1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "code/icBuffer.hpp" 27 #include "gc_implementation/g1/bufferingOopClosure.hpp" 28 #include "gc_implementation/g1/concurrentG1Refine.hpp" 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp" 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp" 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 35 #include "gc_implementation/g1/g1EvacFailure.hpp" 36 #include "gc_implementation/g1/g1Log.hpp" 37 #include "gc_implementation/g1/g1MarkSweep.hpp" 38 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 39 #include "gc_implementation/g1/g1RemSet.inline.hpp" 40 #include "gc_implementation/g1/heapRegion.inline.hpp" 41 #include "gc_implementation/g1/heapRegionRemSet.hpp" 42 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 43 #include "gc_implementation/g1/vm_operations_g1.hpp" 44 #include "gc_implementation/shared/isGCActiveMark.hpp" 45 #include "memory/gcLocker.inline.hpp" 46 #include "memory/genOopClosures.inline.hpp" 47 #include "memory/generationSpec.hpp" 48 #include "memory/referenceProcessor.hpp" 49 #include "oops/oop.inline.hpp" 50 #include "oops/oop.pcgc.inline.hpp" 51 #include "runtime/aprofiler.hpp" 52 #include "runtime/vmThread.hpp" 53 54 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 55 56 // turn it on so that the contents of the young list (scan-only / 57 // to-be-collected) are printed at "strategic" points before / during 58 // / after the collection --- this is useful for debugging 59 #define YOUNG_LIST_VERBOSE 0 60 // CURRENT STATUS 61 // This file is under construction. Search for "FIXME". 62 63 // INVARIANTS/NOTES 64 // 65 // All allocation activity covered by the G1CollectedHeap interface is 66 // serialized by acquiring the HeapLock. This happens in mem_allocate 67 // and allocate_new_tlab, which are the "entry" points to the 68 // allocation code from the rest of the JVM. (Note that this does not 69 // apply to TLAB allocation, which is not part of this interface: it 70 // is done by clients of this interface.) 71 72 // Notes on implementation of parallelism in different tasks. 73 // 74 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 75 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 76 // It does use run_task() which sets _n_workers in the task. 77 // G1ParTask executes g1_process_strong_roots() -> 78 // SharedHeap::process_strong_roots() which calls eventuall to 79 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 80 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 81 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 82 // 83 84 // Local to this file. 85 86 class RefineCardTableEntryClosure: public CardTableEntryClosure { 87 SuspendibleThreadSet* _sts; 88 G1RemSet* _g1rs; 89 ConcurrentG1Refine* _cg1r; 90 bool _concurrent; 91 public: 92 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 93 G1RemSet* g1rs, 94 ConcurrentG1Refine* cg1r) : 95 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 96 {} 97 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 98 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false); 99 // This path is executed by the concurrent refine or mutator threads, 100 // concurrently, and so we do not care if card_ptr contains references 101 // that point into the collection set. 102 assert(!oops_into_cset, "should be"); 103 104 if (_concurrent && _sts->should_yield()) { 105 // Caller will actually yield. 106 return false; 107 } 108 // Otherwise, we finished successfully; return true. 109 return true; 110 } 111 void set_concurrent(bool b) { _concurrent = b; } 112 }; 113 114 115 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 116 int _calls; 117 G1CollectedHeap* _g1h; 118 CardTableModRefBS* _ctbs; 119 int _histo[256]; 120 public: 121 ClearLoggedCardTableEntryClosure() : 122 _calls(0) 123 { 124 _g1h = G1CollectedHeap::heap(); 125 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 126 for (int i = 0; i < 256; i++) _histo[i] = 0; 127 } 128 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 129 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 130 _calls++; 131 unsigned char* ujb = (unsigned char*)card_ptr; 132 int ind = (int)(*ujb); 133 _histo[ind]++; 134 *card_ptr = -1; 135 } 136 return true; 137 } 138 int calls() { return _calls; } 139 void print_histo() { 140 gclog_or_tty->print_cr("Card table value histogram:"); 141 for (int i = 0; i < 256; i++) { 142 if (_histo[i] != 0) { 143 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 144 } 145 } 146 } 147 }; 148 149 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 150 int _calls; 151 G1CollectedHeap* _g1h; 152 CardTableModRefBS* _ctbs; 153 public: 154 RedirtyLoggedCardTableEntryClosure() : 155 _calls(0) 156 { 157 _g1h = G1CollectedHeap::heap(); 158 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 159 } 160 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 161 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 162 _calls++; 163 *card_ptr = 0; 164 } 165 return true; 166 } 167 int calls() { return _calls; } 168 }; 169 170 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 171 public: 172 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 173 *card_ptr = CardTableModRefBS::dirty_card_val(); 174 return true; 175 } 176 }; 177 178 YoungList::YoungList(G1CollectedHeap* g1h) : 179 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 180 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 181 guarantee(check_list_empty(false), "just making sure..."); 182 } 183 184 void YoungList::push_region(HeapRegion *hr) { 185 assert(!hr->is_young(), "should not already be young"); 186 assert(hr->get_next_young_region() == NULL, "cause it should!"); 187 188 hr->set_next_young_region(_head); 189 _head = hr; 190 191 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 192 ++_length; 193 } 194 195 void YoungList::add_survivor_region(HeapRegion* hr) { 196 assert(hr->is_survivor(), "should be flagged as survivor region"); 197 assert(hr->get_next_young_region() == NULL, "cause it should!"); 198 199 hr->set_next_young_region(_survivor_head); 200 if (_survivor_head == NULL) { 201 _survivor_tail = hr; 202 } 203 _survivor_head = hr; 204 ++_survivor_length; 205 } 206 207 void YoungList::empty_list(HeapRegion* list) { 208 while (list != NULL) { 209 HeapRegion* next = list->get_next_young_region(); 210 list->set_next_young_region(NULL); 211 list->uninstall_surv_rate_group(); 212 list->set_not_young(); 213 list = next; 214 } 215 } 216 217 void YoungList::empty_list() { 218 assert(check_list_well_formed(), "young list should be well formed"); 219 220 empty_list(_head); 221 _head = NULL; 222 _length = 0; 223 224 empty_list(_survivor_head); 225 _survivor_head = NULL; 226 _survivor_tail = NULL; 227 _survivor_length = 0; 228 229 _last_sampled_rs_lengths = 0; 230 231 assert(check_list_empty(false), "just making sure..."); 232 } 233 234 bool YoungList::check_list_well_formed() { 235 bool ret = true; 236 237 uint length = 0; 238 HeapRegion* curr = _head; 239 HeapRegion* last = NULL; 240 while (curr != NULL) { 241 if (!curr->is_young()) { 242 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 243 "incorrectly tagged (y: %d, surv: %d)", 244 curr->bottom(), curr->end(), 245 curr->is_young(), curr->is_survivor()); 246 ret = false; 247 } 248 ++length; 249 last = curr; 250 curr = curr->get_next_young_region(); 251 } 252 ret = ret && (length == _length); 253 254 if (!ret) { 255 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 256 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 257 length, _length); 258 } 259 260 return ret; 261 } 262 263 bool YoungList::check_list_empty(bool check_sample) { 264 bool ret = true; 265 266 if (_length != 0) { 267 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 268 _length); 269 ret = false; 270 } 271 if (check_sample && _last_sampled_rs_lengths != 0) { 272 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 273 ret = false; 274 } 275 if (_head != NULL) { 276 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 277 ret = false; 278 } 279 if (!ret) { 280 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 281 } 282 283 return ret; 284 } 285 286 void 287 YoungList::rs_length_sampling_init() { 288 _sampled_rs_lengths = 0; 289 _curr = _head; 290 } 291 292 bool 293 YoungList::rs_length_sampling_more() { 294 return _curr != NULL; 295 } 296 297 void 298 YoungList::rs_length_sampling_next() { 299 assert( _curr != NULL, "invariant" ); 300 size_t rs_length = _curr->rem_set()->occupied(); 301 302 _sampled_rs_lengths += rs_length; 303 304 // The current region may not yet have been added to the 305 // incremental collection set (it gets added when it is 306 // retired as the current allocation region). 307 if (_curr->in_collection_set()) { 308 // Update the collection set policy information for this region 309 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 310 } 311 312 _curr = _curr->get_next_young_region(); 313 if (_curr == NULL) { 314 _last_sampled_rs_lengths = _sampled_rs_lengths; 315 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 316 } 317 } 318 319 void 320 YoungList::reset_auxilary_lists() { 321 guarantee( is_empty(), "young list should be empty" ); 322 assert(check_list_well_formed(), "young list should be well formed"); 323 324 // Add survivor regions to SurvRateGroup. 325 _g1h->g1_policy()->note_start_adding_survivor_regions(); 326 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 327 328 int young_index_in_cset = 0; 329 for (HeapRegion* curr = _survivor_head; 330 curr != NULL; 331 curr = curr->get_next_young_region()) { 332 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 333 334 // The region is a non-empty survivor so let's add it to 335 // the incremental collection set for the next evacuation 336 // pause. 337 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 338 young_index_in_cset += 1; 339 } 340 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 341 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 342 343 _head = _survivor_head; 344 _length = _survivor_length; 345 if (_survivor_head != NULL) { 346 assert(_survivor_tail != NULL, "cause it shouldn't be"); 347 assert(_survivor_length > 0, "invariant"); 348 _survivor_tail->set_next_young_region(NULL); 349 } 350 351 // Don't clear the survivor list handles until the start of 352 // the next evacuation pause - we need it in order to re-tag 353 // the survivor regions from this evacuation pause as 'young' 354 // at the start of the next. 355 356 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 357 358 assert(check_list_well_formed(), "young list should be well formed"); 359 } 360 361 void YoungList::print() { 362 HeapRegion* lists[] = {_head, _survivor_head}; 363 const char* names[] = {"YOUNG", "SURVIVOR"}; 364 365 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 366 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 367 HeapRegion *curr = lists[list]; 368 if (curr == NULL) 369 gclog_or_tty->print_cr(" empty"); 370 while (curr != NULL) { 371 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, " 372 "age: %4d, y: %d, surv: %d", 373 curr->bottom(), curr->end(), 374 curr->top(), 375 curr->prev_top_at_mark_start(), 376 curr->next_top_at_mark_start(), 377 curr->top_at_conc_mark_count(), 378 curr->age_in_surv_rate_group_cond(), 379 curr->is_young(), 380 curr->is_survivor()); 381 curr = curr->get_next_young_region(); 382 } 383 } 384 385 gclog_or_tty->print_cr(""); 386 } 387 388 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 389 { 390 // Claim the right to put the region on the dirty cards region list 391 // by installing a self pointer. 392 HeapRegion* next = hr->get_next_dirty_cards_region(); 393 if (next == NULL) { 394 HeapRegion* res = (HeapRegion*) 395 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 396 NULL); 397 if (res == NULL) { 398 HeapRegion* head; 399 do { 400 // Put the region to the dirty cards region list. 401 head = _dirty_cards_region_list; 402 next = (HeapRegion*) 403 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 404 if (next == head) { 405 assert(hr->get_next_dirty_cards_region() == hr, 406 "hr->get_next_dirty_cards_region() != hr"); 407 if (next == NULL) { 408 // The last region in the list points to itself. 409 hr->set_next_dirty_cards_region(hr); 410 } else { 411 hr->set_next_dirty_cards_region(next); 412 } 413 } 414 } while (next != head); 415 } 416 } 417 } 418 419 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 420 { 421 HeapRegion* head; 422 HeapRegion* hr; 423 do { 424 head = _dirty_cards_region_list; 425 if (head == NULL) { 426 return NULL; 427 } 428 HeapRegion* new_head = head->get_next_dirty_cards_region(); 429 if (head == new_head) { 430 // The last region. 431 new_head = NULL; 432 } 433 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 434 head); 435 } while (hr != head); 436 assert(hr != NULL, "invariant"); 437 hr->set_next_dirty_cards_region(NULL); 438 return hr; 439 } 440 441 void G1CollectedHeap::stop_conc_gc_threads() { 442 _cg1r->stop(); 443 _cmThread->stop(); 444 } 445 446 #ifdef ASSERT 447 // A region is added to the collection set as it is retired 448 // so an address p can point to a region which will be in the 449 // collection set but has not yet been retired. This method 450 // therefore is only accurate during a GC pause after all 451 // regions have been retired. It is used for debugging 452 // to check if an nmethod has references to objects that can 453 // be move during a partial collection. Though it can be 454 // inaccurate, it is sufficient for G1 because the conservative 455 // implementation of is_scavengable() for G1 will indicate that 456 // all nmethods must be scanned during a partial collection. 457 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 458 HeapRegion* hr = heap_region_containing(p); 459 return hr != NULL && hr->in_collection_set(); 460 } 461 #endif 462 463 // Returns true if the reference points to an object that 464 // can move in an incremental collecction. 465 bool G1CollectedHeap::is_scavengable(const void* p) { 466 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 467 G1CollectorPolicy* g1p = g1h->g1_policy(); 468 HeapRegion* hr = heap_region_containing(p); 469 if (hr == NULL) { 470 // perm gen (or null) 471 return false; 472 } else { 473 return !hr->isHumongous(); 474 } 475 } 476 477 void G1CollectedHeap::check_ct_logs_at_safepoint() { 478 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 479 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 480 481 // Count the dirty cards at the start. 482 CountNonCleanMemRegionClosure count1(this); 483 ct_bs->mod_card_iterate(&count1); 484 int orig_count = count1.n(); 485 486 // First clear the logged cards. 487 ClearLoggedCardTableEntryClosure clear; 488 dcqs.set_closure(&clear); 489 dcqs.apply_closure_to_all_completed_buffers(); 490 dcqs.iterate_closure_all_threads(false); 491 clear.print_histo(); 492 493 // Now ensure that there's no dirty cards. 494 CountNonCleanMemRegionClosure count2(this); 495 ct_bs->mod_card_iterate(&count2); 496 if (count2.n() != 0) { 497 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 498 count2.n(), orig_count); 499 } 500 guarantee(count2.n() == 0, "Card table should be clean."); 501 502 RedirtyLoggedCardTableEntryClosure redirty; 503 JavaThread::dirty_card_queue_set().set_closure(&redirty); 504 dcqs.apply_closure_to_all_completed_buffers(); 505 dcqs.iterate_closure_all_threads(false); 506 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 507 clear.calls(), orig_count); 508 guarantee(redirty.calls() == clear.calls(), 509 "Or else mechanism is broken."); 510 511 CountNonCleanMemRegionClosure count3(this); 512 ct_bs->mod_card_iterate(&count3); 513 if (count3.n() != orig_count) { 514 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 515 orig_count, count3.n()); 516 guarantee(count3.n() >= orig_count, "Should have restored them all."); 517 } 518 519 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 520 } 521 522 // Private class members. 523 524 G1CollectedHeap* G1CollectedHeap::_g1h; 525 526 // Private methods. 527 528 HeapRegion* 529 G1CollectedHeap::new_region_try_secondary_free_list() { 530 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 531 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 532 if (!_secondary_free_list.is_empty()) { 533 if (G1ConcRegionFreeingVerbose) { 534 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 535 "secondary_free_list has %u entries", 536 _secondary_free_list.length()); 537 } 538 // It looks as if there are free regions available on the 539 // secondary_free_list. Let's move them to the free_list and try 540 // again to allocate from it. 541 append_secondary_free_list(); 542 543 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 544 "empty we should have moved at least one entry to the free_list"); 545 HeapRegion* res = _free_list.remove_head(); 546 if (G1ConcRegionFreeingVerbose) { 547 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 548 "allocated "HR_FORMAT" from secondary_free_list", 549 HR_FORMAT_PARAMS(res)); 550 } 551 return res; 552 } 553 554 // Wait here until we get notifed either when (a) there are no 555 // more free regions coming or (b) some regions have been moved on 556 // the secondary_free_list. 557 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 558 } 559 560 if (G1ConcRegionFreeingVerbose) { 561 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 562 "could not allocate from secondary_free_list"); 563 } 564 return NULL; 565 } 566 567 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 568 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 569 "the only time we use this to allocate a humongous region is " 570 "when we are allocating a single humongous region"); 571 572 HeapRegion* res; 573 if (G1StressConcRegionFreeing) { 574 if (!_secondary_free_list.is_empty()) { 575 if (G1ConcRegionFreeingVerbose) { 576 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 577 "forced to look at the secondary_free_list"); 578 } 579 res = new_region_try_secondary_free_list(); 580 if (res != NULL) { 581 return res; 582 } 583 } 584 } 585 res = _free_list.remove_head_or_null(); 586 if (res == NULL) { 587 if (G1ConcRegionFreeingVerbose) { 588 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 589 "res == NULL, trying the secondary_free_list"); 590 } 591 res = new_region_try_secondary_free_list(); 592 } 593 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 594 // Currently, only attempts to allocate GC alloc regions set 595 // do_expand to true. So, we should only reach here during a 596 // safepoint. If this assumption changes we might have to 597 // reconsider the use of _expand_heap_after_alloc_failure. 598 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 599 600 ergo_verbose1(ErgoHeapSizing, 601 "attempt heap expansion", 602 ergo_format_reason("region allocation request failed") 603 ergo_format_byte("allocation request"), 604 word_size * HeapWordSize); 605 if (expand(word_size * HeapWordSize)) { 606 // Given that expand() succeeded in expanding the heap, and we 607 // always expand the heap by an amount aligned to the heap 608 // region size, the free list should in theory not be empty. So 609 // it would probably be OK to use remove_head(). But the extra 610 // check for NULL is unlikely to be a performance issue here (we 611 // just expanded the heap!) so let's just be conservative and 612 // use remove_head_or_null(). 613 res = _free_list.remove_head_or_null(); 614 } else { 615 _expand_heap_after_alloc_failure = false; 616 } 617 } 618 return res; 619 } 620 621 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 622 size_t word_size) { 623 assert(isHumongous(word_size), "word_size should be humongous"); 624 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 625 626 uint first = G1_NULL_HRS_INDEX; 627 if (num_regions == 1) { 628 // Only one region to allocate, no need to go through the slower 629 // path. The caller will attempt the expasion if this fails, so 630 // let's not try to expand here too. 631 HeapRegion* hr = new_region(word_size, false /* do_expand */); 632 if (hr != NULL) { 633 first = hr->hrs_index(); 634 } else { 635 first = G1_NULL_HRS_INDEX; 636 } 637 } else { 638 // We can't allocate humongous regions while cleanupComplete() is 639 // running, since some of the regions we find to be empty might not 640 // yet be added to the free list and it is not straightforward to 641 // know which list they are on so that we can remove them. Note 642 // that we only need to do this if we need to allocate more than 643 // one region to satisfy the current humongous allocation 644 // request. If we are only allocating one region we use the common 645 // region allocation code (see above). 646 wait_while_free_regions_coming(); 647 append_secondary_free_list_if_not_empty_with_lock(); 648 649 if (free_regions() >= num_regions) { 650 first = _hrs.find_contiguous(num_regions); 651 if (first != G1_NULL_HRS_INDEX) { 652 for (uint i = first; i < first + num_regions; ++i) { 653 HeapRegion* hr = region_at(i); 654 assert(hr->is_empty(), "sanity"); 655 assert(is_on_master_free_list(hr), "sanity"); 656 hr->set_pending_removal(true); 657 } 658 _free_list.remove_all_pending(num_regions); 659 } 660 } 661 } 662 return first; 663 } 664 665 HeapWord* 666 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 667 uint num_regions, 668 size_t word_size) { 669 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 670 assert(isHumongous(word_size), "word_size should be humongous"); 671 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 672 673 // Index of last region in the series + 1. 674 uint last = first + num_regions; 675 676 // We need to initialize the region(s) we just discovered. This is 677 // a bit tricky given that it can happen concurrently with 678 // refinement threads refining cards on these regions and 679 // potentially wanting to refine the BOT as they are scanning 680 // those cards (this can happen shortly after a cleanup; see CR 681 // 6991377). So we have to set up the region(s) carefully and in 682 // a specific order. 683 684 // The word size sum of all the regions we will allocate. 685 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 686 assert(word_size <= word_size_sum, "sanity"); 687 688 // This will be the "starts humongous" region. 689 HeapRegion* first_hr = region_at(first); 690 // The header of the new object will be placed at the bottom of 691 // the first region. 692 HeapWord* new_obj = first_hr->bottom(); 693 // This will be the new end of the first region in the series that 694 // should also match the end of the last region in the seriers. 695 HeapWord* new_end = new_obj + word_size_sum; 696 // This will be the new top of the first region that will reflect 697 // this allocation. 698 HeapWord* new_top = new_obj + word_size; 699 700 // First, we need to zero the header of the space that we will be 701 // allocating. When we update top further down, some refinement 702 // threads might try to scan the region. By zeroing the header we 703 // ensure that any thread that will try to scan the region will 704 // come across the zero klass word and bail out. 705 // 706 // NOTE: It would not have been correct to have used 707 // CollectedHeap::fill_with_object() and make the space look like 708 // an int array. The thread that is doing the allocation will 709 // later update the object header to a potentially different array 710 // type and, for a very short period of time, the klass and length 711 // fields will be inconsistent. This could cause a refinement 712 // thread to calculate the object size incorrectly. 713 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 714 715 // We will set up the first region as "starts humongous". This 716 // will also update the BOT covering all the regions to reflect 717 // that there is a single object that starts at the bottom of the 718 // first region. 719 first_hr->set_startsHumongous(new_top, new_end); 720 721 // Then, if there are any, we will set up the "continues 722 // humongous" regions. 723 HeapRegion* hr = NULL; 724 for (uint i = first + 1; i < last; ++i) { 725 hr = region_at(i); 726 hr->set_continuesHumongous(first_hr); 727 } 728 // If we have "continues humongous" regions (hr != NULL), then the 729 // end of the last one should match new_end. 730 assert(hr == NULL || hr->end() == new_end, "sanity"); 731 732 // Up to this point no concurrent thread would have been able to 733 // do any scanning on any region in this series. All the top 734 // fields still point to bottom, so the intersection between 735 // [bottom,top] and [card_start,card_end] will be empty. Before we 736 // update the top fields, we'll do a storestore to make sure that 737 // no thread sees the update to top before the zeroing of the 738 // object header and the BOT initialization. 739 OrderAccess::storestore(); 740 741 // Now that the BOT and the object header have been initialized, 742 // we can update top of the "starts humongous" region. 743 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 744 "new_top should be in this region"); 745 first_hr->set_top(new_top); 746 if (_hr_printer.is_active()) { 747 HeapWord* bottom = first_hr->bottom(); 748 HeapWord* end = first_hr->orig_end(); 749 if ((first + 1) == last) { 750 // the series has a single humongous region 751 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 752 } else { 753 // the series has more than one humongous regions 754 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 755 } 756 } 757 758 // Now, we will update the top fields of the "continues humongous" 759 // regions. The reason we need to do this is that, otherwise, 760 // these regions would look empty and this will confuse parts of 761 // G1. For example, the code that looks for a consecutive number 762 // of empty regions will consider them empty and try to 763 // re-allocate them. We can extend is_empty() to also include 764 // !continuesHumongous(), but it is easier to just update the top 765 // fields here. The way we set top for all regions (i.e., top == 766 // end for all regions but the last one, top == new_top for the 767 // last one) is actually used when we will free up the humongous 768 // region in free_humongous_region(). 769 hr = NULL; 770 for (uint i = first + 1; i < last; ++i) { 771 hr = region_at(i); 772 if ((i + 1) == last) { 773 // last continues humongous region 774 assert(hr->bottom() < new_top && new_top <= hr->end(), 775 "new_top should fall on this region"); 776 hr->set_top(new_top); 777 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 778 } else { 779 // not last one 780 assert(new_top > hr->end(), "new_top should be above this region"); 781 hr->set_top(hr->end()); 782 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 783 } 784 } 785 // If we have continues humongous regions (hr != NULL), then the 786 // end of the last one should match new_end and its top should 787 // match new_top. 788 assert(hr == NULL || 789 (hr->end() == new_end && hr->top() == new_top), "sanity"); 790 791 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 792 _summary_bytes_used += first_hr->used(); 793 _humongous_set.add(first_hr); 794 795 return new_obj; 796 } 797 798 // If could fit into free regions w/o expansion, try. 799 // Otherwise, if can expand, do so. 800 // Otherwise, if using ex regions might help, try with ex given back. 801 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 802 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 803 804 verify_region_sets_optional(); 805 806 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 807 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 808 uint x_num = expansion_regions(); 809 uint fs = _hrs.free_suffix(); 810 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 811 if (first == G1_NULL_HRS_INDEX) { 812 // The only thing we can do now is attempt expansion. 