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.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_word_size(buf_size); 252 plab->set_buf(buf_space); 253 record_survivor_plab(buf_space, buf_size); 254 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes); 255 // Note that we cannot compare buf_size < word_sz below 256 // because of AlignmentReserve (see PLAB::allocate()). 257 assert(obj != NULL || plab->words_remaining() < word_sz, 258 "Else should have been able to allocate"); 259 // It's conceivable that we may be able to use the 260 // buffer we just grabbed for subsequent small requests 261 // even if not for this one. 262 } else { 263 // We're used up. 264 _to_space_full = true; 265 } 266 267 } else { 268 // Too large; allocate the object individually. 269 obj = sp->par_allocate(word_sz); 270 } 271 } 272 return obj; 273 } 274 275 276 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) { 277 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 278 } 279 280 void ParScanThreadState::print_promotion_failure_size() { 281 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) { 282 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", 283 _thread_num, _promotion_failed_info.first_size()); 284 } 285 } 286 287 class ParScanThreadStateSet: private ResourceArray { 288 public: 289 // Initializes states for the specified number of threads; 290 ParScanThreadStateSet(int num_threads, 291 Space& to_space, 292 ParNewGeneration& gen, 293 Generation& old_gen, 294 ObjToScanQueueSet& queue_set, 295 Stack<oop, mtGC>* overflow_stacks_, 296 size_t desired_plab_sz, 297 ParallelTaskTerminator& term); 298 299 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 300 301 inline ParScanThreadState& thread_state(int i); 302 303 void trace_promotion_failed(const YoungGCTracer* gc_tracer); 304 void reset(uint active_workers, bool promotion_failed); 305 void flush(); 306 307 #if TASKQUEUE_STATS 308 static void 309 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); 310 void print_termination_stats(outputStream* const st = gclog_or_tty); 311 static void 312 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); 313 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); 314 void reset_stats(); 315 #endif // TASKQUEUE_STATS 316 317 private: 318 ParallelTaskTerminator& _term; 319 ParNewGeneration& _gen; 320 Generation& _old_gen; 321 public: 322 bool is_valid(int id) const { return id < length(); } 323 ParallelTaskTerminator* terminator() { return &_term; } 324 }; 325 326 327 ParScanThreadStateSet::ParScanThreadStateSet( 328 int num_threads, Space& to_space, ParNewGeneration& gen, 329 Generation& old_gen, ObjToScanQueueSet& queue_set, 330 Stack<oop, mtGC>* overflow_stacks, 331 size_t desired_plab_sz, ParallelTaskTerminator& term) 332 : ResourceArray(sizeof(ParScanThreadState), num_threads), 333 _gen(gen), _old_gen(old_gen), _term(term) 334 { 335 assert(num_threads > 0, "sanity check!"); 336 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 337 "overflow_stack allocation mismatch"); 338 // Initialize states. 339 for (int i = 0; i < num_threads; ++i) { 340 new ((ParScanThreadState*)_data + i) 341 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, 342 overflow_stacks, desired_plab_sz, term); 343 } 344 } 345 346 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) 347 { 348 assert(i >= 0 && i < length(), "sanity check!"); 349 return ((ParScanThreadState*)_data)[i]; 350 } 351 352 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) { 353 for (int i = 0; i < length(); ++i) { 354 if (thread_state(i).promotion_failed()) { 355 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info()); 356 thread_state(i).promotion_failed_info().reset(); 357 } 358 } 359 } 360 361 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) 362 { 363 _term.reset_for_reuse(active_threads); 364 if (promotion_failed) { 365 for (int i = 0; i < length(); ++i) { 366 thread_state(i).print_promotion_failure_size(); 367 } 368 } 369 } 370 371 #if TASKQUEUE_STATS 372 void 373 ParScanThreadState::reset_stats() 374 { 375 taskqueue_stats().reset(); 376 _term_attempts = 0; 377 _overflow_refills = 0; 378 _overflow_refill_objs = 0; 379 } 380 381 void ParScanThreadStateSet::reset_stats() 382 { 383 for (int i = 0; i < length(); ++i) { 384 thread_state(i).reset_stats(); 385 } 386 } 387 388 void 389 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) 390 { 391 st->print_raw_cr("GC Termination Stats"); 392 st->print_raw_cr(" elapsed --strong roots-- " 393 "-------termination-------"); 394 st->print_raw_cr("thr ms ms % " 395 " ms % attempts"); 396 st->print_raw_cr("--- --------- --------- ------ " 397 "--------- ------ --------"); 398 } 399 400 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) 401 { 402 print_termination_stats_hdr(st); 403 404 for (int i = 0; i < length(); ++i) { 405 const ParScanThreadState & pss = thread_state(i); 406 const double elapsed_ms = pss.elapsed_time() * 1000.0; 407 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 408 const double term_ms = pss.term_time() * 1000.0; 409 st->print_cr("%3d %9.2f %9.2f %6.2f " 410 "%9.2f %6.2f " SIZE_FORMAT_W(8), 411 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 412 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); 413 } 414 } 415 416 // Print stats related to work queue activity. 