1 /* 2 * Copyright (c) 2001, 2016, 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 "gc/cms/compactibleFreeListSpace.hpp" 27 #include "gc/cms/concurrentMarkSweepGeneration.hpp" 28 #include "gc/cms/parNewGeneration.inline.hpp" 29 #include "gc/cms/parOopClosures.inline.hpp" 30 #include "gc/serial/defNewGeneration.inline.hpp" 31 #include "gc/shared/adaptiveSizePolicy.hpp" 32 #include "gc/shared/ageTable.inline.hpp" 33 #include "gc/shared/copyFailedInfo.hpp" 34 #include "gc/shared/gcHeapSummary.hpp" 35 #include "gc/shared/gcTimer.hpp" 36 #include "gc/shared/gcTrace.hpp" 37 #include "gc/shared/gcTraceTime.inline.hpp" 38 #include "gc/shared/genCollectedHeap.hpp" 39 #include "gc/shared/genOopClosures.inline.hpp" 40 #include "gc/shared/generation.hpp" 41 #include "gc/shared/plab.inline.hpp" 42 #include "gc/shared/preservedMarks.inline.hpp" 43 #include "gc/shared/referencePolicy.hpp" 44 #include "gc/shared/space.hpp" 45 #include "gc/shared/spaceDecorator.hpp" 46 #include "gc/shared/strongRootsScope.hpp" 47 #include "gc/shared/taskqueue.inline.hpp" 48 #include "gc/shared/workgroup.hpp" 49 #include "logging/log.hpp" 50 #include "memory/resourceArea.hpp" 51 #include "oops/objArrayOop.hpp" 52 #include "oops/oop.inline.hpp" 53 #include "runtime/atomic.hpp" 54 #include "runtime/handles.hpp" 55 #include "runtime/handles.inline.hpp" 56 #include "runtime/java.hpp" 57 #include "runtime/thread.inline.hpp" 58 #include "utilities/copy.hpp" 59 #include "utilities/globalDefinitions.hpp" 60 #include "utilities/stack.inline.hpp" 61 62 ParScanThreadState::ParScanThreadState(Space* to_space_, 63 ParNewGeneration* young_gen_, 64 Generation* old_gen_, 65 int thread_num_, 66 ObjToScanQueueSet* work_queue_set_, 67 Stack<oop, mtGC>* overflow_stacks_, 68 PreservedMarks* preserved_marks_, 69 size_t desired_plab_sz_, 70 ParallelTaskTerminator& term_) : 71 _to_space(to_space_), 72 _old_gen(old_gen_), 73 _young_gen(young_gen_), 74 _thread_num(thread_num_), 75 _work_queue(work_queue_set_->queue(thread_num_)), 76 _to_space_full(false), 77 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), 78 _preserved_marks(preserved_marks_), 79 _ageTable(false), // false ==> not the global age table, no perf data. 80 _to_space_alloc_buffer(desired_plab_sz_), 81 _to_space_closure(young_gen_, this), 82 _old_gen_closure(young_gen_, this), 83 _to_space_root_closure(young_gen_, this), 84 _old_gen_root_closure(young_gen_, this), 85 _older_gen_closure(young_gen_, this), 86 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, 87 &_to_space_root_closure, young_gen_, &_old_gen_root_closure, 88 work_queue_set_, &term_), 89 _is_alive_closure(young_gen_), 90 _scan_weak_ref_closure(young_gen_, this), 91 _keep_alive_closure(&_scan_weak_ref_closure), 92 _strong_roots_time(0.0), 93 _term_time(0.0) 94 { 95 #if TASKQUEUE_STATS 96 _term_attempts = 0; 97 _overflow_refills = 0; 98 _overflow_refill_objs = 0; 99 #endif // TASKQUEUE_STATS 100 101 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num()); 102 _hash_seed = 17; // Might want to take time-based random value. 103 _start = os::elapsedTime(); 104 _old_gen_closure.set_generation(old_gen_); 105 _old_gen_root_closure.set_generation(old_gen_); 106 } 107 108 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, 109 size_t plab_word_size) { 110 ChunkArray* sca = survivor_chunk_array(); 111 if (sca != NULL) { 112 // A non-null SCA implies that we want the PLAB data recorded. 113 sca->record_sample(plab_start, plab_word_size); 114 } 115 } 116 117 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { 118 return new_obj->is_objArray() && 119 arrayOop(new_obj)->length() > ParGCArrayScanChunk && 120 new_obj != old_obj; 121 } 122 123 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { 124 assert(old->is_objArray(), "must be obj array"); 125 assert(old->is_forwarded(), "must be forwarded"); 126 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap."); 127 assert(!old_gen()->is_in(old), "must be in young generation."); 128 129 objArrayOop obj = objArrayOop(old->forwardee()); 130 // Process ParGCArrayScanChunk elements now 131 // and push the remainder back onto queue 132 int start = arrayOop(old)->length(); 133 int end = obj->length(); 134 int remainder = end - start; 135 assert(start <= end, "just checking"); 136 if (remainder > 2 * ParGCArrayScanChunk) { 137 // Test above combines last partial chunk with a full chunk 138 end = start + ParGCArrayScanChunk; 139 arrayOop(old)->set_length(end); 140 // Push remainder. 141 bool ok = work_queue()->push(old); 142 assert(ok, "just popped, push must be okay"); 143 } else { 144 // Restore length so that it can be used if there 145 // is a promotion failure and forwarding pointers 146 // must be removed. 147 arrayOop(old)->set_length(end); 148 } 149 150 // process our set of indices (include header in first chunk) 151 // should make sure end is even (aligned to HeapWord in case of compressed oops) 152 if ((HeapWord *)obj < young_old_boundary()) { 153 // object is in to_space 154 obj->oop_iterate_range(&_to_space_closure, start, end); 155 } else { 156 // object is in old generation 157 obj->oop_iterate_range(&_old_gen_closure, start, end); 158 } 159 } 160 161 void ParScanThreadState::trim_queues(int max_size) { 162 ObjToScanQueue* queue = work_queue(); 163 do { 164 while (queue->size() > (juint)max_size) { 165 oop obj_to_scan; 166 if (queue->pop_local(obj_to_scan)) { 167 if ((HeapWord *)obj_to_scan < young_old_boundary()) { 168 if (obj_to_scan->is_objArray() && 169 obj_to_scan->is_forwarded() && 170 obj_to_scan->forwardee() != obj_to_scan) { 171 scan_partial_array_and_push_remainder(obj_to_scan); 172 } else { 173 // object is in to_space 174 obj_to_scan->oop_iterate(&_to_space_closure); 175 } 176 } else { 177 // object is in old generation 178 obj_to_scan->oop_iterate(&_old_gen_closure); 179 } 180 } 181 } 182 // For the case of compressed oops, we have a private, non-shared 183 // overflow stack, so we eagerly drain it so as to more evenly 184 // distribute load early. Note: this may be good to do in 185 // general rather than delay for the final stealing phase. 186 // If applicable, we'll transfer a set of objects over to our 187 // work queue, allowing them to be stolen and draining our 188 // private overflow stack. 189 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); 190 } 191 192 bool ParScanThreadState::take_from_overflow_stack() { 193 assert(ParGCUseLocalOverflow, "Else should not call"); 194 assert(young_gen()->overflow_list() == NULL, "Error"); 195 ObjToScanQueue* queue = work_queue(); 196 Stack<oop, mtGC>* const of_stack = overflow_stack(); 197 const size_t num_overflow_elems = of_stack->size(); 198 const size_t space_available = queue->max_elems() - queue->size(); 199 const size_t num_take_elems = MIN3(space_available / 4, 200 ParGCDesiredObjsFromOverflowList, 201 num_overflow_elems); 202 // Transfer the most recent num_take_elems from the overflow 203 // stack to our work queue. 