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_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 27 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp" 28 #include "gc_implementation/parNew/parNewGeneration.hpp" 29 #include "gc_implementation/parNew/parOopClosures.inline.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/shared/ageTable.hpp" 32 #include "gc_implementation/shared/copyFailedInfo.hpp" 33 #include "gc_implementation/shared/gcHeapSummary.hpp" 34 #include "gc_implementation/shared/gcTimer.hpp" 35 #include "gc_implementation/shared/gcTrace.hpp" 36 #include "gc_implementation/shared/gcTraceTime.hpp" 37 #include "gc_implementation/shared/parGCAllocBuffer.inline.hpp" 38 #include "gc_implementation/shared/spaceDecorator.hpp" 39 #include "memory/defNewGeneration.inline.hpp" 40 #include "memory/genCollectedHeap.hpp" 41 #include "memory/genOopClosures.inline.hpp" 42 #include "memory/generation.hpp" 43 #include "memory/referencePolicy.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "memory/strongRootsScope.hpp" 46 #include "memory/space.hpp" 47 #include "oops/objArrayOop.hpp" 48 #include "oops/oop.inline.hpp" 49 #include "runtime/atomic.inline.hpp" 50 #include "runtime/handles.hpp" 51 #include "runtime/handles.inline.hpp" 52 #include "runtime/java.hpp" 53 #include "runtime/thread.inline.hpp" 54 #include "utilities/copy.hpp" 55 #include "utilities/globalDefinitions.hpp" 56 #include "utilities/workgroup.hpp" 57 58 #ifdef _MSC_VER 59 #pragma warning( push ) 60 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 61 #endif 62 ParScanThreadState::ParScanThreadState(Space* to_space_, 63 ParNewGeneration* gen_, 64 Generation* old_gen_, 65 int thread_num_, 66 ObjToScanQueueSet* work_queue_set_, 67 Stack<oop, mtGC>* overflow_stacks_, 68 size_t desired_plab_sz_, 69 ParallelTaskTerminator& term_) : 70 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), 71 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), 72 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), 73 _ageTable(false), // false ==> not the global age table, no perf data. 74 _to_space_alloc_buffer(desired_plab_sz_), 75 _to_space_closure(gen_, this), _old_gen_closure(gen_, this), 76 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), 77 _older_gen_closure(gen_, this), 78 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, 79 &_to_space_root_closure, gen_, &_old_gen_root_closure, 80 work_queue_set_, &term_), 81 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), 82 _keep_alive_closure(&_scan_weak_ref_closure), 83 _strong_roots_time(0.0), _term_time(0.0) 84 { 85 #if TASKQUEUE_STATS 86 _term_attempts = 0; 87 _overflow_refills = 0; 88 _overflow_refill_objs = 0; 89 #endif // TASKQUEUE_STATS 90 91 _survivor_chunk_array = 92 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); 93 _hash_seed = 17; // Might want to take time-based random value. 94 _start = os::elapsedTime(); 95 _old_gen_closure.set_generation(old_gen_); 96 _old_gen_root_closure.set_generation(old_gen_); 97 } 98 #ifdef _MSC_VER 99 #pragma warning( pop ) 100 #endif 101 102 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, 103 size_t plab_word_size) { 104 ChunkArray* sca = survivor_chunk_array(); 105 if (sca != NULL) { 106 // A non-null SCA implies that we want the PLAB data recorded. 107 sca->record_sample(plab_start, plab_word_size); 108 } 109 } 110 111 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { 112 return new_obj->is_objArray() && 113 arrayOop(new_obj)->length() > ParGCArrayScanChunk && 114 new_obj != old_obj; 115 } 116 117 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { 118 assert(old->is_objArray(), "must be obj array"); 119 assert(old->is_forwarded(), "must be forwarded"); 120 assert(Universe::heap()->is_in_reserved(old), "must be in heap."); 121 assert(!old_gen()->is_in(old), "must be in young generation."); 122 123 objArrayOop obj = objArrayOop(old->forwardee()); 124 // Process ParGCArrayScanChunk elements now 125 // and push the remainder back onto queue 126 int start = arrayOop(old)->length(); 127 int end = obj->length(); 128 int remainder = end - start; 129 assert(start <= end, "just checking"); 130 if (remainder > 2 * ParGCArrayScanChunk) { 131 // Test above combines last partial chunk with a full chunk 132 end = start + ParGCArrayScanChunk; 133 arrayOop(old)->set_length(end); 134 // Push remainder. 135 bool ok = work_queue()->push(old); 136 assert(ok, "just popped, push must be okay"); 137 } else { 138 // Restore length so that it can be used if there 139 // is a promotion failure and forwarding pointers 140 // must be removed. 141 arrayOop(old)->set_length(end); 142 } 143 144 // process our set of indices (include header in first chunk) 145 // should make sure end is even (aligned to HeapWord in case of compressed oops) 146 if ((HeapWord *)obj < young_old_boundary()) { 147 // object is in to_space 148 obj->oop_iterate_range(&_to_space_closure, start, end); 149 } else { 150 // object is in old generation 151 obj->oop_iterate_range(&_old_gen_closure, start, end); 152 } 153 } 154 155 156 void ParScanThreadState::trim_queues(int max_size) { 157 ObjToScanQueue* queue = work_queue(); 158 do { 159 while (queue->size() > (juint)max_size) { 160 oop obj_to_scan; 161 if (queue->pop_local(obj_to_scan)) { 162 if ((HeapWord *)obj_to_scan < young_old_boundary()) { 163 if (obj_to_scan->is_objArray() && 164 obj_to_scan->is_forwarded() && 165 obj_to_scan->forwardee() != obj_to_scan) { 166 scan_partial_array_and_push_remainder(obj_to_scan); 167 } else { 168 // object is in to_space 169 obj_to_scan->oop_iterate(&_to_space_closure); 170 } 171 } else { 172 // object is in old generation 173 obj_to_scan->oop_iterate(&_old_gen_closure); 174 } 175 } 176 } 177 // For the case of compressed oops, we have a private, non-shared 178 // overflow stack, so we eagerly drain it so as to more evenly 179 // distribute load early. Note: this may be good to do in 180 // general rather than delay for the final stealing phase. 