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/g1GCPhaseTimes.hpp" 37 #include "gc_implementation/g1/g1Log.hpp" 38 #include "gc_implementation/g1/g1MarkSweep.hpp" 39 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 40 #include "gc_implementation/g1/g1RemSet.inline.hpp" 41 #include "gc_implementation/g1/heapRegion.inline.hpp" 42 #include "gc_implementation/g1/heapRegionRemSet.hpp" 43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 44 #include "gc_implementation/g1/vm_operations_g1.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "memory/gcLocker.inline.hpp" 47 #include "memory/genOopClosures.inline.hpp" 48 #include "memory/generationSpec.hpp" 49 #include "memory/referenceProcessor.hpp" 50 #include "oops/oop.inline.hpp" 51 #include "oops/oop.pcgc.inline.hpp" 52 #include "runtime/aprofiler.hpp" 53 #include "runtime/vmThread.hpp" 54 55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 56 57 // turn it on so that the contents of the young list (scan-only / 58 // to-be-collected) are printed at "strategic" points before / during 59 // / after the collection --- this is useful for debugging 60 #define YOUNG_LIST_VERBOSE 0 61 // CURRENT STATUS 62 // This file is under construction. Search for "FIXME". 63 64 // INVARIANTS/NOTES 65 // 66 // All allocation activity covered by the G1CollectedHeap interface is 67 // serialized by acquiring the HeapLock. This happens in mem_allocate 68 // and allocate_new_tlab, which are the "entry" points to the 69 // allocation code from the rest of the JVM. (Note that this does not 70 // apply to TLAB allocation, which is not part of this interface: it 71 // is done by clients of this interface.) 72 73 // Notes on implementation of parallelism in different tasks. 74 // 75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 76 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 77 // It does use run_task() which sets _n_workers in the task. 78 // G1ParTask executes g1_process_strong_roots() -> 79 // SharedHeap::process_strong_roots() which calls eventuall to 80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 83 // 84 85 // Local to this file. 86 87 class RefineCardTableEntryClosure: public CardTableEntryClosure { 88 SuspendibleThreadSet* _sts; 89 G1RemSet* _g1rs; 90 ConcurrentG1Refine* _cg1r; 91 bool _concurrent; 92 public: 93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 94 G1RemSet* g1rs, 95 ConcurrentG1Refine* cg1r) : 96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 97 {} 98 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 99 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false); 100 // This path is executed by the concurrent refine or mutator threads, 101 // concurrently, and so we do not care if card_ptr contains references 102 // that point into the collection set. 103 assert(!oops_into_cset, "should be"); 104 105 if (_concurrent && _sts->should_yield()) { 106 // Caller will actually yield. 107 return false; 108 } 109 // Otherwise, we finished successfully; return true. 110 return true; 111 } 112 void set_concurrent(bool b) { _concurrent = b; } 113 }; 114 115 116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 117 int _calls; 118 G1CollectedHeap* _g1h; 119 CardTableModRefBS* _ctbs; 120 int _histo[256]; 121 public: 122 ClearLoggedCardTableEntryClosure() : 123 _calls(0) 124 { 125 _g1h = G1CollectedHeap::heap(); 126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 127 for (int i = 0; i < 256; i++) _histo[i] = 0; 128 } 129 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 131 _calls++; 132 unsigned char* ujb = (unsigned char*)card_ptr; 133 int ind = (int)(*ujb); 134 _histo[ind]++; 135 *card_ptr = -1; 136 } 137 return true; 138 } 139 int calls() { return _calls; } 140 void print_histo() { 141 gclog_or_tty->print_cr("Card table value histogram:"); 142 for (int i = 0; i < 256; i++) { 143 if (_histo[i] != 0) { 144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 145 } 146 } 147 } 148 }; 149 150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 151 int _calls; 152 G1CollectedHeap* _g1h; 153 CardTableModRefBS* _ctbs; 154 public: 155 RedirtyLoggedCardTableEntryClosure() : 156 _calls(0) 157 { 158 _g1h = G1CollectedHeap::heap(); 159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 160 } 161 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 163 _calls++; 164 *card_ptr = 0; 165 } 166 return true; 167 } 168 int calls() { return _calls; } 169 }; 170 171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 172 public: 173 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 174 *card_ptr = CardTableModRefBS::dirty_card_val(); 175 return true; 176 } 177 }; 178 179 YoungList::YoungList(G1CollectedHeap* g1h) : 180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 182 guarantee(check_list_empty(false), "just making sure..."); 183 } 184 185 void YoungList::push_region(HeapRegion *hr) { 186 assert(!hr->is_young(), "should not already be young"); 187 assert(hr->get_next_young_region() == NULL, "cause it should!"); 188 189 hr->set_next_young_region(_head); 190 _head = hr; 191 192 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 193 ++_length; 194 } 195 196 void YoungList::add_survivor_region(HeapRegion* hr) { 197 assert(hr->is_survivor(), "should be flagged as survivor region"); 198 assert(hr->get_next_young_region() == NULL, "cause it should!"); 199 200 hr->set_next_young_region(_survivor_head); 201 if (_survivor_head == NULL) { 202 _survivor_tail = hr; 203 } 204 _survivor_head = hr; 205 ++_survivor_length; 206 } 207 208 void YoungList::empty_list(HeapRegion* list) { 209 while (list != NULL) { 210 HeapRegion* next = list->get_next_young_region(); 211 list->set_next_young_region(NULL); 212 list->uninstall_surv_rate_group(); 213 list->set_not_young(); 214 list = next; 215 } 216 } 217 218 void YoungList::empty_list() { 219 assert(check_list_well_formed(), "young list should be well formed"); 220 221 empty_list(_head); 222 _head = NULL; 223 _length = 0; 224 225 empty_list(_survivor_head); 226 _survivor_head = NULL; 227 _survivor_tail = NULL; 228 _survivor_length = 0; 229 230 _last_sampled_rs_lengths = 0; 231 232 assert(check_list_empty(false), "just making sure..."); 233 } 234 235 bool YoungList::check_list_well_formed() { 236 bool ret = true; 237 238 uint length = 0; 239 HeapRegion* curr = _head; 240 HeapRegion* last = NULL; 241 while (curr != NULL) { 242 if (!curr->is_young()) { 243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 244 "incorrectly tagged (y: %d, surv: %d)", 245 curr->bottom(), curr->end(), 246 curr->is_young(), curr->is_survivor()); 247 ret = false; 248 } 249 ++length; 250 last = curr; 251 curr = curr->get_next_young_region(); 252 } 253 ret = ret && (length == _length); 254 255 if (!ret) { 256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 257 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 258 length, _length); 259 } 260 261 return ret; 262 } 263 264 bool YoungList::check_list_empty(bool check_sample) { 265 bool ret = true; 266 267 if (_length != 0) { 268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 269 _length); 270 ret = false; 271 } 272 if (check_sample && _last_sampled_rs_lengths != 0) { 273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 274 ret = false; 275 } 276 if (_head != NULL) { 277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 278 ret = false; 279 } 280 if (!ret) { 281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 282 } 283 284 return ret; 285 } 286 287 void 288 YoungList::rs_length_sampling_init() { 289 _sampled_rs_lengths = 0; 290 _curr = _head; 291 } 292 293 bool 294 YoungList::rs_length_sampling_more() { 295 return _curr != NULL; 296 } 297 298 void 299 YoungList::rs_length_sampling_next() { 300 assert( _curr != NULL, "invariant" ); 301 size_t rs_length = _curr->rem_set()->occupied(); 302 303 _sampled_rs_lengths += rs_length; 304 305 // The current region may not yet have been added to the 306 // incremental collection set (it gets added when it is 307 // retired as the current allocation region). 308 if (_curr->in_collection_set()) { 309 // Update the collection set policy information for this region 310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 311 } 312 313 _curr = _curr->get_next_young_region(); 314 if (_curr == NULL) { 315 _last_sampled_rs_lengths = _sampled_rs_lengths; 316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 317 } 318 } 319 320 void 321 YoungList::reset_auxilary_lists() { 322 guarantee( is_empty(), "young list should be empty" ); 323 assert(check_list_well_formed(), "young list should be well formed"); 324 325 // Add survivor regions to SurvRateGroup. 326 _g1h->g1_policy()->note_start_adding_survivor_regions(); 327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 328 329 int young_index_in_cset = 0; 330 for (HeapRegion* curr = _survivor_head; 331 curr != NULL; 332 curr = curr->get_next_young_region()) { 333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 334 335 // The region is a non-empty survivor so let's add it to 336 // the incremental collection set for the next evacuation 337 // pause. 338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 339 young_index_in_cset += 1; 340 } 341 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 342 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 343 344 _head = _survivor_head; 345 _length = _survivor_length; 346 if (_survivor_head != NULL) { 347 assert(_survivor_tail != NULL, "cause it shouldn't be"); 348 assert(_survivor_length > 0, "invariant"); 349 _survivor_tail->set_next_young_region(NULL); 350 } 351 352 // Don't clear the survivor list handles until the start of 353 // the next evacuation pause - we need it in order to re-tag 354 // the survivor regions from this evacuation pause as 'young' 355 // at the start of the next. 356 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 358 359 assert(check_list_well_formed(), "young list should be well formed"); 360 } 361 362 void YoungList::print() { 363 HeapRegion* lists[] = {_head, _survivor_head}; 364 const char* names[] = {"YOUNG", "SURVIVOR"}; 365 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 368 HeapRegion *curr = lists[list]; 369 if (curr == NULL) 370 gclog_or_tty->print_cr(" empty"); 371 while (curr != NULL) { 372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 373 HR_FORMAT_PARAMS(curr), 374 curr->prev_top_at_mark_start(), 375 curr->next_top_at_mark_start(), 376 curr->age_in_surv_rate_group_cond()); 377 curr = curr->get_next_young_region(); 378 } 379 } 380 381 gclog_or_tty->print_cr(""); 382 } 383 384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 385 { 386 // Claim the right to put the region on the dirty cards region list 387 // by installing a self pointer. 388 HeapRegion* next = hr->get_next_dirty_cards_region(); 389 if (next == NULL) { 390 HeapRegion* res = (HeapRegion*) 391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 392 NULL); 393 if (res == NULL) { 394 HeapRegion* head; 395 do { 396 // Put the region to the dirty cards region list. 397 head = _dirty_cards_region_list; 398 next = (HeapRegion*) 399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 400 if (next == head) { 401 assert(hr->get_next_dirty_cards_region() == hr, 402 "hr->get_next_dirty_cards_region() != hr"); 403 if (next == NULL) { 404 // The last region in the list points to itself. 405 hr->set_next_dirty_cards_region(hr); 406 } else { 407 hr->set_next_dirty_cards_region(next); 408 } 409 } 410 } while (next != head); 411 } 412 } 413 } 414 415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 416 { 417 HeapRegion* head; 418 HeapRegion* hr; 419 do { 420 head = _dirty_cards_region_list; 421 if (head == NULL) { 422 return NULL; 423 } 424 HeapRegion* new_head = head->get_next_dirty_cards_region(); 425 if (head == new_head) { 426 // The last region. 427 new_head = NULL; 428 } 429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 430 head); 431 } while (hr != head); 432 assert(hr != NULL, "invariant"); 433 hr->set_next_dirty_cards_region(NULL); 434 return hr; 435 } 436 437 void G1CollectedHeap::stop_conc_gc_threads() { 438 _cg1r->stop(); 439 _cmThread->stop(); 440 } 441 442 #ifdef ASSERT 443 // A region is added to the collection set as it is retired 444 // so an address p can point to a region which will be in the 445 // collection set but has not yet been retired. This method 446 // therefore is only accurate during a GC pause after all 447 // regions have been retired. It is used for debugging 448 // to check if an nmethod has references to objects that can 449 // be move during a partial collection. Though it can be 450 // inaccurate, it is sufficient for G1 because the conservative 451 // implementation of is_scavengable() for G1 will indicate that 452 // all nmethods must be scanned during a partial collection. 453 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 454 HeapRegion* hr = heap_region_containing(p); 455 return hr != NULL && hr->in_collection_set(); 456 } 457 #endif 458 459 // Returns true if the reference points to an object that 460 // can move in an incremental collecction. 461 bool G1CollectedHeap::is_scavengable(const void* p) { 462 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 463 G1CollectorPolicy* g1p = g1h->g1_policy(); 464 HeapRegion* hr = heap_region_containing(p); 465 if (hr == NULL) { 466 // null 467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p)); 468 return false; 469 } else { 470 return !hr->isHumongous(); 471 } 472 } 473 474 void G1CollectedHeap::check_ct_logs_at_safepoint() { 475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 477 478 // Count the dirty cards at the start. 479 CountNonCleanMemRegionClosure count1(this); 480 ct_bs->mod_card_iterate(&count1); 481 int orig_count = count1.n(); 482 483 // First clear the logged cards. 484 ClearLoggedCardTableEntryClosure clear; 485 dcqs.set_closure(&clear); 486 dcqs.apply_closure_to_all_completed_buffers(); 487 dcqs.iterate_closure_all_threads(false); 488 clear.print_histo(); 489 490 // Now ensure that there's no dirty cards. 491 CountNonCleanMemRegionClosure count2(this); 492 ct_bs->mod_card_iterate(&count2); 493 if (count2.n() != 0) { 494 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 495 count2.n(), orig_count); 496 } 497 guarantee(count2.n() == 0, "Card table should be clean."); 498 499 RedirtyLoggedCardTableEntryClosure redirty; 500 JavaThread::dirty_card_queue_set().set_closure(&redirty); 501 dcqs.apply_closure_to_all_completed_buffers(); 502 dcqs.iterate_closure_all_threads(false); 503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 504 clear.calls(), orig_count); 505 guarantee(redirty.calls() == clear.calls(), 506 "Or else mechanism is broken."); 507 508 CountNonCleanMemRegionClosure count3(this); 509 ct_bs->mod_card_iterate(&count3); 510 if (count3.n() != orig_count) { 511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 512 orig_count, count3.n()); 513 guarantee(count3.n() >= orig_count, "Should have restored them all."); 514 } 515 516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 517 } 518 519 // Private class members. 520 521 G1CollectedHeap* G1CollectedHeap::_g1h; 522 523 // Private methods. 524 525 HeapRegion* 526 G1CollectedHeap::new_region_try_secondary_free_list() { 527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 528 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 529 if (!_secondary_free_list.is_empty()) { 530 if (G1ConcRegionFreeingVerbose) { 531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 532 "secondary_free_list has %u entries", 533 _secondary_free_list.length()); 534 } 535 // It looks as if there are free regions available on the 536 // secondary_free_list. Let's move them to the free_list and try 537 // again to allocate from it. 538 append_secondary_free_list(); 539 540 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 541 "empty we should have moved at least one entry to the free_list"); 542 HeapRegion* res = _free_list.remove_head(); 543 if (G1ConcRegionFreeingVerbose) { 544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 545 "allocated "HR_FORMAT" from secondary_free_list", 546 HR_FORMAT_PARAMS(res)); 547 } 548 return res; 549 } 550 551 // Wait here until we get notifed either when (a) there are no 552 // more free regions coming or (b) some regions have been moved on 553 // the secondary_free_list. 554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 555 } 556 557 if (G1ConcRegionFreeingVerbose) { 558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 559 "could not allocate from secondary_free_list"); 560 } 561 return NULL; 562 } 563 564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 566 "the only time we use this to allocate a humongous region is " 567 "when we are allocating a single humongous region"); 568 569 HeapRegion* res; 570 if (G1StressConcRegionFreeing) { 571 if (!_secondary_free_list.is_empty()) { 572 if (G1ConcRegionFreeingVerbose) { 573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 574 "forced to look at the secondary_free_list"); 575 } 576 res = new_region_try_secondary_free_list(); 577 if (res != NULL) { 578 return res; 579 } 580 } 581 } 582 res = _free_list.remove_head_or_null(); 583 if (res == NULL) { 584 if (G1ConcRegionFreeingVerbose) { 585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 586 "res == NULL, trying the secondary_free_list"); 587 } 588 res = new_region_try_secondary_free_list(); 589 } 590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 591 // Currently, only attempts to allocate GC alloc regions set 592 // do_expand to true. So, we should only reach here during a 593 // safepoint. If this assumption changes we might have to 594 // reconsider the use of _expand_heap_after_alloc_failure. 595 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 596 597 ergo_verbose1(ErgoHeapSizing, 598 "attempt heap expansion", 599 ergo_format_reason("region allocation request failed") 600 ergo_format_byte("allocation request"), 601 word_size * HeapWordSize); 602 if (expand(word_size * HeapWordSize)) { 603 // Given that expand() succeeded in expanding the heap, and we 604 // always expand the heap by an amount aligned to the heap 605 // region size, the free list should in theory not be empty. So 606 // it would probably be OK to use remove_head(). But the extra 607 // check for NULL is unlikely to be a performance issue here (we 608 // just expanded the heap!) so let's just be conservative and 609 // use remove_head_or_null(). 610 res = _free_list.remove_head_or_null(); 611 } else { 612 _expand_heap_after_alloc_failure = false; 613 } 614 } 615 return res; 616 } 617 618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 619 size_t word_size) { 620 assert(isHumongous(word_size), "word_size should be humongous"); 621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 622 623 uint first = G1_NULL_HRS_INDEX; 624 if (num_regions == 1) { 625 // Only one region to allocate, no need to go through the slower 626 // path. The caller will attempt the expasion if this fails, so 627 // let's not try to expand here too. 628 HeapRegion* hr = new_region(word_size, false /* do_expand */); 629 if (hr != NULL) { 630 first = hr->hrs_index(); 631 } else { 632 first = G1_NULL_HRS_INDEX; 633 } 634 } else { 635 // We can't allocate humongous regions while cleanupComplete() is 636 // running, since some of the regions we find to be empty might not 637 // yet be added to the free list and it is not straightforward to 638 // know which list they are on so that we can remove them. Note 639 // that we only need to do this if we need to allocate more than 640 // one region to satisfy the current humongous allocation 641 // request. If we are only allocating one region we use the common 642 // region allocation code (see above). 643 wait_while_free_regions_coming(); 644 append_secondary_free_list_if_not_empty_with_lock(); 645 646 if (free_regions() >= num_regions) { 647 first = _hrs.find_contiguous(num_regions); 648 if (first != G1_NULL_HRS_INDEX) { 649 for (uint i = first; i < first + num_regions; ++i) { 650 HeapRegion* hr = region_at(i); 651 assert(hr->is_empty(), "sanity"); 652 assert(is_on_master_free_list(hr), "sanity"); 653 hr->set_pending_removal(true); 654 } 655 _free_list.remove_all_pending(num_regions); 656 } 657 } 658 } 659 return first; 660 } 661 662 HeapWord* 663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 664 uint num_regions, 665 size_t word_size) { 666 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 667 assert(isHumongous(word_size), "word_size should be humongous"); 668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 669 670 // Index of last region in the series + 1. 671 uint last = first + num_regions; 672 673 // We need to initialize the region(s) we just discovered. This is 674 // a bit tricky given that it can happen concurrently with 675 // refinement threads refining cards on these regions and 676 // potentially wanting to refine the BOT as they are scanning 677 // those cards (this can happen shortly after a cleanup; see CR 678 // 6991377). So we have to set up the region(s) carefully and in 679 // a specific order. 680 681 // The word size sum of all the regions we will allocate. 682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 683 assert(word_size <= word_size_sum, "sanity"); 684 685 // This will be the "starts humongous" region. 686 HeapRegion* first_hr = region_at(first); 687 // The header of the new object will be placed at the bottom of 688 // the first region. 689 HeapWord* new_obj = first_hr->bottom(); 690 // This will be the new end of the first region in the series that 691 // should also match the end of the last region in the seriers. 692 HeapWord* new_end = new_obj + word_size_sum; 693 // This will be the new top of the first region that will reflect 694 // this allocation. 695 HeapWord* new_top = new_obj + word_size; 696 697 // First, we need to zero the header of the space that we will be 698 // allocating. When we update top further down, some refinement 699 // threads might try to scan the region. By zeroing the header we 700 // ensure that any thread that will try to scan the region will 701 // come across the zero klass word and bail out. 702 // 703 // NOTE: It would not have been correct to have used 704 // CollectedHeap::fill_with_object() and make the space look like 705 // an int array. The thread that is doing the allocation will 706 // later update the object header to a potentially different array 707 // type and, for a very short period of time, the klass and length 708 // fields will be inconsistent. This could cause a refinement 709 // thread to calculate the object size incorrectly. 710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 711 712 // We will set up the first region as "starts humongous". This 713 // will also update the BOT covering all the regions to reflect 714 // that there is a single object that starts at the bottom of the 715 // first region. 716 first_hr->set_startsHumongous(new_top, new_end); 717 718 // Then, if there are any, we will set up the "continues 719 // humongous" regions. 720 HeapRegion* hr = NULL; 721 for (uint i = first + 1; i < last; ++i) { 722 hr = region_at(i); 723 hr->set_continuesHumongous(first_hr); 724 } 725 // If we have "continues humongous" regions (hr != NULL), then the 726 // end of the last one should match new_end. 727 assert(hr == NULL || hr->end() == new_end, "sanity"); 728 729 // Up to this point no concurrent thread would have been able to 730 // do any scanning on any region in this series. All the top 731 // fields still point to bottom, so the intersection between 732 // [bottom,top] and [card_start,card_end] will be empty. Before we 733 // update the top fields, we'll do a storestore to make sure that 734 // no thread sees the update to top before the zeroing of the 735 // object header and the BOT initialization. 736 OrderAccess::storestore(); 737 738 // Now that the BOT and the object header have been initialized, 739 // we can update top of the "starts humongous" region. 740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 741 "new_top should be in this region"); 742 first_hr->set_top(new_top); 743 if (_hr_printer.is_active()) { 744 HeapWord* bottom = first_hr->bottom(); 745 HeapWord* end = first_hr->orig_end(); 746 if ((first + 1) == last) { 747 // the series has a single humongous region 748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 749 } else { 750 // the series has more than one humongous regions 751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 752 } 753 } 754 755 // Now, we will update the top fields of the "continues humongous" 756 // regions. The reason we need to do this is that, otherwise, 757 // these regions would look empty and this will confuse parts of 758 // G1. For example, the code that looks for a consecutive number 759 // of empty regions will consider them empty and try to 760 // re-allocate them. We can extend is_empty() to also include 761 // !continuesHumongous(), but it is easier to just update the top 762 // fields here. The way we set top for all regions (i.e., top == 763 // end for all regions but the last one, top == new_top for the 764 // last one) is actually used when we will free up the humongous 765 // region in free_humongous_region(). 766 hr = NULL; 767 for (uint i = first + 1; i < last; ++i) { 768 hr = region_at(i); 769 if ((i + 1) == last) { 770 // last continues humongous region 771 assert(hr->bottom() < new_top && new_top <= hr->end(), 772 "new_top should fall on this region"); 773 hr->set_top(new_top); 774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 775 } else { 776 // not last one 777 assert(new_top > hr->end(), "new_top should be above this region"); 778 hr->set_top(hr->end()); 779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 780 } 781 } 782 // If we have continues humongous regions (hr != NULL), then the 783 // end of the last one should match new_end and its top should 784 // match new_top. 785 assert(hr == NULL || 786 (hr->end() == new_end && hr->top() == new_top), "sanity"); 787 788 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 789 _summary_bytes_used += first_hr->used(); 790 _humongous_set.add(first_hr); 791 792 return new_obj; 793 } 794 795 // If could fit into free regions w/o expansion, try. 796 // Otherwise, if can expand, do so. 797 // Otherwise, if using ex regions might help, try with ex given back. 798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 800 801 verify_region_sets_optional(); 802 803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 805 uint x_num = expansion_regions(); 806 uint fs = _hrs.free_suffix(); 807 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 808 if (first == G1_NULL_HRS_INDEX) { 809 // The only thing we can do now is attempt expansion. 810 if (fs + x_num >= num_regions) { 811 // If the number of regions we're trying to allocate for this 812 // object is at most the number of regions in the free suffix, 813 // then the call to humongous_obj_allocate_find_first() above 814 // should have succeeded and we wouldn't be here. 815 // 816 // We should only be trying to expand when the free suffix is 817 // not sufficient for the object _and_ we have some expansion 818 // room available. 819 assert(num_regions > fs, "earlier allocation should have succeeded"); 820 821 ergo_verbose1(ErgoHeapSizing, 822 "attempt heap expansion", 823 ergo_format_reason("humongous allocation request failed") 824 ergo_format_byte("allocation request"), 825 word_size * HeapWordSize); 826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 827 // Even though the heap was expanded, it might not have 828 // reached the desired size. So, we cannot assume that the 829 // allocation will succeed. 