered reference lists of the 4958 // concurrent mark ref processor as roots and keep entries 4959 // (which are added by the marking threads) on them live 4960 // until they can be processed at the end of marking. 4961 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4962 } 4963 4964 // Finish up any enqueued closure apps (attributed as object copy time). 4965 buf_scan_non_heap_roots.done(); 4966 buf_scan_perm.done(); 4967 4968 double ext_roots_end = os::elapsedTime(); 4969 4970 g1_policy()->reset_obj_copy_time(worker_i); 4971 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4972 buf_scan_non_heap_roots.closure_app_seconds(); 4973 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4974 4975 double ext_root_time_ms = 4976 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4977 4978 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4979 4980 // During conc marking we have to filter the per-thread SATB buffers 4981 // to make sure we remove any oops into the CSet (which will show up 4982 // as implicitly live). 4983 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 4984 if (mark_in_progress()) { 4985 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 4986 } 4987 } 4988 double satb_filtering_ms = (os::elapsedTime() - ext_roots_end) * 1000.0; 4989 g1_policy()->record_satb_filtering_time(worker_i, satb_filtering_ms); 4990 4991 // Now scan the complement of the collection set. 4992 if (scan_rs != NULL) { 4993 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4994 } 4995 4996 _process_strong_tasks->all_tasks_completed(); 4997 } 4998 4999 void 5000 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 5001 OopClosure* non_root_closure) { 5002 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5003 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 5004 } 5005 5006 // Weak Reference Processing support 5007 5008 // An always "is_alive" closure that is used to preserve referents. 5009 // If the object is non-null then it's alive. Used in the preservation 5010 // of referent objects that are pointed to by reference objects 5011 // discovered by the CM ref processor. 5012 class G1AlwaysAliveClosure: public BoolObjectClosure { 5013 G1CollectedHeap* _g1; 5014 public: 5015 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5016 void do_object(oop p) { assert(false, "Do not call."); } 5017 bool do_object_b(oop p) { 5018 if (p != NULL) { 5019 return true; 5020 } 5021 return false; 5022 } 5023 }; 5024 5025 bool G1STWIsAliveClosure::do_object_b(oop p) { 5026 // An object is reachable if it is outside the collection set, 5027 // or is inside and copied. 5028 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5029 } 5030 5031 // Non Copying Keep Alive closure 5032 class G1KeepAliveClosure: public OopClosure { 5033 G1CollectedHeap* _g1; 5034 public: 5035 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5036 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5037 void do_oop( oop* p) { 5038 oop obj = *p; 5039 5040 if (_g1->obj_in_cs(obj)) { 5041 assert( obj->is_forwarded(), "invariant" ); 5042 *p = obj->forwardee(); 5043 } 5044 } 5045 }; 5046 5047 // Copying Keep Alive closure - can be called from both 5048 // serial and parallel code as long as different worker 5049 // threads utilize different G1ParScanThreadState instances 5050 // and different queues. 5051 5052 class G1CopyingKeepAliveClosure: public OopClosure { 5053 G1CollectedHeap* _g1h; 5054 OopClosure* _copy_non_heap_obj_cl; 5055 OopsInHeapRegionClosure* _copy_perm_obj_cl; 5056 G1ParScanThreadState* _par_scan_state; 5057 5058 public: 5059 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5060 OopClosure* non_heap_obj_cl, 5061 OopsInHeapRegionClosure* perm_obj_cl, 5062 G1ParScanThreadState* pss): 5063 _g1h(g1h), 5064 _copy_non_heap_obj_cl(non_heap_obj_cl), 5065 _copy_perm_obj_cl(perm_obj_cl), 5066 _par_scan_state(pss) 5067 {} 5068 5069 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5070 virtual void do_oop( oop* p) { do_oop_work(p); } 5071 5072 template <class T> void do_oop_work(T* p) { 5073 oop obj = oopDesc::load_decode_heap_oop(p); 5074 5075 if (_g1h->obj_in_cs(obj)) { 5076 // If the referent object has been forwarded (either copied 5077 // to a new location or to itself in the event of an 5078 // evacuation failure) then we need to update the reference 5079 // field and, if both reference and referent are in the G1 5080 // heap, update the RSet for the referent. 5081 // 5082 // If the referent has not been forwarded then we have to keep 5083 // it alive by policy. Therefore we have copy the referent. 5084 // 5085 // If the reference field is in the G1 heap then we can push 5086 // on the PSS queue. When the queue is drained (after each 5087 // phase of reference processing) the object and it's followers 5088 // will be copied, the reference field set to point to the 5089 // new location, and the RSet updated. Otherwise we need to 5090 // use the the non-heap or perm closures directly to copy 5091 // the refernt object and update the pointer, while avoiding 5092 // updating the RSet. 5093 5094 if (_g1h->is_in_g1_reserved(p)) { 5095 _par_scan_state->push_on_queue(p); 5096 } else { 5097 // The reference field is not in the G1 heap. 5098 if (_g1h->perm_gen()->is_in(p)) { 5099 _copy_perm_obj_cl->do_oop(p); 5100 } else { 5101 _copy_non_heap_obj_cl->do_oop(p); 5102 } 5103 } 5104 } 5105 } 5106 }; 5107 5108 // Serial drain queue closure. Called as the 'complete_gc' 5109 // closure for each discovered list in some of the 5110 // reference processing phases. 5111 5112 class G1STWDrainQueueClosure: public VoidClosure { 5113 protected: 5114 G1CollectedHeap* _g1h; 5115 G1ParScanThreadState* _par_scan_state; 5116 5117 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5118 5119 public: 5120 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5121 _g1h(g1h), 5122 _par_scan_state(pss) 5123 { } 5124 5125 void do_void() { 5126 G1ParScanThreadState* const pss = par_scan_state(); 5127 pss->trim_queue(); 5128 } 5129 }; 5130 5131 // Parallel Reference Processing closures 5132 5133 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5134 // processing during G1 evacuation pauses. 