204 for (size_t i = 0; i != num_take_elems; i++) { 205 oop cur = of_stack->pop(); 206 oop obj_to_push = cur->forwardee(); 207 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap"); 208 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); 209 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap"); 210 if (should_be_partially_scanned(obj_to_push, cur)) { 211 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 212 obj_to_push = cur; 213 } 214 bool ok = queue->push(obj_to_push); 215 assert(ok, "Should have succeeded"); 216 } 217 assert(young_gen()->overflow_list() == NULL, "Error"); 218 return num_take_elems > 0; // was something transferred? 219 } 220 221 void ParScanThreadState::push_on_overflow_stack(oop p) { 222 assert(ParGCUseLocalOverflow, "Else should not call"); 223 overflow_stack()->push(p); 224 assert(young_gen()->overflow_list() == NULL, "Error"); 225 } 226 227 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { 228 // If the object is small enough, try to reallocate the buffer. 229 HeapWord* obj = NULL; 230 if (!_to_space_full) { 231 PLAB* const plab = to_space_alloc_buffer(); 232 Space* const sp = to_space(); 233 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) { 234 // Is small enough; abandon this buffer and start a new one. 235 plab->retire(); 236 // The minimum size has to be twice SurvivorAlignmentInBytes to 237 // allow for padding used in the alignment of 1 word. A padding 238 // of 1 is too small for a filler word so the padding size will 239 // be increased by SurvivorAlignmentInBytes. 240 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize); 241 size_t buf_size = MAX2(plab->word_sz(), min_usable_size); 242 HeapWord* buf_space = sp->par_allocate(buf_size); 243 if (buf_space == NULL) { 244 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize; 245 size_t free_bytes = sp->free(); 246 while(buf_space == NULL && free_bytes >= min_bytes) { 247 buf_size = free_bytes >> LogHeapWordSize; 248 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant"); 249 buf_space = sp->par_allocate(buf_size); 250 free_bytes = sp->free(); 251 } 252 } 253 if (buf_space != NULL) { 254 plab->set_buf(buf_space, buf_size); 255 record_survivor_plab(buf_space, buf_size); 256 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes); 257 // Note that we cannot compare buf_size < word_sz below 258 // because of AlignmentReserve (see PLAB::allocate()). 259 assert(obj != NULL || plab->words_remaining() < word_sz, 260 "Else should have been able to allocate requested object size " 261 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes " 262 SIZE_FORMAT ", words_remaining " SIZE_FORMAT, 263 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining()); 264 // It's conceivable that we may be able to use the 265 // buffer we just grabbed for subsequent small requests 266 // even if not for this one. 267 } else { 268 // We're used up. 269 _to_space_full = true; 270 } 271 } else { 272 // Too large; allocate the object individually. 273 obj = sp->par_allocate(word_sz); 274 } 275 } 276 return obj; 277 } 278 279 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) { 280 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 281 } 282 283 void ParScanThreadState::print_promotion_failure_size() { 284 if (_promotion_failed_info.has_failed()) { 285 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ", 286 _thread_num, _promotion_failed_info.first_size()); 287 } 288 } 289 290 class ParScanThreadStateSet: StackObj { 291 public: 292 // Initializes states for the specified number of threads; 293 ParScanThreadStateSet(int num_threads, 294 Space& to_space, 295 ParNewGeneration& young_gen, 296 Generation& old_gen, 297 ObjToScanQueueSet& queue_set, 298 Stack<oop, mtGC>* overflow_stacks_, 299 PreservedMarksSet& preserved_marks_set, 300 size_t desired_plab_sz, 301 ParallelTaskTerminator& term); 302 303 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 304 305 inline ParScanThreadState& thread_state(int i); 306 307 void trace_promotion_failed(const YoungGCTracer* gc_tracer); 308 void reset(uint active_workers, bool promotion_failed); 309 void flush(); 310 311 #if TASKQUEUE_STATS 312 static void 313 print_termination_stats_hdr(outputStream* const st); 314 void print_termination_stats(); 315 static void 316 print_taskqueue_stats_hdr(outputStream* const st); 317 void print_taskqueue_stats(); 318 void reset_stats(); 319 #endif // TASKQUEUE_STATS 320 321 private: 322 ParallelTaskTerminator& _term; 323 ParNewGeneration& _young_gen; 324 Generation& _old_gen; 325 ParScanThreadState* _per_thread_states; 326 const int _num_threads; 327 public: 328 bool is_valid(int id) const { return id < _num_threads; } 329 ParallelTaskTerminator* terminator() { return &_term; } 330 }; 331 332 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads, 333 Space& to_space, 334 ParNewGeneration& young_gen, 335 Generation& old_gen, 336 ObjToScanQueueSet& queue_set, 337 Stack<oop, mtGC>* overflow_stacks, 338 PreservedMarksSet& preserved_marks_set, 339 size_t desired_plab_sz, 340 ParallelTaskTerminator& term) 341 : _young_gen(young_gen), 342 _old_gen(old_gen), 343 _term(term), 344 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)), 345 _num_threads(num_threads) 346 { 347 assert(num_threads > 0, "sanity check!"); 348 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 349 "overflow_stack allocation mismatch"); 350 // Initialize states. 