181 // If applicable, we'll transfer a set of objects over to our 182 // work queue, allowing them to be stolen and draining our 183 // private overflow stack. 184 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); 185 } 186 187 bool ParScanThreadState::take_from_overflow_stack() { 188 assert(ParGCUseLocalOverflow, "Else should not call"); 189 assert(young_gen()->overflow_list() == NULL, "Error"); 190 ObjToScanQueue* queue = work_queue(); 191 Stack<oop, mtGC>* const of_stack = overflow_stack(); 192 const size_t num_overflow_elems = of_stack->size(); 193 const size_t space_available = queue->max_elems() - queue->size(); 194 const size_t num_take_elems = MIN3(space_available / 4, 195 ParGCDesiredObjsFromOverflowList, 196 num_overflow_elems); 197 // Transfer the most recent num_take_elems from the overflow 198 // stack to our work queue. 199 for (size_t i = 0; i != num_take_elems; i++) { 200 oop cur = of_stack->pop(); 201 oop obj_to_push = cur->forwardee(); 202 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap"); 203 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); 204 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap"); 205 if (should_be_partially_scanned(obj_to_push, cur)) { 206 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 207 obj_to_push = cur; 208 } 209 bool ok = queue->push(obj_to_push); 210 assert(ok, "Should have succeeded"); 211 } 212 assert(young_gen()->overflow_list() == NULL, "Error"); 213 return num_take_elems > 0; // was something transferred? 214 } 215 216 void ParScanThreadState::push_on_overflow_stack(oop p) { 217 assert(ParGCUseLocalOverflow, "Else should not call"); 218 overflow_stack()->push(p); 219 assert(young_gen()->overflow_list() == NULL, "Error"); 220 } 221 222 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { 223 224 // Otherwise, if the object is small enough, try to reallocate the 225 // buffer. 226 HeapWord* obj = NULL; 227 if (!_to_space_full) { 228 ParGCAllocBuffer* const plab = to_space_alloc_buffer(); 229 Space* const sp = to_space(); 230 if (word_sz * 100 < 231 ParallelGCBufferWastePct * plab->word_sz()) { 232 // Is small enough; abandon this buffer and start a new one. 233 plab->retire(); 234 size_t buf_size = plab->word_sz(); 235 HeapWord* buf_space = sp->par_allocate(buf_size); 236 if (buf_space == NULL) { 237 const size_t min_bytes = 238 ParGCAllocBuffer::min_size() << LogHeapWordSize; 239 size_t free_bytes = sp->free(); 240 while(buf_space == NULL && free_bytes >= min_bytes) { 241 buf_size = free_bytes >> LogHeapWordSize; 242 assert(buf_size == (size_t)align_object_size(buf_size), 243 "Invariant"); 244 buf_space = sp->par_allocate(buf_size); 245 free_bytes = sp->free(); 246 } 247 } 248 if (buf_space != NULL) { 249 plab->set_word_size(buf_size); 250 plab->set_buf(buf_space); 251 record_survivor_plab(buf_space, buf_size); 252 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes); 253 // Note that we cannot compare buf_size < word_sz below 254 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). 255 assert(obj != NULL || plab->words_remaining() < word_sz, 256 "Else should have been able to allocate"); 257 // It's conceivable that we may be able to use the 258 // buffer we just grabbed for subsequent small requests 259 // even if not for this one. 260 } else { 261 // We're used up. 262 _to_space_full = true; 263 } 264 265 } else { 266 // Too large; allocate the object individually. 267 obj = sp->par_allocate(word_sz); 268 } 269 } 270 return obj; 271 } 272 273 274 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, 275 size_t word_sz) { 276 // Is the alloc in the current alloc buffer? 277 if (to_space_alloc_buffer()->contains(obj)) { 278 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), 279 "Should contain whole object."); 280 to_space_alloc_buffer()->undo_allocation(obj, word_sz); 281 } else { 282 CollectedHeap::fill_with_object(obj, word_sz); 283 } 284 } 285 286 void ParScanThreadState::print_promotion_failure_size() { 287 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) { 288 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", 289 _thread_num, _promotion_failed_info.first_size()); 290 } 291 } 292 293 class ParScanThreadStateSet: private ResourceArray { 294 public: 295 // Initializes states for the specified number of threads; 296 ParScanThreadStateSet(int num_threads, 297 Space& to_space, 298 ParNewGeneration& gen, 299 Generation& old_gen, 300 ObjToScanQueueSet& queue_set, 301 Stack<oop, mtGC>* overflow_stacks_, 302 size_t desired_plab_sz, 303 ParallelTaskTerminator& term); 304 305 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } 306 307 inline ParScanThreadState& thread_state(int i); 308 309 void trace_promotion_failed(const YoungGCTracer* gc_tracer); 310 void reset(int active_workers, bool promotion_failed); 311 void flush(); 312 313 #if TASKQUEUE_STATS 314 static void 315 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); 316 void print_termination_stats(outputStream* const st = gclog_or_tty); 317 static void 318 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); 319 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); 