1 /* 2 * Copyright (c) 2001, 2015, 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.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.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/referencePolicy.hpp" 43 #include "gc/shared/space.hpp" 44 #include "gc/shared/spaceDecorator.hpp" 45 #include "gc/shared/strongRootsScope.hpp" 46 #include "gc/shared/taskqueue.inline.hpp" 47 #include "gc/shared/workgroup.hpp" 48 #include "memory/resourceArea.hpp" 49 #include "oops/objArrayOop.hpp" 50 #include "oops/oop.inline.hpp" 51 #include "runtime/atomic.inline.hpp" 52 #include "runtime/handles.hpp" 53 #include "runtime/handles.inline.hpp" 54 #include "runtime/java.hpp" 55 #include "runtime/thread.inline.hpp" 56 #include "utilities/copy.hpp" 57 #include "utilities/globalDefinitions.hpp" 58 #include "utilities/stack.inline.hpp" 59 60 #ifdef _MSC_VER 61 #pragma warning( push ) 62 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 63 #endif 64 ParScanThreadState::ParScanThreadState(Space* to_space_, 65 ParNewGeneration* young_gen_, 66 Generation* old_gen_, 67 int thread_num_, 68 ObjToScanQueueSet* work_queue_set_, 69 Stack<oop, mtGC>* overflow_stacks_, 70 size_t desired_plab_sz_, 71 ParallelTaskTerminator& term_) : 72 _to_space(to_space_), _old_gen(old_gen_), _young_gen(young_gen_), _thread_num(thread_num_), 73 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), 74 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), 75 _ageTable(false), // false ==> not the global age table, no perf data. 76 _to_space_alloc_buffer(desired_plab_sz_), 77 _to_space_closure(young_gen_, this), _old_gen_closure(young_gen_, this), 78 _to_space_root_closure(young_gen_, this), _old_gen_root_closure(young_gen_, this), 79 _older_gen_closure(young_gen_, this), 80 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, 81 &_to_space_root_closure, young_gen_, &_old_gen_root_closure, 82 work_queue_set_, &term_), 83 _is_alive_closure(young_gen_), _scan_weak_ref_closure(young_gen_, this), 84 _keep_alive_closure(&_scan_weak_ref_closure), 85 _strong_roots_time(0.0), _term_time(0.0) 86 { 87 #if TASKQUEUE_STATS 88 _term_attempts = 0; 89 _overflow_refills = 0; 90 _overflow_refill_objs = 0; 91 #endif // TASKQUEUE_STATS 92 93 _survivor_chunk_array = 94 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); 95 _hash_seed = 17; // Might want to take time-based random value. 96 _start = os::elapsedTime(); 97 _old_gen_closure.set_generation(old_gen_); 98 _old_gen_root_closure.set_generation(old_gen_); 99 } 100 #ifdef _MSC_VER 101 #pragma warning( pop ) 102 #endif 103 104 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, 105 size_t plab_word_size) { 106 ChunkArray* sca = survivor_chunk_array(); 107 if (sca != NULL) { 108 // A non-null SCA implies that we want the PLAB data recorded. 109 sca->record_sample(plab_start, plab_word_size); 110 } 111 } 112 113 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { 114 return new_obj->is_objArray() && 115 arrayOop(new_obj)->length() > ParGCArrayScanChunk && 116 new_obj != old_obj; 117 } 118 119 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { 120 assert(old->is_objArray(), "must be obj array"); 121 assert(old->is_forwarded(), "must be forwarded"); 122 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap."); 123 assert(!old_gen()->is_in(old), "must be in young generation."); 124 125 objArrayOop obj = objArrayOop(old->forwardee()); 126 // Process ParGCArrayScanChunk elements now 127 // and push the remainder back onto queue 128 int start = arrayOop(old)->length(); 129 int end = obj->length(); 130 int remainder = end - start; 131 assert(start <= end, "just checking"); 132 if (remainder > 2 * ParGCArrayScanChunk) { 133 // Test above combines last partial chunk with a full chunk 134 end = start + ParGCArrayScanChunk; 135 arrayOop(old)->set_length(end); 136 // Push remainder. 137 bool ok = work_queue()->push(old); 138 assert(ok, "just popped, push must be okay"); 139 } else { 140 // Restore length so that it can be used if there 141 // is a promotion failure and forwarding pointers 142 // must be removed. 143 arrayOop(old)->set_length(end); 144 } 145 146 // process our set of indices (include header in first chunk) 147 // should make sure end is even (aligned to HeapWord in case of compressed oops) 148 if ((HeapWord *)obj < young_old_boundary()) { 149 // object is in to_space 150 obj->oop_iterate_range(&_to_space_closure, start, end); 151 } else { 152 // object is in old generation 153 obj->oop_iterate_range(&_old_gen_closure, start, end); 154 } 155 } 156 157 158 void ParScanThreadState::trim_queues(int max_size) { 159 ObjToScanQueue* queue = work_queue(); 160 do { 161 while (queue->size() > (juint)max_size) { 162 oop obj_to_scan; 163 if (queue->pop_local(obj_to_scan)) { 164 if ((HeapWord *)obj_to_scan < young_old_boundary()) { 165 if (obj_to_scan->is_objArray() && 166 obj_to_scan->is_forwarded() && 167 obj_to_scan->forwardee() != obj_to_scan) { 168 scan_partial_array_and_push_remainder(obj_to_scan); 169 } else { 170 // object is in to_space 171 obj_to_scan->oop_iterate(&_to_space_closure); 172 } 173 } else { 174 // object is in old generation 175 obj_to_scan->oop_iterate(&_old_gen_closure); 176 } 177 } 178 } 179 // For the case of compressed oops, we have a private, non-shared 180 // overflow stack, so we eagerly drain it so as to more evenly 181 // distribute load early. Note: this may be good to do in 182 // general rather than delay for the final stealing phase. 183 // If applicable, we'll transfer a set of objects over to our 184 // work queue, allowing them to be stolen and draining our 185 // private overflow stack. 186 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); 187 } 188 189 bool ParScanThreadState::take_from_overflow_stack() { 190 assert(ParGCUseLocalOverflow, "Else should not call"); 191 assert(young_gen()->overflow_list() == NULL, "Error"); 192 ObjToScanQueue* queue = work_queue(); 193 Stack<oop, mtGC>* const of_stack = overflow_stack(); 194 const size_t num_overflow_elems = of_stack->size(); 195 const size_t space_available = queue->max_elems() - queue->size(); 196 const size_t num_take_elems = MIN3(space_available / 4, 197 ParGCDesiredObjsFromOverflowList, 198 num_overflow_elems); 199 // Transfer the most recent num_take_elems from the overflow 200 // stack to our work queue. 