351 for (int i = 0; i < num_threads; ++i) { 352 new(_per_thread_states + i) 353 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set, 354 overflow_stacks, preserved_marks_set.get(i), 355 desired_plab_sz, term); 356 } 357 } 358 359 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) { 360 assert(i >= 0 && i < _num_threads, "sanity check!"); 361 return _per_thread_states[i]; 362 } 363 364 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) { 365 for (int i = 0; i < _num_threads; ++i) { 366 if (thread_state(i).promotion_failed()) { 367 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info()); 368 thread_state(i).promotion_failed_info().reset(); 369 } 370 } 371 } 372 373 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) { 374 _term.reset_for_reuse(active_threads); 375 if (promotion_failed) { 376 for (int i = 0; i < _num_threads; ++i) { 377 thread_state(i).print_promotion_failure_size(); 378 } 379 } 380 } 381 382 #if TASKQUEUE_STATS 383 void ParScanThreadState::reset_stats() { 384 taskqueue_stats().reset(); 385 _term_attempts = 0; 386 _overflow_refills = 0; 387 _overflow_refill_objs = 0; 388 } 389 390 void ParScanThreadStateSet::reset_stats() { 391 for (int i = 0; i < _num_threads; ++i) { 392 thread_state(i).reset_stats(); 393 } 394 } 395 396 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) { 397 st->print_raw_cr("GC Termination Stats"); 398 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"); 399 st->print_raw_cr("thr ms ms % ms % attempts"); 400 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"); 401 } 402 403 void ParScanThreadStateSet::print_termination_stats() { 404 Log(gc, task, stats) log; 405 if (!log.is_debug()) { 406 return; 407 } 408 409 ResourceMark rm; 410 outputStream* st = log.debug_stream(); 411 412 print_termination_stats_hdr(st); 413 414 for (int i = 0; i < _num_threads; ++i) { 415 const ParScanThreadState & pss = thread_state(i); 416 const double elapsed_ms = pss.elapsed_time() * 1000.0; 417 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 418 const double term_ms = pss.term_time() * 1000.0; 419 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8), 420 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 421 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); 422 } 423 } 424 425 // Print stats related to work queue activity. 426 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) { 427 st->print_raw_cr("GC Task Stats"); 428 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 429 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 430 } 431 432 void ParScanThreadStateSet::print_taskqueue_stats() { 433 if (!log_develop_is_enabled(Trace, gc, task, stats)) { 434 return; 435 } 436 Log(gc, task, stats) log; 437 ResourceMark rm; 438 outputStream* st = log.trace_stream(); 439 print_taskqueue_stats_hdr(st); 440 441 TaskQueueStats totals; 442 for (int i = 0; i < _num_threads; ++i) { 443 const ParScanThreadState & pss = thread_state(i); 444 const TaskQueueStats & stats = pss.taskqueue_stats(); 445 st->print("%3d ", i); stats.print(st); st->cr(); 446 totals += stats; 447 448 if (pss.overflow_refills() > 0) { 449 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 450 SIZE_FORMAT_W(10) " overflow objects", 451 pss.overflow_refills(), pss.overflow_refill_objs()); 452 } 453 } 454 st->print("tot "); totals.print(st); st->cr(); 455 456 DEBUG_ONLY(totals.verify()); 457 } 458 #endif // TASKQUEUE_STATS 459 460 void ParScanThreadStateSet::flush() { 461 // Work in this loop should be kept as lightweight as 462 // possible since this might otherwise become a bottleneck 463 // to scaling. Should we add heavy-weight work into this 464 // loop, consider parallelizing the loop into the worker threads. 465 for (int i = 0; i < _num_threads; ++i) { 466 ParScanThreadState& par_scan_state = thread_state(i); 467 468 // Flush stats related to To-space PLAB activity and 469 // retire the last buffer. 470 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats()); 471 472 // Every thread has its own age table. We need to merge 473 // them all into one. 474 AgeTable *local_table = par_scan_state.age_table(); 475 _young_gen.age_table()->merge(local_table); 476 477 // Inform old gen that we're done. 478 _old_gen.par_promote_alloc_done(i); 479 } 480 481 if (UseConcMarkSweepGC) { 482 // We need to call this even when ResizeOldPLAB is disabled 483 // so as to avoid breaking some asserts. While we may be able 484 // to avoid this by reorganizing the code a bit, I am loathe 485 // to do that unless we find cases where ergo leads to bad 486 // performance. 487 CompactibleFreeListSpaceLAB::compute_desired_plab_size(); 488 } 489 } 490 491 ParScanClosure::ParScanClosure(ParNewGeneration* g, 492 ParScanThreadState* par_scan_state) : 493 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) { 494 _boundary = _g->reserved().end(); 495 } 496 497 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } 498 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } 499 500 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } 501 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } 502 503 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 504 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 505 506 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 507 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 508 509 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 510 ParScanThreadState* par_scan_state) 511 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 512 {} 513 514 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } 515 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } 516 517 #ifdef WIN32 518 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ 519 #endif 520 521 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( 522 ParScanThreadState* par_scan_state_, 523 ParScanWithoutBarrierClosure* to_space_closure_, 524 ParScanWithBarrierClosure* old_gen_closure_, 525 ParRootScanWithoutBarrierClosure* to_space_root_closure_, 526 ParNewGeneration* par_gen_, 527 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, 528 ObjToScanQueueSet* task_queues_, 529 ParallelTaskTerminator* terminator_) : 530 531 _par_scan_state(par_scan_state_), 532 _to_space_closure(to_space_closure_), 533 _old_gen_closure(old_gen_closure_), 534 _to_space_root_closure(to_space_root_closure_), 535 _old_gen_root_closure(old_gen_root_closure_), 536 _par_gen(par_gen_), 537 _task_queues(task_queues_), 538 _terminator(terminator_) 539 {} 540 541 void ParEvacuateFollowersClosure::do_void() { 542 ObjToScanQueue* work_q = par_scan_state()->work_queue(); 543 544 while (true) { 545 // Scan to-space and old-gen objs until we run out of both. 546 oop obj_to_scan; 547 par_scan_state()->trim_queues(0); 548 549 // We have no local work, attempt to steal from other threads. 550 551 // Attempt to steal work from promoted. 552 if (task_queues()->steal(par_scan_state()->thread_num(), 553 par_scan_state()->hash_seed(), 554 obj_to_scan)) { 555 bool res = work_q->push(obj_to_scan); 556 assert(res, "Empty queue should have room for a push."); 557 558 // If successful, goto Start. 559 continue; 560 561 // Try global overflow list. 562 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 563 continue; 564 } 565 566 // Otherwise, offer termination. 567 par_scan_state()->start_term_time(); 568 if (terminator()->offer_termination()) break; 569 par_scan_state()->end_term_time(); 570 } 571 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 572 "Broken overflow list?"); 573 // Finish the last termination pause. 