320 void reset_stats(); 321 #endif // TASKQUEUE_STATS 322 323 private: 324 ParallelTaskTerminator& _term; 325 ParNewGeneration& _gen; 326 Generation& _old_gen; 327 public: 328 bool is_valid(int id) const { return id < length(); } 329 ParallelTaskTerminator* terminator() { return &_term; } 330 }; 331 332 333 ParScanThreadStateSet::ParScanThreadStateSet( 334 int num_threads, Space& to_space, ParNewGeneration& gen, 335 Generation& old_gen, ObjToScanQueueSet& queue_set, 336 Stack<oop, mtGC>* overflow_stacks, 337 size_t desired_plab_sz, ParallelTaskTerminator& term) 338 : ResourceArray(sizeof(ParScanThreadState), num_threads), 339 _gen(gen), _old_gen(old_gen), _term(term) 340 { 341 assert(num_threads > 0, "sanity check!"); 342 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), 343 "overflow_stack allocation mismatch"); 344 // Initialize states. 345 for (int i = 0; i < num_threads; ++i) { 346 new ((ParScanThreadState*)_data + i) 347 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, 348 overflow_stacks, desired_plab_sz, term); 349 } 350 } 351 352 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) 353 { 354 assert(i >= 0 && i < length(), "sanity check!"); 355 return ((ParScanThreadState*)_data)[i]; 356 } 357 358 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) { 359 for (int i = 0; i < length(); ++i) { 360 if (thread_state(i).promotion_failed()) { 361 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info()); 362 thread_state(i).promotion_failed_info().reset(); 363 } 364 } 365 } 366 367 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) 368 { 369 _term.reset_for_reuse(active_threads); 370 if (promotion_failed) { 371 for (int i = 0; i < length(); ++i) { 372 thread_state(i).print_promotion_failure_size(); 373 } 374 } 375 } 376 377 #if TASKQUEUE_STATS 378 void 379 ParScanThreadState::reset_stats() 380 { 381 taskqueue_stats().reset(); 382 _term_attempts = 0; 383 _overflow_refills = 0; 384 _overflow_refill_objs = 0; 385 } 386 387 void ParScanThreadStateSet::reset_stats() 388 { 389 for (int i = 0; i < length(); ++i) { 390 thread_state(i).reset_stats(); 391 } 392 } 393 394 void 395 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) 396 { 397 st->print_raw_cr("GC Termination Stats"); 398 st->print_raw_cr(" elapsed --strong roots-- " 399 "-------termination-------"); 400 st->print_raw_cr("thr ms ms % " 401 " ms % attempts"); 402 st->print_raw_cr("--- --------- --------- ------ " 403 "--------- ------ --------"); 404 } 405 406 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) 407 { 408 print_termination_stats_hdr(st); 409 410 for (int i = 0; i < length(); ++i) { 411 const ParScanThreadState & pss = thread_state(i); 412 const double elapsed_ms = pss.elapsed_time() * 1000.0; 413 const double s_roots_ms = pss.strong_roots_time() * 1000.0; 414 const double term_ms = pss.term_time() * 1000.0; 415 st->print_cr("%3d %9.2f %9.2f %6.2f " 416 "%9.2f %6.2f " SIZE_FORMAT_W(8), 417 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 418 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); 419 } 420 } 421 422 // Print stats related to work queue activity. 423 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) 424 { 425 st->print_raw_cr("GC Task Stats"); 426 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 427 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 428 } 429 430 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) 431 { 432 print_taskqueue_stats_hdr(st); 433 434 TaskQueueStats totals; 435 for (int i = 0; i < length(); ++i) { 436 const ParScanThreadState & pss = thread_state(i); 437 const TaskQueueStats & stats = pss.taskqueue_stats(); 438 st->print("%3d ", i); stats.print(st); st->cr(); 439 totals += stats; 440 441 if (pss.overflow_refills() > 0) { 442 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " 443 SIZE_FORMAT_W(10) " overflow objects", 444 pss.overflow_refills(), pss.overflow_refill_objs()); 445 } 446 } 447 st->print("tot "); totals.print(st); st->cr(); 448 449 DEBUG_ONLY(totals.verify()); 450 } 451 #endif // TASKQUEUE_STATS 452 453 void ParScanThreadStateSet::flush() 454 { 455 // Work in this loop should be kept as lightweight as 456 // possible since this might otherwise become a bottleneck 457 // to scaling. Should we add heavy-weight work into this 458 // loop, consider parallelizing the loop into the worker threads. 459 for (int i = 0; i < length(); ++i) { 460 ParScanThreadState& par_scan_state = thread_state(i); 461 462 // Flush stats related to To-space PLAB activity and 463 // retire the last buffer. 464 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats()); 465 466 // Every thread has its own age table. We need to merge 467 // them all into one. 468 ageTable *local_table = par_scan_state.age_table(); 469 _gen.age_table()->merge(local_table); 470 471 // Inform old gen that we're done. 472 _old_gen.par_promote_alloc_done(i); 473 _old_gen.par_oop_since_save_marks_iterate_done(i); 474 } 475 476 if (UseConcMarkSweepGC) { 477 // We need to call this even when ResizeOldPLAB is disabled 478 // so as to avoid breaking some asserts. While we may be able 479 // to avoid this by reorganizing the code a bit, I am loathe 480 // to do that unless we find cases where ergo leads to bad 481 // performance. 