201 for (size_t i = 0; i != num_take_elems; i++) { 202 oop cur = of_stack->pop(); 203 oop obj_to_push = cur->forwardee(); 204 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap"); 205 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); 206 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap"); 207 if (should_be_partially_scanned(obj_to_push, cur)) { 208 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 209 obj_to_push = cur; 210 } 211 bool ok = queue->push(obj_to_push); 212 assert(ok, "Should have succeeded"); 213 } 214 assert(young_gen()->overflow_list() == NULL, "Error"); 215 return num_take_elems > 0; // was something transferred? 216 } 217 218 void ParScanThreadState::push_on_overflow_stack(oop p) { 219 assert(ParGCUseLocalOverflow, "Else should not call"); 220 overflow_stack()->push(p); 221 assert(young_gen()->overflow_list() == NULL, "Error"); 222 } 223 224 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { 225 226 // Otherwise, if the object is small enough, try to reallocate the 227 // buffer. 228 HeapWord* obj = NULL; 229 if (!_to_space_full) { 230 PLAB* const plab = to_space_alloc_buffer(); 231 Space* const sp = to_space(); 232 if (word_sz * 100 < 233 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), 245 "Invariant"); 246 buf_space = sp->par_allocate(buf_size); 247 free_bytes = sp->free(); 248 } 249 } 250 if (buf_space != NULL) { 251 plab->set_buf(buf_space, buf_size); 252 record_survivor_plab(buf_space, buf_size); 253 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes); 254 // Note that we cannot compare buf_size < word_sz below 255 // because of AlignmentReserve (see PLAB::allocate()). 256 assert(obj != NULL || plab->words_remaining() < word_sz, 257 "Else should have been able to allocate"); 258 // It's conceivable that we may be able to use the 259 // buffer we just grabbed for subsequent small requests 260 // even if not for this one. 261 } else { 262 // We're used up. 263 _to_space_full = true; 264 } 265 266 } else { 267 // Too large; allocate the object individually. 268 obj = sp->par_allocate(word_sz); 269 } 270 } 271 return obj; 272 } 273 274 275 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) { 276 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 277 } 278 279 void ParScanThreadState::print_promotion_failure_size() { 280 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) { 281 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", 282 _thread_num, _promotion_failed_info.first_size()); 283 } 284 } 285 286 class ParScanThreadStateSet: private ResourceArray { 287 public: 288 // Initializes states for the specified number of threads; 289 ParScanThreadStateSet(int num_threads, 290 Space& to_space, 291 ParNewGeneration& gen, 292 Generation& old_gen, 293 ObjToScanQueueSet& queue_set, 294 Stack<oop, mtGC>* overflow_stacks_, 295 size_t desired_plab_sz, 296 ParallelTaskTerminator& term); 297 298 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 299 300 inline ParScanThreadState& thread_state(int i); 301 302 void trace_promotion_failed(const YoungGCTracer* gc_tracer); 303 void reset(uint active_workers, bool promotion_failed); 304 void flush(); 305 306 #if TASKQUEUE_STATS 307 static void 308 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); 309 void print_termination_stats(outputStream* const st = gclog_or_tty); 310 static void 311 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); 312 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); 313 void reset_stats(); 314 #endif // TASKQUEUE_STATS 315 316 private: 317 ParallelTaskTerminator& _term; 318 ParNewGeneration& _gen; 319 Generation& _old_gen; 320 public: 321 bool is_valid(int id) const { return id < length(); } 322 ParallelTaskTerminator* terminator() { return &_term; } 323 }; 324 325 326 ParScanThreadStateSet::ParScanThreadStateSet( 327 int num_threads, Space& to_space, ParNewGeneration& gen, 328 Generation& old_gen, ObjToScanQueueSet& queue_set, 329 Stack<oop, mtGC>* overflow_stacks, 330 size_t desired_plab_sz, ParallelTaskTerminator& term) 331 : ResourceArray(sizeof(ParScanThreadState), num_threads), 332 _gen(gen), _old_gen(old_gen), _term(term) 333 { 334 assert(num_threads > 0, "sanity check!"); 335 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 336 "overflow_stack allocation mismatch"); 337 // Initialize states. 338 for (int i = 0; i < num_threads; ++i) { 339 new ((ParScanThreadState*)_data + i) 340 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, 341 overflow_stacks, desired_plab_sz, term); 342 } 343 } 344 345 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) 346 { 347 assert(i >= 0 && i < length(), "sanity check!"); 348 return ((ParScanThreadState*)_data)[i]; 349 } 350 351 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) { 352 for (int i = 0; i < length(); ++i) { 353 if (thread_state(i).promotion_failed()) { 354 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info()); 355 thread_state(i).promotion_failed_info().reset(); 356 } 357 } 358 } 359 360 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) 361 { 362 _term.reset_for_reuse(active_threads); 363 if (promotion_failed) { 364 for (int i = 0; i < length(); ++i) { 365 thread_state(i).print_promotion_failure_size(); 366 } 367 } 368 } 369 370 #if TASKQUEUE_STATS 371 void 372 ParScanThreadState::reset_stats() 373 { 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 { 382 for (int i = 0; i < length(); ++i) { 383 thread_state(i).reset_stats(); 384 } 385 } 386 387 void 388 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) 389 { 390 st->print_raw_cr("GC Termination Stats"); 391 st->print_raw_cr(" elapsed --strong roots-- " 392 "-------termination-------"); 393 st->print_raw_cr("thr ms ms % " 394 " ms % attempts"); 395 st->print_raw_cr("--- --------- --------- ------ " 396 "--------- ------ --------"); 397 } 398 399 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) 400 { 401 print_termination_stats_hdr(st); 402 403 for (int i = 0; i < length(); ++i) { 404 const ParScanThreadState & pss = thread_state(i); 405 const double elapsed_ms = pss.elapsed_time() * 1000.0; 406 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 407 const double term_ms = pss.term_time() * 1000.0; 408 st->print_cr("%3d %9.2f %9.2f %6.2f " 409 "%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 { 418 st->print_raw_cr("GC Task Stats"); 419 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 