417 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) 418 { 419 st->print_raw_cr("GC Task Stats"); 420 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 421 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 422 } 423 424 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) 425 { 426 print_taskqueue_stats_hdr(st); 427 428 TaskQueueStats totals; 429 for (int i = 0; i < length(); ++i) { 430 const ParScanThreadState & pss = thread_state(i); 431 const TaskQueueStats & stats = pss.taskqueue_stats(); 432 st->print("%3d ", i); stats.print(st); st->cr(); 433 totals += stats; 434 435 if (pss.overflow_refills() > 0) { 436 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 437 SIZE_FORMAT_W(10) " overflow objects", 438 pss.overflow_refills(), pss.overflow_refill_objs()); 439 } 440 } 441 st->print("tot "); totals.print(st); st->cr(); 442 443 DEBUG_ONLY(totals.verify()); 444 } 445 #endif // TASKQUEUE_STATS 446 447 void ParScanThreadStateSet::flush() 448 { 449 // Work in this loop should be kept as lightweight as 450 // possible since this might otherwise become a bottleneck 451 // to scaling. Should we add heavy-weight work into this 452 // loop, consider parallelizing the loop into the worker threads. 453 for (int i = 0; i < length(); ++i) { 454 ParScanThreadState& par_scan_state = thread_state(i); 455 456 // Flush stats related to To-space PLAB activity and 457 // retire the last buffer. 458 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats()); 459 460 // Every thread has its own age table. We need to merge 461 // them all into one. 462 ageTable *local_table = par_scan_state.age_table(); 463 _gen.age_table()->merge(local_table); 464 465 // Inform old gen that we're done. 466 _old_gen.par_promote_alloc_done(i); 467 _old_gen.par_oop_since_save_marks_iterate_done(i); 468 } 469 470 if (UseConcMarkSweepGC) { 471 // We need to call this even when ResizeOldPLAB is disabled 472 // so as to avoid breaking some asserts. While we may be able 473 // to avoid this by reorganizing the code a bit, I am loathe 474 // to do that unless we find cases where ergo leads to bad 475 // performance. 476 CFLS_LAB::compute_desired_plab_size(); 477 } 478 } 479 480 ParScanClosure::ParScanClosure(ParNewGeneration* g, 481 ParScanThreadState* par_scan_state) : 482 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) 483 { 484 _boundary = _g->reserved().end(); 485 } 486 487 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } 488 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } 489 490 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } 491 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } 492 493 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 494 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 495 496 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 497 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 498 499 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 500 ParScanThreadState* par_scan_state) 501 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 502 {} 503 504 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } 505 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } 506 507 #ifdef WIN32 508 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ 509 #endif 510 511 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( 512 ParScanThreadState* par_scan_state_, 513 ParScanWithoutBarrierClosure* to_space_closure_, 514 ParScanWithBarrierClosure* old_gen_closure_, 515 ParRootScanWithoutBarrierClosure* to_space_root_closure_, 516 ParNewGeneration* par_gen_, 517 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, 518 ObjToScanQueueSet* task_queues_, 519 ParallelTaskTerminator* terminator_) : 520 521 _par_scan_state(par_scan_state_), 522 _to_space_closure(to_space_closure_), 523 _old_gen_closure(old_gen_closure_), 524 _to_space_root_closure(to_space_root_closure_), 525 _old_gen_root_closure(old_gen_root_closure_), 526 _par_gen(par_gen_), 527 _task_queues(task_queues_), 528 _terminator(terminator_) 529 {} 530 531 void ParEvacuateFollowersClosure::do_void() { 532 ObjToScanQueue* work_q = par_scan_state()->work_queue(); 533 534 while (true) { 535 536 // Scan to-space and old-gen objs until we run out of both. 537 oop obj_to_scan; 538 par_scan_state()->trim_queues(0); 539 540 // We have no local work, attempt to steal from other threads. 541 542 // attempt to steal work from promoted. 543 if (task_queues()->steal(par_scan_state()->thread_num(), 544 par_scan_state()->hash_seed(), 545 obj_to_scan)) { 546 bool res = work_q->push(obj_to_scan); 547 assert(res, "Empty queue should have room for a push."); 548 549 // if successful, goto Start. 550 continue; 551 552 // try global overflow list. 553 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 554 continue; 555 } 556 557 // Otherwise, offer termination. 558 par_scan_state()->start_term_time(); 559 if (terminator()->offer_termination()) break; 560 par_scan_state()->end_term_time(); 561 } 562 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 563 "Broken overflow list?"); 564 // Finish the last termination pause. 565 par_scan_state()->end_term_time(); 566 } 567 568 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen, Generation* old_gen, 569 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set, 570 StrongRootsScope* strong_roots_scope) : 571 AbstractGangTask("ParNewGeneration collection"), 572 _young_gen(young_gen), _old_gen(old_gen), 573 _young_old_boundary(young_old_boundary), 574 _state_set(state_set), 575 _strong_roots_scope(strong_roots_scope) 576 {} 577 578 void ParNewGenTask::work(uint worker_id) { 579 GenCollectedHeap* gch = GenCollectedHeap::heap(); 580 // Since this is being done in a separate thread, need new resource 581 // and handle marks. 