1 /* 2 * Copyright (c) 2001, 2013, 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_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp" 27 #include "gc_implementation/parNew/parNewGeneration.hpp" 28 #include "gc_implementation/parNew/parOopClosures.inline.hpp" 29 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 30 #include "gc_implementation/shared/ageTable.hpp" 31 #include "gc_implementation/shared/parGCAllocBuffer.hpp" 32 #include "gc_implementation/shared/gcHeapSummary.hpp" 33 #include "gc_implementation/shared/gcTimer.hpp" 34 #include "gc_implementation/shared/gcTrace.hpp" 35 #include "gc_implementation/shared/gcTraceTime.hpp" 36 #include "gc_implementation/shared/copyFailedInfo.hpp" 37 #include "gc_implementation/shared/spaceDecorator.hpp" 38 #include "memory/defNewGeneration.inline.hpp" 39 #include "memory/genCollectedHeap.hpp" 40 #include "memory/genOopClosures.inline.hpp" 41 #include "memory/generation.hpp" 42 #include "memory/generation.inline.hpp" 43 #include "memory/referencePolicy.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "memory/sharedHeap.hpp" 46 #include "memory/space.hpp" 47 #include "oops/objArrayOop.hpp" 48 #include "oops/oop.inline.hpp" 49 #include "oops/oop.pcgc.inline.hpp" 50 #include "runtime/handles.hpp" 51 #include "runtime/handles.inline.hpp" 52 #include "runtime/java.hpp" 53 #include "runtime/thread.hpp" 54 #include "utilities/copy.hpp" 55 #include "utilities/globalDefinitions.hpp" 56 #include "utilities/workgroup.hpp" 57 58 #ifdef _MSC_VER 59 #pragma warning( push ) 60 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 61 #endif 62 ParScanThreadState::ParScanThreadState(Space* to_space_, 63 ParNewGeneration* gen_, 64 Generation* old_gen_, 65 int thread_num_, 66 ObjToScanQueueSet* work_queue_set_, 67 Stack<oop, mtGC>* overflow_stacks_, 68 size_t desired_plab_sz_, 69 ParallelTaskTerminator& term_) : 70 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), 71 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), 72 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), 73 _ageTable(false), // false ==> not the global age table, no perf data. 74 _to_space_alloc_buffer(desired_plab_sz_), 75 _to_space_closure(gen_, this), _old_gen_closure(gen_, this), 76 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), 77 _older_gen_closure(gen_, this), 78 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, 79 &_to_space_root_closure, gen_, &_old_gen_root_closure, 80 work_queue_set_, &term_), 81 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), 82 _keep_alive_closure(&_scan_weak_ref_closure), 83 _strong_roots_time(0.0), _term_time(0.0) 84 { 85 #if TASKQUEUE_STATS 86 _term_attempts = 0; 87 _overflow_refills = 0; 88 _overflow_refill_objs = 0; 89 #endif // TASKQUEUE_STATS 90 91 _survivor_chunk_array = 92 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); 93 _hash_seed = 17; // Might want to take time-based random value. 94 _start = os::elapsedTime(); 95 _old_gen_closure.set_generation(old_gen_); 96 _old_gen_root_closure.set_generation(old_gen_); 97 } 98 #ifdef _MSC_VER 99 #pragma warning( pop ) 100 #endif 101 102 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, 103 size_t plab_word_size) { 104 ChunkArray* sca = survivor_chunk_array(); 105 if (sca != NULL) { 106 // A non-null SCA implies that we want the PLAB data recorded. 107 sca->record_sample(plab_start, plab_word_size); 108 } 109 } 110 111 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { 112 return new_obj->is_objArray() && 113 arrayOop(new_obj)->length() > ParGCArrayScanChunk && 114 new_obj != old_obj; 115 } 116 117 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { 118 assert(old->is_objArray(), "must be obj array"); 119 assert(old->is_forwarded(), "must be forwarded"); 120 assert(Universe::heap()->is_in_reserved(old), "must be in heap."); 121 assert(!old_gen()->is_in(old), "must be in young generation."); 122 123 objArrayOop obj = objArrayOop(old->forwardee()); 124 // Process ParGCArrayScanChunk elements now 125 // and push the remainder back onto queue 126 int start = arrayOop(old)->length(); 127 int end = obj->length(); 128 int remainder = end - start; 129 assert(start <= end, "just checking"); 130 if (remainder > 2 * ParGCArrayScanChunk) { 131 // Test above combines last partial chunk with a full chunk 132 end = start + ParGCArrayScanChunk; 133 arrayOop(old)->set_length(end); 134 // Push remainder. 135 bool ok = work_queue()->push(old); 136 assert(ok, "just popped, push must be okay"); 137 } else { 138 // Restore length so that it can be used if there 139 // is a promotion failure and forwarding pointers 140 // must be removed. 141 arrayOop(old)->set_length(end); 142 } 143 144 // process our set of indices (include header in first chunk) 145 // should make sure end is even (aligned to HeapWord in case of compressed oops) 146 if ((HeapWord *)obj < young_old_boundary()) { 147 // object is in to_space 148 obj->oop_iterate_range(&_to_space_closure, start, end); 149 } else { 150 // object is in old generation 151 obj->oop_iterate_range(&_old_gen_closure, start, end); 152 } 153 } 154 155 156 void ParScanThreadState::trim_queues(int max_size) { 157 ObjToScanQueue* queue = work_queue(); 158 do { 159 while (queue->size() > (juint)max_size) { 160 oop obj_to_scan; 161 if (queue->pop_local(obj_to_scan)) { 162 if ((HeapWord *)obj_to_scan < young_old_boundary()) { 163 if (obj_to_scan->is_objArray() && 164 obj_to_scan->is_forwarded() && 165 obj_to_scan->forwardee() != obj_to_scan) { 166 scan_partial_array_and_push_remainder(obj_to_scan); 167 } else { 168 // object is in to_space 169 obj_to_scan->oop_iterate(&_to_space_closure); 170 } 171 } else { 172 // object is in old generation 173 obj_to_scan->oop_iterate(&_old_gen_closure); 174 } 175 } 176 } 177 // For the case of compressed oops, we have a private, non-shared 178 // overflow stack, so we eagerly drain it so as to more evenly 179 // distribute load early. Note: this may be good to do in 180 // general rather than delay for the final stealing phase. 181 // If applicable, we'll transfer a set of objects over to our 182 // work queue, allowing them to be stolen and draining our 183 // private overflow stack. 184 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); 185 } 186 187 bool ParScanThreadState::take_from_overflow_stack() { 188 assert(ParGCUseLocalOverflow, "Else should not call"); 189 assert(young_gen()->overflow_list() == NULL, "Error"); 190 ObjToScanQueue* queue = work_queue(); 191 Stack<oop, mtGC>* const of_stack = overflow_stack(); 192 const size_t num_overflow_elems = of_stack->size(); 193 const size_t space_available = queue->max_elems() - queue->size(); 194 const size_t num_take_elems = MIN3(space_available / 4, 195 ParGCDesiredObjsFromOverflowList, 196 num_overflow_elems); 197 // Transfer the most recent num_take_elems from the overflow 198 // stack to our work queue. 199 for (size_t i = 0; i != num_take_elems; i++) { 200 oop cur = of_stack->pop(); 201 oop obj_to_push = cur->forwardee(); 202 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap"); 203 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); 204 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap"); 205 if (should_be_partially_scanned(obj_to_push, cur)) { 206 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 207 obj_to_push = cur; 208 } 209 bool ok = queue->push(obj_to_push); 210 assert(ok, "Should have succeeded"); 211 } 212 assert(young_gen()->overflow_list() == NULL, "Error"); 213 return num_take_elems > 0; // was something transferred? 