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