582 ResourceMark rm; 583 HandleMark hm; 584 585 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 586 assert(_state_set->is_valid(worker_id), "Should not have been called"); 587 588 par_scan_state.set_young_old_boundary(_young_old_boundary); 589 590 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), 591 gch->rem_set()->klass_rem_set()); 592 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure, 593 &par_scan_state.to_space_root_closure(), 594 false); 595 596 par_scan_state.start_strong_roots(); 597 gch->gen_process_roots(_strong_roots_scope, 598 Generation::Young, 599 true, // Process younger gens, if any, 600 // as strong roots. 601 GenCollectedHeap::SO_ScavengeCodeCache, 602 GenCollectedHeap::StrongAndWeakRoots, 603 &par_scan_state.to_space_root_closure(), 604 &par_scan_state.older_gen_closure(), 605 &cld_scan_closure); 606 607 par_scan_state.end_strong_roots(); 608 609 // "evacuate followers". 610 par_scan_state.evacuate_followers_closure().do_void(); 611 } 612 613 #ifdef _MSC_VER 614 #pragma warning( push ) 615 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 616 #endif 617 ParNewGeneration:: 618 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size) 619 : DefNewGeneration(rs, initial_byte_size, "PCopy"), 620 _overflow_list(NULL), 621 _is_alive_closure(this), 622 _plab_stats(YoungPLABSize, PLABWeight) 623 { 624 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 625 NOT_PRODUCT(_num_par_pushes = 0;) 626 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 627 guarantee(_task_queues != NULL, "task_queues allocation failure."); 628 629 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { 630 ObjToScanQueue *q = new ObjToScanQueue(); 631 guarantee(q != NULL, "work_queue Allocation failure."); 632 _task_queues->register_queue(i1, q); 633 } 634 635 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) 636 _task_queues->queue(i2)->initialize(); 637 638 _overflow_stacks = NULL; 639 if (ParGCUseLocalOverflow) { 640 641 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal 642 // with ',' 643 typedef Stack<oop, mtGC> GCOopStack; 644 645 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 646 for (size_t i = 0; i < ParallelGCThreads; ++i) { 647 new (_overflow_stacks + i) Stack<oop, mtGC>(); 648 } 649 } 650 651 if (UsePerfData) { 652 EXCEPTION_MARK; 653 ResourceMark rm; 654 655 const char* cname = 656 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 657 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 658 ParallelGCThreads, CHECK); 659 } 660 } 661 #ifdef _MSC_VER 662 #pragma warning( pop ) 663 #endif 664 665 // ParNewGeneration:: 666 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 667 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 668 669 template <class T> 670 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 671 #ifdef ASSERT 672 { 673 assert(!oopDesc::is_null(*p), "expected non-null ref"); 674 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 675 // We never expect to see a null reference being processed 676 // as a weak reference. 677 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 678 } 679 #endif // ASSERT 680 681 _par_cl->do_oop_nv(p); 682 683 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 684 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 685 _rs->write_ref_field_gc_par(p, obj); 686 } 687 } 688 689 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 690 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 691 692 // ParNewGeneration:: 693 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 694 DefNewGeneration::KeepAliveClosure(cl) {} 695 696 template <class T> 697 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 698 #ifdef ASSERT 699 { 700 assert(!oopDesc::is_null(*p), "expected non-null ref"); 701 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 702 // We never expect to see a null reference being processed 703 // as a weak reference. 704 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 705 } 706 #endif // ASSERT 707 708 _cl->do_oop_nv(p); 709 710 if (GenCollectedHeap::heap()->is_in_reserved(p)) { 711 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 712 _rs->write_ref_field_gc_par(p, obj); 713 } 714 } 715 716 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 717 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 718 719 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 720 T heap_oop = oopDesc::load_heap_oop(p); 721 if (!oopDesc::is_null(heap_oop)) { 722 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 723 if ((HeapWord*)obj < _boundary) { 724 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 725 oop new_obj = obj->is_forwarded() 726 ? obj->forwardee() 727 : _g->DefNewGeneration::copy_to_survivor_space(obj); 728 oopDesc::encode_store_heap_oop_not_null(p, new_obj); 729 } 730 if (_gc_barrier) { 731 // If p points to a younger generation, mark the card. 732 if ((HeapWord*)obj < _gen_boundary) { 733 _rs->write_ref_field_gc_par(p, obj); 734 } 735 } 736 } 737 } 738 739 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 740 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 741 742 class ParNewRefProcTaskProxy: public AbstractGangTask { 743 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 744 public: 745 ParNewRefProcTaskProxy(ProcessTask& task, 