1 /* 2 * Copyright (c) 2001, 2019, 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 "classfile/stringTable.hpp" 27 #include "gc/cms/cmsHeap.inline.hpp" 28 #include "gc/cms/compactibleFreeListSpace.hpp" 29 #include "gc/cms/concurrentMarkSweepGeneration.hpp" 30 #include "gc/cms/parNewGeneration.inline.hpp" 31 #include "gc/cms/parOopClosures.inline.hpp" 32 #include "gc/serial/defNewGeneration.inline.hpp" 33 #include "gc/shared/adaptiveSizePolicy.hpp" 34 #include "gc/shared/ageTable.inline.hpp" 35 #include "gc/shared/copyFailedInfo.hpp" 36 #include "gc/shared/gcHeapSummary.hpp" 37 #include "gc/shared/gcTimer.hpp" 38 #include "gc/shared/gcTrace.hpp" 39 #include "gc/shared/gcTraceTime.inline.hpp" 40 #include "gc/shared/genOopClosures.inline.hpp" 41 #include "gc/shared/generation.hpp" 42 #include "gc/shared/plab.inline.hpp" 43 #include "gc/shared/preservedMarks.inline.hpp" 44 #include "gc/shared/referencePolicy.hpp" 45 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 46 #include "gc/shared/space.hpp" 47 #include "gc/shared/spaceDecorator.hpp" 48 #include "gc/shared/strongRootsScope.hpp" 49 #include "gc/shared/taskqueue.inline.hpp" 50 #include "gc/shared/weakProcessor.hpp" 51 #include "gc/shared/workgroup.hpp" 52 #include "gc/shared/workerPolicy.hpp" 53 #include "logging/log.hpp" 54 #include "logging/logStream.hpp" 55 #include "memory/iterator.inline.hpp" 56 #include "memory/resourceArea.hpp" 57 #include "oops/access.inline.hpp" 58 #include "oops/compressedOops.inline.hpp" 59 #include "oops/objArrayOop.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "runtime/atomic.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/thread.inline.hpp" 65 #include "utilities/copy.hpp" 66 #include "utilities/globalDefinitions.hpp" 67 #include "utilities/stack.inline.hpp" 68 69 ParScanThreadState::ParScanThreadState(Space* to_space_, 70 ParNewGeneration* young_gen_, 71 Generation* old_gen_, 72 int thread_num_, 73 ObjToScanQueueSet* work_queue_set_, 74 Stack<oop, mtGC>* overflow_stacks_, 75 PreservedMarks* preserved_marks_, 76 size_t desired_plab_sz_, 77 TaskTerminator& term_) : 78 _work_queue(work_queue_set_->queue(thread_num_)), 79 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), 80 _preserved_marks(preserved_marks_), 81 _to_space_alloc_buffer(desired_plab_sz_), 82 _to_space_closure(young_gen_, this), 83 _old_gen_closure(young_gen_, this), 84 _to_space_root_closure(young_gen_, this), 85 _older_gen_closure(young_gen_, this), 86 _old_gen_root_closure(young_gen_, this), 87 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, 88 &_to_space_root_closure, young_gen_, &_old_gen_root_closure, 89 work_queue_set_, term_.terminator()), 90 _is_alive_closure(young_gen_), 91 _scan_weak_ref_closure(young_gen_, this), 92 _keep_alive_closure(&_scan_weak_ref_closure), 93 _to_space(to_space_), 94 _young_gen(young_gen_), 95 _old_gen(old_gen_), 96 _young_old_boundary(NULL), 97 _thread_num(thread_num_), 98 _ageTable(false), // false ==> not the global age table, no perf data. 99 _to_space_full(false), 100 _strong_roots_time(0.0), 101 _term_time(0.0) 102 { 103 #if TASKQUEUE_STATS 104 _term_attempts = 0; 105 _overflow_refills = 0; 106 _overflow_refill_objs = 0; 107 #endif // TASKQUEUE_STATS 108 109 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num()); 110 _start = os::elapsedTime(); 111 _old_gen_closure.set_generation(old_gen_); 112 _old_gen_root_closure.set_generation(old_gen_); 113 } 114 115 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, 116 size_t plab_word_size) { 117 ChunkArray* sca = survivor_chunk_array(); 118 if (sca != NULL) { 119 // A non-null SCA implies that we want the PLAB data recorded. 120 sca->record_sample(plab_start, plab_word_size); 121 } 122 } 123 124 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { 125 return new_obj->is_objArray() && 126 arrayOop(new_obj)->length() > ParGCArrayScanChunk && 127 new_obj != old_obj; 128 } 129 130 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { 131 assert(old->is_objArray(), "must be obj array"); 132 assert(old->is_forwarded(), "must be forwarded"); 133 assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap."); 134 assert(!old_gen()->is_in(old), "must be in young generation."); 135 136 objArrayOop obj = objArrayOop(old->forwardee()); 137 // Process ParGCArrayScanChunk elements now 138 // and push the remainder back onto queue 139 int start = arrayOop(old)->length(); 140 int end = obj->length(); 141 int remainder = end - start; 142 assert(start <= end, "just checking"); 143 if (remainder > 2 * ParGCArrayScanChunk) { 144 // Test above combines last partial chunk with a full chunk 145 end = start + ParGCArrayScanChunk; 146 arrayOop(old)->set_length(end); 147 // Push remainder. 148 bool ok = work_queue()->push(old); 149 assert(ok, "just popped, push must be okay"); 150 } else { 151 // Restore length so that it can be used if there 152 // is a promotion failure and forwarding pointers 153 // must be removed. 154 arrayOop(old)->set_length(end); 155 } 156 157 // process our set of indices (include header in first chunk) 158 // should make sure end is even (aligned to HeapWord in case of compressed oops) 159 if ((HeapWord *)obj < young_old_boundary()) { 160 // object is in to_space 161 obj->oop_iterate_range(&_to_space_closure, start, end); 162 } else { 163 // object is in old generation 164 obj->oop_iterate_range(&_old_gen_closure, start, end); 165 } 166 } 167 168 void ParScanThreadState::trim_queues(int max_size) { 169 ObjToScanQueue* queue = work_queue(); 170 do { 171 while (queue->size() > (juint)max_size) { 172 oop obj_to_scan; 173 if (queue->pop_local(obj_to_scan)) { 174 if ((HeapWord *)obj_to_scan < young_old_boundary()) { 175 if (obj_to_scan->is_objArray() && 176 obj_to_scan->is_forwarded() && 177 obj_to_scan->forwardee() != obj_to_scan) { 178 scan_partial_array_and_push_remainder(obj_to_scan); 179 } else { 180 // object is in to_space 181 obj_to_scan->oop_iterate(&_to_space_closure); 182 } 183 } else { 184 // object is in old generation 185 obj_to_scan->oop_iterate(&_old_gen_closure); 186 } 187 } 188 } 189 // For the case of compressed oops, we have a private, non-shared 190 // overflow stack, so we eagerly drain it so as to more evenly 191 // distribute load early. Note: this may be good to do in 192 // general rather than delay for the final stealing phase. 193 // If applicable, we'll transfer a set of objects over to our 194 // work queue, allowing them to be stolen and draining our 195 // private overflow stack. 