420 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 421 } 422 423 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) 424 { 425 print_taskqueue_stats_hdr(st); 426 427 TaskQueueStats totals; 428 for (int i = 0; i < length(); ++i) { 429 const ParScanThreadState & pss = thread_state(i); 430 const TaskQueueStats & stats = pss.taskqueue_stats(); 431 st->print("%3d ", i); stats.print(st); st->cr(); 432 totals += stats; 433 434 if (pss.overflow_refills() > 0) { 435 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 436 SIZE_FORMAT_W(10) " overflow objects", 437 pss.overflow_refills(), pss.overflow_refill_objs()); 438 } 439 } 440 st->print("tot "); totals.print(st); st->cr(); 441 442 DEBUG_ONLY(totals.verify()); 443 } 444 #endif // TASKQUEUE_STATS 445 446 void ParScanThreadStateSet::flush() 447 { 448 // Work in this loop should be kept as lightweight as 449 // possible since this might otherwise become a bottleneck 450 // to scaling. Should we add heavy-weight work into this 451 // loop, consider parallelizing the loop into the worker threads. 452 for (int i = 0; i < length(); ++i) { 453 ParScanThreadState& par_scan_state = thread_state(i); 454 455 // Flush stats related to To-space PLAB activity and 456 // retire the last buffer. 457 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats()); 458 459 // Every thread has its own age table. We need to merge 460 // them all into one. 461 ageTable *local_table = par_scan_state.age_table(); 462 _gen.age_table()->merge(local_table); 463 464 // Inform old gen that we're done. 465 _old_gen.par_promote_alloc_done(i); 466 _old_gen.par_oop_since_save_marks_iterate_done(i); 467 } 468 469 if (UseConcMarkSweepGC) { 470 // We need to call this even when ResizeOldPLAB is disabled 471 // so as to avoid breaking some asserts. While we may be able 472 // to avoid this by reorganizing the code a bit, I am loathe 473 // to do that unless we find cases where ergo leads to bad 474 // performance. 475 CFLS_LAB::compute_desired_plab_size(); 476 } 477 } 478 479 ParScanClosure::ParScanClosure(ParNewGeneration* g, 480 ParScanThreadState* par_scan_state) : 481 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) 482 { 483 _boundary = _g->reserved().end(); 484 } 485 486 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } 487 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } 488 489 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } 490 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } 491 492 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 493 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 494 495 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 496 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 497 498 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 499 ParScanThreadState* par_scan_state) 500 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 501 {} 502 503 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } 504 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } 505 506 #ifdef WIN32 507 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ 508 #endif 509 510 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( 511 ParScanThreadState* par_scan_state_, 512 ParScanWithoutBarrierClosure* to_space_closure_, 513 ParScanWithBarrierClosure* old_gen_closure_, 514 ParRootScanWithoutBarrierClosure* to_space_root_closure_, 515 ParNewGeneration* par_gen_, 516 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, 517 ObjToScanQueueSet* task_queues_, 518 ParallelTaskTerminator* terminator_) : 519 520 _par_scan_state(par_scan_state_), 521 _to_space_closure(to_space_closure_), 522 _old_gen_closure(old_gen_closure_), 523 _to_space_root_closure(to_space_root_closure_), 524 _old_gen_root_closure(old_gen_root_closure_), 525 _par_gen(par_gen_), 526 _task_queues(task_queues_), 527 _terminator(terminator_) 528 {} 529 530 void ParEvacuateFollowersClosure::do_void() { 531 ObjToScanQueue* work_q = par_scan_state()->work_queue(); 532 533 while (true) { 534 535 // Scan to-space and old-gen objs until we run out of both. 536 oop obj_to_scan; 537 par_scan_state()->trim_queues(0); 538 539 // We have no local work, attempt to steal from other threads. 540 541 // attempt to steal work from promoted. 542 if (task_queues()->steal(par_scan_state()->thread_num(), 543 par_scan_state()->hash_seed(), 544 obj_to_scan)) { 545 bool res = work_q->push(obj_to_scan); 546 assert(res, "Empty queue should have room for a push."); 547 548 // if successful, goto Start. 549 continue; 550 551 // try global overflow list. 552 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 553 continue; 554 } 555 556 // Otherwise, offer termination. 557 par_scan_state()->start_term_time(); 558 if (terminator()->offer_termination()) break; 559 par_scan_state()->end_term_time(); 560 } 561 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 562 "Broken overflow list?"); 563 // Finish the last termination pause. 564 par_scan_state()->end_term_time(); 565 } 566 567 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen, Generation* old_gen, 568 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set, 569 StrongRootsScope* strong_roots_scope) : 570 AbstractGangTask("ParNewGeneration collection"), 571 _young_gen(young_gen), _old_gen(old_gen), 572 _young_old_boundary(young_old_boundary), 573 _state_set(state_set), 574 _strong_roots_scope(strong_roots_scope) 575 {} 576 577 void ParNewGenTask::work(uint worker_id) { 578 GenCollectedHeap* gch = GenCollectedHeap::heap(); 579 // Since this is being done in a separate thread, need new resource 580 // and handle marks. 581 ResourceMark rm; 582 HandleMark hm; 583 584 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 585 assert(_state_set->is_valid(worker_id), "Should not have been called"); 586 587 par_scan_state.set_young_old_boundary(_young_old_boundary); 588 589 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), 590 gch->rem_set()->klass_rem_set()); 591 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure, 592 &par_scan_state.to_space_root_closure(), 593 false); 594 595 par_scan_state.start_strong_roots(); 596 gch->gen_process_roots(_strong_roots_scope, 597 GenCollectedHeap::YoungGen, 598 true, // Process younger gens, if any, 599 // as strong roots. 