746 ParNewGeneration& gen, 747 Generation& old_gen, 748 HeapWord* young_old_boundary, 749 ParScanThreadStateSet& state_set); 750 751 private: 752 virtual void work(uint worker_id); 753 private: 754 ParNewGeneration& _young_gen; 755 ProcessTask& _task; 756 Generation& _old_gen; 757 HeapWord* _young_old_boundary; 758 ParScanThreadStateSet& _state_set; 759 }; 760 761 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task, 762 ParNewGeneration& young_gen, 763 Generation& old_gen, 764 HeapWord* young_old_boundary, 765 ParScanThreadStateSet& state_set) 766 : AbstractGangTask("ParNewGeneration parallel reference processing"), 767 _young_gen(young_gen), 768 _task(task), 769 _old_gen(old_gen), 770 _young_old_boundary(young_old_boundary), 771 _state_set(state_set) 772 { 773 } 774 775 void ParNewRefProcTaskProxy::work(uint worker_id) 776 { 777 ResourceMark rm; 778 HandleMark hm; 779 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 780 par_scan_state.set_young_old_boundary(_young_old_boundary); 781 _task.work(worker_id, par_scan_state.is_alive_closure(), 782 par_scan_state.keep_alive_closure(), 783 par_scan_state.evacuate_followers_closure()); 784 } 785 786 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { 787 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 788 EnqueueTask& _task; 789 790 public: 791 ParNewRefEnqueueTaskProxy(EnqueueTask& task) 792 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), 793 _task(task) 794 { } 795 796 virtual void work(uint worker_id) 797 { 798 _task.work(worker_id); 799 } 800 }; 801 802 803 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) 804 { 805 GenCollectedHeap* gch = GenCollectedHeap::heap(); 806 FlexibleWorkGang* workers = gch->workers(); 807 assert(workers != NULL, "Need parallel worker threads."); 808 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed()); 809 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen, 810 _young_gen.reserved().end(), _state_set); 811 workers->run_task(&rp_task); 812 _state_set.reset(0 /* bad value in debug if not reset */, 813 _young_gen.promotion_failed()); 814 } 815 816 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) 817 { 818 GenCollectedHeap* gch = GenCollectedHeap::heap(); 819 FlexibleWorkGang* workers = gch->workers(); 820 assert(workers != NULL, "Need parallel worker threads."); 821 ParNewRefEnqueueTaskProxy enq_task(task); 822 workers->run_task(&enq_task); 823 } 824 825 void ParNewRefProcTaskExecutor::set_single_threaded_mode() 826 { 827 _state_set.flush(); 828 GenCollectedHeap* gch = GenCollectedHeap::heap(); 829 gch->save_marks(); 830 } 831 832 ScanClosureWithParBarrier:: 833 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 834 ScanClosure(g, gc_barrier) {} 835 836 EvacuateFollowersClosureGeneral:: 837 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, 838 OopsInGenClosure* cur, 839 OopsInGenClosure* older) : 840 _gch(gch), 841 _scan_cur_or_nonheap(cur), _scan_older(older) 842 {} 843 844 void EvacuateFollowersClosureGeneral::do_void() { 845 do { 846 // Beware: this call will lead to closure applications via virtual 847 // calls. 848 _gch->oop_since_save_marks_iterate(Generation::Young, 849 _scan_cur_or_nonheap, 850 _scan_older); 851 } while (!_gch->no_allocs_since_save_marks(true /* include_young */)); 852 } 853 854 855 // A Generation that does parallel young-gen collection. 856 857 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) { 858 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 859 _promo_failure_scan_stack.clear(true); // Clear cached segments. 860 861 remove_forwarding_pointers(); 862 if (PrintGCDetails) { 863 gclog_or_tty->print(" (promotion failed)"); 864 } 865 // All the spaces are in play for mark-sweep. 866 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 867 from()->set_next_compaction_space(to()); 868 gch->set_incremental_collection_failed(); 869 // Inform the next generation that a promotion failure occurred. 870 _old_gen->promotion_failure_occurred(); 871 872 // Trace promotion failure in the parallel GC threads 873 thread_state_set.trace_promotion_failed(gc_tracer()); 874 // Single threaded code may have reported promotion failure to the global state 875 if (_promotion_failed_info.has_failed()) { 876 _gc_tracer.report_promotion_failed(_promotion_failed_info); 877 } 878 // Reset the PromotionFailureALot counters. 879 NOT_PRODUCT(gch->reset_promotion_should_fail();) 880 } 881 882 void ParNewGeneration::collect(bool full, 883 bool clear_all_soft_refs, 884 size_t size, 885 bool is_tlab) { 886 assert(full || size > 0, "otherwise we don't want to collect"); 887 888 GenCollectedHeap* gch = GenCollectedHeap::heap(); 889 890 _gc_timer->register_gc_start(); 891 892 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); 893 FlexibleWorkGang* workers = gch->workers(); 894 assert(workers != NULL, "Need workgang for parallel work"); 895 uint active_workers = 896 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), 897 workers->active_workers(), 898 Threads::number_of_non_daemon_threads()); 899 workers->set_active_workers(active_workers); 900 _old_gen = gch->old_gen(); 901 902 // If the next generation is too full to accommodate worst-case promotion 903 // from this generation, pass on collection; let the next generation 904 // do it. 905 if (!collection_attempt_is_safe()) { 906 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 907 return; 908 } 909 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 910 911 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 912 gch->trace_heap_before_gc(gc_tracer()); 913 914 init_assuming_no_promotion_failure(); 915 916 if (UseAdaptiveSizePolicy) { 917 set_survivor_overflow(false); 918 size_policy->minor_collection_begin(); 919 } 920 921 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id()); 922 // Capture heap used before collection (for printing). 