830 first = humongous_obj_allocate_find_first(num_regions, word_size); 831 } 832 } 833 } 834 835 HeapWord* result = NULL; 836 if (first != G1_NULL_HRS_INDEX) { 837 result = 838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 839 assert(result != NULL, "it should always return a valid result"); 840 841 // A successful humongous object allocation changes the used space 842 // information of the old generation so we need to recalculate the 843 // sizes and update the jstat counters here. 844 g1mm()->update_sizes(); 845 } 846 847 verify_region_sets_optional(); 848 849 return result; 850 } 851 852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 853 assert_heap_not_locked_and_not_at_safepoint(); 854 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 855 856 unsigned int dummy_gc_count_before; 857 return attempt_allocation(word_size, &dummy_gc_count_before); 858 } 859 860 HeapWord* 861 G1CollectedHeap::mem_allocate(size_t word_size, 862 bool* gc_overhead_limit_was_exceeded) { 863 assert_heap_not_locked_and_not_at_safepoint(); 864 865 // Loop until the allocation is satisified, or unsatisfied after GC. 866 for (int try_count = 1; /* we'll return */; try_count += 1) { 867 unsigned int gc_count_before; 868 869 HeapWord* result = NULL; 870 if (!isHumongous(word_size)) { 871 result = attempt_allocation(word_size, &gc_count_before); 872 } else { 873 result = attempt_allocation_humongous(word_size, &gc_count_before); 874 } 875 if (result != NULL) { 876 return result; 877 } 878 879 // Create the garbage collection operation... 880 VM_G1CollectForAllocation op(gc_count_before, word_size); 881 // ...and get the VM thread to execute it. 882 VMThread::execute(&op); 883 884 if (op.prologue_succeeded() && op.pause_succeeded()) { 885 // If the operation was successful we'll return the result even 886 // if it is NULL. If the allocation attempt failed immediately 887 // after a Full GC, it's unlikely we'll be able to allocate now. 888 HeapWord* result = op.result(); 889 if (result != NULL && !isHumongous(word_size)) { 890 // Allocations that take place on VM operations do not do any 891 // card dirtying and we have to do it here. We only have to do 892 // this for non-humongous allocations, though. 893 dirty_young_block(result, word_size); 894 } 895 return result; 896 } else { 897 assert(op.result() == NULL, 898 "the result should be NULL if the VM op did not succeed"); 899 } 900 901 // Give a warning if we seem to be looping forever. 902 if ((QueuedAllocationWarningCount > 0) && 903 (try_count % QueuedAllocationWarningCount == 0)) { 904 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 905 } 906 } 907 908 ShouldNotReachHere(); 909 return NULL; 910 } 911 912 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 913 unsigned int *gc_count_before_ret) { 914 // Make sure you read the note in attempt_allocation_humongous(). 915 916 assert_heap_not_locked_and_not_at_safepoint(); 917 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 918 "be called for humongous allocation requests"); 919 920 // We should only get here after the first-level allocation attempt 921 // (attempt_allocation()) failed to allocate. 922 923 // We will loop until a) we manage to successfully perform the 924 // allocation or b) we successfully schedule a collection which 925 // fails to perform the allocation. b) is the only case when we'll 926 // return NULL. 927 HeapWord* result = NULL; 928 for (int try_count = 1; /* we'll return */; try_count += 1) { 929 bool should_try_gc; 930 unsigned int gc_count_before; 931 932 { 933 MutexLockerEx x(Heap_lock); 934 935 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 936 false /* bot_updates */); 937 if (result != NULL) { 938 return result; 939 } 940 941 // If we reach here, attempt_allocation_locked() above failed to 942 // allocate a new region. So the mutator alloc region should be NULL. 943 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 944 945 if (GC_locker::is_active_and_needs_gc()) { 946 if (g1_policy()->can_expand_young_list()) { 947 // No need for an ergo verbose message here, 948 // can_expand_young_list() does this when it returns true. 949 result = _mutator_alloc_region.attempt_allocation_force(word_size, 950 false /* bot_updates */); 951 if (result != NULL) { 952 return result; 953 } 954 } 955 should_try_gc = false; 956 } else { 957 // The GCLocker may not be active but the GCLocker initiated 958 // GC may not yet have been performed (GCLocker::needs_gc() 959 // returns true). In this case we do not try this GC and 960 // wait until the GCLocker initiated GC is performed, and 961 // then retry the allocation. 962 if (GC_locker::needs_gc()) { 963 should_try_gc = false; 964 } else { 965 // Read the GC count while still holding the Heap_lock. 966 gc_count_before = total_collections(); 967 should_try_gc = true; 968 } 969 } 970 } 971 972 if (should_try_gc) { 973 bool succeeded; 974 result = do_collection_pause(word_size, gc_count_before, &succeeded); 975 if (result != NULL) { 976 assert(succeeded, "only way to get back a non-NULL result"); 977 return result; 978 } 979 980 if (succeeded) { 981 // If we get here we successfully scheduled a collection which 982 // failed to allocate. No point in trying to allocate 983 // further. We'll just return NULL. 984 MutexLockerEx x(Heap_lock); 985 *gc_count_before_ret = total_collections(); 986 return NULL; 987 } 988 } else { 989 // The GCLocker is either active or the GCLocker initiated 990 // GC has not yet been performed. Stall until it is and 991 // then retry the allocation. 992 GC_locker::stall_until_clear(); 993 } 994 995 // We can reach here if we were unsuccessul in scheduling a 996 // collection (because another thread beat us to it) or if we were 997 // stalled due to the GC locker. In either can we should retry the 998 // allocation attempt in case another thread successfully 999 // performed a collection and reclaimed enough space. We do the 1000 // first attempt (without holding the Heap_lock) here and the 1001 // follow-on attempt will be at the start of the next loop 1002 // iteration (after taking the Heap_lock). 1003 result = _mutator_alloc_region.attempt_allocation(word_size, 1004 false /* bot_updates */); 1005 if (result != NULL) { 1006 return result; 1007 } 1008 1009 // Give a warning if we seem to be looping forever. 1010 if ((QueuedAllocationWarningCount > 0) && 1011 (try_count % QueuedAllocationWarningCount == 0)) { 1012 warning("G1CollectedHeap::attempt_allocation_slow() " 1013 "retries %d times", try_count); 1014 } 1015 } 1016 1017 ShouldNotReachHere(); 1018 return NULL; 1019 } 1020 1021 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1022 unsigned int * gc_count_before_ret) { 1023 // The structure of this method has a lot of similarities to 1024 // attempt_allocation_slow(). The reason these two were not merged 1025 // into a single one is that such a method would require several "if 1026 // allocation is not humongous do this, otherwise do that" 1027 // conditional paths which would obscure its flow. In fact, an early 1028 // version of this code did use a unified method which was harder to 1029 // follow and, as a result, it had subtle bugs that were hard to 1030 // track down. So keeping these two methods separate allows each to 1031 // be more readable. It will be good to keep these two in sync as 1032 // much as possible. 1033 1034 assert_heap_not_locked_and_not_at_safepoint(); 1035 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1036 "should only be called for humongous allocations"); 1037 1038 // Humongous objects can exhaust the heap quickly, so we should check if we 1039 // need to start a marking cycle at each humongous object allocation. We do 1040 // the check before we do the actual allocation. The reason for doing it 1041 // before the allocation is that we avoid having to keep track of the newly 1042 // allocated memory while we do a GC. 1043 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1044 word_size)) { 1045 collect(GCCause::_g1_humongous_allocation); 1046 } 1047 1048 // We will loop until a) we manage to successfully perform the 1049 // allocation or b) we successfully schedule a collection which 1050 // fails to perform the allocation. b) is the only case when we'll 1051 // return NULL. 1052 HeapWord* result = NULL; 1053 for (int try_count = 1; /* we'll return */; try_count += 1) { 1054 bool should_try_gc; 1055 unsigned int gc_count_before; 1056 1057 { 1058 MutexLockerEx x(Heap_lock); 1059 1060 // Given that humongous objects are not allocated in young 1061 // regions, we'll first try to do the allocation without doing a 1062 // collection hoping that there's enough space in the heap. 1063 result = humongous_obj_allocate(word_size); 1064 if (result != NULL) { 1065 return result; 1066 } 1067 1068 if (GC_locker::is_active_and_needs_gc()) { 1069 should_try_gc = false; 1070 } else { 1071 // The GCLocker may not be active but the GCLocker initiated 1072 // GC may not yet have been performed (GCLocker::needs_gc() 1073 // returns true). In this case we do not try this GC and 1074 // wait until the GCLocker initiated GC is performed, and 1075 // then retry the allocation. 1076 if (GC_locker::needs_gc()) { 1077 should_try_gc = false; 1078 } else { 1079 // Read the GC count while still holding the Heap_lock. 1080 gc_count_before = total_collections(); 1081 should_try_gc = true; 1082 } 1083 } 1084 } 1085 1086 if (should_try_gc) { 1087 // If we failed to allocate the humongous object, we should try to 1088 // do a collection pause (if we're allowed) in case it reclaims 1089 // enough space for the allocation to succeed after the pause. 1090 1091 bool succeeded; 1092 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1093 if (result != NULL) { 1094 assert(succeeded, "only way to get back a non-NULL result"); 1095 return result; 1096 } 1097 1098 if (succeeded) { 1099 // If we get here we successfully scheduled a collection which 1100 // failed to allocate. No point in trying to allocate 1101 // further. We'll just return NULL. 1102 MutexLockerEx x(Heap_lock); 1103 *gc_count_before_ret = total_collections(); 1104 return NULL; 1105 } 1106 } else { 1107 // The GCLocker is either active or the GCLocker initiated 1108 // GC has not yet been performed. Stall until it is and 1109 // then retry the allocation. 1110 GC_locker::stall_until_clear(); 1111 } 1112 1113 // We can reach here if we were unsuccessul in scheduling a 1114 // collection (because another thread beat us to it) or if we were 1115 // stalled due to the GC locker. In either can we should retry the 1116 // allocation attempt in case another thread successfully 1117 // performed a collection and reclaimed enough space. Give a 1118 // warning if we seem to be looping forever. 1119 1120 if ((QueuedAllocationWarningCount > 0) && 1121 (try_count % QueuedAllocationWarningCount == 0)) { 1122 warning("G1CollectedHeap::attempt_allocation_humongous() " 1123 "retries %d times", try_count); 1124 } 1125 } 1126 1127 ShouldNotReachHere(); 1128 return NULL; 1129 } 1130 1131 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1132 bool expect_null_mutator_alloc_region) { 1133 assert_at_safepoint(true /* should_be_vm_thread */); 1134 assert(_mutator_alloc_region.get() == NULL || 1135 !expect_null_mutator_alloc_region, 1136 "the current alloc region was unexpectedly found to be non-NULL"); 1137 1138 if (!isHumongous(word_size)) { 1139 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1140 false /* bot_updates */); 1141 } else { 1142 HeapWord* result = humongous_obj_allocate(word_size); 1143 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1144 g1_policy()->set_initiate_conc_mark_if_possible(); 1145 } 1146 return result; 1147 } 1148 1149 ShouldNotReachHere(); 1150 } 1151 1152 class PostMCRemSetClearClosure: public HeapRegionClosure { 1153 G1CollectedHeap* _g1h; 1154 ModRefBarrierSet* _mr_bs; 1155 public: 1156 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) : 1157 _g1h(g1h), _mr_bs(mr_bs) { } 1158 bool doHeapRegion(HeapRegion* r) { 1159 if (r->continuesHumongous()) { 1160 return false; 1161 } 1162 _g1h->reset_gc_time_stamps(r); 1163 HeapRegionRemSet* hrrs = r->rem_set(); 1164 if (hrrs != NULL) hrrs->clear(); 1165 // You might think here that we could clear just the cards 1166 // corresponding to the used region. But no: if we leave a dirty card 1167 // in a region we might allocate into, then it would prevent that card 1168 // from being enqueued, and cause it to be missed. 1169 // Re: the performance cost: we shouldn't be doing full GC anyway! 1170 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1171 return false; 1172 } 1173 }; 1174 1175 void G1CollectedHeap::clear_rsets_post_compaction() { 1176 PostMCRemSetClearClosure rs_clear(this, mr_bs()); 1177 heap_region_iterate(&rs_clear); 1178 } 1179 1180 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1181 G1CollectedHeap* _g1h; 1182 UpdateRSOopClosure _cl; 1183 int _worker_i; 1184 public: 1185 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1186 _cl(g1->g1_rem_set(), worker_i), 1187 _worker_i(worker_i), 1188 _g1h(g1) 1189 { } 1190 1191 bool doHeapRegion(HeapRegion* r) { 1192 if (!r->continuesHumongous()) { 1193 _cl.set_from(r); 1194 r->oop_iterate(&_cl); 1195 } 1196 return false; 1197 } 1198 }; 1199 1200 class ParRebuildRSTask: public AbstractGangTask { 1201 G1CollectedHeap* _g1; 1202 public: 1203 ParRebuildRSTask(G1CollectedHeap* g1) 1204 : AbstractGangTask("ParRebuildRSTask"), 1205 _g1(g1) 1206 { } 1207 1208 void work(uint worker_id) { 1209 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1210 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1211 _g1->workers()->active_workers(), 1212 HeapRegion::RebuildRSClaimValue); 1213 } 1214 }; 1215 1216 class PostCompactionPrinterClosure: public HeapRegionClosure { 1217 private: 1218 G1HRPrinter* _hr_printer; 1219 public: 1220 bool doHeapRegion(HeapRegion* hr) { 1221 assert(!hr->is_young(), "not expecting to find young regions"); 1222 // We only generate output for non-empty regions. 1223 if (!hr->is_empty()) { 1224 if (!hr->isHumongous()) { 1225 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1226 } else if (hr->startsHumongous()) { 1227 if (hr->region_num() == 1) { 1228 // single humongous region 1229 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1230 } else { 1231 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1232 } 1233 } else { 1234 assert(hr->continuesHumongous(), "only way to get here"); 1235 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1236 } 1237 } 1238 return false; 1239 } 1240 1241 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1242 : _hr_printer(hr_printer) { } 1243 }; 1244 1245 void G1CollectedHeap::print_hrs_post_compaction() { 1246 PostCompactionPrinterClosure cl(hr_printer()); 1247 heap_region_iterate(&cl); 1248 } 1249 1250 double G1CollectedHeap::verify(bool guard, const char* msg) { 1251 double verify_time_ms = 0.0; 1252 1253 if (guard && total_collections() >= VerifyGCStartAt) { 1254 double verify_start = os::elapsedTime(); 1255 HandleMark hm; // Discard invalid handles created during verification 1256 gclog_or_tty->print(msg); 1257 prepare_for_verify(); 1258 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking); 1259 verify_time_ms = (os::elapsedTime() - verify_start) * 1000; 1260 } 1261 1262 return verify_time_ms; 1263 } 1264 1265 void G1CollectedHeap::verify_before_gc() { 1266 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:"); 1267 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms); 1268 } 1269 1270 void G1CollectedHeap::verify_after_gc() { 1271 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:"); 1272 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms); 1273 } 1274 1275 bool G1CollectedHeap::do_collection(bool explicit_gc, 1276 bool clear_all_soft_refs, 1277 size_t word_size) { 1278 assert_at_safepoint(true /* should_be_vm_thread */); 1279 1280 if (GC_locker::check_active_before_gc()) { 1281 return false; 1282 } 1283 1284 SvcGCMarker sgcm(SvcGCMarker::FULL); 1285 ResourceMark rm; 1286 1287 print_heap_before_gc(); 1288 1289 size_t metadata_prev_used = MetaspaceAux::used_in_bytes(); 1290 1291 HRSPhaseSetter x(HRSPhaseFullGC); 1292 verify_region_sets_optional(); 1293 1294 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1295 collector_policy()->should_clear_all_soft_refs(); 1296 1297 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1298 1299 { 1300 IsGCActiveMark x; 1301 1302 // Timing 1303 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1304 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1305 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1306 1307 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty); 1308 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1309 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1310 1311 double start = os::elapsedTime(); 1312 g1_policy()->record_full_collection_start(); 1313 1314 // Note: When we have a more flexible GC logging framework that 1315 // allows us to add optional attributes to a GC log record we 1316 // could consider timing and reporting how long we wait in the 1317 // following two methods. 1318 wait_while_free_regions_coming(); 1319 // If we start the compaction before the CM threads finish 1320 // scanning the root regions we might trip them over as we'll 1321 // be moving objects / updating references. So let's wait until 1322 // they are done. By telling them to abort, they should complete 1323 // early. 1324 _cm->root_regions()->abort(); 1325 _cm->root_regions()->wait_until_scan_finished(); 1326 append_secondary_free_list_if_not_empty_with_lock(); 1327 1328 gc_prologue(true); 1329 increment_total_collections(true /* full gc */); 1330 increment_old_marking_cycles_started(); 1331 1332 size_t g1h_prev_used = used(); 1333 assert(used() == recalculate_used(), "Should be equal"); 1334 1335 verify_before_gc(); 1336 1337 pre_full_gc_dump(); 1338 1339 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1340 1341 // Disable discovery and empty the discovered lists 1342 // for the CM ref processor. 1343 ref_processor_cm()->disable_discovery(); 1344 ref_processor_cm()->abandon_partial_discovery(); 1345 ref_processor_cm()->verify_no_references_recorded(); 1346 1347 // Abandon current iterations of concurrent marking and concurrent 1348 // refinement, if any are in progress. We have to do this before 1349 // wait_until_scan_finished() below. 1350 concurrent_mark()->abort(); 1351 1352 // Make sure we'll choose a new allocation region afterwards. 1353 release_mutator_alloc_region(); 1354 abandon_gc_alloc_regions(); 1355 g1_rem_set()->cleanupHRRS(); 1356 1357 // We should call this after we retire any currently active alloc 1358 // regions so that all the ALLOC / RETIRE events are generated 1359 // before the start GC event. 1360 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1361 1362 // We may have added regions to the current incremental collection 1363 // set between the last GC or pause and now. We need to clear the 1364 // incremental collection set and then start rebuilding it afresh 1365 // after this full GC. 1366 abandon_collection_set(g1_policy()->inc_cset_head()); 1367 g1_policy()->clear_incremental_cset(); 1368 g1_policy()->stop_incremental_cset_building(); 1369 1370 tear_down_region_sets(false /* free_list_only */); 1371 g1_policy()->set_gcs_are_young(true); 1372 1373 // See the comments in g1CollectedHeap.hpp and 1374 // G1CollectedHeap::ref_processing_init() about 1375 // how reference processing currently works in G1. 1376 1377 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1378 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1379 1380 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1381 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1382 1383 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1384 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1385 1386 // Do collection work 1387 { 1388 HandleMark hm; // Discard invalid handles created during gc 1389 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1390 } 1391 1392 assert(free_regions() == 0, "we should not have added any free regions"); 1393 rebuild_region_sets(false /* free_list_only */); 1394 1395 // Enqueue any discovered reference objects that have 1396 // not been removed from the discovered lists. 1397 ref_processor_stw()->enqueue_discovered_references(); 1398 1399 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1400 1401 MemoryService::track_memory_usage(); 1402 1403 verify_after_gc(); 1404 1405 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1406 ref_processor_stw()->verify_no_references_recorded(); 1407 1408 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1409 ClassLoaderDataGraph::purge(); 1410 1411 // Note: since we've just done a full GC, concurrent 1412 // marking is no longer active. Therefore we need not 1413 // re-enable reference discovery for the CM ref processor. 1414 // That will be done at the start of the next marking cycle. 1415 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1416 ref_processor_cm()->verify_no_references_recorded(); 1417 1418 reset_gc_time_stamp(); 1419 // Since everything potentially moved, we will clear all remembered 1420 // sets, and clear all cards. Later we will rebuild remebered 1421 // sets. We will also reset the GC time stamps of the regions. 1422 clear_rsets_post_compaction(); 1423 check_gc_time_stamps(); 1424 1425 // Resize the heap if necessary. 1426 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1427 1428 if (_hr_printer.is_active()) { 1429 // We should do this after we potentially resize the heap so 1430 // that all the COMMIT / UNCOMMIT events are generated before 1431 // the end GC event. 1432 1433 print_hrs_post_compaction(); 1434 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1435 } 1436 1437 if (_cg1r->use_cache()) { 1438 _cg1r->clear_and_record_card_counts(); 1439 _cg1r->clear_hot_cache(); 1440 } 1441 1442 // Rebuild remembered sets of all regions. 1443 if (G1CollectedHeap::use_parallel_gc_threads()) { 1444 uint n_workers = 1445 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1446 workers()->active_workers(), 1447 Threads::number_of_non_daemon_threads()); 1448 assert(UseDynamicNumberOfGCThreads || 1449 n_workers == workers()->total_workers(), 1450 "If not dynamic should be using all the workers"); 1451 workers()->set_active_workers(n_workers); 1452 // Set parallel threads in the heap (_n_par_threads) only 1453 // before a parallel phase and always reset it to 0 after 1454 // the phase so that the number of parallel threads does 1455 // no get carried forward to a serial phase where there 1456 // may be code that is "possibly_parallel". 1457 set_par_threads(n_workers); 1458 1459 ParRebuildRSTask rebuild_rs_task(this); 1460 assert(check_heap_region_claim_values( 1461 HeapRegion::InitialClaimValue), "sanity check"); 1462 assert(UseDynamicNumberOfGCThreads || 1463 workers()->active_workers() == workers()->total_workers(), 1464 "Unless dynamic should use total workers"); 1465 // Use the most recent number of active workers 1466 assert(workers()->active_workers() > 0, 1467 "Active workers not properly set"); 1468 set_par_threads(workers()->active_workers()); 1469 workers()->run_task(&rebuild_rs_task); 1470 set_par_threads(0); 1471 assert(check_heap_region_claim_values( 1472 HeapRegion::RebuildRSClaimValue), "sanity check"); 1473 reset_heap_region_claim_values(); 1474 } else { 1475 RebuildRSOutOfRegionClosure rebuild_rs(this); 1476 heap_region_iterate(&rebuild_rs); 1477 } 1478 1479 if (G1Log::fine()) { 1480 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1481 } 1482 1483 if (true) { // FIXME 1484 MetaspaceGC::compute_new_size(); 1485 } 1486 1487 // Start a new incremental collection set for the next pause 1488 assert(g1_policy()->collection_set() == NULL, "must be"); 1489 g1_policy()->start_incremental_cset_building(); 1490 1491 // Clear the _cset_fast_test bitmap in anticipation of adding 1492 // regions to the incremental collection set for the next 1493 // evacuation pause. 1494 clear_cset_fast_test(); 1495 1496 init_mutator_alloc_region(); 1497 1498 double end = os::elapsedTime(); 1499 g1_policy()->record_full_collection_end(); 1500 1501 #ifdef TRACESPINNING 1502 ParallelTaskTerminator::print_termination_counts(); 1503 #endif 1504 1505 gc_epilogue(true); 1506 1507 // Discard all rset updates 1508 JavaThread::dirty_card_queue_set().abandon_logs(); 1509 assert(!G1DeferredRSUpdate 1510 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1511 1512 _young_list->reset_sampled_info(); 1513 // At this point there should be no regions in the 1514 // entire heap tagged as young. 1515 assert( check_young_list_empty(true /* check_heap */), 1516 "young list should be empty at this point"); 1517 1518 // Update the number of full collections that have been completed. 1519 increment_old_marking_cycles_completed(false /* concurrent */); 1520 1521 _hrs.verify_optional(); 1522 verify_region_sets_optional(); 1523 1524 print_heap_after_gc(); 1525 1526 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1527 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1528 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1529 // before any GC notifications are raised. 1530 g1mm()->update_sizes(); 1531 } 1532 1533 post_full_gc_dump(); 1534 1535 return true; 1536 } 1537 1538 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1539 // do_collection() will return whether it succeeded in performing 1540 // the GC. Currently, there is no facility on the 1541 // do_full_collection() API to notify the caller than the collection 1542 // did not succeed (e.g., because it was locked out by the GC 1543 // locker). So, right now, we'll ignore the return value. 1544 bool dummy = do_collection(true, /* explicit_gc */ 1545 clear_all_soft_refs, 1546 0 /* word_size */); 1547 } 1548 1549 // This code is mostly copied from TenuredGeneration. 1550 void 1551 G1CollectedHeap:: 1552 resize_if_necessary_after_full_collection(size_t word_size) { 1553 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1554 1555 // Include the current allocation, if any, and bytes that will be 1556 // pre-allocated to support collections, as "used". 1557 const size_t used_after_gc = used(); 1558 const size_t capacity_after_gc = capacity(); 1559 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1560 1561 // This is enforced in arguments.cpp. 1562 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1563 "otherwise the code below doesn't make sense"); 1564 1565 // We don't have floating point command-line arguments 1566 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1567 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1568 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1569 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1570 1571 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1572 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1573 1574 // We have to be careful here as these two calculations can overflow 1575 // 32-bit size_t's. 1576 double used_after_gc_d = (double) used_after_gc; 1577 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1578 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1579 1580 // Let's make sure that they are both under the max heap size, which 1581 // by default will make them fit into a size_t. 1582 double desired_capacity_upper_bound = (double) max_heap_size; 1583 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1584 desired_capacity_upper_bound); 1585 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1586 desired_capacity_upper_bound); 1587 1588 // We can now safely turn them into size_t's. 1589 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1590 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1591 1592 // This assert only makes sense here, before we adjust them 1593 // with respect to the min and max heap size. 1594 assert(minimum_desired_capacity <= maximum_desired_capacity, 1595 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1596 "maximum_desired_capacity = "SIZE_FORMAT, 1597 minimum_desired_capacity, maximum_desired_capacity)); 1598 1599 // Should not be greater than the heap max size. No need to adjust 1600 // it with respect to the heap min size as it's a lower bound (i.e., 1601 // we'll try to make the capacity larger than it, not smaller). 1602 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1603 // Should not be less than the heap min size. No need to adjust it 1604 // with respect to the heap max size as it's an upper bound (i.e., 1605 // we'll try to make the capacity smaller than it, not greater). 