574 par_scan_state()->end_term_time(); 575 } 576 577 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen, 578 Generation* old_gen, 579 HeapWord* young_old_boundary, 580 ParScanThreadStateSet* state_set, 581 StrongRootsScope* strong_roots_scope) : 582 AbstractGangTask("ParNewGeneration collection"), 583 _young_gen(young_gen), _old_gen(old_gen), 584 _young_old_boundary(young_old_boundary), 585 _state_set(state_set), 586 _strong_roots_scope(strong_roots_scope) 587 {} 588 589 void ParNewGenTask::work(uint worker_id) { 590 GenCollectedHeap* gch = GenCollectedHeap::heap(); 591 // Since this is being done in a separate thread, need new resource 592 // and handle marks. 593 ResourceMark rm; 594 HandleMark hm; 595 596 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 597 assert(_state_set->is_valid(worker_id), "Should not have been called"); 598 599 par_scan_state.set_young_old_boundary(_young_old_boundary); 600 601 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), 602 gch->rem_set()->klass_rem_set()); 603 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure, 604 &par_scan_state.to_space_root_closure(), 605 false); 606 607 par_scan_state.start_strong_roots(); 608 gch->gen_process_roots(_strong_roots_scope, 609 GenCollectedHeap::YoungGen, 610 true, // Process younger gens, if any, as strong roots. 611 GenCollectedHeap::SO_ScavengeCodeCache, 612 GenCollectedHeap::StrongAndWeakRoots, 613 &par_scan_state.to_space_root_closure(), 614 &par_scan_state.older_gen_closure(), 615 &cld_scan_closure); 616 617 par_scan_state.end_strong_roots(); 618 619 // "evacuate followers". 620 par_scan_state.evacuate_followers_closure().do_void(); 621 622 // This will collapse this worker's promoted object list that's 623 // created during the main ParNew parallel phase of ParNew. This has 624 // to be called after all workers have finished promoting objects 625 // and scanning promoted objects. It should be safe calling it from 626 // here, given that we can only reach here after all thread have 627 // offered termination, i.e., after there is no more work to be 628 // done. It will also disable promotion tracking for the rest of 629 // this GC as it's not necessary to be on during reference processing. 630 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id); 631 } 632 633 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size) 634 : DefNewGeneration(rs, initial_byte_size, "PCopy"), 635 _overflow_list(NULL), 636 _is_alive_closure(this), 637 _plab_stats("Young", YoungPLABSize, PLABWeight) 638 { 639 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 640 NOT_PRODUCT(_num_par_pushes = 0;) 641 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 642 guarantee(_task_queues != NULL, "task_queues allocation failure."); 643 644 for (uint i = 0; i < ParallelGCThreads; i++) { 645 ObjToScanQueue *q = new ObjToScanQueue(); 646 guarantee(q != NULL, "work_queue Allocation failure."); 647 _task_queues->register_queue(i, q); 648 } 649 650 for (uint i = 0; i < ParallelGCThreads; i++) { 651 _task_queues->queue(i)->initialize(); 652 } 653 654 _overflow_stacks = NULL; 655 if (ParGCUseLocalOverflow) { 656 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ',' 657 typedef Stack<oop, mtGC> GCOopStack; 658 659 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 660 for (size_t i = 0; i < ParallelGCThreads; ++i) { 661 new (_overflow_stacks + i) Stack<oop, mtGC>(); 662 } 663 } 664 665 if (UsePerfData) { 666 EXCEPTION_MARK; 667 ResourceMark rm; 668 669 const char* cname = 670 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 671 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 672 ParallelGCThreads, CHECK); 673 } 674 } 675 676 // ParNewGeneration:: 677 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 678 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 679 680 template <class T> 681 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 682 #ifdef ASSERT 683 { 684 assert(!oopDesc::is_null(*p), "expected non-null ref"); 685 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 686 // We never expect to see a null reference being processed 687 // as a weak reference. 688 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 689 } 690 #endif // ASSERT 691 692 _par_cl->do_oop_nv(p); 693 694 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 695 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 696 _rs->write_ref_field_gc_par(p, obj); 697 } 698 } 699 700 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 701 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 702 703 // ParNewGeneration:: 704 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 705 DefNewGeneration::KeepAliveClosure(cl) {} 706 707 template <class T> 708 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 709 #ifdef ASSERT 710 { 711 assert(!oopDesc::is_null(*p), "expected non-null ref"); 712 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 713 // We never expect to see a null reference being processed 714 // as a weak reference. 715 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 716 } 717 #endif // ASSERT 718 719 _cl->do_oop_nv(p); 720 721 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 722 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 723 _rs->write_ref_field_gc_par(p, obj); 724 } 725 } 726 727 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 728 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 729 730 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 731 T heap_oop = oopDesc::load_heap_oop(p); 732 if (!oopDesc::is_null(heap_oop)) { 733 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 734 if ((HeapWord*)obj < _boundary) { 735 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 736 oop new_obj = obj->is_forwarded() 737 ? obj->forwardee() 738 : _g->DefNewGeneration::copy_to_survivor_space(obj); 739 oopDesc::encode_store_heap_oop_not_null(p, new_obj); 740 } 741 if (_gc_barrier) { 742 // If p points to a younger generation, mark the card. 743 if ((HeapWord*)obj < _gen_boundary) { 744 _rs->write_ref_field_gc_par(p, obj); 745 } 746 } 747 } 748 } 749 750 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 751 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 752 753 class ParNewRefProcTaskProxy: public AbstractGangTask { 754 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 755 public: 756 ParNewRefProcTaskProxy(ProcessTask& task, 757 ParNewGeneration& young_gen, 758 Generation& old_gen, 759 HeapWord* young_old_boundary, 760 ParScanThreadStateSet& state_set); 761 762 private: 763 virtual void work(uint worker_id); 764 private: 765 ParNewGeneration& _young_gen; 766 ProcessTask& _task; 767 Generation& _old_gen; 768 HeapWord* _young_old_boundary; 