923 size_t gch_prev_used = gch->used(); 924 925 age_table()->clear(); 926 to()->clear(SpaceDecorator::Mangle); 927 928 gch->save_marks(); 929 930 // Set the correct parallelism (number of queues) in the reference processor 931 ref_processor()->set_active_mt_degree(active_workers); 932 933 // Always set the terminator for the active number of workers 934 // because only those workers go through the termination protocol. 935 ParallelTaskTerminator _term(active_workers, task_queues()); 936 ParScanThreadStateSet thread_state_set(active_workers, 937 *to(), *this, *_old_gen, *task_queues(), 938 _overflow_stacks, desired_plab_sz(), _term); 939 940 thread_state_set.reset(active_workers, promotion_failed()); 941 942 { 943 StrongRootsScope srs(active_workers); 944 945 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs); 946 gch->rem_set()->prepare_for_younger_refs_iterate(true); 947 // It turns out that even when we're using 1 thread, doing the work in a 948 // separate thread causes wide variance in run times. We can't help this 949 // in the multi-threaded case, but we special-case n=1 here to get 950 // repeatable measurements of the 1-thread overhead of the parallel code. 951 if (active_workers > 1) { 952 workers->run_task(&tsk); 953 } else { 954 tsk.work(0); 955 } 956 } 957 958 thread_state_set.reset(0 /* Bad value in debug if not reset */, 959 promotion_failed()); 960 961 // Trace and reset failed promotion info. 962 if (promotion_failed()) { 963 thread_state_set.trace_promotion_failed(gc_tracer()); 964 } 965 966 // Process (weak) reference objects found during scavenge. 967 ReferenceProcessor* rp = ref_processor(); 968 IsAliveClosure is_alive(this); 969 ScanWeakRefClosure scan_weak_ref(this); 970 KeepAliveClosure keep_alive(&scan_weak_ref); 971 ScanClosure scan_without_gc_barrier(this, false); 972 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 973 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 974 EvacuateFollowersClosureGeneral evacuate_followers(gch, 975 &scan_without_gc_barrier, &scan_with_gc_barrier); 976 rp->setup_policy(clear_all_soft_refs); 977 // Can the mt_degree be set later (at run_task() time would be best)? 978 rp->set_active_mt_degree(active_workers); 979 ReferenceProcessorStats stats; 980 if (rp->processing_is_mt()) { 981 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 982 stats = rp->process_discovered_references(&is_alive, &keep_alive, 983 &evacuate_followers, &task_executor, 984 _gc_timer, _gc_tracer.gc_id()); 985 } else { 986 thread_state_set.flush(); 987 gch->save_marks(); 988 stats = rp->process_discovered_references(&is_alive, &keep_alive, 989 &evacuate_followers, NULL, 990 _gc_timer, _gc_tracer.gc_id()); 991 } 992 _gc_tracer.report_gc_reference_stats(stats); 993 if (!promotion_failed()) { 994 // Swap the survivor spaces. 995 eden()->clear(SpaceDecorator::Mangle); 996 from()->clear(SpaceDecorator::Mangle); 997 if (ZapUnusedHeapArea) { 998 // This is now done here because of the piece-meal mangling which 999 // can check for valid mangling at intermediate points in the 1000 // collection(s). When a minor collection fails to collect 1001 // sufficient space resizing of the young generation can occur 1002 // an redistribute the spaces in the young generation. Mangle 1003 // here so that unzapped regions don't get distributed to 1004 // other spaces. 1005 to()->mangle_unused_area(); 1006 } 1007 swap_spaces(); 1008 1009 // A successful scavenge should restart the GC time limit count which is 1010 // for full GC's. 1011 size_policy->reset_gc_overhead_limit_count(); 1012 1013 assert(to()->is_empty(), "to space should be empty now"); 1014 1015 adjust_desired_tenuring_threshold(); 1016 } else { 1017 handle_promotion_failed(gch, thread_state_set); 1018 } 1019 // set new iteration safe limit for the survivor spaces 1020 from()->set_concurrent_iteration_safe_limit(from()->top()); 1021 to()->set_concurrent_iteration_safe_limit(to()->top()); 1022 1023 if (ResizePLAB) { 1024 plab_stats()->adjust_desired_plab_sz(active_workers); 1025 } 1026 1027 if (PrintGC && !PrintGCDetails) { 1028 gch->print_heap_change(gch_prev_used); 1029 } 1030 1031 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats()); 1032 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats()); 1033 1034 if (UseAdaptiveSizePolicy) { 1035 size_policy->minor_collection_end(gch->gc_cause()); 1036 size_policy->avg_survived()->sample(from()->used()); 1037 } 1038 1039 // We need to use a monotonically non-decreasing time in ms 1040 // or we will see time-warp warnings and os::javaTimeMillis() 1041 // does not guarantee monotonicity. 