1606 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1607 1608 if (capacity_after_gc < minimum_desired_capacity) { 1609 // Don't expand unless it's significant 1610 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1611 ergo_verbose4(ErgoHeapSizing, 1612 "attempt heap expansion", 1613 ergo_format_reason("capacity lower than " 1614 "min desired capacity after Full GC") 1615 ergo_format_byte("capacity") 1616 ergo_format_byte("occupancy") 1617 ergo_format_byte_perc("min desired capacity"), 1618 capacity_after_gc, used_after_gc, 1619 minimum_desired_capacity, (double) MinHeapFreeRatio); 1620 expand(expand_bytes); 1621 1622 // No expansion, now see if we want to shrink 1623 } else if (capacity_after_gc > maximum_desired_capacity) { 1624 // Capacity too large, compute shrinking size 1625 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1626 ergo_verbose4(ErgoHeapSizing, 1627 "attempt heap shrinking", 1628 ergo_format_reason("capacity higher than " 1629 "max desired capacity after Full GC") 1630 ergo_format_byte("capacity") 1631 ergo_format_byte("occupancy") 1632 ergo_format_byte_perc("max desired capacity"), 1633 capacity_after_gc, used_after_gc, 1634 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1635 shrink(shrink_bytes); 1636 } 1637 } 1638 1639 1640 HeapWord* 1641 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1642 bool* succeeded) { 1643 assert_at_safepoint(true /* should_be_vm_thread */); 1644 1645 *succeeded = true; 1646 // Let's attempt the allocation first. 1647 HeapWord* result = 1648 attempt_allocation_at_safepoint(word_size, 1649 false /* expect_null_mutator_alloc_region */); 1650 if (result != NULL) { 1651 assert(*succeeded, "sanity"); 1652 return result; 1653 } 1654 1655 // In a G1 heap, we're supposed to keep allocation from failing by 1656 // incremental pauses. Therefore, at least for now, we'll favor 1657 // expansion over collection. (This might change in the future if we can 1658 // do something smarter than full collection to satisfy a failed alloc.) 1659 result = expand_and_allocate(word_size); 1660 if (result != NULL) { 1661 assert(*succeeded, "sanity"); 1662 return result; 1663 } 1664 1665 // Expansion didn't work, we'll try to do a Full GC. 1666 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1667 false, /* clear_all_soft_refs */ 1668 word_size); 1669 if (!gc_succeeded) { 1670 *succeeded = false; 1671 return NULL; 1672 } 1673 1674 // Retry the allocation 1675 result = attempt_allocation_at_safepoint(word_size, 1676 true /* expect_null_mutator_alloc_region */); 1677 if (result != NULL) { 1678 assert(*succeeded, "sanity"); 1679 return result; 1680 } 1681 1682 // Then, try a Full GC that will collect all soft references. 1683 gc_succeeded = do_collection(false, /* explicit_gc */ 1684 true, /* clear_all_soft_refs */ 1685 word_size); 1686 if (!gc_succeeded) { 1687 *succeeded = false; 1688 return NULL; 1689 } 1690 1691 // Retry the allocation once more 1692 result = attempt_allocation_at_safepoint(word_size, 1693 true /* expect_null_mutator_alloc_region */); 1694 if (result != NULL) { 1695 assert(*succeeded, "sanity"); 1696 return result; 1697 } 1698 1699 assert(!collector_policy()->should_clear_all_soft_refs(), 1700 "Flag should have been handled and cleared prior to this point"); 1701 1702 // What else? We might try synchronous finalization later. If the total 1703 // space available is large enough for the allocation, then a more 1704 // complete compaction phase than we've tried so far might be 1705 // appropriate. 1706 assert(*succeeded, "sanity"); 1707 return NULL; 1708 } 1709 1710 // Attempting to expand the heap sufficiently 1711 // to support an allocation of the given "word_size". If 1712 // successful, perform the allocation and return the address of the 1713 // allocated block, or else "NULL". 1714 1715 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1716 assert_at_safepoint(true /* should_be_vm_thread */); 1717 1718 verify_region_sets_optional(); 1719 1720 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1721 ergo_verbose1(ErgoHeapSizing, 1722 "attempt heap expansion", 1723 ergo_format_reason("allocation request failed") 1724 ergo_format_byte("allocation request"), 1725 word_size * HeapWordSize); 1726 if (expand(expand_bytes)) { 1727 _hrs.verify_optional(); 1728 verify_region_sets_optional(); 1729 return attempt_allocation_at_safepoint(word_size, 1730 false /* expect_null_mutator_alloc_region */); 1731 } 1732 return NULL; 1733 } 1734 1735 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1736 HeapWord* new_end) { 1737 assert(old_end != new_end, "don't call this otherwise"); 1738 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1739 1740 // Update the committed mem region. 1741 _g1_committed.set_end(new_end); 1742 // Tell the card table about the update. 1743 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1744 // Tell the BOT about the update. 1745 _bot_shared->resize(_g1_committed.word_size()); 1746 } 1747 1748 bool G1CollectedHeap::expand(size_t expand_bytes) { 1749 size_t old_mem_size = _g1_storage.committed_size(); 1750 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1751 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1752 HeapRegion::GrainBytes); 1753 ergo_verbose2(ErgoHeapSizing, 1754 "expand the heap", 1755 ergo_format_byte("requested expansion amount") 1756 ergo_format_byte("attempted expansion amount"), 1757 expand_bytes, aligned_expand_bytes); 1758 1759 // First commit the memory. 1760 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1761 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1762 if (successful) { 1763 // Then propagate this update to the necessary data structures. 1764 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1765 update_committed_space(old_end, new_end); 1766 1767 FreeRegionList expansion_list("Local Expansion List"); 1768 MemRegion mr = _hrs.expand_to(new_end, &expansion_list); 1769 assert(mr.start() == old_end, "post-condition"); 1770 // mr might be a smaller region than what was requested if 1771 // expand_by() was unable to allocate the HeapRegion instances 1772 assert(mr.end() <= new_end, "post-condition"); 1773 1774 size_t actual_expand_bytes = mr.byte_size(); 1775 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1776 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1777 "post-condition"); 1778 if (actual_expand_bytes < aligned_expand_bytes) { 1779 // We could not expand _hrs to the desired size. In this case we 1780 // need to shrink the committed space accordingly. 1781 assert(mr.end() < new_end, "invariant"); 1782 1783 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1784 // First uncommit the memory. 1785 _g1_storage.shrink_by(diff_bytes); 1786 // Then propagate this update to the necessary data structures. 1787 update_committed_space(new_end, mr.end()); 1788 } 1789 _free_list.add_as_tail(&expansion_list); 1790 1791 if (_hr_printer.is_active()) { 1792 HeapWord* curr = mr.start(); 1793 while (curr < mr.end()) { 1794 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1795 _hr_printer.commit(curr, curr_end); 1796 curr = curr_end; 1797 } 1798 assert(curr == mr.end(), "post-condition"); 1799 } 1800 g1_policy()->record_new_heap_size(n_regions()); 1801 } else { 1802 ergo_verbose0(ErgoHeapSizing, 1803 "did not expand the heap", 1804 ergo_format_reason("heap expansion operation failed")); 1805 // The expansion of the virtual storage space was unsuccessful. 1806 // Let's see if it was because we ran out of swap. 1807 if (G1ExitOnExpansionFailure && 1808 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1809 // We had head room... 1810 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1811 } 1812 } 1813 return successful; 1814 } 1815 1816 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1817 size_t old_mem_size = _g1_storage.committed_size(); 1818 size_t aligned_shrink_bytes = 1819 ReservedSpace::page_align_size_down(shrink_bytes); 1820 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1821 HeapRegion::GrainBytes); 1822 uint num_regions_deleted = 0; 1823 HeapWord* desired_new_end = 1824 (HeapWord*) _g1_storage.high() - (aligned_shrink_bytes / HeapWordSize); 1825 MemRegion mr = _hrs.shrink_to(desired_new_end, &num_regions_deleted); 1826 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1827 assert(mr.end() == old_end, "post-condition"); 1828 1829 ergo_verbose3(ErgoHeapSizing, 1830 "shrink the heap", 1831 ergo_format_byte("requested shrinking amount") 1832 ergo_format_byte("aligned shrinking amount") 1833 ergo_format_byte("attempted shrinking amount"), 1834 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1835 if (mr.byte_size() > 0) { 1836 if (_hr_printer.is_active()) { 1837 HeapWord* curr = mr.end(); 1838 while (curr > mr.start()) { 1839 HeapWord* curr_end = curr; 1840 curr -= HeapRegion::GrainWords; 1841 _hr_printer.uncommit(curr, curr_end); 1842 } 1843 assert(curr == mr.start(), "post-condition"); 1844 } 1845 1846 _g1_storage.shrink_by(mr.byte_size()); 1847 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1848 assert(mr.start() == new_end, "post-condition"); 1849 1850 _expansion_regions += num_regions_deleted; 1851 update_committed_space(old_end, new_end); 1852 HeapRegionRemSet::shrink_heap(n_regions()); 1853 g1_policy()->record_new_heap_size(n_regions()); 1854 } else { 1855 ergo_verbose0(ErgoHeapSizing, 1856 "did not shrink the heap", 1857 ergo_format_reason("heap shrinking operation failed")); 1858 } 1859 } 1860 1861 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1862 verify_region_sets_optional(); 1863 1864 // We should only reach here at the end of a Full GC which means we 1865 // should not not be holding to any GC alloc regions. The method 1866 // below will make sure of that and do any remaining clean up. 1867 abandon_gc_alloc_regions(); 1868 1869 // Instead of tearing down / rebuilding the free lists here, we 1870 // could instead use the remove_all_pending() method on free_list to 1871 // remove only the ones that we need to remove. 1872 tear_down_region_sets(true /* free_list_only */); 1873 shrink_helper(shrink_bytes); 1874 rebuild_region_sets(true /* free_list_only */); 1875 1876 _hrs.verify_optional(); 1877 verify_region_sets_optional(); 1878 } 1879 1880 // Public methods. 1881 1882 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1883 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1884 #endif // _MSC_VER 1885 1886 1887 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1888 SharedHeap(policy_), 1889 _g1_policy(policy_), 1890 _dirty_card_queue_set(false), 1891 _into_cset_dirty_card_queue_set(false), 1892 _is_alive_closure_cm(this), 1893 _is_alive_closure_stw(this), 1894 _ref_processor_cm(NULL), 1895 _ref_processor_stw(NULL), 1896 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1897 _bot_shared(NULL), 1898 _evac_failure_scan_stack(NULL) , 1899 _mark_in_progress(false), 1900 _cg1r(NULL), _summary_bytes_used(0), 1901 _g1mm(NULL), 1902 _refine_cte_cl(NULL), 1903 _full_collection(false), 1904 _free_list("Master Free List"), 1905 _secondary_free_list("Secondary Free List"), 1906 _old_set("Old Set"), 1907 _humongous_set("Master Humongous Set"), 1908 _free_regions_coming(false), 1909 _young_list(new YoungList(this)), 1910 _gc_time_stamp(0), 1911 _retained_old_gc_alloc_region(NULL), 1912 _survivor_plab_stats(YoungPLABSize, PLABWeight), 1913 _old_plab_stats(OldPLABSize, PLABWeight), 1914 _expand_heap_after_alloc_failure(true), 1915 _surviving_young_words(NULL), 1916 _old_marking_cycles_started(0), 1917 _old_marking_cycles_completed(0), 1918 _in_cset_fast_test(NULL), 1919 _in_cset_fast_test_base(NULL), 1920 _dirty_cards_region_list(NULL), 1921 _worker_cset_start_region(NULL), 1922 _worker_cset_start_region_time_stamp(NULL) { 1923 _g1h = this; // To catch bugs. 1924 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1925 vm_exit_during_initialization("Failed necessary allocation."); 1926 } 1927 1928 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1929 1930 int n_queues = MAX2((int)ParallelGCThreads, 1); 1931 _task_queues = new RefToScanQueueSet(n_queues); 1932 1933 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1934 assert(n_rem_sets > 0, "Invariant."); 1935 1936 HeapRegionRemSetIterator** iter_arr = 1937 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC); 1938 for (int i = 0; i < n_queues; i++) { 1939 iter_arr[i] = new HeapRegionRemSetIterator(); 1940 } 1941 _rem_set_iterator = iter_arr; 1942 1943 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1944 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1945 1946 for (int i = 0; i < n_queues; i++) { 1947 RefToScanQueue* q = new RefToScanQueue(); 1948 q->initialize(); 1949 _task_queues->register_queue(i, q); 1950 } 1951 1952 clear_cset_start_regions(); 1953 1954 // Initialize the G1EvacuationFailureALot counters and flags. 1955 NOT_PRODUCT(reset_evacuation_should_fail();) 1956 1957 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1958 } 1959 1960 jint G1CollectedHeap::initialize() { 1961 CollectedHeap::pre_initialize(); 1962 os::enable_vtime(); 1963 1964 G1Log::init(); 1965 1966 // Necessary to satisfy locking discipline assertions. 1967 1968 MutexLocker x(Heap_lock); 1969 1970 // We have to initialize the printer before committing the heap, as 1971 // it will be used then. 1972 _hr_printer.set_active(G1PrintHeapRegions); 1973 1974 // While there are no constraints in the GC code that HeapWordSize 1975 // be any particular value, there are multiple other areas in the 1976 // system which believe this to be true (e.g. oop->object_size in some 1977 // cases incorrectly returns the size in wordSize units rather than 1978 // HeapWordSize). 1979 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1980 1981 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1982 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1983 1984 // Ensure that the sizes are properly aligned. 1985 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1986 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1987 1988 _cg1r = new ConcurrentG1Refine(); 1989 1990 // Reserve the maximum. 1991 1992 // When compressed oops are enabled, the preferred heap base 1993 // is calculated by subtracting the requested size from the 1994 // 32Gb boundary and using the result as the base address for 1995 // heap reservation. If the requested size is not aligned to 1996 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1997 // into the ReservedHeapSpace constructor) then the actual 1998 // base of the reserved heap may end up differing from the 1999 // address that was requested (i.e. the preferred heap base). 2000 // If this happens then we could end up using a non-optimal 2001 // compressed oops mode. 2002 2003 // Since max_byte_size is aligned to the size of a heap region (checked 2004 // above). 2005 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 2006 2007 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size, 2008 HeapRegion::GrainBytes); 2009 2010 // It is important to do this in a way such that concurrent readers can't 2011 // temporarily think somethings in the heap. (I've actually seen this 2012 // happen in asserts: DLD.) 2013 _reserved.set_word_size(0); 2014 _reserved.set_start((HeapWord*)heap_rs.base()); 2015 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2016 2017 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2018 2019 // Create the gen rem set (and barrier set) for the entire reserved region. 2020 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2021 set_barrier_set(rem_set()->bs()); 2022 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2023 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2024 } else { 2025 vm_exit_during_initialization("G1 requires a mod ref bs."); 2026 return JNI_ENOMEM; 2027 } 2028 2029 // Also create a G1 rem set. 2030 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2031 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2032 } else { 2033 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2034 return JNI_ENOMEM; 2035 } 2036 2037 // Carve out the G1 part of the heap. 2038 2039 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2040 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2041 g1_rs.size()/HeapWordSize); 2042 2043 _g1_storage.initialize(g1_rs, 0); 2044 _g1_committed = MemRegion((HeapWord*) _g1_storage.low(), (size_t) 0); 2045 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2046 (HeapWord*) _g1_reserved.end()); 2047 assert(_hrs.max_length() == _expansion_regions, 2048 err_msg("max length: %u expansion regions: %u", 2049 _hrs.max_length(), _expansion_regions)); 2050 2051 // 6843694 - ensure that the maximum region index can fit 2052 // in the remembered set structures. 2053 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2054 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2055 2056 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2057 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2058 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2059 "too many cards per region"); 2060 2061 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2062 2063 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2064 heap_word_size(init_byte_size)); 2065 2066 _g1h = this; 2067 2068 _in_cset_fast_test_length = max_regions(); 2069 _in_cset_fast_test_base = 2070 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2071 2072 // We're biasing _in_cset_fast_test to avoid subtracting the 2073 // beginning of the heap every time we want to index; basically 2074 // it's the same with what we do with the card table. 2075 _in_cset_fast_test = _in_cset_fast_test_base - 2076 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2077 2078 // Clear the _cset_fast_test bitmap in anticipation of adding 2079 // regions to the incremental collection set for the first 2080 // evacuation pause. 2081 clear_cset_fast_test(); 2082 2083 // Create the ConcurrentMark data structure and thread. 2084 // (Must do this late, so that "max_regions" is defined.) 2085 _cm = new ConcurrentMark(this, heap_rs); 2086 if (_cm == NULL || !_cm->completed_initialization()) { 2087 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark"); 2088 return JNI_ENOMEM; 2089 } 2090 _cmThread = _cm->cmThread(); 2091 2092 // Initialize the from_card cache structure of HeapRegionRemSet. 2093 HeapRegionRemSet::init_heap(max_regions()); 2094 2095 // Now expand into the initial heap size. 2096 if (!expand(init_byte_size)) { 2097 vm_shutdown_during_initialization("Failed to allocate initial heap."); 2098 return JNI_ENOMEM; 2099 } 2100 2101 // Perform any initialization actions delegated to the policy. 2102 g1_policy()->init(); 2103 2104 _refine_cte_cl = 2105 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2106 g1_rem_set(), 2107 concurrent_g1_refine()); 2108 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2109 2110 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2111 SATB_Q_FL_lock, 2112 G1SATBProcessCompletedThreshold, 2113 Shared_SATB_Q_lock); 2114 2115 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2116 DirtyCardQ_FL_lock, 2117 concurrent_g1_refine()->yellow_zone(), 2118 concurrent_g1_refine()->red_zone(), 2119 Shared_DirtyCardQ_lock); 2120 2121 if (G1DeferredRSUpdate) { 2122 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2123 DirtyCardQ_FL_lock, 2124 -1, // never trigger processing 2125 -1, // no limit on length 2126 Shared_DirtyCardQ_lock, 2127 &JavaThread::dirty_card_queue_set()); 2128 } 2129 2130 // Initialize the card queue set used to hold cards containing 2131 // references into the collection set. 2132 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2133 DirtyCardQ_FL_lock, 2134 -1, // never trigger processing 2135 -1, // no limit on length 2136 Shared_DirtyCardQ_lock, 2137 &JavaThread::dirty_card_queue_set()); 2138 2139 // In case we're keeping closure specialization stats, initialize those 2140 // counts and that mechanism. 2141 SpecializationStats::clear(); 2142 2143 // Do later initialization work for concurrent refinement. 2144 _cg1r->init(); 2145 2146 // Here we allocate the dummy full region that is required by the 2147 // G1AllocRegion class. If we don't pass an address in the reserved 2148 // space here, lots of asserts fire. 2149 2150 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2151 _g1_reserved.start()); 2152 // We'll re-use the same region whether the alloc region will 2153 // require BOT updates or not and, if it doesn't, then a non-young 2154 // region will complain that it cannot support allocations without 2155 // BOT updates. So we'll tag the dummy region as young to avoid that. 2156 dummy_region->set_young(); 2157 // Make sure it's full. 2158 dummy_region->set_top(dummy_region->end()); 2159 G1AllocRegion::setup(this, dummy_region); 2160 2161 init_mutator_alloc_region(); 2162 2163 // Do create of the monitoring and management support so that 2164 // values in the heap have been properly initialized. 2165 _g1mm = new G1MonitoringSupport(this); 2166 2167 return JNI_OK; 2168 } 2169 2170 void G1CollectedHeap::ref_processing_init() { 2171 // Reference processing in G1 currently works as follows: 2172 // 2173 // * There are two reference processor instances. One is 2174 // used to record and process discovered references 2175 // during concurrent marking; the other is used to 2176 // record and process references during STW pauses 2177 // (both full and incremental). 2178 // * Both ref processors need to 'span' the entire heap as 2179 // the regions in the collection set may be dotted around. 2180 // 2181 // * For the concurrent marking ref processor: 2182 // * Reference discovery is enabled at initial marking. 2183 // * Reference discovery is disabled and the discovered 2184 // references processed etc during remarking. 2185 // * Reference discovery is MT (see below). 2186 // * Reference discovery requires a barrier (see below). 2187 // * Reference processing may or may not be MT 2188 // (depending on the value of ParallelRefProcEnabled 2189 // and ParallelGCThreads). 2190 // * A full GC disables reference discovery by the CM 2191 // ref processor and abandons any entries on it's 2192 // discovered lists. 2193 // 2194 // * For the STW processor: 2195 // * Non MT discovery is enabled at the start of a full GC. 2196 // * Processing and enqueueing during a full GC is non-MT. 2197 // * During a full GC, references are processed after marking. 2198 // 2199 // * Discovery (may or may not be MT) is enabled at the start 2200 // of an incremental evacuation pause. 2201 // * References are processed near the end of a STW evacuation pause. 2202 // * For both types of GC: 2203 // * Discovery is atomic - i.e. not concurrent. 2204 // * Reference discovery will not need a barrier. 2205 2206 SharedHeap::ref_processing_init(); 2207 MemRegion mr = reserved_region(); 2208 2209 // Concurrent Mark ref processor 2210 _ref_processor_cm = 2211 new ReferenceProcessor(mr, // span 2212 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2213 // mt processing 2214 (int) ParallelGCThreads, 2215 // degree of mt processing 2216 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2217 // mt discovery 2218 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2219 // degree of mt discovery 2220 false, 2221 // Reference discovery is not atomic 2222 &_is_alive_closure_cm, 2223 // is alive closure 2224 // (for efficiency/performance) 2225 true); 2226 // Setting next fields of discovered 2227 // lists requires a barrier. 2228 2229 // STW ref processor 2230 _ref_processor_stw = 2231 new ReferenceProcessor(mr, // span 2232 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2233 // mt processing 2234 MAX2((int)ParallelGCThreads, 1), 2235 // degree of mt processing 2236 (ParallelGCThreads > 1), 2237 // mt discovery 2238 MAX2((int)ParallelGCThreads, 1), 2239 // degree of mt discovery 2240 true, 2241 // Reference discovery is atomic 2242 &_is_alive_closure_stw, 2243 // is alive closure 2244 // (for efficiency/performance) 2245 false); 2246 // Setting next fields of discovered 2247 // lists requires a barrier. 2248 } 2249 2250 size_t G1CollectedHeap::capacity() const { 2251 return _g1_committed.byte_size(); 2252 } 2253 2254 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) { 2255 assert(!hr->continuesHumongous(), "pre-condition"); 2256 hr->reset_gc_time_stamp(); 2257 if (hr->startsHumongous()) { 2258 uint first_index = hr->hrs_index() + 1; 2259 uint last_index = hr->last_hc_index(); 2260 for (uint i = first_index; i < last_index; i += 1) { 2261 HeapRegion* chr = region_at(i); 2262 assert(chr->continuesHumongous(), "sanity"); 2263 chr->reset_gc_time_stamp(); 2264 } 2265 } 2266 } 2267 2268 #ifndef PRODUCT 2269 class CheckGCTimeStampsHRClosure : public HeapRegionClosure { 2270 private: 2271 unsigned _gc_time_stamp; 2272 bool _failures; 2273 2274 public: 2275 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) : 2276 _gc_time_stamp(gc_time_stamp), _failures(false) { } 2277 2278 virtual bool doHeapRegion(HeapRegion* hr) { 2279 unsigned region_gc_time_stamp = hr->get_gc_time_stamp(); 2280 if (_gc_time_stamp != region_gc_time_stamp) { 2281 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, " 2282 "expected %d", HR_FORMAT_PARAMS(hr), 2283 region_gc_time_stamp, _gc_time_stamp); 2284 _failures = true; 2285 } 2286 return false; 2287 } 2288 2289 bool failures() { return _failures; } 2290 }; 2291 2292 void G1CollectedHeap::check_gc_time_stamps() { 2293 CheckGCTimeStampsHRClosure cl(_gc_time_stamp); 2294 heap_region_iterate(&cl); 2295 guarantee(!cl.failures(), "all GC time stamps should have been reset"); 2296 } 2297 #endif // PRODUCT 2298 2299 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2300 DirtyCardQueue* into_cset_dcq, 2301 bool concurrent, 2302 int worker_i) { 2303 // Clean cards in the hot card cache 2304 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2305 2306 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2307 int n_completed_buffers = 0; 2308 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2309 n_completed_buffers++; 2310 } 2311 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2312 dcqs.clear_n_completed_buffers(); 2313 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2314 } 2315 2316 2317 // Computes the sum of the storage used by the various regions. 2318 2319 size_t G1CollectedHeap::used() const { 2320 assert(Heap_lock->owner() != NULL, 2321 "Should be owned on this thread's behalf."); 2322 size_t result = _summary_bytes_used; 2323 // Read only once in case it is set to NULL concurrently 2324 HeapRegion* hr = _mutator_alloc_region.get(); 2325 if (hr != NULL) 2326 result += hr->used(); 2327 return result; 2328 } 2329 2330 size_t G1CollectedHeap::used_unlocked() const { 2331 size_t result = _summary_bytes_used; 2332 return result; 2333 } 2334 2335 class SumUsedClosure: public HeapRegionClosure { 2336 size_t _used; 2337 public: 2338 SumUsedClosure() : _used(0) {} 2339 bool doHeapRegion(HeapRegion* r) { 2340 if (!r->continuesHumongous()) { 2341 _used += r->used(); 2342 } 2343 return false; 2344 } 2345 size_t result() { return _used; } 2346 }; 2347 2348 size_t G1CollectedHeap::recalculate_used() const { 2349 SumUsedClosure blk; 2350 heap_region_iterate(&blk); 2351 return blk.result(); 2352 } 2353 2354 size_t G1CollectedHeap::unsafe_max_alloc() { 2355 if (free_regions() > 0) return HeapRegion::GrainBytes; 2356 // otherwise, is there space in the current allocation region? 2357 2358 // We need to store the current allocation region in a local variable 2359 // here. The problem is that this method doesn't take any locks and 2360 // there may be other threads which overwrite the current allocation 2361 // region field. attempt_allocation(), for example, sets it to NULL 2362 // and this can happen *after* the NULL check here but before the call 2363 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2364 // to be a problem in the optimized build, since the two loads of the 2365 // current allocation region field are optimized away. 2366 HeapRegion* hr = _mutator_alloc_region.get(); 2367 if (hr == NULL) { 2368 return 0; 2369 } 2370 return hr->free(); 2371 } 2372 2373 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2374 switch (cause) { 2375 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2376 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2377 case GCCause::_g1_humongous_allocation: return true; 2378 default: return false; 2379 } 2380 } 2381 2382 #ifndef PRODUCT 2383 void G1CollectedHeap::allocate_dummy_regions() { 2384 // Let's fill up most of the region 2385 size_t word_size = HeapRegion::GrainWords - 1024; 2386 // And as a result the region we'll allocate will be humongous. 