600 GenCollectedHeap::SO_ScavengeCodeCache, 601 GenCollectedHeap::StrongAndWeakRoots, 602 &par_scan_state.to_space_root_closure(), 603 &par_scan_state.older_gen_closure(), 604 &cld_scan_closure); 605 606 par_scan_state.end_strong_roots(); 607 608 // "evacuate followers". 609 par_scan_state.evacuate_followers_closure().do_void(); 610 } 611 612 #ifdef _MSC_VER 613 #pragma warning( push ) 614 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 615 #endif 616 ParNewGeneration:: 617 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size) 618 : DefNewGeneration(rs, initial_byte_size, "PCopy"), 619 _overflow_list(NULL), 620 _is_alive_closure(this), 621 _plab_stats(YoungPLABSize, PLABWeight) 622 { 623 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 624 NOT_PRODUCT(_num_par_pushes = 0;) 625 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 626 guarantee(_task_queues != NULL, "task_queues allocation failure."); 627 628 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { 629 ObjToScanQueue *q = new ObjToScanQueue(); 630 guarantee(q != NULL, "work_queue Allocation failure."); 631 _task_queues->register_queue(i1, q); 632 } 633 634 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) 635 _task_queues->queue(i2)->initialize(); 636 637 _overflow_stacks = NULL; 638 if (ParGCUseLocalOverflow) { 639 640 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal 641 // with ',' 642 typedef Stack<oop, mtGC> GCOopStack; 643 644 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 645 for (size_t i = 0; i < ParallelGCThreads; ++i) { 646 new (_overflow_stacks + i) Stack<oop, mtGC>(); 647 } 648 } 649 650 if (UsePerfData) { 651 EXCEPTION_MARK; 652 ResourceMark rm; 653 654 const char* cname = 655 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 656 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 657 ParallelGCThreads, CHECK); 658 } 659 } 660 #ifdef _MSC_VER 661 #pragma warning( pop ) 662 #endif 663 664 // ParNewGeneration:: 665 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 666 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 667 668 template <class T> 669 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 670 #ifdef ASSERT 671 { 672 assert(!oopDesc::is_null(*p), "expected non-null ref"); 673 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 674 // We never expect to see a null reference being processed 675 // as a weak reference. 676 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 677 } 678 #endif // ASSERT 679 680 _par_cl->do_oop_nv(p); 681 682 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 683 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 684 _rs->write_ref_field_gc_par(p, obj); 685 } 686 } 687 688 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 689 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 690 691 // ParNewGeneration:: 692 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 693 DefNewGeneration::KeepAliveClosure(cl) {} 694 695 template <class T> 696 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 697 #ifdef ASSERT 698 { 699 assert(!oopDesc::is_null(*p), "expected non-null ref"); 700 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 701 // We never expect to see a null reference being processed 702 // as a weak reference. 703 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 704 } 705 #endif // ASSERT 706 707 _cl->do_oop_nv(p); 708 709 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 710 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 711 _rs->write_ref_field_gc_par(p, obj); 712 } 713 } 714 715 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 716 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 717 718 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 719 T heap_oop = oopDesc::load_heap_oop(p); 720 if (!oopDesc::is_null(heap_oop)) { 721 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 722 if ((HeapWord*)obj < _boundary) { 723 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 724 oop new_obj = obj->is_forwarded() 725 ? obj->forwardee() 726 : _g->DefNewGeneration::copy_to_survivor_space(obj); 727 oopDesc::encode_store_heap_oop_not_null(p, new_obj); 728 } 729 if (_gc_barrier) { 730 // If p points to a younger generation, mark the card. 731 if ((HeapWord*)obj < _gen_boundary) { 732 _rs->write_ref_field_gc_par(p, obj); 733 } 734 } 735 } 736 } 737 738 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 739 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 740 741 class ParNewRefProcTaskProxy: public AbstractGangTask { 742 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 743 public: 744 ParNewRefProcTaskProxy(ProcessTask& task, 745 ParNewGeneration& gen, 746 Generation& old_gen, 747 HeapWord* young_old_boundary, 748 ParScanThreadStateSet& state_set); 749 750 private: 751 virtual void work(uint worker_id); 752 private: 753 ParNewGeneration& _young_gen; 754 ProcessTask& _task; 755 Generation& _old_gen; 756 HeapWord* _young_old_boundary; 757 ParScanThreadStateSet& _state_set; 758 }; 759 760 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task, 761 ParNewGeneration& young_gen, 762 Generation& old_gen, 763 HeapWord* young_old_boundary, 764 ParScanThreadStateSet& state_set) 765 : AbstractGangTask("ParNewGeneration parallel reference processing"), 766 _young_gen(young_gen), 767 _task(task), 768 _old_gen(old_gen), 769 _young_old_boundary(young_old_boundary), 770 _state_set(state_set) 771 { 772 } 773 774 void ParNewRefProcTaskProxy::work(uint worker_id) 775 { 776 ResourceMark rm; 777 HandleMark hm; 778 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 779 par_scan_state.set_young_old_boundary(_young_old_boundary); 780 _task.work(worker_id, par_scan_state.is_alive_closure(), 781 par_scan_state.keep_alive_closure(), 782 par_scan_state.evacuate_followers_closure()); 783 } 784 785 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { 786 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 787 EnqueueTask& _task; 788 789 public: 790 ParNewRefEnqueueTaskProxy(EnqueueTask& task) 791 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), 792 _task(task) 793 { } 794 795 virtual void work(uint worker_id) 796 { 797 _task.work(worker_id); 798 } 799 }; 800 801 802 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) 803 { 804 GenCollectedHeap* gch = GenCollectedHeap::heap(); 805 WorkGang* workers = gch->workers(); 806 assert(workers != NULL, "Need parallel worker threads."); 807 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed()); 