214 } 215 216 void ParScanThreadState::push_on_overflow_stack(oop p) { 217 assert(ParGCUseLocalOverflow, "Else should not call"); 218 overflow_stack()->push(p); 219 assert(young_gen()->overflow_list() == NULL, "Error"); 220 } 221 222 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { 223 224 // Otherwise, if the object is small enough, try to reallocate the 225 // buffer. 226 HeapWord* obj = NULL; 227 if (!_to_space_full) { 228 ParGCAllocBuffer* const plab = to_space_alloc_buffer(); 229 Space* const sp = to_space(); 230 if (word_sz * 100 < 231 ParallelGCBufferWastePct * plab->word_sz()) { 232 // Is small enough; abandon this buffer and start a new one. 233 plab->retire(false, false); 234 size_t buf_size = plab->word_sz(); 235 HeapWord* buf_space = sp->par_allocate(buf_size); 236 if (buf_space == NULL) { 237 const size_t min_bytes = 238 ParGCAllocBuffer::min_size() << LogHeapWordSize; 239 size_t free_bytes = sp->free(); 240 while(buf_space == NULL && free_bytes >= min_bytes) { 241 buf_size = free_bytes >> LogHeapWordSize; 242 assert(buf_size == (size_t)align_object_size(buf_size), 243 "Invariant"); 244 buf_space = sp->par_allocate(buf_size); 245 free_bytes = sp->free(); 246 } 247 } 248 if (buf_space != NULL) { 249 plab->set_word_size(buf_size); 250 plab->set_buf(buf_space); 251 record_survivor_plab(buf_space, buf_size); 252 obj = plab->allocate(word_sz); 253 // Note that we cannot compare buf_size < word_sz below 254 // because of AlignmentReserve (see ParGCAllocBuffer::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 } 264 265 } else { 266 // Too large; allocate the object individually. 267 obj = sp->par_allocate(word_sz); 268 } 269 } 270 return obj; 271 } 272 273 274 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, 275 size_t word_sz) { 276 // Is the alloc in the current alloc buffer? 277 if (to_space_alloc_buffer()->contains(obj)) { 278 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), 279 "Should contain whole object."); 280 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 281 } else { 282 CollectedHeap::fill_with_object(obj, word_sz); 283 } 284 } 285 286 void ParScanThreadState::print_promotion_failure_size() { 287 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) { 288 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", 289 _thread_num, _promotion_failed_info.first_size()); 290 } 291 } 292 293 class ParScanThreadStateSet: private ResourceArray { 294 public: 295 // Initializes states for the specified number of threads; 296 ParScanThreadStateSet(int num_threads, 297 Space& to_space, 298 ParNewGeneration& gen, 299 Generation& old_gen, 300 ObjToScanQueueSet& queue_set, 301 Stack<oop, mtGC>* overflow_stacks_, 302 size_t desired_plab_sz, 303 ParallelTaskTerminator& term); 304 305 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 306 307 inline ParScanThreadState& thread_state(int i); 308 309 void trace_promotion_failed(YoungGCTracer& gc_tracer); 310 void reset(int active_workers, bool promotion_failed); 311 void flush(); 312 313 #if TASKQUEUE_STATS 314 static void 315 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); 316 void print_termination_stats(outputStream* const st = gclog_or_tty); 317 static void 318 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); 319 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); 320 void reset_stats(); 321 #endif // TASKQUEUE_STATS 322 323 private: 324 ParallelTaskTerminator& _term; 325 ParNewGeneration& _gen; 326 Generation& _next_gen; 327 public: 328 bool is_valid(int id) const { return id < length(); } 329 ParallelTaskTerminator* terminator() { return &_term; } 330 }; 331 332 333 ParScanThreadStateSet::ParScanThreadStateSet( 334 int num_threads, Space& to_space, ParNewGeneration& gen, 335 Generation& old_gen, ObjToScanQueueSet& queue_set, 336 Stack<oop, mtGC>* overflow_stacks, 337 size_t desired_plab_sz, ParallelTaskTerminator& term) 338 : ResourceArray(sizeof(ParScanThreadState), num_threads), 339 _gen(gen), _next_gen(old_gen), _term(term) 340 { 341 assert(num_threads > 0, "sanity check!"); 342 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 343 "overflow_stack allocation mismatch"); 344 // Initialize states. 345 for (int i = 0; i < num_threads; ++i) { 346 new ((ParScanThreadState*)_data + i) 347 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, 348 overflow_stacks, desired_plab_sz, term); 349 } 350 } 351 352 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) 353 { 354 assert(i >= 0 && i < length(), "sanity check!"); 355 return ((ParScanThreadState*)_data)[i]; 356 } 357 358 void ParScanThreadStateSet::trace_promotion_failed(YoungGCTracer& gc_tracer) { 359 for (int i = 0; i < length(); ++i) { 360 if (thread_state(i).promotion_failed()) { 361 gc_tracer.report_promotion_failed(thread_state(i).promotion_failed_info()); 362 thread_state(i).promotion_failed_info().reset(); 363 } 364 } 365 } 366 367 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) 368 { 369 _term.reset_for_reuse(active_threads); 370 if (promotion_failed) { 371 for (int i = 0; i < length(); ++i) { 372 thread_state(i).print_promotion_failure_size(); 373 } 374 } 375 } 376 377 #if TASKQUEUE_STATS 378 void 379 ParScanThreadState::reset_stats() 380 { 381 taskqueue_stats().reset(); 382 _term_attempts = 0; 383 _overflow_refills = 0; 384 _overflow_refill_objs = 0; 385 } 386 387 void ParScanThreadStateSet::reset_stats() 388 { 389 for (int i = 0; i < length(); ++i) { 390 thread_state(i).reset_stats(); 391 } 392 } 393 394 void 395 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) 396 { 397 st->print_raw_cr("GC Termination Stats"); 398 st->print_raw_cr(" elapsed --strong roots-- " 399 "-------termination-------"); 400 st->print_raw_cr("thr ms ms % " 401 " ms % attempts"); 402 st->print_raw_cr("--- --------- --------- ------ " 403 "--------- ------ --------"); 404 } 405 406 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) 407 { 408 print_termination_stats_hdr(st); 409 410 for (int i = 0; i < length(); ++i) { 411 const ParScanThreadState & pss = thread_state(i); 412 const double elapsed_ms = pss.elapsed_time() * 1000.0; 413 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 414 const double term_ms = pss.term_time() * 1000.0; 415 st->print_cr("%3d %9.2f %9.2f %6.2f " 416 "%9.2f %6.2f " SIZE_FORMAT_W(8), 417 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 418 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); 419 } 420 } 421 422 // Print stats related to work queue activity. 