196 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); 197 } 198 199 bool ParScanThreadState::take_from_overflow_stack() { 200 assert(ParGCUseLocalOverflow, "Else should not call"); 201 assert(young_gen()->overflow_list() == NULL, "Error"); 202 ObjToScanQueue* queue = work_queue(); 203 Stack<oop, mtGC>* const of_stack = overflow_stack(); 204 const size_t num_overflow_elems = of_stack->size(); 205 const size_t space_available = queue->max_elems() - queue->size(); 206 const size_t num_take_elems = MIN3(space_available / 4, 207 (size_t)ParGCDesiredObjsFromOverflowList, 208 num_overflow_elems); 209 // Transfer the most recent num_take_elems from the overflow 210 // stack to our work queue. 211 for (size_t i = 0; i != num_take_elems; i++) { 212 oop cur = of_stack->pop(); 213 oop obj_to_push = cur->forwardee(); 214 assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap"); 215 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); 216 assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap"); 217 if (should_be_partially_scanned(obj_to_push, cur)) { 218 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 219 obj_to_push = cur; 220 } 221 bool ok = queue->push(obj_to_push); 222 assert(ok, "Should have succeeded"); 223 } 224 assert(young_gen()->overflow_list() == NULL, "Error"); 225 return num_take_elems > 0; // was something transferred? 226 } 227 228 void ParScanThreadState::push_on_overflow_stack(oop p) { 229 assert(ParGCUseLocalOverflow, "Else should not call"); 230 overflow_stack()->push(p); 231 assert(young_gen()->overflow_list() == NULL, "Error"); 232 } 233 234 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { 235 // If the object is small enough, try to reallocate the buffer. 236 HeapWord* obj = NULL; 237 if (!_to_space_full) { 238 PLAB* const plab = to_space_alloc_buffer(); 239 Space* const sp = to_space(); 240 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) { 241 // Is small enough; abandon this buffer and start a new one. 242 plab->retire(); 243 // The minimum size has to be twice SurvivorAlignmentInBytes to 244 // allow for padding used in the alignment of 1 word. A padding 245 // of 1 is too small for a filler word so the padding size will 246 // be increased by SurvivorAlignmentInBytes. 247 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize); 248 size_t buf_size = MAX2(plab->word_sz(), min_usable_size); 249 HeapWord* buf_space = sp->par_allocate(buf_size); 250 if (buf_space == NULL) { 251 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize; 252 size_t free_bytes = sp->free(); 253 while(buf_space == NULL && free_bytes >= min_bytes) { 254 buf_size = free_bytes >> LogHeapWordSize; 255 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant"); 256 buf_space = sp->par_allocate(buf_size); 257 free_bytes = sp->free(); 258 } 259 } 260 if (buf_space != NULL) { 261 plab->set_buf(buf_space, buf_size); 262 record_survivor_plab(buf_space, buf_size); 263 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes); 264 // Note that we cannot compare buf_size < word_sz below 265 // because of AlignmentReserve (see PLAB::allocate()). 266 assert(obj != NULL || plab->words_remaining() < word_sz, 267 "Else should have been able to allocate requested object size " 268 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes " 269 SIZE_FORMAT ", words_remaining " SIZE_FORMAT, 270 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining()); 271 // It's conceivable that we may be able to use the 272 // buffer we just grabbed for subsequent small requests 273 // even if not for this one. 274 } else { 275 // We're used up. 276 _to_space_full = true; 277 } 278 } else { 279 // Too large; allocate the object individually. 280 obj = sp->par_allocate(word_sz); 281 } 282 } 283 return obj; 284 } 285 286 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) { 287 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 288 } 289 290 void ParScanThreadState::print_promotion_failure_size() { 291 if (_promotion_failed_info.has_failed()) { 292 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ", 293 _thread_num, _promotion_failed_info.first_size()); 294 } 295 } 296 297 class ParScanThreadStateSet: StackObj { 298 public: 299 // Initializes states for the specified number of threads; 300 ParScanThreadStateSet(int num_threads, 301 Space& to_space, 302 ParNewGeneration& young_gen, 303 Generation& old_gen, 304 ObjToScanQueueSet& queue_set, 305 Stack<oop, mtGC>* overflow_stacks_, 306 PreservedMarksSet& preserved_marks_set, 307 size_t desired_plab_sz, 308 TaskTerminator& term); 309 310 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 311 312 inline ParScanThreadState& thread_state(int i); 313 314 void trace_promotion_failed(const YoungGCTracer* gc_tracer); 315 void reset(uint active_workers, bool promotion_failed); 316 void flush(); 317 318 #if TASKQUEUE_STATS 319 static void 320 print_termination_stats_hdr(outputStream* const st); 321 void print_termination_stats(); 322 static void 323 print_taskqueue_stats_hdr(outputStream* const st); 324 void print_taskqueue_stats(); 325 void reset_stats(); 326 #endif // TASKQUEUE_STATS 327 328 private: 329 TaskTerminator& _term; 330 ParNewGeneration& _young_gen; 331 Generation& _old_gen; 332 ParScanThreadState* _per_thread_states; 333 const int _num_threads; 334 public: 335 bool is_valid(int id) const { return id < _num_threads; } 336 ParallelTaskTerminator* terminator() { return _term.terminator(); } 337 }; 338 339 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads, 340 Space& to_space, 341 ParNewGeneration& young_gen, 342 Generation& old_gen, 343 ObjToScanQueueSet& queue_set, 344 Stack<oop, mtGC>* overflow_stacks, 345 PreservedMarksSet& preserved_marks_set, 346 size_t desired_plab_sz, 347 TaskTerminator& term) 348 : _term(term), 349 _young_gen(young_gen), 350 _old_gen(old_gen), 351 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)), 352 _num_threads(num_threads) 353 { 354 assert(num_threads > 0, "sanity check!"); 355 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 356 "overflow_stack allocation mismatch"); 357 // Initialize states. 