2387 guarantee(isHumongous(word_size), "sanity"); 2388 2389 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2390 // Let's use the existing mechanism for the allocation 2391 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2392 if (dummy_obj != NULL) { 2393 MemRegion mr(dummy_obj, word_size); 2394 CollectedHeap::fill_with_object(mr); 2395 } else { 2396 // If we can't allocate once, we probably cannot allocate 2397 // again. Let's get out of the loop. 2398 break; 2399 } 2400 } 2401 } 2402 #endif // !PRODUCT 2403 2404 void G1CollectedHeap::increment_old_marking_cycles_started() { 2405 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2406 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2407 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2408 _old_marking_cycles_started, _old_marking_cycles_completed)); 2409 2410 _old_marking_cycles_started++; 2411 } 2412 2413 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2414 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2415 2416 // We assume that if concurrent == true, then the caller is a 2417 // concurrent thread that was joined the Suspendible Thread 2418 // Set. If there's ever a cheap way to check this, we should add an 2419 // assert here. 2420 2421 // Given that this method is called at the end of a Full GC or of a 2422 // concurrent cycle, and those can be nested (i.e., a Full GC can 2423 // interrupt a concurrent cycle), the number of full collections 2424 // completed should be either one (in the case where there was no 2425 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2426 // behind the number of full collections started. 2427 2428 // This is the case for the inner caller, i.e. a Full GC. 2429 assert(concurrent || 2430 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2431 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2432 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2433 "is inconsistent with _old_marking_cycles_completed = %u", 2434 _old_marking_cycles_started, _old_marking_cycles_completed)); 2435 2436 // This is the case for the outer caller, i.e. the concurrent cycle. 2437 assert(!concurrent || 2438 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2439 err_msg("for outer caller (concurrent cycle): " 2440 "_old_marking_cycles_started = %u " 2441 "is inconsistent with _old_marking_cycles_completed = %u", 2442 _old_marking_cycles_started, _old_marking_cycles_completed)); 2443 2444 _old_marking_cycles_completed += 1; 2445 2446 // We need to clear the "in_progress" flag in the CM thread before 2447 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2448 // is set) so that if a waiter requests another System.gc() it doesn't 2449 // incorrectly see that a marking cyle is still in progress. 2450 if (concurrent) { 2451 _cmThread->clear_in_progress(); 2452 } 2453 2454 // This notify_all() will ensure that a thread that called 2455 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2456 // and it's waiting for a full GC to finish will be woken up. It is 2457 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2458 FullGCCount_lock->notify_all(); 2459 } 2460 2461 void G1CollectedHeap::collect(GCCause::Cause cause) { 2462 assert_heap_not_locked(); 2463 2464 unsigned int gc_count_before; 2465 unsigned int old_marking_count_before; 2466 bool retry_gc; 2467 2468 do { 2469 retry_gc = false; 2470 2471 { 2472 MutexLocker ml(Heap_lock); 2473 2474 // Read the GC count while holding the Heap_lock 2475 gc_count_before = total_collections(); 2476 old_marking_count_before = _old_marking_cycles_started; 2477 } 2478 2479 if (should_do_concurrent_full_gc(cause)) { 2480 // Schedule an initial-mark evacuation pause that will start a 2481 // concurrent cycle. We're setting word_size to 0 which means that 2482 // we are not requesting a post-GC allocation. 2483 VM_G1IncCollectionPause op(gc_count_before, 2484 0, /* word_size */ 2485 true, /* should_initiate_conc_mark */ 2486 g1_policy()->max_pause_time_ms(), 2487 cause); 2488 2489 VMThread::execute(&op); 2490 if (!op.pause_succeeded()) { 2491 if (old_marking_count_before == _old_marking_cycles_started) { 2492 retry_gc = op.should_retry_gc(); 2493 } else { 2494 // A Full GC happened while we were trying to schedule the 2495 // initial-mark GC. No point in starting a new cycle given 2496 // that the whole heap was collected anyway. 2497 } 2498 2499 if (retry_gc) { 2500 if (GC_locker::is_active_and_needs_gc()) { 2501 GC_locker::stall_until_clear(); 2502 } 2503 } 2504 } 2505 } else { 2506 if (cause == GCCause::_gc_locker 2507 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2508 2509 // Schedule a standard evacuation pause. We're setting word_size 2510 // to 0 which means that we are not requesting a post-GC allocation. 2511 VM_G1IncCollectionPause op(gc_count_before, 2512 0, /* word_size */ 2513 false, /* should_initiate_conc_mark */ 2514 g1_policy()->max_pause_time_ms(), 2515 cause); 2516 VMThread::execute(&op); 2517 } else { 2518 // Schedule a Full GC. 2519 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2520 VMThread::execute(&op); 2521 } 2522 } 2523 } while (retry_gc); 2524 } 2525 2526 bool G1CollectedHeap::is_in(const void* p) const { 2527 if (_g1_committed.contains(p)) { 2528 // Given that we know that p is in the committed space, 2529 // heap_region_containing_raw() should successfully 2530 // return the containing region. 2531 HeapRegion* hr = heap_region_containing_raw(p); 2532 return hr->is_in(p); 2533 } else { 2534 return false; 2535 } 2536 } 2537 2538 // Iteration functions. 2539 2540 // Iterates an OopClosure over all ref-containing fields of objects 2541 // within a HeapRegion. 2542 2543 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2544 MemRegion _mr; 2545 ExtendedOopClosure* _cl; 2546 public: 2547 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl) 2548 : _mr(mr), _cl(cl) {} 2549 bool doHeapRegion(HeapRegion* r) { 2550 if (!r->continuesHumongous()) { 2551 r->oop_iterate(_cl); 2552 } 2553 return false; 2554 } 2555 }; 2556 2557 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) { 2558 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2559 heap_region_iterate(&blk); 2560 } 2561 2562 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 2563 IterateOopClosureRegionClosure blk(mr, cl); 2564 heap_region_iterate(&blk); 2565 } 2566 2567 // Iterates an ObjectClosure over all objects within a HeapRegion. 2568 2569 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2570 ObjectClosure* _cl; 2571 public: 2572 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2573 bool doHeapRegion(HeapRegion* r) { 2574 if (! r->continuesHumongous()) { 2575 r->object_iterate(_cl); 2576 } 2577 return false; 2578 } 2579 }; 2580 2581 void G1CollectedHeap::object_iterate(ObjectClosure* cl) { 2582 IterateObjectClosureRegionClosure blk(cl); 2583 heap_region_iterate(&blk); 2584 } 2585 2586 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2587 // FIXME: is this right? 2588 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2589 } 2590 2591 // Calls a SpaceClosure on a HeapRegion. 2592 2593 class SpaceClosureRegionClosure: public HeapRegionClosure { 2594 SpaceClosure* _cl; 2595 public: 2596 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2597 bool doHeapRegion(HeapRegion* r) { 2598 _cl->do_space(r); 2599 return false; 2600 } 2601 }; 2602 2603 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2604 SpaceClosureRegionClosure blk(cl); 2605 heap_region_iterate(&blk); 2606 } 2607 2608 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2609 _hrs.iterate(cl); 2610 } 2611 2612 void 2613 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2614 uint worker_id, 2615 uint no_of_par_workers, 2616 jint claim_value) { 2617 const uint regions = n_regions(); 2618 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2619 no_of_par_workers : 2620 1); 2621 assert(UseDynamicNumberOfGCThreads || 2622 no_of_par_workers == workers()->total_workers(), 2623 "Non dynamic should use fixed number of workers"); 2624 // try to spread out the starting points of the workers 2625 const HeapRegion* start_hr = 2626 start_region_for_worker(worker_id, no_of_par_workers); 2627 const uint start_index = start_hr->hrs_index(); 2628 2629 // each worker will actually look at all regions 2630 for (uint count = 0; count < regions; ++count) { 2631 const uint index = (start_index + count) % regions; 2632 assert(0 <= index && index < regions, "sanity"); 2633 HeapRegion* r = region_at(index); 2634 // we'll ignore "continues humongous" regions (we'll process them 2635 // when we come across their corresponding "start humongous" 2636 // region) and regions already claimed 2637 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2638 continue; 2639 } 2640 // OK, try to claim it 2641 if (r->claimHeapRegion(claim_value)) { 2642 // success! 2643 assert(!r->continuesHumongous(), "sanity"); 2644 if (r->startsHumongous()) { 2645 // If the region is "starts humongous" we'll iterate over its 2646 // "continues humongous" first; in fact we'll do them 2647 // first. The order is important. In on case, calling the 2648 // closure on the "starts humongous" region might de-allocate 2649 // and clear all its "continues humongous" regions and, as a 2650 // result, we might end up processing them twice. So, we'll do 2651 // them first (notice: most closures will ignore them anyway) and 2652 // then we'll do the "starts humongous" region. 2653 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2654 HeapRegion* chr = region_at(ch_index); 2655 2656 // if the region has already been claimed or it's not 2657 // "continues humongous" we're done 2658 if (chr->claim_value() == claim_value || 2659 !chr->continuesHumongous()) { 2660 break; 2661 } 2662 2663 // Noone should have claimed it directly. We can given 2664 // that we claimed its "starts humongous" region. 2665 assert(chr->claim_value() != claim_value, "sanity"); 2666 assert(chr->humongous_start_region() == r, "sanity"); 2667 2668 if (chr->claimHeapRegion(claim_value)) { 2669 // we should always be able to claim it; noone else should 2670 // be trying to claim this region 2671 2672 bool res2 = cl->doHeapRegion(chr); 2673 assert(!res2, "Should not abort"); 2674 2675 // Right now, this holds (i.e., no closure that actually 2676 // does something with "continues humongous" regions 2677 // clears them). We might have to weaken it in the future, 2678 // but let's leave these two asserts here for extra safety. 2679 assert(chr->continuesHumongous(), "should still be the case"); 2680 assert(chr->humongous_start_region() == r, "sanity"); 2681 } else { 2682 guarantee(false, "we should not reach here"); 2683 } 2684 } 2685 } 2686 2687 assert(!r->continuesHumongous(), "sanity"); 2688 bool res = cl->doHeapRegion(r); 2689 assert(!res, "Should not abort"); 2690 } 2691 } 2692 } 2693 2694 class ResetClaimValuesClosure: public HeapRegionClosure { 2695 public: 2696 bool doHeapRegion(HeapRegion* r) { 2697 r->set_claim_value(HeapRegion::InitialClaimValue); 2698 return false; 2699 } 2700 }; 2701 2702 void G1CollectedHeap::reset_heap_region_claim_values() { 2703 ResetClaimValuesClosure blk; 2704 heap_region_iterate(&blk); 2705 } 2706 2707 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2708 ResetClaimValuesClosure blk; 2709 collection_set_iterate(&blk); 2710 } 2711 2712 #ifdef ASSERT 2713 // This checks whether all regions in the heap have the correct claim 2714 // value. I also piggy-backed on this a check to ensure that the 2715 // humongous_start_region() information on "continues humongous" 2716 // regions is correct. 2717 2718 class CheckClaimValuesClosure : public HeapRegionClosure { 2719 private: 2720 jint _claim_value; 2721 uint _failures; 2722 HeapRegion* _sh_region; 2723 2724 public: 2725 CheckClaimValuesClosure(jint claim_value) : 2726 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2727 bool doHeapRegion(HeapRegion* r) { 2728 if (r->claim_value() != _claim_value) { 2729 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2730 "claim value = %d, should be %d", 2731 HR_FORMAT_PARAMS(r), 2732 r->claim_value(), _claim_value); 2733 ++_failures; 2734 } 2735 if (!r->isHumongous()) { 2736 _sh_region = NULL; 2737 } else if (r->startsHumongous()) { 2738 _sh_region = r; 2739 } else if (r->continuesHumongous()) { 2740 if (r->humongous_start_region() != _sh_region) { 2741 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2742 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2743 HR_FORMAT_PARAMS(r), 2744 r->humongous_start_region(), 2745 _sh_region); 2746 ++_failures; 2747 } 2748 } 2749 return false; 2750 } 2751 uint failures() { return _failures; } 2752 }; 2753 2754 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2755 CheckClaimValuesClosure cl(claim_value); 2756 heap_region_iterate(&cl); 2757 return cl.failures() == 0; 2758 } 2759 2760 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2761 private: 2762 jint _claim_value; 2763 uint _failures; 2764 2765 public: 2766 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2767 _claim_value(claim_value), _failures(0) { } 2768 2769 uint failures() { return _failures; } 2770 2771 bool doHeapRegion(HeapRegion* hr) { 2772 assert(hr->in_collection_set(), "how?"); 2773 assert(!hr->isHumongous(), "H-region in CSet"); 2774 if (hr->claim_value() != _claim_value) { 2775 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2776 "claim value = %d, should be %d", 2777 HR_FORMAT_PARAMS(hr), 2778 hr->claim_value(), _claim_value); 2779 _failures += 1; 2780 } 2781 return false; 2782 } 2783 }; 2784 2785 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2786 CheckClaimValuesInCSetHRClosure cl(claim_value); 2787 collection_set_iterate(&cl); 2788 return cl.failures() == 0; 2789 } 2790 #endif // ASSERT 2791 2792 // Clear the cached CSet starting regions and (more importantly) 2793 // the time stamps. Called when we reset the GC time stamp. 2794 void G1CollectedHeap::clear_cset_start_regions() { 2795 assert(_worker_cset_start_region != NULL, "sanity"); 2796 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2797 2798 int n_queues = MAX2((int)ParallelGCThreads, 1); 2799 for (int i = 0; i < n_queues; i++) { 2800 _worker_cset_start_region[i] = NULL; 2801 _worker_cset_start_region_time_stamp[i] = 0; 2802 } 2803 } 2804 2805 // Given the id of a worker, obtain or calculate a suitable 2806 // starting region for iterating over the current collection set. 2807 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2808 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2809 2810 HeapRegion* result = NULL; 2811 unsigned gc_time_stamp = get_gc_time_stamp(); 2812 2813 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2814 // Cached starting region for current worker was set 2815 // during the current pause - so it's valid. 2816 // Note: the cached starting heap region may be NULL 2817 // (when the collection set is empty). 2818 result = _worker_cset_start_region[worker_i]; 2819 assert(result == NULL || result->in_collection_set(), "sanity"); 2820 return result; 2821 } 2822 2823 // The cached entry was not valid so let's calculate 2824 // a suitable starting heap region for this worker. 2825 2826 // We want the parallel threads to start their collection 2827 // set iteration at different collection set regions to 2828 // avoid contention. 2829 // If we have: 2830 // n collection set regions 2831 // p threads 2832 // Then thread t will start at region floor ((t * n) / p) 2833 2834 result = g1_policy()->collection_set(); 2835 if (G1CollectedHeap::use_parallel_gc_threads()) { 2836 uint cs_size = g1_policy()->cset_region_length(); 2837 uint active_workers = workers()->active_workers(); 2838 assert(UseDynamicNumberOfGCThreads || 2839 active_workers == workers()->total_workers(), 2840 "Unless dynamic should use total workers"); 2841 2842 uint end_ind = (cs_size * worker_i) / active_workers; 2843 uint start_ind = 0; 2844 2845 if (worker_i > 0 && 2846 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2847 // Previous workers starting region is valid 2848 // so let's iterate from there 2849 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2850 result = _worker_cset_start_region[worker_i - 1]; 2851 } 2852 2853 for (uint i = start_ind; i < end_ind; i++) { 2854 result = result->next_in_collection_set(); 2855 } 2856 } 2857 2858 // Note: the calculated starting heap region may be NULL 2859 // (when the collection set is empty). 2860 assert(result == NULL || result->in_collection_set(), "sanity"); 2861 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2862 "should be updated only once per pause"); 2863 _worker_cset_start_region[worker_i] = result; 2864 OrderAccess::storestore(); 2865 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2866 return result; 2867 } 2868 2869 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i, 2870 uint no_of_par_workers) { 2871 uint worker_num = 2872 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U; 2873 assert(UseDynamicNumberOfGCThreads || 2874 no_of_par_workers == workers()->total_workers(), 2875 "Non dynamic should use fixed number of workers"); 2876 const uint start_index = n_regions() * worker_i / worker_num; 2877 return region_at(start_index); 2878 } 2879 2880 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2881 HeapRegion* r = g1_policy()->collection_set(); 2882 while (r != NULL) { 2883 HeapRegion* next = r->next_in_collection_set(); 2884 if (cl->doHeapRegion(r)) { 2885 cl->incomplete(); 2886 return; 2887 } 2888 r = next; 2889 } 2890 } 2891 2892 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2893 HeapRegionClosure *cl) { 2894 if (r == NULL) { 2895 // The CSet is empty so there's nothing to do. 2896 return; 2897 } 2898 2899 assert(r->in_collection_set(), 2900 "Start region must be a member of the collection set."); 2901 HeapRegion* cur = r; 2902 while (cur != NULL) { 2903 HeapRegion* next = cur->next_in_collection_set(); 2904 if (cl->doHeapRegion(cur) && false) { 2905 cl->incomplete(); 2906 return; 2907 } 2908 cur = next; 2909 } 2910 cur = g1_policy()->collection_set(); 2911 while (cur != r) { 2912 HeapRegion* next = cur->next_in_collection_set(); 2913 if (cl->doHeapRegion(cur) && false) { 2914 cl->incomplete(); 2915 return; 2916 } 2917 cur = next; 2918 } 2919 } 2920 2921 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2922 return n_regions() > 0 ? region_at(0) : NULL; 2923 } 2924 2925 2926 Space* G1CollectedHeap::space_containing(const void* addr) const { 2927 Space* res = heap_region_containing(addr); 2928 return res; 2929 } 2930 2931 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2932 Space* sp = space_containing(addr); 2933 if (sp != NULL) { 2934 return sp->block_start(addr); 2935 } 2936 return NULL; 2937 } 2938 2939 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2940 Space* sp = space_containing(addr); 2941 assert(sp != NULL, "block_size of address outside of heap"); 2942 return sp->block_size(addr); 2943 } 2944 2945 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2946 Space* sp = space_containing(addr); 2947 return sp->block_is_obj(addr); 2948 } 2949 2950 bool G1CollectedHeap::supports_tlab_allocation() const { 2951 return true; 2952 } 2953 2954 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2955 return HeapRegion::GrainBytes; 2956 } 2957 2958 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2959 // Return the remaining space in the cur alloc region, but not less than 2960 // the min TLAB size. 2961 2962 // Also, this value can be at most the humongous object threshold, 2963 // since we can't allow tlabs to grow big enough to accomodate 2964 // humongous objects. 2965 2966 HeapRegion* hr = _mutator_alloc_region.get(); 2967 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2968 if (hr == NULL) { 2969 return max_tlab_size; 2970 } else { 2971 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2972 } 2973 } 2974 2975 size_t G1CollectedHeap::max_capacity() const { 2976 return _g1_reserved.byte_size(); 2977 } 2978 2979 jlong G1CollectedHeap::millis_since_last_gc() { 2980 // assert(false, "NYI"); 2981 return 0; 2982 } 2983 2984 void G1CollectedHeap::prepare_for_verify() { 2985 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2986 ensure_parsability(false); 2987 } 2988 g1_rem_set()->prepare_for_verify(); 2989 } 2990 2991 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr, 2992 VerifyOption vo) { 2993 switch (vo) { 2994 case VerifyOption_G1UsePrevMarking: 2995 return hr->obj_allocated_since_prev_marking(obj); 2996 case VerifyOption_G1UseNextMarking: 2997 return hr->obj_allocated_since_next_marking(obj); 2998 case VerifyOption_G1UseMarkWord: 2999 return false; 3000 default: 3001 ShouldNotReachHere(); 3002 } 3003 return false; // keep some compilers happy 3004 } 3005 3006 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) { 3007 switch (vo) { 3008 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start(); 3009 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start(); 3010 case VerifyOption_G1UseMarkWord: return NULL; 3011 default: ShouldNotReachHere(); 3012 } 3013 return NULL; // keep some compilers happy 3014 } 3015 3016 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) { 3017 switch (vo) { 3018 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj); 3019 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj); 3020 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked(); 3021 default: ShouldNotReachHere(); 3022 } 3023 return false; // keep some compilers happy 3024 } 3025 3026 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) { 3027 switch (vo) { 3028 case VerifyOption_G1UsePrevMarking: return "PTAMS"; 3029 case VerifyOption_G1UseNextMarking: return "NTAMS"; 3030 case VerifyOption_G1UseMarkWord: return "NONE"; 3031 default: ShouldNotReachHere(); 3032 } 3033 return NULL; // keep some compilers happy 3034 } 3035 3036 class VerifyLivenessOopClosure: public OopClosure { 3037 G1CollectedHeap* _g1h; 3038 VerifyOption _vo; 3039 public: 3040 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 3041 _g1h(g1h), _vo(vo) 3042 { } 3043 void do_oop(narrowOop *p) { do_oop_work(p); } 3044 void do_oop( oop *p) { do_oop_work(p); } 3045 3046 template <class T> void do_oop_work(T *p) { 3047 oop obj = oopDesc::load_decode_heap_oop(p); 3048 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 3049 "Dead object referenced by a not dead object"); 3050 } 3051 }; 3052 3053 class VerifyObjsInRegionClosure: public ObjectClosure { 3054 private: 3055 G1CollectedHeap* _g1h; 3056 size_t _live_bytes; 3057 HeapRegion *_hr; 3058 VerifyOption _vo; 3059 public: 3060 // _vo == UsePrevMarking -> use "prev" marking information, 3061 // _vo == UseNextMarking -> use "next" marking information, 3062 // _vo == UseMarkWord -> use mark word from object header. 3063 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3064 : _live_bytes(0), _hr(hr), _vo(vo) { 3065 _g1h = G1CollectedHeap::heap(); 3066 } 3067 void do_object(oop o) { 3068 VerifyLivenessOopClosure isLive(_g1h, _vo); 3069 assert(o != NULL, "Huh?"); 3070 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3071 // If the object is alive according to the mark word, 3072 // then verify that the marking information agrees. 3073 // Note we can't verify the contra-positive of the 3074 // above: if the object is dead (according to the mark 3075 // word), it may not be marked, or may have been marked 3076 // but has since became dead, or may have been allocated 3077 // since the last marking. 3078 if (_vo == VerifyOption_G1UseMarkWord) { 3079 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3080 } 3081 3082 o->oop_iterate_no_header(&isLive); 3083 if (!_hr->obj_allocated_since_prev_marking(o)) { 3084 size_t obj_size = o->size(); // Make sure we don't overflow 3085 _live_bytes += (obj_size * HeapWordSize); 3086 } 3087 } 3088 } 3089 size_t live_bytes() { return _live_bytes; } 3090 }; 3091 3092 class PrintObjsInRegionClosure : public ObjectClosure { 3093 HeapRegion *_hr; 3094 G1CollectedHeap *_g1; 3095 public: 3096 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3097 _g1 = G1CollectedHeap::heap(); 3098 }; 3099 3100 void do_object(oop o) { 3101 if (o != NULL) { 3102 HeapWord *start = (HeapWord *) o; 3103 size_t word_sz = o->size(); 3104 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3105 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3106 (void*) o, word_sz, 3107 _g1->isMarkedPrev(o), 3108 _g1->isMarkedNext(o), 3109 _hr->obj_allocated_since_prev_marking(o)); 3110 HeapWord *end = start + word_sz; 3111 HeapWord *cur; 3112 int *val; 3113 for (cur = start; cur < end; cur++) { 3114 val = (int *) cur; 3115 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3116 } 3117 } 3118 } 3119 }; 3120 3121 class VerifyRegionClosure: public HeapRegionClosure { 3122 private: 3123 bool _par; 3124 VerifyOption _vo; 3125 bool _failures; 3126 public: 3127 // _vo == UsePrevMarking -> use "prev" marking information, 3128 // _vo == UseNextMarking -> use "next" marking information, 3129 // _vo == UseMarkWord -> use mark word from object header. 3130 VerifyRegionClosure(bool par, VerifyOption vo) 3131 : _par(par), 3132 _vo(vo), 3133 _failures(false) {} 3134 3135 bool failures() { 3136 return _failures; 3137 } 3138 3139 bool doHeapRegion(HeapRegion* r) { 3140 if (!r->continuesHumongous()) { 3141 bool failures = false; 3142 r->verify(_vo, &failures); 3143 if (failures) { 3144 _failures = true; 3145 } else { 3146 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3147 r->object_iterate(¬_dead_yet_cl); 3148 if (_vo != VerifyOption_G1UseNextMarking) { 3149 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3150 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3151 "max_live_bytes "SIZE_FORMAT" " 3152 "< calculated "SIZE_FORMAT, 3153 r->bottom(), r->end(), 3154 r->max_live_bytes(), 3155 not_dead_yet_cl.live_bytes()); 3156 _failures = true; 3157 } 3158 } else { 3159 // When vo == UseNextMarking we cannot currently do a sanity 3160 // check on the live bytes as the calculation has not been 3161 // finalized yet. 3162 } 3163 } 3164 } 3165 return false; // stop the region iteration if we hit a failure 3166 } 3167 }; 3168 3169 class YoungRefCounterClosure : public OopClosure { 3170 G1CollectedHeap* _g1h; 3171 int _count; 3172 public: 3173 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {} 3174 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } } 3175 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3176 3177 int count() { return _count; } 3178 void reset_count() { _count = 0; }; 3179 }; 3180 3181 class VerifyKlassClosure: public KlassClosure { 3182 YoungRefCounterClosure _young_ref_counter_closure; 3183 OopClosure *_oop_closure; 3184 public: 3185 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {} 3186 void do_klass(Klass* k) { 3187 k->oops_do(_oop_closure); 3188 3189 _young_ref_counter_closure.reset_count(); 3190 k->oops_do(&_young_ref_counter_closure); 3191 if (_young_ref_counter_closure.count() > 0) { 3192 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k)); 3193 } 3194 } 3195 }; 3196 3197 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can 3198 // pass it as the perm_blk to SharedHeap::process_strong_roots. 3199 // When process_strong_roots stop calling perm_blk->younger_refs_iterate 3200 // we can change this closure to extend the simpler OopClosure. 3201 class VerifyRootsClosure: public OopsInGenClosure { 3202 private: 3203 G1CollectedHeap* _g1h; 3204 VerifyOption _vo; 3205 bool _failures; 3206 public: 3207 // _vo == UsePrevMarking -> use "prev" marking information, 3208 // _vo == UseNextMarking -> use "next" marking information, 3209 // _vo == UseMarkWord -> use mark word from object header. 3210 VerifyRootsClosure(VerifyOption vo) : 3211 _g1h(G1CollectedHeap::heap()), 3212 _vo(vo), 3213 _failures(false) { } 3214 3215 bool failures() { return _failures; } 3216 3217 template <class T> void do_oop_nv(T* p) { 3218 T heap_oop = oopDesc::load_heap_oop(p); 3219 if (!oopDesc::is_null(heap_oop)) { 3220 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3221 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3222 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3223 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3224 if (_vo == VerifyOption_G1UseMarkWord) { 3225 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3226 } 3227 obj->print_on(gclog_or_tty); 3228 _failures = true; 3229 } 3230 } 3231 } 3232 3233 void do_oop(oop* p) { do_oop_nv(p); } 3234 void do_oop(narrowOop* p) { do_oop_nv(p); } 3235 }; 3236 3237 // This is the task used for parallel heap verification. 