423 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) 424 { 425 st->print_raw_cr("GC Task Stats"); 426 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 427 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 428 } 429 430 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) 431 { 432 print_taskqueue_stats_hdr(st); 433 434 TaskQueueStats totals; 435 for (int i = 0; i < length(); ++i) { 436 const ParScanThreadState & pss = thread_state(i); 437 const TaskQueueStats & stats = pss.taskqueue_stats(); 438 st->print("%3d ", i); stats.print(st); st->cr(); 439 totals += stats; 440 441 if (pss.overflow_refills() > 0) { 442 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 443 SIZE_FORMAT_W(10) " overflow objects", 444 pss.overflow_refills(), pss.overflow_refill_objs()); 445 } 446 } 447 st->print("tot "); totals.print(st); st->cr(); 448 449 DEBUG_ONLY(totals.verify()); 450 } 451 #endif // TASKQUEUE_STATS 452 453 void ParScanThreadStateSet::flush() 454 { 455 // Work in this loop should be kept as lightweight as 456 // possible since this might otherwise become a bottleneck 457 // to scaling. Should we add heavy-weight work into this 458 // loop, consider parallelizing the loop into the worker threads. 459 for (int i = 0; i < length(); ++i) { 460 ParScanThreadState& par_scan_state = thread_state(i); 461 462 // Flush stats related to To-space PLAB activity and 463 // retire the last buffer. 464 par_scan_state.to_space_alloc_buffer()-> 465 flush_stats_and_retire(_gen.plab_stats(), 466 true /* end_of_gc */, 467 false /* retain */); 468 469 // Every thread has its own age table. We need to merge 470 // them all into one. 471 ageTable *local_table = par_scan_state.age_table(); 472 _gen.age_table()->merge(local_table); 473 474 // Inform old gen that we're done. 475 _next_gen.par_promote_alloc_done(i); 476 _next_gen.par_oop_since_save_marks_iterate_done(i); 477 } 478 479 if (UseConcMarkSweepGC && ParallelGCThreads > 0) { 480 // We need to call this even when ResizeOldPLAB is disabled 481 // so as to avoid breaking some asserts. While we may be able 482 // to avoid this by reorganizing the code a bit, I am loathe 483 // to do that unless we find cases where ergo leads to bad 484 // performance. 485 CFLS_LAB::compute_desired_plab_size(); 486 } 487 } 488 489 ParScanClosure::ParScanClosure(ParNewGeneration* g, 490 ParScanThreadState* par_scan_state) : 491 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) 492 { 493 assert(_g->level() == 0, "Optimized for youngest generation"); 494 _boundary = _g->reserved().end(); 495 } 496 497 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } 498 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } 499 500 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } 501 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } 502 503 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 504 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 505 506 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 507 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 508 509 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 510 ParScanThreadState* par_scan_state) 511 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 512 {} 513 514 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } 515 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } 516 517 #ifdef WIN32 518 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ 519 #endif 520 521 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( 522 ParScanThreadState* par_scan_state_, 523 ParScanWithoutBarrierClosure* to_space_closure_, 524 ParScanWithBarrierClosure* old_gen_closure_, 525 ParRootScanWithoutBarrierClosure* to_space_root_closure_, 526 ParNewGeneration* par_gen_, 527 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, 528 ObjToScanQueueSet* task_queues_, 529 ParallelTaskTerminator* terminator_) : 530 531 _par_scan_state(par_scan_state_), 532 _to_space_closure(to_space_closure_), 533 _old_gen_closure(old_gen_closure_), 534 _to_space_root_closure(to_space_root_closure_), 535 _old_gen_root_closure(old_gen_root_closure_), 536 _par_gen(par_gen_), 537 _task_queues(task_queues_), 538 _terminator(terminator_) 539 {} 540 541 void ParEvacuateFollowersClosure::do_void() { 542 ObjToScanQueue* work_q = par_scan_state()->work_queue(); 543 544 while (true) { 545 546 // Scan to-space and old-gen objs until we run out of both. 547 oop obj_to_scan; 548 par_scan_state()->trim_queues(0); 549 550 // We have no local work, attempt to steal from other threads. 551 552 // attempt to steal work from promoted. 553 if (task_queues()->steal(par_scan_state()->thread_num(), 554 par_scan_state()->hash_seed(), 555 obj_to_scan)) { 556 bool res = work_q->push(obj_to_scan); 557 assert(res, "Empty queue should have room for a push."); 558 559 // if successful, goto Start. 560 continue; 561 562 // try global overflow list. 563 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 564 continue; 565 } 566 567 // Otherwise, offer termination. 568 par_scan_state()->start_term_time(); 569 if (terminator()->offer_termination()) break; 570 par_scan_state()->end_term_time(); 571 } 572 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 573 "Broken overflow list?"); 574 // Finish the last termination pause. 575 par_scan_state()->end_term_time(); 576 } 577 578 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, 579 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : 580 AbstractGangTask("ParNewGeneration collection"), 581 _gen(gen), _next_gen(next_gen), 582 _young_old_boundary(young_old_boundary), 583 _state_set(state_set) 584 {} 585 586 // Reset the terminator for the given number of 587 // active threads. 588 void ParNewGenTask::set_for_termination(int active_workers) { 589 _state_set->reset(active_workers, _gen->promotion_failed()); 590 // Should the heap be passed in? There's only 1 for now so 591 // grab it instead. 592 GenCollectedHeap* gch = GenCollectedHeap::heap(); 593 gch->set_n_termination(active_workers); 594 } 595 596 void ParNewGenTask::work(uint worker_id) { 597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 598 // Since this is being done in a separate thread, need new resource 599 // and handle marks. 600 ResourceMark rm; 601 HandleMark hm; 602 // We would need multiple old-gen queues otherwise. 603 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); 604 605 Generation* old_gen = gch->next_gen(_gen); 606 607 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 608 assert(_state_set->is_valid(worker_id), "Should not have been called"); 609 610 par_scan_state.set_young_old_boundary(_young_old_boundary); 611 612 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), 613 gch->rem_set()->klass_rem_set()); 614 615 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 616 617 par_scan_state.start_strong_roots(); 618 gch->gen_process_strong_roots(_gen->level(), 619 true, // Process younger gens, if any, 620 // as strong roots. 621 false, // no scope; this is parallel code 622 true, // is scavenging 623 SharedHeap::ScanningOption(so), 624 &par_scan_state.to_space_root_closure(), 625 true, // walk *all* scavengable nmethods 626 &par_scan_state.older_gen_closure(), 627 &klass_scan_closure); 628 par_scan_state.end_strong_roots(); 629 630 // "evacuate followers". 