358 for (int i = 0; i < num_threads; ++i) { 359 new(_per_thread_states + i) 360 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set, 361 overflow_stacks, preserved_marks_set.get(i), 362 desired_plab_sz, term); 363 } 364 } 365 366 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) { 367 assert(i >= 0 && i < _num_threads, "sanity check!"); 368 return _per_thread_states[i]; 369 } 370 371 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) { 372 for (int i = 0; i < _num_threads; ++i) { 373 if (thread_state(i).promotion_failed()) { 374 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info()); 375 thread_state(i).promotion_failed_info().reset(); 376 } 377 } 378 } 379 380 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) { 381 _term.terminator()->reset_for_reuse(active_threads); 382 if (promotion_failed) { 383 for (int i = 0; i < _num_threads; ++i) { 384 thread_state(i).print_promotion_failure_size(); 385 } 386 } 387 } 388 389 #if TASKQUEUE_STATS 390 void ParScanThreadState::reset_stats() { 391 taskqueue_stats().reset(); 392 _term_attempts = 0; 393 _overflow_refills = 0; 394 _overflow_refill_objs = 0; 395 } 396 397 void ParScanThreadStateSet::reset_stats() { 398 for (int i = 0; i < _num_threads; ++i) { 399 thread_state(i).reset_stats(); 400 } 401 } 402 403 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) { 404 st->print_raw_cr("GC Termination Stats"); 405 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"); 406 st->print_raw_cr("thr ms ms % ms % attempts"); 407 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"); 408 } 409 410 void ParScanThreadStateSet::print_termination_stats() { 411 Log(gc, task, stats) log; 412 if (!log.is_debug()) { 413 return; 414 } 415 416 ResourceMark rm; 417 LogStream ls(log.debug()); 418 outputStream* st = &ls; 419 420 print_termination_stats_hdr(st); 421 422 for (int i = 0; i < _num_threads; ++i) { 423 const ParScanThreadState & pss = thread_state(i); 424 const double elapsed_ms = pss.elapsed_time() * 1000.0; 425 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 426 const double term_ms = pss.term_time() * 1000.0; 427 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8), 428 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 429 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); 430 } 431 } 432 433 // Print stats related to work queue activity. 434 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) { 435 st->print_raw_cr("GC Task Stats"); 436 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 437 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 438 } 439 440 void ParScanThreadStateSet::print_taskqueue_stats() { 441 if (!log_is_enabled(Trace, gc, task, stats)) { 442 return; 443 } 444 Log(gc, task, stats) log; 445 ResourceMark rm; 446 LogStream ls(log.trace()); 447 outputStream* st = &ls; 448 print_taskqueue_stats_hdr(st); 449 450 TaskQueueStats totals; 451 for (int i = 0; i < _num_threads; ++i) { 452 const ParScanThreadState & pss = thread_state(i); 453 const TaskQueueStats & stats = pss.taskqueue_stats(); 454 st->print("%3d ", i); stats.print(st); st->cr(); 455 totals += stats; 456 457 if (pss.overflow_refills() > 0) { 458 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 459 SIZE_FORMAT_W(10) " overflow objects", 460 pss.overflow_refills(), pss.overflow_refill_objs()); 461 } 462 } 463 st->print("tot "); totals.print(st); st->cr(); 464 465 DEBUG_ONLY(totals.verify()); 466 } 467 #endif // TASKQUEUE_STATS 468 469 void ParScanThreadStateSet::flush() { 470 // Work in this loop should be kept as lightweight as 471 // possible since this might otherwise become a bottleneck 472 // to scaling. Should we add heavy-weight work into this 473 // loop, consider parallelizing the loop into the worker threads. 474 for (int i = 0; i < _num_threads; ++i) { 475 ParScanThreadState& par_scan_state = thread_state(i); 476 477 // Flush stats related to To-space PLAB activity and 478 // retire the last buffer. 479 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats()); 480 481 // Every thread has its own age table. We need to merge 482 // them all into one. 483 AgeTable *local_table = par_scan_state.age_table(); 484 _young_gen.age_table()->merge(local_table); 485 486 // Inform old gen that we're done. 487 _old_gen.par_promote_alloc_done(i); 488 } 489 490 if (UseConcMarkSweepGC) { 491 // We need to call this even when ResizeOldPLAB is disabled 492 // so as to avoid breaking some asserts. While we may be able 493 // to avoid this by reorganizing the code a bit, I am loathe 494 // to do that unless we find cases where ergo leads to bad 495 // performance. 496 CompactibleFreeListSpaceLAB::compute_desired_plab_size(); 497 } 498 } 499 500 ParScanClosure::ParScanClosure(ParNewGeneration* g, 501 ParScanThreadState* par_scan_state) : 502 OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) { 503 _boundary = _g->reserved().end(); 504 } 505 506 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 507 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 508 509 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 510 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 511 512 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 513 ParScanThreadState* par_scan_state) 514 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 515 {} 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 _to_space_root_closure(to_space_root_closure_), 534 _old_gen_closure(old_gen_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 obj_to_scan)) { 554 bool res = work_q->push(obj_to_scan); 555 assert(res, "Empty queue should have room for a push."); 556 557 // If successful, goto Start. 558 continue; 559 560 // Try global overflow list. 561 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 562 continue; 563 } 564 565 // Otherwise, offer termination. 566 par_scan_state()->start_term_time(); 567 if (terminator()->offer_termination()) break; 568 par_scan_state()->end_term_time(); 569 } 570 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 571 "Broken overflow list?"); 572 // Finish the last termination pause. 573 par_scan_state()->end_term_time(); 574 } 575 576 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen, 577 Generation* old_gen, 578 HeapWord* young_old_boundary, 579 ParScanThreadStateSet* state_set, 580 StrongRootsScope* strong_roots_scope) : 581 AbstractGangTask("ParNewGeneration collection"), 582 _young_gen(young_gen), _old_gen(old_gen), 583 _young_old_boundary(young_old_boundary), 584 _state_set(state_set), 585 _strong_roots_scope(strong_roots_scope) 586 {} 587 588 void ParNewGenTask::work(uint worker_id) { 589 CMSHeap* heap = CMSHeap::heap(); 590 // Since this is being done in a separate thread, need new resource 591 // and handle marks. 