3238 3239 class G1ParVerifyTask: public AbstractGangTask { 3240 private: 3241 G1CollectedHeap* _g1h; 3242 VerifyOption _vo; 3243 bool _failures; 3244 3245 public: 3246 // _vo == UsePrevMarking -> use "prev" marking information, 3247 // _vo == UseNextMarking -> use "next" marking information, 3248 // _vo == UseMarkWord -> use mark word from object header. 3249 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3250 AbstractGangTask("Parallel verify task"), 3251 _g1h(g1h), 3252 _vo(vo), 3253 _failures(false) { } 3254 3255 bool failures() { 3256 return _failures; 3257 } 3258 3259 void work(uint worker_id) { 3260 HandleMark hm; 3261 VerifyRegionClosure blk(true, _vo); 3262 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3263 _g1h->workers()->active_workers(), 3264 HeapRegion::ParVerifyClaimValue); 3265 if (blk.failures()) { 3266 _failures = true; 3267 } 3268 } 3269 }; 3270 3271 void G1CollectedHeap::verify(bool silent) { 3272 verify(silent, VerifyOption_G1UsePrevMarking); 3273 } 3274 3275 void G1CollectedHeap::verify(bool silent, 3276 VerifyOption vo) { 3277 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3278 if (!silent) { gclog_or_tty->print("Roots "); } 3279 VerifyRootsClosure rootsCl(vo); 3280 3281 assert(Thread::current()->is_VM_thread(), 3282 "Expected to be executed serially by the VM thread at this point"); 3283 3284 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3285 VerifyKlassClosure klassCl(this, &rootsCl); 3286 3287 // We apply the relevant closures to all the oops in the 3288 // system dictionary, the string table and the code cache. 3289 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3290 3291 // Need cleared claim bits for the strong roots processing 3292 ClassLoaderDataGraph::clear_claimed_marks(); 3293 3294 process_strong_roots(true, // activate StrongRootsScope 3295 false, // we set "is scavenging" to false, 3296 // so we don't reset the dirty cards. 3297 ScanningOption(so), // roots scanning options 3298 &rootsCl, 3299 &blobsCl, 3300 &klassCl 3301 ); 3302 3303 bool failures = rootsCl.failures(); 3304 3305 if (vo != VerifyOption_G1UseMarkWord) { 3306 // If we're verifying during a full GC then the region sets 3307 // will have been torn down at the start of the GC. Therefore 3308 // verifying the region sets will fail. So we only verify 3309 // the region sets when not in a full GC. 3310 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3311 verify_region_sets(); 3312 } 3313 3314 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3315 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3316 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3317 "sanity check"); 3318 3319 G1ParVerifyTask task(this, vo); 3320 assert(UseDynamicNumberOfGCThreads || 3321 workers()->active_workers() == workers()->total_workers(), 3322 "If not dynamic should be using all the workers"); 3323 int n_workers = workers()->active_workers(); 3324 set_par_threads(n_workers); 3325 workers()->run_task(&task); 3326 set_par_threads(0); 3327 if (task.failures()) { 3328 failures = true; 3329 } 3330 3331 // Checks that the expected amount of parallel work was done. 3332 // The implication is that n_workers is > 0. 3333 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3334 "sanity check"); 3335 3336 reset_heap_region_claim_values(); 3337 3338 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3339 "sanity check"); 3340 } else { 3341 VerifyRegionClosure blk(false, vo); 3342 heap_region_iterate(&blk); 3343 if (blk.failures()) { 3344 failures = true; 3345 } 3346 } 3347 if (!silent) gclog_or_tty->print("RemSet "); 3348 rem_set()->verify(); 3349 3350 if (failures) { 3351 gclog_or_tty->print_cr("Heap:"); 3352 // It helps to have the per-region information in the output to 3353 // help us track down what went wrong. This is why we call 3354 // print_extended_on() instead of print_on(). 3355 print_extended_on(gclog_or_tty); 3356 gclog_or_tty->print_cr(""); 3357 #ifndef PRODUCT 3358 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3359 concurrent_mark()->print_reachable("at-verification-failure", 3360 vo, false /* all */); 3361 } 3362 #endif 3363 gclog_or_tty->flush(); 3364 } 3365 guarantee(!failures, "there should not have been any failures"); 3366 } else { 3367 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3368 } 3369 } 3370 3371 class PrintRegionClosure: public HeapRegionClosure { 3372 outputStream* _st; 3373 public: 3374 PrintRegionClosure(outputStream* st) : _st(st) {} 3375 bool doHeapRegion(HeapRegion* r) { 3376 r->print_on(_st); 3377 return false; 3378 } 3379 }; 3380 3381 void G1CollectedHeap::print_on(outputStream* st) const { 3382 st->print(" %-20s", "garbage-first heap"); 3383 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3384 capacity()/K, used_unlocked()/K); 3385 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3386 _g1_storage.low_boundary(), 3387 _g1_storage.high(), 3388 _g1_storage.high_boundary()); 3389 st->cr(); 3390 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3391 uint young_regions = _young_list->length(); 3392 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3393 (size_t) young_regions * HeapRegion::GrainBytes / K); 3394 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3395 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3396 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3397 st->cr(); 3398 MetaspaceAux::print_on(st); 3399 } 3400 3401 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3402 print_on(st); 3403 3404 // Print the per-region information. 3405 st->cr(); 3406 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3407 "HS=humongous(starts), HC=humongous(continues), " 3408 "CS=collection set, F=free, TS=gc time stamp, " 3409 "PTAMS=previous top-at-mark-start, " 3410 "NTAMS=next top-at-mark-start)"); 3411 PrintRegionClosure blk(st); 3412 heap_region_iterate(&blk); 3413 } 3414 3415 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3416 if (G1CollectedHeap::use_parallel_gc_threads()) { 3417 workers()->print_worker_threads_on(st); 3418 } 3419 _cmThread->print_on(st); 3420 st->cr(); 3421 _cm->print_worker_threads_on(st); 3422 _cg1r->print_worker_threads_on(st); 3423 } 3424 3425 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3426 if (G1CollectedHeap::use_parallel_gc_threads()) { 3427 workers()->threads_do(tc); 3428 } 3429 tc->do_thread(_cmThread); 3430 _cg1r->threads_do(tc); 3431 } 3432 3433 void G1CollectedHeap::print_tracing_info() const { 3434 // We'll overload this to mean "trace GC pause statistics." 3435 if (TraceGen0Time || TraceGen1Time) { 3436 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3437 // to that. 3438 g1_policy()->print_tracing_info(); 3439 } 3440 if (G1SummarizeRSetStats) { 3441 g1_rem_set()->print_summary_info(); 3442 } 3443 if (G1SummarizeConcMark) { 3444 concurrent_mark()->print_summary_info(); 3445 } 3446 g1_policy()->print_yg_surv_rate_info(); 3447 SpecializationStats::print(); 3448 } 3449 3450 #ifndef PRODUCT 3451 // Helpful for debugging RSet issues. 3452 3453 class PrintRSetsClosure : public HeapRegionClosure { 3454 private: 3455 const char* _msg; 3456 size_t _occupied_sum; 3457 3458 public: 3459 bool doHeapRegion(HeapRegion* r) { 3460 HeapRegionRemSet* hrrs = r->rem_set(); 3461 size_t occupied = hrrs->occupied(); 3462 _occupied_sum += occupied; 3463 3464 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3465 HR_FORMAT_PARAMS(r)); 3466 if (occupied == 0) { 3467 gclog_or_tty->print_cr(" RSet is empty"); 3468 } else { 3469 hrrs->print(); 3470 } 3471 gclog_or_tty->print_cr("----------"); 3472 return false; 3473 } 3474 3475 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3476 gclog_or_tty->cr(); 3477 gclog_or_tty->print_cr("========================================"); 3478 gclog_or_tty->print_cr(msg); 3479 gclog_or_tty->cr(); 3480 } 3481 3482 ~PrintRSetsClosure() { 3483 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3484 gclog_or_tty->print_cr("========================================"); 3485 gclog_or_tty->cr(); 3486 } 3487 }; 3488 3489 void G1CollectedHeap::print_cset_rsets() { 3490 PrintRSetsClosure cl("Printing CSet RSets"); 3491 collection_set_iterate(&cl); 3492 } 3493 3494 void G1CollectedHeap::print_all_rsets() { 3495 PrintRSetsClosure cl("Printing All RSets");; 3496 heap_region_iterate(&cl); 3497 } 3498 #endif // PRODUCT 3499 3500 G1CollectedHeap* G1CollectedHeap::heap() { 3501 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3502 "not a garbage-first heap"); 3503 return _g1h; 3504 } 3505 3506 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3507 // always_do_update_barrier = false; 3508 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3509 // Call allocation profiler 3510 AllocationProfiler::iterate_since_last_gc(); 3511 // Fill TLAB's and such 3512 ensure_parsability(true); 3513 } 3514 3515 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3516 // FIXME: what is this about? 3517 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3518 // is set. 3519 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3520 "derived pointer present")); 3521 // always_do_update_barrier = true; 3522 3523 // We have just completed a GC. Update the soft reference 3524 // policy with the new heap occupancy 3525 Universe::update_heap_info_at_gc(); 3526 } 3527 3528 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3529 unsigned int gc_count_before, 3530 bool* succeeded) { 3531 assert_heap_not_locked_and_not_at_safepoint(); 3532 g1_policy()->record_stop_world_start(); 3533 VM_G1IncCollectionPause op(gc_count_before, 3534 word_size, 3535 false, /* should_initiate_conc_mark */ 3536 g1_policy()->max_pause_time_ms(), 3537 GCCause::_g1_inc_collection_pause); 3538 VMThread::execute(&op); 3539 3540 HeapWord* result = op.result(); 3541 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3542 assert(result == NULL || ret_succeeded, 3543 "the result should be NULL if the VM did not succeed"); 3544 *succeeded = ret_succeeded; 3545 3546 assert_heap_not_locked(); 3547 return result; 3548 } 3549 3550 void 3551 G1CollectedHeap::doConcurrentMark() { 3552 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3553 if (!_cmThread->in_progress()) { 3554 _cmThread->set_started(); 3555 CGC_lock->notify(); 3556 } 3557 } 3558 3559 size_t G1CollectedHeap::pending_card_num() { 3560 size_t extra_cards = 0; 3561 JavaThread *curr = Threads::first(); 3562 while (curr != NULL) { 3563 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3564 extra_cards += dcq.size(); 3565 curr = curr->next(); 3566 } 3567 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3568 size_t buffer_size = dcqs.buffer_size(); 3569 size_t buffer_num = dcqs.completed_buffers_num(); 3570 3571 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes 3572 // in bytes - not the number of 'entries'. We need to convert 3573 // into a number of cards. 3574 return (buffer_size * buffer_num + extra_cards) / oopSize; 3575 } 3576 3577 size_t G1CollectedHeap::cards_scanned() { 3578 return g1_rem_set()->cardsScanned(); 3579 } 3580 3581 void 3582 G1CollectedHeap::setup_surviving_young_words() { 3583 assert(_surviving_young_words == NULL, "pre-condition"); 3584 uint array_length = g1_policy()->young_cset_region_length(); 3585 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3586 if (_surviving_young_words == NULL) { 3587 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3588 "Not enough space for young surv words summary."); 3589 } 3590 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3591 #ifdef ASSERT 3592 for (uint i = 0; i < array_length; ++i) { 3593 assert( _surviving_young_words[i] == 0, "memset above" ); 3594 } 3595 #endif // !ASSERT 3596 } 3597 3598 void 3599 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3600 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3601 uint array_length = g1_policy()->young_cset_region_length(); 3602 for (uint i = 0; i < array_length; ++i) { 3603 _surviving_young_words[i] += surv_young_words[i]; 3604 } 3605 } 3606 3607 void 3608 G1CollectedHeap::cleanup_surviving_young_words() { 3609 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3610 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3611 _surviving_young_words = NULL; 3612 } 3613 3614 #ifdef ASSERT 3615 class VerifyCSetClosure: public HeapRegionClosure { 3616 public: 3617 bool doHeapRegion(HeapRegion* hr) { 3618 // Here we check that the CSet region's RSet is ready for parallel 3619 // iteration. The fields that we'll verify are only manipulated 3620 // when the region is part of a CSet and is collected. Afterwards, 3621 // we reset these fields when we clear the region's RSet (when the 3622 // region is freed) so they are ready when the region is 3623 // re-allocated. The only exception to this is if there's an 3624 // evacuation failure and instead of freeing the region we leave 3625 // it in the heap. In that case, we reset these fields during 3626 // evacuation failure handling. 3627 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3628 3629 // Here's a good place to add any other checks we'd like to 3630 // perform on CSet regions. 3631 return false; 3632 } 3633 }; 3634 #endif // ASSERT 3635 3636 #if TASKQUEUE_STATS 3637 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3638 st->print_raw_cr("GC Task Stats"); 3639 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3640 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3641 } 3642 3643 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3644 print_taskqueue_stats_hdr(st); 3645 3646 TaskQueueStats totals; 3647 const int n = workers() != NULL ? workers()->total_workers() : 1; 3648 for (int i = 0; i < n; ++i) { 3649 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3650 totals += task_queue(i)->stats; 3651 } 3652 st->print_raw("tot "); totals.print(st); st->cr(); 3653 3654 DEBUG_ONLY(totals.verify()); 3655 } 3656 3657 void G1CollectedHeap::reset_taskqueue_stats() { 3658 const int n = workers() != NULL ? workers()->total_workers() : 1; 3659 for (int i = 0; i < n; ++i) { 3660 task_queue(i)->stats.reset(); 3661 } 3662 } 3663 #endif // TASKQUEUE_STATS 3664 3665 void G1CollectedHeap::log_gc_header() { 3666 if (!G1Log::fine()) { 3667 return; 3668 } 3669 3670 gclog_or_tty->date_stamp(PrintGCDateStamps); 3671 gclog_or_tty->stamp(PrintGCTimeStamps); 3672 3673 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause()) 3674 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)") 3675 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : ""); 3676 3677 gclog_or_tty->print("[%s", (const char*)gc_cause_str); 3678 } 3679 3680 void G1CollectedHeap::log_gc_footer(double pause_time_sec) { 3681 if (!G1Log::fine()) { 3682 return; 3683 } 3684 3685 if (G1Log::finer()) { 3686 if (evacuation_failed()) { 3687 gclog_or_tty->print(" (to-space exhausted)"); 3688 } 3689 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3690 g1_policy()->phase_times()->note_gc_end(); 3691 g1_policy()->phase_times()->print(pause_time_sec); 3692 g1_policy()->print_detailed_heap_transition(); 3693 } else { 3694 if (evacuation_failed()) { 3695 gclog_or_tty->print("--"); 3696 } 3697 g1_policy()->print_heap_transition(); 3698 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec); 3699 } 3700 gclog_or_tty->flush(); 3701 } 3702 3703 bool 3704 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3705 assert_at_safepoint(true /* should_be_vm_thread */); 3706 guarantee(!is_gc_active(), "collection is not reentrant"); 3707 3708 if (GC_locker::check_active_before_gc()) { 3709 return false; 3710 } 3711 3712 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3713 ResourceMark rm; 3714 3715 print_heap_before_gc(); 3716 3717 HRSPhaseSetter x(HRSPhaseEvacuation); 3718 verify_region_sets_optional(); 3719 verify_dirty_young_regions(); 3720 3721 // This call will decide whether this pause is an initial-mark 3722 // pause. If it is, during_initial_mark_pause() will return true 3723 // for the duration of this pause. 3724 g1_policy()->decide_on_conc_mark_initiation(); 3725 3726 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3727 assert(!g1_policy()->during_initial_mark_pause() || 3728 g1_policy()->gcs_are_young(), "sanity"); 3729 3730 // We also do not allow mixed GCs during marking. 3731 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3732 3733 // Record whether this pause is an initial mark. When the current 3734 // thread has completed its logging output and it's safe to signal 3735 // the CM thread, the flag's value in the policy has been reset. 3736 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3737 3738 // Inner scope for scope based logging, timers, and stats collection 3739 { 3740 if (g1_policy()->during_initial_mark_pause()) { 3741 // We are about to start a marking cycle, so we increment the 3742 // full collection counter. 3743 increment_old_marking_cycles_started(); 3744 } 3745 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3746 3747 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3748 workers()->active_workers() : 1); 3749 double pause_start_sec = os::elapsedTime(); 3750 g1_policy()->phase_times()->note_gc_start(active_workers); 3751 log_gc_header(); 3752 3753 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3754 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3755 3756 // If the secondary_free_list is not empty, append it to the 3757 // free_list. No need to wait for the cleanup operation to finish; 3758 // the region allocation code will check the secondary_free_list 3759 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3760 // set, skip this step so that the region allocation code has to 3761 // get entries from the secondary_free_list. 3762 if (!G1StressConcRegionFreeing) { 3763 append_secondary_free_list_if_not_empty_with_lock(); 3764 } 3765 3766 assert(check_young_list_well_formed(), 3767 "young list should be well formed"); 3768 3769 // Don't dynamically change the number of GC threads this early. A value of 3770 // 0 is used to indicate serial work. When parallel work is done, 3771 // it will be set. 3772 3773 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3774 IsGCActiveMark x; 3775 3776 gc_prologue(false); 3777 increment_total_collections(false /* full gc */); 3778 increment_gc_time_stamp(); 3779 3780 verify_before_gc(); 3781 3782 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3783 3784 // Please see comment in g1CollectedHeap.hpp and 3785 // G1CollectedHeap::ref_processing_init() to see how 3786 // reference processing currently works in G1. 3787 3788 // Enable discovery in the STW reference processor 3789 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3790 true /*verify_no_refs*/); 3791 3792 { 3793 // We want to temporarily turn off discovery by the 3794 // CM ref processor, if necessary, and turn it back on 3795 // on again later if we do. Using a scoped 3796 // NoRefDiscovery object will do this. 3797 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3798 3799 // Forget the current alloc region (we might even choose it to be part 3800 // of the collection set!). 3801 release_mutator_alloc_region(); 3802 3803 // We should call this after we retire the mutator alloc 3804 // region(s) so that all the ALLOC / RETIRE events are generated 3805 // before the start GC event. 3806 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3807 3808 // This timing is only used by the ergonomics to handle our pause target. 3809 // It is unclear why this should not include the full pause. We will 3810 // investigate this in CR 7178365. 3811 // 3812 // Preserving the old comment here if that helps the investigation: 3813 // 3814 // The elapsed time induced by the start time below deliberately elides 3815 // the possible verification above. 3816 double sample_start_time_sec = os::elapsedTime(); 3817 size_t start_used_bytes = used(); 3818 3819 #if YOUNG_LIST_VERBOSE 3820 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3821 _young_list->print(); 3822 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3823 #endif // YOUNG_LIST_VERBOSE 3824 3825 g1_policy()->record_collection_pause_start(sample_start_time_sec, 3826 start_used_bytes); 3827 3828 double scan_wait_start = os::elapsedTime(); 3829 // We have to wait until the CM threads finish scanning the 3830 // root regions as it's the only way to ensure that all the 3831 // objects on them have been correctly scanned before we start 3832 // moving them during the GC. 3833 bool waited = _cm->root_regions()->wait_until_scan_finished(); 3834 double wait_time_ms = 0.0; 3835 if (waited) { 3836 double scan_wait_end = os::elapsedTime(); 3837 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 3838 } 3839 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 3840 3841 #if YOUNG_LIST_VERBOSE 3842 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3843 _young_list->print(); 3844 #endif // YOUNG_LIST_VERBOSE 3845 3846 if (g1_policy()->during_initial_mark_pause()) { 3847 concurrent_mark()->checkpointRootsInitialPre(); 3848 } 3849 3850 #if YOUNG_LIST_VERBOSE 3851 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3852 _young_list->print(); 3853 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3854 #endif // YOUNG_LIST_VERBOSE 3855 3856 g1_policy()->finalize_cset(target_pause_time_ms); 3857 3858 _cm->note_start_of_gc(); 3859 // We should not verify the per-thread SATB buffers given that 3860 // we have not filtered them yet (we'll do so during the 3861 // GC). We also call this after finalize_cset() to 3862 // ensure that the CSet has been finalized. 3863 _cm->verify_no_cset_oops(true /* verify_stacks */, 3864 true /* verify_enqueued_buffers */, 3865 false /* verify_thread_buffers */, 3866 true /* verify_fingers */); 3867 3868 if (_hr_printer.is_active()) { 3869 HeapRegion* hr = g1_policy()->collection_set(); 3870 while (hr != NULL) { 3871 G1HRPrinter::RegionType type; 3872 if (!hr->is_young()) { 3873 type = G1HRPrinter::Old; 3874 } else if (hr->is_survivor()) { 3875 type = G1HRPrinter::Survivor; 3876 } else { 3877 type = G1HRPrinter::Eden; 3878 } 3879 _hr_printer.cset(hr); 3880 hr = hr->next_in_collection_set(); 3881 } 3882 } 3883 3884 #ifdef ASSERT 3885 VerifyCSetClosure cl; 3886 collection_set_iterate(&cl); 3887 #endif // ASSERT 3888 3889 setup_surviving_young_words(); 3890 3891 // Initialize the GC alloc regions. 3892 init_gc_alloc_regions(); 3893 3894 // Actually do the work... 3895 evacuate_collection_set(); 3896 3897 // We do this to mainly verify the per-thread SATB buffers 3898 // (which have been filtered by now) since we didn't verify 3899 // them earlier. No point in re-checking the stacks / enqueued 3900 // buffers given that the CSet has not changed since last time 3901 // we checked. 3902 _cm->verify_no_cset_oops(false /* verify_stacks */, 3903 false /* verify_enqueued_buffers */, 3904 true /* verify_thread_buffers */, 3905 true /* verify_fingers */); 3906 3907 free_collection_set(g1_policy()->collection_set()); 3908 g1_policy()->clear_collection_set(); 3909 3910 cleanup_surviving_young_words(); 3911 3912 // Start a new incremental collection set for the next pause. 3913 g1_policy()->start_incremental_cset_building(); 3914 3915 // Clear the _cset_fast_test bitmap in anticipation of adding 3916 // regions to the incremental collection set for the next 3917 // evacuation pause. 3918 clear_cset_fast_test(); 3919 3920 _young_list->reset_sampled_info(); 3921 3922 // Don't check the whole heap at this point as the 3923 // GC alloc regions from this pause have been tagged 3924 // as survivors and moved on to the survivor list. 3925 // Survivor regions will fail the !is_young() check. 3926 assert(check_young_list_empty(false /* check_heap */), 3927 "young list should be empty"); 3928 3929 #if YOUNG_LIST_VERBOSE 3930 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3931 _young_list->print(); 3932 #endif // YOUNG_LIST_VERBOSE 3933 3934 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3935 _young_list->first_survivor_region(), 3936 _young_list->last_survivor_region()); 3937 3938 _young_list->reset_auxilary_lists(); 3939 3940 if (evacuation_failed()) { 3941 _summary_bytes_used = recalculate_used(); 3942 } else { 3943 // The "used" of the the collection set have already been subtracted 3944 // when they were freed. Add in the bytes evacuated. 3945 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3946 } 3947 3948 if (g1_policy()->during_initial_mark_pause()) { 3949 // We have to do this before we notify the CM threads that 3950 // they can start working to make sure that all the 3951 // appropriate initialization is done on the CM object. 3952 concurrent_mark()->checkpointRootsInitialPost(); 3953 set_marking_started(); 3954 // Note that we don't actually trigger the CM thread at 3955 // this point. We do that later when we're sure that 3956 // the current thread has completed its logging output. 3957 } 3958 3959 allocate_dummy_regions(); 3960 3961 #if YOUNG_LIST_VERBOSE 3962 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3963 _young_list->print(); 3964 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3965 #endif // YOUNG_LIST_VERBOSE 3966 3967 init_mutator_alloc_region(); 3968 3969 { 3970 size_t expand_bytes = g1_policy()->expansion_amount(); 3971 if (expand_bytes > 0) { 3972 size_t bytes_before = capacity(); 3973 // No need for an ergo verbose message here, 3974 // expansion_amount() does this when it returns a value > 0. 3975 if (!expand(expand_bytes)) { 3976 // We failed to expand the heap so let's verify that 3977 // committed/uncommitted amount match the backing store 3978 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3979 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3980 } 3981 } 3982 } 3983 3984 // We redo the verificaiton but now wrt to the new CSet which 3985 // has just got initialized after the previous CSet was freed. 3986 _cm->verify_no_cset_oops(true /* verify_stacks */, 3987 true /* verify_enqueued_buffers */, 3988 true /* verify_thread_buffers */, 3989 true /* verify_fingers */); 3990 _cm->note_end_of_gc(); 3991 3992 // This timing is only used by the ergonomics to handle our pause target. 3993 // It is unclear why this should not include the full pause. We will 3994 // investigate this in CR 7178365. 3995 double sample_end_time_sec = os::elapsedTime(); 3996 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 3997 g1_policy()->record_collection_pause_end(pause_time_ms); 3998 3999 MemoryService::track_memory_usage(); 4000 4001 // In prepare_for_verify() below we'll need to scan the deferred 4002 // update buffers to bring the RSets up-to-date if 4003 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 4004 // the update buffers we'll probably need to scan cards on the 4005 // regions we just allocated to (i.e., the GC alloc 4006 // regions). However, during the last GC we called 4007 // set_saved_mark() on all the GC alloc regions, so card 4008 // scanning might skip the [saved_mark_word()...top()] area of 4009 // those regions (i.e., the area we allocated objects into 4010 // during the last GC). But it shouldn't. Given that 4011 // saved_mark_word() is conditional on whether the GC time stamp 4012 // on the region is current or not, by incrementing the GC time 4013 // stamp here we invalidate all the GC time stamps on all the 4014 // regions and saved_mark_word() will simply return top() for 4015 // all the regions. This is a nicer way of ensuring this rather 4016 // than iterating over the regions and fixing them. In fact, the 4017 // GC time stamp increment here also ensures that 4018 // saved_mark_word() will return top() between pauses, i.e., 4019 // during concurrent refinement. So we don't need the 4020 // is_gc_active() check to decided which top to use when 4021 // scanning cards (see CR 7039627). 4022 increment_gc_time_stamp(); 4023 4024 verify_after_gc(); 4025 4026 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 4027 ref_processor_stw()->verify_no_references_recorded(); 4028 4029 // CM reference discovery will be re-enabled if necessary. 4030 } 4031 4032 // We should do this after we potentially expand the heap so 4033 // that all the COMMIT events are generated before the end GC 4034 // event, and after we retire the GC alloc regions so that all 4035 // RETIRE events are generated before the end GC event. 