813 if (fs + x_num >= num_regions) { 814 // If the number of regions we're trying to allocate for this 815 // object is at most the number of regions in the free suffix, 816 // then the call to humongous_obj_allocate_find_first() above 817 // should have succeeded and we wouldn't be here. 818 // 819 // We should only be trying to expand when the free suffix is 820 // not sufficient for the object _and_ we have some expansion 821 // room available. 822 assert(num_regions > fs, "earlier allocation should have succeeded"); 823 824 ergo_verbose1(ErgoHeapSizing, 825 "attempt heap expansion", 826 ergo_format_reason("humongous allocation request failed") 827 ergo_format_byte("allocation request"), 828 word_size * HeapWordSize); 829 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 830 // Even though the heap was expanded, it might not have 831 // reached the desired size. So, we cannot assume that the 832 // allocation will succeed. 833 first = humongous_obj_allocate_find_first(num_regions, word_size); 834 } 835 } 836 } 837 838 HeapWord* result = NULL; 839 if (first != G1_NULL_HRS_INDEX) { 840 result = 841 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 842 assert(result != NULL, "it should always return a valid result"); 843 844 // A successful humongous object allocation changes the used space 845 // information of the old generation so we need to recalculate the 846 // sizes and update the jstat counters here. 847 g1mm()->update_sizes(); 848 } 849 850 verify_region_sets_optional(); 851 852 return result; 853 } 854 855 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 856 assert_heap_not_locked_and_not_at_safepoint(); 857 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 858 859 unsigned int dummy_gc_count_before; 860 return attempt_allocation(word_size, &dummy_gc_count_before); 861 } 862 863 HeapWord* 864 G1CollectedHeap::mem_allocate(size_t word_size, 865 bool* gc_overhead_limit_was_exceeded) { 866 assert_heap_not_locked_and_not_at_safepoint(); 867 868 // Loop until the allocation is satisified, or unsatisfied after GC. 869 for (int try_count = 1; /* we'll return */; try_count += 1) { 870 unsigned int gc_count_before; 871 872 HeapWord* result = NULL; 873 if (!isHumongous(word_size)) { 874 result = attempt_allocation(word_size, &gc_count_before); 875 } else { 876 result = attempt_allocation_humongous(word_size, &gc_count_before); 877 } 878 if (result != NULL) { 879 return result; 880 } 881 882 // Create the garbage collection operation... 883 VM_G1CollectForAllocation op(gc_count_before, word_size); 884 // ...and get the VM thread to execute it. 885 VMThread::execute(&op); 886 887 if (op.prologue_succeeded() && op.pause_succeeded()) { 888 // If the operation was successful we'll return the result even 889 // if it is NULL. If the allocation attempt failed immediately 890 // after a Full GC, it's unlikely we'll be able to allocate now. 891 HeapWord* result = op.result(); 892 if (result != NULL && !isHumongous(word_size)) { 893 // Allocations that take place on VM operations do not do any 894 // card dirtying and we have to do it here. We only have to do 895 // this for non-humongous allocations, though. 896 dirty_young_block(result, word_size); 897 } 898 return result; 899 } else { 900 assert(op.result() == NULL, 901 "the result should be NULL if the VM op did not succeed"); 902 } 903 904 // Give a warning if we seem to be looping forever. 905 if ((QueuedAllocationWarningCount > 0) && 906 (try_count % QueuedAllocationWarningCount == 0)) { 907 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 908 } 909 } 910 911 ShouldNotReachHere(); 912 return NULL; 913 } 914 915 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 916 unsigned int *gc_count_before_ret) { 917 // Make sure you read the note in attempt_allocation_humongous(). 918 919 assert_heap_not_locked_and_not_at_safepoint(); 920 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 921 "be called for humongous allocation requests"); 922 923 // We should only get here after the first-level allocation attempt 924 // (attempt_allocation()) failed to allocate. 925 926 // We will loop until a) we manage to successfully perform the 927 // allocation or b) we successfully schedule a collection which 928 // fails to perform the allocation. b) is the only case when we'll 929 // return NULL. 930 HeapWord* result = NULL; 931 for (int try_count = 1; /* we'll return */; try_count += 1) { 932 bool should_try_gc; 933 unsigned int gc_count_before; 934 935 { 936 MutexLockerEx x(Heap_lock); 937 938 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 939 false /* bot_updates */); 940 if (result != NULL) { 941 return result; 942 } 943 944 // If we reach here, attempt_allocation_locked() above failed to 945 // allocate a new region. So the mutator alloc region should be NULL. 946 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 947 948 if (GC_locker::is_active_and_needs_gc()) { 949 if (g1_policy()->can_expand_young_list()) { 950 // No need for an ergo verbose message here, 951 // can_expand_young_list() does this when it returns true. 952 result = _mutator_alloc_region.attempt_allocation_force(word_size, 953 false /* bot_updates */); 954 if (result != NULL) { 955 return result; 956 } 957 } 958 should_try_gc = false; 959 } else { 960 // Read the GC count while still holding the Heap_lock. 961 gc_count_before = total_collections(); 962 should_try_gc = true; 963 } 964 } 965 966 if (should_try_gc) { 967 bool succeeded; 968 result = do_collection_pause(word_size, gc_count_before, &succeeded); 969 if (result != NULL) { 970 assert(succeeded, "only way to get back a non-NULL result"); 971 return result; 972 } 973 974 if (succeeded) { 975 // If we get here we successfully scheduled a collection which 976 // failed to allocate. No point in trying to allocate 977 // further. We'll just return NULL. 978 MutexLockerEx x(Heap_lock); 979 *gc_count_before_ret = total_collections(); 980 return NULL; 981 } 982 } else { 983 GC_locker::stall_until_clear(); 984 } 985 986 // We can reach here if we were unsuccessul in scheduling a 987 // collection (because another thread beat us to it) or if we were 988 // stalled due to the GC locker. In either can we should retry the 989 // allocation attempt in case another thread successfully 990 // performed a collection and reclaimed enough space. We do the 991 // first attempt (without holding the Heap_lock) here and the 992 // follow-on attempt will be at the start of the next loop 993 // iteration (after taking the Heap_lock). 994 result = _mutator_alloc_region.attempt_allocation(word_size, 995 false /* bot_updates */); 996 if (result != NULL) { 997 return result; 998 } 999 1000 // Give a warning if we seem to be looping forever. 1001 if ((QueuedAllocationWarningCount > 0) && 1002 (try_count % QueuedAllocationWarningCount == 0)) { 1003 warning("G1CollectedHeap::attempt_allocation_slow() " 1004 "retries %d times", try_count); 1005 } 1006 } 1007 1008 ShouldNotReachHere(); 1009 return NULL; 1010 } 1011 1012 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1013 unsigned int * gc_count_before_ret) { 1014 // The structure of this method has a lot of similarities to 1015 // attempt_allocation_slow(). The reason these two were not merged 1016 // into a single one is that such a method would require several "if 1017 // allocation is not humongous do this, otherwise do that" 1018 // conditional paths which would obscure its flow. In fact, an early 1019 // version of this code did use a unified method which was harder to 1020 // follow and, as a result, it had subtle bugs that were hard to 1021 // track down. So keeping these two methods separate allows each to 1022 // be more readable. It will be good to keep these two in sync as 1023 // much as possible. 1024 1025 assert_heap_not_locked_and_not_at_safepoint(); 1026 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1027 "should only be called for humongous allocations"); 1028 1029 // Humongous objects can exhaust the heap quickly, so we should check if we 1030 // need to start a marking cycle at each humongous object allocation. We do 1031 // the check before we do the actual allocation. The reason for doing it 1032 // before the allocation is that we avoid having to keep track of the newly 1033 // allocated memory while we do a GC. 1034 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1035 word_size)) { 1036 collect(GCCause::_g1_humongous_allocation); 1037 } 1038 1039 // We will loop until a) we manage to successfully perform the 1040 // allocation or b) we successfully schedule a collection which 1041 // fails to perform the allocation. b) is the only case when we'll 1042 // return NULL. 1043 HeapWord* result = NULL; 1044 for (int try_count = 1; /* we'll return */; try_count += 1) { 1045 bool should_try_gc; 1046 unsigned int gc_count_before; 1047 1048 { 1049 MutexLockerEx x(Heap_lock); 1050 1051 // Given that humongous objects are not allocated in young 1052 // regions, we'll first try to do the allocation without doing a 1053 // collection hoping that there's enough space in the heap. 1054 result = humongous_obj_allocate(word_size); 1055 if (result != NULL) { 1056 return result; 1057 } 1058 1059 if (GC_locker::is_active_and_needs_gc()) { 1060 should_try_gc = false; 1061 } else { 1062 // Read the GC count while still holding the Heap_lock. 1063 gc_count_before = total_collections(); 1064 should_try_gc = true; 1065 } 1066 } 1067 1068 if (should_try_gc) { 1069 // If we failed to allocate the humongous object, we should try to 1070 // do a collection pause (if we're allowed) in case it reclaims 1071 // enough space for the allocation to succeed after the pause. 1072 1073 bool succeeded; 1074 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1075 if (result != NULL) { 1076 assert(succeeded, "only way to get back a non-NULL result"); 1077 return result; 1078 } 1079 1080 if (succeeded) { 1081 // If we get here we successfully scheduled a collection which 1082 // failed to allocate. No point in trying to allocate 1083 // further. We'll just return NULL. 1084 MutexLockerEx x(Heap_lock); 1085 *gc_count_before_ret = total_collections(); 1086 return NULL; 1087 } 1088 } else { 1089 GC_locker::stall_until_clear(); 1090 } 1091 1092 // We can reach here if we were unsuccessul in scheduling a 1093 // collection (because another thread beat us to it) or if we were 1094 // stalled due to the GC locker. In either can we should retry the 1095 // allocation attempt in case another thread successfully 1096 // performed a collection and reclaimed enough space. Give a 1097 // warning if we seem to be looping forever. 1098 1099 if ((QueuedAllocationWarningCount > 0) && 1100 (try_count % QueuedAllocationWarningCount == 0)) { 1101 warning("G1CollectedHeap::attempt_allocation_humongous() " 1102 "retries %d times", try_count); 1103 } 1104 } 1105 1106 ShouldNotReachHere(); 1107 return NULL; 1108 } 1109 1110 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1111 bool expect_null_mutator_alloc_region) { 1112 assert_at_safepoint(true /* should_be_vm_thread */); 1113 assert(_mutator_alloc_region.get() == NULL || 1114 !expect_null_mutator_alloc_region, 1115 "the current alloc region was unexpectedly found to be non-NULL"); 1116 1117 if (!isHumongous(word_size)) { 1118 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1119 false /* bot_updates */); 1120 } else { 1121 HeapWord* result = humongous_obj_allocate(word_size); 1122 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1123 g1_policy()->set_initiate_conc_mark_if_possible(); 1124 } 1125 return result; 1126 } 1127 1128 ShouldNotReachHere(); 1129 } 1130 1131 class PostMCRemSetClearClosure: public HeapRegionClosure { 1132 ModRefBarrierSet* _mr_bs; 1133 public: 1134 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1135 bool doHeapRegion(HeapRegion* r) { 1136 r->reset_gc_time_stamp(); 1137 if (r->continuesHumongous()) 1138 return false; 1139 HeapRegionRemSet* hrrs = r->rem_set(); 1140 if (hrrs != NULL) hrrs->clear(); 1141 // You might think here that we could clear just the cards 1142 // corresponding to the used region. But no: if we leave a dirty card 1143 // in a region we might allocate into, then it would prevent that card 1144 // from being enqueued, and cause it to be missed. 1145 // Re: the performance cost: we shouldn't be doing full GC anyway! 1146 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1147 return false; 1148 } 1149 }; 1150 1151 1152 class PostMCRemSetInvalidateClosure: public HeapRegionClosure { 1153 ModRefBarrierSet* _mr_bs; 1154 public: 1155 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1156 bool doHeapRegion(HeapRegion* r) { 1157 if (r->continuesHumongous()) return false; 1158 if (r->used_region().word_size() != 0) { 1159 _mr_bs->invalidate(r->used_region(), true /*whole heap*/); 1160 } 1161 return false; 1162 } 1163 }; 1164 1165 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1166 G1CollectedHeap* _g1h; 1167 UpdateRSOopClosure _cl; 1168 int _worker_i; 1169 public: 1170 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1171 _cl(g1->g1_rem_set(), worker_i), 1172 _worker_i(worker_i), 1173 _g1h(g1) 1174 { } 1175 1176 bool doHeapRegion(HeapRegion* r) { 1177 if (!r->continuesHumongous()) { 1178 _cl.set_from(r); 1179 r->oop_iterate(&_cl); 1180 } 1181 return false; 1182 } 1183 }; 1184 1185 class ParRebuildRSTask: public AbstractGangTask { 1186 G1CollectedHeap* _g1; 1187 public: 1188 ParRebuildRSTask(G1CollectedHeap* g1) 1189 : AbstractGangTask("ParRebuildRSTask"), 1190 _g1(g1) 1191 { } 1192 1193 void work(uint worker_id) { 1194 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1195 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1196 _g1->workers()->active_workers(), 1197 HeapRegion::RebuildRSClaimValue); 1198 } 1199 }; 1200 1201 class PostCompactionPrinterClosure: public HeapRegionClosure { 1202 private: 1203 G1HRPrinter* _hr_printer; 1204 public: 1205 bool doHeapRegion(HeapRegion* hr) { 1206 assert(!hr->is_young(), "not expecting to find young regions"); 1207 // We only generate output for non-empty regions. 1208 if (!hr->is_empty()) { 1209 if (!hr->isHumongous()) { 1210 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1211 } else if (hr->startsHumongous()) { 1212 if (hr->capacity() == HeapRegion::GrainBytes) { 1213 // single humongous region 1214 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1215 } else { 1216 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1217 } 1218 } else { 1219 assert(hr->continuesHumongous(), "only way to get here"); 1220 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1221 } 1222 } 1223 return false; 1224 } 1225 1226 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1227 : _hr_printer(hr_printer) { } 1228 }; 1229 1230 bool G1CollectedHeap::do_collection(bool explicit_gc, 1231 bool clear_all_soft_refs, 1232 size_t word_size) { 1233 assert_at_safepoint(true /* should_be_vm_thread */); 1234 1235 if (GC_locker::check_active_before_gc()) { 1236 return false; 1237 } 1238 1239 SvcGCMarker sgcm(SvcGCMarker::FULL); 1240 ResourceMark rm; 1241 1242 print_heap_before_gc(); 1243 1244 HRSPhaseSetter x(HRSPhaseFullGC); 1245 verify_region_sets_optional(); 1246 1247 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1248 collector_policy()->should_clear_all_soft_refs(); 1249 1250 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1251 1252 { 1253 IsGCActiveMark x; 1254 1255 // Timing 1256 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc); 1257 assert(!system_gc || explicit_gc, "invariant"); 1258 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1259 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1260 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC", 1261 G1Log::fine(), true, gclog_or_tty); 1262 1263 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1264 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1265 1266 double start = os::elapsedTime(); 1267 g1_policy()->record_full_collection_start(); 1268 1269 // Note: When we have a more flexible GC logging framework that 1270 // allows us to add optional attributes to a GC log record we 1271 // could consider timing and reporting how long we wait in the 1272 // following two methods. 1273 wait_while_free_regions_coming(); 1274 // If we start the compaction before the CM threads finish 1275 // scanning the root regions we might trip them over as we'll 1276 // be moving objects / updating references. So let's wait until 1277 // they are done. By telling them to abort, they should complete 1278 // early. 1279 _cm->root_regions()->abort(); 1280 _cm->root_regions()->wait_until_scan_finished(); 1281 append_secondary_free_list_if_not_empty_with_lock(); 1282 1283 gc_prologue(true); 1284 increment_total_collections(true /* full gc */); 1285 1286 size_t g1h_prev_used = used(); 1287 assert(used() == recalculate_used(), "Should be equal"); 1288 1289 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 1290 HandleMark hm; // Discard invalid handles created during verification 1291 gclog_or_tty->print(" VerifyBeforeGC:"); 1292 prepare_for_verify(); 1293 Universe::verify(/* silent */ false, 1294 /* option */ VerifyOption_G1UsePrevMarking); 1295 1296 } 1297 pre_full_gc_dump(); 1298 1299 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1300 1301 // Disable discovery and empty the discovered lists 1302 // for the CM ref processor. 1303 ref_processor_cm()->disable_discovery(); 1304 ref_processor_cm()->abandon_partial_discovery(); 1305 ref_processor_cm()->verify_no_references_recorded(); 1306 1307 // Abandon current iterations of concurrent marking and concurrent 1308 // refinement, if any are in progress. We have to do this before 1309 // wait_until_scan_finished() below. 1310 concurrent_mark()->abort(); 1311 1312 // Make sure we'll choose a new allocation region afterwards. 1313 release_mutator_alloc_region(); 1314 abandon_gc_alloc_regions(); 1315 g1_rem_set()->cleanupHRRS(); 1316 1317 // We should call this after we retire any currently active alloc 1318 // regions so that all the ALLOC / RETIRE events are generated 1319 // before the start GC event. 1320 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1321 1322 // We may have added regions to the current incremental collection 1323 // set between the last GC or pause and now. We need to clear the 1324 // incremental collection set and then start rebuilding it afresh 1325 // after this full GC. 1326 abandon_collection_set(g1_policy()->inc_cset_head()); 1327 g1_policy()->clear_incremental_cset(); 1328 g1_policy()->stop_incremental_cset_building(); 1329 1330 tear_down_region_sets(false /* free_list_only */); 1331 g1_policy()->set_gcs_are_young(true); 1332 1333 // See the comments in g1CollectedHeap.hpp and 1334 // G1CollectedHeap::ref_processing_init() about 1335 // how reference processing currently works in G1. 1336 1337 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1338 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1339 1340 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1341 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1342 1343 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1344 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1345 1346 // Do collection work 1347 { 1348 HandleMark hm; // Discard invalid handles created during gc 1349 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1350 } 1351 1352 assert(free_regions() == 0, "we should not have added any free regions"); 1353 rebuild_region_sets(false /* free_list_only */); 1354 1355 // Enqueue any discovered reference objects that have 1356 // not been removed from the discovered lists. 1357 ref_processor_stw()->enqueue_discovered_references(); 1358 1359 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1360 1361 MemoryService::track_memory_usage(); 1362 1363 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 1364 HandleMark hm; // Discard invalid handles created during verification 1365 gclog_or_tty->print(" VerifyAfterGC:"); 1366 prepare_for_verify(); 1367 Universe::verify(/* silent */ false, 1368 /* option */ VerifyOption_G1UsePrevMarking); 1369 1370 } 1371 1372 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1373 ref_processor_stw()->verify_no_references_recorded(); 1374 1375 // Note: since we've just done a full GC, concurrent 1376 // marking is no longer active. Therefore we need not 1377 // re-enable reference discovery for the CM ref processor. 1378 // That will be done at the start of the next marking cycle. 1379 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1380 ref_processor_cm()->verify_no_references_recorded(); 1381 1382 reset_gc_time_stamp(); 1383 // Since everything potentially moved, we will clear all remembered 1384 // sets, and clear all cards. Later we will rebuild remebered 1385 // sets. We will also reset the GC time stamps of the regions. 1386 PostMCRemSetClearClosure rs_clear(mr_bs()); 1387 heap_region_iterate(&rs_clear); 1388 1389 // Resize the heap if necessary. 1390 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1391 1392 if (_hr_printer.is_active()) { 1393 // We should do this after we potentially resize the heap so 1394 // that all the COMMIT / UNCOMMIT events are generated before 1395 // the end GC event. 1396 1397 PostCompactionPrinterClosure cl(hr_printer()); 1398 heap_region_iterate(&cl); 1399 1400 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1401 } 1402 1403 if (_cg1r->use_cache()) { 1404 _cg1r->clear_and_record_card_counts(); 1405 _cg1r->clear_hot_cache(); 1406 } 1407 1408 // Rebuild remembered sets of all regions. 1409 if (G1CollectedHeap::use_parallel_gc_threads()) { 1410 uint n_workers = 1411 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1412 workers()->active_workers(), 1413 Threads::number_of_non_daemon_threads()); 1414 assert(UseDynamicNumberOfGCThreads || 1415 n_workers == workers()->total_workers(), 1416 "If not dynamic should be using all the workers"); 1417 workers()->set_active_workers(n_workers); 1418 // Set parallel threads in the heap (_n_par_threads) only 1419 // before a parallel phase and always reset it to 0 after 1420 // the phase so that the number of parallel threads does 1421 // no get carried forward to a serial phase where there 1422 // may be code that is "possibly_parallel". 1423 set_par_threads(n_workers); 1424 1425 ParRebuildRSTask rebuild_rs_task(this); 1426 assert(check_heap_region_claim_values( 1427 HeapRegion::InitialClaimValue), "sanity check"); 1428 assert(UseDynamicNumberOfGCThreads || 1429 workers()->active_workers() == workers()->total_workers(), 1430 "Unless dynamic should use total workers"); 1431 // Use the most recent number of active workers 1432 assert(workers()->active_workers() > 0, 1433 "Active workers not properly set"); 1434 set_par_threads(workers()->active_workers()); 1435 workers()->run_task(&rebuild_rs_task); 1436 set_par_threads(0); 1437 assert(check_heap_region_claim_values( 1438 HeapRegion::RebuildRSClaimValue), "sanity check"); 1439 reset_heap_region_claim_values(); 1440 } else { 1441 RebuildRSOutOfRegionClosure rebuild_rs(this); 1442 heap_region_iterate(&rebuild_rs); 1443 } 1444 1445 if (G1Log::fine()) { 1446 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1447 } 1448 1449 if (true) { // FIXME 1450 // Ask the permanent generation to adjust size for full collections 1451 perm()->compute_new_size(); 1452 } 1453 1454 // Start a new incremental collection set for the next pause 1455 assert(g1_policy()->collection_set() == NULL, "must be"); 1456 g1_policy()->start_incremental_cset_building(); 1457 1458 // Clear the _cset_fast_test bitmap in anticipation of adding 1459 // regions to the incremental collection set for the next 1460 // evacuation pause. 1461 clear_cset_fast_test(); 1462 1463 init_mutator_alloc_region(); 1464 1465 double end = os::elapsedTime(); 1466 g1_policy()->record_full_collection_end(); 1467 1468 #ifdef TRACESPINNING 1469 ParallelTaskTerminator::print_termination_counts(); 1470 #endif 1471 1472 gc_epilogue(true); 1473 1474 // Discard all rset updates 1475 JavaThread::dirty_card_queue_set().abandon_logs(); 1476 assert(!G1DeferredRSUpdate 1477 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1478 } 1479 1480 _young_list->reset_sampled_info(); 1481 // At this point there should be no regions in the 1482 // entire heap tagged as young. 1483 assert( check_young_list_empty(true /* check_heap */), 1484 "young list should be empty at this point"); 1485 1486 // Update the number of full collections that have been completed. 1487 increment_full_collections_completed(false /* concurrent */); 1488 1489 _hrs.verify_optional(); 1490 verify_region_sets_optional(); 1491 1492 print_heap_after_gc(); 1493 g1mm()->update_sizes(); 1494 post_full_gc_dump(); 1495 1496 return true; 1497 } 1498 1499 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1500 // do_collection() will return whether it succeeded in performing 1501 // the GC. Currently, there is no facility on the 1502 // do_full_collection() API to notify the caller than the collection 1503 // did not succeed (e.g., because it was locked out by the GC 1504 // locker). So, right now, we'll ignore the return value. 1505 bool dummy = do_collection(true, /* explicit_gc */ 1506 clear_all_soft_refs, 1507 0 /* word_size */); 1508 } 1509 1510 // This code is mostly copied from TenuredGeneration. 1511 void 1512 G1CollectedHeap:: 1513 resize_if_necessary_after_full_collection(size_t word_size) { 1514 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1515 1516 // Include the current allocation, if any, and bytes that will be 1517 // pre-allocated to support collections, as "used". 1518 const size_t used_after_gc = used(); 1519 const size_t capacity_after_gc = capacity(); 1520 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1521 1522 // This is enforced in arguments.cpp. 1523 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1524 "otherwise the code below doesn't make sense"); 1525 1526 // We don't have floating point command-line arguments 1527 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1528 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1529 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1530 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1531 1532 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1533 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1534 1535 // We have to be careful here as these two calculations can overflow 1536 // 32-bit size_t's. 1537 double used_after_gc_d = (double) used_after_gc; 1538 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1539 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1540 1541 // Let's make sure that they are both under the max heap size, which 1542 // by default will make them fit into a size_t. 1543 double desired_capacity_upper_bound = (double) max_heap_size; 1544 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1545 desired_capacity_upper_bound); 1546 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1547 desired_capacity_upper_bound); 1548 1549 // We can now safely turn them into size_t's. 1550 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1551 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1552 1553 // This assert only makes sense here, before we adjust them 1554 // with respect to the min and max heap size. 1555 assert(minimum_desired_capacity <= maximum_desired_capacity, 1556 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1557 "maximum_desired_capacity = "SIZE_FORMAT, 1558 minimum_desired_capacity, maximum_desired_capacity)); 1559 1560 // Should not be greater than the heap max size. No need to adjust 1561 // it with respect to the heap min size as it's a lower bound (i.e., 1562 // we'll try to make the capacity larger than it, not smaller). 