592 ResourceMark rm; 593 HandleMark hm; 594 595 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 596 assert(_state_set->is_valid(worker_id), "Should not have been called"); 597 598 par_scan_state.set_young_old_boundary(_young_old_boundary); 599 600 CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(), 601 heap->rem_set()->cld_rem_set()->accumulate_modified_oops()); 602 603 par_scan_state.start_strong_roots(); 604 heap->young_process_roots(_strong_roots_scope, 605 &par_scan_state.to_space_root_closure(), 606 &par_scan_state.older_gen_closure(), 607 &cld_scan_closure); 608 609 par_scan_state.end_strong_roots(); 610 611 // "evacuate followers". 612 par_scan_state.evacuate_followers_closure().do_void(); 613 614 // This will collapse this worker's promoted object list that's 615 // created during the main ParNew parallel phase of ParNew. This has 616 // to be called after all workers have finished promoting objects 617 // and scanning promoted objects. It should be safe calling it from 618 // here, given that we can only reach here after all thread have 619 // offered termination, i.e., after there is no more work to be 620 // done. It will also disable promotion tracking for the rest of 621 // this GC as it's not necessary to be on during reference processing. 622 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id); 623 } 624 625 ParNewGeneration::ParNewGeneration(ReservedSpace rs, 626 size_t initial_byte_size, 627 size_t min_byte_size, 628 size_t max_byte_size) 629 : DefNewGeneration(rs, initial_byte_size, min_byte_size, max_byte_size, "CMS young collection pauses"), 630 _plab_stats("Young", YoungPLABSize, PLABWeight), 631 _overflow_list(NULL), 632 _is_alive_closure(this) 633 { 634 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 635 NOT_PRODUCT(_num_par_pushes = 0;) 636 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 637 guarantee(_task_queues != NULL, "task_queues allocation failure."); 638 639 for (uint i = 0; i < ParallelGCThreads; i++) { 640 ObjToScanQueue *q = new ObjToScanQueue(); 641 guarantee(q != NULL, "work_queue Allocation failure."); 642 _task_queues->register_queue(i, q); 643 } 644 645 for (uint i = 0; i < ParallelGCThreads; i++) { 646 _task_queues->queue(i)->initialize(); 647 } 648 649 _overflow_stacks = NULL; 650 if (ParGCUseLocalOverflow) { 651 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ',' 652 typedef Stack<oop, mtGC> GCOopStack; 653 654 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 655 for (size_t i = 0; i < ParallelGCThreads; ++i) { 656 new (_overflow_stacks + i) Stack<oop, mtGC>(); 657 } 658 } 659 660 if (UsePerfData) { 661 EXCEPTION_MARK; 662 ResourceMark rm; 663 664 const char* cname = 665 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 666 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 667 ParallelGCThreads, CHECK); 668 } 669 } 670 671 // ParNewGeneration:: 672 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 673 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 674 675 template <class T> 676 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 677 #ifdef ASSERT 678 { 679 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p); 680 // We never expect to see a null reference being processed 681 // as a weak reference. 682 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs"); 683 } 684 #endif // ASSERT 685 686 Devirtualizer::do_oop_no_verify(_par_cl, p); 687 688 if (CMSHeap::heap()->is_in_reserved(p)) { 689 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);; 690 _rs->write_ref_field_gc_par(p, obj); 691 } 692 } 693 694 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 695 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 696 697 // ParNewGeneration:: 698 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 699 DefNewGeneration::KeepAliveClosure(cl) {} 700 701 template <class T> 702 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 703 #ifdef ASSERT 704 { 705 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p); 706 // We never expect to see a null reference being processed 707 // as a weak reference. 708 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs"); 709 } 710 #endif // ASSERT 711 712 Devirtualizer::do_oop_no_verify(_cl, p); 713 714 if (CMSHeap::heap()->is_in_reserved(p)) { 715 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p); 716 _rs->write_ref_field_gc_par(p, obj); 717 } 718 } 719 720 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 721 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 722 723 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 724 T heap_oop = RawAccess<>::oop_load(p); 725 if (!CompressedOops::is_null(heap_oop)) { 726 oop obj = CompressedOops::decode_not_null(heap_oop); 727 if ((HeapWord*)obj < _boundary) { 728 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 729 oop new_obj = obj->is_forwarded() 730 ? obj->forwardee() 731 : _g->DefNewGeneration::copy_to_survivor_space(obj); 732 RawAccess<IS_NOT_NULL>::oop_store(p, new_obj); 733 } 734 if (_gc_barrier) { 735 // If p points to a younger generation, mark the card. 736 if ((HeapWord*)obj < _gen_boundary) { 737 _rs->write_ref_field_gc_par(p, obj); 738 } 739 } 740 } 741 } 742 743 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 744 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 745 746 class ParNewRefProcTaskProxy: public AbstractGangTask { 747 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 748 public: 749 ParNewRefProcTaskProxy(ProcessTask& task, 750 ParNewGeneration& young_gen, 751 Generation& old_gen, 752 HeapWord* young_old_boundary, 753 ParScanThreadStateSet& state_set); 754 755 private: 756 virtual void work(uint worker_id); 757 private: 758 ParNewGeneration& _young_gen; 759 ProcessTask& _task; 760 Generation& _old_gen; 761 HeapWord* _young_old_boundary; 762 ParScanThreadStateSet& _state_set; 763 }; 764 765 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task, 766 ParNewGeneration& young_gen, 767 Generation& old_gen, 768 HeapWord* young_old_boundary, 769 ParScanThreadStateSet& state_set) 770 : AbstractGangTask("ParNewGeneration parallel reference processing"), 771 _young_gen(young_gen), 772 _task(task), 773 _old_gen(old_gen), 774 _young_old_boundary(young_old_boundary), 775 _state_set(state_set) 776 { } 777 778 void ParNewRefProcTaskProxy::work(uint worker_id) { 779 ResourceMark rm; 780 HandleMark hm; 781 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 782 