4036 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 4037 4038 if (mark_in_progress()) { 4039 concurrent_mark()->update_g1_committed(); 4040 } 4041 4042 #ifdef TRACESPINNING 4043 ParallelTaskTerminator::print_termination_counts(); 4044 #endif 4045 4046 gc_epilogue(false); 4047 } 4048 4049 // Print the remainder of the GC log output. 4050 log_gc_footer(os::elapsedTime() - pause_start_sec); 4051 4052 // It is not yet to safe to tell the concurrent mark to 4053 // start as we have some optional output below. We don't want the 4054 // output from the concurrent mark thread interfering with this 4055 // logging output either. 4056 4057 _hrs.verify_optional(); 4058 verify_region_sets_optional(); 4059 4060 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 4061 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 4062 4063 print_heap_after_gc(); 4064 4065 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 4066 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 4067 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 4068 // before any GC notifications are raised. 4069 g1mm()->update_sizes(); 4070 } 4071 4072 if (G1SummarizeRSetStats && 4073 (G1SummarizeRSetStatsPeriod > 0) && 4074 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 4075 g1_rem_set()->print_summary_info(); 4076 } 4077 4078 // It should now be safe to tell the concurrent mark thread to start 4079 // without its logging output interfering with the logging output 4080 // that came from the pause. 4081 4082 if (should_start_conc_mark) { 4083 // CAUTION: after the doConcurrentMark() call below, 4084 // the concurrent marking thread(s) could be running 4085 // concurrently with us. Make sure that anything after 4086 // this point does not assume that we are the only GC thread 4087 // running. Note: of course, the actual marking work will 4088 // not start until the safepoint itself is released in 4089 // ConcurrentGCThread::safepoint_desynchronize(). 4090 doConcurrentMark(); 4091 } 4092 4093 return true; 4094 } 4095 4096 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4097 { 4098 size_t gclab_word_size; 4099 switch (purpose) { 4100 case GCAllocForSurvived: 4101 gclab_word_size = _survivor_plab_stats.desired_plab_sz(); 4102 break; 4103 case GCAllocForTenured: 4104 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4105 break; 4106 default: 4107 assert(false, "unknown GCAllocPurpose"); 4108 gclab_word_size = _old_plab_stats.desired_plab_sz(); 4109 break; 4110 } 4111 4112 // Prevent humongous PLAB sizes for two reasons: 4113 // * PLABs are allocated using a similar paths as oops, but should 4114 // never be in a humongous region 4115 // * Allowing humongous PLABs needlessly churns the region free lists 4116 return MIN2(_humongous_object_threshold_in_words, gclab_word_size); 4117 } 4118 4119 void G1CollectedHeap::init_mutator_alloc_region() { 4120 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4121 _mutator_alloc_region.init(); 4122 } 4123 4124 void G1CollectedHeap::release_mutator_alloc_region() { 4125 _mutator_alloc_region.release(); 4126 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4127 } 4128 4129 void G1CollectedHeap::init_gc_alloc_regions() { 4130 assert_at_safepoint(true /* should_be_vm_thread */); 4131 4132 _survivor_gc_alloc_region.init(); 4133 _old_gc_alloc_region.init(); 4134 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4135 _retained_old_gc_alloc_region = NULL; 4136 4137 // We will discard the current GC alloc region if: 4138 // a) it's in the collection set (it can happen!), 4139 // b) it's already full (no point in using it), 4140 // c) it's empty (this means that it was emptied during 4141 // a cleanup and it should be on the free list now), or 4142 // d) it's humongous (this means that it was emptied 4143 // during a cleanup and was added to the free list, but 4144 // has been subseqently used to allocate a humongous 4145 // object that may be less than the region size). 4146 if (retained_region != NULL && 4147 !retained_region->in_collection_set() && 4148 !(retained_region->top() == retained_region->end()) && 4149 !retained_region->is_empty() && 4150 !retained_region->isHumongous()) { 4151 retained_region->set_saved_mark(); 4152 // The retained region was added to the old region set when it was 4153 // retired. We have to remove it now, since we don't allow regions 4154 // we allocate to in the region sets. We'll re-add it later, when 4155 // it's retired again. 4156 _old_set.remove(retained_region); 4157 bool during_im = g1_policy()->during_initial_mark_pause(); 4158 retained_region->note_start_of_copying(during_im); 4159 _old_gc_alloc_region.set(retained_region); 4160 _hr_printer.reuse(retained_region); 4161 } 4162 } 4163 4164 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) { 4165 _survivor_gc_alloc_region.release(); 4166 // If we have an old GC alloc region to release, we'll save it in 4167 // _retained_old_gc_alloc_region. If we don't 4168 // _retained_old_gc_alloc_region will become NULL. This is what we 4169 // want either way so no reason to check explicitly for either 4170 // condition. 4171 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4172 4173 if (ResizePLAB) { 4174 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4175 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers); 4176 } 4177 } 4178 4179 void G1CollectedHeap::abandon_gc_alloc_regions() { 4180 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4181 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4182 _retained_old_gc_alloc_region = NULL; 4183 } 4184 4185 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4186 _drain_in_progress = false; 4187 set_evac_failure_closure(cl); 4188 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4189 } 4190 4191 void G1CollectedHeap::finalize_for_evac_failure() { 4192 assert(_evac_failure_scan_stack != NULL && 4193 _evac_failure_scan_stack->length() == 0, 4194 "Postcondition"); 4195 assert(!_drain_in_progress, "Postcondition"); 4196 delete _evac_failure_scan_stack; 4197 _evac_failure_scan_stack = NULL; 4198 } 4199 4200 void G1CollectedHeap::remove_self_forwarding_pointers() { 4201 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4202 4203 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4204 4205 if (G1CollectedHeap::use_parallel_gc_threads()) { 4206 set_par_threads(); 4207 workers()->run_task(&rsfp_task); 4208 set_par_threads(0); 4209 } else { 4210 rsfp_task.work(0); 4211 } 4212 4213 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4214 4215 // Reset the claim values in the regions in the collection set. 4216 reset_cset_heap_region_claim_values(); 4217 4218 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4219 4220 // Now restore saved marks, if any. 4221 assert(_objs_with_preserved_marks.size() == 4222 _preserved_marks_of_objs.size(), "Both or none."); 4223 while (!_objs_with_preserved_marks.is_empty()) { 4224 oop obj = _objs_with_preserved_marks.pop(); 4225 markOop m = _preserved_marks_of_objs.pop(); 4226 obj->set_mark(m); 4227 } 4228 _objs_with_preserved_marks.clear(true); 4229 _preserved_marks_of_objs.clear(true); 4230 } 4231 4232 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4233 _evac_failure_scan_stack->push(obj); 4234 } 4235 4236 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4237 assert(_evac_failure_scan_stack != NULL, "precondition"); 4238 4239 while (_evac_failure_scan_stack->length() > 0) { 4240 oop obj = _evac_failure_scan_stack->pop(); 4241 _evac_failure_closure->set_region(heap_region_containing(obj)); 4242 obj->oop_iterate_backwards(_evac_failure_closure); 4243 } 4244 } 4245 4246 oop 4247 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4248 oop old) { 4249 assert(obj_in_cs(old), 4250 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4251 (HeapWord*) old)); 4252 markOop m = old->mark(); 4253 oop forward_ptr = old->forward_to_atomic(old); 4254 if (forward_ptr == NULL) { 4255 // Forward-to-self succeeded. 4256 4257 if (_evac_failure_closure != cl) { 4258 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4259 assert(!_drain_in_progress, 4260 "Should only be true while someone holds the lock."); 4261 // Set the global evac-failure closure to the current thread's. 4262 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4263 set_evac_failure_closure(cl); 4264 // Now do the common part. 4265 handle_evacuation_failure_common(old, m); 4266 // Reset to NULL. 4267 set_evac_failure_closure(NULL); 4268 } else { 4269 // The lock is already held, and this is recursive. 4270 assert(_drain_in_progress, "This should only be the recursive case."); 4271 handle_evacuation_failure_common(old, m); 4272 } 4273 return old; 4274 } else { 4275 // Forward-to-self failed. Either someone else managed to allocate 4276 // space for this object (old != forward_ptr) or they beat us in 4277 // self-forwarding it (old == forward_ptr). 4278 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4279 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4280 "should not be in the CSet", 4281 (HeapWord*) old, (HeapWord*) forward_ptr)); 4282 return forward_ptr; 4283 } 4284 } 4285 4286 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4287 set_evacuation_failed(true); 4288 4289 preserve_mark_if_necessary(old, m); 4290 4291 HeapRegion* r = heap_region_containing(old); 4292 if (!r->evacuation_failed()) { 4293 r->set_evacuation_failed(true); 4294 _hr_printer.evac_failure(r); 4295 } 4296 4297 push_on_evac_failure_scan_stack(old); 4298 4299 if (!_drain_in_progress) { 4300 // prevent recursion in copy_to_survivor_space() 4301 _drain_in_progress = true; 4302 drain_evac_failure_scan_stack(); 4303 _drain_in_progress = false; 4304 } 4305 } 4306 4307 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4308 assert(evacuation_failed(), "Oversaving!"); 4309 // We want to call the "for_promotion_failure" version only in the 4310 // case of a promotion failure. 4311 if (m->must_be_preserved_for_promotion_failure(obj)) { 4312 _objs_with_preserved_marks.push(obj); 4313 _preserved_marks_of_objs.push(m); 4314 } 4315 } 4316 4317 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4318 size_t word_size) { 4319 if (purpose == GCAllocForSurvived) { 4320 HeapWord* result = survivor_attempt_allocation(word_size); 4321 if (result != NULL) { 4322 return result; 4323 } else { 4324 // Let's try to allocate in the old gen in case we can fit the 4325 // object there. 4326 return old_attempt_allocation(word_size); 4327 } 4328 } else { 4329 assert(purpose == GCAllocForTenured, "sanity"); 4330 HeapWord* result = old_attempt_allocation(word_size); 4331 if (result != NULL) { 4332 return result; 4333 } else { 4334 // Let's try to allocate in the survivors in case we can fit the 4335 // object there. 4336 return survivor_attempt_allocation(word_size); 4337 } 4338 } 4339 4340 ShouldNotReachHere(); 4341 // Trying to keep some compilers happy. 4342 return NULL; 4343 } 4344 4345 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4346 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4347 4348 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4349 : _g1h(g1h), 4350 _refs(g1h->task_queue(queue_num)), 4351 _dcq(&g1h->dirty_card_queue_set()), 4352 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4353 _g1_rem(g1h->g1_rem_set()), 4354 _hash_seed(17), _queue_num(queue_num), 4355 _term_attempts(0), 4356 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4357 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4358 _age_table(false), 4359 _strong_roots_time(0), _term_time(0), 4360 _alloc_buffer_waste(0), _undo_waste(0) { 4361 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4362 // we "sacrifice" entry 0 to keep track of surviving bytes for 4363 // non-young regions (where the age is -1) 4364 // We also add a few elements at the beginning and at the end in 4365 // an attempt to eliminate cache contention 4366 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4367 uint array_length = PADDING_ELEM_NUM + 4368 real_length + 4369 PADDING_ELEM_NUM; 4370 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4371 if (_surviving_young_words_base == NULL) 4372 vm_exit_out_of_memory(array_length * sizeof(size_t), 4373 "Not enough space for young surv histo."); 4374 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4375 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4376 4377 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4378 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4379 4380 _start = os::elapsedTime(); 4381 } 4382 4383 void 4384 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4385 { 4386 st->print_raw_cr("GC Termination Stats"); 4387 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4388 " ------waste (KiB)------"); 4389 st->print_raw_cr("thr ms ms % ms % attempts" 4390 " total alloc undo"); 4391 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4392 " ------- ------- -------"); 4393 } 4394 4395 void 4396 G1ParScanThreadState::print_termination_stats(int i, 4397 outputStream* const st) const 4398 { 4399 const double elapsed_ms = elapsed_time() * 1000.0; 4400 const double s_roots_ms = strong_roots_time() * 1000.0; 4401 const double term_ms = term_time() * 1000.0; 4402 st->print_cr("%3d %9.2f %9.2f %6.2f " 4403 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4404 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4405 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4406 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4407 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4408 alloc_buffer_waste() * HeapWordSize / K, 4409 undo_waste() * HeapWordSize / K); 4410 } 4411 4412 #ifdef ASSERT 4413 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4414 assert(ref != NULL, "invariant"); 4415 assert(UseCompressedOops, "sanity"); 4416 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4417 oop p = oopDesc::load_decode_heap_oop(ref); 4418 assert(_g1h->is_in_g1_reserved(p), 4419 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4420 return true; 4421 } 4422 4423 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4424 assert(ref != NULL, "invariant"); 4425 if (has_partial_array_mask(ref)) { 4426 // Must be in the collection set--it's already been copied. 4427 oop p = clear_partial_array_mask(ref); 4428 assert(_g1h->obj_in_cs(p), 4429 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4430 } else { 4431 oop p = oopDesc::load_decode_heap_oop(ref); 4432 assert(_g1h->is_in_g1_reserved(p), 4433 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4434 } 4435 return true; 4436 } 4437 4438 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4439 if (ref.is_narrow()) { 4440 return verify_ref((narrowOop*) ref); 4441 } else { 4442 return verify_ref((oop*) ref); 4443 } 4444 } 4445 #endif // ASSERT 4446 4447 void G1ParScanThreadState::trim_queue() { 4448 assert(_evac_cl != NULL, "not set"); 4449 assert(_evac_failure_cl != NULL, "not set"); 4450 assert(_partial_scan_cl != NULL, "not set"); 4451 4452 StarTask ref; 4453 do { 4454 // Drain the overflow stack first, so other threads can steal. 4455 while (refs()->pop_overflow(ref)) { 4456 deal_with_reference(ref); 4457 } 4458 4459 while (refs()->pop_local(ref)) { 4460 deal_with_reference(ref); 4461 } 4462 } while (!refs()->is_empty()); 4463 } 4464 4465 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4466 G1ParScanThreadState* par_scan_state) : 4467 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4468 _par_scan_state(par_scan_state), 4469 _worker_id(par_scan_state->queue_num()), 4470 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4471 _mark_in_progress(_g1->mark_in_progress()) { } 4472 4473 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4474 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4475 #ifdef ASSERT 4476 HeapRegion* hr = _g1->heap_region_containing(obj); 4477 assert(hr != NULL, "sanity"); 4478 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4479 #endif // ASSERT 4480 4481 // We know that the object is not moving so it's safe to read its size. 4482 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4483 } 4484 4485 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4486 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4487 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4488 #ifdef ASSERT 4489 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4490 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4491 assert(from_obj != to_obj, "should not be self-forwarded"); 4492 4493 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4494 assert(from_hr != NULL, "sanity"); 4495 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4496 4497 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4498 assert(to_hr != NULL, "sanity"); 4499 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4500 #endif // ASSERT 4501 4502 // The object might be in the process of being copied by another 4503 // worker so we cannot trust that its to-space image is 4504 // well-formed. So we have to read its size from its from-space 4505 // image which we know should not be changing. 4506 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4507 } 4508 4509 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4510 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4511 ::copy_to_survivor_space(oop old) { 4512 size_t word_sz = old->size(); 4513 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4514 // +1 to make the -1 indexes valid... 4515 int young_index = from_region->young_index_in_cset()+1; 4516 assert( (from_region->is_young() && young_index > 0) || 4517 (!from_region->is_young() && young_index == 0), "invariant" ); 4518 G1CollectorPolicy* g1p = _g1->g1_policy(); 4519 markOop m = old->mark(); 4520 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4521 : m->age(); 4522 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4523 word_sz); 4524 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4525 #ifndef PRODUCT 4526 // Should this evacuation fail? 4527 if (_g1->evacuation_should_fail()) { 4528 if (obj_ptr != NULL) { 4529 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4530 obj_ptr = NULL; 4531 } 4532 } 4533 #endif // !PRODUCT 4534 4535 if (obj_ptr == NULL) { 4536 // This will either forward-to-self, or detect that someone else has 4537 // installed a forwarding pointer. 4538 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4539 return _g1->handle_evacuation_failure_par(cl, old); 4540 } 4541 4542 oop obj = oop(obj_ptr); 4543 4544 // We're going to allocate linearly, so might as well prefetch ahead. 4545 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4546 4547 oop forward_ptr = old->forward_to_atomic(obj); 4548 if (forward_ptr == NULL) { 4549 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4550 if (g1p->track_object_age(alloc_purpose)) { 4551 // We could simply do obj->incr_age(). However, this causes a 4552 // performance issue. obj->incr_age() will first check whether 4553 // the object has a displaced mark by checking its mark word; 4554 // getting the mark word from the new location of the object 4555 // stalls. So, given that we already have the mark word and we 4556 // are about to install it anyway, it's better to increase the 4557 // age on the mark word, when the object does not have a 4558 // displaced mark word. We're not expecting many objects to have 4559 // a displaced marked word, so that case is not optimized 4560 // further (it could be...) and we simply call obj->incr_age(). 4561 4562 if (m->has_displaced_mark_helper()) { 4563 // in this case, we have to install the mark word first, 4564 // otherwise obj looks to be forwarded (the old mark word, 4565 // which contains the forward pointer, was copied) 4566 obj->set_mark(m); 4567 obj->incr_age(); 4568 } else { 4569 m = m->incr_age(); 4570 obj->set_mark(m); 4571 } 4572 _par_scan_state->age_table()->add(obj, word_sz); 4573 } else { 4574 obj->set_mark(m); 4575 } 4576 4577 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4578 surv_young_words[young_index] += word_sz; 4579 4580 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4581 // We keep track of the next start index in the length field of 4582 // the to-space object. The actual length can be found in the 4583 // length field of the from-space object. 4584 arrayOop(obj)->set_length(0); 4585 oop* old_p = set_partial_array_mask(old); 4586 _par_scan_state->push_on_queue(old_p); 4587 } else { 4588 // No point in using the slower heap_region_containing() method, 4589 // given that we know obj is in the heap. 4590 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4591 obj->oop_iterate_backwards(&_scanner); 4592 } 4593 } else { 4594 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4595 obj = forward_ptr; 4596 } 4597 return obj; 4598 } 4599 4600 template <class T> 4601 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) { 4602 if (_g1->heap_region_containing_raw(new_obj)->is_young()) { 4603 _scanned_klass->record_modified_oops(); 4604 } 4605 } 4606 4607 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4608 template <class T> 4609 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4610 ::do_oop_work(T* p) { 4611 oop obj = oopDesc::load_decode_heap_oop(p); 4612 assert(barrier != G1BarrierRS || obj != NULL, 4613 "Precondition: G1BarrierRS implies obj is non-NULL"); 4614 4615 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4616 4617 // here the null check is implicit in the cset_fast_test() test 4618 if (_g1->in_cset_fast_test(obj)) { 4619 oop forwardee; 4620 if (obj->is_forwarded()) { 4621 forwardee = obj->forwardee(); 4622 } else { 4623 forwardee = copy_to_survivor_space(obj); 4624 } 4625 assert(forwardee != NULL, "forwardee should not be NULL"); 4626 oopDesc::encode_store_heap_oop(p, forwardee); 4627 if (do_mark_object && forwardee != obj) { 4628 // If the object is self-forwarded we don't need to explicitly 4629 // mark it, the evacuation failure protocol will do so. 4630 mark_forwarded_object(obj, forwardee); 4631 } 4632 4633 // When scanning the RS, we only care about objs in CS. 4634 if (barrier == G1BarrierRS) { 4635 _par_scan_state->update_rs(_from, p, _worker_id); 4636 } else if (barrier == G1BarrierKlass) { 4637 do_klass_barrier(p, forwardee); 4638 } 4639 } else { 4640 // The object is not in collection set. If we're a root scanning 4641 // closure during an initial mark pause (i.e. do_mark_object will 4642 // be true) then attempt to mark the object. 4643 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4644 mark_object(obj); 4645 } 4646 } 4647 4648 if (barrier == G1BarrierEvac && obj != NULL) { 4649 _par_scan_state->update_rs(_from, p, _worker_id); 4650 } 4651 4652 if (do_gen_barrier && obj != NULL) { 4653 par_do_barrier(p); 4654 } 4655 } 4656 4657 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4658 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4659 4660 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4661 assert(has_partial_array_mask(p), "invariant"); 4662 oop from_obj = clear_partial_array_mask(p); 4663 4664 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4665 assert(from_obj->is_objArray(), "must be obj array"); 4666 objArrayOop from_obj_array = objArrayOop(from_obj); 4667 // The from-space object contains the real length. 4668 int length = from_obj_array->length(); 4669 4670 assert(from_obj->is_forwarded(), "must be forwarded"); 4671 oop to_obj = from_obj->forwardee(); 4672 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4673 objArrayOop to_obj_array = objArrayOop(to_obj); 4674 // We keep track of the next start index in the length field of the 4675 // to-space object. 4676 int next_index = to_obj_array->length(); 4677 assert(0 <= next_index && next_index < length, 4678 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4679 4680 int start = next_index; 4681 int end = length; 4682 int remainder = end - start; 4683 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4684 if (remainder > 2 * ParGCArrayScanChunk) { 4685 end = start + ParGCArrayScanChunk; 4686 to_obj_array->set_length(end); 4687 // Push the remainder before we process the range in case another 4688 // worker has run out of things to do and can steal it. 4689 oop* from_obj_p = set_partial_array_mask(from_obj); 4690 _par_scan_state->push_on_queue(from_obj_p); 4691 } else { 4692 assert(length == end, "sanity"); 4693 // We'll process the final range for this object. Restore the length 4694 // so that the heap remains parsable in case of evacuation failure. 4695 to_obj_array->set_length(end); 4696 } 4697 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4698 // Process indexes [start,end). It will also process the header 4699 // along with the first chunk (i.e., the chunk with start == 0). 4700 // Note that at this point the length field of to_obj_array is not 4701 // correct given that we are using it to keep track of the next 4702 // start index. oop_iterate_range() (thankfully!) ignores the length 4703 // field and only relies on the start / end parameters. It does 4704 // however return the size of the object which will be incorrect. So 4705 // we have to ignore it even if we wanted to use it. 4706 to_obj_array->oop_iterate_range(&_scanner, start, end); 4707 } 4708 4709 class G1ParEvacuateFollowersClosure : public VoidClosure { 4710 protected: 4711 G1CollectedHeap* _g1h; 4712 G1ParScanThreadState* _par_scan_state; 4713 RefToScanQueueSet* _queues; 4714 ParallelTaskTerminator* _terminator; 4715 4716 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4717 RefToScanQueueSet* queues() { return _queues; } 4718 ParallelTaskTerminator* terminator() { return _terminator; } 4719 4720 public: 4721 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4722 G1ParScanThreadState* par_scan_state, 4723 RefToScanQueueSet* queues, 4724 ParallelTaskTerminator* terminator) 4725 : _g1h(g1h), _par_scan_state(par_scan_state), 4726 _queues(queues), _terminator(terminator) {} 4727 4728 void do_void(); 4729 4730 private: 4731 inline bool offer_termination(); 4732 }; 4733 4734 bool G1ParEvacuateFollowersClosure::offer_termination() { 4735 G1ParScanThreadState* const pss = par_scan_state(); 4736 pss->start_term_time(); 4737 const bool res = terminator()->offer_termination(); 4738 pss->end_term_time(); 4739 return res; 4740 } 4741 4742 void G1ParEvacuateFollowersClosure::do_void() { 4743 StarTask stolen_task; 4744 G1ParScanThreadState* const pss = par_scan_state(); 4745 pss->trim_queue(); 4746 4747 do { 4748 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4749 assert(pss->verify_task(stolen_task), "sanity"); 4750 if (stolen_task.is_narrow()) { 4751 pss->deal_with_reference((narrowOop*) stolen_task); 4752 } else { 4753 pss->deal_with_reference((oop*) stolen_task); 4754 } 4755 4756 // We've just processed a reference and we might have made 4757 // available new entries on the queues. So we have to make sure 4758 // we drain the queues as necessary. 4759 pss->trim_queue(); 4760 } 4761 } while (!offer_termination()); 4762 4763 pss->retire_alloc_buffers(); 4764 } 4765 4766 class G1KlassScanClosure : public KlassClosure { 4767 G1ParCopyHelper* _closure; 4768 bool _process_only_dirty; 4769 int _count; 4770 public: 4771 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty) 4772 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {} 4773 void do_klass(Klass* klass) { 4774 // If the klass has not been dirtied we know that there's 4775 // no references into the young gen and we can skip it. 4776 if (!_process_only_dirty || klass->has_modified_oops()) { 4777 // Clean the klass since we're going to scavenge all the metadata. 