1563 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1564 // Should not be less than the heap min size. No need to adjust it 1565 // with respect to the heap max size as it's an upper bound (i.e., 1566 // we'll try to make the capacity smaller than it, not greater). 1567 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1568 1569 if (capacity_after_gc < minimum_desired_capacity) { 1570 // Don't expand unless it's significant 1571 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1572 ergo_verbose4(ErgoHeapSizing, 1573 "attempt heap expansion", 1574 ergo_format_reason("capacity lower than " 1575 "min desired capacity after Full GC") 1576 ergo_format_byte("capacity") 1577 ergo_format_byte("occupancy") 1578 ergo_format_byte_perc("min desired capacity"), 1579 capacity_after_gc, used_after_gc, 1580 minimum_desired_capacity, (double) MinHeapFreeRatio); 1581 expand(expand_bytes); 1582 1583 // No expansion, now see if we want to shrink 1584 } else if (capacity_after_gc > maximum_desired_capacity) { 1585 // Capacity too large, compute shrinking size 1586 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1587 ergo_verbose4(ErgoHeapSizing, 1588 "attempt heap shrinking", 1589 ergo_format_reason("capacity higher than " 1590 "max desired capacity after Full GC") 1591 ergo_format_byte("capacity") 1592 ergo_format_byte("occupancy") 1593 ergo_format_byte_perc("max desired capacity"), 1594 capacity_after_gc, used_after_gc, 1595 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1596 shrink(shrink_bytes); 1597 } 1598 } 1599 1600 1601 HeapWord* 1602 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1603 bool* succeeded) { 1604 assert_at_safepoint(true /* should_be_vm_thread */); 1605 1606 *succeeded = true; 1607 // Let's attempt the allocation first. 1608 HeapWord* result = 1609 attempt_allocation_at_safepoint(word_size, 1610 false /* expect_null_mutator_alloc_region */); 1611 if (result != NULL) { 1612 assert(*succeeded, "sanity"); 1613 return result; 1614 } 1615 1616 // In a G1 heap, we're supposed to keep allocation from failing by 1617 // incremental pauses. Therefore, at least for now, we'll favor 1618 // expansion over collection. (This might change in the future if we can 1619 // do something smarter than full collection to satisfy a failed alloc.) 1620 result = expand_and_allocate(word_size); 1621 if (result != NULL) { 1622 assert(*succeeded, "sanity"); 1623 return result; 1624 } 1625 1626 // Expansion didn't work, we'll try to do a Full GC. 1627 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1628 false, /* clear_all_soft_refs */ 1629 word_size); 1630 if (!gc_succeeded) { 1631 *succeeded = false; 1632 return NULL; 1633 } 1634 1635 // Retry the allocation 1636 result = attempt_allocation_at_safepoint(word_size, 1637 true /* expect_null_mutator_alloc_region */); 1638 if (result != NULL) { 1639 assert(*succeeded, "sanity"); 1640 return result; 1641 } 1642 1643 // Then, try a Full GC that will collect all soft references. 1644 gc_succeeded = do_collection(false, /* explicit_gc */ 1645 true, /* clear_all_soft_refs */ 1646 word_size); 1647 if (!gc_succeeded) { 1648 *succeeded = false; 1649 return NULL; 1650 } 1651 1652 // Retry the allocation once more 1653 result = attempt_allocation_at_safepoint(word_size, 1654 true /* expect_null_mutator_alloc_region */); 1655 if (result != NULL) { 1656 assert(*succeeded, "sanity"); 1657 return result; 1658 } 1659 1660 assert(!collector_policy()->should_clear_all_soft_refs(), 1661 "Flag should have been handled and cleared prior to this point"); 1662 1663 // What else? We might try synchronous finalization later. If the total 1664 // space available is large enough for the allocation, then a more 1665 // complete compaction phase than we've tried so far might be 1666 // appropriate. 1667 assert(*succeeded, "sanity"); 1668 return NULL; 1669 } 1670 1671 // Attempting to expand the heap sufficiently 1672 // to support an allocation of the given "word_size". If 1673 // successful, perform the allocation and return the address of the 1674 // allocated block, or else "NULL". 1675 1676 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1677 assert_at_safepoint(true /* should_be_vm_thread */); 1678 1679 verify_region_sets_optional(); 1680 1681 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1682 ergo_verbose1(ErgoHeapSizing, 1683 "attempt heap expansion", 1684 ergo_format_reason("allocation request failed") 1685 ergo_format_byte("allocation request"), 1686 word_size * HeapWordSize); 1687 if (expand(expand_bytes)) { 1688 _hrs.verify_optional(); 1689 verify_region_sets_optional(); 1690 return attempt_allocation_at_safepoint(word_size, 1691 false /* expect_null_mutator_alloc_region */); 1692 } 1693 return NULL; 1694 } 1695 1696 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1697 HeapWord* new_end) { 1698 assert(old_end != new_end, "don't call this otherwise"); 1699 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1700 1701 // Update the committed mem region. 1702 _g1_committed.set_end(new_end); 1703 // Tell the card table about the update. 1704 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1705 // Tell the BOT about the update. 1706 _bot_shared->resize(_g1_committed.word_size()); 1707 } 1708 1709 bool G1CollectedHeap::expand(size_t expand_bytes) { 1710 size_t old_mem_size = _g1_storage.committed_size(); 1711 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1712 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1713 HeapRegion::GrainBytes); 1714 ergo_verbose2(ErgoHeapSizing, 1715 "expand the heap", 1716 ergo_format_byte("requested expansion amount") 1717 ergo_format_byte("attempted expansion amount"), 1718 expand_bytes, aligned_expand_bytes); 1719 1720 // First commit the memory. 1721 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1722 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1723 if (successful) { 1724 // Then propagate this update to the necessary data structures. 1725 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1726 update_committed_space(old_end, new_end); 1727 1728 FreeRegionList expansion_list("Local Expansion List"); 1729 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1730 assert(mr.start() == old_end, "post-condition"); 1731 // mr might be a smaller region than what was requested if 1732 // expand_by() was unable to allocate the HeapRegion instances 1733 assert(mr.end() <= new_end, "post-condition"); 1734 1735 size_t actual_expand_bytes = mr.byte_size(); 1736 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1737 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1738 "post-condition"); 1739 if (actual_expand_bytes < aligned_expand_bytes) { 1740 // We could not expand _hrs to the desired size. In this case we 1741 // need to shrink the committed space accordingly. 1742 assert(mr.end() < new_end, "invariant"); 1743 1744 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1745 // First uncommit the memory. 1746 _g1_storage.shrink_by(diff_bytes); 1747 // Then propagate this update to the necessary data structures. 1748 update_committed_space(new_end, mr.end()); 1749 } 1750 _free_list.add_as_tail(&expansion_list); 1751 1752 if (_hr_printer.is_active()) { 1753 HeapWord* curr = mr.start(); 1754 while (curr < mr.end()) { 1755 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1756 _hr_printer.commit(curr, curr_end); 1757 curr = curr_end; 1758 } 1759 assert(curr == mr.end(), "post-condition"); 1760 } 1761 g1_policy()->record_new_heap_size(n_regions()); 1762 } else { 1763 ergo_verbose0(ErgoHeapSizing, 1764 "did not expand the heap", 1765 ergo_format_reason("heap expansion operation failed")); 1766 // The expansion of the virtual storage space was unsuccessful. 1767 // Let's see if it was because we ran out of swap. 1768 if (G1ExitOnExpansionFailure && 1769 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1770 // We had head room... 1771 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1772 } 1773 } 1774 return successful; 1775 } 1776 1777 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1778 size_t old_mem_size = _g1_storage.committed_size(); 1779 size_t aligned_shrink_bytes = 1780 ReservedSpace::page_align_size_down(shrink_bytes); 1781 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1782 HeapRegion::GrainBytes); 1783 uint num_regions_deleted = 0; 1784 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted); 1785 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1786 assert(mr.end() == old_end, "post-condition"); 1787 1788 ergo_verbose3(ErgoHeapSizing, 1789 "shrink the heap", 1790 ergo_format_byte("requested shrinking amount") 1791 ergo_format_byte("aligned shrinking amount") 1792 ergo_format_byte("attempted shrinking amount"), 1793 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1794 if (mr.byte_size() > 0) { 1795 if (_hr_printer.is_active()) { 1796 HeapWord* curr = mr.end(); 1797 while (curr > mr.start()) { 1798 HeapWord* curr_end = curr; 1799 curr -= HeapRegion::GrainWords; 1800 _hr_printer.uncommit(curr, curr_end); 1801 } 1802 assert(curr == mr.start(), "post-condition"); 1803 } 1804 1805 _g1_storage.shrink_by(mr.byte_size()); 1806 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1807 assert(mr.start() == new_end, "post-condition"); 1808 1809 _expansion_regions += num_regions_deleted; 1810 update_committed_space(old_end, new_end); 1811 HeapRegionRemSet::shrink_heap(n_regions()); 1812 g1_policy()->record_new_heap_size(n_regions()); 1813 } else { 1814 ergo_verbose0(ErgoHeapSizing, 1815 "did not shrink the heap", 1816 ergo_format_reason("heap shrinking operation failed")); 1817 } 1818 } 1819 1820 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1821 verify_region_sets_optional(); 1822 1823 // We should only reach here at the end of a Full GC which means we 1824 // should not not be holding to any GC alloc regions. The method 1825 // below will make sure of that and do any remaining clean up. 1826 abandon_gc_alloc_regions(); 1827 1828 // Instead of tearing down / rebuilding the free lists here, we 1829 // could instead use the remove_all_pending() method on free_list to 1830 // remove only the ones that we need to remove. 1831 tear_down_region_sets(true /* free_list_only */); 1832 shrink_helper(shrink_bytes); 1833 rebuild_region_sets(true /* free_list_only */); 1834 1835 _hrs.verify_optional(); 1836 verify_region_sets_optional(); 1837 } 1838 1839 // Public methods. 1840 1841 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1842 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1843 #endif // _MSC_VER 1844 1845 1846 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1847 SharedHeap(policy_), 1848 _g1_policy(policy_), 1849 _dirty_card_queue_set(false), 1850 _into_cset_dirty_card_queue_set(false), 1851 _is_alive_closure_cm(this), 1852 _is_alive_closure_stw(this), 1853 _ref_processor_cm(NULL), 1854 _ref_processor_stw(NULL), 1855 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1856 _bot_shared(NULL), 1857 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL), 1858 _evac_failure_scan_stack(NULL) , 1859 _mark_in_progress(false), 1860 _cg1r(NULL), _summary_bytes_used(0), 1861 _g1mm(NULL), 1862 _refine_cte_cl(NULL), 1863 _full_collection(false), 1864 _free_list("Master Free List"), 1865 _secondary_free_list("Secondary Free List"), 1866 _old_set("Old Set"), 1867 _humongous_set("Master Humongous Set"), 1868 _free_regions_coming(false), 1869 _young_list(new YoungList(this)), 1870 _gc_time_stamp(0), 1871 _retained_old_gc_alloc_region(NULL), 1872 _expand_heap_after_alloc_failure(true), 1873 _surviving_young_words(NULL), 1874 _full_collections_completed(0), 1875 _in_cset_fast_test(NULL), 1876 _in_cset_fast_test_base(NULL), 1877 _dirty_cards_region_list(NULL), 1878 _worker_cset_start_region(NULL), 1879 _worker_cset_start_region_time_stamp(NULL) { 1880 _g1h = this; // To catch bugs. 1881 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1882 vm_exit_during_initialization("Failed necessary allocation."); 1883 } 1884 1885 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1886 1887 int n_queues = MAX2((int)ParallelGCThreads, 1); 1888 _task_queues = new RefToScanQueueSet(n_queues); 1889 1890 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1891 assert(n_rem_sets > 0, "Invariant."); 1892 1893 HeapRegionRemSetIterator** iter_arr = 1894 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues); 1895 for (int i = 0; i < n_queues; i++) { 1896 iter_arr[i] = new HeapRegionRemSetIterator(); 1897 } 1898 _rem_set_iterator = iter_arr; 1899 1900 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues); 1901 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues); 1902 1903 for (int i = 0; i < n_queues; i++) { 1904 RefToScanQueue* q = new RefToScanQueue(); 1905 q->initialize(); 1906 _task_queues->register_queue(i, q); 1907 } 1908 1909 clear_cset_start_regions(); 1910 1911 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1912 } 1913 1914 jint G1CollectedHeap::initialize() { 1915 CollectedHeap::pre_initialize(); 1916 os::enable_vtime(); 1917 1918 G1Log::init(); 1919 1920 // Necessary to satisfy locking discipline assertions. 1921 1922 MutexLocker x(Heap_lock); 1923 1924 // We have to initialize the printer before committing the heap, as 1925 // it will be used then. 1926 _hr_printer.set_active(G1PrintHeapRegions); 1927 1928 // While there are no constraints in the GC code that HeapWordSize 1929 // be any particular value, there are multiple other areas in the 1930 // system which believe this to be true (e.g. oop->object_size in some 1931 // cases incorrectly returns the size in wordSize units rather than 1932 // HeapWordSize). 1933 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1934 1935 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1936 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1937 1938 // Ensure that the sizes are properly aligned. 1939 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1940 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1941 1942 _cg1r = new ConcurrentG1Refine(); 1943 1944 // Reserve the maximum. 1945 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1946 // Includes the perm-gen. 1947 1948 // When compressed oops are enabled, the preferred heap base 1949 // is calculated by subtracting the requested size from the 1950 // 32Gb boundary and using the result as the base address for 1951 // heap reservation. If the requested size is not aligned to 1952 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1953 // into the ReservedHeapSpace constructor) then the actual 1954 // base of the reserved heap may end up differing from the 1955 // address that was requested (i.e. the preferred heap base). 1956 // If this happens then we could end up using a non-optimal 1957 // compressed oops mode. 1958 1959 // Since max_byte_size is aligned to the size of a heap region (checked 1960 // above), we also need to align the perm gen size as it might not be. 1961 const size_t total_reserved = max_byte_size + 1962 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1963 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1964 1965 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1966 1967 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1968 UseLargePages, addr); 1969 1970 if (UseCompressedOops) { 1971 if (addr != NULL && !heap_rs.is_reserved()) { 1972 // Failed to reserve at specified address - the requested memory 1973 // region is taken already, for example, by 'java' launcher. 1974 // Try again to reserver heap higher. 1975 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 1976 1977 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 1978 UseLargePages, addr); 1979 1980 if (addr != NULL && !heap_rs0.is_reserved()) { 1981 // Failed to reserve at specified address again - give up. 1982 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 1983 assert(addr == NULL, ""); 1984 1985 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 1986 UseLargePages, addr); 1987 heap_rs = heap_rs1; 1988 } else { 1989 heap_rs = heap_rs0; 1990 } 1991 } 1992 } 1993 1994 if (!heap_rs.is_reserved()) { 1995 vm_exit_during_initialization("Could not reserve enough space for object heap"); 1996 return JNI_ENOMEM; 1997 } 1998 1999 // It is important to do this in a way such that concurrent readers can't 2000 // temporarily think somethings in the heap. (I've actually seen this 2001 // happen in asserts: DLD.) 2002 _reserved.set_word_size(0); 2003 _reserved.set_start((HeapWord*)heap_rs.base()); 2004 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2005 2006 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2007 2008 // Create the gen rem set (and barrier set) for the entire reserved region. 2009 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2010 set_barrier_set(rem_set()->bs()); 2011 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2012 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2013 } else { 2014 vm_exit_during_initialization("G1 requires a mod ref bs."); 2015 return JNI_ENOMEM; 2016 } 2017 2018 // Also create a G1 rem set. 2019 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2020 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2021 } else { 2022 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2023 return JNI_ENOMEM; 2024 } 2025 2026 // Carve out the G1 part of the heap. 2027 2028 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2029 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2030 g1_rs.size()/HeapWordSize); 2031 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 2032 2033 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2034 2035 _g1_storage.initialize(g1_rs, 0); 2036 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2037 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2038 (HeapWord*) _g1_reserved.end(), 2039 _expansion_regions); 2040 2041 // 6843694 - ensure that the maximum region index can fit 2042 // in the remembered set structures. 2043 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2044 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2045 2046 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2047 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2048 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2049 "too many cards per region"); 2050 2051 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2052 2053 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2054 heap_word_size(init_byte_size)); 2055 2056 _g1h = this; 2057 2058 _in_cset_fast_test_length = max_regions(); 2059 _in_cset_fast_test_base = 2060 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length); 2061 2062 // We're biasing _in_cset_fast_test to avoid subtracting the 2063 // beginning of the heap every time we want to index; basically 2064 // it's the same with what we do with the card table. 2065 _in_cset_fast_test = _in_cset_fast_test_base - 2066 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2067 2068 // Clear the _cset_fast_test bitmap in anticipation of adding 2069 // regions to the incremental collection set for the first 2070 // evacuation pause. 2071 clear_cset_fast_test(); 2072 2073 // Create the ConcurrentMark data structure and thread. 2074 // (Must do this late, so that "max_regions" is defined.) 2075 _cm = new ConcurrentMark(heap_rs, max_regions()); 2076 _cmThread = _cm->cmThread(); 2077 2078 // Initialize the from_card cache structure of HeapRegionRemSet. 2079 HeapRegionRemSet::init_heap(max_regions()); 2080 2081 // Now expand into the initial heap size. 2082 if (!expand(init_byte_size)) { 2083 vm_exit_during_initialization("Failed to allocate initial heap."); 2084 return JNI_ENOMEM; 2085 } 2086 2087 // Perform any initialization actions delegated to the policy. 2088 g1_policy()->init(); 2089 2090 _refine_cte_cl = 2091 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2092 g1_rem_set(), 2093 concurrent_g1_refine()); 2094 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2095 2096 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2097 SATB_Q_FL_lock, 2098 G1SATBProcessCompletedThreshold, 2099 Shared_SATB_Q_lock); 2100 2101 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2102 DirtyCardQ_FL_lock, 2103 concurrent_g1_refine()->yellow_zone(), 2104 concurrent_g1_refine()->red_zone(), 2105 Shared_DirtyCardQ_lock); 2106 2107 if (G1DeferredRSUpdate) { 2108 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2109 DirtyCardQ_FL_lock, 2110 -1, // never trigger processing 2111 -1, // no limit on length 2112 Shared_DirtyCardQ_lock, 2113 &JavaThread::dirty_card_queue_set()); 2114 } 2115 2116 // Initialize the card queue set used to hold cards containing 2117 // references into the collection set. 2118 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2119 DirtyCardQ_FL_lock, 2120 -1, // never trigger processing 2121 -1, // no limit on length 2122 Shared_DirtyCardQ_lock, 2123 &JavaThread::dirty_card_queue_set()); 2124 2125 // In case we're keeping closure specialization stats, initialize those 2126 // counts and that mechanism. 2127 SpecializationStats::clear(); 2128 2129 // Do later initialization work for concurrent refinement. 2130 _cg1r->init(); 2131 2132 // Here we allocate the dummy full region that is required by the 2133 // G1AllocRegion class. If we don't pass an address in the reserved 2134 // space here, lots of asserts fire. 2135 2136 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2137 _g1_reserved.start()); 2138 // We'll re-use the same region whether the alloc region will 2139 // require BOT updates or not and, if it doesn't, then a non-young 2140 // region will complain that it cannot support allocations without 2141 // BOT updates. So we'll tag the dummy region as young to avoid that. 2142 dummy_region->set_young(); 2143 // Make sure it's full. 2144 dummy_region->set_top(dummy_region->end()); 2145 G1AllocRegion::setup(this, dummy_region); 2146 2147 init_mutator_alloc_region(); 2148 2149 // Do create of the monitoring and management support so that 2150 // values in the heap have been properly initialized. 2151 _g1mm = new G1MonitoringSupport(this); 2152 2153 return JNI_OK; 2154 } 2155 2156 void G1CollectedHeap::ref_processing_init() { 2157 // Reference processing in G1 currently works as follows: 2158 // 2159 // * There are two reference processor instances. One is 2160 // used to record and process discovered references 2161 // during concurrent marking; the other is used to 2162 // record and process references during STW pauses 2163 // (both full and incremental). 2164 // * Both ref processors need to 'span' the entire heap as 2165 // the regions in the collection set may be dotted around. 2166 // 2167 // * For the concurrent marking ref processor: 2168 // * Reference discovery is enabled at initial marking. 2169 // * Reference discovery is disabled and the discovered 2170 // references processed etc during remarking. 2171 // * Reference discovery is MT (see below). 2172 // * Reference discovery requires a barrier (see below). 2173 // * Reference processing may or may not be MT 2174 // (depending on the value of ParallelRefProcEnabled 2175 // and ParallelGCThreads). 2176 // * A full GC disables reference discovery by the CM 2177 // ref processor and abandons any entries on it's 2178 // discovered lists. 2179 // 2180 // * For the STW processor: 2181 // * Non MT discovery is enabled at the start of a full GC. 2182 // * Processing and enqueueing during a full GC is non-MT. 2183 // * During a full GC, references are processed after marking. 2184 // 2185 // * Discovery (may or may not be MT) is enabled at the start 2186 // of an incremental evacuation pause. 2187 // * References are processed near the end of a STW evacuation pause. 2188 // * For both types of GC: 2189 // * Discovery is atomic - i.e. not concurrent. 2190 // * Reference discovery will not need a barrier. 2191 2192 SharedHeap::ref_processing_init(); 2193 MemRegion mr = reserved_region(); 2194 2195 // Concurrent Mark ref processor 2196 _ref_processor_cm = 2197 new ReferenceProcessor(mr, // span 2198 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2199 // mt processing 2200 (int) ParallelGCThreads, 2201 // degree of mt processing 2202 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2203 // mt discovery 2204 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2205 // degree of mt discovery 2206 false, 2207 // Reference discovery is not atomic 2208 &_is_alive_closure_cm, 2209 // is alive closure 2210 // (for efficiency/performance) 2211 true); 2212 // Setting next fields of discovered 2213 // lists requires a barrier. 2214 2215 // STW ref processor 2216 _ref_processor_stw = 2217 new ReferenceProcessor(mr, // span 2218 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2219 // mt processing 2220 MAX2((int)ParallelGCThreads, 1), 2221 // degree of mt processing 2222 (ParallelGCThreads > 1), 2223 // mt discovery 2224 MAX2((int)ParallelGCThreads, 1), 2225 // degree of mt discovery 2226 true, 2227 // Reference discovery is atomic 2228 &_is_alive_closure_stw, 2229 // is alive closure 2230 // (for efficiency/performance) 2231 false); 2232 // Setting next fields of discovered 2233 // lists requires a barrier. 2234 } 2235 2236 size_t G1CollectedHeap::capacity() const { 2237 return _g1_committed.byte_size(); 2238 } 2239 2240 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2241 DirtyCardQueue* into_cset_dcq, 2242 bool concurrent, 2243 int worker_i) { 2244 // Clean cards in the hot card cache 2245 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2246 2247 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2248 int n_completed_buffers = 0; 2249 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2250 n_completed_buffers++; 2251 } 2252 g1_policy()->record_update_rs_processed_buffers(worker_i, 2253 (double) n_completed_buffers); 2254 dcqs.clear_n_completed_buffers(); 2255 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2256 } 2257 2258 2259 // Computes the sum of the storage used by the various regions. 2260 2261 size_t G1CollectedHeap::used() const { 2262 assert(Heap_lock->owner() != NULL, 2263 "Should be owned on this thread's behalf."); 2264 size_t result = _summary_bytes_used; 2265 // Read only once in case it is set to NULL concurrently 2266 HeapRegion* hr = _mutator_alloc_region.get(); 2267 if (hr != NULL) 2268 result += hr->used(); 2269 return result; 2270 } 2271 2272 size_t G1CollectedHeap::used_unlocked() const { 2273 size_t result = _summary_bytes_used; 2274 return result; 2275 } 2276 2277 class SumUsedClosure: public HeapRegionClosure { 2278 size_t _used; 2279 public: 2280 SumUsedClosure() : _used(0) {} 2281 bool doHeapRegion(HeapRegion* r) { 2282 if (!r->continuesHumongous()) { 2283 _used += r->used(); 2284 } 2285 return false; 2286 } 2287 size_t result() { return _used; } 2288 }; 2289 2290 size_t G1CollectedHeap::recalculate_used() const { 2291 SumUsedClosure blk; 2292 heap_region_iterate(&blk); 2293 return blk.result(); 2294 } 2295 2296 size_t G1CollectedHeap::unsafe_max_alloc() { 2297 if (free_regions() > 0) return HeapRegion::GrainBytes; 2298 // otherwise, is there space in the current allocation region? 