769 ParScanThreadStateSet& _state_set; 770 }; 771 772 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task, 773 ParNewGeneration& young_gen, 774 Generation& old_gen, 775 HeapWord* young_old_boundary, 776 ParScanThreadStateSet& state_set) 777 : AbstractGangTask("ParNewGeneration parallel reference processing"), 778 _young_gen(young_gen), 779 _task(task), 780 _old_gen(old_gen), 781 _young_old_boundary(young_old_boundary), 782 _state_set(state_set) 783 { } 784 785 void ParNewRefProcTaskProxy::work(uint worker_id) { 786 ResourceMark rm; 787 HandleMark hm; 788 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 789 par_scan_state.set_young_old_boundary(_young_old_boundary); 790 _task.work(worker_id, par_scan_state.is_alive_closure(), 791 par_scan_state.keep_alive_closure(), 792 par_scan_state.evacuate_followers_closure()); 793 } 794 795 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { 796 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 797 EnqueueTask& _task; 798 799 public: 800 ParNewRefEnqueueTaskProxy(EnqueueTask& task) 801 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), 802 _task(task) 803 { } 804 805 virtual void work(uint worker_id) { 806 _task.work(worker_id); 807 } 808 }; 809 810 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) { 811 GenCollectedHeap* gch = GenCollectedHeap::heap(); 812 WorkGang* workers = gch->workers(); 813 assert(workers != NULL, "Need parallel worker threads."); 814 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed()); 815 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen, 816 _young_gen.reserved().end(), _state_set); 817 workers->run_task(&rp_task); 818 _state_set.reset(0 /* bad value in debug if not reset */, 819 _young_gen.promotion_failed()); 820 } 821 822 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) { 823 GenCollectedHeap* gch = GenCollectedHeap::heap(); 824 WorkGang* workers = gch->workers(); 825 assert(workers != NULL, "Need parallel worker threads."); 826 ParNewRefEnqueueTaskProxy enq_task(task); 827 workers->run_task(&enq_task); 828 } 829 830 void ParNewRefProcTaskExecutor::set_single_threaded_mode() { 831 _state_set.flush(); 832 GenCollectedHeap* gch = GenCollectedHeap::heap(); 833 gch->save_marks(); 834 } 835 836 ScanClosureWithParBarrier:: 837 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 838 ScanClosure(g, gc_barrier) 839 { } 840 841 EvacuateFollowersClosureGeneral:: 842 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, 843 OopsInGenClosure* cur, 844 OopsInGenClosure* older) : 845 _gch(gch), 846 _scan_cur_or_nonheap(cur), _scan_older(older) 847 { } 848 849 void EvacuateFollowersClosureGeneral::do_void() { 850 do { 851 // Beware: this call will lead to closure applications via virtual 852 // calls. 853 _gch->oop_since_save_marks_iterate(GenCollectedHeap::YoungGen, 854 _scan_cur_or_nonheap, 855 _scan_older); 856 } while (!_gch->no_allocs_since_save_marks()); 857 } 858 859 // A Generation that does parallel young-gen collection. 860 861 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) { 862 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 863 _promo_failure_scan_stack.clear(true); // Clear cached segments. 864 865 remove_forwarding_pointers(); 866 log_info(gc, promotion)("Promotion failed"); 867 // All the spaces are in play for mark-sweep. 868 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 869 from()->set_next_compaction_space(to()); 870 gch->set_incremental_collection_failed(); 871 // Inform the next generation that a promotion failure occurred. 872 _old_gen->promotion_failure_occurred(); 873 874 // Trace promotion failure in the parallel GC threads 875 thread_state_set.trace_promotion_failed(gc_tracer()); 876 // Single threaded code may have reported promotion failure to the global state 877 if (_promotion_failed_info.has_failed()) { 878 _gc_tracer.report_promotion_failed(_promotion_failed_info); 879 } 880 // Reset the PromotionFailureALot counters. 881 NOT_PRODUCT(gch->reset_promotion_should_fail();) 882 } 883 884 void ParNewGeneration::collect(bool full, 885 bool clear_all_soft_refs, 886 size_t size, 887 bool is_tlab) { 888 assert(full || size > 0, "otherwise we don't want to collect"); 889 890 GenCollectedHeap* gch = GenCollectedHeap::heap(); 891 892 _gc_timer->register_gc_start(); 893 894 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); 895 WorkGang* workers = gch->workers(); 896 assert(workers != NULL, "Need workgang for parallel work"); 897 uint active_workers = 898 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), 899 workers->active_workers(), 900 Threads::number_of_non_daemon_threads()); 901 active_workers = workers->update_active_workers(active_workers); 902 _old_gen = gch->old_gen(); 903 904 // If the next generation is too full to accommodate worst-case promotion 905 // from this generation, pass on collection; let the next generation 906 // do it. 907 if (!collection_attempt_is_safe()) { 908 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 909 return; 910 } 911 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 912 913 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 914 gch->trace_heap_before_gc(gc_tracer()); 915 916 init_assuming_no_promotion_failure(); 917 918 if (UseAdaptiveSizePolicy) { 919 set_survivor_overflow(false); 920 size_policy->minor_collection_begin(); 921 } 922 923 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause()); 924 925 age_table()->clear(); 926 to()->clear(SpaceDecorator::Mangle); 927 928 gch->save_marks(); 929 930 // Set the correct parallelism (number of queues) in the reference processor 931 ref_processor()->set_active_mt_degree(active_workers); 932 933 // Need to initialize the preserved marks before the ThreadStateSet c'tor. 934 _preserved_marks_set.init(active_workers); 935 936 // Always set the terminator for the active number of workers 937 // because only those workers go through the termination protocol. 938 ParallelTaskTerminator _term(active_workers, task_queues()); 939 ParScanThreadStateSet thread_state_set(active_workers, 940 *to(), *this, *_old_gen, *task_queues(), 941 _overflow_stacks, _preserved_marks_set, 942 desired_plab_sz(), _term); 943 944 thread_state_set.reset(active_workers, promotion_failed()); 945 946 { 947 StrongRootsScope srs(active_workers); 948 949 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs); 950 gch->rem_set()->prepare_for_younger_refs_iterate(true); 951 // It turns out that even when we're using 1 thread, doing the work in a 952 // separate thread causes wide variance in run times. We can't help this 953 // in the multi-threaded case, but we special-case n=1 here to get 954 // repeatable measurements of the 1-thread overhead of the parallel code. 955 // Might multiple workers ever be used? If yes, initialization 956 // has been done such that the single threaded path should not be used. 957 if (workers->total_workers() > 1) { 958 workers->run_task(&tsk); 959 } else { 960 tsk.work(0); 961 } 962 } 963 964 thread_state_set.reset(0 /* Bad value in debug if not reset */, 965 promotion_failed()); 966 967 // Trace and reset failed promotion info. 968 if (promotion_failed()) { 969 thread_state_set.trace_promotion_failed(gc_tracer()); 970 } 971 972 // Process (weak) reference objects found during scavenge. 