4778 klass->clear_modified_oops(); 4779 4780 // Tell the closure that this klass is the Klass to scavenge 4781 // and is the one to dirty if oops are left pointing into the young gen. 4782 _closure->set_scanned_klass(klass); 4783 4784 klass->oops_do(_closure); 4785 4786 _closure->set_scanned_klass(NULL); 4787 } 4788 _count++; 4789 } 4790 }; 4791 4792 class G1ParTask : public AbstractGangTask { 4793 protected: 4794 G1CollectedHeap* _g1h; 4795 RefToScanQueueSet *_queues; 4796 ParallelTaskTerminator _terminator; 4797 uint _n_workers; 4798 4799 Mutex _stats_lock; 4800 Mutex* stats_lock() { return &_stats_lock; } 4801 4802 size_t getNCards() { 4803 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4804 / G1BlockOffsetSharedArray::N_bytes; 4805 } 4806 4807 public: 4808 G1ParTask(G1CollectedHeap* g1h, 4809 RefToScanQueueSet *task_queues) 4810 : AbstractGangTask("G1 collection"), 4811 _g1h(g1h), 4812 _queues(task_queues), 4813 _terminator(0, _queues), 4814 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4815 {} 4816 4817 RefToScanQueueSet* queues() { return _queues; } 4818 4819 RefToScanQueue *work_queue(int i) { 4820 return queues()->queue(i); 4821 } 4822 4823 ParallelTaskTerminator* terminator() { return &_terminator; } 4824 4825 virtual void set_for_termination(int active_workers) { 4826 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4827 // in the young space (_par_seq_tasks) in the G1 heap 4828 // for SequentialSubTasksDone. 4829 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4830 // both of which need setting by set_n_termination(). 4831 _g1h->SharedHeap::set_n_termination(active_workers); 4832 _g1h->set_n_termination(active_workers); 4833 terminator()->reset_for_reuse(active_workers); 4834 _n_workers = active_workers; 4835 } 4836 4837 void work(uint worker_id) { 4838 if (worker_id >= _n_workers) return; // no work needed this round 4839 4840 double start_time_ms = os::elapsedTime() * 1000.0; 4841 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 4842 4843 { 4844 ResourceMark rm; 4845 HandleMark hm; 4846 4847 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4848 4849 G1ParScanThreadState pss(_g1h, worker_id); 4850 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4851 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4852 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4853 4854 pss.set_evac_closure(&scan_evac_cl); 4855 pss.set_evac_failure_closure(&evac_failure_cl); 4856 pss.set_partial_scan_closure(&partial_scan_cl); 4857 4858 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4859 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp); 4860 4861 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4862 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp); 4863 4864 bool only_young = _g1h->g1_policy()->gcs_are_young(); 4865 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false); 4866 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young); 4867 4868 OopClosure* scan_root_cl = &only_scan_root_cl; 4869 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s; 4870 4871 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4872 // We also need to mark copied objects. 4873 scan_root_cl = &scan_mark_root_cl; 4874 scan_klasses_cl = &scan_mark_klasses_cl_s; 4875 } 4876 4877 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4878 4879 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 4880 4881 pss.start_strong_roots(); 4882 _g1h->g1_process_strong_roots(/* is scavenging */ true, 4883 SharedHeap::ScanningOption(so), 4884 scan_root_cl, 4885 &push_heap_rs_cl, 4886 scan_klasses_cl, 4887 worker_id); 4888 pss.end_strong_roots(); 4889 4890 { 4891 double start = os::elapsedTime(); 4892 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4893 evac.do_void(); 4894 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4895 double term_ms = pss.term_time()*1000.0; 4896 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 4897 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 4898 } 4899 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4900 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4901 4902 if (ParallelGCVerbose) { 4903 MutexLocker x(stats_lock()); 4904 pss.print_termination_stats(worker_id); 4905 } 4906 4907 assert(pss.refs()->is_empty(), "should be empty"); 4908 4909 // Close the inner scope so that the ResourceMark and HandleMark 4910 // destructors are executed here and are included as part of the 4911 // "GC Worker Time". 4912 } 4913 4914 double end_time_ms = os::elapsedTime() * 1000.0; 4915 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 4916 } 4917 }; 4918 4919 // *** Common G1 Evacuation Stuff 4920 4921 // Closures that support the filtering of CodeBlobs scanned during 4922 // external root scanning. 4923 4924 // Closure applied to reference fields in code blobs (specifically nmethods) 4925 // to determine whether an nmethod contains references that point into 4926 // the collection set. Used as a predicate when walking code roots so 4927 // that only nmethods that point into the collection set are added to the 4928 // 'marked' list. 4929 4930 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure { 4931 4932 class G1PointsIntoCSOopClosure : public OopClosure { 4933 G1CollectedHeap* _g1; 4934 bool _points_into_cs; 4935 public: 4936 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) : 4937 _g1(g1), _points_into_cs(false) { } 4938 4939 bool points_into_cs() const { return _points_into_cs; } 4940 4941 template <class T> 4942 void do_oop_nv(T* p) { 4943 if (!_points_into_cs) { 4944 T heap_oop = oopDesc::load_heap_oop(p); 4945 if (!oopDesc::is_null(heap_oop) && 4946 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) { 4947 _points_into_cs = true; 4948 } 4949 } 4950 } 4951 4952 virtual void do_oop(oop* p) { do_oop_nv(p); } 4953 virtual void do_oop(narrowOop* p) { do_oop_nv(p); } 4954 }; 4955 4956 G1CollectedHeap* _g1; 4957 4958 public: 4959 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) : 4960 CodeBlobToOopClosure(cl, true), _g1(g1) { } 4961 4962 virtual void do_code_blob(CodeBlob* cb) { 4963 nmethod* nm = cb->as_nmethod_or_null(); 4964 if (nm != NULL && !(nm->test_oops_do_mark())) { 4965 G1PointsIntoCSOopClosure predicate_cl(_g1); 4966 nm->oops_do(&predicate_cl); 4967 4968 if (predicate_cl.points_into_cs()) { 4969 // At least one of the reference fields or the oop relocations 4970 // in the nmethod points into the collection set. We have to 4971 // 'mark' this nmethod. 4972 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob() 4973 // or MarkingCodeBlobClosure::do_code_blob() change. 4974 if (!nm->test_set_oops_do_mark()) { 4975 do_newly_marked_nmethod(nm); 4976 } 4977 } 4978 } 4979 } 4980 }; 4981 4982 // This method is run in a GC worker. 4983 4984 void 4985 G1CollectedHeap:: 4986 g1_process_strong_roots(bool is_scavenging, 4987 ScanningOption so, 4988 OopClosure* scan_non_heap_roots, 4989 OopsInHeapRegionClosure* scan_rs, 4990 G1KlassScanClosure* scan_klasses, 4991 int worker_i) { 4992 4993 // First scan the strong roots 4994 double ext_roots_start = os::elapsedTime(); 4995 double closure_app_time_sec = 0.0; 4996 4997 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4998 4999 // Walk the code cache w/o buffering, because StarTask cannot handle 5000 // unaligned oop locations. 5001 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots); 5002 5003 process_strong_roots(false, // no scoping; this is parallel code 5004 is_scavenging, so, 5005 &buf_scan_non_heap_roots, 5006 &eager_scan_code_roots, 5007 scan_klasses 5008 ); 5009 5010 // Now the CM ref_processor roots. 5011 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 5012 // We need to treat the discovered reference lists of the 5013 // concurrent mark ref processor as roots and keep entries 5014 // (which are added by the marking threads) on them live 5015 // until they can be processed at the end of marking. 5016 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 5017 } 5018 5019 // Finish up any enqueued closure apps (attributed as object copy time). 5020 buf_scan_non_heap_roots.done(); 5021 5022 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds(); 5023 5024 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 5025 5026 double ext_root_time_ms = 5027 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 5028 5029 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 5030 5031 // During conc marking we have to filter the per-thread SATB buffers 5032 // to make sure we remove any oops into the CSet (which will show up 5033 // as implicitly live). 5034 double satb_filtering_ms = 0.0; 5035 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 5036 if (mark_in_progress()) { 5037 double satb_filter_start = os::elapsedTime(); 5038 5039 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 5040 5041 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 5042 } 5043 } 5044 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 5045 5046 // Now scan the complement of the collection set. 5047 if (scan_rs != NULL) { 5048 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 5049 } 5050 _process_strong_tasks->all_tasks_completed(); 5051 } 5052 5053 void 5054 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 5055 OopClosure* non_root_closure) { 5056 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 5057 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 5058 } 5059 5060 // Weak Reference Processing support 5061 5062 // An always "is_alive" closure that is used to preserve referents. 5063 // If the object is non-null then it's alive. Used in the preservation 5064 // of referent objects that are pointed to by reference objects 5065 // discovered by the CM ref processor. 5066 class G1AlwaysAliveClosure: public BoolObjectClosure { 5067 G1CollectedHeap* _g1; 5068 public: 5069 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5070 void do_object(oop p) { assert(false, "Do not call."); } 5071 bool do_object_b(oop p) { 5072 if (p != NULL) { 5073 return true; 5074 } 5075 return false; 5076 } 5077 }; 5078 5079 bool G1STWIsAliveClosure::do_object_b(oop p) { 5080 // An object is reachable if it is outside the collection set, 5081 // or is inside and copied. 5082 return !_g1->obj_in_cs(p) || p->is_forwarded(); 5083 } 5084 5085 // Non Copying Keep Alive closure 5086 class G1KeepAliveClosure: public OopClosure { 5087 G1CollectedHeap* _g1; 5088 public: 5089 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 5090 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 5091 void do_oop( oop* p) { 5092 oop obj = *p; 5093 5094 if (_g1->obj_in_cs(obj)) { 5095 assert( obj->is_forwarded(), "invariant" ); 5096 *p = obj->forwardee(); 5097 } 5098 } 5099 }; 5100 5101 // Copying Keep Alive closure - can be called from both 5102 // serial and parallel code as long as different worker 5103 // threads utilize different G1ParScanThreadState instances 5104 // and different queues. 5105 5106 class G1CopyingKeepAliveClosure: public OopClosure { 5107 G1CollectedHeap* _g1h; 5108 OopClosure* _copy_non_heap_obj_cl; 5109 OopsInHeapRegionClosure* _copy_metadata_obj_cl; 5110 G1ParScanThreadState* _par_scan_state; 5111 5112 public: 5113 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 5114 OopClosure* non_heap_obj_cl, 5115 OopsInHeapRegionClosure* metadata_obj_cl, 5116 G1ParScanThreadState* pss): 5117 _g1h(g1h), 5118 _copy_non_heap_obj_cl(non_heap_obj_cl), 5119 _copy_metadata_obj_cl(metadata_obj_cl), 5120 _par_scan_state(pss) 5121 {} 5122 5123 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 5124 virtual void do_oop( oop* p) { do_oop_work(p); } 5125 5126 template <class T> void do_oop_work(T* p) { 5127 oop obj = oopDesc::load_decode_heap_oop(p); 5128 5129 if (_g1h->obj_in_cs(obj)) { 5130 // If the referent object has been forwarded (either copied 5131 // to a new location or to itself in the event of an 5132 // evacuation failure) then we need to update the reference 5133 // field and, if both reference and referent are in the G1 5134 // heap, update the RSet for the referent. 5135 // 5136 // If the referent has not been forwarded then we have to keep 5137 // it alive by policy. Therefore we have copy the referent. 5138 // 5139 // If the reference field is in the G1 heap then we can push 5140 // on the PSS queue. When the queue is drained (after each 5141 // phase of reference processing) the object and it's followers 5142 // will be copied, the reference field set to point to the 5143 // new location, and the RSet updated. Otherwise we need to 5144 // use the the non-heap or metadata closures directly to copy 5145 // the refernt object and update the pointer, while avoiding 5146 // updating the RSet. 5147 5148 if (_g1h->is_in_g1_reserved(p)) { 5149 _par_scan_state->push_on_queue(p); 5150 } else { 5151 assert(!ClassLoaderDataGraph::contains((address)p), 5152 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) " 5153 PTR_FORMAT, p)); 5154 _copy_non_heap_obj_cl->do_oop(p); 5155 } 5156 } 5157 } 5158 }; 5159 5160 // Serial drain queue closure. Called as the 'complete_gc' 5161 // closure for each discovered list in some of the 5162 // reference processing phases. 5163 5164 class G1STWDrainQueueClosure: public VoidClosure { 5165 protected: 5166 G1CollectedHeap* _g1h; 5167 G1ParScanThreadState* _par_scan_state; 5168 5169 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5170 5171 public: 5172 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5173 _g1h(g1h), 5174 _par_scan_state(pss) 5175 { } 5176 5177 void do_void() { 5178 G1ParScanThreadState* const pss = par_scan_state(); 5179 pss->trim_queue(); 5180 } 5181 }; 5182 5183 // Parallel Reference Processing closures 5184 5185 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5186 // processing during G1 evacuation pauses. 5187 5188 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5189 private: 5190 G1CollectedHeap* _g1h; 5191 RefToScanQueueSet* _queues; 5192 FlexibleWorkGang* _workers; 5193 int _active_workers; 5194 5195 public: 5196 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5197 FlexibleWorkGang* workers, 5198 RefToScanQueueSet *task_queues, 5199 int n_workers) : 5200 _g1h(g1h), 5201 _queues(task_queues), 5202 _workers(workers), 5203 _active_workers(n_workers) 5204 { 5205 assert(n_workers > 0, "shouldn't call this otherwise"); 5206 } 5207 5208 // Executes the given task using concurrent marking worker threads. 5209 virtual void execute(ProcessTask& task); 5210 virtual void execute(EnqueueTask& task); 5211 }; 5212 5213 // Gang task for possibly parallel reference processing 5214 5215 class G1STWRefProcTaskProxy: public AbstractGangTask { 5216 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5217 ProcessTask& _proc_task; 5218 G1CollectedHeap* _g1h; 5219 RefToScanQueueSet *_task_queues; 5220 ParallelTaskTerminator* _terminator; 5221 5222 public: 5223 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5224 G1CollectedHeap* g1h, 5225 RefToScanQueueSet *task_queues, 5226 ParallelTaskTerminator* terminator) : 5227 AbstractGangTask("Process reference objects in parallel"), 5228 _proc_task(proc_task), 5229 _g1h(g1h), 5230 _task_queues(task_queues), 5231 _terminator(terminator) 5232 {} 5233 5234 virtual void work(uint worker_id) { 5235 // The reference processing task executed by a single worker. 5236 ResourceMark rm; 5237 HandleMark hm; 5238 5239 G1STWIsAliveClosure is_alive(_g1h); 5240 5241 G1ParScanThreadState pss(_g1h, worker_id); 5242 5243 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5244 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5245 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5246 5247 pss.set_evac_closure(&scan_evac_cl); 5248 pss.set_evac_failure_closure(&evac_failure_cl); 5249 pss.set_partial_scan_closure(&partial_scan_cl); 5250 5251 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5252 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5253 5254 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5255 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5256 5257 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5258 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5259 5260 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5261 // We also need to mark copied objects. 5262 copy_non_heap_cl = ©_mark_non_heap_cl; 5263 copy_metadata_cl = ©_mark_metadata_cl; 5264 } 5265 5266 // Keep alive closure. 5267 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5268 5269 // Complete GC closure 5270 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5271 5272 // Call the reference processing task's work routine. 5273 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5274 5275 // Note we cannot assert that the refs array is empty here as not all 5276 // of the processing tasks (specifically phase2 - pp2_work) execute 5277 // the complete_gc closure (which ordinarily would drain the queue) so 5278 // the queue may not be empty. 5279 } 5280 }; 5281 5282 // Driver routine for parallel reference processing. 5283 // Creates an instance of the ref processing gang 5284 // task and has the worker threads execute it. 5285 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5286 assert(_workers != NULL, "Need parallel worker threads."); 5287 5288 ParallelTaskTerminator terminator(_active_workers, _queues); 5289 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5290 5291 _g1h->set_par_threads(_active_workers); 5292 _workers->run_task(&proc_task_proxy); 5293 _g1h->set_par_threads(0); 5294 } 5295 5296 // Gang task for parallel reference enqueueing. 5297 5298 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5299 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5300 EnqueueTask& _enq_task; 5301 5302 public: 5303 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5304 AbstractGangTask("Enqueue reference objects in parallel"), 5305 _enq_task(enq_task) 5306 { } 5307 5308 virtual void work(uint worker_id) { 5309 _enq_task.work(worker_id); 5310 } 5311 }; 5312 5313 // Driver routine for parallel reference enqueing. 5314 // Creates an instance of the ref enqueueing gang 5315 // task and has the worker threads execute it. 5316 5317 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5318 assert(_workers != NULL, "Need parallel worker threads."); 5319 5320 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5321 5322 _g1h->set_par_threads(_active_workers); 5323 _workers->run_task(&enq_task_proxy); 5324 _g1h->set_par_threads(0); 5325 } 5326 5327 // End of weak reference support closures 5328 5329 // Abstract task used to preserve (i.e. copy) any referent objects 5330 // that are in the collection set and are pointed to by reference 5331 // objects discovered by the CM ref processor. 5332 5333 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5334 protected: 5335 G1CollectedHeap* _g1h; 5336 RefToScanQueueSet *_queues; 5337 ParallelTaskTerminator _terminator; 5338 uint _n_workers; 5339 5340 public: 5341 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5342 AbstractGangTask("ParPreserveCMReferents"), 5343 _g1h(g1h), 5344 _queues(task_queues), 5345 _terminator(workers, _queues), 5346 _n_workers(workers) 5347 { } 5348 5349 void work(uint worker_id) { 5350 ResourceMark rm; 5351 HandleMark hm; 5352 5353 G1ParScanThreadState pss(_g1h, worker_id); 5354 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5355 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5356 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5357 5358 pss.set_evac_closure(&scan_evac_cl); 5359 pss.set_evac_failure_closure(&evac_failure_cl); 5360 pss.set_partial_scan_closure(&partial_scan_cl); 5361 5362 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5363 5364 5365 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5366 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL); 5367 5368 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5369 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL); 5370 5371 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5372 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5373 5374 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5375 // We also need to mark copied objects. 5376 copy_non_heap_cl = ©_mark_non_heap_cl; 5377 copy_metadata_cl = ©_mark_metadata_cl; 5378 } 5379 5380 // Is alive closure 5381 G1AlwaysAliveClosure always_alive(_g1h); 5382 5383 // Copying keep alive closure. Applied to referent objects that need 5384 // to be copied. 5385 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss); 5386 5387 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5388 5389 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5390 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5391 5392 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5393 // So this must be true - but assert just in case someone decides to 5394 // change the worker ids. 5395 assert(0 <= worker_id && worker_id < limit, "sanity"); 5396 assert(!rp->discovery_is_atomic(), "check this code"); 5397 5398 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5399 for (uint idx = worker_id; idx < limit; idx += stride) { 5400 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5401 5402 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5403 while (iter.has_next()) { 5404 // Since discovery is not atomic for the CM ref processor, we 5405 // can see some null referent objects. 5406 iter.load_ptrs(DEBUG_ONLY(true)); 5407 oop ref = iter.obj(); 5408 5409 // This will filter nulls. 5410 if (iter.is_referent_alive()) { 5411 iter.make_referent_alive(); 5412 } 5413 iter.move_to_next(); 5414 } 5415 } 5416 5417 // Drain the queue - which may cause stealing 5418 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5419 drain_queue.do_void(); 5420 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5421 assert(pss.refs()->is_empty(), "should be"); 5422 } 5423 }; 5424 5425 // Weak Reference processing during an evacuation pause (part 1). 5426 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) { 5427 double ref_proc_start = os::elapsedTime(); 5428 5429 ReferenceProcessor* rp = _ref_processor_stw; 5430 assert(rp->discovery_enabled(), "should have been enabled"); 5431 5432 // Any reference objects, in the collection set, that were 'discovered' 5433 // by the CM ref processor should have already been copied (either by 5434 // applying the external root copy closure to the discovered lists, or 5435 // by following an RSet entry). 5436 // 5437 // But some of the referents, that are in the collection set, that these 5438 // reference objects point to may not have been copied: the STW ref 5439 // processor would have seen that the reference object had already 5440 // been 'discovered' and would have skipped discovering the reference, 5441 // but would not have treated the reference object as a regular oop. 5442 // As a reult the copy closure would not have been applied to the 5443 // referent object. 5444 // 5445 // We need to explicitly copy these referent objects - the references 5446 // will be processed at the end of remarking. 5447 // 5448 // We also need to do this copying before we process the reference 5449 // objects discovered by the STW ref processor in case one of these 5450 // referents points to another object which is also referenced by an 5451 // object discovered by the STW ref processor. 5452 5453 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5454 no_of_gc_workers == workers()->active_workers(), 5455 "Need to reset active GC workers"); 5456 5457 set_par_threads(no_of_gc_workers); 5458 G1ParPreserveCMReferentsTask keep_cm_referents(this, 5459 no_of_gc_workers, 5460 _task_queues); 5461 5462 if (G1CollectedHeap::use_parallel_gc_threads()) { 5463 workers()->run_task(&keep_cm_referents); 5464 } else { 5465 keep_cm_referents.work(0); 5466 } 5467 5468 set_par_threads(0); 5469 5470 // Closure to test whether a referent is alive. 5471 G1STWIsAliveClosure is_alive(this); 5472 5473 // Even when parallel reference processing is enabled, the processing 5474 // of JNI refs is serial and performed serially by the current thread 5475 // rather than by a worker. The following PSS will be used for processing 5476 // JNI refs. 5477 5478 // Use only a single queue for this PSS. 5479 G1ParScanThreadState pss(this, 0); 5480 5481 // We do not embed a reference processor in the copying/scanning 5482 // closures while we're actually processing the discovered 5483 // reference objects. 5484 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5485 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5486 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5487 5488 pss.set_evac_closure(&scan_evac_cl); 5489 pss.set_evac_failure_closure(&evac_failure_cl); 5490 pss.set_partial_scan_closure(&partial_scan_cl); 5491 5492 assert(pss.refs()->is_empty(), "pre-condition"); 5493 5494 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5495 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL); 5496 5497 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5498 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL); 5499 5500 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5501 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl; 5502 5503 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5504 // We also need to mark copied objects. 5505 copy_non_heap_cl = ©_mark_non_heap_cl; 5506 copy_metadata_cl = ©_mark_metadata_cl; 5507 } 5508 5509 // Keep alive closure. 5510 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss); 5511 5512 // Serial Complete GC closure 5513 G1STWDrainQueueClosure drain_queue(this, &pss); 5514 5515 // Setup the soft refs policy... 5516 rp->setup_policy(false); 5517 5518 if (!rp->processing_is_mt()) { 5519 // Serial reference processing... 5520 rp->process_discovered_references(&is_alive, 5521 &keep_alive, 5522 &drain_queue, 5523 NULL); 5524 } else { 5525 // Parallel reference processing 5526 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5527 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5528 5529 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5530 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5531 } 5532 5533 // We have completed copying any necessary live referent objects 5534 // (that were not copied during the actual pause) so we can 5535 // retire any active alloc buffers 5536 pss.retire_alloc_buffers(); 5537 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5538 5539 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5540 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5541 } 5542 5543 // Weak Reference processing during an evacuation pause (part 2). 5544 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) { 5545 double ref_enq_start = os::elapsedTime(); 5546 5547 ReferenceProcessor* rp = _ref_processor_stw; 5548 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5549 5550 // Now enqueue any remaining on the discovered lists on to 5551 // the pending list. 5552 if (!rp->processing_is_mt()) { 5553 // Serial reference processing... 5554 rp->enqueue_discovered_references(); 5555 } else { 5556 // Parallel reference enqueuing 5557 5558 assert(no_of_gc_workers == workers()->active_workers(), 5559 "Need to reset active workers"); 5560 assert(rp->num_q() == no_of_gc_workers, "sanity"); 5561 assert(no_of_gc_workers <= rp->max_num_q(), "sanity"); 5562 5563 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers); 5564 rp->enqueue_discovered_references(&par_task_executor); 5565 } 5566 5567 rp->verify_no_references_recorded(); 5568 assert(!rp->discovery_enabled(), "should have been disabled"); 5569 5570 // FIXME 5571 // CM's reference processing also cleans up the string and symbol tables. 5572 // Should we do that here also? We could, but it is a serial operation 5573 // and could signicantly increase the pause time. 5574 5575 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5576 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5577 } 5578 5579 void G1CollectedHeap::evacuate_collection_set() { 5580 _expand_heap_after_alloc_failure = true; 5581 set_evacuation_failed(false); 5582 5583 // Should G1EvacuationFailureALot be in effect for this GC? 5584 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();) 5585 5586 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5587 concurrent_g1_refine()->set_use_cache(false); 5588 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5589 5590 uint n_workers; 5591 if (G1CollectedHeap::use_parallel_gc_threads()) { 5592 n_workers = 5593 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5594 workers()->active_workers(), 5595 Threads::number_of_non_daemon_threads()); 5596 assert(UseDynamicNumberOfGCThreads || 5597 n_workers == workers()->total_workers(), 5598 "If not dynamic should be using all the workers"); 5599 workers()->set_active_workers(n_workers); 5600 set_par_threads(n_workers); 5601 } else { 5602 assert(n_par_threads() == 0, 5603 "Should be the original non-parallel value"); 5604 n_workers = 1; 5605 } 5606 5607 G1ParTask g1_par_task(this, _task_queues); 5608 5609 init_for_evac_failure(NULL); 5610 5611 rem_set()->prepare_for_younger_refs_iterate(true); 5612 5613 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5614 double start_par_time_sec = os::elapsedTime(); 5615 double end_par_time_sec; 5616 5617 { 5618 StrongRootsScope srs(this); 5619 5620 if (G1CollectedHeap::use_parallel_gc_threads()) { 5621 // The individual threads will set their evac-failure closures. 