482 CFLS_LAB::compute_desired_plab_size(); 483 } 484 } 485 486 ParScanClosure::ParScanClosure(ParNewGeneration* g, 487 ParScanThreadState* par_scan_state) : 488 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) 489 { 490 assert(_g->level() == 0, "Optimized for youngest generation"); 491 _boundary = _g->reserved().end(); 492 } 493 494 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } 495 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } 496 497 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } 498 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } 499 500 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } 501 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } 502 503 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } 504 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } 505 506 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, 507 ParScanThreadState* par_scan_state) 508 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) 509 {} 510 511 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } 512 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } 513 514 #ifdef WIN32 515 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ 516 #endif 517 518 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( 519 ParScanThreadState* par_scan_state_, 520 ParScanWithoutBarrierClosure* to_space_closure_, 521 ParScanWithBarrierClosure* old_gen_closure_, 522 ParRootScanWithoutBarrierClosure* to_space_root_closure_, 523 ParNewGeneration* par_gen_, 524 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, 525 ObjToScanQueueSet* task_queues_, 526 ParallelTaskTerminator* terminator_) : 527 528 _par_scan_state(par_scan_state_), 529 _to_space_closure(to_space_closure_), 530 _old_gen_closure(old_gen_closure_), 531 _to_space_root_closure(to_space_root_closure_), 532 _old_gen_root_closure(old_gen_root_closure_), 533 _par_gen(par_gen_), 534 _task_queues(task_queues_), 535 _terminator(terminator_) 536 {} 537 538 void ParEvacuateFollowersClosure::do_void() { 539 ObjToScanQueue* work_q = par_scan_state()->work_queue(); 540 541 while (true) { 542 543 // Scan to-space and old-gen objs until we run out of both. 544 oop obj_to_scan; 545 par_scan_state()->trim_queues(0); 546 547 // We have no local work, attempt to steal from other threads. 548 549 // attempt to steal work from promoted. 550 if (task_queues()->steal(par_scan_state()->thread_num(), 551 par_scan_state()->hash_seed(), 552 obj_to_scan)) { 553 bool res = work_q->push(obj_to_scan); 554 assert(res, "Empty queue should have room for a push."); 555 556 // if successful, goto Start. 557 continue; 558 559 // try global overflow list. 560 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { 561 continue; 562 } 563 564 // Otherwise, offer termination. 565 par_scan_state()->start_term_time(); 566 if (terminator()->offer_termination()) break; 567 par_scan_state()->end_term_time(); 568 } 569 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, 570 "Broken overflow list?"); 571 // Finish the last termination pause. 572 par_scan_state()->end_term_time(); 573 } 574 575 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* old_gen, 576 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : 577 AbstractGangTask("ParNewGeneration collection"), 578 _gen(gen), _old_gen(old_gen), 579 _young_old_boundary(young_old_boundary), 580 _state_set(state_set) 581 {} 582 583 // Reset the terminator for the given number of 584 // active threads. 585 void ParNewGenTask::set_for_termination(int active_workers) { 586 _state_set->reset(active_workers, _gen->promotion_failed()); 587 // Should the heap be passed in? There's only 1 for now so 588 // grab it instead. 589 GenCollectedHeap* gch = GenCollectedHeap::heap(); 590 gch->set_n_termination(active_workers); 591 } 592 593 void ParNewGenTask::work(uint worker_id) { 594 GenCollectedHeap* gch = GenCollectedHeap::heap(); 595 // Since this is being done in a separate thread, need new resource 596 // and handle marks. 597 ResourceMark rm; 598 HandleMark hm; 599 // We would need multiple old-gen queues otherwise. 600 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); 601 602 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); 603 assert(_state_set->is_valid(worker_id), "Should not have been called"); 604 605 par_scan_state.set_young_old_boundary(_young_old_boundary); 606 607 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), 608 gch->rem_set()->klass_rem_set()); 609 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure, 610 &par_scan_state.to_space_root_closure(), 611 false); 612 613 par_scan_state.start_strong_roots(); 614 gch->gen_process_roots(_gen->level(), 615 true, // Process younger gens, if any, 616 // as strong roots. 617 false, // no scope; this is parallel code 618 GenCollectedHeap::SO_ScavengeCodeCache, 619 GenCollectedHeap::StrongAndWeakRoots, 620 &par_scan_state.to_space_root_closure(), 621 &par_scan_state.older_gen_closure(), 622 &cld_scan_closure); 623 624 par_scan_state.end_strong_roots(); 625 626 // "evacuate followers". 