5135 5136 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5137 private: 5138 G1CollectedHeap* _g1h; 5139 RefToScanQueueSet* _queues; 5140 FlexibleWorkGang* _workers; 5141 int _active_workers; 5142 5143 public: 5144 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5145 FlexibleWorkGang* workers, 5146 RefToScanQueueSet *task_queues, 5147 int n_workers) : 5148 _g1h(g1h), 5149 _queues(task_queues), 5150 _workers(workers), 5151 _active_workers(n_workers) 5152 { 5153 assert(n_workers > 0, "shouldn't call this otherwise"); 5154 } 5155 5156 // Executes the given task using concurrent marking worker threads. 5157 virtual void execute(ProcessTask& task); 5158 virtual void execute(EnqueueTask& task); 5159 }; 5160 5161 // Gang task for possibly parallel reference processing 5162 5163 class G1STWRefProcTaskProxy: public AbstractGangTask { 5164 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5165 ProcessTask& _proc_task; 5166 G1CollectedHeap* _g1h; 5167 RefToScanQueueSet *_task_queues; 5168 ParallelTaskTerminator* _terminator; 5169 5170 public: 5171 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5172 G1CollectedHeap* g1h, 5173 RefToScanQueueSet *task_queues, 5174 ParallelTaskTerminator* terminator) : 5175 AbstractGangTask("Process reference objects in parallel"), 5176 _proc_task(proc_task), 5177 _g1h(g1h), 5178 _task_queues(task_queues), 5179 _terminator(terminator) 5180 {} 5181 5182 virtual void work(uint worker_id) { 5183 // The reference processing task executed by a single worker. 5184 ResourceMark rm; 5185 HandleMark hm; 5186 5187 G1STWIsAliveClosure is_alive(_g1h); 5188 5189 G1ParScanThreadState pss(_g1h, worker_id); 5190 5191 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5192 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5193 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5194 5195 pss.set_evac_closure(&scan_evac_cl); 5196 pss.set_evac_failure_closure(&evac_failure_cl); 5197 pss.set_partial_scan_closure(&partial_scan_cl); 5198 5199 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5200 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5201 5202 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5203 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5204 5205 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5206 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5207 5208 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5209 // We also need to mark copied objects. 5210 copy_non_heap_cl = ©_mark_non_heap_cl; 5211 copy_perm_cl = ©_mark_perm_cl; 5212 } 5213 5214 // Keep alive closure. 5215 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5216 5217 // Complete GC closure 5218 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5219 5220 // Call the reference processing task's work routine. 5221 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5222 5223 // Note we cannot assert that the refs array is empty here as not all 5224 // of the processing tasks (specifically phase2 - pp2_work) execute 5225 // the complete_gc closure (which ordinarily would drain the queue) so 5226 // the queue may not be empty. 5227 } 5228 }; 5229 5230 // Driver routine for parallel reference processing. 5231 // Creates an instance of the ref processing gang 5232 // task and has the worker threads execute it. 5233 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5234 assert(_workers != NULL, "Need parallel worker threads."); 5235 5236 ParallelTaskTerminator terminator(_active_workers, _queues); 5237 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5238 5239 _g1h->set_par_threads(_active_workers); 5240 _workers->run_task(&proc_task_proxy); 5241 _g1h->set_par_threads(0); 5242 } 5243 5244 // Gang task for parallel reference enqueueing. 5245 5246 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5247 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5248 EnqueueTask& _enq_task; 5249 5250 public: 5251 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5252 AbstractGangTask("Enqueue reference objects in parallel"), 5253 _enq_task(enq_task) 5254 { } 5255 5256 virtual void work(uint worker_id) { 5257 _enq_task.work(worker_id); 5258 } 5259 }; 5260 5261 // Driver routine for parallel reference enqueing. 5262 // Creates an instance of the ref enqueueing gang 5263 // task and has the worker threads execute it. 5264 5265 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5266 assert(_workers != NULL, "Need parallel worker threads."); 5267 5268 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5269 5270 _g1h->set_par_threads(_active_workers); 5271 _workers->run_task(&enq_task_proxy); 5272 _g1h->set_par_threads(0); 5273 } 5274 5275 // End of weak reference support closures 5276 5277 // Abstract task used to preserve (i.e. copy) any referent objects 5278 // that are in the collection set and are pointed to by reference 5279 // objects discovered by the CM ref processor. 5280 5281 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5282 protected: 5283 G1CollectedHeap* _g1h; 5284 RefToScanQueueSet *_queues; 5285 ParallelTaskTerminator _terminator; 5286 uint _n_workers; 5287 5288 public: 5289 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5290 AbstractGangTask("ParPreserveCMReferents"), 5291 _g1h(g1h), 5292 _queues(task_queues), 5293 _terminator(workers, _queues), 5294 _n_workers(workers) 5295 { } 5296 5297 void work(uint worker_id) { 5298 ResourceMark rm; 5299 HandleMark hm; 5300 5301 G1ParScanThreadState pss(_g1h, worker_id); 5302 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5303 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5304 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5305 5306 pss.set_evac_closure(&scan_evac_cl); 5307 pss.set_evac_failure_closure(&evac_failure_cl); 5308 pss.set_partial_scan_closure(&partial_scan_cl); 5309 5310 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5311 5312 5313 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5314 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5315 5316 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5317 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5318 5319 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5320 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5321 5322 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5323 // We also need to mark copied objects. 