631 par_scan_state.evacuate_followers_closure().do_void(); 632 } 633 634 #ifdef _MSC_VER 635 #pragma warning( push ) 636 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 637 #endif 638 ParNewGeneration:: 639 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) 640 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), 641 _overflow_list(NULL), 642 _is_alive_closure(this), 643 _plab_stats(YoungPLABSize, PLABWeight) 644 { 645 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 646 NOT_PRODUCT(_num_par_pushes = 0;) 647 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 648 guarantee(_task_queues != NULL, "task_queues allocation failure."); 649 650 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { 651 ObjToScanQueue *q = new ObjToScanQueue(); 652 guarantee(q != NULL, "work_queue Allocation failure."); 653 _task_queues->register_queue(i1, q); 654 } 655 656 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) 657 _task_queues->queue(i2)->initialize(); 658 659 _overflow_stacks = NULL; 660 if (ParGCUseLocalOverflow) { 661 662 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal 663 // with ',' 664 typedef Stack<oop, mtGC> GCOopStack; 665 666 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 667 for (size_t i = 0; i < ParallelGCThreads; ++i) { 668 new (_overflow_stacks + i) Stack<oop, mtGC>(); 669 } 670 } 671 672 if (UsePerfData) { 673 EXCEPTION_MARK; 674 ResourceMark rm; 675 676 const char* cname = 677 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 678 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 679 ParallelGCThreads, CHECK); 680 } 681 } 682 #ifdef _MSC_VER 683 #pragma warning( pop ) 684 #endif 685 686 // ParNewGeneration:: 687 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 688 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 689 690 template <class T> 691 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 692 #ifdef ASSERT 693 { 694 assert(!oopDesc::is_null(*p), "expected non-null ref"); 695 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 696 // We never expect to see a null reference being processed 697 // as a weak reference. 698 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 699 } 700 #endif // ASSERT 701 702 _par_cl->do_oop_nv(p); 703 704 if (Universe::heap()->is_in_reserved(p)) { 705 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 706 _rs->write_ref_field_gc_par(p, obj); 707 } 708 } 709 710 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 711 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 712 713 // ParNewGeneration:: 714 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 715 DefNewGeneration::KeepAliveClosure(cl) {} 716 717 template <class T> 718 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 719 #ifdef ASSERT 720 { 721 assert(!oopDesc::is_null(*p), "expected non-null ref"); 722 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 723 // We never expect to see a null reference being processed 724 // as a weak reference. 725 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 726 } 727 #endif // ASSERT 728 729 _cl->do_oop_nv(p); 730 731 if (Universe::heap()->is_in_reserved(p)) { 732 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 733 _rs->write_ref_field_gc_par(p, obj); 734 } 735 } 736 737 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 738 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 739 740 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 741 T heap_oop = oopDesc::load_heap_oop(p); 742 if (!oopDesc::is_null(heap_oop)) { 743 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 744 if ((HeapWord*)obj < _boundary) { 745 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 746 oop new_obj = obj->is_forwarded() 747 ? obj->forwardee() 748 : _g->DefNewGeneration::copy_to_survivor_space(obj); 749 oopDesc::encode_store_heap_oop_not_null(p, new_obj); 750 } 751 if (_gc_barrier) { 752 // If p points to a younger generation, mark the card. 753 if ((HeapWord*)obj < _gen_boundary) { 754 _rs->write_ref_field_gc_par(p, obj); 755 } 756 } 757 } 758 } 759 760 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 761 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 762 763 class ParNewRefProcTaskProxy: public AbstractGangTask { 764 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 765 public: 766 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, 767 Generation& next_gen, 768 HeapWord* young_old_boundary, 769 ParScanThreadStateSet& state_set); 770 771 private: 772 virtual void work(uint worker_id); 773 virtual void set_for_termination(int active_workers) { 774 _state_set.terminator()->reset_for_reuse(active_workers); 775 } 776 private: 777 ParNewGeneration& _gen; 778 ProcessTask& _task; 779 Generation& _next_gen; 780 HeapWord* _young_old_boundary; 781 ParScanThreadStateSet& _state_set; 782 }; 783 784 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( 785 ProcessTask& task, ParNewGeneration& gen, 786 Generation& next_gen, 787 HeapWord* young_old_boundary, 788 ParScanThreadStateSet& state_set) 789 : AbstractGangTask("ParNewGeneration parallel reference processing"), 790 _gen(gen), 791 _task(task), 792 _next_gen(next_gen), 793 _young_old_boundary(young_old_boundary), 794 _state_set(state_set) 795 { 796 } 797 798 void ParNewRefProcTaskProxy::work(uint worker_id) 799 { 800 ResourceMark rm; 801 HandleMark hm; 802 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 803 par_scan_state.set_young_old_boundary(_young_old_boundary); 804 _task.work(worker_id, par_scan_state.is_alive_closure(), 805 par_scan_state.keep_alive_closure(), 806 par_scan_state.evacuate_followers_closure()); 807 } 808 809 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { 810 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 811 EnqueueTask& _task; 812 813 public: 814 ParNewRefEnqueueTaskProxy(EnqueueTask& task) 815 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), 816 _task(task) 817 { } 818 819 virtual void work(uint worker_id) 820 { 821 _task.work(worker_id); 822 } 823 }; 824 825 826 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) 827 { 828 GenCollectedHeap* gch = GenCollectedHeap::heap(); 829 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 830 "not a generational heap"); 831 FlexibleWorkGang* workers = gch->workers(); 832 assert(workers != NULL, "Need parallel worker threads."); 833 _state_set.reset(workers->active_workers(), _generation.promotion_failed()); 834 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), 835 _generation.reserved().end(), _state_set); 836 workers->run_task(&rp_task); 837 _state_set.reset(0 /* bad value in debug if not reset */, 838 _generation.promotion_failed()); 839 } 840 841 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) 842 { 843 GenCollectedHeap* gch = GenCollectedHeap::heap(); 844 FlexibleWorkGang* workers = gch->workers(); 845 assert(workers != NULL, "Need parallel worker threads."); 846 ParNewRefEnqueueTaskProxy enq_task(task); 847 workers->run_task(&enq_task); 848 } 849 850 void ParNewRefProcTaskExecutor::set_single_threaded_mode() 851 { 852 _state_set.flush(); 853 GenCollectedHeap* gch = GenCollectedHeap::heap(); 854 gch->set_par_threads(0); // 0 ==> non-parallel. 855 gch->save_marks(); 856 } 857 858 ScanClosureWithParBarrier:: 859 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 860 ScanClosure(g, gc_barrier) {} 861 862 EvacuateFollowersClosureGeneral:: 863 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, 864 OopsInGenClosure* cur, 865 OopsInGenClosure* older) : 866 _gch(gch), _level(level), 867 _scan_cur_or_nonheap(cur), _scan_older(older) 868 {} 869 870 void EvacuateFollowersClosureGeneral::do_void() { 871 do { 872 // Beware: this call will lead to closure applications via virtual 873 // calls. 874 _gch->oop_since_save_marks_iterate(_level, 875 _scan_cur_or_nonheap, 876 _scan_older); 877 } while (!_gch->no_allocs_since_save_marks(_level)); 878 } 879 880 881 // A Generation that does parallel young-gen collection. 