627 par_scan_state.evacuate_followers_closure().do_void(); 628 } 629 630 #ifdef _MSC_VER 631 #pragma warning( push ) 632 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 633 #endif 634 ParNewGeneration:: 635 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) 636 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), 637 _overflow_list(NULL), 638 _is_alive_closure(this), 639 _plab_stats(YoungPLABSize, PLABWeight) 640 { 641 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) 642 NOT_PRODUCT(_num_par_pushes = 0;) 643 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); 644 guarantee(_task_queues != NULL, "task_queues allocation failure."); 645 646 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { 647 ObjToScanQueue *q = new ObjToScanQueue(); 648 guarantee(q != NULL, "work_queue Allocation failure."); 649 _task_queues->register_queue(i1, q); 650 } 651 652 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) 653 _task_queues->queue(i2)->initialize(); 654 655 _overflow_stacks = NULL; 656 if (ParGCUseLocalOverflow) { 657 658 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal 659 // with ',' 660 typedef Stack<oop, mtGC> GCOopStack; 661 662 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); 663 for (size_t i = 0; i < ParallelGCThreads; ++i) { 664 new (_overflow_stacks + i) Stack<oop, mtGC>(); 665 } 666 } 667 668 if (UsePerfData) { 669 EXCEPTION_MARK; 670 ResourceMark rm; 671 672 const char* cname = 673 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); 674 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, 675 ParallelGCThreads, CHECK); 676 } 677 } 678 #ifdef _MSC_VER 679 #pragma warning( pop ) 680 #endif 681 682 // ParNewGeneration:: 683 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : 684 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} 685 686 template <class T> 687 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { 688 #ifdef ASSERT 689 { 690 assert(!oopDesc::is_null(*p), "expected non-null ref"); 691 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 692 // We never expect to see a null reference being processed 693 // as a weak reference. 694 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 695 } 696 #endif // ASSERT 697 698 _par_cl->do_oop_nv(p); 699 700 if (Universe::heap()->is_in_reserved(p)) { 701 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 702 _rs->write_ref_field_gc_par(p, obj); 703 } 704 } 705 706 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } 707 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } 708 709 // ParNewGeneration:: 710 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : 711 DefNewGeneration::KeepAliveClosure(cl) {} 712 713 template <class T> 714 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { 715 #ifdef ASSERT 716 { 717 assert(!oopDesc::is_null(*p), "expected non-null ref"); 718 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 719 // We never expect to see a null reference being processed 720 // as a weak reference. 721 assert(obj->is_oop(), "expected an oop while scanning weak refs"); 722 } 723 #endif // ASSERT 724 725 _cl->do_oop_nv(p); 726 727 if (Universe::heap()->is_in_reserved(p)) { 728 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 729 _rs->write_ref_field_gc_par(p, obj); 730 } 731 } 732 733 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } 734 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } 735 736 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { 737 T heap_oop = oopDesc::load_heap_oop(p); 738 if (!oopDesc::is_null(heap_oop)) { 739 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 740 if ((HeapWord*)obj < _boundary) { 741 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); 742 oop new_obj = obj->is_forwarded() 743 ? obj->forwardee() 744 : _g->DefNewGeneration::copy_to_survivor_space(obj); 745 oopDesc::encode_store_heap_oop_not_null(p, new_obj); 746 } 747 if (_gc_barrier) { 748 // If p points to a younger generation, mark the card. 749 if ((HeapWord*)obj < _gen_boundary) { 750 _rs->write_ref_field_gc_par(p, obj); 751 } 752 } 753 } 754 } 755 756 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 757 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } 758 759 class ParNewRefProcTaskProxy: public AbstractGangTask { 760 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 761 public: 762 ParNewRefProcTaskProxy(ProcessTask& task, 763 ParNewGeneration& gen, 764 Generation& old_gen, 765 HeapWord* young_old_boundary, 766 ParScanThreadStateSet& state_set); 767 768 private: 769 virtual void work(uint worker_id); 770 virtual void set_for_termination(int active_workers) { 771 _state_set.terminator()->reset_for_reuse(active_workers); 772 } 773 private: 774 ParNewGeneration& _gen; 775 ProcessTask& _task; 776 Generation& _old_gen; 777 HeapWord* _young_old_boundary; 778 ParScanThreadStateSet& _state_set; 779 }; 780 781 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task, 782 ParNewGeneration& gen, 783 Generation& old_gen, 784 HeapWord* young_old_boundary, 785 ParScanThreadStateSet& state_set) 786 : AbstractGangTask("ParNewGeneration parallel reference processing"), 787 _gen(gen), 788 _task(task), 789 _old_gen(old_gen), 790 _young_old_boundary(young_old_boundary), 791 _state_set(state_set) 792 { 793 } 794 795 void ParNewRefProcTaskProxy::work(uint worker_id) 796 { 797 ResourceMark rm; 798 HandleMark hm; 799 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); 800 par_scan_state.set_young_old_boundary(_young_old_boundary); 801 _task.work(worker_id, par_scan_state.is_alive_closure(), 802 par_scan_state.keep_alive_closure(), 803 par_scan_state.evacuate_followers_closure()); 804 } 805 806 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { 807 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 808 EnqueueTask& _task; 809 810 public: 811 ParNewRefEnqueueTaskProxy(EnqueueTask& task) 812 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), 813 _task(task) 814 { } 815 816 virtual void work(uint worker_id) 817 { 818 _task.work(worker_id); 819 } 820 }; 821 822 823 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) 824 { 825 GenCollectedHeap* gch = GenCollectedHeap::heap(); 826 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 827 "not a generational heap"); 828 FlexibleWorkGang* workers = gch->workers(); 829 assert(workers != NULL, "Need parallel worker threads."); 830 _state_set.reset(workers->active_workers(), _generation.promotion_failed()); 831 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), 832 _generation.reserved().end(), _state_set); 833 workers->run_task(&rp_task); 834 _state_set.reset(0 /* bad value in debug if not reset */, 835 _generation.promotion_failed()); 836 } 837 838 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) 839 { 840 GenCollectedHeap* gch = GenCollectedHeap::heap(); 841 FlexibleWorkGang* workers = gch->workers(); 842 assert(workers != NULL, "Need parallel worker threads."); 843 ParNewRefEnqueueTaskProxy enq_task(task); 844 workers->run_task(&enq_task); 845 } 846 847 void ParNewRefProcTaskExecutor::set_single_threaded_mode() 848 { 849 _state_set.flush(); 850 GenCollectedHeap* gch = GenCollectedHeap::heap(); 851 gch->set_par_threads(0); // 0 ==> non-parallel. 