5324 copy_non_heap_cl = ©_mark_non_heap_cl; 5325 copy_perm_cl = ©_mark_perm_cl; 5326 } 5327 5328 // Is alive closure 5329 G1AlwaysAliveClosure always_alive(_g1h); 5330 5331 // Copying keep alive closure. Applied to referent objects that need 5332 // to be copied. 5333 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5334 5335 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5336 5337 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5338 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5339 5340 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5341 // So this must be true - but assert just in case someone decides to 5342 // change the worker ids. 5343 assert(0 <= worker_id && worker_id < limit, "sanity"); 5344 assert(!rp->discovery_is_atomic(), "check this code"); 5345 5346 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5347 for (uint idx = worker_id; idx < limit; idx += stride) { 5348 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5349 5350 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5351 while (iter.has_next()) { 5352 // Since discovery is not atomic for the CM ref processor, we 5353 // can see some null referent objects. 5354 iter.load_ptrs(DEBUG_ONLY(true)); 5355 oop ref = iter.obj(); 5356 5357 // This will filter nulls. 5358 if (iter.is_referent_alive()) { 5359 iter.make_referent_alive(); 5360 } 5361 iter.move_to_next(); 5362 } 5363 } 5364 5365 // Drain the queue - which may cause stealing 5366 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5367 drain_queue.do_void(); 5368 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5369 assert(pss.refs()->is_empty(), "should be"); 5370 } 5371 }; 5372 5373 // Weak Reference processing during an evacuation pause (part 1). 5374 void G1CollectedHeap::process_discovered_references() { 5375 double ref_proc_start = os::elapsedTime(); 5376 5377 ReferenceProcessor* rp = _ref_processor_stw; 5378 assert(rp->discovery_enabled(), "should have been enabled"); 5379 5380 // Any reference objects, in the collection set, that were 'discovered' 5381 // by the CM ref processor should have already been copied (either by 5382 // applying the external root copy closure to the discovered lists, or 5383 // by following an RSet entry). 5384 // 5385 // But some of the referents, that are in the collection set, that these 5386 // reference objects point to may not have been copied: the STW ref 5387 // processor would have seen that the reference object had already 5388 // been 'discovered' and would have skipped discovering the reference, 5389 // but would not have treated the reference object as a regular oop. 5390 // As a reult the copy closure would not have been applied to the 5391 // referent object. 5392 // 5393 // We need to explicitly copy these referent objects - the references 5394 // will be processed at the end of remarking. 5395 // 5396 // We also need to do this copying before we process the reference 5397 // objects discovered by the STW ref processor in case one of these 5398 // referents points to another object which is also referenced by an 5399 // object discovered by the STW ref processor. 5400 5401 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5402 workers()->active_workers() : 1); 5403 5404 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5405 active_workers == workers()->active_workers(), 5406 "Need to reset active_workers"); 5407 5408 set_par_threads(active_workers); 5409 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5410 5411 if (G1CollectedHeap::use_parallel_gc_threads()) { 5412 workers()->run_task(&keep_cm_referents); 5413 } else { 5414 keep_cm_referents.work(0); 5415 } 5416 5417 set_par_threads(0); 5418 5419 // Closure to test whether a referent is alive. 5420 G1STWIsAliveClosure is_alive(this); 5421 5422 // Even when parallel reference processing is enabled, the processing 5423 // of JNI refs is serial and performed serially by the current thread 5424 // rather than by a worker. The following PSS will be used for processing 5425 // JNI refs. 5426 5427 // Use only a single queue for this PSS. 5428 G1ParScanThreadState pss(this, 0); 5429 5430 // We do not embed a reference processor in the copying/scanning 5431 // closures while we're actually processing the discovered 5432 // reference objects. 5433 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5434 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5435 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5436 5437 pss.set_evac_closure(&scan_evac_cl); 5438 pss.set_evac_failure_closure(&evac_failure_cl); 5439 pss.set_partial_scan_closure(&partial_scan_cl); 5440 5441 assert(pss.refs()->is_empty(), "pre-condition"); 5442 5443 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5444 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5445 5446 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5447 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5448 5449 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5450 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5451 5452 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5453 // We also need to mark copied objects. 5454 copy_non_heap_cl = ©_mark_non_heap_cl; 5455 copy_perm_cl = ©_mark_perm_cl; 5456 } 5457 5458 // Keep alive closure. 5459 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5460 5461 // Serial Complete GC closure 5462 G1STWDrainQueueClosure drain_queue(this, &pss); 5463 5464 // Setup the soft refs policy... 5465 rp->setup_policy(false); 5466 5467 if (!rp->processing_is_mt()) { 5468 // Serial reference processing... 5469 rp->process_discovered_references(&is_alive, 5470 &keep_alive, 5471 &drain_queue, 5472 NULL); 5473 } else { 5474 // Parallel reference processing 5475 assert(rp->num_q() == active_workers, "sanity"); 5476 assert(active_workers <= rp->max_num_q(), "sanity"); 5477 5478 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5479 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5480 } 5481 5482 // We have completed copying any necessary live referent objects 5483 // (that were not copied during the actual pause) so we can 5484 // retire any active alloc buffers 5485 pss.retire_alloc_buffers(); 5486 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5487 5488 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5489 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0); 5490 } 5491 5492 // Weak Reference processing during an evacuation pause (part 2). 