973 ReferenceProcessor* rp = ref_processor(); 974 IsAliveClosure is_alive(this); 975 ScanWeakRefClosure scan_weak_ref(this); 976 KeepAliveClosure keep_alive(&scan_weak_ref); 977 ScanClosure scan_without_gc_barrier(this, false); 978 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 979 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 980 EvacuateFollowersClosureGeneral evacuate_followers(gch, 981 &scan_without_gc_barrier, &scan_with_gc_barrier); 982 rp->setup_policy(clear_all_soft_refs); 983 // Can the mt_degree be set later (at run_task() time would be best)? 984 rp->set_active_mt_degree(active_workers); 985 ReferenceProcessorStats stats; 986 if (rp->processing_is_mt()) { 987 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 988 stats = rp->process_discovered_references(&is_alive, &keep_alive, 989 &evacuate_followers, &task_executor, 990 _gc_timer); 991 } else { 992 thread_state_set.flush(); 993 gch->save_marks(); 994 stats = rp->process_discovered_references(&is_alive, &keep_alive, 995 &evacuate_followers, NULL, 996 _gc_timer); 997 } 998 _gc_tracer.report_gc_reference_stats(stats); 999 _gc_tracer.report_tenuring_threshold(tenuring_threshold()); 1000 1001 if (!promotion_failed()) { 1002 // Swap the survivor spaces. 1003 eden()->clear(SpaceDecorator::Mangle); 1004 from()->clear(SpaceDecorator::Mangle); 1005 if (ZapUnusedHeapArea) { 1006 // This is now done here because of the piece-meal mangling which 1007 // can check for valid mangling at intermediate points in the 1008 // collection(s). When a young collection fails to collect 1009 // sufficient space resizing of the young generation can occur 1010 // and redistribute the spaces in the young generation. Mangle 1011 // here so that unzapped regions don't get distributed to 1012 // other spaces. 1013 to()->mangle_unused_area(); 1014 } 1015 swap_spaces(); 1016 1017 // A successful scavenge should restart the GC time limit count which is 1018 // for full GC's. 1019 size_policy->reset_gc_overhead_limit_count(); 1020 1021 assert(to()->is_empty(), "to space should be empty now"); 1022 1023 adjust_desired_tenuring_threshold(); 1024 } else { 1025 handle_promotion_failed(gch, thread_state_set); 1026 } 1027 _preserved_marks_set.reclaim(); 1028 // set new iteration safe limit for the survivor spaces 1029 from()->set_concurrent_iteration_safe_limit(from()->top()); 1030 to()->set_concurrent_iteration_safe_limit(to()->top()); 1031 1032 plab_stats()->adjust_desired_plab_sz(); 1033 1034 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); 1035 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); 1036 1037 if (UseAdaptiveSizePolicy) { 1038 size_policy->minor_collection_end(gch->gc_cause()); 1039 size_policy->avg_survived()->sample(from()->used()); 1040 } 1041 1042 // We need to use a monotonically non-decreasing time in ms 1043 // or we will see time-warp warnings and os::javaTimeMillis() 1044 // does not guarantee monotonicity. 1045 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1046 update_time_of_last_gc(now); 1047 1048 rp->set_enqueuing_is_done(true); 1049 if (rp->processing_is_mt()) { 1050 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 1051 rp->enqueue_discovered_references(&task_executor); 1052 } else { 1053 rp->enqueue_discovered_references(NULL); 1054 } 1055 rp->verify_no_references_recorded(); 1056 1057 gch->trace_heap_after_gc(gc_tracer()); 1058 1059 _gc_timer->register_gc_end(); 1060 1061 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1062 } 1063 1064 size_t ParNewGeneration::desired_plab_sz() { 1065 return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers()); 1066 } 1067 1068 static int sum; 1069 void ParNewGeneration::waste_some_time() { 1070 for (int i = 0; i < 100; i++) { 1071 sum += i; 1072 } 1073 } 1074 1075 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); 1076 1077 // Because of concurrency, there are times where an object for which 1078 // "is_forwarded()" is true contains an "interim" forwarding pointer 1079 // value. Such a value will soon be overwritten with a real value. 1080 // This method requires "obj" to have a forwarding pointer, and waits, if 1081 // necessary for a real one to be inserted, and returns it. 1082 1083 oop ParNewGeneration::real_forwardee(oop obj) { 1084 oop forward_ptr = obj->forwardee(); 1085 if (forward_ptr != ClaimedForwardPtr) { 1086 return forward_ptr; 1087 } else { 1088 return real_forwardee_slow(obj); 1089 } 1090 } 1091 1092 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1093 // Spin-read if it is claimed but not yet written by another thread. 1094 oop forward_ptr = obj->forwardee(); 1095 while (forward_ptr == ClaimedForwardPtr) { 1096 waste_some_time(); 1097 assert(obj->is_forwarded(), "precondition"); 1098 forward_ptr = obj->forwardee(); 1099 } 1100 return forward_ptr; 1101 } 1102 1103 // Multiple GC threads may try to promote an object. If the object 1104 // is successfully promoted, a forwarding pointer will be installed in 1105 // the object in the young generation. This method claims the right 1106 // to install the forwarding pointer before it copies the object, 1107 // thus avoiding the need to undo the copy as in 1108 // copy_to_survivor_space_avoiding_with_undo. 1109 1110 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state, 1111 oop old, 1112 size_t sz, 1113 markOop m) { 1114 // In the sequential version, this assert also says that the object is 1115 // not forwarded. That might not be the case here. It is the case that 1116 // the caller observed it to be not forwarded at some time in the past. 1117 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1118 1119 // The sequential code read "old->age()" below. That doesn't work here, 1120 // since the age is in the mark word, and that might be overwritten with 1121 // a forwarding pointer by a parallel thread. So we must save the mark 1122 // word in a local and then analyze it. 1123 oopDesc dummyOld; 1124 dummyOld.set_mark(m); 1125 assert(!dummyOld.is_forwarded(), 1126 "should not be called with forwarding pointer mark word."); 1127 1128 oop new_obj = NULL; 1129 oop forward_ptr; 1130 1131 // Try allocating obj in to-space (unless too old) 1132 if (dummyOld.age() < tenuring_threshold()) { 1133 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1134 if (new_obj == NULL) { 1135 set_survivor_overflow(true); 1136 } 1137 } 1138 1139 if (new_obj == NULL) { 1140 // Either to-space is full or we decided to promote try allocating obj tenured 1141 1142 // Attempt to install a null forwarding pointer (atomically), 1143 // to claim the right to install the real forwarding pointer. 