882 883 bool ParNewGeneration::_avoid_promotion_undo = false; 884 885 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set, ParNewTracer& gc_tracer) { 886 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 887 _promo_failure_scan_stack.clear(true); // Clear cached segments. 888 889 remove_forwarding_pointers(); 890 if (PrintGCDetails) { 891 gclog_or_tty->print(" (promotion failed)"); 892 } 893 // All the spaces are in play for mark-sweep. 894 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 895 from()->set_next_compaction_space(to()); 896 gch->set_incremental_collection_failed(); 897 // Inform the next generation that a promotion failure occurred. 898 _next_gen->promotion_failure_occurred(); 899 900 // Trace promotion failure in the parallel GC threads 901 thread_state_set.trace_promotion_failed(gc_tracer); 902 // Single threaded code may have reported promotion failure to the global state 903 if (_promotion_failed_info.has_failed()) { 904 gc_tracer.report_promotion_failed(_promotion_failed_info); 905 } 906 // Reset the PromotionFailureALot counters. 907 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) 908 } 909 910 void ParNewGeneration::collect(bool full, 911 bool clear_all_soft_refs, 912 size_t size, 913 bool is_tlab) { 914 assert(full || size > 0, "otherwise we don't want to collect"); 915 916 GenCollectedHeap* gch = GenCollectedHeap::heap(); 917 918 _gc_timer->register_gc_start(os::elapsed_counter()); 919 920 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 921 "not a CMS generational heap"); 922 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); 923 FlexibleWorkGang* workers = gch->workers(); 924 assert(workers != NULL, "Need workgang for parallel work"); 925 int active_workers = 926 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), 927 workers->active_workers(), 928 Threads::number_of_non_daemon_threads()); 929 workers->set_active_workers(active_workers); 930 _next_gen = gch->next_gen(this); 931 assert(_next_gen != NULL, 932 "This must be the youngest gen, and not the only gen"); 933 assert(gch->n_gens() == 2, 934 "Par collection currently only works with single older gen."); 935 // Do we have to avoid promotion_undo? 936 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { 937 set_avoid_promotion_undo(true); 938 } 939 940 // If the next generation is too full to accommodate worst-case promotion 941 // from this generation, pass on collection; let the next generation 942 // do it. 943 if (!collection_attempt_is_safe()) { 944 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 945 return; 946 } 947 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 948 949 ParNewTracer gc_tracer; 950 gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 951 gch->trace_heap_before_gc(&gc_tracer); 952 953 init_assuming_no_promotion_failure(); 954 955 if (UseAdaptiveSizePolicy) { 956 set_survivor_overflow(false); 957 size_policy->minor_collection_begin(); 958 } 959 960 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL); 961 // Capture heap used before collection (for printing). 962 size_t gch_prev_used = gch->used(); 963 964 SpecializationStats::clear(); 965 966 age_table()->clear(); 967 to()->clear(SpaceDecorator::Mangle); 968 969 gch->save_marks(); 970 assert(workers != NULL, "Need parallel worker threads."); 971 int n_workers = active_workers; 972 973 // Set the correct parallelism (number of queues) in the reference processor 974 ref_processor()->set_active_mt_degree(n_workers); 975 976 // Always set the terminator for the active number of workers 977 // because only those workers go through the termination protocol. 978 ParallelTaskTerminator _term(n_workers, task_queues()); 979 ParScanThreadStateSet thread_state_set(workers->active_workers(), 980 *to(), *this, *_next_gen, *task_queues(), 981 _overflow_stacks, desired_plab_sz(), _term); 982 983 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); 984 gch->set_par_threads(n_workers); 985 gch->rem_set()->prepare_for_younger_refs_iterate(true); 986 // It turns out that even when we're using 1 thread, doing the work in a 987 // separate thread causes wide variance in run times. We can't help this 988 // in the multi-threaded case, but we special-case n=1 here to get 989 // repeatable measurements of the 1-thread overhead of the parallel code. 990 if (n_workers > 1) { 991 GenCollectedHeap::StrongRootsScope srs(gch); 992 workers->run_task(&tsk); 993 } else { 994 GenCollectedHeap::StrongRootsScope srs(gch); 995 tsk.work(0); 996 } 997 thread_state_set.reset(0 /* Bad value in debug if not reset */, 998 promotion_failed()); 999 1000 // Process (weak) reference objects found during scavenge. 1001 ReferenceProcessor* rp = ref_processor(); 1002 IsAliveClosure is_alive(this); 1003 ScanWeakRefClosure scan_weak_ref(this); 1004 KeepAliveClosure keep_alive(&scan_weak_ref); 1005 ScanClosure scan_without_gc_barrier(this, false); 1006 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 1007 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 1008 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, 1009 &scan_without_gc_barrier, &scan_with_gc_barrier); 1010 rp->setup_policy(clear_all_soft_refs); 1011 // Can the mt_degree be set later (at run_task() time would be best)? 1012 rp->set_active_mt_degree(active_workers); 1013 ReferenceProcessorStats stats; 1014 if (rp->processing_is_mt()) { 1015 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); 1016 stats = rp->process_discovered_references(&is_alive, &keep_alive, 1017 &evacuate_followers, &task_executor, 1018 _gc_timer); 1019 } else { 1020 thread_state_set.flush(); 1021 gch->set_par_threads(0); // 0 ==> non-parallel. 1022 gch->save_marks(); 1023 stats = rp->process_discovered_references(&is_alive, &keep_alive, 1024 &evacuate_followers, NULL, 1025 _gc_timer); 1026 } 1027 gc_tracer.report_gc_reference_stats(stats); 1028 if (!promotion_failed()) { 1029 // Swap the survivor spaces. 1030 eden()->clear(SpaceDecorator::Mangle); 1031 from()->clear(SpaceDecorator::Mangle); 1032 if (ZapUnusedHeapArea) { 1033 // This is now done here because of the piece-meal mangling which 1034 // can check for valid mangling at intermediate points in the 1035 // collection(s). When a minor collection fails to collect 1036 // sufficient space resizing of the young generation can occur 1037 // an redistribute the spaces in the young generation. Mangle 1038 // here so that unzapped regions don't get distributed to 1039 // other spaces. 1040 to()->mangle_unused_area(); 1041 } 1042 swap_spaces(); 1043 1044 // A successful scavenge should restart the GC time limit count which is 1045 // for full GC's. 1046 size_policy->reset_gc_overhead_limit_count(); 1047 1048 assert(to()->is_empty(), "to space should be empty now"); 1049 1050 adjust_desired_tenuring_threshold(); 1051 } else { 1052 handle_promotion_failed(gch, thread_state_set, gc_tracer); 1053 } 1054 // set new iteration safe limit for the survivor spaces 1055 from()->set_concurrent_iteration_safe_limit(from()->top()); 1056 to()->set_concurrent_iteration_safe_limit(to()->top()); 1057 1058 if (ResizePLAB) { 1059 plab_stats()->adjust_desired_plab_sz(n_workers); 1060 } 1061 1062 if (PrintGC && !PrintGCDetails) { 1063 gch->print_heap_change(gch_prev_used); 1064 } 1065 1066 if (PrintGCDetails && ParallelGCVerbose) { 1067 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); 1068 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); 1069 } 1070 1071 if (UseAdaptiveSizePolicy) { 1072 size_policy->minor_collection_end(gch->gc_cause()); 1073 size_policy->avg_survived()->sample(from()->used()); 1074 } 1075 1076 // We need to use a monotonically non-deccreasing time in ms 1077 // or we will see time-warp warnings and os::javaTimeMillis() 1078 // does not guarantee monotonicity. 