5493 void G1CollectedHeap::enqueue_discovered_references() { 5494 double ref_enq_start = os::elapsedTime(); 5495 5496 ReferenceProcessor* rp = _ref_processor_stw; 5497 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5498 5499 // Now enqueue any remaining on the discovered lists on to 5500 // the pending list. 5501 if (!rp->processing_is_mt()) { 5502 // Serial reference processing... 5503 rp->enqueue_discovered_references(); 5504 } else { 5505 // Parallel reference enqueuing 5506 5507 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5508 assert(active_workers == workers()->active_workers(), 5509 "Need to reset active_workers"); 5510 assert(rp->num_q() == active_workers, "sanity"); 5511 assert(active_workers <= rp->max_num_q(), "sanity"); 5512 5513 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5514 rp->enqueue_discovered_references(&par_task_executor); 5515 } 5516 5517 rp->verify_no_references_recorded(); 5518 assert(!rp->discovery_enabled(), "should have been disabled"); 5519 5520 // FIXME 5521 // CM's reference processing also cleans up the string and symbol tables. 5522 // Should we do that here also? We could, but it is a serial operation 5523 // and could signicantly increase the pause time. 5524 5525 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5526 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0); 5527 } 5528 5529 void G1CollectedHeap::evacuate_collection_set() { 5530 _expand_heap_after_alloc_failure = true; 5531 set_evacuation_failed(false); 5532 5533 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5534 concurrent_g1_refine()->set_use_cache(false); 5535 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5536 5537 uint n_workers; 5538 if (G1CollectedHeap::use_parallel_gc_threads()) { 5539 n_workers = 5540 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5541 workers()->active_workers(), 5542 Threads::number_of_non_daemon_threads()); 5543 assert(UseDynamicNumberOfGCThreads || 5544 n_workers == workers()->total_workers(), 5545 "If not dynamic should be using all the workers"); 5546 workers()->set_active_workers(n_workers); 5547 set_par_threads(n_workers); 5548 } else { 5549 assert(n_par_threads() == 0, 5550 "Should be the original non-parallel value"); 5551 n_workers = 1; 5552 } 5553 5554 G1ParTask g1_par_task(this, _task_queues); 5555 5556 init_for_evac_failure(NULL); 5557 5558 rem_set()->prepare_for_younger_refs_iterate(true); 5559 5560 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5561 double start_par_time_sec = os::elapsedTime(); 5562 double end_par_time_sec; 5563 5564 { 5565 StrongRootsScope srs(this); 5566 5567 if (G1CollectedHeap::use_parallel_gc_threads()) { 5568 // The individual threads will set their evac-failure closures. 5569 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5570 // These tasks use ShareHeap::_process_strong_tasks 5571 assert(UseDynamicNumberOfGCThreads || 5572 workers()->active_workers() == workers()->total_workers(), 5573 "If not dynamic should be using all the workers"); 5574 workers()->run_task(&g1_par_task); 5575 } else { 5576 g1_par_task.set_for_termination(n_workers); 5577 g1_par_task.work(0); 5578 } 5579 end_par_time_sec = os::elapsedTime(); 5580 5581 // Closing the inner scope will execute the destructor 5582 // for the StrongRootsScope object. We record the current 5583 // elapsed time before closing the scope so that time 5584 // taken for the SRS destructor is NOT included in the 5585 // reported parallel time. 5586 } 5587 5588 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5589 g1_policy()->record_par_time(par_time_ms); 5590 5591 double code_root_fixup_time_ms = 5592 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5593 g1_policy()->record_code_root_fixup_time(code_root_fixup_time_ms); 5594 5595 set_par_threads(0); 5596 5597 // Process any discovered reference objects - we have 5598 // to do this _before_ we retire the GC alloc regions 5599 // as we may have to copy some 'reachable' referent 5600 // objects (and their reachable sub-graphs) that were 5601 // not copied during the pause. 5602 process_discovered_references(); 5603 5604 // Weak root processing. 5605 // Note: when JSR 292 is enabled and code blobs can contain 5606 // non-perm oops then we will need to process the code blobs 5607 // here too. 5608 { 5609 G1STWIsAliveClosure is_alive(this); 5610 G1KeepAliveClosure keep_alive(this); 5611 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5612 } 5613 5614 release_gc_alloc_regions(); 5615 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5616 5617 concurrent_g1_refine()->clear_hot_cache(); 5618 concurrent_g1_refine()->set_use_cache(true); 5619 5620 finalize_for_evac_failure(); 5621 5622 if (evacuation_failed()) { 5623 remove_self_forwarding_pointers(); 5624 if (G1Log::finer()) { 5625 gclog_or_tty->print(" (to-space exhausted)"); 5626 } else if (G1Log::fine()) { 5627 gclog_or_tty->print("--"); 5628 } 5629 } 5630 5631 // Enqueue any remaining references remaining on the STW 5632 // reference processor's discovered lists. We need to do 5633 // this after the card table is cleaned (and verified) as 5634 // the act of enqueuing entries on to the pending list 5635 // will log these updates (and dirty their associated 5636 // cards). We need these updates logged to update any 5637 // RSets. 