2299 2300 // We need to store the current allocation region in a local variable 2301 // here. The problem is that this method doesn't take any locks and 2302 // there may be other threads which overwrite the current allocation 2303 // region field. attempt_allocation(), for example, sets it to NULL 2304 // and this can happen *after* the NULL check here but before the call 2305 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2306 // to be a problem in the optimized build, since the two loads of the 2307 // current allocation region field are optimized away. 2308 HeapRegion* hr = _mutator_alloc_region.get(); 2309 if (hr == NULL) { 2310 return 0; 2311 } 2312 return hr->free(); 2313 } 2314 2315 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2316 switch (cause) { 2317 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2318 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2319 case GCCause::_g1_humongous_allocation: return true; 2320 default: return false; 2321 } 2322 } 2323 2324 #ifndef PRODUCT 2325 void G1CollectedHeap::allocate_dummy_regions() { 2326 // Let's fill up most of the region 2327 size_t word_size = HeapRegion::GrainWords - 1024; 2328 // And as a result the region we'll allocate will be humongous. 2329 guarantee(isHumongous(word_size), "sanity"); 2330 2331 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2332 // Let's use the existing mechanism for the allocation 2333 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2334 if (dummy_obj != NULL) { 2335 MemRegion mr(dummy_obj, word_size); 2336 CollectedHeap::fill_with_object(mr); 2337 } else { 2338 // If we can't allocate once, we probably cannot allocate 2339 // again. Let's get out of the loop. 2340 break; 2341 } 2342 } 2343 } 2344 #endif // !PRODUCT 2345 2346 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) { 2347 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2348 2349 // We assume that if concurrent == true, then the caller is a 2350 // concurrent thread that was joined the Suspendible Thread 2351 // Set. If there's ever a cheap way to check this, we should add an 2352 // assert here. 2353 2354 // We have already incremented _total_full_collections at the start 2355 // of the GC, so total_full_collections() represents how many full 2356 // collections have been started. 2357 unsigned int full_collections_started = total_full_collections(); 2358 2359 // Given that this method is called at the end of a Full GC or of a 2360 // concurrent cycle, and those can be nested (i.e., a Full GC can 2361 // interrupt a concurrent cycle), the number of full collections 2362 // completed should be either one (in the case where there was no 2363 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2364 // behind the number of full collections started. 2365 2366 // This is the case for the inner caller, i.e. a Full GC. 2367 assert(concurrent || 2368 (full_collections_started == _full_collections_completed + 1) || 2369 (full_collections_started == _full_collections_completed + 2), 2370 err_msg("for inner caller (Full GC): full_collections_started = %u " 2371 "is inconsistent with _full_collections_completed = %u", 2372 full_collections_started, _full_collections_completed)); 2373 2374 // This is the case for the outer caller, i.e. the concurrent cycle. 2375 assert(!concurrent || 2376 (full_collections_started == _full_collections_completed + 1), 2377 err_msg("for outer caller (concurrent cycle): " 2378 "full_collections_started = %u " 2379 "is inconsistent with _full_collections_completed = %u", 2380 full_collections_started, _full_collections_completed)); 2381 2382 _full_collections_completed += 1; 2383 2384 // We need to clear the "in_progress" flag in the CM thread before 2385 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2386 // is set) so that if a waiter requests another System.gc() it doesn't 2387 // incorrectly see that a marking cyle is still in progress. 2388 if (concurrent) { 2389 _cmThread->clear_in_progress(); 2390 } 2391 2392 // This notify_all() will ensure that a thread that called 2393 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2394 // and it's waiting for a full GC to finish will be woken up. It is 2395 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2396 FullGCCount_lock->notify_all(); 2397 } 2398 2399 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2400 assert_at_safepoint(true /* should_be_vm_thread */); 2401 GCCauseSetter gcs(this, cause); 2402 switch (cause) { 2403 case GCCause::_heap_inspection: 2404 case GCCause::_heap_dump: { 2405 HandleMark hm; 2406 do_full_collection(false); // don't clear all soft refs 2407 break; 2408 } 2409 default: // XXX FIX ME 2410 ShouldNotReachHere(); // Unexpected use of this function 2411 } 2412 } 2413 2414 void G1CollectedHeap::collect(GCCause::Cause cause) { 2415 assert_heap_not_locked(); 2416 2417 unsigned int gc_count_before; 2418 unsigned int full_gc_count_before; 2419 bool retry_gc; 2420 2421 do { 2422 retry_gc = false; 2423 2424 { 2425 MutexLocker ml(Heap_lock); 2426 2427 // Read the GC count while holding the Heap_lock 2428 gc_count_before = total_collections(); 2429 full_gc_count_before = total_full_collections(); 2430 } 2431 2432 if (should_do_concurrent_full_gc(cause)) { 2433 // Schedule an initial-mark evacuation pause that will start a 2434 // concurrent cycle. We're setting word_size to 0 which means that 2435 // we are not requesting a post-GC allocation. 2436 VM_G1IncCollectionPause op(gc_count_before, 2437 0, /* word_size */ 2438 true, /* should_initiate_conc_mark */ 2439 g1_policy()->max_pause_time_ms(), 2440 cause); 2441 2442 VMThread::execute(&op); 2443 if (!op.pause_succeeded()) { 2444 if (full_gc_count_before == total_full_collections()) { 2445 retry_gc = op.should_retry_gc(); 2446 } else { 2447 // A Full GC happened while we were trying to schedule the 2448 // initial-mark GC. No point in starting a new cycle given 2449 // that the whole heap was collected anyway. 2450 } 2451 2452 if (retry_gc) { 2453 if (GC_locker::is_active_and_needs_gc()) { 2454 GC_locker::stall_until_clear(); 2455 } 2456 } 2457 } 2458 } else { 2459 if (cause == GCCause::_gc_locker 2460 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2461 2462 // Schedule a standard evacuation pause. We're setting word_size 2463 // to 0 which means that we are not requesting a post-GC allocation. 2464 VM_G1IncCollectionPause op(gc_count_before, 2465 0, /* word_size */ 2466 false, /* should_initiate_conc_mark */ 2467 g1_policy()->max_pause_time_ms(), 2468 cause); 2469 VMThread::execute(&op); 2470 } else { 2471 // Schedule a Full GC. 2472 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause); 2473 VMThread::execute(&op); 2474 } 2475 } 2476 } while (retry_gc); 2477 } 2478 2479 bool G1CollectedHeap::is_in(const void* p) const { 2480 if (_g1_committed.contains(p)) { 2481 // Given that we know that p is in the committed space, 2482 // heap_region_containing_raw() should successfully 2483 // return the containing region. 2484 HeapRegion* hr = heap_region_containing_raw(p); 2485 return hr->is_in(p); 2486 } else { 2487 return _perm_gen->as_gen()->is_in(p); 2488 } 2489 } 2490 2491 // Iteration functions. 2492 2493 // Iterates an OopClosure over all ref-containing fields of objects 2494 // within a HeapRegion. 2495 2496 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2497 MemRegion _mr; 2498 OopClosure* _cl; 2499 public: 2500 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2501 : _mr(mr), _cl(cl) {} 2502 bool doHeapRegion(HeapRegion* r) { 2503 if (! r->continuesHumongous()) { 2504 r->oop_iterate(_cl); 2505 } 2506 return false; 2507 } 2508 }; 2509 2510 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2511 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2512 heap_region_iterate(&blk); 2513 if (do_perm) { 2514 perm_gen()->oop_iterate(cl); 2515 } 2516 } 2517 2518 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2519 IterateOopClosureRegionClosure blk(mr, cl); 2520 heap_region_iterate(&blk); 2521 if (do_perm) { 2522 perm_gen()->oop_iterate(cl); 2523 } 2524 } 2525 2526 // Iterates an ObjectClosure over all objects within a HeapRegion. 2527 2528 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2529 ObjectClosure* _cl; 2530 public: 2531 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2532 bool doHeapRegion(HeapRegion* r) { 2533 if (! r->continuesHumongous()) { 2534 r->object_iterate(_cl); 2535 } 2536 return false; 2537 } 2538 }; 2539 2540 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2541 IterateObjectClosureRegionClosure blk(cl); 2542 heap_region_iterate(&blk); 2543 if (do_perm) { 2544 perm_gen()->object_iterate(cl); 2545 } 2546 } 2547 2548 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2549 // FIXME: is this right? 2550 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2551 } 2552 2553 // Calls a SpaceClosure on a HeapRegion. 2554 2555 class SpaceClosureRegionClosure: public HeapRegionClosure { 2556 SpaceClosure* _cl; 2557 public: 2558 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2559 bool doHeapRegion(HeapRegion* r) { 2560 _cl->do_space(r); 2561 return false; 2562 } 2563 }; 2564 2565 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2566 SpaceClosureRegionClosure blk(cl); 2567 heap_region_iterate(&blk); 2568 } 2569 2570 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2571 _hrs.iterate(cl); 2572 } 2573 2574 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r, 2575 HeapRegionClosure* cl) const { 2576 _hrs.iterate_from(r, cl); 2577 } 2578 2579 void 2580 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2581 uint worker, 2582 uint no_of_par_workers, 2583 jint claim_value) { 2584 const uint regions = n_regions(); 2585 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2586 no_of_par_workers : 2587 1); 2588 assert(UseDynamicNumberOfGCThreads || 2589 no_of_par_workers == workers()->total_workers(), 2590 "Non dynamic should use fixed number of workers"); 2591 // try to spread out the starting points of the workers 2592 const uint start_index = regions / max_workers * worker; 2593 2594 // each worker will actually look at all regions 2595 for (uint count = 0; count < regions; ++count) { 2596 const uint index = (start_index + count) % regions; 2597 assert(0 <= index && index < regions, "sanity"); 2598 HeapRegion* r = region_at(index); 2599 // we'll ignore "continues humongous" regions (we'll process them 2600 // when we come across their corresponding "start humongous" 2601 // region) and regions already claimed 2602 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2603 continue; 2604 } 2605 // OK, try to claim it 2606 if (r->claimHeapRegion(claim_value)) { 2607 // success! 2608 assert(!r->continuesHumongous(), "sanity"); 2609 if (r->startsHumongous()) { 2610 // If the region is "starts humongous" we'll iterate over its 2611 // "continues humongous" first; in fact we'll do them 2612 // first. The order is important. In on case, calling the 2613 // closure on the "starts humongous" region might de-allocate 2614 // and clear all its "continues humongous" regions and, as a 2615 // result, we might end up processing them twice. So, we'll do 2616 // them first (notice: most closures will ignore them anyway) and 2617 // then we'll do the "starts humongous" region. 2618 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2619 HeapRegion* chr = region_at(ch_index); 2620 2621 // if the region has already been claimed or it's not 2622 // "continues humongous" we're done 2623 if (chr->claim_value() == claim_value || 2624 !chr->continuesHumongous()) { 2625 break; 2626 } 2627 2628 // Noone should have claimed it directly. We can given 2629 // that we claimed its "starts humongous" region. 2630 assert(chr->claim_value() != claim_value, "sanity"); 2631 assert(chr->humongous_start_region() == r, "sanity"); 2632 2633 if (chr->claimHeapRegion(claim_value)) { 2634 // we should always be able to claim it; noone else should 2635 // be trying to claim this region 2636 2637 bool res2 = cl->doHeapRegion(chr); 2638 assert(!res2, "Should not abort"); 2639 2640 // Right now, this holds (i.e., no closure that actually 2641 // does something with "continues humongous" regions 2642 // clears them). We might have to weaken it in the future, 2643 // but let's leave these two asserts here for extra safety. 2644 assert(chr->continuesHumongous(), "should still be the case"); 2645 assert(chr->humongous_start_region() == r, "sanity"); 2646 } else { 2647 guarantee(false, "we should not reach here"); 2648 } 2649 } 2650 } 2651 2652 assert(!r->continuesHumongous(), "sanity"); 2653 bool res = cl->doHeapRegion(r); 2654 assert(!res, "Should not abort"); 2655 } 2656 } 2657 } 2658 2659 class ResetClaimValuesClosure: public HeapRegionClosure { 2660 public: 2661 bool doHeapRegion(HeapRegion* r) { 2662 r->set_claim_value(HeapRegion::InitialClaimValue); 2663 return false; 2664 } 2665 }; 2666 2667 void G1CollectedHeap::reset_heap_region_claim_values() { 2668 ResetClaimValuesClosure blk; 2669 heap_region_iterate(&blk); 2670 } 2671 2672 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2673 ResetClaimValuesClosure blk; 2674 collection_set_iterate(&blk); 2675 } 2676 2677 #ifdef ASSERT 2678 // This checks whether all regions in the heap have the correct claim 2679 // value. I also piggy-backed on this a check to ensure that the 2680 // humongous_start_region() information on "continues humongous" 2681 // regions is correct. 2682 2683 class CheckClaimValuesClosure : public HeapRegionClosure { 2684 private: 2685 jint _claim_value; 2686 uint _failures; 2687 HeapRegion* _sh_region; 2688 2689 public: 2690 CheckClaimValuesClosure(jint claim_value) : 2691 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2692 bool doHeapRegion(HeapRegion* r) { 2693 if (r->claim_value() != _claim_value) { 2694 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2695 "claim value = %d, should be %d", 2696 HR_FORMAT_PARAMS(r), 2697 r->claim_value(), _claim_value); 2698 ++_failures; 2699 } 2700 if (!r->isHumongous()) { 2701 _sh_region = NULL; 2702 } else if (r->startsHumongous()) { 2703 _sh_region = r; 2704 } else if (r->continuesHumongous()) { 2705 if (r->humongous_start_region() != _sh_region) { 2706 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2707 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2708 HR_FORMAT_PARAMS(r), 2709 r->humongous_start_region(), 2710 _sh_region); 2711 ++_failures; 2712 } 2713 } 2714 return false; 2715 } 2716 uint failures() { return _failures; } 2717 }; 2718 2719 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2720 CheckClaimValuesClosure cl(claim_value); 2721 heap_region_iterate(&cl); 2722 return cl.failures() == 0; 2723 } 2724 2725 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2726 private: 2727 jint _claim_value; 2728 uint _failures; 2729 2730 public: 2731 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2732 _claim_value(claim_value), _failures(0) { } 2733 2734 uint failures() { return _failures; } 2735 2736 bool doHeapRegion(HeapRegion* hr) { 2737 assert(hr->in_collection_set(), "how?"); 2738 assert(!hr->isHumongous(), "H-region in CSet"); 2739 if (hr->claim_value() != _claim_value) { 2740 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2741 "claim value = %d, should be %d", 2742 HR_FORMAT_PARAMS(hr), 2743 hr->claim_value(), _claim_value); 2744 _failures += 1; 2745 } 2746 return false; 2747 } 2748 }; 2749 2750 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2751 CheckClaimValuesInCSetHRClosure cl(claim_value); 2752 collection_set_iterate(&cl); 2753 return cl.failures() == 0; 2754 } 2755 #endif // ASSERT 2756 2757 // Clear the cached CSet starting regions and (more importantly) 2758 // the time stamps. Called when we reset the GC time stamp. 2759 void G1CollectedHeap::clear_cset_start_regions() { 2760 assert(_worker_cset_start_region != NULL, "sanity"); 2761 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2762 2763 int n_queues = MAX2((int)ParallelGCThreads, 1); 2764 for (int i = 0; i < n_queues; i++) { 2765 _worker_cset_start_region[i] = NULL; 2766 _worker_cset_start_region_time_stamp[i] = 0; 2767 } 2768 } 2769 2770 // Given the id of a worker, obtain or calculate a suitable 2771 // starting region for iterating over the current collection set. 2772 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2773 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2774 2775 HeapRegion* result = NULL; 2776 unsigned gc_time_stamp = get_gc_time_stamp(); 2777 2778 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2779 // Cached starting region for current worker was set 2780 // during the current pause - so it's valid. 2781 // Note: the cached starting heap region may be NULL 2782 // (when the collection set is empty). 2783 result = _worker_cset_start_region[worker_i]; 2784 assert(result == NULL || result->in_collection_set(), "sanity"); 2785 return result; 2786 } 2787 2788 // The cached entry was not valid so let's calculate 2789 // a suitable starting heap region for this worker. 2790 2791 // We want the parallel threads to start their collection 2792 // set iteration at different collection set regions to 2793 // avoid contention. 2794 // If we have: 2795 // n collection set regions 2796 // p threads 2797 // Then thread t will start at region floor ((t * n) / p) 2798 2799 result = g1_policy()->collection_set(); 2800 if (G1CollectedHeap::use_parallel_gc_threads()) { 2801 uint cs_size = g1_policy()->cset_region_length(); 2802 uint active_workers = workers()->active_workers(); 2803 assert(UseDynamicNumberOfGCThreads || 2804 active_workers == workers()->total_workers(), 2805 "Unless dynamic should use total workers"); 2806 2807 uint end_ind = (cs_size * worker_i) / active_workers; 2808 uint start_ind = 0; 2809 2810 if (worker_i > 0 && 2811 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2812 // Previous workers starting region is valid 2813 // so let's iterate from there 2814 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2815 result = _worker_cset_start_region[worker_i - 1]; 2816 } 2817 2818 for (uint i = start_ind; i < end_ind; i++) { 2819 result = result->next_in_collection_set(); 2820 } 2821 } 2822 2823 // Note: the calculated starting heap region may be NULL 2824 // (when the collection set is empty). 2825 assert(result == NULL || result->in_collection_set(), "sanity"); 2826 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2827 "should be updated only once per pause"); 2828 _worker_cset_start_region[worker_i] = result; 2829 OrderAccess::storestore(); 2830 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2831 return result; 2832 } 2833 2834 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2835 HeapRegion* r = g1_policy()->collection_set(); 2836 while (r != NULL) { 2837 HeapRegion* next = r->next_in_collection_set(); 2838 if (cl->doHeapRegion(r)) { 2839 cl->incomplete(); 2840 return; 2841 } 2842 r = next; 2843 } 2844 } 2845 2846 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2847 HeapRegionClosure *cl) { 2848 if (r == NULL) { 2849 // The CSet is empty so there's nothing to do. 2850 return; 2851 } 2852 2853 assert(r->in_collection_set(), 2854 "Start region must be a member of the collection set."); 2855 HeapRegion* cur = r; 2856 while (cur != NULL) { 2857 HeapRegion* next = cur->next_in_collection_set(); 2858 if (cl->doHeapRegion(cur) && false) { 2859 cl->incomplete(); 2860 return; 2861 } 2862 cur = next; 2863 } 2864 cur = g1_policy()->collection_set(); 2865 while (cur != r) { 2866 HeapRegion* next = cur->next_in_collection_set(); 2867 if (cl->doHeapRegion(cur) && false) { 2868 cl->incomplete(); 2869 return; 2870 } 2871 cur = next; 2872 } 2873 } 2874 2875 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2876 return n_regions() > 0 ? region_at(0) : NULL; 2877 } 2878 2879 2880 Space* G1CollectedHeap::space_containing(const void* addr) const { 2881 Space* res = heap_region_containing(addr); 2882 if (res == NULL) 2883 res = perm_gen()->space_containing(addr); 2884 return res; 2885 } 2886 2887 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2888 Space* sp = space_containing(addr); 2889 if (sp != NULL) { 2890 return sp->block_start(addr); 2891 } 2892 return NULL; 2893 } 2894 2895 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2896 Space* sp = space_containing(addr); 2897 assert(sp != NULL, "block_size of address outside of heap"); 2898 return sp->block_size(addr); 2899 } 2900 2901 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2902 Space* sp = space_containing(addr); 2903 return sp->block_is_obj(addr); 2904 } 2905 2906 bool G1CollectedHeap::supports_tlab_allocation() const { 2907 return true; 2908 } 2909 2910 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2911 return HeapRegion::GrainBytes; 2912 } 2913 2914 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2915 // Return the remaining space in the cur alloc region, but not less than 2916 // the min TLAB size. 2917 2918 // Also, this value can be at most the humongous object threshold, 2919 // since we can't allow tlabs to grow big enough to accomodate 2920 // humongous objects. 2921 2922 HeapRegion* hr = _mutator_alloc_region.get(); 2923 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2924 if (hr == NULL) { 2925 return max_tlab_size; 2926 } else { 2927 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2928 } 2929 } 2930 2931 size_t G1CollectedHeap::max_capacity() const { 2932 return _g1_reserved.byte_size(); 2933 } 2934 2935 jlong G1CollectedHeap::millis_since_last_gc() { 2936 // assert(false, "NYI"); 2937 return 0; 2938 } 2939 2940 void G1CollectedHeap::prepare_for_verify() { 2941 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2942 ensure_parsability(false); 2943 } 2944 g1_rem_set()->prepare_for_verify(); 2945 } 2946 2947 class VerifyLivenessOopClosure: public OopClosure { 2948 G1CollectedHeap* _g1h; 2949 VerifyOption _vo; 2950 public: 2951 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 2952 _g1h(g1h), _vo(vo) 2953 { } 2954 void do_oop(narrowOop *p) { do_oop_work(p); } 2955 void do_oop( oop *p) { do_oop_work(p); } 2956 2957 template <class T> void do_oop_work(T *p) { 2958 oop obj = oopDesc::load_decode_heap_oop(p); 2959 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 2960 "Dead object referenced by a not dead object"); 2961 } 2962 }; 2963 2964 class VerifyObjsInRegionClosure: public ObjectClosure { 2965 private: 2966 G1CollectedHeap* _g1h; 2967 size_t _live_bytes; 2968 HeapRegion *_hr; 2969 VerifyOption _vo; 2970 public: 2971 // _vo == UsePrevMarking -> use "prev" marking information, 2972 // _vo == UseNextMarking -> use "next" marking information, 2973 // _vo == UseMarkWord -> use mark word from object header. 2974 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 2975 : _live_bytes(0), _hr(hr), _vo(vo) { 2976 _g1h = G1CollectedHeap::heap(); 2977 } 2978 void do_object(oop o) { 2979 VerifyLivenessOopClosure isLive(_g1h, _vo); 2980 assert(o != NULL, "Huh?"); 2981 if (!_g1h->is_obj_dead_cond(o, _vo)) { 2982 // If the object is alive according to the mark word, 2983 // then verify that the marking information agrees. 2984 // Note we can't verify the contra-positive of the 2985 // above: if the object is dead (according to the mark 2986 // word), it may not be marked, or may have been marked 2987 // but has since became dead, or may have been allocated 2988 // since the last marking. 2989 if (_vo == VerifyOption_G1UseMarkWord) { 2990 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 2991 } 2992 2993 o->oop_iterate(&isLive); 2994 if (!_hr->obj_allocated_since_prev_marking(o)) { 2995 size_t obj_size = o->size(); // Make sure we don't overflow 2996 _live_bytes += (obj_size * HeapWordSize); 2997 } 2998 } 2999 } 3000 size_t live_bytes() { return _live_bytes; } 3001 }; 3002 3003 class PrintObjsInRegionClosure : public ObjectClosure { 3004 HeapRegion *_hr; 3005 G1CollectedHeap *_g1; 3006 public: 3007 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3008 _g1 = G1CollectedHeap::heap(); 3009 }; 3010 3011 void do_object(oop o) { 3012 if (o != NULL) { 3013 HeapWord *start = (HeapWord *) o; 3014 size_t word_sz = o->size(); 3015 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3016 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3017 (void*) o, word_sz, 3018 _g1->isMarkedPrev(o), 3019 _g1->isMarkedNext(o), 3020 _hr->obj_allocated_since_prev_marking(o)); 3021 HeapWord *end = start + word_sz; 3022 HeapWord *cur; 3023 int *val; 3024 for (cur = start; cur < end; cur++) { 3025 val = (int *) cur; 3026 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3027 } 3028 } 3029 } 3030 }; 3031 3032 class VerifyRegionClosure: public HeapRegionClosure { 3033 private: 3034 bool _par; 3035 VerifyOption _vo; 3036 bool _failures; 3037 public: 3038 // _vo == UsePrevMarking -> use "prev" marking information, 3039 // _vo == UseNextMarking -> use "next" marking information, 3040 // _vo == UseMarkWord -> use mark word from object header. 3041 VerifyRegionClosure(bool par, VerifyOption vo) 3042 : _par(par), 3043 _vo(vo), 3044 _failures(false) {} 3045 3046 bool failures() { 3047 return _failures; 3048 } 3049 3050 bool doHeapRegion(HeapRegion* r) { 3051 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue, 3052 "Should be unclaimed at verify points."); 3053 if (!r->continuesHumongous()) { 3054 bool failures = false; 3055 r->verify(_vo, &failures); 3056 if (failures) { 3057 _failures = true; 3058 } else { 3059 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3060 r->object_iterate(¬_dead_yet_cl); 3061 if (_vo != VerifyOption_G1UseNextMarking) { 3062 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3063 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3064 "max_live_bytes "SIZE_FORMAT" " 3065 "< calculated "SIZE_FORMAT, 3066 r->bottom(), r->end(), 3067 r->max_live_bytes(), 3068 not_dead_yet_cl.live_bytes()); 3069 _failures = true; 3070 } 3071 } else { 3072 // When vo == UseNextMarking we cannot currently do a sanity 3073 // check on the live bytes as the calculation has not been 3074 // finalized yet. 3075 } 3076 } 3077 } 3078 return false; // stop the region iteration if we hit a failure 3079 } 3080 }; 3081 3082 class VerifyRootsClosure: public OopsInGenClosure { 3083 private: 3084 G1CollectedHeap* _g1h; 3085 VerifyOption _vo; 3086 bool _failures; 3087 public: 3088 // _vo == UsePrevMarking -> use "prev" marking information, 3089 // _vo == UseNextMarking -> use "next" marking information, 3090 // _vo == UseMarkWord -> use mark word from object header. 3091 VerifyRootsClosure(VerifyOption vo) : 3092 _g1h(G1CollectedHeap::heap()), 3093 _vo(vo), 3094 _failures(false) { } 3095 3096 bool failures() { return _failures; } 3097 3098 template <class T> void do_oop_nv(T* p) { 3099 T heap_oop = oopDesc::load_heap_oop(p); 3100 if (!oopDesc::is_null(heap_oop)) { 3101 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3102 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3103 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3104 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3105 if (_vo == VerifyOption_G1UseMarkWord) { 3106 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3107 } 3108 obj->print_on(gclog_or_tty); 3109 _failures = true; 3110 } 3111 } 3112 } 3113 3114 void do_oop(oop* p) { do_oop_nv(p); } 3115 void do_oop(narrowOop* p) { do_oop_nv(p); } 3116 }; 3117 3118 // This is the task used for parallel heap verification. 3119 3120 class G1ParVerifyTask: public AbstractGangTask { 3121 private: 3122 G1CollectedHeap* _g1h; 3123 VerifyOption _vo; 3124 bool _failures; 3125 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 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3131 AbstractGangTask("Parallel verify task"), 3132 _g1h(g1h), 3133 _vo(vo), 3134 _failures(false) { } 3135 3136 bool failures() { 3137 return _failures; 3138 } 3139 3140 void work(uint worker_id) { 3141 HandleMark hm; 3142 VerifyRegionClosure blk(true, _vo); 3143 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3144 _g1h->workers()->active_workers(), 3145 HeapRegion::ParVerifyClaimValue); 3146 if (blk.failures()) { 3147 _failures = true; 3148 } 3149 } 3150 }; 3151 3152 void G1CollectedHeap::verify(bool silent) { 3153 verify(silent, VerifyOption_G1UsePrevMarking); 3154 } 3155 3156 void G1CollectedHeap::verify(bool silent, 3157 VerifyOption vo) { 3158 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3159 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3160 VerifyRootsClosure rootsCl(vo); 3161 3162 assert(Thread::current()->is_VM_thread(), 3163 "Expected to be executed serially by the VM thread at this point"); 3164 3165 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3166 3167 // We apply the relevant closures to all the oops in the 3168 // system dictionary, the string table and the code cache. 