1042 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1043 update_time_of_last_gc(now); 1044 1045 rp->set_enqueuing_is_done(true); 1046 if (rp->processing_is_mt()) { 1047 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set); 1048 rp->enqueue_discovered_references(&task_executor); 1049 } else { 1050 rp->enqueue_discovered_references(NULL); 1051 } 1052 rp->verify_no_references_recorded(); 1053 1054 gch->trace_heap_after_gc(gc_tracer()); 1055 _gc_tracer.report_tenuring_threshold(tenuring_threshold()); 1056 1057 _gc_timer->register_gc_end(); 1058 1059 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1060 } 1061 1062 static int sum; 1063 void ParNewGeneration::waste_some_time() { 1064 for (int i = 0; i < 100; i++) { 1065 sum += i; 1066 } 1067 } 1068 1069 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); 1070 1071 // Because of concurrency, there are times where an object for which 1072 // "is_forwarded()" is true contains an "interim" forwarding pointer 1073 // value. Such a value will soon be overwritten with a real value. 1074 // This method requires "obj" to have a forwarding pointer, and waits, if 1075 // necessary for a real one to be inserted, and returns it. 1076 1077 oop ParNewGeneration::real_forwardee(oop obj) { 1078 oop forward_ptr = obj->forwardee(); 1079 if (forward_ptr != ClaimedForwardPtr) { 1080 return forward_ptr; 1081 } else { 1082 return real_forwardee_slow(obj); 1083 } 1084 } 1085 1086 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1087 // Spin-read if it is claimed but not yet written by another thread. 1088 oop forward_ptr = obj->forwardee(); 1089 while (forward_ptr == ClaimedForwardPtr) { 1090 waste_some_time(); 1091 assert(obj->is_forwarded(), "precondition"); 1092 forward_ptr = obj->forwardee(); 1093 } 1094 return forward_ptr; 1095 } 1096 1097 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { 1098 if (m->must_be_preserved_for_promotion_failure(obj)) { 1099 // We should really have separate per-worker stacks, rather 1100 // than use locking of a common pair of stacks. 1101 MutexLocker ml(ParGCRareEvent_lock); 1102 preserve_mark(obj, m); 1103 } 1104 } 1105 1106 // Multiple GC threads may try to promote an object. If the object 1107 // is successfully promoted, a forwarding pointer will be installed in 1108 // the object in the young generation. This method claims the right 1109 // to install the forwarding pointer before it copies the object, 1110 // thus avoiding the need to undo the copy as in 1111 // copy_to_survivor_space_avoiding_with_undo. 1112 1113 oop ParNewGeneration::copy_to_survivor_space( 1114 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { 1115 // In the sequential version, this assert also says that the object is 1116 // not forwarded. That might not be the case here. It is the case that 1117 // the caller observed it to be not forwarded at some time in the past. 1118 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1119 1120 // The sequential code read "old->age()" below. That doesn't work here, 1121 // since the age is in the mark word, and that might be overwritten with 1122 // a forwarding pointer by a parallel thread. So we must save the mark 1123 // word in a local and then analyze it. 1124 oopDesc dummyOld; 1125 dummyOld.set_mark(m); 1126 assert(!dummyOld.is_forwarded(), 1127 "should not be called with forwarding pointer mark word."); 1128 1129 oop new_obj = NULL; 1130 oop forward_ptr; 1131 1132 // Try allocating obj in to-space (unless too old) 1133 if (dummyOld.age() < tenuring_threshold()) { 1134 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1135 if (new_obj == NULL) { 1136 set_survivor_overflow(true); 1137 } 1138 } 1139 1140 if (new_obj == NULL) { 1141 // Either to-space is full or we decided to promote 1142 // try allocating obj tenured 1143 1144 // Attempt to install a null forwarding pointer (atomically), 1145 // to claim the right to install the real forwarding pointer. 1146 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); 1147 if (forward_ptr != NULL) { 1148 // someone else beat us to it. 1149 return real_forwardee(old); 1150 } 1151 1152 if (!_promotion_failed) { 1153 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), 1154 old, m, sz); 1155 } 1156 1157 if (new_obj == NULL) { 1158 // promotion failed, forward to self 1159 _promotion_failed = true; 1160 new_obj = old; 1161 1162 preserve_mark_if_necessary(old, m); 1163 par_scan_state->register_promotion_failure(sz); 1164 } 1165 1166 old->forward_to(new_obj); 1167 forward_ptr = NULL; 1168 } else { 1169 // Is in to-space; do copying ourselves. 1170 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1171 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value."); 1172 forward_ptr = old->forward_to_atomic(new_obj); 1173 // Restore the mark word copied above. 1174 new_obj->set_mark(m); 1175 // Increment age if obj still in new generation 1176 new_obj->incr_age(); 1177 par_scan_state->age_table()->add(new_obj, sz); 1178 } 1179 assert(new_obj != NULL, "just checking"); 1180 1181 #ifndef PRODUCT 1182 // This code must come after the CAS test, or it will print incorrect 1183 // information. 1184 if (TraceScavenge) { 1185 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1186 is_in_reserved(new_obj) ? "copying" : "tenuring", 1187 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size()); 1188 } 1189 #endif 1190 1191 if (forward_ptr == NULL) { 1192 oop obj_to_push = new_obj; 1193 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1194 // Length field used as index of next element to be scanned. 1195 // Real length can be obtained from real_forwardee() 1196 arrayOop(old)->set_length(0); 1197 obj_to_push = old; 1198 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1199 "push forwarded object"); 1200 } 1201 // Push it on one of the queues of to-be-scanned objects. 