1079 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1080 update_time_of_last_gc(now); 1081 1082 SpecializationStats::print(); 1083 1084 rp->set_enqueuing_is_done(true); 1085 if (rp->processing_is_mt()) { 1086 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); 1087 rp->enqueue_discovered_references(&task_executor); 1088 } else { 1089 rp->enqueue_discovered_references(NULL); 1090 } 1091 rp->verify_no_references_recorded(); 1092 1093 gch->trace_heap_after_gc(&gc_tracer); 1094 gc_tracer.report_tenuring_threshold(tenuring_threshold()); 1095 1096 _gc_timer->register_gc_end(os::elapsed_counter()); 1097 1098 gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1099 } 1100 1101 static int sum; 1102 void ParNewGeneration::waste_some_time() { 1103 for (int i = 0; i < 100; i++) { 1104 sum += i; 1105 } 1106 } 1107 1108 static const oop ClaimedForwardPtr = oop(0x4); 1109 1110 // Because of concurrency, there are times where an object for which 1111 // "is_forwarded()" is true contains an "interim" forwarding pointer 1112 // value. Such a value will soon be overwritten with a real value. 1113 // This method requires "obj" to have a forwarding pointer, and waits, if 1114 // necessary for a real one to be inserted, and returns it. 1115 1116 oop ParNewGeneration::real_forwardee(oop obj) { 1117 oop forward_ptr = obj->forwardee(); 1118 if (forward_ptr != ClaimedForwardPtr) { 1119 return forward_ptr; 1120 } else { 1121 return real_forwardee_slow(obj); 1122 } 1123 } 1124 1125 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1126 // Spin-read if it is claimed but not yet written by another thread. 1127 oop forward_ptr = obj->forwardee(); 1128 while (forward_ptr == ClaimedForwardPtr) { 1129 waste_some_time(); 1130 assert(obj->is_forwarded(), "precondition"); 1131 forward_ptr = obj->forwardee(); 1132 } 1133 return forward_ptr; 1134 } 1135 1136 #ifdef ASSERT 1137 bool ParNewGeneration::is_legal_forward_ptr(oop p) { 1138 return 1139 (_avoid_promotion_undo && p == ClaimedForwardPtr) 1140 || Universe::heap()->is_in_reserved(p); 1141 } 1142 #endif 1143 1144 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { 1145 if (m->must_be_preserved_for_promotion_failure(obj)) { 1146 // We should really have separate per-worker stacks, rather 1147 // than use locking of a common pair of stacks. 1148 MutexLocker ml(ParGCRareEvent_lock); 1149 preserve_mark(obj, m); 1150 } 1151 } 1152 1153 // Multiple GC threads may try to promote an object. If the object 1154 // is successfully promoted, a forwarding pointer will be installed in 1155 // the object in the young generation. This method claims the right 1156 // to install the forwarding pointer before it copies the object, 1157 // thus avoiding the need to undo the copy as in 1158 // copy_to_survivor_space_avoiding_with_undo. 1159 1160 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( 1161 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { 1162 // In the sequential version, this assert also says that the object is 1163 // not forwarded. That might not be the case here. It is the case that 1164 // the caller observed it to be not forwarded at some time in the past. 1165 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1166 1167 // The sequential code read "old->age()" below. That doesn't work here, 1168 // since the age is in the mark word, and that might be overwritten with 1169 // a forwarding pointer by a parallel thread. So we must save the mark 1170 // word in a local and then analyze it. 1171 oopDesc dummyOld; 1172 dummyOld.set_mark(m); 1173 assert(!dummyOld.is_forwarded(), 1174 "should not be called with forwarding pointer mark word."); 1175 1176 oop new_obj = NULL; 1177 oop forward_ptr; 1178 1179 // Try allocating obj in to-space (unless too old) 1180 if (dummyOld.age() < tenuring_threshold()) { 1181 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1182 if (new_obj == NULL) { 1183 set_survivor_overflow(true); 1184 } 1185 } 1186 1187 if (new_obj == NULL) { 1188 // Either to-space is full or we decided to promote 1189 // try allocating obj tenured 1190 1191 // Attempt to install a null forwarding pointer (atomically), 1192 // to claim the right to install the real forwarding pointer. 1193 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); 1194 if (forward_ptr != NULL) { 1195 // someone else beat us to it. 1196 return real_forwardee(old); 1197 } 1198 1199 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), 1200 old, m, sz); 1201 1202 if (new_obj == NULL) { 1203 // promotion failed, forward to self 1204 _promotion_failed = true; 1205 new_obj = old; 1206 1207 preserve_mark_if_necessary(old, m); 1208 par_scan_state->register_promotion_failure(sz); 1209 } 1210 1211 old->forward_to(new_obj); 1212 forward_ptr = NULL; 1213 } else { 1214 // Is in to-space; do copying ourselves. 1215 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1216 forward_ptr = old->forward_to_atomic(new_obj); 1217 // Restore the mark word copied above. 1218 new_obj->set_mark(m); 1219 // Increment age if obj still in new generation 1220 new_obj->incr_age(); 1221 par_scan_state->age_table()->add(new_obj, sz); 1222 } 1223 assert(new_obj != NULL, "just checking"); 1224 1225 #ifndef PRODUCT 1226 // This code must come after the CAS test, or it will print incorrect 1227 // information. 1228 if (TraceScavenge) { 1229 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1230 is_in_reserved(new_obj) ? "copying" : "tenuring", 1231 new_obj->klass()->internal_name(), old, new_obj, new_obj->size()); 1232 } 1233 #endif 1234 1235 if (forward_ptr == NULL) { 1236 oop obj_to_push = new_obj; 1237 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1238 // Length field used as index of next element to be scanned. 1239 // Real length can be obtained from real_forwardee() 1240 arrayOop(old)->set_length(0); 1241 obj_to_push = old; 1242 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1243 "push forwarded object"); 1244 } 1245 // Push it on one of the queues of to-be-scanned objects. 1246 bool simulate_overflow = false; 1247 NOT_PRODUCT( 1248 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1249 // simulate a stack overflow 1250 simulate_overflow = true; 1251 } 1252 ) 1253 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1254 // Add stats for overflow pushes. 1255 if (Verbose && PrintGCDetails) { 1256 gclog_or_tty->print("queue overflow!\n"); 1257 } 1258 push_on_overflow_list(old, par_scan_state); 1259 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1260 } 1261 1262 return new_obj; 1263 } 1264 1265 // Oops. Someone beat us to it. Undo the allocation. Where did we 1266 // allocate it? 1267 if (is_in_reserved(new_obj)) { 1268 // Must be in to_space. 1269 assert(to()->is_in_reserved(new_obj), "Checking"); 1270 if (forward_ptr == ClaimedForwardPtr) { 1271 // Wait to get the real forwarding pointer value. 