1144 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); 1145 if (forward_ptr != NULL) { 1146 // someone else beat us to it. 1147 return real_forwardee(old); 1148 } 1149 1150 if (!_promotion_failed) { 1151 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), 1152 old, m, sz); 1153 } 1154 1155 if (new_obj == NULL) { 1156 // promotion failed, forward to self 1157 _promotion_failed = true; 1158 new_obj = old; 1159 1160 par_scan_state->preserved_marks()->push_if_necessary(old, m); 1161 par_scan_state->register_promotion_failure(sz); 1162 } 1163 1164 old->forward_to(new_obj); 1165 forward_ptr = NULL; 1166 } else { 1167 // Is in to-space; do copying ourselves. 1168 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1169 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value."); 1170 forward_ptr = old->forward_to_atomic(new_obj); 1171 // Restore the mark word copied above. 1172 new_obj->set_mark(m); 1173 // Increment age if obj still in new generation 1174 new_obj->incr_age(); 1175 par_scan_state->age_table()->add(new_obj, sz); 1176 } 1177 assert(new_obj != NULL, "just checking"); 1178 1179 // This code must come after the CAS test, or it will print incorrect 1180 // information. 1181 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1182 is_in_reserved(new_obj) ? "copying" : "tenuring", 1183 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size()); 1184 1185 if (forward_ptr == NULL) { 1186 oop obj_to_push = new_obj; 1187 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1188 // Length field used as index of next element to be scanned. 1189 // Real length can be obtained from real_forwardee() 1190 arrayOop(old)->set_length(0); 1191 obj_to_push = old; 1192 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1193 "push forwarded object"); 1194 } 1195 // Push it on one of the queues of to-be-scanned objects. 1196 bool simulate_overflow = false; 1197 NOT_PRODUCT( 1198 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1199 // simulate a stack overflow 1200 simulate_overflow = true; 1201 } 1202 ) 1203 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1204 // Add stats for overflow pushes. 1205 log_develop_trace(gc)("Queue Overflow"); 1206 push_on_overflow_list(old, par_scan_state); 1207 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1208 } 1209 1210 return new_obj; 1211 } 1212 1213 // Oops. Someone beat us to it. Undo the allocation. Where did we 1214 // allocate it? 1215 if (is_in_reserved(new_obj)) { 1216 // Must be in to_space. 1217 assert(to()->is_in_reserved(new_obj), "Checking"); 1218 if (forward_ptr == ClaimedForwardPtr) { 1219 // Wait to get the real forwarding pointer value. 1220 forward_ptr = real_forwardee(old); 1221 } 1222 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1223 } 1224 1225 return forward_ptr; 1226 } 1227 1228 #ifndef PRODUCT 1229 // It's OK to call this multi-threaded; the worst thing 1230 // that can happen is that we'll get a bunch of closely 1231 // spaced simulated overflows, but that's OK, in fact 1232 // probably good as it would exercise the overflow code 1233 // under contention. 1234 bool ParNewGeneration::should_simulate_overflow() { 1235 if (_overflow_counter-- <= 0) { // just being defensive 1236 _overflow_counter = ParGCWorkQueueOverflowInterval; 1237 return true; 1238 } else { 1239 return false; 1240 } 1241 } 1242 #endif 1243 1244 // In case we are using compressed oops, we need to be careful. 1245 // If the object being pushed is an object array, then its length 1246 // field keeps track of the "grey boundary" at which the next 1247 // incremental scan will be done (see ParGCArrayScanChunk). 1248 // When using compressed oops, this length field is kept in the 1249 // lower 32 bits of the erstwhile klass word and cannot be used 1250 // for the overflow chaining pointer (OCP below). As such the OCP 1251 // would itself need to be compressed into the top 32-bits in this 1252 // case. Unfortunately, see below, in the event that we have a 1253 // promotion failure, the node to be pushed on the list can be 1254 // outside of the Java heap, so the heap-based pointer compression 1255 // would not work (we would have potential aliasing between C-heap 1256 // and Java-heap pointers). For this reason, when using compressed 1257 // oops, we simply use a worker-thread-local, non-shared overflow 1258 // list in the form of a growable array, with a slightly different 1259 // overflow stack draining strategy. If/when we start using fat 1260 // stacks here, we can go back to using (fat) pointer chains 1261 // (although some performance comparisons would be useful since 1262 // single global lists have their own performance disadvantages 1263 // as we were made painfully aware not long ago, see 6786503). 1264 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 1265 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1266 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1267 if (ParGCUseLocalOverflow) { 1268 // In the case of compressed oops, we use a private, not-shared 1269 // overflow stack. 1270 par_scan_state->push_on_overflow_stack(from_space_obj); 1271 } else { 1272 assert(!UseCompressedOops, "Error"); 1273 // if the object has been forwarded to itself, then we cannot 1274 // use the klass pointer for the linked list. Instead we have 1275 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1276 // XXX This is horribly inefficient when a promotion failure occurs 1277 // and should be fixed. XXX FIX ME !!! 