852 gch->save_marks(); 853 } 854 855 ScanClosureWithParBarrier:: 856 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : 857 ScanClosure(g, gc_barrier) {} 858 859 EvacuateFollowersClosureGeneral:: 860 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, 861 OopsInGenClosure* cur, 862 OopsInGenClosure* older) : 863 _gch(gch), _level(level), 864 _scan_cur_or_nonheap(cur), _scan_older(older) 865 {} 866 867 void EvacuateFollowersClosureGeneral::do_void() { 868 do { 869 // Beware: this call will lead to closure applications via virtual 870 // calls. 871 _gch->oop_since_save_marks_iterate(_level, 872 _scan_cur_or_nonheap, 873 _scan_older); 874 } while (!_gch->no_allocs_since_save_marks(_level)); 875 } 876 877 878 // A Generation that does parallel young-gen collection. 879 880 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) { 881 assert(_promo_failure_scan_stack.is_empty(), "post condition"); 882 _promo_failure_scan_stack.clear(true); // Clear cached segments. 883 884 remove_forwarding_pointers(); 885 if (PrintGCDetails) { 886 gclog_or_tty->print(" (promotion failed)"); 887 } 888 // All the spaces are in play for mark-sweep. 889 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. 890 from()->set_next_compaction_space(to()); 891 gch->set_incremental_collection_failed(); 892 // Inform the next generation that a promotion failure occurred. 893 _old_gen->promotion_failure_occurred(); 894 895 // Trace promotion failure in the parallel GC threads 896 thread_state_set.trace_promotion_failed(gc_tracer()); 897 // Single threaded code may have reported promotion failure to the global state 898 if (_promotion_failed_info.has_failed()) { 899 _gc_tracer.report_promotion_failed(_promotion_failed_info); 900 } 901 // Reset the PromotionFailureALot counters. 902 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) 903 } 904 905 void ParNewGeneration::collect(bool full, 906 bool clear_all_soft_refs, 907 size_t size, 908 bool is_tlab) { 909 assert(full || size > 0, "otherwise we don't want to collect"); 910 911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 912 913 _gc_timer->register_gc_start(); 914 915 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 916 "not a CMS generational heap"); 917 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); 918 FlexibleWorkGang* workers = gch->workers(); 919 assert(workers != NULL, "Need workgang for parallel work"); 920 int active_workers = 921 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), 922 workers->active_workers(), 923 Threads::number_of_non_daemon_threads()); 924 workers->set_active_workers(active_workers); 925 assert(gch->n_gens() == 2, 926 "Par collection currently only works with single older gen."); 927 _old_gen = gch->old_gen(); 928 929 // If the next generation is too full to accommodate worst-case promotion 930 // from this generation, pass on collection; let the next generation 931 // do it. 932 if (!collection_attempt_is_safe()) { 933 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one 934 return; 935 } 936 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); 937 938 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); 939 gch->trace_heap_before_gc(gc_tracer()); 940 941 init_assuming_no_promotion_failure(); 942 943 if (UseAdaptiveSizePolicy) { 944 set_survivor_overflow(false); 945 size_policy->minor_collection_begin(); 946 } 947 948 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id()); 949 // Capture heap used before collection (for printing). 950 size_t gch_prev_used = gch->used(); 951 952 age_table()->clear(); 953 to()->clear(SpaceDecorator::Mangle); 954 955 gch->save_marks(); 956 assert(workers != NULL, "Need parallel worker threads."); 957 int n_workers = active_workers; 958 959 // Set the correct parallelism (number of queues) in the reference processor 960 ref_processor()->set_active_mt_degree(n_workers); 961 962 // Always set the terminator for the active number of workers 963 // because only those workers go through the termination protocol. 964 ParallelTaskTerminator _term(n_workers, task_queues()); 965 ParScanThreadStateSet thread_state_set(workers->active_workers(), 966 *to(), *this, *_old_gen, *task_queues(), 967 _overflow_stacks, desired_plab_sz(), _term); 968 969 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set); 970 gch->set_par_threads(n_workers); 971 gch->rem_set()->prepare_for_younger_refs_iterate(true); 972 // It turns out that even when we're using 1 thread, doing the work in a 973 // separate thread causes wide variance in run times. We can't help this 974 // in the multi-threaded case, but we special-case n=1 here to get 975 // repeatable measurements of the 1-thread overhead of the parallel code. 976 if (n_workers > 1) { 977 StrongRootsScope srs; 978 workers->run_task(&tsk); 979 } else { 980 StrongRootsScope srs; 981 tsk.work(0); 982 } 983 thread_state_set.reset(0 /* Bad value in debug if not reset */, 984 promotion_failed()); 985 986 // Trace and reset failed promotion info. 987 if (promotion_failed()) { 988 thread_state_set.trace_promotion_failed(gc_tracer()); 989 } 990 991 // Process (weak) reference objects found during scavenge. 