1202 bool simulate_overflow = false; 1203 NOT_PRODUCT( 1204 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1205 // simulate a stack overflow 1206 simulate_overflow = true; 1207 } 1208 ) 1209 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1210 // Add stats for overflow pushes. 1211 if (Verbose && PrintGCDetails) { 1212 gclog_or_tty->print("queue overflow!\n"); 1213 } 1214 push_on_overflow_list(old, par_scan_state); 1215 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1216 } 1217 1218 return new_obj; 1219 } 1220 1221 // Oops. Someone beat us to it. Undo the allocation. Where did we 1222 // allocate it? 1223 if (is_in_reserved(new_obj)) { 1224 // Must be in to_space. 1225 assert(to()->is_in_reserved(new_obj), "Checking"); 1226 if (forward_ptr == ClaimedForwardPtr) { 1227 // Wait to get the real forwarding pointer value. 1228 forward_ptr = real_forwardee(old); 1229 } 1230 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1231 } 1232 1233 return forward_ptr; 1234 } 1235 1236 #ifndef PRODUCT 1237 // It's OK to call this multi-threaded; the worst thing 1238 // that can happen is that we'll get a bunch of closely 1239 // spaced simulated overflows, but that's OK, in fact 1240 // probably good as it would exercise the overflow code 1241 // under contention. 1242 bool ParNewGeneration::should_simulate_overflow() { 1243 if (_overflow_counter-- <= 0) { // just being defensive 1244 _overflow_counter = ParGCWorkQueueOverflowInterval; 1245 return true; 1246 } else { 1247 return false; 1248 } 1249 } 1250 #endif 1251 1252 // In case we are using compressed oops, we need to be careful. 1253 // If the object being pushed is an object array, then its length 1254 // field keeps track of the "grey boundary" at which the next 1255 // incremental scan will be done (see ParGCArrayScanChunk). 1256 // When using compressed oops, this length field is kept in the 1257 // lower 32 bits of the erstwhile klass word and cannot be used 1258 // for the overflow chaining pointer (OCP below). As such the OCP 1259 // would itself need to be compressed into the top 32-bits in this 1260 // case. Unfortunately, see below, in the event that we have a 1261 // promotion failure, the node to be pushed on the list can be 1262 // outside of the Java heap, so the heap-based pointer compression 1263 // would not work (we would have potential aliasing between C-heap 1264 // and Java-heap pointers). For this reason, when using compressed 1265 // oops, we simply use a worker-thread-local, non-shared overflow 1266 // list in the form of a growable array, with a slightly different 1267 // overflow stack draining strategy. If/when we start using fat 1268 // stacks here, we can go back to using (fat) pointer chains 1269 // (although some performance comparisons would be useful since 1270 // single global lists have their own performance disadvantages 1271 // as we were made painfully aware not long ago, see 6786503). 1272 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 1273 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1274 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1275 if (ParGCUseLocalOverflow) { 1276 // In the case of compressed oops, we use a private, not-shared 1277 // overflow stack. 1278 par_scan_state->push_on_overflow_stack(from_space_obj); 1279 } else { 1280 assert(!UseCompressedOops, "Error"); 1281 // if the object has been forwarded to itself, then we cannot 1282 // use the klass pointer for the linked list. Instead we have 1283 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1284 // XXX This is horribly inefficient when a promotion failure occurs 1285 // and should be fixed. XXX FIX ME !!! 1286 #ifndef PRODUCT 1287 Atomic::inc_ptr(&_num_par_pushes); 1288 assert(_num_par_pushes > 0, "Tautology"); 1289 #endif 1290 if (from_space_obj->forwardee() == from_space_obj) { 1291 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); 1292 listhead->forward_to(from_space_obj); 1293 from_space_obj = listhead; 1294 } 1295 oop observed_overflow_list = _overflow_list; 1296 oop cur_overflow_list; 1297 do { 1298 cur_overflow_list = observed_overflow_list; 1299 if (cur_overflow_list != BUSY) { 1300 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1301 } else { 1302 from_space_obj->set_klass_to_list_ptr(NULL); 1303 } 1304 observed_overflow_list = 1305 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); 1306 } while (cur_overflow_list != observed_overflow_list); 1307 } 1308 } 1309 1310 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1311 bool res; 1312 1313 if (ParGCUseLocalOverflow) { 1314 res = par_scan_state->take_from_overflow_stack(); 1315 } else { 1316 assert(!UseCompressedOops, "Error"); 1317 res = take_from_overflow_list_work(par_scan_state); 1318 } 1319 return res; 1320 } 1321 1322 1323 // *NOTE*: The overflow list manipulation code here and 1324 // in CMSCollector:: are very similar in shape, 1325 // except that in the CMS case we thread the objects 1326 // directly into the list via their mark word, and do 1327 // not need to deal with special cases below related 1328 // to chunking of object arrays and promotion failure 1329 // handling. 1330 // CR 6797058 has been filed to attempt consolidation of 1331 // the common code. 1332 // Because of the common code, if you make any changes in 1333 // the code below, please check the CMS version to see if 1334 // similar changes might be needed. 1335 // See CMSCollector::par_take_from_overflow_list() for 1336 // more extensive documentation comments. 