1272 forward_ptr = real_forwardee(old); 1273 } 1274 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1275 } 1276 1277 return forward_ptr; 1278 } 1279 1280 1281 // Multiple GC threads may try to promote the same object. If two 1282 // or more GC threads copy the object, only one wins the race to install 1283 // the forwarding pointer. The other threads have to undo their copy. 1284 1285 oop ParNewGeneration::copy_to_survivor_space_with_undo( 1286 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { 1287 1288 // In the sequential version, this assert also says that the object is 1289 // not forwarded. That might not be the case here. It is the case that 1290 // the caller observed it to be not forwarded at some time in the past. 1291 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1292 1293 // The sequential code read "old->age()" below. That doesn't work here, 1294 // since the age is in the mark word, and that might be overwritten with 1295 // a forwarding pointer by a parallel thread. So we must save the mark 1296 // word here, install it in a local oopDesc, and then analyze it. 1297 oopDesc dummyOld; 1298 dummyOld.set_mark(m); 1299 assert(!dummyOld.is_forwarded(), 1300 "should not be called with forwarding pointer mark word."); 1301 1302 bool failed_to_promote = false; 1303 oop new_obj = NULL; 1304 oop forward_ptr; 1305 1306 // Try allocating obj in to-space (unless too old) 1307 if (dummyOld.age() < tenuring_threshold()) { 1308 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1309 if (new_obj == NULL) { 1310 set_survivor_overflow(true); 1311 } 1312 } 1313 1314 if (new_obj == NULL) { 1315 // Either to-space is full or we decided to promote 1316 // try allocating obj tenured 1317 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), 1318 old, m, sz); 1319 1320 if (new_obj == NULL) { 1321 // promotion failed, forward to self 1322 forward_ptr = old->forward_to_atomic(old); 1323 new_obj = old; 1324 1325 if (forward_ptr != NULL) { 1326 return forward_ptr; // someone else succeeded 1327 } 1328 1329 _promotion_failed = true; 1330 failed_to_promote = true; 1331 1332 preserve_mark_if_necessary(old, m); 1333 par_scan_state->register_promotion_failure(sz); 1334 } 1335 } else { 1336 // Is in to-space; do copying ourselves. 1337 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1338 // Restore the mark word copied above. 1339 new_obj->set_mark(m); 1340 // Increment age if new_obj still in new generation 1341 new_obj->incr_age(); 1342 par_scan_state->age_table()->add(new_obj, sz); 1343 } 1344 assert(new_obj != NULL, "just checking"); 1345 1346 #ifndef PRODUCT 1347 // This code must come after the CAS test, or it will print incorrect 1348 // information. 1349 if (TraceScavenge) { 1350 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1351 is_in_reserved(new_obj) ? "copying" : "tenuring", 1352 new_obj->klass()->internal_name(), old, new_obj, new_obj->size()); 1353 } 1354 #endif 1355 1356 // Now attempt to install the forwarding pointer (atomically). 1357 // We have to copy the mark word before overwriting with forwarding 1358 // ptr, so we can restore it below in the copy. 1359 if (!failed_to_promote) { 1360 forward_ptr = old->forward_to_atomic(new_obj); 1361 } 1362 1363 if (forward_ptr == NULL) { 1364 oop obj_to_push = new_obj; 1365 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1366 // Length field used as index of next element to be scanned. 1367 // Real length can be obtained from real_forwardee() 1368 arrayOop(old)->set_length(0); 1369 obj_to_push = old; 1370 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1371 "push forwarded object"); 1372 } 1373 // Push it on one of the queues of to-be-scanned objects. 1374 bool simulate_overflow = false; 1375 NOT_PRODUCT( 1376 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1377 // simulate a stack overflow 1378 simulate_overflow = true; 1379 } 1380 ) 1381 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1382 // Add stats for overflow pushes. 1383 push_on_overflow_list(old, par_scan_state); 1384 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1385 } 1386 1387 return new_obj; 1388 } 1389 1390 // Oops. Someone beat us to it. Undo the allocation. Where did we 1391 // allocate it? 1392 if (is_in_reserved(new_obj)) { 1393 // Must be in to_space. 1394 assert(to()->is_in_reserved(new_obj), "Checking"); 1395 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1396 } else { 1397 assert(!_avoid_promotion_undo, "Should not be here if avoiding."); 1398 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), 1399 (HeapWord*)new_obj, sz); 1400 } 1401 1402 return forward_ptr; 1403 } 1404 1405 #ifndef PRODUCT 1406 // It's OK to call this multi-threaded; the worst thing 1407 // that can happen is that we'll get a bunch of closely 1408 // spaced simulated oveflows, but that's OK, in fact 1409 // probably good as it would exercise the overflow code 1410 // under contention. 1411 bool ParNewGeneration::should_simulate_overflow() { 1412 if (_overflow_counter-- <= 0) { // just being defensive 1413 _overflow_counter = ParGCWorkQueueOverflowInterval; 1414 return true; 1415 } else { 1416 return false; 1417 } 1418 } 1419 #endif 1420 1421 // In case we are using compressed oops, we need to be careful. 1422 // If the object being pushed is an object array, then its length 1423 // field keeps track of the "grey boundary" at which the next 1424 // incremental scan will be done (see ParGCArrayScanChunk). 1425 // When using compressed oops, this length field is kept in the 1426 // lower 32 bits of the erstwhile klass word and cannot be used 1427 // for the overflow chaining pointer (OCP below). As such the OCP 1428 // would itself need to be compressed into the top 32-bits in this 1429 // case. Unfortunately, see below, in the event that we have a 1430 // promotion failure, the node to be pushed on the list can be 1431 // outside of the Java heap, so the heap-based pointer compression 1432 // would not work (we would have potential aliasing between C-heap 1433 // and Java-heap pointers). For this reason, when using compressed 1434 // oops, we simply use a worker-thread-local, non-shared overflow 1435 // list in the form of a growable array, with a slightly different 1436 // overflow stack draining strategy. If/when we start using fat 1437 // stacks here, we can go back to using (fat) pointer chains 1438 // (although some performance comparisons would be useful since 1439 // single global lists have their own performance disadvantages 1440 // as we were made painfully aware not long ago, see 6786503). 1441 #define BUSY (oop(0x1aff1aff)) 1442 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1443 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1444 if (ParGCUseLocalOverflow) { 1445 // In the case of compressed oops, we use a private, not-shared 1446 // overflow stack. 1447 par_scan_state->push_on_overflow_stack(from_space_obj); 1448 } else { 1449 assert(!UseCompressedOops, "Error"); 1450 // if the object has been forwarded to itself, then we cannot 1451 // use the klass pointer for the linked list. Instead we have 1452 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1453 // XXX This is horribly inefficient when a promotion failure occurs 1454 // and should be fixed. XXX FIX ME !!! 