992 ReferenceProcessor* rp = ref_processor(); 993 IsAliveClosure is_alive(this); 994 ScanWeakRefClosure scan_weak_ref(this); 995 KeepAliveClosure keep_alive(&scan_weak_ref); 996 ScanClosure scan_without_gc_barrier(this, false); 997 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); 998 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); 999 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, 1000 &scan_without_gc_barrier, &scan_with_gc_barrier); 1001 rp->setup_policy(clear_all_soft_refs); 1002 // Can the mt_degree be set later (at run_task() time would be best)? 1003 rp->set_active_mt_degree(active_workers); 1004 ReferenceProcessorStats stats; 1005 if (rp->processing_is_mt()) { 1006 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); 1007 stats = rp->process_discovered_references(&is_alive, &keep_alive, 1008 &evacuate_followers, &task_executor, 1009 _gc_timer, _gc_tracer.gc_id()); 1010 } else { 1011 thread_state_set.flush(); 1012 gch->set_par_threads(0); // 0 ==> non-parallel. 1013 gch->save_marks(); 1014 stats = rp->process_discovered_references(&is_alive, &keep_alive, 1015 &evacuate_followers, NULL, 1016 _gc_timer, _gc_tracer.gc_id()); 1017 } 1018 _gc_tracer.report_gc_reference_stats(stats); 1019 if (!promotion_failed()) { 1020 // Swap the survivor spaces. 1021 eden()->clear(SpaceDecorator::Mangle); 1022 from()->clear(SpaceDecorator::Mangle); 1023 if (ZapUnusedHeapArea) { 1024 // This is now done here because of the piece-meal mangling which 1025 // can check for valid mangling at intermediate points in the 1026 // collection(s). When a minor collection fails to collect 1027 // sufficient space resizing of the young generation can occur 1028 // an redistribute the spaces in the young generation. Mangle 1029 // here so that unzapped regions don't get distributed to 1030 // other spaces. 1031 to()->mangle_unused_area(); 1032 } 1033 swap_spaces(); 1034 1035 // A successful scavenge should restart the GC time limit count which is 1036 // for full GC's. 1037 size_policy->reset_gc_overhead_limit_count(); 1038 1039 assert(to()->is_empty(), "to space should be empty now"); 1040 1041 adjust_desired_tenuring_threshold(); 1042 } else { 1043 handle_promotion_failed(gch, thread_state_set); 1044 } 1045 // set new iteration safe limit for the survivor spaces 1046 from()->set_concurrent_iteration_safe_limit(from()->top()); 1047 to()->set_concurrent_iteration_safe_limit(to()->top()); 1048 1049 if (ResizePLAB) { 1050 plab_stats()->adjust_desired_plab_sz(n_workers); 1051 } 1052 1053 if (PrintGC && !PrintGCDetails) { 1054 gch->print_heap_change(gch_prev_used); 1055 } 1056 1057 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats()); 1058 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats()); 1059 1060 if (UseAdaptiveSizePolicy) { 1061 size_policy->minor_collection_end(gch->gc_cause()); 1062 size_policy->avg_survived()->sample(from()->used()); 1063 } 1064 1065 // We need to use a monotonically non-decreasing time in ms 1066 // or we will see time-warp warnings and os::javaTimeMillis() 1067 // does not guarantee monotonicity. 1068 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 1069 update_time_of_last_gc(now); 1070 1071 rp->set_enqueuing_is_done(true); 1072 if (rp->processing_is_mt()) { 1073 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); 1074 rp->enqueue_discovered_references(&task_executor); 1075 } else { 1076 rp->enqueue_discovered_references(NULL); 1077 } 1078 rp->verify_no_references_recorded(); 1079 1080 gch->trace_heap_after_gc(gc_tracer()); 1081 _gc_tracer.report_tenuring_threshold(tenuring_threshold()); 1082 1083 _gc_timer->register_gc_end(); 1084 1085 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); 1086 } 1087 1088 static int sum; 1089 void ParNewGeneration::waste_some_time() { 1090 for (int i = 0; i < 100; i++) { 1091 sum += i; 1092 } 1093 } 1094 1095 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); 1096 1097 // Because of concurrency, there are times where an object for which 1098 // "is_forwarded()" is true contains an "interim" forwarding pointer 1099 // value. Such a value will soon be overwritten with a real value. 1100 // This method requires "obj" to have a forwarding pointer, and waits, if 1101 // necessary for a real one to be inserted, and returns it. 1102 1103 oop ParNewGeneration::real_forwardee(oop obj) { 1104 oop forward_ptr = obj->forwardee(); 1105 if (forward_ptr != ClaimedForwardPtr) { 1106 return forward_ptr; 1107 } else { 1108 return real_forwardee_slow(obj); 1109 } 1110 } 1111 1112 oop ParNewGeneration::real_forwardee_slow(oop obj) { 1113 // Spin-read if it is claimed but not yet written by another thread. 1114 oop forward_ptr = obj->forwardee(); 1115 while (forward_ptr == ClaimedForwardPtr) { 1116 waste_some_time(); 1117 assert(obj->is_forwarded(), "precondition"); 1118 forward_ptr = obj->forwardee(); 1119 } 1120 return forward_ptr; 1121 } 1122 1123 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { 1124 if (m->must_be_preserved_for_promotion_failure(obj)) { 1125 // We should really have separate per-worker stacks, rather 1126 // than use locking of a common pair of stacks. 