5638 enqueue_discovered_references(); 5639 5640 if (G1DeferredRSUpdate) { 5641 RedirtyLoggedCardTableEntryFastClosure redirty; 5642 dirty_card_queue_set().set_closure(&redirty); 5643 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5644 5645 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5646 dcq.merge_bufferlists(&dirty_card_queue_set()); 5647 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5648 } 5649 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5650 } 5651 5652 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5653 size_t* pre_used, 5654 FreeRegionList* free_list, 5655 OldRegionSet* old_proxy_set, 5656 HumongousRegionSet* humongous_proxy_set, 5657 HRRSCleanupTask* hrrs_cleanup_task, 5658 bool par) { 5659 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5660 if (hr->isHumongous()) { 5661 assert(hr->startsHumongous(), "we should only see starts humongous"); 5662 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5663 } else { 5664 _old_set.remove_with_proxy(hr, old_proxy_set); 5665 free_region(hr, pre_used, free_list, par); 5666 } 5667 } else { 5668 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5669 } 5670 } 5671 5672 void G1CollectedHeap::free_region(HeapRegion* hr, 5673 size_t* pre_used, 5674 FreeRegionList* free_list, 5675 bool par) { 5676 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5677 assert(!hr->is_empty(), "the region should not be empty"); 5678 assert(free_list != NULL, "pre-condition"); 5679 5680 *pre_used += hr->used(); 5681 hr->hr_clear(par, true /* clear_space */); 5682 free_list->add_as_head(hr); 5683 } 5684 5685 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5686 size_t* pre_used, 5687 FreeRegionList* free_list, 5688 HumongousRegionSet* humongous_proxy_set, 5689 bool par) { 5690 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5691 assert(free_list != NULL, "pre-condition"); 5692 assert(humongous_proxy_set != NULL, "pre-condition"); 5693 5694 size_t hr_used = hr->used(); 5695 size_t hr_capacity = hr->capacity(); 5696 size_t hr_pre_used = 0; 5697 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5698 // We need to read this before we make the region non-humongous, 5699 // otherwise the information will be gone. 5700 uint last_index = hr->last_hc_index(); 5701 hr->set_notHumongous(); 5702 free_region(hr, &hr_pre_used, free_list, par); 5703 5704 uint i = hr->hrs_index() + 1; 5705 while (i < last_index) { 5706 HeapRegion* curr_hr = region_at(i); 5707 assert(curr_hr->continuesHumongous(), "invariant"); 5708 curr_hr->set_notHumongous(); 5709 free_region(curr_hr, &hr_pre_used, free_list, par); 5710 i += 1; 5711 } 5712 assert(hr_pre_used == hr_used, 5713 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5714 "should be the same", hr_pre_used, hr_used)); 5715 *pre_used += hr_pre_used; 5716 } 5717 5718 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5719 FreeRegionList* free_list, 5720 OldRegionSet* old_proxy_set, 5721 HumongousRegionSet* humongous_proxy_set, 5722 bool par) { 5723 if (pre_used > 0) { 5724 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5725 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5726 assert(_summary_bytes_used >= pre_used, 5727 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5728 "should be >= pre_used: "SIZE_FORMAT, 5729 _summary_bytes_used, pre_used)); 5730 _summary_bytes_used -= pre_used; 5731 } 5732 if (free_list != NULL && !free_list->is_empty()) { 5733 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5734 _free_list.add_as_head(free_list); 5735 } 5736 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5737 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5738 _old_set.update_from_proxy(old_proxy_set); 5739 } 5740 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5741 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5742 _humongous_set.update_from_proxy(humongous_proxy_set); 5743 } 5744 } 5745 5746 class G1ParCleanupCTTask : public AbstractGangTask { 5747 CardTableModRefBS* _ct_bs; 5748 G1CollectedHeap* _g1h; 5749 HeapRegion* volatile _su_head; 5750 public: 5751 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5752 G1CollectedHeap* g1h) : 5753 AbstractGangTask("G1 Par Cleanup CT Task"), 5754 _ct_bs(ct_bs), _g1h(g1h) { } 5755 5756 void work(uint worker_id) { 5757 HeapRegion* r; 5758 while (r = _g1h->pop_dirty_cards_region()) { 5759 clear_cards(r); 5760 } 5761 } 5762 5763 void clear_cards(HeapRegion* r) { 5764 // Cards of the survivors should have already been dirtied. 5765 if (!r->is_survivor()) { 5766 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5767 } 5768 } 5769 }; 5770 5771 #ifndef PRODUCT 5772 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5773 G1CollectedHeap* _g1h; 5774 CardTableModRefBS* _ct_bs; 5775 public: 5776 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5777 : _g1h(g1h), _ct_bs(ct_bs) { } 5778 virtual bool doHeapRegion(HeapRegion* r) { 5779 if (r->is_survivor()) { 5780 _g1h->verify_dirty_region(r); 5781 } else { 5782 _g1h->verify_not_dirty_region(r); 5783 } 5784 return false; 5785 } 5786 }; 5787 5788 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5789 // All of the region should be clean. 5790 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5791 MemRegion mr(hr->bottom(), hr->end()); 5792 ct_bs->verify_not_dirty_region(mr); 5793 } 5794 5795 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5796 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5797 // dirty allocated blocks as they allocate them. The thread that 5798 // retires each region and replaces it with a new one will do a 5799 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5800 // not dirty that area (one less thing to have to do while holding 5801 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5802 // is dirty. 5803 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5804 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5805 ct_bs->verify_dirty_region(mr); 5806 } 5807 5808 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5809 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5810 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5811 verify_dirty_region(hr); 5812 } 5813 } 5814 5815 void G1CollectedHeap::verify_dirty_young_regions() { 5816 verify_dirty_young_list(_young_list->first_region()); 5817 } 5818 #endif 5819 5820 void G1CollectedHeap::cleanUpCardTable() { 5821 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5822 double start = os::elapsedTime(); 5823 5824 { 5825 // Iterate over the dirty cards region list. 5826 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5827 5828 if (G1CollectedHeap::use_parallel_gc_threads()) { 5829 set_par_threads(); 5830 workers()->run_task(&cleanup_task); 5831 set_par_threads(0); 5832 } else { 5833 while (_dirty_cards_region_list) { 5834 HeapRegion* r = _dirty_cards_region_list; 5835 cleanup_task.clear_cards(r); 5836 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5837 if (_dirty_cards_region_list == r) { 5838 // The last region. 