par_scan_state.set_young_old_boundary(_young_old_boundary); 783 _task.work(worker_id, par_scan_state.is_alive_closure(), 784 par_scan_state.keep_alive_closure(), 785 par_scan_state.evacuate_followers_closure()); 786 } 787 788 void ParNewRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 789 CMSHeap* gch = CMSHeap::heap(); 790 WorkGang* workers = gch->workers(); 791 assert(workers != NULL, "Need parallel worker threads."); 792 assert(workers->active_workers() == ergo_workers, 793 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 794 ergo_workers, workers->active_workers()); 795 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed()); 796 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen, 797 _young_gen.reserved().end(), _state_set); 798 workers->run_task(&rp_task, workers->active_workers()); 799 _state_set.reset(0 /* bad value in debug if not reset */, 800 _young_gen.promotion_failed()); 801 } 802 803 void ParNewRefProcTaskExecutor::set_single_threaded_mode() { 804 _state_set.flush(); 805 CMSHeap* heap = CMSHeap::heap(); 806 heap->save_marks(); 807 } 808 809 ScanClosureWithParBarrier:: 810 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 811 OopsInClassLoaderDataOrGenClosure(g), _g(g), _boundary(g->reserved().end()), _gc_barrier(gc_barrier) 812 { } 813 814 template <typename OopClosureType1, typename OopClosureType2> 815 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>:: 816 EvacuateFollowersClosureGeneral(CMSHeap* heap, 817 OopClosureType1* cur, 818 OopClosureType2* older) : 819 _heap(heap), 820 _scan_cur_or_nonheap(cur), _scan_older(older) 821 { } 822 823 template <typename OopClosureType1, typename OopClosureType2> 824 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() { 825 do { 826 _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap, 827 _scan_older); 828 } while (!_heap->no_allocs_since_save_marks()); 829 } 830 831 // A Generation that does parallel young-gen collection. 832 833 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) { 834 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 835 _promo_failure_scan_stack.clear(true); // Clear cached segments. 836 837 remove_forwarding_pointers(); 838 log_info(gc, promotion)("Promotion failed"); 839 // All the spaces are in play for mark-sweep. 840 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 841 from()->set_next_compaction_space(to()); 842 gch->set_incremental_collection_failed(); 843 // Inform the next generation that a promotion failure occurred. 844 _old_gen->promotion_failure_occurred(); 845 846 // Trace promotion failure in the parallel GC threads 847 thread_state_set.trace_promotion_failed(gc_tracer()); 848 // Single threaded code may have reported promotion failure to the global state 849 if (_promotion_failed_info.has_failed()) { 850 _gc_tracer.report_promotion_failed(_promotion_failed_info); 851 } 852 // Reset the PromotionFailureALot counters. 853 NOT_PRODUCT(gch->reset_promotion_should_fail();) 854 } 855 856 void ParNewGeneration::collect(bool full, 857 bool clear_all_soft_refs, 858 size_t size, 859 bool is_tlab) { 860 assert(full || size > 0, "otherwise we don't want to collect"); 861 862 CMSHeap* gch = CMSHeap::heap(); 863 864 _gc_timer->register_gc_start(); 865 866 AdaptiveSizePolicy* size_policy = gch->size_policy(); 867 WorkGang* workers = gch->workers(); 868 assert(workers != NULL, "Need workgang for parallel work"); 869 uint active_workers = 870 WorkerPolicy::calc_active_workers(workers->total_workers(), 871 workers->active_workers(), 872 Threads::number_of_non_daemon_threads()); 873 active_workers = workers->update_active_workers(active_workers); 874 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers()); 875 876 _old_gen = gch->old_gen(); 877 878 // If the next generation is too full to accommodate worst-case promotion 879 // from this generation, pass on collection; let the next generation 880 // do it. 881 if (!collection_attempt_is_safe()) { 882 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 883 return; 884 } 885 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 886 887 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 888 gch->trace_heap_before_gc(gc_tracer()); 889 890 init_assuming_no_promotion_failure(); 891 892 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause()); 893 894 age_table()->clear(); 895 to()->clear(SpaceDecorator::Mangle); 896 897 gch->save_marks(); 898 899 // Set the correct parallelism (number of queues) in the reference processor 900 ref_processor()->set_active_mt_degree(active_workers); 901 902 // Need to initialize the preserved marks before the ThreadStateSet c'tor. 903 _preserved_marks_set.init(active_workers); 904 905 // Always set the terminator for the active number of workers 906 // because only those workers go through the termination protocol. 907 TaskTerminator _term(active_workers, task_queues()); 908 ParScanThreadStateSet thread_state_set(active_workers, 909 *to(), *this, *_old_gen, *task_queues(), 910 _overflow_stacks, _preserved_marks_set, 911 desired_plab_sz(), _term); 912 913 thread_state_set.reset(active_workers, promotion_failed()); 914 915 { 916 StrongRootsScope srs(active_workers); 917 918 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs); 919 gch->rem_set()->prepare_for_younger_refs_iterate(true); 920 // It turns out that even when we're using 1 thread, doing the work in a 921 // separate thread causes wide variance in run times. We can't help this 922 // in the multi-threaded case, but we special-case n=1 here to get 923 // repeatable measurements of the 1-thread overhead of the parallel code. 924 // Might multiple workers ever be used? If yes, initialization 925 // has been done such that the single threaded path should not be used. 926 if (workers->total_workers() > 1) { 927 workers->run_task(&tsk); 928 } else { 929 tsk.work(0); 930 } 931 } 932 933 thread_state_set.reset(0 /* Bad value in debug if not reset */, 934 promotion_failed()); 935 936 // Trace and reset failed promotion info. 937 if (promotion_failed()) { 938 thread_state_set.trace_promotion_failed(gc_tracer()); 939 } 940 941 // Process (weak) reference objects found during scavenge. 942 ReferenceProcessor* rp = ref_processor(); 943 IsAliveClosure is_alive(this); 944 ScanWeakRefClosure scan_weak_ref(this); 945 KeepAliveClosure keep_alive(&scan_weak_ref); 946 ScanClosure scan_without_gc_barrier(this, false); 947 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 948 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 949 EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers( 950 gch, &scan_without_gc_barrier, &scan_with_gc_barrier); 951 rp->setup_policy(clear_all_soft_refs); 952 // Can the mt_degree be set later (at run_task() time would be best)? 