3169 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3170 3171 process_strong_roots(true, // activate StrongRootsScope 3172 true, // we set "collecting perm gen" to true, 3173 // so we don't reset the dirty cards in the perm gen. 3174 ScanningOption(so), // roots scanning options 3175 &rootsCl, 3176 &blobsCl, 3177 &rootsCl); 3178 3179 // If we're verifying after the marking phase of a Full GC then we can't 3180 // treat the perm gen as roots into the G1 heap. Some of the objects in 3181 // the perm gen may be dead and hence not marked. If one of these dead 3182 // objects is considered to be a root then we may end up with a false 3183 // "Root location <x> points to dead ob <y>" failure. 3184 if (vo != VerifyOption_G1UseMarkWord) { 3185 // Since we used "collecting_perm_gen" == true above, we will not have 3186 // checked the refs from perm into the G1-collected heap. We check those 3187 // references explicitly below. Whether the relevant cards are dirty 3188 // is checked further below in the rem set verification. 3189 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3190 perm_gen()->oop_iterate(&rootsCl); 3191 } 3192 bool failures = rootsCl.failures(); 3193 3194 if (vo != VerifyOption_G1UseMarkWord) { 3195 // If we're verifying during a full GC then the region sets 3196 // will have been torn down at the start of the GC. Therefore 3197 // verifying the region sets will fail. So we only verify 3198 // the region sets when not in a full GC. 3199 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3200 verify_region_sets(); 3201 } 3202 3203 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3204 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3205 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3206 "sanity check"); 3207 3208 G1ParVerifyTask task(this, vo); 3209 assert(UseDynamicNumberOfGCThreads || 3210 workers()->active_workers() == workers()->total_workers(), 3211 "If not dynamic should be using all the workers"); 3212 int n_workers = workers()->active_workers(); 3213 set_par_threads(n_workers); 3214 workers()->run_task(&task); 3215 set_par_threads(0); 3216 if (task.failures()) { 3217 failures = true; 3218 } 3219 3220 // Checks that the expected amount of parallel work was done. 3221 // The implication is that n_workers is > 0. 3222 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3223 "sanity check"); 3224 3225 reset_heap_region_claim_values(); 3226 3227 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3228 "sanity check"); 3229 } else { 3230 VerifyRegionClosure blk(false, vo); 3231 heap_region_iterate(&blk); 3232 if (blk.failures()) { 3233 failures = true; 3234 } 3235 } 3236 if (!silent) gclog_or_tty->print("RemSet "); 3237 rem_set()->verify(); 3238 3239 if (failures) { 3240 gclog_or_tty->print_cr("Heap:"); 3241 // It helps to have the per-region information in the output to 3242 // help us track down what went wrong. This is why we call 3243 // print_extended_on() instead of print_on(). 3244 print_extended_on(gclog_or_tty); 3245 gclog_or_tty->print_cr(""); 3246 #ifndef PRODUCT 3247 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3248 concurrent_mark()->print_reachable("at-verification-failure", 3249 vo, false /* all */); 3250 } 3251 #endif 3252 gclog_or_tty->flush(); 3253 } 3254 guarantee(!failures, "there should not have been any failures"); 3255 } else { 3256 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3257 } 3258 } 3259 3260 class PrintRegionClosure: public HeapRegionClosure { 3261 outputStream* _st; 3262 public: 3263 PrintRegionClosure(outputStream* st) : _st(st) {} 3264 bool doHeapRegion(HeapRegion* r) { 3265 r->print_on(_st); 3266 return false; 3267 } 3268 }; 3269 3270 void G1CollectedHeap::print_on(outputStream* st) const { 3271 st->print(" %-20s", "garbage-first heap"); 3272 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3273 capacity()/K, used_unlocked()/K); 3274 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3275 _g1_storage.low_boundary(), 3276 _g1_storage.high(), 3277 _g1_storage.high_boundary()); 3278 st->cr(); 3279 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3280 uint young_regions = _young_list->length(); 3281 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3282 (size_t) young_regions * HeapRegion::GrainBytes / K); 3283 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3284 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3285 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3286 st->cr(); 3287 perm()->as_gen()->print_on(st); 3288 } 3289 3290 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3291 print_on(st); 3292 3293 // Print the per-region information. 3294 st->cr(); 3295 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3296 "HS=humongous(starts), HC=humongous(continues), " 3297 "CS=collection set, F=free, TS=gc time stamp, " 3298 "PTAMS=previous top-at-mark-start, " 3299 "NTAMS=next top-at-mark-start)"); 3300 PrintRegionClosure blk(st); 3301 heap_region_iterate(&blk); 3302 } 3303 3304 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3305 if (G1CollectedHeap::use_parallel_gc_threads()) { 3306 workers()->print_worker_threads_on(st); 3307 } 3308 _cmThread->print_on(st); 3309 st->cr(); 3310 _cm->print_worker_threads_on(st); 3311 _cg1r->print_worker_threads_on(st); 3312 st->cr(); 3313 } 3314 3315 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3316 if (G1CollectedHeap::use_parallel_gc_threads()) { 3317 workers()->threads_do(tc); 3318 } 3319 tc->do_thread(_cmThread); 3320 _cg1r->threads_do(tc); 3321 } 3322 3323 void G1CollectedHeap::print_tracing_info() const { 3324 // We'll overload this to mean "trace GC pause statistics." 3325 if (TraceGen0Time || TraceGen1Time) { 3326 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3327 // to that. 3328 g1_policy()->print_tracing_info(); 3329 } 3330 if (G1SummarizeRSetStats) { 3331 g1_rem_set()->print_summary_info(); 3332 } 3333 if (G1SummarizeConcMark) { 3334 concurrent_mark()->print_summary_info(); 3335 } 3336 g1_policy()->print_yg_surv_rate_info(); 3337 SpecializationStats::print(); 3338 } 3339 3340 #ifndef PRODUCT 3341 // Helpful for debugging RSet issues. 3342 3343 class PrintRSetsClosure : public HeapRegionClosure { 3344 private: 3345 const char* _msg; 3346 size_t _occupied_sum; 3347 3348 public: 3349 bool doHeapRegion(HeapRegion* r) { 3350 HeapRegionRemSet* hrrs = r->rem_set(); 3351 size_t occupied = hrrs->occupied(); 3352 _occupied_sum += occupied; 3353 3354 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3355 HR_FORMAT_PARAMS(r)); 3356 if (occupied == 0) { 3357 gclog_or_tty->print_cr(" RSet is empty"); 3358 } else { 3359 hrrs->print(); 3360 } 3361 gclog_or_tty->print_cr("----------"); 3362 return false; 3363 } 3364 3365 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3366 gclog_or_tty->cr(); 3367 gclog_or_tty->print_cr("========================================"); 3368 gclog_or_tty->print_cr(msg); 3369 gclog_or_tty->cr(); 3370 } 3371 3372 ~PrintRSetsClosure() { 3373 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3374 gclog_or_tty->print_cr("========================================"); 3375 gclog_or_tty->cr(); 3376 } 3377 }; 3378 3379 void G1CollectedHeap::print_cset_rsets() { 3380 PrintRSetsClosure cl("Printing CSet RSets"); 3381 collection_set_iterate(&cl); 3382 } 3383 3384 void G1CollectedHeap::print_all_rsets() { 3385 PrintRSetsClosure cl("Printing All RSets");; 3386 heap_region_iterate(&cl); 3387 } 3388 #endif // PRODUCT 3389 3390 G1CollectedHeap* G1CollectedHeap::heap() { 3391 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3392 "not a garbage-first heap"); 3393 return _g1h; 3394 } 3395 3396 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3397 // always_do_update_barrier = false; 3398 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3399 // Call allocation profiler 3400 AllocationProfiler::iterate_since_last_gc(); 3401 // Fill TLAB's and such 3402 ensure_parsability(true); 3403 } 3404 3405 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3406 // FIXME: what is this about? 3407 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3408 // is set. 3409 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3410 "derived pointer present")); 3411 // always_do_update_barrier = true; 3412 3413 // We have just completed a GC. Update the soft reference 3414 // policy with the new heap occupancy 3415 Universe::update_heap_info_at_gc(); 3416 } 3417 3418 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3419 unsigned int gc_count_before, 3420 bool* succeeded) { 3421 assert_heap_not_locked_and_not_at_safepoint(); 3422 g1_policy()->record_stop_world_start(); 3423 VM_G1IncCollectionPause op(gc_count_before, 3424 word_size, 3425 false, /* should_initiate_conc_mark */ 3426 g1_policy()->max_pause_time_ms(), 3427 GCCause::_g1_inc_collection_pause); 3428 VMThread::execute(&op); 3429 3430 HeapWord* result = op.result(); 3431 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3432 assert(result == NULL || ret_succeeded, 3433 "the result should be NULL if the VM did not succeed"); 3434 *succeeded = ret_succeeded; 3435 3436 assert_heap_not_locked(); 3437 return result; 3438 } 3439 3440 void 3441 G1CollectedHeap::doConcurrentMark() { 3442 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3443 if (!_cmThread->in_progress()) { 3444 _cmThread->set_started(); 3445 CGC_lock->notify(); 3446 } 3447 } 3448 3449 size_t G1CollectedHeap::pending_card_num() { 3450 size_t extra_cards = 0; 3451 JavaThread *curr = Threads::first(); 3452 while (curr != NULL) { 3453 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3454 extra_cards += dcq.size(); 3455 curr = curr->next(); 3456 } 3457 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3458 size_t buffer_size = dcqs.buffer_size(); 3459 size_t buffer_num = dcqs.completed_buffers_num(); 3460 return buffer_size * buffer_num + extra_cards; 3461 } 3462 3463 size_t G1CollectedHeap::max_pending_card_num() { 3464 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3465 size_t buffer_size = dcqs.buffer_size(); 3466 size_t buffer_num = dcqs.completed_buffers_num(); 3467 int thread_num = Threads::number_of_threads(); 3468 return (buffer_num + thread_num) * buffer_size; 3469 } 3470 3471 size_t G1CollectedHeap::cards_scanned() { 3472 return g1_rem_set()->cardsScanned(); 3473 } 3474 3475 void 3476 G1CollectedHeap::setup_surviving_young_words() { 3477 assert(_surviving_young_words == NULL, "pre-condition"); 3478 uint array_length = g1_policy()->young_cset_region_length(); 3479 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length); 3480 if (_surviving_young_words == NULL) { 3481 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3482 "Not enough space for young surv words summary."); 3483 } 3484 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3485 #ifdef ASSERT 3486 for (uint i = 0; i < array_length; ++i) { 3487 assert( _surviving_young_words[i] == 0, "memset above" ); 3488 } 3489 #endif // !ASSERT 3490 } 3491 3492 void 3493 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3494 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3495 uint array_length = g1_policy()->young_cset_region_length(); 3496 for (uint i = 0; i < array_length; ++i) { 3497 _surviving_young_words[i] += surv_young_words[i]; 3498 } 3499 } 3500 3501 void 3502 G1CollectedHeap::cleanup_surviving_young_words() { 3503 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3504 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words); 3505 _surviving_young_words = NULL; 3506 } 3507 3508 #ifdef ASSERT 3509 class VerifyCSetClosure: public HeapRegionClosure { 3510 public: 3511 bool doHeapRegion(HeapRegion* hr) { 3512 // Here we check that the CSet region's RSet is ready for parallel 3513 // iteration. The fields that we'll verify are only manipulated 3514 // when the region is part of a CSet and is collected. Afterwards, 3515 // we reset these fields when we clear the region's RSet (when the 3516 // region is freed) so they are ready when the region is 3517 // re-allocated. The only exception to this is if there's an 3518 // evacuation failure and instead of freeing the region we leave 3519 // it in the heap. In that case, we reset these fields during 3520 // evacuation failure handling. 3521 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3522 3523 // Here's a good place to add any other checks we'd like to 3524 // perform on CSet regions. 3525 return false; 3526 } 3527 }; 3528 #endif // ASSERT 3529 3530 #if TASKQUEUE_STATS 3531 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3532 st->print_raw_cr("GC Task Stats"); 3533 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3534 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3535 } 3536 3537 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3538 print_taskqueue_stats_hdr(st); 3539 3540 TaskQueueStats totals; 3541 const int n = workers() != NULL ? workers()->total_workers() : 1; 3542 for (int i = 0; i < n; ++i) { 3543 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3544 totals += task_queue(i)->stats; 3545 } 3546 st->print_raw("tot "); totals.print(st); st->cr(); 3547 3548 DEBUG_ONLY(totals.verify()); 3549 } 3550 3551 void G1CollectedHeap::reset_taskqueue_stats() { 3552 const int n = workers() != NULL ? workers()->total_workers() : 1; 3553 for (int i = 0; i < n; ++i) { 3554 task_queue(i)->stats.reset(); 3555 } 3556 } 3557 #endif // TASKQUEUE_STATS 3558 3559 bool 3560 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3561 assert_at_safepoint(true /* should_be_vm_thread */); 3562 guarantee(!is_gc_active(), "collection is not reentrant"); 3563 3564 if (GC_locker::check_active_before_gc()) { 3565 return false; 3566 } 3567 3568 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3569 ResourceMark rm; 3570 3571 print_heap_before_gc(); 3572 3573 HRSPhaseSetter x(HRSPhaseEvacuation); 3574 verify_region_sets_optional(); 3575 verify_dirty_young_regions(); 3576 3577 // This call will decide whether this pause is an initial-mark 3578 // pause. If it is, during_initial_mark_pause() will return true 3579 // for the duration of this pause. 3580 g1_policy()->decide_on_conc_mark_initiation(); 3581 3582 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3583 assert(!g1_policy()->during_initial_mark_pause() || 3584 g1_policy()->gcs_are_young(), "sanity"); 3585 3586 // We also do not allow mixed GCs during marking. 3587 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3588 3589 // Record whether this pause is an initial mark. When the current 3590 // thread has completed its logging output and it's safe to signal 3591 // the CM thread, the flag's value in the policy has been reset. 3592 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3593 3594 // Inner scope for scope based logging, timers, and stats collection 3595 { 3596 char verbose_str[128]; 3597 sprintf(verbose_str, "GC pause "); 3598 if (g1_policy()->gcs_are_young()) { 3599 strcat(verbose_str, "(young)"); 3600 } else { 3601 strcat(verbose_str, "(mixed)"); 3602 } 3603 if (g1_policy()->during_initial_mark_pause()) { 3604 strcat(verbose_str, " (initial-mark)"); 3605 // We are about to start a marking cycle, so we increment the 3606 // full collection counter. 3607 increment_total_full_collections(); 3608 } 3609 3610 // if the log level is "finer" is on, we'll print long statistics information 3611 // in the collector policy code, so let's not print this as the output 3612 // is messy if we do. 3613 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 3614 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3615 TraceTime t(verbose_str, G1Log::fine() && !G1Log::finer(), true, gclog_or_tty); 3616 3617 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3618 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3619 3620 // If the secondary_free_list is not empty, append it to the 3621 // free_list. No need to wait for the cleanup operation to finish; 3622 // the region allocation code will check the secondary_free_list 3623 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3624 // set, skip this step so that the region allocation code has to 3625 // get entries from the secondary_free_list. 3626 if (!G1StressConcRegionFreeing) { 3627 append_secondary_free_list_if_not_empty_with_lock(); 3628 } 3629 3630 assert(check_young_list_well_formed(), 3631 "young list should be well formed"); 3632 3633 // Don't dynamically change the number of GC threads this early. A value of 3634 // 0 is used to indicate serial work. When parallel work is done, 3635 // it will be set. 3636 3637 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3638 IsGCActiveMark x; 3639 3640 gc_prologue(false); 3641 increment_total_collections(false /* full gc */); 3642 increment_gc_time_stamp(); 3643 3644 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3645 HandleMark hm; // Discard invalid handles created during verification 3646 gclog_or_tty->print(" VerifyBeforeGC:"); 3647 prepare_for_verify(); 3648 Universe::verify(/* silent */ false, 3649 /* option */ VerifyOption_G1UsePrevMarking); 3650 } 3651 3652 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3653 3654 // Please see comment in g1CollectedHeap.hpp and 3655 // G1CollectedHeap::ref_processing_init() to see how 3656 // reference processing currently works in G1. 3657 3658 // Enable discovery in the STW reference processor 3659 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3660 true /*verify_no_refs*/); 3661 3662 { 3663 // We want to temporarily turn off discovery by the 3664 // CM ref processor, if necessary, and turn it back on 3665 // on again later if we do. Using a scoped 3666 // NoRefDiscovery object will do this. 3667 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3668 3669 // Forget the current alloc region (we might even choose it to be part 3670 // of the collection set!). 3671 release_mutator_alloc_region(); 3672 3673 // We should call this after we retire the mutator alloc 3674 // region(s) so that all the ALLOC / RETIRE events are generated 3675 // before the start GC event. 3676 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3677 3678 // The elapsed time induced by the start time below deliberately elides 3679 // the possible verification above. 3680 double start_time_sec = os::elapsedTime(); 3681 size_t start_used_bytes = used(); 3682 3683 #if YOUNG_LIST_VERBOSE 3684 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3685 _young_list->print(); 3686 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3687 #endif // YOUNG_LIST_VERBOSE 3688 3689 g1_policy()->record_collection_pause_start(start_time_sec, 3690 start_used_bytes); 3691 3692 double scan_wait_start = os::elapsedTime(); 3693 // We have to wait until the CM threads finish scanning the 3694 // root regions as it's the only way to ensure that all the 3695 // objects on them have been correctly scanned before we start 3696 // moving them during the GC. 3697 bool waited = _cm->root_regions()->wait_until_scan_finished(); 3698 if (waited) { 3699 double scan_wait_end = os::elapsedTime(); 3700 double wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 3701 g1_policy()->record_root_region_scan_wait_time(wait_time_ms); 3702 } 3703 3704 #if YOUNG_LIST_VERBOSE 3705 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3706 _young_list->print(); 3707 #endif // YOUNG_LIST_VERBOSE 3708 3709 if (g1_policy()->during_initial_mark_pause()) { 3710 concurrent_mark()->checkpointRootsInitialPre(); 3711 } 3712 perm_gen()->save_marks(); 3713 3714 #if YOUNG_LIST_VERBOSE 3715 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3716 _young_list->print(); 3717 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3718 #endif // YOUNG_LIST_VERBOSE 3719 3720 g1_policy()->finalize_cset(target_pause_time_ms); 3721 3722 _cm->note_start_of_gc(); 3723 // We should not verify the per-thread SATB buffers given that 3724 // we have not filtered them yet (we'll do so during the 3725 // GC). We also call this after finalize_cset() to 3726 // ensure that the CSet has been finalized. 3727 _cm->verify_no_cset_oops(true /* verify_stacks */, 3728 true /* verify_enqueued_buffers */, 3729 false /* verify_thread_buffers */, 3730 true /* verify_fingers */); 3731 3732 if (_hr_printer.is_active()) { 3733 HeapRegion* hr = g1_policy()->collection_set(); 3734 while (hr != NULL) { 3735 G1HRPrinter::RegionType type; 3736 if (!hr->is_young()) { 3737 type = G1HRPrinter::Old; 3738 } else if (hr->is_survivor()) { 3739 type = G1HRPrinter::Survivor; 3740 } else { 3741 type = G1HRPrinter::Eden; 3742 } 3743 _hr_printer.cset(hr); 3744 hr = hr->next_in_collection_set(); 3745 } 3746 } 3747 3748 #ifdef ASSERT 3749 VerifyCSetClosure cl; 3750 collection_set_iterate(&cl); 3751 #endif // ASSERT 3752 3753 setup_surviving_young_words(); 3754 3755 // Initialize the GC alloc regions. 3756 init_gc_alloc_regions(); 3757 3758 // Actually do the work... 3759 evacuate_collection_set(); 3760 3761 // We do this to mainly verify the per-thread SATB buffers 3762 // (which have been filtered by now) since we didn't verify 3763 // them earlier. No point in re-checking the stacks / enqueued 3764 // buffers given that the CSet has not changed since last time 3765 // we checked. 3766 _cm->verify_no_cset_oops(false /* verify_stacks */, 3767 false /* verify_enqueued_buffers */, 3768 true /* verify_thread_buffers */, 3769 true /* verify_fingers */); 3770 3771 free_collection_set(g1_policy()->collection_set()); 3772 g1_policy()->clear_collection_set(); 3773 3774 cleanup_surviving_young_words(); 3775 3776 // Start a new incremental collection set for the next pause. 3777 g1_policy()->start_incremental_cset_building(); 3778 3779 // Clear the _cset_fast_test bitmap in anticipation of adding 3780 // regions to the incremental collection set for the next 3781 // evacuation pause. 3782 clear_cset_fast_test(); 3783 3784 _young_list->reset_sampled_info(); 3785 3786 // Don't check the whole heap at this point as the 3787 // GC alloc regions from this pause have been tagged 3788 // as survivors and moved on to the survivor list. 3789 // Survivor regions will fail the !is_young() check. 3790 assert(check_young_list_empty(false /* check_heap */), 3791 "young list should be empty"); 3792 3793 #if YOUNG_LIST_VERBOSE 3794 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3795 _young_list->print(); 3796 #endif // YOUNG_LIST_VERBOSE 3797 3798 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3799 _young_list->first_survivor_region(), 3800 _young_list->last_survivor_region()); 3801 3802 _young_list->reset_auxilary_lists(); 3803 3804 if (evacuation_failed()) { 3805 _summary_bytes_used = recalculate_used(); 3806 } else { 3807 // The "used" of the the collection set have already been subtracted 3808 // when they were freed. Add in the bytes evacuated. 3809 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3810 } 3811 3812 if (g1_policy()->during_initial_mark_pause()) { 3813 // We have to do this before we notify the CM threads that 3814 // they can start working to make sure that all the 3815 // appropriate initialization is done on the CM object. 3816 concurrent_mark()->checkpointRootsInitialPost(); 3817 set_marking_started(); 3818 // Note that we don't actually trigger the CM thread at 3819 // this point. We do that later when we're sure that 3820 // the current thread has completed its logging output. 3821 } 3822 3823 allocate_dummy_regions(); 3824 3825 #if YOUNG_LIST_VERBOSE 3826 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3827 _young_list->print(); 3828 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3829 #endif // YOUNG_LIST_VERBOSE 3830 3831 init_mutator_alloc_region(); 3832 3833 { 3834 size_t expand_bytes = g1_policy()->expansion_amount(); 3835 if (expand_bytes > 0) { 3836 size_t bytes_before = capacity(); 3837 // No need for an ergo verbose message here, 3838 // expansion_amount() does this when it returns a value > 0. 3839 if (!expand(expand_bytes)) { 3840 // We failed to expand the heap so let's verify that 3841 // committed/uncommitted amount match the backing store 3842 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3843 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3844 } 3845 } 3846 } 3847 3848 // We redo the verificaiton but now wrt to the new CSet which 3849 // has just got initialized after the previous CSet was freed. 3850 _cm->verify_no_cset_oops(true /* verify_stacks */, 3851 true /* verify_enqueued_buffers */, 3852 true /* verify_thread_buffers */, 3853 true /* verify_fingers */); 3854 _cm->note_end_of_gc(); 3855 3856 double end_time_sec = os::elapsedTime(); 3857 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS; 3858 g1_policy()->record_pause_time_ms(pause_time_ms); 3859 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3860 workers()->active_workers() : 1); 3861 g1_policy()->record_collection_pause_end(active_workers); 3862 3863 MemoryService::track_memory_usage(); 3864 3865 // In prepare_for_verify() below we'll need to scan the deferred 3866 // update buffers to bring the RSets up-to-date if 3867 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3868 // the update buffers we'll probably need to scan cards on the 3869 // regions we just allocated to (i.e., the GC alloc 3870 // regions). However, during the last GC we called 3871 // set_saved_mark() on all the GC alloc regions, so card 3872 // scanning might skip the [saved_mark_word()...top()] area of 3873 // those regions (i.e., the area we allocated objects into 3874 // during the last GC). But it shouldn't. Given that 3875 // saved_mark_word() is conditional on whether the GC time stamp 3876 // on the region is current or not, by incrementing the GC time 3877 // stamp here we invalidate all the GC time stamps on all the 3878 // regions and saved_mark_word() will simply return top() for 3879 // all the regions. This is a nicer way of ensuring this rather 3880 // than iterating over the regions and fixing them. In fact, the 3881 // GC time stamp increment here also ensures that 3882 // saved_mark_word() will return top() between pauses, i.e., 3883 // during concurrent refinement. So we don't need the 3884 // is_gc_active() check to decided which top to use when 3885 // scanning cards (see CR 7039627). 3886 increment_gc_time_stamp(); 3887 3888 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 3889 HandleMark hm; // Discard invalid handles created during verification 3890 gclog_or_tty->print(" VerifyAfterGC:"); 3891 prepare_for_verify(); 3892 Universe::verify(/* silent */ false, 3893 /* option */ VerifyOption_G1UsePrevMarking); 3894 } 3895 3896 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 3897 ref_processor_stw()->verify_no_references_recorded(); 3898 3899 // CM reference discovery will be re-enabled if necessary. 3900 } 3901 3902 // We should do this after we potentially expand the heap so 3903 // that all the COMMIT events are generated before the end GC 3904 // event, and after we retire the GC alloc regions so that all 3905 // RETIRE events are generated before the end GC event. 