808 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen, 809 _young_gen.reserved().end(), _state_set); 810 workers->run_task(&rp_task); 811 _state_set.reset(0 /* bad value in debug if not reset */, 812 _young_gen.promotion_failed()); 813 } 814 815 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) 816 { 817 GenCollectedHeap* gch = GenCollectedHeap::heap(); 818 WorkGang* workers = gch->workers(); 819 assert(workers != NULL, "Need parallel worker threads."); 820 ParNewRefEnqueueTaskProxy enq_task(task); 821 workers->run_task(&enq_task); 822 } 823 824 void ParNewRefProcTaskExecutor::set_single_threaded_mode() 825 { 826 _state_set.flush(); 827 GenCollectedHeap* gch = GenCollectedHeap::heap(); 828 gch->save_marks(); 829 } 830 831 ScanClosureWithParBarrier:: 832 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 833 ScanClosure(g, gc_barrier) {} 834 835 EvacuateFollowersClosureGeneral:: 836 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, 837 OopsInGenClosure* cur, 838 OopsInGenClosure* older) : 839 _gch(gch), 840 _scan_cur_or_nonheap(cur), _scan_older(older) 841 {} 842 843 void EvacuateFollowersClosureGeneral::do_void() { 844 do { 845 // Beware: this call will lead to closure applications via virtual 846 // calls. 847 _gch->oop_since_save_marks_iterate(GenCollectedHeap::YoungGen, 848 _scan_cur_or_nonheap, 849 _scan_older); 850 } while (!_gch->no_allocs_since_save_marks()); 851 } 852 853 854 // A Generation that does parallel young-gen collection. 855 856 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) { 857 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 858 _promo_failure_scan_stack.clear(true); // Clear cached segments. 859 860 remove_forwarding_pointers(); 861 if (PrintGCDetails) { 862 gclog_or_tty->print(" (promotion failed)"); 863 } 864 // All the spaces are in play for mark-sweep. 865 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 866 from()->set_next_compaction_space(to()); 867 gch->set_incremental_collection_failed(); 868 // Inform the next generation that a promotion failure occurred. 869 _old_gen->promotion_failure_occurred(); 870 871 // Trace promotion failure in the parallel GC threads 872 thread_state_set.trace_promotion_failed(gc_tracer()); 873 // Single threaded code may have reported promotion failure to the global state 874 if (_promotion_failed_info.has_failed()) { 875 _gc_tracer.report_promotion_failed(_promotion_failed_info); 876 } 877 // Reset the PromotionFailureALot counters. 878 NOT_PRODUCT(gch->reset_promotion_should_fail();) 879 } 880 881 void ParNewGeneration::collect(bool full, 882 bool clear_all_soft_refs, 883 size_t size, 884 bool is_tlab) { 885 assert(full || size > 0, "otherwise we don't want to collect"); 886 887 GenCollectedHeap* gch = GenCollectedHeap::heap(); 888 889 _gc_timer->register_gc_start(); 890 891 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); 892 WorkGang* workers = gch->workers(); 893 assert(workers != NULL, "Need workgang for parallel work"); 894 uint active_workers = 895 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), 896 workers->active_workers(), 897 Threads::number_of_non_daemon_threads()); 898 workers->set_active_workers(active_workers); 899 _old_gen = gch->old_gen(); 900 901 // If the next generation is too full to accommodate worst-case promotion 902 // from this generation, pass on collection; let the next generation 903 // do it. 904 if (!collection_attempt_is_safe()) { 905 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 906 return; 907 } 908 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 909 910 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 911 gch->trace_heap_before_gc(gc_tracer()); 912 913 init_assuming_no_promotion_failure(); 914 915 if (UseAdaptiveSizePolicy) { 916 set_survivor_overflow(false); 917 size_policy->minor_collection_begin(); 918 } 919 920 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id()); 921 // Capture heap used before collection (for printing). 922 size_t gch_prev_used = gch->used(); 923 924 age_table()->clear(); 925 to()->clear(SpaceDecorator::Mangle); 926 927 gch->save_marks(); 928 929 // Set the correct parallelism (number of queues) in the reference processor 930 ref_processor()->set_active_mt_degree(active_workers); 931 932 // Always set the terminator for the active number of workers 933 // because only those workers go through the termination protocol. 934 ParallelTaskTerminator _term(active_workers, task_queues()); 935 ParScanThreadStateSet thread_state_set(active_workers, 936 *to(), *this, *_old_gen, *task_queues(), 937 _overflow_stacks, desired_plab_sz(), _term); 938 939 thread_state_set.reset(active_workers, promotion_failed()); 940 941 { 942 StrongRootsScope srs(active_workers); 943 944 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs); 945 gch->rem_set()->prepare_for_younger_refs_iterate(true); 946 // It turns out that even when we're using 1 thread, doing the work in a 947 // separate thread causes wide variance in run times. We can't help this 948 // in the multi-threaded case, but we special-case n=1 here to get 949 // repeatable measurements of the 1-thread overhead of the parallel code. 950 if (active_workers > 1) { 951 workers->run_task(&tsk); 952 } else { 953 tsk.work(0); 954 } 955 } 956 957 thread_state_set.reset(0 /* Bad value in debug if not reset */, 958 promotion_failed()); 959 960 // Trace and reset failed promotion info. 961 if (promotion_failed()) { 962 thread_state_set.trace_promotion_failed(gc_tracer()); 963 } 964 965 // Process (weak) reference objects found during scavenge. 966 ReferenceProcessor* rp = ref_processor(); 967 IsAliveClosure is_alive(this); 968 ScanWeakRefClosure scan_weak_ref(this); 969 KeepAliveClosure keep_alive(&scan_weak_ref); 970 ScanClosure scan_without_gc_barrier(this, false); 971 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 972 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 973 EvacuateFollowersClosureGeneral evacuate_followers(gch, 974 &scan_without_gc_barrier, &scan_with_gc_barrier); 975 rp->setup_policy(clear_all_soft_refs); 976 // Can the mt_degree be set later (at run_task() time would be best)? 