5839 _dirty_cards_region_list = NULL; 5840 } 5841 r->set_next_dirty_cards_region(NULL); 5842 } 5843 } 5844 #ifndef PRODUCT 5845 if (G1VerifyCTCleanup || VerifyAfterGC) { 5846 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5847 heap_region_iterate(&cleanup_verifier); 5848 } 5849 #endif 5850 } 5851 5852 double elapsed = os::elapsedTime() - start; 5853 g1_policy()->record_clear_ct_time(elapsed * 1000.0); 5854 } 5855 5856 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5857 size_t pre_used = 0; 5858 FreeRegionList local_free_list("Local List for CSet Freeing"); 5859 5860 double young_time_ms = 0.0; 5861 double non_young_time_ms = 0.0; 5862 5863 // Since the collection set is a superset of the the young list, 5864 // all we need to do to clear the young list is clear its 5865 // head and length, and unlink any young regions in the code below 5866 _young_list->clear(); 5867 5868 G1CollectorPolicy* policy = g1_policy(); 5869 5870 double start_sec = os::elapsedTime(); 5871 bool non_young = true; 5872 5873 HeapRegion* cur = cs_head; 5874 int age_bound = -1; 5875 size_t rs_lengths = 0; 5876 5877 while (cur != NULL) { 5878 assert(!is_on_master_free_list(cur), "sanity"); 5879 if (non_young) { 5880 if (cur->is_young()) { 5881 double end_sec = os::elapsedTime(); 5882 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5883 non_young_time_ms += elapsed_ms; 5884 5885 start_sec = os::elapsedTime(); 5886 non_young = false; 5887 } 5888 } else { 5889 if (!cur->is_young()) { 5890 double end_sec = os::elapsedTime(); 5891 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5892 young_time_ms += elapsed_ms; 5893 5894 start_sec = os::elapsedTime(); 5895 non_young = true; 5896 } 5897 } 5898 5899 rs_lengths += cur->rem_set()->occupied(); 5900 5901 HeapRegion* next = cur->next_in_collection_set(); 5902 assert(cur->in_collection_set(), "bad CS"); 5903 cur->set_next_in_collection_set(NULL); 5904 cur->set_in_collection_set(false); 5905 5906 if (cur->is_young()) { 5907 int index = cur->young_index_in_cset(); 5908 assert(index != -1, "invariant"); 5909 assert((uint) index < policy->young_cset_region_length(), "invariant"); 5910 size_t words_survived = _surviving_young_words[index]; 5911 cur->record_surv_words_in_group(words_survived); 5912 5913 // At this point the we have 'popped' cur from the collection set 5914 // (linked via next_in_collection_set()) but it is still in the 5915 // young list (linked via next_young_region()). Clear the 5916 // _next_young_region field. 5917 cur->set_next_young_region(NULL); 5918 } else { 5919 int index = cur->young_index_in_cset(); 5920 assert(index == -1, "invariant"); 5921 } 5922 5923 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5924 (!cur->is_young() && cur->young_index_in_cset() == -1), 5925 "invariant" ); 5926 5927 if (!cur->evacuation_failed()) { 5928 MemRegion used_mr = cur->used_region(); 5929 5930 // And the region is empty. 5931 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5932 free_region(cur, &pre_used, &local_free_list, false /* par */); 5933 } else { 5934 cur->uninstall_surv_rate_group(); 5935 if (cur->is_young()) { 5936 cur->set_young_index_in_cset(-1); 5937 } 5938 cur->set_not_young(); 5939 cur->set_evacuation_failed(false); 5940 // The region is now considered to be old. 5941 _old_set.add(cur); 5942 } 5943 cur = next; 5944 } 5945 5946 policy->record_max_rs_lengths(rs_lengths); 5947 policy->cset_regions_freed(); 5948 5949 double end_sec = os::elapsedTime(); 5950 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5951 5952 if (non_young) { 5953 non_young_time_ms += elapsed_ms; 5954 } else { 5955 young_time_ms += elapsed_ms; 5956 } 5957 5958 update_sets_after_freeing_regions(pre_used, &local_free_list, 5959 NULL /* old_proxy_set */, 5960 NULL /* humongous_proxy_set */, 5961 false /* par */); 5962 policy->record_young_free_cset_time_ms(young_time_ms); 5963 policy->record_non_young_free_cset_time_ms(non_young_time_ms); 5964 } 5965 5966 // This routine is similar to the above but does not record 5967 // any policy statistics or update free lists; we are abandoning 5968 // the current incremental collection set in preparation of a 5969 // full collection. After the full GC we will start to build up 5970 // the incremental collection set again. 5971 // This is only called when we're doing a full collection 5972 // and is immediately followed by the tearing down of the young list. 5973 5974 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5975 HeapRegion* cur = cs_head; 5976 5977 while (cur != NULL) { 5978 HeapRegion* next = cur->next_in_collection_set(); 5979 assert(cur->in_collection_set(), "bad CS"); 5980 cur->set_next_in_collection_set(NULL); 5981 cur->set_in_collection_set(false); 5982 cur->set_young_index_in_cset(-1); 5983 cur = next; 5984 } 5985 } 5986 5987 void G1CollectedHeap::set_free_regions_coming() { 5988 if (G1ConcRegionFreeingVerbose) { 5989 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5990 "setting free regions coming"); 5991 } 5992 5993 assert(!free_regions_coming(), "pre-condition"); 5994 _free_regions_coming = true; 5995 } 5996 5997 void G1CollectedHeap::reset_free_regions_coming() { 5998 assert(free_regions_coming(), "pre-condition"); 5999 6000 { 6001 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6002 _free_regions_coming = false; 6003 SecondaryFreeList_lock->notify_all(); 6004 } 6005 6006 if (G1ConcRegionFreeingVerbose) { 6007 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6008 "reset free regions coming"); 6009 } 6010 } 6011 6012 void G1CollectedHeap::wait_while_free_regions_coming() { 6013 // Most of the time we won't have to wait, so let's do a quick test 6014 // first before we take the lock. 6015 if (!free_regions_coming()) { 6016 return; 6017 } 6018 6019 if (G1ConcRegionFreeingVerbose) { 6020 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6021 "waiting for free regions"); 6022 } 6023 6024 { 6025 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6026 