5622 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5623 // These tasks use ShareHeap::_process_strong_tasks 5624 assert(UseDynamicNumberOfGCThreads || 5625 workers()->active_workers() == workers()->total_workers(), 5626 "If not dynamic should be using all the workers"); 5627 workers()->run_task(&g1_par_task); 5628 } else { 5629 g1_par_task.set_for_termination(n_workers); 5630 g1_par_task.work(0); 5631 } 5632 end_par_time_sec = os::elapsedTime(); 5633 5634 // Closing the inner scope will execute the destructor 5635 // for the StrongRootsScope object. We record the current 5636 // elapsed time before closing the scope so that time 5637 // taken for the SRS destructor is NOT included in the 5638 // reported parallel time. 5639 } 5640 5641 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5642 g1_policy()->phase_times()->record_par_time(par_time_ms); 5643 5644 double code_root_fixup_time_ms = 5645 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5646 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5647 5648 set_par_threads(0); 5649 5650 // Process any discovered reference objects - we have 5651 // to do this _before_ we retire the GC alloc regions 5652 // as we may have to copy some 'reachable' referent 5653 // objects (and their reachable sub-graphs) that were 5654 // not copied during the pause. 5655 process_discovered_references(n_workers); 5656 5657 // Weak root processing. 5658 // Note: when JSR 292 is enabled and code blobs can contain 5659 // non-perm oops then we will need to process the code blobs 5660 // here too. 5661 { 5662 G1STWIsAliveClosure is_alive(this); 5663 G1KeepAliveClosure keep_alive(this); 5664 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5665 } 5666 5667 release_gc_alloc_regions(n_workers); 5668 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5669 5670 concurrent_g1_refine()->clear_hot_cache(); 5671 concurrent_g1_refine()->set_use_cache(true); 5672 5673 finalize_for_evac_failure(); 5674 5675 if (evacuation_failed()) { 5676 remove_self_forwarding_pointers(); 5677 5678 // Reset the G1EvacuationFailureALot counters and flags 5679 // Note: the values are reset only when an actual 5680 // evacuation failure occurs. 5681 NOT_PRODUCT(reset_evacuation_should_fail();) 5682 } 5683 5684 // Enqueue any remaining references remaining on the STW 5685 // reference processor's discovered lists. We need to do 5686 // this after the card table is cleaned (and verified) as 5687 // the act of enqueuing entries on to the pending list 5688 // will log these updates (and dirty their associated 5689 // cards). We need these updates logged to update any 5690 // RSets. 5691 enqueue_discovered_references(n_workers); 5692 5693 if (G1DeferredRSUpdate) { 5694 RedirtyLoggedCardTableEntryFastClosure redirty; 5695 dirty_card_queue_set().set_closure(&redirty); 5696 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5697 5698 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5699 dcq.merge_bufferlists(&dirty_card_queue_set()); 5700 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5701 } 5702 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5703 } 5704 5705 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5706 size_t* pre_used, 5707 FreeRegionList* free_list, 5708 OldRegionSet* old_proxy_set, 5709 HumongousRegionSet* humongous_proxy_set, 5710 HRRSCleanupTask* hrrs_cleanup_task, 5711 bool par) { 5712 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5713 if (hr->isHumongous()) { 5714 assert(hr->startsHumongous(), "we should only see starts humongous"); 5715 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5716 } else { 5717 _old_set.remove_with_proxy(hr, old_proxy_set); 5718 free_region(hr, pre_used, free_list, par); 5719 } 5720 } else { 5721 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5722 } 5723 } 5724 5725 void G1CollectedHeap::free_region(HeapRegion* hr, 5726 size_t* pre_used, 5727 FreeRegionList* free_list, 5728 bool par) { 5729 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5730 assert(!hr->is_empty(), "the region should not be empty"); 5731 assert(free_list != NULL, "pre-condition"); 5732 5733 *pre_used += hr->used(); 5734 hr->hr_clear(par, true /* clear_space */); 5735 free_list->add_as_head(hr); 5736 } 5737 5738 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5739 size_t* pre_used, 5740 FreeRegionList* free_list, 5741 HumongousRegionSet* humongous_proxy_set, 5742 bool par) { 5743 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5744 assert(free_list != NULL, "pre-condition"); 5745 assert(humongous_proxy_set != NULL, "pre-condition"); 5746 5747 size_t hr_used = hr->used(); 5748 size_t hr_capacity = hr->capacity(); 5749 size_t hr_pre_used = 0; 5750 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5751 // We need to read this before we make the region non-humongous, 5752 // otherwise the information will be gone. 5753 uint last_index = hr->last_hc_index(); 5754 hr->set_notHumongous(); 5755 free_region(hr, &hr_pre_used, free_list, par); 5756 5757 uint i = hr->hrs_index() + 1; 5758 while (i < last_index) { 5759 HeapRegion* curr_hr = region_at(i); 5760 assert(curr_hr->continuesHumongous(), "invariant"); 5761 curr_hr->set_notHumongous(); 5762 free_region(curr_hr, &hr_pre_used, free_list, par); 5763 i += 1; 5764 } 5765 assert(hr_pre_used == hr_used, 5766 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5767 "should be the same", hr_pre_used, hr_used)); 5768 *pre_used += hr_pre_used; 5769 } 5770 5771 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5772 FreeRegionList* free_list, 5773 OldRegionSet* old_proxy_set, 5774 HumongousRegionSet* humongous_proxy_set, 5775 bool par) { 5776 if (pre_used > 0) { 5777 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5778 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5779 assert(_summary_bytes_used >= pre_used, 5780 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5781 "should be >= pre_used: "SIZE_FORMAT, 5782 _summary_bytes_used, pre_used)); 5783 _summary_bytes_used -= pre_used; 5784 } 5785 if (free_list != NULL && !free_list->is_empty()) { 5786 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5787 _free_list.add_as_head(free_list); 5788 } 5789 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5790 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5791 _old_set.update_from_proxy(old_proxy_set); 5792 } 5793 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5794 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5795 _humongous_set.update_from_proxy(humongous_proxy_set); 5796 } 5797 } 5798 5799 class G1ParCleanupCTTask : public AbstractGangTask { 5800 CardTableModRefBS* _ct_bs; 5801 G1CollectedHeap* _g1h; 5802 HeapRegion* volatile _su_head; 5803 public: 5804 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5805 G1CollectedHeap* g1h) : 5806 AbstractGangTask("G1 Par Cleanup CT Task"), 5807 _ct_bs(ct_bs), _g1h(g1h) { } 5808 5809 void work(uint worker_id) { 5810 HeapRegion* r; 5811 while (r = _g1h->pop_dirty_cards_region()) { 5812 clear_cards(r); 5813 } 5814 } 5815 5816 void clear_cards(HeapRegion* r) { 5817 // Cards of the survivors should have already been dirtied. 5818 if (!r->is_survivor()) { 5819 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5820 } 5821 } 5822 }; 5823 5824 #ifndef PRODUCT 5825 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5826 G1CollectedHeap* _g1h; 5827 CardTableModRefBS* _ct_bs; 5828 public: 5829 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5830 : _g1h(g1h), _ct_bs(ct_bs) { } 5831 virtual bool doHeapRegion(HeapRegion* r) { 5832 if (r->is_survivor()) { 5833 _g1h->verify_dirty_region(r); 5834 } else { 5835 _g1h->verify_not_dirty_region(r); 5836 } 5837 return false; 5838 } 5839 }; 5840 5841 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5842 // All of the region should be clean. 5843 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5844 MemRegion mr(hr->bottom(), hr->end()); 5845 ct_bs->verify_not_dirty_region(mr); 5846 } 5847 5848 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5849 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5850 // dirty allocated blocks as they allocate them. The thread that 5851 // retires each region and replaces it with a new one will do a 5852 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5853 // not dirty that area (one less thing to have to do while holding 5854 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5855 // is dirty. 5856 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5857 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5858 ct_bs->verify_dirty_region(mr); 5859 } 5860 5861 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5862 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5863 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5864 verify_dirty_region(hr); 5865 } 5866 } 5867 5868 void G1CollectedHeap::verify_dirty_young_regions() { 5869 verify_dirty_young_list(_young_list->first_region()); 5870 } 5871 #endif 5872 5873 void G1CollectedHeap::cleanUpCardTable() { 5874 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5875 double start = os::elapsedTime(); 5876 5877 { 5878 // Iterate over the dirty cards region list. 5879 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5880 5881 if (G1CollectedHeap::use_parallel_gc_threads()) { 5882 set_par_threads(); 5883 workers()->run_task(&cleanup_task); 5884 set_par_threads(0); 5885 } else { 5886 while (_dirty_cards_region_list) { 5887 HeapRegion* r = _dirty_cards_region_list; 5888 cleanup_task.clear_cards(r); 5889 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5890 if (_dirty_cards_region_list == r) { 5891 // The last region. 5892 _dirty_cards_region_list = NULL; 5893 } 5894 r->set_next_dirty_cards_region(NULL); 5895 } 5896 } 5897 #ifndef PRODUCT 5898 if (G1VerifyCTCleanup || VerifyAfterGC) { 5899 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5900 heap_region_iterate(&cleanup_verifier); 5901 } 5902 #endif 5903 } 5904 5905 double elapsed = os::elapsedTime() - start; 5906 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 5907 } 5908 5909 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5910 size_t pre_used = 0; 5911 FreeRegionList local_free_list("Local List for CSet Freeing"); 5912 5913 double young_time_ms = 0.0; 5914 double non_young_time_ms = 0.0; 5915 5916 // Since the collection set is a superset of the the young list, 5917 // all we need to do to clear the young list is clear its 5918 // head and length, and unlink any young regions in the code below 5919 _young_list->clear(); 5920 5921 G1CollectorPolicy* policy = g1_policy(); 5922 5923 double start_sec = os::elapsedTime(); 5924 bool non_young = true; 5925 5926 HeapRegion* cur = cs_head; 5927 int age_bound = -1; 5928 size_t rs_lengths = 0; 5929 5930 while (cur != NULL) { 5931 assert(!is_on_master_free_list(cur), "sanity"); 5932 if (non_young) { 5933 if (cur->is_young()) { 5934 double end_sec = os::elapsedTime(); 5935 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5936 non_young_time_ms += elapsed_ms; 5937 5938 start_sec = os::elapsedTime(); 5939 non_young = false; 5940 } 5941 } else { 5942 if (!cur->is_young()) { 5943 double end_sec = os::elapsedTime(); 5944 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5945 young_time_ms += elapsed_ms; 5946 5947 start_sec = os::elapsedTime(); 5948 non_young = true; 5949 } 5950 } 5951 5952 rs_lengths += cur->rem_set()->occupied(); 5953 5954 HeapRegion* next = cur->next_in_collection_set(); 5955 assert(cur->in_collection_set(), "bad CS"); 5956 cur->set_next_in_collection_set(NULL); 5957 cur->set_in_collection_set(false); 5958 5959 if (cur->is_young()) { 5960 int index = cur->young_index_in_cset(); 5961 assert(index != -1, "invariant"); 5962 assert((uint) index < policy->young_cset_region_length(), "invariant"); 5963 size_t words_survived = _surviving_young_words[index]; 5964 cur->record_surv_words_in_group(words_survived); 5965 5966 // At this point the we have 'popped' cur from the collection set 5967 // (linked via next_in_collection_set()) but it is still in the 5968 // young list (linked via next_young_region()). Clear the 5969 // _next_young_region field. 5970 cur->set_next_young_region(NULL); 5971 } else { 5972 int index = cur->young_index_in_cset(); 5973 assert(index == -1, "invariant"); 5974 } 5975 5976 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5977 (!cur->is_young() && cur->young_index_in_cset() == -1), 5978 "invariant" ); 5979 5980 if (!cur->evacuation_failed()) { 5981 MemRegion used_mr = cur->used_region(); 5982 5983 // And the region is empty. 5984 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5985 free_region(cur, &pre_used, &local_free_list, false /* par */); 5986 } else { 5987 cur->uninstall_surv_rate_group(); 5988 if (cur->is_young()) { 5989 cur->set_young_index_in_cset(-1); 5990 } 5991 cur->set_not_young(); 5992 cur->set_evacuation_failed(false); 5993 // The region is now considered to be old. 5994 _old_set.add(cur); 5995 } 5996 cur = next; 5997 } 5998 5999 policy->record_max_rs_lengths(rs_lengths); 6000 policy->cset_regions_freed(); 6001 6002 double end_sec = os::elapsedTime(); 6003 double elapsed_ms = (end_sec - start_sec) * 1000.0; 6004 6005 if (non_young) { 6006 non_young_time_ms += elapsed_ms; 6007 } else { 6008 young_time_ms += elapsed_ms; 6009 } 6010 6011 update_sets_after_freeing_regions(pre_used, &local_free_list, 6012 NULL /* old_proxy_set */, 6013 NULL /* humongous_proxy_set */, 6014 false /* par */); 6015 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 6016 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 6017 } 6018 6019 // This routine is similar to the above but does not record 6020 // any policy statistics or update free lists; we are abandoning 6021 // the current incremental collection set in preparation of a 6022 // full collection. After the full GC we will start to build up 6023 // the incremental collection set again. 6024 // This is only called when we're doing a full collection 6025 // and is immediately followed by the tearing down of the young list. 6026 6027 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 6028 HeapRegion* cur = cs_head; 6029 6030 while (cur != NULL) { 6031 HeapRegion* next = cur->next_in_collection_set(); 6032 assert(cur->in_collection_set(), "bad CS"); 6033 cur->set_next_in_collection_set(NULL); 6034 cur->set_in_collection_set(false); 6035 cur->set_young_index_in_cset(-1); 6036 cur = next; 6037 } 6038 } 6039 6040 void G1CollectedHeap::set_free_regions_coming() { 6041 if (G1ConcRegionFreeingVerbose) { 6042 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6043 "setting free regions coming"); 6044 } 6045 6046 assert(!free_regions_coming(), "pre-condition"); 6047 _free_regions_coming = true; 6048 } 6049 6050 void G1CollectedHeap::reset_free_regions_coming() { 6051 assert(free_regions_coming(), "pre-condition"); 6052 6053 { 6054 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6055 _free_regions_coming = false; 6056 SecondaryFreeList_lock->notify_all(); 6057 } 6058 6059 if (G1ConcRegionFreeingVerbose) { 6060 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 6061 "reset free regions coming"); 6062 } 6063 } 6064 6065 void G1CollectedHeap::wait_while_free_regions_coming() { 6066 // Most of the time we won't have to wait, so let's do a quick test 6067 // first before we take the lock. 6068 if (!free_regions_coming()) { 6069 return; 6070 } 6071 6072 if (G1ConcRegionFreeingVerbose) { 6073 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6074 "waiting for free regions"); 6075 } 6076 6077 { 6078 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6079 while (free_regions_coming()) { 6080 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 6081 } 6082 } 6083 6084 if (G1ConcRegionFreeingVerbose) { 6085 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 6086 "done waiting for free regions"); 6087 } 6088 } 6089 6090 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 6091 assert(heap_lock_held_for_gc(), 6092 "the heap lock should already be held by or for this thread"); 6093 _young_list->push_region(hr); 6094 } 6095 6096 class NoYoungRegionsClosure: public HeapRegionClosure { 6097 private: 6098 bool _success; 6099 public: 6100 NoYoungRegionsClosure() : _success(true) { } 6101 bool doHeapRegion(HeapRegion* r) { 6102 if (r->is_young()) { 6103 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 6104 r->bottom(), r->end()); 6105 _success = false; 6106 } 6107 return false; 6108 } 6109 bool success() { return _success; } 6110 }; 6111 6112 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 6113 bool ret = _young_list->check_list_empty(check_sample); 6114 6115 if (check_heap) { 6116 NoYoungRegionsClosure closure; 6117 heap_region_iterate(&closure); 6118 ret = ret && closure.success(); 6119 } 6120 6121 return ret; 6122 } 6123 6124 class TearDownRegionSetsClosure : public HeapRegionClosure { 6125 private: 6126 OldRegionSet *_old_set; 6127 6128 public: 6129 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 6130 6131 bool doHeapRegion(HeapRegion* r) { 6132 if (r->is_empty()) { 6133 // We ignore empty regions, we'll empty the free list afterwards 6134 } else if (r->is_young()) { 6135 // We ignore young regions, we'll empty the young list afterwards 6136 } else if (r->isHumongous()) { 6137 // We ignore humongous regions, we're not tearing down the 6138 // humongous region set 6139 } else { 6140 // The rest should be old 6141 _old_set->remove(r); 6142 } 6143 return false; 6144 } 6145 6146 ~TearDownRegionSetsClosure() { 6147 assert(_old_set->is_empty(), "post-condition"); 6148 } 6149 }; 6150 6151 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6152 assert_at_safepoint(true /* should_be_vm_thread */); 6153 6154 if (!free_list_only) { 6155 TearDownRegionSetsClosure cl(&_old_set); 6156 heap_region_iterate(&cl); 6157 6158 // Need to do this after the heap iteration to be able to 6159 // recognize the young regions and ignore them during the iteration. 6160 _young_list->empty_list(); 6161 } 6162 _free_list.remove_all(); 6163 } 6164 6165 class RebuildRegionSetsClosure : public HeapRegionClosure { 6166 private: 6167 bool _free_list_only; 6168 OldRegionSet* _old_set; 6169 FreeRegionList* _free_list; 6170 size_t _total_used; 6171 6172 public: 6173 RebuildRegionSetsClosure(bool free_list_only, 6174 OldRegionSet* old_set, FreeRegionList* free_list) : 6175 _free_list_only(free_list_only), 6176 _old_set(old_set), _free_list(free_list), _total_used(0) { 6177 assert(_free_list->is_empty(), "pre-condition"); 6178 if (!free_list_only) { 6179 assert(_old_set->is_empty(), "pre-condition"); 6180 } 6181 } 6182 6183 bool doHeapRegion(HeapRegion* r) { 6184 if (r->continuesHumongous()) { 6185 return false; 6186 } 6187 6188 if (r->is_empty()) { 6189 // Add free regions to the free list 6190 _free_list->add_as_tail(r); 6191 } else if (!_free_list_only) { 6192 assert(!r->is_young(), "we should not come across young regions"); 6193 6194 if (r->isHumongous()) { 6195 // We ignore humongous regions, we left the humongous set unchanged 6196 } else { 6197 // The rest should be old, add them to the old set 6198 _old_set->add(r); 6199 } 6200 _total_used += r->used(); 6201 } 6202 6203 return false; 6204 } 6205 6206 size_t total_used() { 6207 return _total_used; 6208 } 6209 }; 6210 6211 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6212 assert_at_safepoint(true /* should_be_vm_thread */); 6213 6214 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6215 heap_region_iterate(&cl); 6216 6217 if (!free_list_only) { 6218 _summary_bytes_used = cl.total_used(); 6219 } 6220 assert(_summary_bytes_used == recalculate_used(), 6221 err_msg("inconsistent _summary_bytes_used, " 6222 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6223 _summary_bytes_used, recalculate_used())); 6224 } 6225 6226 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6227 _refine_cte_cl->set_concurrent(concurrent); 6228 } 6229 6230 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6231 HeapRegion* hr = heap_region_containing(p); 6232 if (hr == NULL) { 6233 return false; 6234 } else { 6235 return hr->is_in(p); 6236 } 6237 } 6238 6239 // Methods for the mutator alloc region 6240 6241 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6242 bool force) { 6243 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6244 assert(!force || g1_policy()->can_expand_young_list(), 6245 "if force is true we should be able to expand the young list"); 6246 bool young_list_full = g1_policy()->is_young_list_full(); 6247 if (force || !young_list_full) { 6248 HeapRegion* new_alloc_region = new_region(word_size, 6249 false /* do_expand */); 6250 if (new_alloc_region != NULL) { 6251 set_region_short_lived_locked(new_alloc_region); 6252 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6253 return new_alloc_region; 6254 } 6255 } 6256 return NULL; 6257 } 6258 6259 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6260 size_t allocated_bytes) { 6261 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6262 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6263 6264 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6265 _summary_bytes_used += allocated_bytes; 6266 _hr_printer.retire(alloc_region); 6267 // We update the eden sizes here, when the region is retired, 6268 // instead of when it's allocated, since this is the point that its 6269 // used space has been recored in _summary_bytes_used. 6270 g1mm()->update_eden_size(); 6271 } 6272 6273 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6274 bool force) { 6275 return _g1h->new_mutator_alloc_region(word_size, force); 6276 } 6277 6278 void G1CollectedHeap::set_par_threads() { 6279 // Don't change the number of workers. Use the value previously set 6280 // in the workgroup. 6281 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6282 uint n_workers = workers()->active_workers(); 6283 assert(UseDynamicNumberOfGCThreads || 6284 n_workers == workers()->total_workers(), 6285 "Otherwise should be using the total number of workers"); 6286 if (n_workers == 0) { 6287 assert(false, "Should have been set in prior evacuation pause."); 6288 n_workers = ParallelGCThreads; 6289 workers()->set_active_workers(n_workers); 6290 } 6291 set_par_threads(n_workers); 6292 } 6293 6294 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6295 size_t allocated_bytes) { 6296 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6297 } 6298 6299 // Methods for the GC alloc regions 6300 6301 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6302 uint count, 6303 GCAllocPurpose ap) { 6304 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6305 6306 if (count < g1_policy()->max_regions(ap)) { 6307 HeapRegion* new_alloc_region = new_region(word_size, 6308 true /* do_expand */); 6309 if (new_alloc_region != NULL) { 6310 // We really only need to do this for old regions given that we 6311 // should never scan survivors. But it doesn't hurt to do it 6312 // for survivors too. 6313 new_alloc_region->set_saved_mark(); 6314 if (ap == GCAllocForSurvived) { 6315 new_alloc_region->set_survivor(); 6316 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6317 } else { 6318 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6319 } 6320 bool during_im = g1_policy()->during_initial_mark_pause(); 6321 new_alloc_region->note_start_of_copying(during_im); 6322 return new_alloc_region; 6323 } else { 6324 g1_policy()->note_alloc_region_limit_reached(ap); 6325 } 6326 } 6327 return NULL; 6328 } 6329 6330 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6331 size_t allocated_bytes, 6332 GCAllocPurpose ap) { 6333 bool during_im = g1_policy()->during_initial_mark_pause(); 6334 alloc_region->note_end_of_copying(during_im); 6335 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6336 if (ap == GCAllocForSurvived) { 6337 young_list()->add_survivor_region(alloc_region); 6338 } else { 6339 _old_set.add(alloc_region); 6340 } 6341 _hr_printer.retire(alloc_region); 6342 } 6343 6344 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6345 bool force) { 6346 assert(!force, "not supported for GC alloc regions"); 6347 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6348 } 6349 6350 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6351 size_t allocated_bytes) { 6352 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6353 GCAllocForSurvived); 6354 } 6355 6356 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6357 bool force) { 6358 assert(!force, "not supported for GC alloc regions"); 6359 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6360 } 6361 6362 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6363 size_t allocated_bytes) { 6364 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6365 GCAllocForTenured); 6366 } 6367 // Heap region set verification 6368 6369 class VerifyRegionListsClosure : public HeapRegionClosure { 6370 private: 6371 FreeRegionList* _free_list; 6372 OldRegionSet* _old_set; 6373 HumongousRegionSet* _humongous_set; 6374 uint _region_count; 6375 6376 public: 6377 VerifyRegionListsClosure(OldRegionSet* old_set, 6378 HumongousRegionSet* humongous_set, 6379 FreeRegionList* free_list) : 6380 _old_set(old_set), _humongous_set(humongous_set), 6381 _free_list(free_list), _region_count(0) { } 6382 6383 uint region_count() { return _region_count; } 6384 6385 bool doHeapRegion(HeapRegion* hr) { 6386 _region_count += 1; 6387 6388 if (hr->continuesHumongous()) { 6389 return false; 6390 } 6391 6392 if (hr->is_young()) { 6393 // TODO 6394 } else if (hr->startsHumongous()) { 6395 _humongous_set->verify_next_region(hr); 6396 } else if (hr->is_empty()) { 6397 _free_list->verify_next_region(hr); 6398 } else { 6399 _old_set->verify_next_region(hr); 6400 } 6401 return false; 6402 } 6403 }; 6404 6405 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6406 HeapWord* bottom) { 6407 HeapWord* end = bottom + HeapRegion::GrainWords; 6408 MemRegion mr(bottom, end); 6409 assert(_g1_reserved.contains(mr), "invariant"); 6410 // This might return NULL if the allocation fails 6411 return new HeapRegion(hrs_index, _bot_shared, mr); 6412 } 6413 6414 void G1CollectedHeap::verify_region_sets() { 6415 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6416 6417 // First, check the explicit lists. 6418 _free_list.verify(); 6419 { 6420 // Given that a concurrent operation might be adding regions to 6421 // the secondary free list we have to take the lock before 6422 // verifying it. 6423 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6424 _secondary_free_list.verify(); 6425 } 6426 _old_set.verify(); 6427 _humongous_set.verify(); 6428 6429 // If a concurrent region freeing operation is in progress it will 6430 // be difficult to correctly attributed any free regions we come 6431 // across to the correct free list given that they might belong to 6432 // one of several (free_list, secondary_free_list, any local lists, 6433 // etc.). So, if that's the case we will skip the rest of the 6434 // verification operation. Alternatively, waiting for the concurrent 6435 // operation to complete will have a non-trivial effect on the GC's 6436 // operation (no concurrent operation will last longer than the 6437 // interval between two calls to verification) and it might hide 6438 // any issues that we would like to catch during testing. 6439 if (free_regions_coming()) { 6440 return; 6441 } 6442 6443 // Make sure we append the secondary_free_list on the free_list so 6444 // that all free regions we will come across can be safely 6445 // attributed to the free_list. 6446 append_secondary_free_list_if_not_empty_with_lock(); 6447 6448 // Finally, make sure that the region accounting in the lists is 6449 // consistent with what we see in the heap. 6450 _old_set.verify_start(); 6451 _humongous_set.verify_start(); 6452 _free_list.verify_start(); 6453 6454 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6455 heap_region_iterate(&cl); 6456 6457 _old_set.verify_end(); 6458 _humongous_set.verify_end(); 6459 _free_list.verify_end(); 6460 }