977 rp->set_active_mt_degree(active_workers); 978 ReferenceProcessorStats stats; 979 if (rp->processing_is_mt()) { 980 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 981 stats = rp->process_discovered_references(&is_alive, &keep_alive, 982 &evacuate_followers, &task_executor, 983 _gc_timer, _gc_tracer.gc_id()); 984 } else { 985 thread_state_set.flush(); 986 gch->save_marks(); 987 stats = rp->process_discovered_references(&is_alive, &keep_alive, 988 &evacuate_followers, NULL, 989 _gc_timer, _gc_tracer.gc_id()); 990 } 991 _gc_tracer.report_gc_reference_stats(stats); 992 if (!promotion_failed()) { 993 // Swap the survivor spaces. 994 eden()->clear(SpaceDecorator::Mangle); 995 from()->clear(SpaceDecorator::Mangle); 996 if (ZapUnusedHeapArea) { 997 // This is now done here because of the piece-meal mangling which 998 // can check for valid mangling at intermediate points in the 999 // collection(s). When a minor collection fails to collect 1000 // sufficient space resizing of the young generation can occur 1001 // an redistribute the spaces in the young generation. Mangle 1002 // here so that unzapped regions don't get distributed to 1003 // other spaces. 1004 to()->mangle_unused_area(); 1005 } 1006 swap_spaces(); 1007 1008 // A successful scavenge should restart the GC time limit count which is 1009 // for full GC's. 1010 size_policy->reset_gc_overhead_limit_count(); 1011 1012 assert(to()->is_empty(), "to space should be empty now"); 1013 1014 adjust_desired_tenuring_threshold(); 1015 } else { 1016 handle_promotion_failed(gch, thread_state_set); 1017 } 1018 // set new iteration safe limit for the survivor spaces 1019 from()->set_concurrent_iteration_safe_limit(from()->top()); 1020 to()->set_concurrent_iteration_safe_limit(to()->top()); 1021 1022 if (ResizePLAB) { 1023 plab_stats()->adjust_desired_plab_sz(); 1024 } 1025 1026 if (PrintGC && !PrintGCDetails) { 1027 gch->print_heap_change(gch_prev_used); 1028 } 1029 1030 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats()); 1031 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats()); 1032 1033 if (UseAdaptiveSizePolicy) { 1034 size_policy->minor_collection_end(gch->gc_cause()); 1035 size_policy->avg_survived()->sample(from()->used()); 1036 } 1037 1038 // We need to use a monotonically non-decreasing time in ms 1039 // or we will see time-warp warnings and os::javaTimeMillis() 1040 // does not guarantee monotonicity. 1041 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1042 update_time_of_last_gc(now); 1043 1044 rp->set_enqueuing_is_done(true); 1045 if (rp->processing_is_mt()) { 1046 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 1047 rp->enqueue_discovered_references(&task_executor); 1048 } else { 1049 rp->enqueue_discovered_references(NULL); 1050 } 1051 rp->verify_no_references_recorded(); 1052 1053 gch->trace_heap_after_gc(gc_tracer()); 1054 _gc_tracer.report_tenuring_threshold(tenuring_threshold()); 1055 1056 _gc_timer->register_gc_end(); 1057 1058 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1059 } 1060 1061 size_t ParNewGeneration::desired_plab_sz() { 1062 return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers()); 1063 } 1064 1065 static int sum; 1066 void ParNewGeneration::waste_some_time() { 1067 for (int i = 0; i < 100; i++) { 1068 sum += i; 1069 } 1070 } 1071 1072 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); 1073 1074 // Because of concurrency, there are times where an object for which 1075 // "is_forwarded()" is true contains an "interim" forwarding pointer 1076 // value. Such a value will soon be overwritten with a real value. 1077 // This method requires "obj" to have a forwarding pointer, and waits, if 1078 // necessary for a real one to be inserted, and returns it. 1079 1080 oop ParNewGeneration::real_forwardee(oop obj) { 1081 oop forward_ptr = obj->forwardee(); 1082 if (forward_ptr != ClaimedForwardPtr) { 1083 return forward_ptr; 1084 } else { 1085 return real_forwardee_slow(obj); 1086 } 1087 } 1088 1089 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1090 // Spin-read if it is claimed but not yet written by another thread. 1091 oop forward_ptr = obj->forwardee(); 1092 while (forward_ptr == ClaimedForwardPtr) { 1093 waste_some_time(); 1094 assert(obj->is_forwarded(), "precondition"); 1095 forward_ptr = obj->forwardee(); 1096 } 1097 return forward_ptr; 1098 } 1099 1100 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { 1101 if (m->must_be_preserved_for_promotion_failure(obj)) { 1102 // We should really have separate per-worker stacks, rather 1103 // than use locking of a common pair of stacks. 1104 MutexLocker ml(ParGCRareEvent_lock); 1105 preserve_mark(obj, m); 1106 } 1107 } 1108 1109 // Multiple GC threads may try to promote an object. If the object 1110 // is successfully promoted, a forwarding pointer will be installed in 1111 // the object in the young generation. This method claims the right 1112 // to install the forwarding pointer before it copies the object, 1113 // thus avoiding the need to undo the copy as in 1114 // copy_to_survivor_space_avoiding_with_undo. 1115 1116 oop ParNewGeneration::copy_to_survivor_space( 1117 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { 1118 // In the sequential version, this assert also says that the object is 1119 // not forwarded. That might not be the case here. It is the case that 1120 // the caller observed it to be not forwarded at some time in the past. 1121 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1122 1123 // The sequential code read "old->age()" below. That doesn't work here, 1124 // since the age is in the mark word, and that might be overwritten with 1125 // a forwarding pointer by a parallel thread. So we must save the mark 1126 // word in a local and then analyze it. 1127 oopDesc dummyOld; 1128 dummyOld.set_mark(m); 1129 assert(!dummyOld.is_forwarded(), 1130 "should not be called with forwarding pointer mark word."); 1131 1132 oop new_obj = NULL; 1133 oop forward_ptr; 1134 1135 // Try allocating obj in to-space (unless too old) 1136 if (dummyOld.age() < tenuring_threshold()) { 1137 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1138 if (new_obj == NULL) { 1139 set_survivor_overflow(true); 1140 } 1141 } 1142 1143 if (new_obj == NULL) { 1144 // Either to-space is full or we decided to promote 1145 // try allocating obj tenured 1146 1147 // Attempt to install a null forwarding pointer (atomically), 1148 // to claim the right to install the real forwarding pointer. 