while (free_regions_coming()) { 6027 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6028 } 6029 } 6030 6031 if (G1ConcRegionFreeingVerbose) { 6032 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6033 "done waiting for free regions"); 6034 } 6035 } 6036 6037 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6038 assert(heap_lock_held_for_gc(), 6039 "the heap lock should already be held by or for this thread"); 6040 _young_list->push_region(hr); 6041 } 6042 6043 class NoYoungRegionsClosure: public HeapRegionClosure { 6044 private: 6045 bool _success; 6046 public: 6047 NoYoungRegionsClosure() : _success(true) { } 6048 bool doHeapRegion(HeapRegion* r) { 6049 if (r->is_young()) { 6050 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6051 r->bottom(), r->end()); 6052 _success = false; 6053 } 6054 return false; 6055 } 6056 bool success() { return _success; } 6057 }; 6058 6059 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6060 bool ret = _young_list->check_list_empty(check_sample); 6061 6062 if (check_heap) { 6063 NoYoungRegionsClosure closure; 6064 heap_region_iterate(&closure); 6065 ret = ret && closure.success(); 6066 } 6067 6068 return ret; 6069 } 6070 6071 class TearDownRegionSetsClosure : public HeapRegionClosure { 6072 private: 6073 OldRegionSet *_old_set; 6074 6075 public: 6076 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6077 6078 bool doHeapRegion(HeapRegion* r) { 6079 if (r->is_empty()) { 6080 // We ignore empty regions, we'll empty the free list afterwards 6081 } else if (r->is_young()) { 6082 // We ignore young regions, we'll empty the young list afterwards 6083 } else if (r->isHumongous()) { 6084 // We ignore humongous regions, we're not tearing down the 6085 // humongous region set 6086 } else { 6087 // The rest should be old 6088 _old_set->remove(r); 6089 } 6090 return false; 6091 } 6092 6093 ~TearDownRegionSetsClosure() { 6094 assert(_old_set->is_empty(), "post-condition"); 6095 } 6096 }; 6097 6098 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6099 assert_at_safepoint(true /* should_be_vm_thread */); 6100 6101 if (!free_list_only) { 6102 TearDownRegionSetsClosure cl(&_old_set); 6103 heap_region_iterate(&cl); 6104 6105 // Need to do this after the heap iteration to be able to 6106 // recognize the young regions and ignore them during the iteration. 6107 _young_list->empty_list(); 6108 } 6109 _free_list.remove_all(); 6110 } 6111 6112 class RebuildRegionSetsClosure : public HeapRegionClosure { 6113 private: 6114 bool _free_list_only; 6115 OldRegionSet* _old_set; 6116 FreeRegionList* _free_list; 6117 size_t _total_used; 6118 6119 public: 6120 RebuildRegionSetsClosure(bool free_list_only, 6121 OldRegionSet* old_set, FreeRegionList* free_list) : 6122 _free_list_only(free_list_only), 6123 _old_set(old_set), _free_list(free_list), _total_used(0) { 6124 assert(_free_list->is_empty(), "pre-condition"); 6125 if (!free_list_only) { 6126 assert(_old_set->is_empty(), "pre-condition"); 6127 } 6128 } 6129 6130 bool doHeapRegion(HeapRegion* r) { 6131 if (r->continuesHumongous()) { 6132 return false; 6133 } 6134 6135 if (r->is_empty()) { 6136 // Add free regions to the free list 6137 _free_list->add_as_tail(r); 6138 } else if (!_free_list_only) { 6139 assert(!r->is_young(), "we should not come across young regions"); 6140 6141 if (r->isHumongous()) { 6142 // We ignore humongous regions, we left the humongous set unchanged 6143 } else { 6144 // The rest should be old, add them to the old set 6145 _old_set->add(r); 6146 } 6147 _total_used += r->used(); 6148 } 6149 6150 return false; 6151 } 6152 6153 size_t total_used() { 6154 return _total_used; 6155 } 6156 }; 6157 6158 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6159 assert_at_safepoint(true /* should_be_vm_thread */); 6160 6161 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6162 heap_region_iterate(&cl); 6163 6164 if (!free_list_only) { 6165 _summary_bytes_used = cl.total_used(); 6166 } 6167 assert(_summary_bytes_used == recalculate_used(), 6168 err_msg("inconsistent _summary_bytes_used, " 6169 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6170 _summary_bytes_used, recalculate_used())); 6171 } 6172 6173 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6174 _refine_cte_cl->set_concurrent(concurrent); 6175 } 6176 6177 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6178 HeapRegion* hr = heap_region_containing(p); 6179 if (hr == NULL) { 6180 return is_in_permanent(p); 6181 } else { 6182 return hr->is_in(p); 6183 } 6184 } 6185 6186 // Methods for the mutator alloc region 6187 6188 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6189 bool force) { 6190 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6191 assert(!force || g1_policy()->can_expand_young_list(), 6192 "if force is true we should be able to expand the young list"); 6193 bool young_list_full = g1_policy()->is_young_list_full(); 6194 if (force || !young_list_full) { 6195 HeapRegion* new_alloc_region = new_region(word_size, 6196 false /* do_expand */); 6197 if (new_alloc_region != NULL) { 6198 set_region_short_lived_locked(new_alloc_region); 6199 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6200 return new_alloc_region; 6201 } 6202 } 6203 return NULL; 6204 } 6205 6206 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6207 size_t allocated_bytes) { 6208 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6209 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6210 6211 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6212 _summary_bytes_used += allocated_bytes; 6213 _hr_printer.retire(alloc_region); 6214 // We update the eden sizes here, when the region is retired, 6215 // instead of when it's allocated, since this is the point that its 6216 // used space has been recored in _summary_bytes_used. 6217 g1mm()->update_eden_size(); 6218 } 6219 6220 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6221 bool force) { 6222 return _g1h->new_mutator_alloc_region(word_size, force); 6223 } 6224 6225 void G1CollectedHeap::set_par_threads() { 6226 // Don't change the number of workers. Use the value previously set 6227 // in the workgroup. 