1455 #ifndef PRODUCT 1456 Atomic::inc_ptr(&_num_par_pushes); 1457 assert(_num_par_pushes > 0, "Tautology"); 1458 #endif 1459 if (from_space_obj->forwardee() == from_space_obj) { 1460 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); 1461 listhead->forward_to(from_space_obj); 1462 from_space_obj = listhead; 1463 } 1464 oop observed_overflow_list = _overflow_list; 1465 oop cur_overflow_list; 1466 do { 1467 cur_overflow_list = observed_overflow_list; 1468 if (cur_overflow_list != BUSY) { 1469 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1470 } else { 1471 from_space_obj->set_klass_to_list_ptr(NULL); 1472 } 1473 observed_overflow_list = 1474 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); 1475 } while (cur_overflow_list != observed_overflow_list); 1476 } 1477 } 1478 1479 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1480 bool res; 1481 1482 if (ParGCUseLocalOverflow) { 1483 res = par_scan_state->take_from_overflow_stack(); 1484 } else { 1485 assert(!UseCompressedOops, "Error"); 1486 res = take_from_overflow_list_work(par_scan_state); 1487 } 1488 return res; 1489 } 1490 1491 1492 // *NOTE*: The overflow list manipulation code here and 1493 // in CMSCollector:: are very similar in shape, 1494 // except that in the CMS case we thread the objects 1495 // directly into the list via their mark word, and do 1496 // not need to deal with special cases below related 1497 // to chunking of object arrays and promotion failure 1498 // handling. 1499 // CR 6797058 has been filed to attempt consolidation of 1500 // the common code. 1501 // Because of the common code, if you make any changes in 1502 // the code below, please check the CMS version to see if 1503 // similar changes might be needed. 1504 // See CMSCollector::par_take_from_overflow_list() for 1505 // more extensive documentation comments. 1506 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1507 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1508 // How many to take? 1509 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1510 (size_t)ParGCDesiredObjsFromOverflowList); 1511 1512 assert(!UseCompressedOops, "Error"); 1513 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1514 if (_overflow_list == NULL) return false; 1515 1516 // Otherwise, there was something there; try claiming the list. 1517 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 1518 // Trim off a prefix of at most objsFromOverflow items 1519 Thread* tid = Thread::current(); 1520 size_t spin_count = (size_t)ParallelGCThreads; 1521 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1522 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1523 // someone grabbed it before we did ... 1524 // ... we spin for a short while... 1525 os::sleep(tid, sleep_time_millis, false); 1526 if (_overflow_list == NULL) { 1527 // nothing left to take 1528 return false; 1529 } else if (_overflow_list != BUSY) { 1530 // try and grab the prefix 1531 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 1532 } 1533 } 1534 if (prefix == NULL || prefix == BUSY) { 1535 // Nothing to take or waited long enough 1536 if (prefix == NULL) { 1537 // Write back the NULL in case we overwrote it with BUSY above 1538 // and it is still the same value. 1539 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1540 } 1541 return false; 1542 } 1543 assert(prefix != NULL && prefix != BUSY, "Error"); 1544 size_t i = 1; 1545 oop cur = prefix; 1546 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { 1547 i++; cur = cur->list_ptr_from_klass(); 1548 } 1549 1550 // Reattach remaining (suffix) to overflow list 1551 if (cur->klass_or_null() == NULL) { 1552 // Write back the NULL in lieu of the BUSY we wrote 1553 // above and it is still the same value. 1554 if (_overflow_list == BUSY) { 1555 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1556 } 1557 } else { 1558 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); 1559 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list 1560 cur->set_klass_to_list_ptr(NULL); // break off suffix 1561 // It's possible that the list is still in the empty(busy) state 1562 // we left it in a short while ago; in that case we may be 1563 // able to place back the suffix. 1564 oop observed_overflow_list = _overflow_list; 1565 oop cur_overflow_list = observed_overflow_list; 1566 bool attached = false; 1567 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1568 observed_overflow_list = 1569 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1570 if (cur_overflow_list == observed_overflow_list) { 1571 attached = true; 1572 break; 1573 } else cur_overflow_list = observed_overflow_list; 1574 } 1575 if (!attached) { 1576 // Too bad, someone else got in in between; we'll need to do a splice. 1577 // Find the last item of suffix list 1578 oop last = suffix; 1579 while (last->klass_or_null() != NULL) { 1580 last = last->list_ptr_from_klass(); 1581 } 1582 // Atomically prepend suffix to current overflow list 1583 observed_overflow_list = _overflow_list; 1584 do { 1585 cur_overflow_list = observed_overflow_list; 1586 if (cur_overflow_list != BUSY) { 1587 // Do the splice ... 1588 last->set_klass_to_list_ptr(cur_overflow_list); 1589 } else { // cur_overflow_list == BUSY 1590 last->set_klass_to_list_ptr(NULL); 1591 } 1592 observed_overflow_list = 1593 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1594 } while (cur_overflow_list != observed_overflow_list); 1595 } 1596 } 1597 1598 // Push objects on prefix list onto this thread's work queue 1599 assert(prefix != NULL && prefix != BUSY, "program logic"); 1600 cur = prefix; 1601 ssize_t n = 0; 1602 while (cur != NULL) { 1603 oop obj_to_push = cur->forwardee(); 1604 oop next = cur->list_ptr_from_klass(); 1605 cur->set_klass(obj_to_push->klass()); 1606 // This may be an array object that is self-forwarded. In that case, the list pointer 1607 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1608 if (!is_in_reserved(cur)) { 1609 // This can become a scaling bottleneck when there is work queue overflow coincident 1610 // with promotion failure. 1611 oopDesc* f = cur; 1612 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC); 1613 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1614 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1615 obj_to_push = cur; 1616 } 1617 bool ok = work_q->push(obj_to_push); 1618 assert(ok, "Should have succeeded"); 1619 cur = next; 1620 n++; 1621 } 1622 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1623 #ifndef PRODUCT 1624 assert(_num_par_pushes >= n, "Too many pops?"); 1625 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 1626 #endif 1627 return true; 1628 } 1629 #undef BUSY 1630 1631 void ParNewGeneration::ref_processor_init() { 1632 if (_ref_processor == NULL) { 1633 // Allocate and initialize a reference processor 1634 _ref_processor = 1635 new ReferenceProcessor(_reserved, // span 1636 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1637 (int) ParallelGCThreads, // mt processing degree 1638 refs_discovery_is_mt(), // mt discovery 1639 (int) ParallelGCThreads, // mt discovery degree 1640 refs_discovery_is_atomic(), // atomic_discovery 1641 NULL, // is_alive_non_header 1642 false); // write barrier for next field updates 1643 } 1644 } 1645 1646 const char* ParNewGeneration::name() const { 1647 return "par new generation"; 1648 }