1149 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); 1150 if (forward_ptr != NULL) { 1151 // someone else beat us to it. 1152 return real_forwardee(old); 1153 } 1154 1155 if (!_promotion_failed) { 1156 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), 1157 old, m, sz); 1158 } 1159 1160 if (new_obj == NULL) { 1161 // promotion failed, forward to self 1162 _promotion_failed = true; 1163 new_obj = old; 1164 1165 preserve_mark_if_necessary(old, m); 1166 par_scan_state->register_promotion_failure(sz); 1167 } 1168 1169 old->forward_to(new_obj); 1170 forward_ptr = NULL; 1171 } else { 1172 // Is in to-space; do copying ourselves. 1173 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1174 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value."); 1175 forward_ptr = old->forward_to_atomic(new_obj); 1176 // Restore the mark word copied above. 1177 new_obj->set_mark(m); 1178 // Increment age if obj still in new generation 1179 new_obj->incr_age(); 1180 par_scan_state->age_table()->add(new_obj, sz); 1181 } 1182 assert(new_obj != NULL, "just checking"); 1183 1184 #ifndef PRODUCT 1185 // This code must come after the CAS test, or it will print incorrect 1186 // information. 1187 if (TraceScavenge) { 1188 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1189 is_in_reserved(new_obj) ? "copying" : "tenuring", 1190 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size()); 1191 } 1192 #endif 1193 1194 if (forward_ptr == NULL) { 1195 oop obj_to_push = new_obj; 1196 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1197 // Length field used as index of next element to be scanned. 1198 // Real length can be obtained from real_forwardee() 1199 arrayOop(old)->set_length(0); 1200 obj_to_push = old; 1201 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1202 "push forwarded object"); 1203 } 1204 // Push it on one of the queues of to-be-scanned objects. 1205 bool simulate_overflow = false; 1206 NOT_PRODUCT( 1207 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1208 // simulate a stack overflow 1209 simulate_overflow = true; 1210 } 1211 ) 1212 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1213 // Add stats for overflow pushes. 1214 if (Verbose && PrintGCDetails) { 1215 gclog_or_tty->print("queue overflow!\n"); 1216 } 1217 push_on_overflow_list(old, par_scan_state); 1218 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1219 } 1220 1221 return new_obj; 1222 } 1223 1224 // Oops. Someone beat us to it. Undo the allocation. Where did we 1225 // allocate it? 1226 if (is_in_reserved(new_obj)) { 1227 // Must be in to_space. 1228 assert(to()->is_in_reserved(new_obj), "Checking"); 1229 if (forward_ptr == ClaimedForwardPtr) { 1230 // Wait to get the real forwarding pointer value. 1231 forward_ptr = real_forwardee(old); 1232 } 1233 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1234 } 1235 1236 return forward_ptr; 1237 } 1238 1239 #ifndef PRODUCT 1240 // It's OK to call this multi-threaded; the worst thing 1241 // that can happen is that we'll get a bunch of closely 1242 // spaced simulated overflows, but that's OK, in fact 1243 // probably good as it would exercise the overflow code 1244 // under contention. 1245 bool ParNewGeneration::should_simulate_overflow() { 1246 if (_overflow_counter-- <= 0) { // just being defensive 1247 _overflow_counter = ParGCWorkQueueOverflowInterval; 1248 return true; 1249 } else { 1250 return false; 1251 } 1252 } 1253 #endif 1254 1255 // In case we are using compressed oops, we need to be careful. 1256 // If the object being pushed is an object array, then its length 1257 // field keeps track of the "grey boundary" at which the next 1258 // incremental scan will be done (see ParGCArrayScanChunk). 1259 // When using compressed oops, this length field is kept in the 1260 // lower 32 bits of the erstwhile klass word and cannot be used 1261 // for the overflow chaining pointer (OCP below). As such the OCP 1262 // would itself need to be compressed into the top 32-bits in this 1263 // case. Unfortunately, see below, in the event that we have a 1264 // promotion failure, the node to be pushed on the list can be 1265 // outside of the Java heap, so the heap-based pointer compression 1266 // would not work (we would have potential aliasing between C-heap 1267 // and Java-heap pointers). For this reason, when using compressed 1268 // oops, we simply use a worker-thread-local, non-shared overflow 1269 // list in the form of a growable array, with a slightly different 1270 // overflow stack draining strategy. If/when we start using fat 1271 // stacks here, we can go back to using (fat) pointer chains 1272 // (although some performance comparisons would be useful since 1273 // single global lists have their own performance disadvantages 1274 // as we were made painfully aware not long ago, see 6786503). 1275 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 1276 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1277 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1278 if (ParGCUseLocalOverflow) { 1279 // In the case of compressed oops, we use a private, not-shared 1280 // overflow stack. 1281 par_scan_state->push_on_overflow_stack(from_space_obj); 1282 } else { 1283 assert(!UseCompressedOops, "Error"); 1284 // if the object has been forwarded to itself, then we cannot 1285 // use the klass pointer for the linked list. Instead we have 1286 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1287 // XXX This is horribly inefficient when a promotion failure occurs 1288 // and should be fixed. XXX FIX ME !!! 1289 #ifndef PRODUCT 1290 Atomic::inc_ptr(&_num_par_pushes); 1291 assert(_num_par_pushes > 0, "Tautology"); 1292 #endif 1293 if (from_space_obj->forwardee() == from_space_obj) { 1294 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); 1295 listhead->forward_to(from_space_obj); 1296 from_space_obj = listhead; 1297 } 1298 oop observed_overflow_list = _overflow_list; 1299 oop cur_overflow_list; 1300 do { 1301 cur_overflow_list = observed_overflow_list; 1302 if (cur_overflow_list != BUSY) { 1303 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1304 } else { 1305 from_space_obj->set_klass_to_list_ptr(NULL); 1306 } 1307 observed_overflow_list = 1308 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); 1309 } while (cur_overflow_list != observed_overflow_list); 1310 } 1311 } 1312 1313 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1314 bool res; 1315 1316 if (ParGCUseLocalOverflow) { 1317 res = par_scan_state->take_from_overflow_stack(); 1318 } else { 1319 assert(!UseCompressedOops, "Error"); 1320 res = take_from_overflow_list_work(par_scan_state); 1321 } 1322 return res; 1323 } 1324 1325 1326 // *NOTE*: The overflow list manipulation code here and 1327 // in CMSCollector:: are very similar in shape, 1328 // except that in the CMS case we thread the objects 1329 // directly into the list via their mark word, and do 1330 // not need to deal with special cases below related 1331 // to chunking of object arrays and promotion failure 1332 // handling. 