3906 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 3907 3908 // We have to do this after we decide whether to expand the heap or not. 3909 g1_policy()->print_heap_transition(); 3910 3911 if (mark_in_progress()) { 3912 concurrent_mark()->update_g1_committed(); 3913 } 3914 3915 #ifdef TRACESPINNING 3916 ParallelTaskTerminator::print_termination_counts(); 3917 #endif 3918 3919 gc_epilogue(false); 3920 } 3921 3922 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) { 3923 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum); 3924 print_tracing_info(); 3925 vm_exit(-1); 3926 } 3927 } 3928 3929 // The closing of the inner scope, immediately above, will complete 3930 // logging at the "fine" level. The record_collection_pause_end() call 3931 // above will complete logging at the "finer" level. 3932 // 3933 // It is not yet to safe, however, to tell the concurrent mark to 3934 // start as we have some optional output below. We don't want the 3935 // output from the concurrent mark thread interfering with this 3936 // logging output either. 3937 3938 _hrs.verify_optional(); 3939 verify_region_sets_optional(); 3940 3941 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 3942 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 3943 3944 print_heap_after_gc(); 3945 g1mm()->update_sizes(); 3946 3947 if (G1SummarizeRSetStats && 3948 (G1SummarizeRSetStatsPeriod > 0) && 3949 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3950 g1_rem_set()->print_summary_info(); 3951 } 3952 3953 // It should now be safe to tell the concurrent mark thread to start 3954 // without its logging output interfering with the logging output 3955 // that came from the pause. 3956 3957 if (should_start_conc_mark) { 3958 // CAUTION: after the doConcurrentMark() call below, 3959 // the concurrent marking thread(s) could be running 3960 // concurrently with us. Make sure that anything after 3961 // this point does not assume that we are the only GC thread 3962 // running. Note: of course, the actual marking work will 3963 // not start until the safepoint itself is released in 3964 // ConcurrentGCThread::safepoint_desynchronize(). 3965 doConcurrentMark(); 3966 } 3967 3968 return true; 3969 } 3970 3971 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 3972 { 3973 size_t gclab_word_size; 3974 switch (purpose) { 3975 case GCAllocForSurvived: 3976 gclab_word_size = YoungPLABSize; 3977 break; 3978 case GCAllocForTenured: 3979 gclab_word_size = OldPLABSize; 3980 break; 3981 default: 3982 assert(false, "unknown GCAllocPurpose"); 3983 gclab_word_size = OldPLABSize; 3984 break; 3985 } 3986 return gclab_word_size; 3987 } 3988 3989 void G1CollectedHeap::init_mutator_alloc_region() { 3990 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 3991 _mutator_alloc_region.init(); 3992 } 3993 3994 void G1CollectedHeap::release_mutator_alloc_region() { 3995 _mutator_alloc_region.release(); 3996 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 3997 } 3998 3999 void G1CollectedHeap::init_gc_alloc_regions() { 4000 assert_at_safepoint(true /* should_be_vm_thread */); 4001 4002 _survivor_gc_alloc_region.init(); 4003 _old_gc_alloc_region.init(); 4004 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4005 _retained_old_gc_alloc_region = NULL; 4006 4007 // We will discard the current GC alloc region if: 4008 // a) it's in the collection set (it can happen!), 4009 // b) it's already full (no point in using it), 4010 // c) it's empty (this means that it was emptied during 4011 // a cleanup and it should be on the free list now), or 4012 // d) it's humongous (this means that it was emptied 4013 // during a cleanup and was added to the free list, but 4014 // has been subseqently used to allocate a humongous 4015 // object that may be less than the region size). 4016 if (retained_region != NULL && 4017 !retained_region->in_collection_set() && 4018 !(retained_region->top() == retained_region->end()) && 4019 !retained_region->is_empty() && 4020 !retained_region->isHumongous()) { 4021 retained_region->set_saved_mark(); 4022 // The retained region was added to the old region set when it was 4023 // retired. We have to remove it now, since we don't allow regions 4024 // we allocate to in the region sets. We'll re-add it later, when 4025 // it's retired again. 4026 _old_set.remove(retained_region); 4027 bool during_im = g1_policy()->during_initial_mark_pause(); 4028 retained_region->note_start_of_copying(during_im); 4029 _old_gc_alloc_region.set(retained_region); 4030 _hr_printer.reuse(retained_region); 4031 } 4032 } 4033 4034 void G1CollectedHeap::release_gc_alloc_regions() { 4035 _survivor_gc_alloc_region.release(); 4036 // If we have an old GC alloc region to release, we'll save it in 4037 // _retained_old_gc_alloc_region. If we don't 4038 // _retained_old_gc_alloc_region will become NULL. This is what we 4039 // want either way so no reason to check explicitly for either 4040 // condition. 4041 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4042 } 4043 4044 void G1CollectedHeap::abandon_gc_alloc_regions() { 4045 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4046 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4047 _retained_old_gc_alloc_region = NULL; 4048 } 4049 4050 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4051 _drain_in_progress = false; 4052 set_evac_failure_closure(cl); 4053 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4054 } 4055 4056 void G1CollectedHeap::finalize_for_evac_failure() { 4057 assert(_evac_failure_scan_stack != NULL && 4058 _evac_failure_scan_stack->length() == 0, 4059 "Postcondition"); 4060 assert(!_drain_in_progress, "Postcondition"); 4061 delete _evac_failure_scan_stack; 4062 _evac_failure_scan_stack = NULL; 4063 } 4064 4065 void G1CollectedHeap::remove_self_forwarding_pointers() { 4066 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4067 4068 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4069 4070 if (G1CollectedHeap::use_parallel_gc_threads()) { 4071 set_par_threads(); 4072 workers()->run_task(&rsfp_task); 4073 set_par_threads(0); 4074 } else { 4075 rsfp_task.work(0); 4076 } 4077 4078 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4079 4080 // Reset the claim values in the regions in the collection set. 4081 reset_cset_heap_region_claim_values(); 4082 4083 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4084 4085 // Now restore saved marks, if any. 4086 if (_objs_with_preserved_marks != NULL) { 4087 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4088 guarantee(_objs_with_preserved_marks->length() == 4089 _preserved_marks_of_objs->length(), "Both or none."); 4090 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4091 oop obj = _objs_with_preserved_marks->at(i); 4092 markOop m = _preserved_marks_of_objs->at(i); 4093 obj->set_mark(m); 4094 } 4095 4096 // Delete the preserved marks growable arrays (allocated on the C heap). 4097 delete _objs_with_preserved_marks; 4098 delete _preserved_marks_of_objs; 4099 _objs_with_preserved_marks = NULL; 4100 _preserved_marks_of_objs = NULL; 4101 } 4102 } 4103 4104 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4105 _evac_failure_scan_stack->push(obj); 4106 } 4107 4108 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4109 assert(_evac_failure_scan_stack != NULL, "precondition"); 4110 4111 while (_evac_failure_scan_stack->length() > 0) { 4112 oop obj = _evac_failure_scan_stack->pop(); 4113 _evac_failure_closure->set_region(heap_region_containing(obj)); 4114 obj->oop_iterate_backwards(_evac_failure_closure); 4115 } 4116 } 4117 4118 oop 4119 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4120 oop old) { 4121 assert(obj_in_cs(old), 4122 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4123 (HeapWord*) old)); 4124 markOop m = old->mark(); 4125 oop forward_ptr = old->forward_to_atomic(old); 4126 if (forward_ptr == NULL) { 4127 // Forward-to-self succeeded. 4128 4129 if (_evac_failure_closure != cl) { 4130 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4131 assert(!_drain_in_progress, 4132 "Should only be true while someone holds the lock."); 4133 // Set the global evac-failure closure to the current thread's. 4134 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4135 set_evac_failure_closure(cl); 4136 // Now do the common part. 4137 handle_evacuation_failure_common(old, m); 4138 // Reset to NULL. 4139 set_evac_failure_closure(NULL); 4140 } else { 4141 // The lock is already held, and this is recursive. 4142 assert(_drain_in_progress, "This should only be the recursive case."); 4143 handle_evacuation_failure_common(old, m); 4144 } 4145 return old; 4146 } else { 4147 // Forward-to-self failed. Either someone else managed to allocate 4148 // space for this object (old != forward_ptr) or they beat us in 4149 // self-forwarding it (old == forward_ptr). 4150 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4151 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4152 "should not be in the CSet", 4153 (HeapWord*) old, (HeapWord*) forward_ptr)); 4154 return forward_ptr; 4155 } 4156 } 4157 4158 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4159 set_evacuation_failed(true); 4160 4161 preserve_mark_if_necessary(old, m); 4162 4163 HeapRegion* r = heap_region_containing(old); 4164 if (!r->evacuation_failed()) { 4165 r->set_evacuation_failed(true); 4166 _hr_printer.evac_failure(r); 4167 } 4168 4169 push_on_evac_failure_scan_stack(old); 4170 4171 if (!_drain_in_progress) { 4172 // prevent recursion in copy_to_survivor_space() 4173 _drain_in_progress = true; 4174 drain_evac_failure_scan_stack(); 4175 _drain_in_progress = false; 4176 } 4177 } 4178 4179 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4180 assert(evacuation_failed(), "Oversaving!"); 4181 // We want to call the "for_promotion_failure" version only in the 4182 // case of a promotion failure. 4183 if (m->must_be_preserved_for_promotion_failure(obj)) { 4184 if (_objs_with_preserved_marks == NULL) { 4185 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4186 _objs_with_preserved_marks = 4187 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true); 4188 _preserved_marks_of_objs = 4189 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true); 4190 } 4191 _objs_with_preserved_marks->push(obj); 4192 _preserved_marks_of_objs->push(m); 4193 } 4194 } 4195 4196 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4197 size_t word_size) { 4198 if (purpose == GCAllocForSurvived) { 4199 HeapWord* result = survivor_attempt_allocation(word_size); 4200 if (result != NULL) { 4201 return result; 4202 } else { 4203 // Let's try to allocate in the old gen in case we can fit the 4204 // object there. 4205 return old_attempt_allocation(word_size); 4206 } 4207 } else { 4208 assert(purpose == GCAllocForTenured, "sanity"); 4209 HeapWord* result = old_attempt_allocation(word_size); 4210 if (result != NULL) { 4211 return result; 4212 } else { 4213 // Let's try to allocate in the survivors in case we can fit the 4214 // object there. 4215 return survivor_attempt_allocation(word_size); 4216 } 4217 } 4218 4219 ShouldNotReachHere(); 4220 // Trying to keep some compilers happy. 4221 return NULL; 4222 } 4223 4224 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4225 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4226 4227 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4228 : _g1h(g1h), 4229 _refs(g1h->task_queue(queue_num)), 4230 _dcq(&g1h->dirty_card_queue_set()), 4231 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4232 _g1_rem(g1h->g1_rem_set()), 4233 _hash_seed(17), _queue_num(queue_num), 4234 _term_attempts(0), 4235 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4236 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4237 _age_table(false), 4238 _strong_roots_time(0), _term_time(0), 4239 _alloc_buffer_waste(0), _undo_waste(0) { 4240 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4241 // we "sacrifice" entry 0 to keep track of surviving bytes for 4242 // non-young regions (where the age is -1) 4243 // We also add a few elements at the beginning and at the end in 4244 // an attempt to eliminate cache contention 4245 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4246 uint array_length = PADDING_ELEM_NUM + 4247 real_length + 4248 PADDING_ELEM_NUM; 4249 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length); 4250 if (_surviving_young_words_base == NULL) 4251 vm_exit_out_of_memory(array_length * sizeof(size_t), 4252 "Not enough space for young surv histo."); 4253 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4254 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4255 4256 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4257 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4258 4259 _start = os::elapsedTime(); 4260 } 4261 4262 void 4263 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4264 { 4265 st->print_raw_cr("GC Termination Stats"); 4266 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4267 " ------waste (KiB)------"); 4268 st->print_raw_cr("thr ms ms % ms % attempts" 4269 " total alloc undo"); 4270 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4271 " ------- ------- -------"); 4272 } 4273 4274 void 4275 G1ParScanThreadState::print_termination_stats(int i, 4276 outputStream* const st) const 4277 { 4278 const double elapsed_ms = elapsed_time() * 1000.0; 4279 const double s_roots_ms = strong_roots_time() * 1000.0; 4280 const double term_ms = term_time() * 1000.0; 4281 st->print_cr("%3d %9.2f %9.2f %6.2f " 4282 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4283 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4284 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4285 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4286 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4287 alloc_buffer_waste() * HeapWordSize / K, 4288 undo_waste() * HeapWordSize / K); 4289 } 4290 4291 #ifdef ASSERT 4292 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4293 assert(ref != NULL, "invariant"); 4294 assert(UseCompressedOops, "sanity"); 4295 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4296 oop p = oopDesc::load_decode_heap_oop(ref); 4297 assert(_g1h->is_in_g1_reserved(p), 4298 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4299 return true; 4300 } 4301 4302 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4303 assert(ref != NULL, "invariant"); 4304 if (has_partial_array_mask(ref)) { 4305 // Must be in the collection set--it's already been copied. 4306 oop p = clear_partial_array_mask(ref); 4307 assert(_g1h->obj_in_cs(p), 4308 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4309 } else { 4310 oop p = oopDesc::load_decode_heap_oop(ref); 4311 assert(_g1h->is_in_g1_reserved(p), 4312 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4313 } 4314 return true; 4315 } 4316 4317 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4318 if (ref.is_narrow()) { 4319 return verify_ref((narrowOop*) ref); 4320 } else { 4321 return verify_ref((oop*) ref); 4322 } 4323 } 4324 #endif // ASSERT 4325 4326 void G1ParScanThreadState::trim_queue() { 4327 assert(_evac_cl != NULL, "not set"); 4328 assert(_evac_failure_cl != NULL, "not set"); 4329 assert(_partial_scan_cl != NULL, "not set"); 4330 4331 StarTask ref; 4332 do { 4333 // Drain the overflow stack first, so other threads can steal. 4334 while (refs()->pop_overflow(ref)) { 4335 deal_with_reference(ref); 4336 } 4337 4338 while (refs()->pop_local(ref)) { 4339 deal_with_reference(ref); 4340 } 4341 } while (!refs()->is_empty()); 4342 } 4343 4344 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4345 G1ParScanThreadState* par_scan_state) : 4346 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4347 _par_scan_state(par_scan_state), 4348 _worker_id(par_scan_state->queue_num()), 4349 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4350 _mark_in_progress(_g1->mark_in_progress()) { } 4351 4352 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4353 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4354 #ifdef ASSERT 4355 HeapRegion* hr = _g1->heap_region_containing(obj); 4356 assert(hr != NULL, "sanity"); 4357 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4358 #endif // ASSERT 4359 4360 // We know that the object is not moving so it's safe to read its size. 4361 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4362 } 4363 4364 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4365 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4366 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4367 #ifdef ASSERT 4368 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4369 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4370 assert(from_obj != to_obj, "should not be self-forwarded"); 4371 4372 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4373 assert(from_hr != NULL, "sanity"); 4374 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4375 4376 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4377 assert(to_hr != NULL, "sanity"); 4378 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4379 #endif // ASSERT 4380 4381 // The object might be in the process of being copied by another 4382 // worker so we cannot trust that its to-space image is 4383 // well-formed. So we have to read its size from its from-space 4384 // image which we know should not be changing. 4385 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4386 } 4387 4388 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4389 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4390 ::copy_to_survivor_space(oop old) { 4391 size_t word_sz = old->size(); 4392 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4393 // +1 to make the -1 indexes valid... 4394 int young_index = from_region->young_index_in_cset()+1; 4395 assert( (from_region->is_young() && young_index > 0) || 4396 (!from_region->is_young() && young_index == 0), "invariant" ); 4397 G1CollectorPolicy* g1p = _g1->g1_policy(); 4398 markOop m = old->mark(); 4399 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4400 : m->age(); 4401 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4402 word_sz); 4403 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4404 oop obj = oop(obj_ptr); 4405 4406 if (obj_ptr == NULL) { 4407 // This will either forward-to-self, or detect that someone else has 4408 // installed a forwarding pointer. 4409 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4410 return _g1->handle_evacuation_failure_par(cl, old); 4411 } 4412 4413 // We're going to allocate linearly, so might as well prefetch ahead. 4414 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4415 4416 oop forward_ptr = old->forward_to_atomic(obj); 4417 if (forward_ptr == NULL) { 4418 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4419 if (g1p->track_object_age(alloc_purpose)) { 4420 // We could simply do obj->incr_age(). However, this causes a 4421 // performance issue. obj->incr_age() will first check whether 4422 // the object has a displaced mark by checking its mark word; 4423 // getting the mark word from the new location of the object 4424 // stalls. So, given that we already have the mark word and we 4425 // are about to install it anyway, it's better to increase the 4426 // age on the mark word, when the object does not have a 4427 // displaced mark word. We're not expecting many objects to have 4428 // a displaced marked word, so that case is not optimized 4429 // further (it could be...) and we simply call obj->incr_age(). 4430 4431 if (m->has_displaced_mark_helper()) { 4432 // in this case, we have to install the mark word first, 4433 // otherwise obj looks to be forwarded (the old mark word, 4434 // which contains the forward pointer, was copied) 4435 obj->set_mark(m); 4436 obj->incr_age(); 4437 } else { 4438 m = m->incr_age(); 4439 obj->set_mark(m); 4440 } 4441 _par_scan_state->age_table()->add(obj, word_sz); 4442 } else { 4443 obj->set_mark(m); 4444 } 4445 4446 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4447 surv_young_words[young_index] += word_sz; 4448 4449 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4450 // We keep track of the next start index in the length field of 4451 // the to-space object. The actual length can be found in the 4452 // length field of the from-space object. 4453 arrayOop(obj)->set_length(0); 4454 oop* old_p = set_partial_array_mask(old); 4455 _par_scan_state->push_on_queue(old_p); 4456 } else { 4457 // No point in using the slower heap_region_containing() method, 4458 // given that we know obj is in the heap. 4459 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4460 obj->oop_iterate_backwards(&_scanner); 4461 } 4462 } else { 4463 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4464 obj = forward_ptr; 4465 } 4466 return obj; 4467 } 4468 4469 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4470 template <class T> 4471 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4472 ::do_oop_work(T* p) { 4473 oop obj = oopDesc::load_decode_heap_oop(p); 4474 assert(barrier != G1BarrierRS || obj != NULL, 4475 "Precondition: G1BarrierRS implies obj is non-NULL"); 4476 4477 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4478 4479 // here the null check is implicit in the cset_fast_test() test 4480 if (_g1->in_cset_fast_test(obj)) { 4481 oop forwardee; 4482 if (obj->is_forwarded()) { 4483 forwardee = obj->forwardee(); 4484 } else { 4485 forwardee = copy_to_survivor_space(obj); 4486 } 4487 assert(forwardee != NULL, "forwardee should not be NULL"); 4488 oopDesc::encode_store_heap_oop(p, forwardee); 4489 if (do_mark_object && forwardee != obj) { 4490 // If the object is self-forwarded we don't need to explicitly 4491 // mark it, the evacuation failure protocol will do so. 4492 mark_forwarded_object(obj, forwardee); 4493 } 4494 4495 // When scanning the RS, we only care about objs in CS. 4496 if (barrier == G1BarrierRS) { 4497 _par_scan_state->update_rs(_from, p, _worker_id); 4498 } 4499 } else { 4500 // The object is not in collection set. If we're a root scanning 4501 // closure during an initial mark pause (i.e. do_mark_object will 4502 // be true) then attempt to mark the object. 4503 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4504 mark_object(obj); 4505 } 4506 } 4507 4508 if (barrier == G1BarrierEvac && obj != NULL) { 4509 _par_scan_state->update_rs(_from, p, _worker_id); 4510 } 4511 4512 if (do_gen_barrier && obj != NULL) { 4513 par_do_barrier(p); 4514 } 4515 } 4516 4517 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4518 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4519 4520 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4521 assert(has_partial_array_mask(p), "invariant"); 4522 oop from_obj = clear_partial_array_mask(p); 4523 4524 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4525 assert(from_obj->is_objArray(), "must be obj array"); 4526 objArrayOop from_obj_array = objArrayOop(from_obj); 4527 // The from-space object contains the real length. 4528 int length = from_obj_array->length(); 4529 4530 assert(from_obj->is_forwarded(), "must be forwarded"); 4531 oop to_obj = from_obj->forwardee(); 4532 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4533 objArrayOop to_obj_array = objArrayOop(to_obj); 4534 // We keep track of the next start index in the length field of the 4535 // to-space object. 4536 int next_index = to_obj_array->length(); 4537 assert(0 <= next_index && next_index < length, 4538 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4539 4540 int start = next_index; 4541 int end = length; 4542 int remainder = end - start; 4543 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4544 if (remainder > 2 * ParGCArrayScanChunk) { 4545 end = start + ParGCArrayScanChunk; 4546 to_obj_array->set_length(end); 4547 // Push the remainder before we process the range in case another 4548 // worker has run out of things to do and can steal it. 4549 oop* from_obj_p = set_partial_array_mask(from_obj); 4550 _par_scan_state->push_on_queue(from_obj_p); 4551 } else { 4552 assert(length == end, "sanity"); 4553 // We'll process the final range for this object. Restore the length 4554 // so that the heap remains parsable in case of evacuation failure. 4555 to_obj_array->set_length(end); 4556 } 4557 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4558 // Process indexes [start,end). It will also process the header 4559 // along with the first chunk (i.e., the chunk with start == 0). 4560 // Note that at this point the length field of to_obj_array is not 4561 // correct given that we are using it to keep track of the next 4562 // start index. oop_iterate_range() (thankfully!) ignores the length 4563 // field and only relies on the start / end parameters. It does 4564 // however return the size of the object which will be incorrect. So 4565 // we have to ignore it even if we wanted to use it. 4566 to_obj_array->oop_iterate_range(&_scanner, start, end); 4567 } 4568 4569 class G1ParEvacuateFollowersClosure : public VoidClosure { 4570 protected: 4571 G1CollectedHeap* _g1h; 4572 G1ParScanThreadState* _par_scan_state; 4573 RefToScanQueueSet* _queues; 4574 ParallelTaskTerminator* _terminator; 4575 4576 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4577 RefToScanQueueSet* queues() { return _queues; } 4578 ParallelTaskTerminator* terminator() { return _terminator; } 4579 4580 public: 4581 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4582 G1ParScanThreadState* par_scan_state, 4583 RefToScanQueueSet* queues, 4584 ParallelTaskTerminator* terminator) 4585 : _g1h(g1h), _par_scan_state(par_scan_state), 4586 _queues(queues), _terminator(terminator) {} 4587 4588 void do_void(); 4589 4590 private: 4591 inline bool offer_termination(); 4592 }; 4593 4594 bool G1ParEvacuateFollowersClosure::offer_termination() { 4595 G1ParScanThreadState* const pss = par_scan_state(); 4596 pss->start_term_time(); 4597 const bool res = terminator()->offer_termination(); 4598 pss->end_term_time(); 4599 return res; 4600 } 4601 4602 void G1ParEvacuateFollowersClosure::do_void() { 4603 StarTask stolen_task; 4604 G1ParScanThreadState* const pss = par_scan_state(); 4605 pss->trim_queue(); 4606 4607 do { 4608 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4609 assert(pss->verify_task(stolen_task), "sanity"); 4610 if (stolen_task.is_narrow()) { 4611 pss->deal_with_reference((narrowOop*) stolen_task); 4612 } else { 4613 pss->deal_with_reference((oop*) stolen_task); 4614 } 4615 4616 // We've just processed a reference and we might have made 4617 // available new entries on the queues. So we have to make sure 4618 // we drain the queues as necessary. 4619 pss->trim_queue(); 4620 } 4621 } while (!offer_termination()); 4622 4623 pss->retire_alloc_buffers(); 4624 } 4625 4626 class G1ParTask : public AbstractGangTask { 4627 protected: 4628 G1CollectedHeap* _g1h; 4629 RefToScanQueueSet *_queues; 4630 ParallelTaskTerminator _terminator; 4631 uint _n_workers; 4632 4633 Mutex _stats_lock; 4634 Mutex* stats_lock() { return &_stats_lock; } 4635 4636 size_t getNCards() { 4637 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4638 / G1BlockOffsetSharedArray::N_bytes; 4639 } 4640 4641 public: 4642 G1ParTask(G1CollectedHeap* g1h, 4643 RefToScanQueueSet *task_queues) 4644 : AbstractGangTask("G1 collection"), 4645 _g1h(g1h), 4646 _queues(task_queues), 4647 _terminator(0, _queues), 4648 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4649 {} 4650 4651 RefToScanQueueSet* queues() { return _queues; } 4652 4653 RefToScanQueue *work_queue(int i) { 4654 return queues()->queue(i); 4655 } 4656 4657 ParallelTaskTerminator* terminator() { return &_terminator; } 4658 4659 virtual void set_for_termination(int active_workers) { 4660 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4661 // in the young space (_par_seq_tasks) in the G1 heap 4662 // for SequentialSubTasksDone. 4663 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4664 // both of which need setting by set_n_termination(). 