1278 #ifndef PRODUCT 1279 Atomic::inc_ptr(&_num_par_pushes); 1280 assert(_num_par_pushes > 0, "Tautology"); 1281 #endif 1282 if (from_space_obj->forwardee() == from_space_obj) { 1283 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); 1284 listhead->forward_to(from_space_obj); 1285 from_space_obj = listhead; 1286 } 1287 oop observed_overflow_list = _overflow_list; 1288 oop cur_overflow_list; 1289 do { 1290 cur_overflow_list = observed_overflow_list; 1291 if (cur_overflow_list != BUSY) { 1292 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1293 } else { 1294 from_space_obj->set_klass_to_list_ptr(NULL); 1295 } 1296 observed_overflow_list = 1297 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); 1298 } while (cur_overflow_list != observed_overflow_list); 1299 } 1300 } 1301 1302 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1303 bool res; 1304 1305 if (ParGCUseLocalOverflow) { 1306 res = par_scan_state->take_from_overflow_stack(); 1307 } else { 1308 assert(!UseCompressedOops, "Error"); 1309 res = take_from_overflow_list_work(par_scan_state); 1310 } 1311 return res; 1312 } 1313 1314 1315 // *NOTE*: The overflow list manipulation code here and 1316 // in CMSCollector:: are very similar in shape, 1317 // except that in the CMS case we thread the objects 1318 // directly into the list via their mark word, and do 1319 // not need to deal with special cases below related 1320 // to chunking of object arrays and promotion failure 1321 // handling. 1322 // CR 6797058 has been filed to attempt consolidation of 1323 // the common code. 1324 // Because of the common code, if you make any changes in 1325 // the code below, please check the CMS version to see if 1326 // similar changes might be needed. 1327 // See CMSCollector::par_take_from_overflow_list() for 1328 // more extensive documentation comments. 1329 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1330 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1331 // How many to take? 1332 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1333 (size_t)ParGCDesiredObjsFromOverflowList); 1334 1335 assert(!UseCompressedOops, "Error"); 1336 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1337 if (_overflow_list == NULL) return false; 1338 1339 // Otherwise, there was something there; try claiming the list. 1340 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1341 // Trim off a prefix of at most objsFromOverflow items 1342 Thread* tid = Thread::current(); 1343 size_t spin_count = ParallelGCThreads; 1344 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1345 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1346 // someone grabbed it before we did ... 1347 // ... we spin for a short while... 1348 os::sleep(tid, sleep_time_millis, false); 1349 if (_overflow_list == NULL) { 1350 // nothing left to take 1351 return false; 1352 } else if (_overflow_list != BUSY) { 1353 // try and grab the prefix 1354 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1355 } 1356 } 1357 if (prefix == NULL || prefix == BUSY) { 1358 // Nothing to take or waited long enough 1359 if (prefix == NULL) { 1360 // Write back the NULL in case we overwrote it with BUSY above 1361 // and it is still the same value. 1362 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1363 } 1364 return false; 1365 } 1366 assert(prefix != NULL && prefix != BUSY, "Error"); 1367 size_t i = 1; 1368 oop cur = prefix; 1369 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { 1370 i++; cur = cur->list_ptr_from_klass(); 1371 } 1372 1373 // Reattach remaining (suffix) to overflow list 1374 if (cur->klass_or_null() == NULL) { 1375 // Write back the NULL in lieu of the BUSY we wrote 1376 // above and it is still the same value. 1377 if (_overflow_list == BUSY) { 1378 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1379 } 1380 } else { 1381 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); 1382 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list 1383 cur->set_klass_to_list_ptr(NULL); // break off suffix 1384 // It's possible that the list is still in the empty(busy) state 1385 // we left it in a short while ago; in that case we may be 1386 // able to place back the suffix. 1387 oop observed_overflow_list = _overflow_list; 1388 oop cur_overflow_list = observed_overflow_list; 1389 bool attached = false; 1390 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1391 observed_overflow_list = 1392 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1393 if (cur_overflow_list == observed_overflow_list) { 1394 attached = true; 1395 break; 1396 } else cur_overflow_list = observed_overflow_list; 1397 } 1398 if (!attached) { 1399 // Too bad, someone else got in in between; we'll need to do a splice. 1400 // Find the last item of suffix list 1401 oop last = suffix; 1402 while (last->klass_or_null() != NULL) { 1403 last = last->list_ptr_from_klass(); 1404 } 1405 // Atomically prepend suffix to current overflow list 1406 observed_overflow_list = _overflow_list; 1407 do { 1408 cur_overflow_list = observed_overflow_list; 1409 if (cur_overflow_list != BUSY) { 1410 // Do the splice ... 1411 last->set_klass_to_list_ptr(cur_overflow_list); 1412 } else { // cur_overflow_list == BUSY 1413 last->set_klass_to_list_ptr(NULL); 1414 } 1415 observed_overflow_list = 1416 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1417 } while (cur_overflow_list != observed_overflow_list); 1418 } 1419 } 1420 1421 // Push objects on prefix list onto this thread's work queue 1422 assert(prefix != NULL && prefix != BUSY, "program logic"); 1423 cur = prefix; 1424 ssize_t n = 0; 1425 while (cur != NULL) { 1426 oop obj_to_push = cur->forwardee(); 1427 oop next = cur->list_ptr_from_klass(); 1428 cur->set_klass(obj_to_push->klass()); 1429 // This may be an array object that is self-forwarded. In that case, the list pointer 1430 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1431 if (!is_in_reserved(cur)) { 1432 // This can become a scaling bottleneck when there is work queue overflow coincident 1433 // with promotion failure. 1434 oopDesc* f = cur; 1435 FREE_C_HEAP_ARRAY(oopDesc, f); 1436 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1437 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1438 obj_to_push = cur; 1439 } 1440 bool ok = work_q->push(obj_to_push); 1441 assert(ok, "Should have succeeded"); 1442 cur = next; 1443 n++; 1444 } 1445 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1446 #ifndef PRODUCT 1447 assert(_num_par_pushes >= n, "Too many pops?"); 1448 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 1449 #endif 1450 return true; 1451 } 1452 #undef BUSY 1453 1454 void ParNewGeneration::ref_processor_init() { 1455 if (_ref_processor == NULL) { 1456 // Allocate and initialize a reference processor 1457 _ref_processor = 1458 new ReferenceProcessor(_reserved, // span 1459 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1460 ParallelGCThreads, // mt processing degree 1461 refs_discovery_is_mt(), // mt discovery 1462 ParallelGCThreads, // mt discovery degree 1463 refs_discovery_is_atomic(), // atomic_discovery 1464 NULL); // is_alive_non_header 1465 } 1466 } 1467 1468 const char* ParNewGeneration::name() const { 1469 return "par new generation"; 1470 }