953 rp->set_active_mt_degree(active_workers); 954 ReferenceProcessorStats stats; 955 ReferenceProcessorPhaseTimes pt(_gc_timer, rp->max_num_queues()); 956 if (rp->processing_is_mt()) { 957 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 958 stats = rp->process_discovered_references(&is_alive, &keep_alive, 959 &evacuate_followers, &task_executor, 960 &pt); 961 } else { 962 thread_state_set.flush(); 963 gch->save_marks(); 964 stats = rp->process_discovered_references(&is_alive, &keep_alive, 965 &evacuate_followers, NULL, 966 &pt); 967 } 968 _gc_tracer.report_gc_reference_stats(stats); 969 _gc_tracer.report_tenuring_threshold(tenuring_threshold()); 970 pt.print_all_references(); 971 972 assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point"); 973 974 WeakProcessor::weak_oops_do(&is_alive, &keep_alive); 975 976 // Verify that the usage of keep_alive only forwarded 977 // the oops and did not find anything new to copy. 978 assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects"); 979 980 if (!promotion_failed()) { 981 // Swap the survivor spaces. 982 eden()->clear(SpaceDecorator::Mangle); 983 from()->clear(SpaceDecorator::Mangle); 984 if (ZapUnusedHeapArea) { 985 // This is now done here because of the piece-meal mangling which 986 // can check for valid mangling at intermediate points in the 987 // collection(s). When a young collection fails to collect 988 // sufficient space resizing of the young generation can occur 989 // and redistribute the spaces in the young generation. Mangle 990 // here so that unzapped regions don't get distributed to 991 // other spaces. 992 to()->mangle_unused_area(); 993 } 994 swap_spaces(); 995 996 // A successful scavenge should restart the GC time limit count which is 997 // for full GC's. 998 size_policy->reset_gc_overhead_limit_count(); 999 1000 assert(to()->is_empty(), "to space should be empty now"); 1001 1002 adjust_desired_tenuring_threshold(); 1003 } else { 1004 handle_promotion_failed(gch, thread_state_set); 1005 } 1006 _preserved_marks_set.reclaim(); 1007 // set new iteration safe limit for the survivor spaces 1008 from()->set_concurrent_iteration_safe_limit(from()->top()); 1009 to()->set_concurrent_iteration_safe_limit(to()->top()); 1010 1011 plab_stats()->adjust_desired_plab_sz(); 1012 1013 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); 1014 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); 1015 1016 // We need to use a monotonically non-decreasing time in ms 1017 // or we will see time-warp warnings and os::javaTimeMillis() 1018 // does not guarantee monotonicity. 1019 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1020 update_time_of_last_gc(now); 1021 1022 rp->set_enqueuing_is_done(true); 1023 rp->verify_no_references_recorded(); 1024 1025 gch->trace_heap_after_gc(gc_tracer()); 1026 1027 _gc_timer->register_gc_end(); 1028 1029 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1030 } 1031 1032 size_t ParNewGeneration::desired_plab_sz() { 1033 return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers()); 1034 } 1035 1036 static int sum; 1037 void ParNewGeneration::waste_some_time() { 1038 for (int i = 0; i < 100; i++) { 1039 sum += i; 1040 } 1041 } 1042 1043 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); 1044 1045 // Because of concurrency, there are times where an object for which 1046 // "is_forwarded()" is true contains an "interim" forwarding pointer 1047 // value. Such a value will soon be overwritten with a real value. 1048 // This method requires "obj" to have a forwarding pointer, and waits, if 1049 // necessary for a real one to be inserted, and returns it. 1050 1051 oop ParNewGeneration::real_forwardee(oop obj) { 1052 oop forward_ptr = obj->forwardee(); 1053 if (forward_ptr != ClaimedForwardPtr) { 1054 return forward_ptr; 1055 } else { 1056 return real_forwardee_slow(obj); 1057 } 1058 } 1059 1060 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1061 // Spin-read if it is claimed but not yet written by another thread. 1062 oop forward_ptr = obj->forwardee(); 1063 while (forward_ptr == ClaimedForwardPtr) { 1064 waste_some_time(); 1065 assert(obj->is_forwarded(), "precondition"); 1066 forward_ptr = obj->forwardee(); 1067 } 1068 return forward_ptr; 1069 } 1070 1071 // Multiple GC threads may try to promote an object. If the object 1072 // is successfully promoted, a forwarding pointer will be installed in 1073 // the object in the young generation. This method claims the right 1074 // to install the forwarding pointer before it copies the object, 1075 // thus avoiding the need to undo the copy as in 1076 // copy_to_survivor_space_avoiding_with_undo. 1077 1078 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state, 1079 oop old, 1080 size_t sz, 1081 markWord m) { 1082 // In the sequential version, this assert also says that the object is 1083 // not forwarded. That might not be the case here. It is the case that 1084 // the caller observed it to be not forwarded at some time in the past. 1085 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1086 1087 // The sequential code read "old->age()" below. That doesn't work here, 1088 // since the age is in the mark word, and that might be overwritten with 1089 // a forwarding pointer by a parallel thread. So we must save the mark 1090 // word in a local and then analyze it. 1091 oopDesc dummyOld; 1092 dummyOld.set_mark_raw(m); 1093 assert(!dummyOld.is_forwarded(), 1094 "should not be called with forwarding pointer mark word."); 1095 1096 oop new_obj = NULL; 1097 oop forward_ptr; 1098 1099 // Try allocating obj in to-space (unless too old) 1100 if (dummyOld.age() < tenuring_threshold()) { 1101 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1102 } 1103 1104 if (new_obj == NULL) { 1105 // Either to-space is full or we decided to promote try allocating obj tenured 1106 1107 // Attempt to install a null forwarding pointer (atomically), 1108 // to claim the right to install the real forwarding pointer. 1109 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr, m); 1110 if (forward_ptr != NULL) { 1111 // someone else beat us to it. 1112 return real_forwardee(old); 1113 } 1114 1115 if (!