4665 _g1h->SharedHeap::set_n_termination(active_workers); 4666 _g1h->set_n_termination(active_workers); 4667 terminator()->reset_for_reuse(active_workers); 4668 _n_workers = active_workers; 4669 } 4670 4671 void work(uint worker_id) { 4672 if (worker_id >= _n_workers) return; // no work needed this round 4673 4674 double start_time_ms = os::elapsedTime() * 1000.0; 4675 _g1h->g1_policy()->record_gc_worker_start_time(worker_id, start_time_ms); 4676 4677 { 4678 ResourceMark rm; 4679 HandleMark hm; 4680 4681 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4682 4683 G1ParScanThreadState pss(_g1h, worker_id); 4684 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4685 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4686 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4687 4688 pss.set_evac_closure(&scan_evac_cl); 4689 pss.set_evac_failure_closure(&evac_failure_cl); 4690 pss.set_partial_scan_closure(&partial_scan_cl); 4691 4692 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4693 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4694 4695 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4696 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4697 4698 OopClosure* scan_root_cl = &only_scan_root_cl; 4699 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4700 4701 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4702 // We also need to mark copied objects. 4703 scan_root_cl = &scan_mark_root_cl; 4704 scan_perm_cl = &scan_mark_perm_cl; 4705 } 4706 4707 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4708 4709 pss.start_strong_roots(); 4710 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4711 SharedHeap::SO_AllClasses, 4712 scan_root_cl, 4713 &push_heap_rs_cl, 4714 scan_perm_cl, 4715 worker_id); 4716 pss.end_strong_roots(); 4717 4718 { 4719 double start = os::elapsedTime(); 4720 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4721 evac.do_void(); 4722 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4723 double term_ms = pss.term_time()*1000.0; 4724 _g1h->g1_policy()->record_obj_copy_time(worker_id, elapsed_ms-term_ms); 4725 _g1h->g1_policy()->record_termination(worker_id, term_ms, pss.term_attempts()); 4726 } 4727 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4728 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4729 4730 // Clean up any par-expanded rem sets. 4731 HeapRegionRemSet::par_cleanup(); 4732 4733 if (ParallelGCVerbose) { 4734 MutexLocker x(stats_lock()); 4735 pss.print_termination_stats(worker_id); 4736 } 4737 4738 assert(pss.refs()->is_empty(), "should be empty"); 4739 4740 // Close the inner scope so that the ResourceMark and HandleMark 4741 // destructors are executed here and are included as part of the 4742 // "GC Worker Time". 4743 } 4744 4745 double end_time_ms = os::elapsedTime() * 1000.0; 4746 _g1h->g1_policy()->record_gc_worker_end_time(worker_id, end_time_ms); 4747 } 4748 }; 4749 4750 // *** Common G1 Evacuation Stuff 4751 4752 // Closures that support the filtering of CodeBlobs scanned during 4753 // external root scanning. 4754 4755 // Closure applied to reference fields in code blobs (specifically nmethods) 4756 // to determine whether an nmethod contains references that point into 4757 // the collection set. Used as a predicate when walking code roots so 4758 // that only nmethods that point into the collection set are added to the 4759 // 'marked' list. 4760 4761 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure { 4762 4763 class G1PointsIntoCSOopClosure : public OopClosure { 4764 G1CollectedHeap* _g1; 4765 bool _points_into_cs; 4766 public: 4767 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) : 4768 _g1(g1), _points_into_cs(false) { } 4769 4770 bool points_into_cs() const { return _points_into_cs; } 4771 4772 template <class T> 4773 void do_oop_nv(T* p) { 4774 if (!_points_into_cs) { 4775 T heap_oop = oopDesc::load_heap_oop(p); 4776 if (!oopDesc::is_null(heap_oop) && 4777 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) { 4778 _points_into_cs = true; 4779 } 4780 } 4781 } 4782 4783 virtual void do_oop(oop* p) { do_oop_nv(p); } 4784 virtual void do_oop(narrowOop* p) { do_oop_nv(p); } 4785 }; 4786 4787 G1CollectedHeap* _g1; 4788 4789 public: 4790 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) : 4791 CodeBlobToOopClosure(cl, true), _g1(g1) { } 4792 4793 virtual void do_code_blob(CodeBlob* cb) { 4794 nmethod* nm = cb->as_nmethod_or_null(); 4795 if (nm != NULL && !(nm->test_oops_do_mark())) { 4796 G1PointsIntoCSOopClosure predicate_cl(_g1); 4797 nm->oops_do(&predicate_cl); 4798 4799 if (predicate_cl.points_into_cs()) { 4800 // At least one of the reference fields or the oop relocations 4801 // in the nmethod points into the collection set. We have to 4802 // 'mark' this nmethod. 4803 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob() 4804 // or MarkingCodeBlobClosure::do_code_blob() change. 4805 if (!nm->test_set_oops_do_mark()) { 4806 do_newly_marked_nmethod(nm); 4807 } 4808 } 4809 } 4810 } 4811 }; 4812 4813 // This method is run in a GC worker. 4814 4815 void 4816 G1CollectedHeap:: 4817 g1_process_strong_roots(bool collecting_perm_gen, 4818 ScanningOption so, 4819 OopClosure* scan_non_heap_roots, 4820 OopsInHeapRegionClosure* scan_rs, 4821 OopsInGenClosure* scan_perm, 4822 int worker_i) { 4823 4824 // First scan the strong roots, including the perm gen. 4825 double ext_roots_start = os::elapsedTime(); 4826 double closure_app_time_sec = 0.0; 4827 4828 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4829 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4830 buf_scan_perm.set_generation(perm_gen()); 4831 4832 // Walk the code cache w/o buffering, because StarTask cannot handle 4833 // unaligned oop locations. 4834 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots); 4835 4836 process_strong_roots(false, // no scoping; this is parallel code 4837 collecting_perm_gen, so, 4838 &buf_scan_non_heap_roots, 4839 &eager_scan_code_roots, 4840 &buf_scan_perm); 4841 4842 // Now the CM ref_processor roots. 4843 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4844 // We need to treat the discovered reference lists of the 4845 // concurrent mark ref processor as roots and keep entries 4846 // (which are added by the marking threads) on them live 4847 // until they can be processed at the end of marking. 4848 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4849 } 4850 4851 // Finish up any enqueued closure apps (attributed as object copy time). 4852 buf_scan_non_heap_roots.done(); 4853 buf_scan_perm.done(); 4854 4855 double ext_roots_end = os::elapsedTime(); 4856 4857 g1_policy()->reset_obj_copy_time(worker_i); 4858 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4859 buf_scan_non_heap_roots.closure_app_seconds(); 4860 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4861 4862 double ext_root_time_ms = 4863 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4864 4865 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4866 4867 // During conc marking we have to filter the per-thread SATB buffers 4868 // to make sure we remove any oops into the CSet (which will show up 4869 // as implicitly live). 4870 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 4871 if (mark_in_progress()) { 4872 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 4873 } 4874 } 4875 double satb_filtering_ms = (os::elapsedTime() - ext_roots_end) * 1000.0; 4876 g1_policy()->record_satb_filtering_time(worker_i, satb_filtering_ms); 4877 4878 // Now scan the complement of the collection set. 4879 if (scan_rs != NULL) { 4880 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4881 } 4882 4883 _process_strong_tasks->all_tasks_completed(); 4884 } 4885 4886 void 4887 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 4888 OopClosure* non_root_closure) { 4889 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 4890 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 4891 } 4892 4893 // Weak Reference Processing support 4894 4895 // An always "is_alive" closure that is used to preserve referents. 4896 // If the object is non-null then it's alive. Used in the preservation 4897 // of referent objects that are pointed to by reference objects 4898 // discovered by the CM ref processor. 4899 class G1AlwaysAliveClosure: public BoolObjectClosure { 4900 G1CollectedHeap* _g1; 4901 public: 4902 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4903 void do_object(oop p) { assert(false, "Do not call."); } 4904 bool do_object_b(oop p) { 4905 if (p != NULL) { 4906 return true; 4907 } 4908 return false; 4909 } 4910 }; 4911 4912 bool G1STWIsAliveClosure::do_object_b(oop p) { 4913 // An object is reachable if it is outside the collection set, 4914 // or is inside and copied. 4915 return !_g1->obj_in_cs(p) || p->is_forwarded(); 4916 } 4917 4918 // Non Copying Keep Alive closure 4919 class G1KeepAliveClosure: public OopClosure { 4920 G1CollectedHeap* _g1; 4921 public: 4922 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4923 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 4924 void do_oop( oop* p) { 4925 oop obj = *p; 4926 4927 if (_g1->obj_in_cs(obj)) { 4928 assert( obj->is_forwarded(), "invariant" ); 4929 *p = obj->forwardee(); 4930 } 4931 } 4932 }; 4933 4934 // Copying Keep Alive closure - can be called from both 4935 // serial and parallel code as long as different worker 4936 // threads utilize different G1ParScanThreadState instances 4937 // and different queues. 4938 4939 class G1CopyingKeepAliveClosure: public OopClosure { 4940 G1CollectedHeap* _g1h; 4941 OopClosure* _copy_non_heap_obj_cl; 4942 OopsInHeapRegionClosure* _copy_perm_obj_cl; 4943 G1ParScanThreadState* _par_scan_state; 4944 4945 public: 4946 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 4947 OopClosure* non_heap_obj_cl, 4948 OopsInHeapRegionClosure* perm_obj_cl, 4949 G1ParScanThreadState* pss): 4950 _g1h(g1h), 4951 _copy_non_heap_obj_cl(non_heap_obj_cl), 4952 _copy_perm_obj_cl(perm_obj_cl), 4953 _par_scan_state(pss) 4954 {} 4955 4956 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 4957 virtual void do_oop( oop* p) { do_oop_work(p); } 4958 4959 template <class T> void do_oop_work(T* p) { 4960 oop obj = oopDesc::load_decode_heap_oop(p); 4961 4962 if (_g1h->obj_in_cs(obj)) { 4963 // If the referent object has been forwarded (either copied 4964 // to a new location or to itself in the event of an 4965 // evacuation failure) then we need to update the reference 4966 // field and, if both reference and referent are in the G1 4967 // heap, update the RSet for the referent. 4968 // 4969 // If the referent has not been forwarded then we have to keep 4970 // it alive by policy. Therefore we have copy the referent. 4971 // 4972 // If the reference field is in the G1 heap then we can push 4973 // on the PSS queue. When the queue is drained (after each 4974 // phase of reference processing) the object and it's followers 4975 // will be copied, the reference field set to point to the 4976 // new location, and the RSet updated. Otherwise we need to 4977 // use the the non-heap or perm closures directly to copy 4978 // the refernt object and update the pointer, while avoiding 4979 // updating the RSet. 4980 4981 if (_g1h->is_in_g1_reserved(p)) { 4982 _par_scan_state->push_on_queue(p); 4983 } else { 4984 // The reference field is not in the G1 heap. 4985 if (_g1h->perm_gen()->is_in(p)) { 4986 _copy_perm_obj_cl->do_oop(p); 4987 } else { 4988 _copy_non_heap_obj_cl->do_oop(p); 4989 } 4990 } 4991 } 4992 } 4993 }; 4994 4995 // Serial drain queue closure. Called as the 'complete_gc' 4996 // closure for each discovered list in some of the 4997 // reference processing phases. 4998 4999 class G1STWDrainQueueClosure: public VoidClosure { 5000 protected: 5001 G1CollectedHeap* _g1h; 5002 G1ParScanThreadState* _par_scan_state; 5003 5004 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5005 5006 public: 5007 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5008 _g1h(g1h), 5009 _par_scan_state(pss) 5010 { } 5011 5012 void do_void() { 5013 G1ParScanThreadState* const pss = par_scan_state(); 5014 pss->trim_queue(); 5015 } 5016 }; 5017 5018 // Parallel Reference Processing closures 5019 5020 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5021 // processing during G1 evacuation pauses. 5022 5023 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5024 private: 5025 G1CollectedHeap* _g1h; 5026 RefToScanQueueSet* _queues; 5027 FlexibleWorkGang* _workers; 5028 int _active_workers; 5029 5030 public: 5031 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5032 FlexibleWorkGang* workers, 5033 RefToScanQueueSet *task_queues, 5034 int n_workers) : 5035 _g1h(g1h), 5036 _queues(task_queues), 5037 _workers(workers), 5038 _active_workers(n_workers) 5039 { 5040 assert(n_workers > 0, "shouldn't call this otherwise"); 5041 } 5042 5043 // Executes the given task using concurrent marking worker threads. 5044 virtual void execute(ProcessTask& task); 5045 virtual void execute(EnqueueTask& task); 5046 }; 5047 5048 // Gang task for possibly parallel reference processing 5049 5050 class G1STWRefProcTaskProxy: public AbstractGangTask { 5051 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5052 ProcessTask& _proc_task; 5053 G1CollectedHeap* _g1h; 5054 RefToScanQueueSet *_task_queues; 5055 ParallelTaskTerminator* _terminator; 5056 5057 public: 5058 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5059 G1CollectedHeap* g1h, 5060 RefToScanQueueSet *task_queues, 5061 ParallelTaskTerminator* terminator) : 5062 AbstractGangTask("Process reference objects in parallel"), 5063 _proc_task(proc_task), 5064 _g1h(g1h), 5065 _task_queues(task_queues), 5066 _terminator(terminator) 5067 {} 5068 5069 virtual void work(uint worker_id) { 5070 // The reference processing task executed by a single worker. 5071 ResourceMark rm; 5072 HandleMark hm; 5073 5074 G1STWIsAliveClosure is_alive(_g1h); 5075 5076 G1ParScanThreadState pss(_g1h, worker_id); 5077 5078 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5079 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5080 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5081 5082 pss.set_evac_closure(&scan_evac_cl); 5083 pss.set_evac_failure_closure(&evac_failure_cl); 5084 pss.set_partial_scan_closure(&partial_scan_cl); 5085 5086 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5087 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5088 5089 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5090 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5091 5092 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5093 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5094 5095 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5096 // We also need to mark copied objects. 5097 copy_non_heap_cl = ©_mark_non_heap_cl; 5098 copy_perm_cl = ©_mark_perm_cl; 5099 } 5100 5101 // Keep alive closure. 5102 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5103 5104 // Complete GC closure 5105 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5106 5107 // Call the reference processing task's work routine. 5108 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5109 5110 // Note we cannot assert that the refs array is empty here as not all 5111 // of the processing tasks (specifically phase2 - pp2_work) execute 5112 // the complete_gc closure (which ordinarily would drain the queue) so 5113 // the queue may not be empty. 5114 } 5115 }; 5116 5117 // Driver routine for parallel reference processing. 5118 // Creates an instance of the ref processing gang 5119 // task and has the worker threads execute it. 5120 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5121 assert(_workers != NULL, "Need parallel worker threads."); 5122 5123 ParallelTaskTerminator terminator(_active_workers, _queues); 5124 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5125 5126 _g1h->set_par_threads(_active_workers); 5127 _workers->run_task(&proc_task_proxy); 5128 _g1h->set_par_threads(0); 5129 } 5130 5131 // Gang task for parallel reference enqueueing. 5132 5133 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5134 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5135 EnqueueTask& _enq_task; 5136 5137 public: 5138 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5139 AbstractGangTask("Enqueue reference objects in parallel"), 5140 _enq_task(enq_task) 5141 { } 5142 5143 virtual void work(uint worker_id) { 5144 _enq_task.work(worker_id); 5145 } 5146 }; 5147 5148 // Driver routine for parallel reference enqueing. 5149 // Creates an instance of the ref enqueueing gang 5150 // task and has the worker threads execute it. 5151 5152 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5153 assert(_workers != NULL, "Need parallel worker threads."); 5154 5155 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5156 5157 _g1h->set_par_threads(_active_workers); 5158 _workers->run_task(&enq_task_proxy); 5159 _g1h->set_par_threads(0); 5160 } 5161 5162 // End of weak reference support closures 5163 5164 // Abstract task used to preserve (i.e. copy) any referent objects 5165 // that are in the collection set and are pointed to by reference 5166 // objects discovered by the CM ref processor. 5167 5168 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5169 protected: 5170 G1CollectedHeap* _g1h; 5171 RefToScanQueueSet *_queues; 5172 ParallelTaskTerminator _terminator; 5173 uint _n_workers; 5174 5175 public: 5176 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5177 AbstractGangTask("ParPreserveCMReferents"), 5178 _g1h(g1h), 5179 _queues(task_queues), 5180 _terminator(workers, _queues), 5181 _n_workers(workers) 5182 { } 5183 5184 void work(uint worker_id) { 5185 ResourceMark rm; 5186 HandleMark hm; 5187 5188 G1ParScanThreadState pss(_g1h, worker_id); 5189 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5190 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5191 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5192 5193 pss.set_evac_closure(&scan_evac_cl); 5194 pss.set_evac_failure_closure(&evac_failure_cl); 5195 pss.set_partial_scan_closure(&partial_scan_cl); 5196 5197 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5198 5199 5200 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5201 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5202 5203 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5204 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5205 5206 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5207 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5208 5209 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5210 // We also need to mark copied objects. 5211 copy_non_heap_cl = ©_mark_non_heap_cl; 5212 copy_perm_cl = ©_mark_perm_cl; 5213 } 5214 5215 // Is alive closure 5216 G1AlwaysAliveClosure always_alive(_g1h); 5217 5218 // Copying keep alive closure. Applied to referent objects that need 5219 // to be copied. 5220 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5221 5222 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5223 5224 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5225 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5226 5227 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5228 // So this must be true - but assert just in case someone decides to 5229 // change the worker ids. 5230 assert(0 <= worker_id && worker_id < limit, "sanity"); 5231 assert(!rp->discovery_is_atomic(), "check this code"); 5232 5233 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5234 for (uint idx = worker_id; idx < limit; idx += stride) { 5235 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5236 5237 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5238 while (iter.has_next()) { 5239 // Since discovery is not atomic for the CM ref processor, we 5240 // can see some null referent objects. 5241 iter.load_ptrs(DEBUG_ONLY(true)); 5242 oop ref = iter.obj(); 5243 5244 // This will filter nulls. 5245 if (iter.is_referent_alive()) { 5246 iter.make_referent_alive(); 5247 } 5248 iter.move_to_next(); 5249 } 5250 } 5251 5252 // Drain the queue - which may cause stealing 5253 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5254 drain_queue.do_void(); 5255 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5256 assert(pss.refs()->is_empty(), "should be"); 5257 } 5258 }; 5259 5260 // Weak Reference processing during an evacuation pause (part 1). 5261 void G1CollectedHeap::process_discovered_references() { 5262 double ref_proc_start = os::elapsedTime(); 5263 5264 ReferenceProcessor* rp = _ref_processor_stw; 5265 assert(rp->discovery_enabled(), "should have been enabled"); 5266 5267 // Any reference objects, in the collection set, that were 'discovered' 5268 // by the CM ref processor should have already been copied (either by 5269 // applying the external root copy closure to the discovered lists, or 5270 // by following an RSet entry). 5271 // 5272 // But some of the referents, that are in the collection set, that these 5273 // reference objects point to may not have been copied: the STW ref 5274 // processor would have seen that the reference object had already 5275 // been 'discovered' and would have skipped discovering the reference, 5276 // but would not have treated the reference object as a regular oop. 5277 // As a reult the copy closure would not have been applied to the 5278 // referent object. 5279 // 5280 // We need to explicitly copy these referent objects - the references 5281 // will be processed at the end of remarking. 5282 // 5283 // We also need to do this copying before we process the reference 5284 // objects discovered by the STW ref processor in case one of these 5285 // referents points to another object which is also referenced by an 5286 // object discovered by the STW ref processor. 5287 5288 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5289 workers()->active_workers() : 1); 5290 5291 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5292 active_workers == workers()->active_workers(), 5293 "Need to reset active_workers"); 5294 5295 set_par_threads(active_workers); 5296 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5297 5298 if (G1CollectedHeap::use_parallel_gc_threads()) { 5299 workers()->run_task(&keep_cm_referents); 5300 } else { 5301 keep_cm_referents.work(0); 5302 } 5303 5304 set_par_threads(0); 5305 5306 // Closure to test whether a referent is alive. 5307 G1STWIsAliveClosure is_alive(this); 5308 5309 // Even when parallel reference processing is enabled, the processing 5310 // of JNI refs is serial and performed serially by the current thread 5311 // rather than by a worker. The following PSS will be used for processing 5312 // JNI refs. 5313 5314 // Use only a single queue for this PSS. 5315 G1ParScanThreadState pss(this, 0); 5316 5317 // We do not embed a reference processor in the copying/scanning 5318 // closures while we're actually processing the discovered 5319 // reference objects. 5320 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5321 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5322 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5323 5324 pss.set_evac_closure(&scan_evac_cl); 5325 pss.set_evac_failure_closure(&evac_failure_cl); 5326 pss.set_partial_scan_closure(&partial_scan_cl); 5327 5328 assert(pss.refs()->is_empty(), "pre-condition"); 5329 5330 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5331 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5332 5333 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5334 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5335 5336 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5337 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5338 5339 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5340 // We also need to mark copied objects. 5341 copy_non_heap_cl = ©_mark_non_heap_cl; 5342 copy_perm_cl = ©_mark_perm_cl; 5343 } 5344 5345 // Keep alive closure. 5346 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5347 5348 // Serial Complete GC closure 5349 G1STWDrainQueueClosure drain_queue(this, &pss); 5350 5351 // Setup the soft refs policy... 5352 rp->setup_policy(false); 5353 5354 if (!rp->processing_is_mt()) { 5355 // Serial reference processing... 5356 rp->process_discovered_references(&is_alive, 5357 &keep_alive, 5358 &drain_queue, 5359 NULL); 5360 } else { 5361 // Parallel reference processing 5362 assert(rp->num_q() == active_workers, "sanity"); 5363 assert(active_workers <= rp->max_num_q(), "sanity"); 5364 5365 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5366 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5367 } 5368 5369 // We have completed copying any necessary live referent objects 5370 // (that were not copied during the actual pause) so we can 5371 // retire any active alloc buffers 5372 pss.retire_alloc_buffers(); 5373 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5374 5375 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5376 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0); 5377 } 5378 5379 // Weak Reference processing during an evacuation pause (part 2). 5380 void G1CollectedHeap::enqueue_discovered_references() { 5381 double ref_enq_start = os::elapsedTime(); 5382 5383 ReferenceProcessor* rp = _ref_processor_stw; 5384 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5385 5386 // Now enqueue any remaining on the discovered lists on to 5387 // the pending list. 5388 if (!rp->processing_is_mt()) { 5389 // Serial reference processing... 5390 rp->enqueue_discovered_references(); 5391 } else { 5392 // Parallel reference enqueuing 5393 5394 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5395 assert(active_workers == workers()->active_workers(), 5396 "Need to reset active_workers"); 5397 assert(rp->num_q() == active_workers, "sanity"); 5398 assert(active_workers <= rp->max_num_q(), "sanity"); 5399 5400 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5401 rp->enqueue_discovered_references(&par_task_executor); 5402 } 5403 5404 rp->verify_no_references_recorded(); 5405 assert(!rp->discovery_enabled(), "should have been disabled"); 5406 5407 // FIXME 5408 // CM's reference processing also cleans up the string and symbol tables. 5409 // Should we do that here also? We could, but it is a serial operation 5410 // and could signicantly increase the pause time. 5411 5412 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5413 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0); 5414 } 5415 5416 void G1CollectedHeap::evacuate_collection_set() { 5417 _expand_heap_after_alloc_failure = true; 5418 set_evacuation_failed(false); 5419 5420 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5421 concurrent_g1_refine()->set_use_cache(false); 5422 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5423 5424 uint n_workers; 5425 if (G1CollectedHeap::use_parallel_gc_threads()) { 5426 n_workers = 5427 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5428 workers()->active_workers(), 5429 Threads::number_of_non_daemon_threads()); 5430 assert(UseDynamicNumberOfGCThreads || 5431 n_workers == workers()->total_workers(), 5432 "If not dynamic should be using all the workers"); 5433 workers()->set_active_workers(n_workers); 5434 set_par_threads(n_workers); 5435 } else { 5436 assert(n_par_threads() == 0, 5437 "Should be the original non-parallel value"); 5438 n_workers = 1; 5439 } 5440 5441 G1ParTask g1_par_task(this, _task_queues); 5442 5443 init_for_evac_failure(NULL); 5444 5445 rem_set()->prepare_for_younger_refs_iterate(true); 5446 5447 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5448 double start_par_time_sec = os::elapsedTime(); 5449 double end_par_time_sec; 5450 5451 { 5452 StrongRootsScope srs(this); 5453 5454 if (G1CollectedHeap::use_parallel_gc_threads()) { 5455 // The individual threads will set their evac-failure closures. 