6228 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6229 uint n_workers = workers()->active_workers(); 6230 assert(UseDynamicNumberOfGCThreads || 6231 n_workers == workers()->total_workers(), 6232 "Otherwise should be using the total number of workers"); 6233 if (n_workers == 0) { 6234 assert(false, "Should have been set in prior evacuation pause."); 6235 n_workers = ParallelGCThreads; 6236 workers()->set_active_workers(n_workers); 6237 } 6238 set_par_threads(n_workers); 6239 } 6240 6241 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6242 size_t allocated_bytes) { 6243 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6244 } 6245 6246 // Methods for the GC alloc regions 6247 6248 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6249 uint count, 6250 GCAllocPurpose ap) { 6251 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6252 6253 if (count < g1_policy()->max_regions(ap)) { 6254 HeapRegion* new_alloc_region = new_region(word_size, 6255 true /* do_expand */); 6256 if (new_alloc_region != NULL) { 6257 // We really only need to do this for old regions given that we 6258 // should never scan survivors. But it doesn't hurt to do it 6259 // for survivors too. 6260 new_alloc_region->set_saved_mark(); 6261 if (ap == GCAllocForSurvived) { 6262 new_alloc_region->set_survivor(); 6263 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6264 } else { 6265 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6266 } 6267 bool during_im = g1_policy()->during_initial_mark_pause(); 6268 new_alloc_region->note_start_of_copying(during_im); 6269 return new_alloc_region; 6270 } else { 6271 g1_policy()->note_alloc_region_limit_reached(ap); 6272 } 6273 } 6274 return NULL; 6275 } 6276 6277 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6278 size_t allocated_bytes, 6279 GCAllocPurpose ap) { 6280 bool during_im = g1_policy()->during_initial_mark_pause(); 6281 alloc_region->note_end_of_copying(during_im); 6282 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6283 if (ap == GCAllocForSurvived) { 6284 young_list()->add_survivor_region(alloc_region); 6285 } else { 6286 _old_set.add(alloc_region); 6287 } 6288 _hr_printer.retire(alloc_region); 6289 } 6290 6291 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6292 bool force) { 6293 assert(!force, "not supported for GC alloc regions"); 6294 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6295 } 6296 6297 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6298 size_t allocated_bytes) { 6299 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6300 GCAllocForSurvived); 6301 } 6302 6303 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6304 bool force) { 6305 assert(!force, "not supported for GC alloc regions"); 6306 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6307 } 6308 6309 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6310 size_t allocated_bytes) { 6311 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6312 GCAllocForTenured); 6313 } 6314 // Heap region set verification 6315 6316 class VerifyRegionListsClosure : public HeapRegionClosure { 6317 private: 6318 FreeRegionList* _free_list; 6319 OldRegionSet* _old_set; 6320 HumongousRegionSet* _humongous_set; 6321 uint _region_count; 6322 6323 public: 6324 VerifyRegionListsClosure(OldRegionSet* old_set, 6325 HumongousRegionSet* humongous_set, 6326 FreeRegionList* free_list) : 6327 _old_set(old_set), _humongous_set(humongous_set), 6328 _free_list(free_list), _region_count(0) { } 6329 6330 uint region_count() { return _region_count; } 6331 6332 bool doHeapRegion(HeapRegion* hr) { 6333 _region_count += 1; 6334 6335 if (hr->continuesHumongous()) { 6336 return false; 6337 } 6338 6339 if (hr->is_young()) { 6340 // TODO 6341 } else if (hr->startsHumongous()) { 6342 _humongous_set->verify_next_region(hr); 6343 } else if (hr->is_empty()) { 6344 _free_list->verify_next_region(hr); 6345 } else { 6346 _old_set->verify_next_region(hr); 6347 } 6348 return false; 6349 } 6350 }; 6351 6352 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6353 HeapWord* bottom) { 6354 HeapWord* end = bottom + HeapRegion::GrainWords; 6355 MemRegion mr(bottom, end); 6356 assert(_g1_reserved.contains(mr), "invariant"); 6357 // This might return NULL if the allocation fails 6358 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6359 } 6360 6361 void G1CollectedHeap::verify_region_sets() { 6362 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6363 6364 // First, check the explicit lists. 6365 _free_list.verify(); 6366 { 6367 // Given that a concurrent operation might be adding regions to 6368 // the secondary free list we have to take the lock before 6369 // verifying it. 6370 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6371 _secondary_free_list.verify(); 6372 } 6373 _old_set.verify(); 6374 _humongous_set.verify(); 6375 6376 // If a concurrent region freeing operation is in progress it will 6377 // be difficult to correctly attributed any free regions we come 6378 // across to the correct free list given that they might belong to 6379 // one of several (free_list, secondary_free_list, any local lists, 6380 // etc.). So, if that's the case we will skip the rest of the 6381 // verification operation. Alternatively, waiting for the concurrent 6382 // operation to complete will have a non-trivial effect on the GC's 6383 // operation (no concurrent operation will last longer than the 6384 // interval between two calls to verification) and it might hide 6385 // any issues that we would like to catch during testing. 6386 if (free_regions_coming()) { 6387 return; 6388 } 6389 6390 // Make sure we append the secondary_free_list on the free_list so 6391 // that all free regions we will come across can be safely 6392 // attributed to the free_list. 6393 append_secondary_free_list_if_not_empty_with_lock(); 6394 6395 // Finally, make sure that the region accounting in the lists is 6396 // consistent with what we see in the heap. 6397 _old_set.verify_start(); 6398 _humongous_set.verify_start(); 6399 _free_list.verify_start(); 6400 6401 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6402 heap_region_iterate(&cl); 6403 6404 _old_set.verify_end(); 6405 _humongous_set.verify_end(); 6406 _free_list.verify_end(); 6407 }