1333 // CR 6797058 has been filed to attempt consolidation of 1334 // the common code. 1335 // Because of the common code, if you make any changes in 1336 // the code below, please check the CMS version to see if 1337 // similar changes might be needed. 1338 // See CMSCollector::par_take_from_overflow_list() for 1339 // more extensive documentation comments. 1340 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1341 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1342 // How many to take? 1343 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1344 (size_t)ParGCDesiredObjsFromOverflowList); 1345 1346 assert(!UseCompressedOops, "Error"); 1347 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1348 if (_overflow_list == NULL) return false; 1349 1350 // Otherwise, there was something there; try claiming the list. 1351 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1352 // Trim off a prefix of at most objsFromOverflow items 1353 Thread* tid = Thread::current(); 1354 size_t spin_count = ParallelGCThreads; 1355 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1356 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1357 // someone grabbed it before we did ... 1358 // ... we spin for a short while... 1359 os::sleep(tid, sleep_time_millis, false); 1360 if (_overflow_list == NULL) { 1361 // nothing left to take 1362 return false; 1363 } else if (_overflow_list != BUSY) { 1364 // try and grab the prefix 1365 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1366 } 1367 } 1368 if (prefix == NULL || prefix == BUSY) { 1369 // Nothing to take or waited long enough 1370 if (prefix == NULL) { 1371 // Write back the NULL in case we overwrote it with BUSY above 1372 // and it is still the same value. 1373 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1374 } 1375 return false; 1376 } 1377 assert(prefix != NULL && prefix != BUSY, "Error"); 1378 size_t i = 1; 1379 oop cur = prefix; 1380 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { 1381 i++; cur = cur->list_ptr_from_klass(); 1382 } 1383 1384 // Reattach remaining (suffix) to overflow list 1385 if (cur->klass_or_null() == NULL) { 1386 // Write back the NULL in lieu of the BUSY we wrote 1387 // above and it is still the same value. 1388 if (_overflow_list == BUSY) { 1389 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1390 } 1391 } else { 1392 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); 1393 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list 1394 cur->set_klass_to_list_ptr(NULL); // break off suffix 1395 // It's possible that the list is still in the empty(busy) state 1396 // we left it in a short while ago; in that case we may be 1397 // able to place back the suffix. 1398 oop observed_overflow_list = _overflow_list; 1399 oop cur_overflow_list = observed_overflow_list; 1400 bool attached = false; 1401 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1402 observed_overflow_list = 1403 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1404 if (cur_overflow_list == observed_overflow_list) { 1405 attached = true; 1406 break; 1407 } else cur_overflow_list = observed_overflow_list; 1408 } 1409 if (!attached) { 1410 // Too bad, someone else got in in between; we'll need to do a splice. 1411 // Find the last item of suffix list 1412 oop last = suffix; 1413 while (last->klass_or_null() != NULL) { 1414 last = last->list_ptr_from_klass(); 1415 } 1416 // Atomically prepend suffix to current overflow list 1417 observed_overflow_list = _overflow_list; 1418 do { 1419 cur_overflow_list = observed_overflow_list; 1420 if (cur_overflow_list != BUSY) { 1421 // Do the splice ... 1422 last->set_klass_to_list_ptr(cur_overflow_list); 1423 } else { // cur_overflow_list == BUSY 1424 last->set_klass_to_list_ptr(NULL); 1425 } 1426 observed_overflow_list = 1427 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1428 } while (cur_overflow_list != observed_overflow_list); 1429 } 1430 } 1431 1432 // Push objects on prefix list onto this thread's work queue 1433 assert(prefix != NULL && prefix != BUSY, "program logic"); 1434 cur = prefix; 1435 ssize_t n = 0; 1436 while (cur != NULL) { 1437 oop obj_to_push = cur->forwardee(); 1438 oop next = cur->list_ptr_from_klass(); 1439 cur->set_klass(obj_to_push->klass()); 1440 // This may be an array object that is self-forwarded. In that case, the list pointer 1441 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1442 if (!is_in_reserved(cur)) { 1443 // This can become a scaling bottleneck when there is work queue overflow coincident 1444 // with promotion failure. 1445 oopDesc* f = cur; 1446 FREE_C_HEAP_ARRAY(oopDesc, f); 1447 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1448 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1449 obj_to_push = cur; 1450 } 1451 bool ok = work_q->push(obj_to_push); 1452 assert(ok, "Should have succeeded"); 1453 cur = next; 1454 n++; 1455 } 1456 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1457 #ifndef PRODUCT 1458 assert(_num_par_pushes >= n, "Too many pops?"); 1459 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 1460 #endif 1461 return true; 1462 } 1463 #undef BUSY 1464 1465 void ParNewGeneration::ref_processor_init() { 1466 if (_ref_processor == NULL) { 1467 // Allocate and initialize a reference processor 1468 _ref_processor = 1469 new ReferenceProcessor(_reserved, // span 1470 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1471 ParallelGCThreads, // mt processing degree 1472 refs_discovery_is_mt(), // mt discovery 1473 ParallelGCThreads, // mt discovery degree 1474 refs_discovery_is_atomic(), // atomic_discovery 1475 NULL); // is_alive_non_header 1476 } 1477 } 1478 1479 const char* ParNewGeneration::name() const { 1480 return "par new generation"; 1481 }