5456 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5457 // These tasks use ShareHeap::_process_strong_tasks 5458 assert(UseDynamicNumberOfGCThreads || 5459 workers()->active_workers() == workers()->total_workers(), 5460 "If not dynamic should be using all the workers"); 5461 workers()->run_task(&g1_par_task); 5462 } else { 5463 g1_par_task.set_for_termination(n_workers); 5464 g1_par_task.work(0); 5465 } 5466 end_par_time_sec = os::elapsedTime(); 5467 5468 // Closing the inner scope will execute the destructor 5469 // for the StrongRootsScope object. We record the current 5470 // elapsed time before closing the scope so that time 5471 // taken for the SRS destructor is NOT included in the 5472 // reported parallel time. 5473 } 5474 5475 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5476 g1_policy()->record_par_time(par_time_ms); 5477 5478 double code_root_fixup_time_ms = 5479 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5480 g1_policy()->record_code_root_fixup_time(code_root_fixup_time_ms); 5481 5482 set_par_threads(0); 5483 5484 // Process any discovered reference objects - we have 5485 // to do this _before_ we retire the GC alloc regions 5486 // as we may have to copy some 'reachable' referent 5487 // objects (and their reachable sub-graphs) that were 5488 // not copied during the pause. 5489 process_discovered_references(); 5490 5491 // Weak root processing. 5492 // Note: when JSR 292 is enabled and code blobs can contain 5493 // non-perm oops then we will need to process the code blobs 5494 // here too. 5495 { 5496 G1STWIsAliveClosure is_alive(this); 5497 G1KeepAliveClosure keep_alive(this); 5498 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5499 } 5500 5501 release_gc_alloc_regions(); 5502 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5503 5504 concurrent_g1_refine()->clear_hot_cache(); 5505 concurrent_g1_refine()->set_use_cache(true); 5506 5507 finalize_for_evac_failure(); 5508 5509 if (evacuation_failed()) { 5510 remove_self_forwarding_pointers(); 5511 if (G1Log::finer()) { 5512 gclog_or_tty->print(" (to-space exhausted)"); 5513 } else if (G1Log::fine()) { 5514 gclog_or_tty->print("--"); 5515 } 5516 } 5517 5518 // Enqueue any remaining references remaining on the STW 5519 // reference processor's discovered lists. We need to do 5520 // this after the card table is cleaned (and verified) as 5521 // the act of enqueuing entries on to the pending list 5522 // will log these updates (and dirty their associated 5523 // cards). We need these updates logged to update any 5524 // RSets. 5525 enqueue_discovered_references(); 5526 5527 if (G1DeferredRSUpdate) { 5528 RedirtyLoggedCardTableEntryFastClosure redirty; 5529 dirty_card_queue_set().set_closure(&redirty); 5530 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5531 5532 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5533 dcq.merge_bufferlists(&dirty_card_queue_set()); 5534 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5535 } 5536 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5537 } 5538 5539 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5540 size_t* pre_used, 5541 FreeRegionList* free_list, 5542 OldRegionSet* old_proxy_set, 5543 HumongousRegionSet* humongous_proxy_set, 5544 HRRSCleanupTask* hrrs_cleanup_task, 5545 bool par) { 5546 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5547 if (hr->isHumongous()) { 5548 assert(hr->startsHumongous(), "we should only see starts humongous"); 5549 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5550 } else { 5551 _old_set.remove_with_proxy(hr, old_proxy_set); 5552 free_region(hr, pre_used, free_list, par); 5553 } 5554 } else { 5555 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5556 } 5557 } 5558 5559 void G1CollectedHeap::free_region(HeapRegion* hr, 5560 size_t* pre_used, 5561 FreeRegionList* free_list, 5562 bool par) { 5563 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5564 assert(!hr->is_empty(), "the region should not be empty"); 5565 assert(free_list != NULL, "pre-condition"); 5566 5567 *pre_used += hr->used(); 5568 hr->hr_clear(par, true /* clear_space */); 5569 free_list->add_as_head(hr); 5570 } 5571 5572 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5573 size_t* pre_used, 5574 FreeRegionList* free_list, 5575 HumongousRegionSet* humongous_proxy_set, 5576 bool par) { 5577 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5578 assert(free_list != NULL, "pre-condition"); 5579 assert(humongous_proxy_set != NULL, "pre-condition"); 5580 5581 size_t hr_used = hr->used(); 5582 size_t hr_capacity = hr->capacity(); 5583 size_t hr_pre_used = 0; 5584 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5585 hr->set_notHumongous(); 5586 free_region(hr, &hr_pre_used, free_list, par); 5587 5588 uint i = hr->hrs_index() + 1; 5589 uint num = 1; 5590 while (i < n_regions()) { 5591 HeapRegion* curr_hr = region_at(i); 5592 if (!curr_hr->continuesHumongous()) { 5593 break; 5594 } 5595 curr_hr->set_notHumongous(); 5596 free_region(curr_hr, &hr_pre_used, free_list, par); 5597 num += 1; 5598 i += 1; 5599 } 5600 assert(hr_pre_used == hr_used, 5601 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5602 "should be the same", hr_pre_used, hr_used)); 5603 *pre_used += hr_pre_used; 5604 } 5605 5606 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5607 FreeRegionList* free_list, 5608 OldRegionSet* old_proxy_set, 5609 HumongousRegionSet* humongous_proxy_set, 5610 bool par) { 5611 if (pre_used > 0) { 5612 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5613 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5614 assert(_summary_bytes_used >= pre_used, 5615 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5616 "should be >= pre_used: "SIZE_FORMAT, 5617 _summary_bytes_used, pre_used)); 5618 _summary_bytes_used -= pre_used; 5619 } 5620 if (free_list != NULL && !free_list->is_empty()) { 5621 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5622 _free_list.add_as_head(free_list); 5623 } 5624 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5625 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5626 _old_set.update_from_proxy(old_proxy_set); 5627 } 5628 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5629 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5630 _humongous_set.update_from_proxy(humongous_proxy_set); 5631 } 5632 } 5633 5634 class G1ParCleanupCTTask : public AbstractGangTask { 5635 CardTableModRefBS* _ct_bs; 5636 G1CollectedHeap* _g1h; 5637 HeapRegion* volatile _su_head; 5638 public: 5639 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5640 G1CollectedHeap* g1h) : 5641 AbstractGangTask("G1 Par Cleanup CT Task"), 5642 _ct_bs(ct_bs), _g1h(g1h) { } 5643 5644 void work(uint worker_id) { 5645 HeapRegion* r; 5646 while (r = _g1h->pop_dirty_cards_region()) { 5647 clear_cards(r); 5648 } 5649 } 5650 5651 void clear_cards(HeapRegion* r) { 5652 // Cards of the survivors should have already been dirtied. 5653 if (!r->is_survivor()) { 5654 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5655 } 5656 } 5657 }; 5658 5659 #ifndef PRODUCT 5660 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5661 G1CollectedHeap* _g1h; 5662 CardTableModRefBS* _ct_bs; 5663 public: 5664 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5665 : _g1h(g1h), _ct_bs(ct_bs) { } 5666 virtual bool doHeapRegion(HeapRegion* r) { 5667 if (r->is_survivor()) { 5668 _g1h->verify_dirty_region(r); 5669 } else { 5670 _g1h->verify_not_dirty_region(r); 5671 } 5672 return false; 5673 } 5674 }; 5675 5676 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5677 // All of the region should be clean. 5678 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5679 MemRegion mr(hr->bottom(), hr->end()); 5680 ct_bs->verify_not_dirty_region(mr); 5681 } 5682 5683 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5684 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5685 // dirty allocated blocks as they allocate them. The thread that 5686 // retires each region and replaces it with a new one will do a 5687 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5688 // not dirty that area (one less thing to have to do while holding 5689 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5690 // is dirty. 5691 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5692 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5693 ct_bs->verify_dirty_region(mr); 5694 } 5695 5696 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5697 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5698 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5699 verify_dirty_region(hr); 5700 } 5701 } 5702 5703 void G1CollectedHeap::verify_dirty_young_regions() { 5704 verify_dirty_young_list(_young_list->first_region()); 5705 verify_dirty_young_list(_young_list->first_survivor_region()); 5706 } 5707 #endif 5708 5709 void G1CollectedHeap::cleanUpCardTable() { 5710 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5711 double start = os::elapsedTime(); 5712 5713 { 5714 // Iterate over the dirty cards region list. 5715 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5716 5717 if (G1CollectedHeap::use_parallel_gc_threads()) { 5718 set_par_threads(); 5719 workers()->run_task(&cleanup_task); 5720 set_par_threads(0); 5721 } else { 5722 while (_dirty_cards_region_list) { 5723 HeapRegion* r = _dirty_cards_region_list; 5724 cleanup_task.clear_cards(r); 5725 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5726 if (_dirty_cards_region_list == r) { 5727 // The last region. 5728 _dirty_cards_region_list = NULL; 5729 } 5730 r->set_next_dirty_cards_region(NULL); 5731 } 5732 } 5733 #ifndef PRODUCT 5734 if (G1VerifyCTCleanup || VerifyAfterGC) { 5735 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5736 heap_region_iterate(&cleanup_verifier); 5737 } 5738 #endif 5739 } 5740 5741 double elapsed = os::elapsedTime() - start; 5742 g1_policy()->record_clear_ct_time(elapsed * 1000.0); 5743 } 5744 5745 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5746 size_t pre_used = 0; 5747 FreeRegionList local_free_list("Local List for CSet Freeing"); 5748 5749 double young_time_ms = 0.0; 5750 double non_young_time_ms = 0.0; 5751 5752 // Since the collection set is a superset of the the young list, 5753 // all we need to do to clear the young list is clear its 5754 // head and length, and unlink any young regions in the code below 5755 _young_list->clear(); 5756 5757 G1CollectorPolicy* policy = g1_policy(); 5758 5759 double start_sec = os::elapsedTime(); 5760 bool non_young = true; 5761 5762 HeapRegion* cur = cs_head; 5763 int age_bound = -1; 5764 size_t rs_lengths = 0; 5765 5766 while (cur != NULL) { 5767 assert(!is_on_master_free_list(cur), "sanity"); 5768 if (non_young) { 5769 if (cur->is_young()) { 5770 double end_sec = os::elapsedTime(); 5771 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5772 non_young_time_ms += elapsed_ms; 5773 5774 start_sec = os::elapsedTime(); 5775 non_young = false; 5776 } 5777 } else { 5778 if (!cur->is_young()) { 5779 double end_sec = os::elapsedTime(); 5780 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5781 young_time_ms += elapsed_ms; 5782 5783 start_sec = os::elapsedTime(); 5784 non_young = true; 5785 } 5786 } 5787 5788 rs_lengths += cur->rem_set()->occupied(); 5789 5790 HeapRegion* next = cur->next_in_collection_set(); 5791 assert(cur->in_collection_set(), "bad CS"); 5792 cur->set_next_in_collection_set(NULL); 5793 cur->set_in_collection_set(false); 5794 5795 if (cur->is_young()) { 5796 int index = cur->young_index_in_cset(); 5797 assert(index != -1, "invariant"); 5798 assert((uint) index < policy->young_cset_region_length(), "invariant"); 5799 size_t words_survived = _surviving_young_words[index]; 5800 cur->record_surv_words_in_group(words_survived); 5801 5802 // At this point the we have 'popped' cur from the collection set 5803 // (linked via next_in_collection_set()) but it is still in the 5804 // young list (linked via next_young_region()). Clear the 5805 // _next_young_region field. 5806 cur->set_next_young_region(NULL); 5807 } else { 5808 int index = cur->young_index_in_cset(); 5809 assert(index == -1, "invariant"); 5810 } 5811 5812 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5813 (!cur->is_young() && cur->young_index_in_cset() == -1), 5814 "invariant" ); 5815 5816 if (!cur->evacuation_failed()) { 5817 MemRegion used_mr = cur->used_region(); 5818 5819 // And the region is empty. 5820 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5821 free_region(cur, &pre_used, &local_free_list, false /* par */); 5822 } else { 5823 cur->uninstall_surv_rate_group(); 5824 if (cur->is_young()) { 5825 cur->set_young_index_in_cset(-1); 5826 } 5827 cur->set_not_young(); 5828 cur->set_evacuation_failed(false); 5829 // The region is now considered to be old. 5830 _old_set.add(cur); 5831 } 5832 cur = next; 5833 } 5834 5835 policy->record_max_rs_lengths(rs_lengths); 5836 policy->cset_regions_freed(); 5837 5838 double end_sec = os::elapsedTime(); 5839 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5840 5841 if (non_young) { 5842 non_young_time_ms += elapsed_ms; 5843 } else { 5844 young_time_ms += elapsed_ms; 5845 } 5846 5847 update_sets_after_freeing_regions(pre_used, &local_free_list, 5848 NULL /* old_proxy_set */, 5849 NULL /* humongous_proxy_set */, 5850 false /* par */); 5851 policy->record_young_free_cset_time_ms(young_time_ms); 5852 policy->record_non_young_free_cset_time_ms(non_young_time_ms); 5853 } 5854 5855 // This routine is similar to the above but does not record 5856 // any policy statistics or update free lists; we are abandoning 5857 // the current incremental collection set in preparation of a 5858 // full collection. After the full GC we will start to build up 5859 // the incremental collection set again. 5860 // This is only called when we're doing a full collection 5861 // and is immediately followed by the tearing down of the young list. 5862 5863 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5864 HeapRegion* cur = cs_head; 5865 5866 while (cur != NULL) { 5867 HeapRegion* next = cur->next_in_collection_set(); 5868 assert(cur->in_collection_set(), "bad CS"); 5869 cur->set_next_in_collection_set(NULL); 5870 cur->set_in_collection_set(false); 5871 cur->set_young_index_in_cset(-1); 5872 cur = next; 5873 } 5874 } 5875 5876 void G1CollectedHeap::set_free_regions_coming() { 5877 if (G1ConcRegionFreeingVerbose) { 5878 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5879 "setting free regions coming"); 5880 } 5881 5882 assert(!free_regions_coming(), "pre-condition"); 5883 _free_regions_coming = true; 5884 } 5885 5886 void G1CollectedHeap::reset_free_regions_coming() { 5887 assert(free_regions_coming(), "pre-condition"); 5888 5889 { 5890 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5891 _free_regions_coming = false; 5892 SecondaryFreeList_lock->notify_all(); 5893 } 5894 5895 if (G1ConcRegionFreeingVerbose) { 5896 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5897 "reset free regions coming"); 5898 } 5899 } 5900 5901 void G1CollectedHeap::wait_while_free_regions_coming() { 5902 // Most of the time we won't have to wait, so let's do a quick test 5903 // first before we take the lock. 5904 if (!free_regions_coming()) { 5905 return; 5906 } 5907 5908 if (G1ConcRegionFreeingVerbose) { 5909 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5910 "waiting for free regions"); 5911 } 5912 5913 { 5914 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5915 while (free_regions_coming()) { 5916 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 5917 } 5918 } 5919 5920 if (G1ConcRegionFreeingVerbose) { 5921 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5922 "done waiting for free regions"); 5923 } 5924 } 5925 5926 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 5927 assert(heap_lock_held_for_gc(), 5928 "the heap lock should already be held by or for this thread"); 5929 _young_list->push_region(hr); 5930 } 5931 5932 class NoYoungRegionsClosure: public HeapRegionClosure { 5933 private: 5934 bool _success; 5935 public: 5936 NoYoungRegionsClosure() : _success(true) { } 5937 bool doHeapRegion(HeapRegion* r) { 5938 if (r->is_young()) { 5939 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 5940 r->bottom(), r->end()); 5941 _success = false; 5942 } 5943 return false; 5944 } 5945 bool success() { return _success; } 5946 }; 5947 5948 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 5949 bool ret = _young_list->check_list_empty(check_sample); 5950 5951 if (check_heap) { 5952 NoYoungRegionsClosure closure; 5953 heap_region_iterate(&closure); 5954 ret = ret && closure.success(); 5955 } 5956 5957 return ret; 5958 } 5959 5960 class TearDownRegionSetsClosure : public HeapRegionClosure { 5961 private: 5962 OldRegionSet *_old_set; 5963 5964 public: 5965 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 5966 5967 bool doHeapRegion(HeapRegion* r) { 5968 if (r->is_empty()) { 5969 // We ignore empty regions, we'll empty the free list afterwards 5970 } else if (r->is_young()) { 5971 // We ignore young regions, we'll empty the young list afterwards 5972 } else if (r->isHumongous()) { 5973 // We ignore humongous regions, we're not tearing down the 5974 // humongous region set 5975 } else { 5976 // The rest should be old 5977 _old_set->remove(r); 5978 } 5979 return false; 5980 } 5981 5982 ~TearDownRegionSetsClosure() { 5983 assert(_old_set->is_empty(), "post-condition"); 5984 } 5985 }; 5986 5987 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 5988 assert_at_safepoint(true /* should_be_vm_thread */); 5989 5990 if (!free_list_only) { 5991 TearDownRegionSetsClosure cl(&_old_set); 5992 heap_region_iterate(&cl); 5993 5994 // Need to do this after the heap iteration to be able to 5995 // recognize the young regions and ignore them during the iteration. 5996 _young_list->empty_list(); 5997 } 5998 _free_list.remove_all(); 5999 } 6000 6001 class RebuildRegionSetsClosure : public HeapRegionClosure { 6002 private: 6003 bool _free_list_only; 6004 OldRegionSet* _old_set; 6005 FreeRegionList* _free_list; 6006 size_t _total_used; 6007 6008 public: 6009 RebuildRegionSetsClosure(bool free_list_only, 6010 OldRegionSet* old_set, FreeRegionList* free_list) : 6011 _free_list_only(free_list_only), 6012 _old_set(old_set), _free_list(free_list), _total_used(0) { 6013 assert(_free_list->is_empty(), "pre-condition"); 6014 if (!free_list_only) { 6015 assert(_old_set->is_empty(), "pre-condition"); 6016 } 6017 } 6018 6019 bool doHeapRegion(HeapRegion* r) { 6020 if (r->continuesHumongous()) { 6021 return false; 6022 } 6023 6024 if (r->is_empty()) { 6025 // Add free regions to the free list 6026 _free_list->add_as_tail(r); 6027 } else if (!_free_list_only) { 6028 assert(!r->is_young(), "we should not come across young regions"); 6029 6030 if (r->isHumongous()) { 6031 // We ignore humongous regions, we left the humongous set unchanged 6032 } else { 6033 // The rest should be old, add them to the old set 6034 _old_set->add(r); 6035 } 6036 _total_used += r->used(); 6037 } 6038 6039 return false; 6040 } 6041 6042 size_t total_used() { 6043 return _total_used; 6044 } 6045 }; 6046 6047 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6048 assert_at_safepoint(true /* should_be_vm_thread */); 6049 6050 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6051 heap_region_iterate(&cl); 6052 6053 if (!free_list_only) { 6054 _summary_bytes_used = cl.total_used(); 6055 } 6056 assert(_summary_bytes_used == recalculate_used(), 6057 err_msg("inconsistent _summary_bytes_used, " 6058 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6059 _summary_bytes_used, recalculate_used())); 6060 } 6061 6062 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6063 _refine_cte_cl->set_concurrent(concurrent); 6064 } 6065 6066 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6067 HeapRegion* hr = heap_region_containing(p); 6068 if (hr == NULL) { 6069 return is_in_permanent(p); 6070 } else { 6071 return hr->is_in(p); 6072 } 6073 } 6074 6075 // Methods for the mutator alloc region 6076 6077 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6078 bool force) { 6079 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6080 assert(!force || g1_policy()->can_expand_young_list(), 6081 "if force is true we should be able to expand the young list"); 6082 bool young_list_full = g1_policy()->is_young_list_full(); 6083 if (force || !young_list_full) { 6084 HeapRegion* new_alloc_region = new_region(word_size, 6085 false /* do_expand */); 6086 if (new_alloc_region != NULL) { 6087 set_region_short_lived_locked(new_alloc_region); 6088 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6089 return new_alloc_region; 6090 } 6091 } 6092 return NULL; 6093 } 6094 6095 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6096 size_t allocated_bytes) { 6097 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6098 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6099 6100 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6101 _summary_bytes_used += allocated_bytes; 6102 _hr_printer.retire(alloc_region); 6103 // We update the eden sizes here, when the region is retired, 6104 // instead of when it's allocated, since this is the point that its 6105 // used space has been recored in _summary_bytes_used. 6106 g1mm()->update_eden_size(); 6107 } 6108 6109 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6110 bool force) { 6111 return _g1h->new_mutator_alloc_region(word_size, force); 6112 } 6113 6114 void G1CollectedHeap::set_par_threads() { 6115 // Don't change the number of workers. Use the value previously set 6116 // in the workgroup. 6117 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6118 uint n_workers = workers()->active_workers(); 6119 assert(UseDynamicNumberOfGCThreads || 6120 n_workers == workers()->total_workers(), 6121 "Otherwise should be using the total number of workers"); 6122 if (n_workers == 0) { 6123 assert(false, "Should have been set in prior evacuation pause."); 6124 n_workers = ParallelGCThreads; 6125 workers()->set_active_workers(n_workers); 6126 } 6127 set_par_threads(n_workers); 6128 } 6129 6130 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6131 size_t allocated_bytes) { 6132 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6133 } 6134 6135 // Methods for the GC alloc regions 6136 6137 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6138 uint count, 6139 GCAllocPurpose ap) { 6140 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6141 6142 if (count < g1_policy()->max_regions(ap)) { 6143 HeapRegion* new_alloc_region = new_region(word_size, 6144 true /* do_expand */); 6145 if (new_alloc_region != NULL) { 6146 // We really only need to do this for old regions given that we 6147 // should never scan survivors. But it doesn't hurt to do it 6148 // for survivors too. 6149 new_alloc_region->set_saved_mark(); 6150 if (ap == GCAllocForSurvived) { 6151 new_alloc_region->set_survivor(); 6152 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6153 } else { 6154 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6155 } 6156 bool during_im = g1_policy()->during_initial_mark_pause(); 6157 new_alloc_region->note_start_of_copying(during_im); 6158 return new_alloc_region; 6159 } else { 6160 g1_policy()->note_alloc_region_limit_reached(ap); 6161 } 6162 } 6163 return NULL; 6164 } 6165 6166 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6167 size_t allocated_bytes, 6168 GCAllocPurpose ap) { 6169 bool during_im = g1_policy()->during_initial_mark_pause(); 6170 alloc_region->note_end_of_copying(during_im); 6171 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6172 if (ap == GCAllocForSurvived) { 6173 young_list()->add_survivor_region(alloc_region); 6174 } else { 6175 _old_set.add(alloc_region); 6176 } 6177 _hr_printer.retire(alloc_region); 6178 } 6179 6180 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6181 bool force) { 6182 assert(!force, "not supported for GC alloc regions"); 6183 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6184 } 6185 6186 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6187 size_t allocated_bytes) { 6188 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6189 GCAllocForSurvived); 6190 } 6191 6192 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6193 bool force) { 6194 assert(!force, "not supported for GC alloc regions"); 6195 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6196 } 6197 6198 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6199 size_t allocated_bytes) { 6200 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6201 GCAllocForTenured); 6202 } 6203 // Heap region set verification 6204 6205 class VerifyRegionListsClosure : public HeapRegionClosure { 6206 private: 6207 FreeRegionList* _free_list; 6208 OldRegionSet* _old_set; 6209 HumongousRegionSet* _humongous_set; 6210 uint _region_count; 6211 6212 public: 6213 VerifyRegionListsClosure(OldRegionSet* old_set, 6214 HumongousRegionSet* humongous_set, 6215 FreeRegionList* free_list) : 6216 _old_set(old_set), _humongous_set(humongous_set), 6217 _free_list(free_list), _region_count(0) { } 6218 6219 uint region_count() { return _region_count; } 6220 6221 bool doHeapRegion(HeapRegion* hr) { 6222 _region_count += 1; 6223 6224 if (hr->continuesHumongous()) { 6225 return false; 6226 } 6227 6228 if (hr->is_young()) { 6229 // TODO 6230 } else if (hr->startsHumongous()) { 6231 _humongous_set->verify_next_region(hr); 6232 } else if (hr->is_empty()) { 6233 _free_list->verify_next_region(hr); 6234 } else { 6235 _old_set->verify_next_region(hr); 6236 } 6237 return false; 6238 } 6239 }; 6240 6241 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6242 HeapWord* bottom) { 6243 HeapWord* end = bottom + HeapRegion::GrainWords; 6244 MemRegion mr(bottom, end); 6245 assert(_g1_reserved.contains(mr), "invariant"); 6246 // This might return NULL if the allocation fails 6247 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6248 } 6249 6250 void G1CollectedHeap::verify_region_sets() { 6251 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6252 6253 // First, check the explicit lists. 6254 _free_list.verify(); 6255 { 6256 // Given that a concurrent operation might be adding regions to 6257 // the secondary free list we have to take the lock before 6258 // verifying it. 6259 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6260 _secondary_free_list.verify(); 6261 } 6262 _old_set.verify(); 6263 _humongous_set.verify(); 6264 6265 // If a concurrent region freeing operation is in progress it will 6266 // be difficult to correctly attributed any free regions we come 6267 // across to the correct free list given that they might belong to 6268 // one of several (free_list, secondary_free_list, any local lists, 6269 // etc.). So, if that's the case we will skip the rest of the 6270 // verification operation. Alternatively, waiting for the concurrent 6271 // operation to complete will have a non-trivial effect on the GC's 6272 // operation (no concurrent operation will last longer than the 6273 // interval between two calls to verification) and it might hide 6274 // any issues that we would like to catch during testing. 6275 if (free_regions_coming()) { 6276 return; 6277 } 6278 6279 // Make sure we append the secondary_free_list on the free_list so 6280 // that all free regions we will come across can be safely 6281 // attributed to the free_list. 6282 append_secondary_free_list_if_not_empty_with_lock(); 6283 6284 // Finally, make sure that the region accounting in the lists is 6285 // consistent with what we see in the heap. 6286 _old_set.verify_start(); 6287 _humongous_set.verify_start(); 6288 _free_list.verify_start(); 6289 6290 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6291 heap_region_iterate(&cl); 6292 6293 _old_set.verify_end(); 6294 _humongous_set.verify_end(); 6295 _free_list.verify_end(); 6296 }