_promotion_failed) { 1116 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), 1117 old, m, sz); 1118 } 1119 1120 if (new_obj == NULL) { 1121 // promotion failed, forward to self 1122 _promotion_failed = true; 1123 new_obj = old; 1124 1125 par_scan_state->preserved_marks()->push_if_necessary(old, m); 1126 par_scan_state->register_promotion_failure(sz); 1127 } 1128 1129 old->forward_to(new_obj); 1130 forward_ptr = NULL; 1131 } else { 1132 // Is in to-space; do copying ourselves. 1133 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1134 assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value."); 1135 forward_ptr = old->forward_to_atomic(new_obj, m); 1136 // Restore the mark word copied above. 1137 new_obj->set_mark_raw(m); 1138 // Increment age if obj still in new generation 1139 new_obj->incr_age(); 1140 par_scan_state->age_table()->add(new_obj, sz); 1141 } 1142 assert(new_obj != NULL, "just checking"); 1143 1144 // This code must come after the CAS test, or it will print incorrect 1145 // information. 1146 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1147 is_in_reserved(new_obj) ? "copying" : "tenuring", 1148 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size()); 1149 1150 if (forward_ptr == NULL) { 1151 oop obj_to_push = new_obj; 1152 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1153 // Length field used as index of next element to be scanned. 1154 // Real length can be obtained from real_forwardee() 1155 arrayOop(old)->set_length(0); 1156 obj_to_push = old; 1157 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1158 "push forwarded object"); 1159 } 1160 // Push it on one of the queues of to-be-scanned objects. 1161 bool simulate_overflow = false; 1162 NOT_PRODUCT( 1163 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1164 // simulate a stack overflow 1165 simulate_overflow = true; 1166 } 1167 ) 1168 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1169 // Add stats for overflow pushes. 1170 log_develop_trace(gc)("Queue Overflow"); 1171 push_on_overflow_list(old, par_scan_state); 1172 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1173 } 1174 1175 return new_obj; 1176 } 1177 1178 // Oops. Someone beat us to it. Undo the allocation. Where did we 1179 // allocate it? 1180 if (is_in_reserved(new_obj)) { 1181 // Must be in to_space. 1182 assert(to()->is_in_reserved(new_obj), "Checking"); 1183 if (forward_ptr == ClaimedForwardPtr) { 1184 // Wait to get the real forwarding pointer value. 1185 forward_ptr = real_forwardee(old); 1186 } 1187 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1188 } 1189 1190 return forward_ptr; 1191 } 1192 1193 #ifndef PRODUCT 1194 // It's OK to call this multi-threaded; the worst thing 1195 // that can happen is that we'll get a bunch of closely 1196 // spaced simulated overflows, but that's OK, in fact 1197 // probably good as it would exercise the overflow code 1198 // under contention. 1199 bool ParNewGeneration::should_simulate_overflow() { 1200 if (_overflow_counter-- <= 0) { // just being defensive 1201 _overflow_counter = ParGCWorkQueueOverflowInterval; 1202 return true; 1203 } else { 1204 return false; 1205 } 1206 } 1207 #endif 1208 1209 // In case we are using compressed oops, we need to be careful. 1210 // If the object being pushed is an object array, then its length 1211 // field keeps track of the "grey boundary" at which the next 1212 // incremental scan will be done (see ParGCArrayScanChunk). 1213 // When using compressed oops, this length field is kept in the 1214 // lower 32 bits of the erstwhile klass word and cannot be used 1215 // for the overflow chaining pointer (OCP below). As such the OCP 1216 // would itself need to be compressed into the top 32-bits in this 1217 // case. Unfortunately, see below, in the event that we have a 1218 // promotion failure, the node to be pushed on the list can be 1219 // outside of the Java heap, so the heap-based pointer compression 1220 // would not work (we would have potential aliasing between C-heap 1221 // and Java-heap pointers). For this reason, when using compressed 1222 // oops, we simply use a worker-thread-local, non-shared overflow 1223 // list in the form of a growable array, with a slightly different 1224 // overflow stack draining strategy. If/when we start using fat 1225 // stacks here, we can go back to using (fat) pointer chains 1226 // (although some performance comparisons would be useful since 1227 // single global lists have their own performance disadvantages 1228 // as we were made painfully aware not long ago, see 6786503). 1229 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 1230 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1231 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1232 if (ParGCUseLocalOverflow) { 1233 // In the case of compressed oops, we use a private, not-shared 1234 // overflow stack. 1235 par_scan_state->push_on_overflow_stack(from_space_obj); 1236 } else { 1237 assert(!UseCompressedOops, "Error"); 1238 // if the object has been forwarded to itself, then we cannot 1239 // use the klass pointer for the linked list. Instead we have 1240 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1241 // XXX This is horribly inefficient when a promotion failure occurs 1242 // and should be fixed. XXX FIX ME !!! 1243 #ifndef PRODUCT 1244 Atomic::inc(&_num_par_pushes); 1245 assert(_num_par_pushes > 0, "Tautology"); 1246 #endif 1247 if (from_space_obj->forwardee() == from_space_obj) { 1248 oopDesc* listhead = NEW_C_HEAP_OBJ(oopDesc, mtGC); 1249 listhead->forward_to(from_space_obj); 1250 from_space_obj = listhead; 1251 } 1252 oop observed_overflow_list = _overflow_list; 1253 oop cur_overflow_list; 1254 do { 1255 cur_overflow_list = observed_overflow_list; 1256 if (cur_overflow_list != BUSY) { 1257 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1258 } else { 1259 from_space_obj->set_klass_to_list_ptr(NULL); 1260 } 1261 observed_overflow_list = 1262 Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list); 1263 } while (cur_overflow_list != observed_overflow_list); 1264 } 1265 } 1266 1267 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1268 bool res; 1269 1270 if (ParGCUseLocalOverflow) { 1271 res = par_scan_state->take_from_overflow_stack(); 1272 } else { 1273 assert(!UseCompressedOops, "Error"); 1274 res = take_from_overflow_list_work(par_scan_state); 1275 } 1276 return res; 1277 } 1278 1279 1280 // *NOTE*: The overflow list manipulation code here and 1281 // in CMSCollector:: are very similar in shape, 1282 // except that in the CMS case we thread the objects 1283 // directly into the list via their mark word, and do 1284 // not need to deal with special cases below related 1285 // to chunking of object arrays and promotion failure 1286 // handling. 