1337 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1338 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1339 // How many to take? 1340 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1341 (size_t)ParGCDesiredObjsFromOverflowList); 1342 1343 assert(!UseCompressedOops, "Error"); 1344 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1345 if (_overflow_list == NULL) return false; 1346 1347 // Otherwise, there was something there; try claiming the list. 1348 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1349 // Trim off a prefix of at most objsFromOverflow items 1350 Thread* tid = Thread::current(); 1351 size_t spin_count = (size_t)ParallelGCThreads; 1352 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1353 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1354 // someone grabbed it before we did ... 1355 // ... we spin for a short while... 1356 os::sleep(tid, sleep_time_millis, false); 1357 if (_overflow_list == NULL) { 1358 // nothing left to take 1359 return false; 1360 } else if (_overflow_list != BUSY) { 1361 // try and grab the prefix 1362 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1363 } 1364 } 1365 if (prefix == NULL || prefix == BUSY) { 1366 // Nothing to take or waited long enough 1367 if (prefix == NULL) { 1368 // Write back the NULL in case we overwrote it with BUSY above 1369 // and it is still the same value. 1370 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1371 } 1372 return false; 1373 } 1374 assert(prefix != NULL && prefix != BUSY, "Error"); 1375 size_t i = 1; 1376 oop cur = prefix; 1377 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { 1378 i++; cur = cur->list_ptr_from_klass(); 1379 } 1380 1381 // Reattach remaining (suffix) to overflow list 1382 if (cur->klass_or_null() == NULL) { 1383 // Write back the NULL in lieu of the BUSY we wrote 1384 // above and it is still the same value. 1385 if (_overflow_list == BUSY) { 1386 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1387 } 1388 } else { 1389 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); 1390 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list 1391 cur->set_klass_to_list_ptr(NULL); // break off suffix 1392 // It's possible that the list is still in the empty(busy) state 1393 // we left it in a short while ago; in that case we may be 1394 // able to place back the suffix. 1395 oop observed_overflow_list = _overflow_list; 1396 oop cur_overflow_list = observed_overflow_list; 1397 bool attached = false; 1398 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1399 observed_overflow_list = 1400 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1401 if (cur_overflow_list == observed_overflow_list) { 1402 attached = true; 1403 break; 1404 } else cur_overflow_list = observed_overflow_list; 1405 } 1406 if (!attached) { 1407 // Too bad, someone else got in in between; we'll need to do a splice. 1408 // Find the last item of suffix list 1409 oop last = suffix; 1410 while (last->klass_or_null() != NULL) { 1411 last = last->list_ptr_from_klass(); 1412 } 1413 // Atomically prepend suffix to current overflow list 1414 observed_overflow_list = _overflow_list; 1415 do { 1416 cur_overflow_list = observed_overflow_list; 1417 if (cur_overflow_list != BUSY) { 1418 // Do the splice ... 1419 last->set_klass_to_list_ptr(cur_overflow_list); 1420 } else { // cur_overflow_list == BUSY 1421 last->set_klass_to_list_ptr(NULL); 1422 } 1423 observed_overflow_list = 1424 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1425 } while (cur_overflow_list != observed_overflow_list); 1426 } 1427 } 1428 1429 // Push objects on prefix list onto this thread's work queue 1430 assert(prefix != NULL && prefix != BUSY, "program logic"); 1431 cur = prefix; 1432 ssize_t n = 0; 1433 while (cur != NULL) { 1434 oop obj_to_push = cur->forwardee(); 1435 oop next = cur->list_ptr_from_klass(); 1436 cur->set_klass(obj_to_push->klass()); 1437 // This may be an array object that is self-forwarded. In that case, the list pointer 1438 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1439 if (!is_in_reserved(cur)) { 1440 // This can become a scaling bottleneck when there is work queue overflow coincident 1441 // with promotion failure. 1442 oopDesc* f = cur; 1443 FREE_C_HEAP_ARRAY(oopDesc, f); 1444 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1445 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1446 obj_to_push = cur; 1447 } 1448 bool ok = work_q->push(obj_to_push); 1449 assert(ok, "Should have succeeded"); 1450 cur = next; 1451 n++; 1452 } 1453 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1454 #ifndef PRODUCT 1455 assert(_num_par_pushes >= n, "Too many pops?"); 1456 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 1457 #endif 1458 return true; 1459 } 1460 #undef BUSY 1461 1462 void ParNewGeneration::ref_processor_init() { 1463 if (_ref_processor == NULL) { 1464 // Allocate and initialize a reference processor 1465 _ref_processor = 1466 new ReferenceProcessor(_reserved, // span 1467 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1468 (uint) ParallelGCThreads, // mt processing degree 1469 refs_discovery_is_mt(), // mt discovery 1470 (uint) ParallelGCThreads, // mt discovery degree 1471 refs_discovery_is_atomic(), // atomic_discovery 1472 NULL); // is_alive_non_header 1473 } 1474 } 1475 1476 const char* ParNewGeneration::name() const { 1477 return "par new generation"; 1478 }