1287 // CR 6797058 has been filed to attempt consolidation of 1288 // the common code. 1289 // Because of the common code, if you make any changes in 1290 // the code below, please check the CMS version to see if 1291 // similar changes might be needed. 1292 // See CMSCollector::par_take_from_overflow_list() for 1293 // more extensive documentation comments. 1294 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1295 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1296 // How many to take? 1297 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1298 (size_t)ParGCDesiredObjsFromOverflowList); 1299 1300 assert(!UseCompressedOops, "Error"); 1301 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1302 if (_overflow_list == NULL) return false; 1303 1304 // Otherwise, there was something there; try claiming the list. 1305 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 1306 // Trim off a prefix of at most objsFromOverflow items 1307 Thread* tid = Thread::current(); 1308 size_t spin_count = ParallelGCThreads; 1309 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1310 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1311 // someone grabbed it before we did ... 1312 // ... we spin for a short while... 1313 os::sleep(tid, sleep_time_millis, false); 1314 if (_overflow_list == NULL) { 1315 // nothing left to take 1316 return false; 1317 } else if (_overflow_list != BUSY) { 1318 // try and grab the prefix 1319 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 1320 } 1321 } 1322 if (prefix == NULL || prefix == BUSY) { 1323 // Nothing to take or waited long enough 1324 if (prefix == NULL) { 1325 // Write back the NULL in case we overwrote it with BUSY above 1326 // and it is still the same value. 1327 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 1328 } 1329 return false; 1330 } 1331 assert(prefix != NULL && prefix != BUSY, "Error"); 1332 oop cur = prefix; 1333 for (size_t i = 1; i < objsFromOverflow; ++i) { 1334 oop next = cur->list_ptr_from_klass(); 1335 if (next == NULL) break; 1336 cur = next; 1337 } 1338 assert(cur != NULL, "Loop postcondition"); 1339 1340 // Reattach remaining (suffix) to overflow list 1341 oop suffix = cur->list_ptr_from_klass(); 1342 if (suffix == NULL) { 1343 // Write back the NULL in lieu of the BUSY we wrote 1344 // above and it is still the same value. 1345 if (_overflow_list == BUSY) { 1346 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 1347 } 1348 } else { 1349 assert(suffix != BUSY, "Error"); 1350 // suffix will be put back on global list 1351 cur->set_klass_to_list_ptr(NULL); // break off suffix 1352 // It's possible that the list is still in the empty(busy) state 1353 // we left it in a short while ago; in that case we may be 1354 // able to place back the suffix. 1355 oop observed_overflow_list = _overflow_list; 1356 oop cur_overflow_list = observed_overflow_list; 1357 bool attached = false; 1358 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1359 observed_overflow_list = 1360 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list); 1361 if (cur_overflow_list == observed_overflow_list) { 1362 attached = true; 1363 break; 1364 } else cur_overflow_list = observed_overflow_list; 1365 } 1366 if (!attached) { 1367 // Too bad, someone else got in in between; we'll need to do a splice. 1368 // Find the last item of suffix list 1369 oop last = suffix; 1370 while (true) { 1371 oop next = last->list_ptr_from_klass(); 1372 if (next == NULL) break; 1373 last = next; 1374 } 1375 // Atomically prepend suffix to current overflow list 1376 observed_overflow_list = _overflow_list; 1377 do { 1378 cur_overflow_list = observed_overflow_list; 1379 if (cur_overflow_list != BUSY) { 1380 // Do the splice ... 1381 last->set_klass_to_list_ptr(cur_overflow_list); 1382 } else { // cur_overflow_list == BUSY 1383 last->set_klass_to_list_ptr(NULL); 1384 } 1385 observed_overflow_list = 1386 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list); 1387 } while (cur_overflow_list != observed_overflow_list); 1388 } 1389 } 1390 1391 // Push objects on prefix list onto this thread's work queue 1392 assert(prefix != NULL && prefix != BUSY, "program logic"); 1393 cur = prefix; 1394 ssize_t n = 0; 1395 while (cur != NULL) { 1396 oop obj_to_push = cur->forwardee(); 1397 oop next = cur->list_ptr_from_klass(); 1398 cur->set_klass(obj_to_push->klass()); 1399 // This may be an array object that is self-forwarded. In that case, the list pointer 1400 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1401 if (!is_in_reserved(cur)) { 1402 // This can become a scaling bottleneck when there is work queue overflow coincident 1403 // with promotion failure. 1404 oopDesc* f = cur; 1405 FREE_C_HEAP_OBJ(f); 1406 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1407 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1408 obj_to_push = cur; 1409 } 1410 bool ok = work_q->push(obj_to_push); 1411 assert(ok, "Should have succeeded"); 1412 cur = next; 1413 n++; 1414 } 1415 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1416 #ifndef PRODUCT 1417 assert(_num_par_pushes >= n, "Too many pops?"); 1418 Atomic::sub(n, &_num_par_pushes); 1419 #endif 1420 return true; 1421 } 1422 #undef BUSY 1423 1424 void ParNewGeneration::ref_processor_init() { 1425 if (_ref_processor == NULL) { 1426 // Allocate and initialize a reference processor 1427 _span_based_discoverer.set_span(_reserved); 1428 _ref_processor = 1429 new ReferenceProcessor(&_span_based_discoverer, // span 1430 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1431 ParallelGCThreads, // mt processing degree 1432 refs_discovery_is_mt(), // mt discovery 1433 ParallelGCThreads, // mt discovery degree 1434 refs_discovery_is_atomic(), // atomic_discovery 1435 NULL, // is_alive_non_header 1436 false); // disable adjusting number of processing threads 1437 } 1438 } 1439 1440 const char* ParNewGeneration::name() const { 1441 return "par new generation"; 1442 } 1443 1444 void ParNewGeneration::restore_preserved_marks() { 1445 SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers()); 1446 _preserved_marks_set.restore(&task_executor); 1447 }