1127 MutexLocker ml(ParGCRareEvent_lock); 1128 preserve_mark(obj, m); 1129 } 1130 } 1131 1132 // Multiple GC threads may try to promote an object. If the object 1133 // is successfully promoted, a forwarding pointer will be installed in 1134 // the object in the young generation. This method claims the right 1135 // to install the forwarding pointer before it copies the object, 1136 // thus avoiding the need to undo the copy as in 1137 // copy_to_survivor_space_avoiding_with_undo. 1138 1139 oop ParNewGeneration::copy_to_survivor_space( 1140 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { 1141 // In the sequential version, this assert also says that the object is 1142 // not forwarded. That might not be the case here. It is the case that 1143 // the caller observed it to be not forwarded at some time in the past. 1144 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); 1145 1146 // The sequential code read "old->age()" below. That doesn't work here, 1147 // since the age is in the mark word, and that might be overwritten with 1148 // a forwarding pointer by a parallel thread. So we must save the mark 1149 // word in a local and then analyze it. 1150 oopDesc dummyOld; 1151 dummyOld.set_mark(m); 1152 assert(!dummyOld.is_forwarded(), 1153 "should not be called with forwarding pointer mark word."); 1154 1155 oop new_obj = NULL; 1156 oop forward_ptr; 1157 1158 // Try allocating obj in to-space (unless too old) 1159 if (dummyOld.age() < tenuring_threshold()) { 1160 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); 1161 if (new_obj == NULL) { 1162 set_survivor_overflow(true); 1163 } 1164 } 1165 1166 if (new_obj == NULL) { 1167 // Either to-space is full or we decided to promote 1168 // try allocating obj tenured 1169 1170 // Attempt to install a null forwarding pointer (atomically), 1171 // to claim the right to install the real forwarding pointer. 1172 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); 1173 if (forward_ptr != NULL) { 1174 // someone else beat us to it. 1175 return real_forwardee(old); 1176 } 1177 1178 if (!_promotion_failed) { 1179 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), 1180 old, m, sz); 1181 } 1182 1183 if (new_obj == NULL) { 1184 // promotion failed, forward to self 1185 _promotion_failed = true; 1186 new_obj = old; 1187 1188 preserve_mark_if_necessary(old, m); 1189 par_scan_state->register_promotion_failure(sz); 1190 } 1191 1192 old->forward_to(new_obj); 1193 forward_ptr = NULL; 1194 } else { 1195 // Is in to-space; do copying ourselves. 1196 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); 1197 assert(Universe::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value."); 1198 forward_ptr = old->forward_to_atomic(new_obj); 1199 // Restore the mark word copied above. 1200 new_obj->set_mark(m); 1201 // Increment age if obj still in new generation 1202 new_obj->incr_age(); 1203 par_scan_state->age_table()->add(new_obj, sz); 1204 } 1205 assert(new_obj != NULL, "just checking"); 1206 1207 #ifndef PRODUCT 1208 // This code must come after the CAS test, or it will print incorrect 1209 // information. 1210 if (TraceScavenge) { 1211 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", 1212 is_in_reserved(new_obj) ? "copying" : "tenuring", 1213 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size()); 1214 } 1215 #endif 1216 1217 if (forward_ptr == NULL) { 1218 oop obj_to_push = new_obj; 1219 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { 1220 // Length field used as index of next element to be scanned. 1221 // Real length can be obtained from real_forwardee() 1222 arrayOop(old)->set_length(0); 1223 obj_to_push = old; 1224 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, 1225 "push forwarded object"); 1226 } 1227 // Push it on one of the queues of to-be-scanned objects. 1228 bool simulate_overflow = false; 1229 NOT_PRODUCT( 1230 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { 1231 // simulate a stack overflow 1232 simulate_overflow = true; 1233 } 1234 ) 1235 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { 1236 // Add stats for overflow pushes. 1237 if (Verbose && PrintGCDetails) { 1238 gclog_or_tty->print("queue overflow!\n"); 1239 } 1240 push_on_overflow_list(old, par_scan_state); 1241 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); 1242 } 1243 1244 return new_obj; 1245 } 1246 1247 // Oops. Someone beat us to it. Undo the allocation. Where did we 1248 // allocate it? 1249 if (is_in_reserved(new_obj)) { 1250 // Must be in to_space. 1251 assert(to()->is_in_reserved(new_obj), "Checking"); 1252 if (forward_ptr == ClaimedForwardPtr) { 1253 // Wait to get the real forwarding pointer value. 1254 forward_ptr = real_forwardee(old); 1255 } 1256 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); 1257 } 1258 1259 return forward_ptr; 1260 } 1261 1262 #ifndef PRODUCT 1263 // It's OK to call this multi-threaded; the worst thing 1264 // that can happen is that we'll get a bunch of closely 1265 // spaced simulated overflows, but that's OK, in fact 1266 // probably good as it would exercise the overflow code 1267 // under contention. 1268 bool ParNewGeneration::should_simulate_overflow() { 1269 if (_overflow_counter-- <= 0) { // just being defensive 1270 _overflow_counter = ParGCWorkQueueOverflowInterval; 1271 return true; 1272 } else { 1273 return false; 1274 } 1275 } 1276 #endif 1277 1278 // In case we are using compressed oops, we need to be careful. 1279 // If the object being pushed is an object array, then its length 1280 // field keeps track of the "grey boundary" at which the next 1281 // incremental scan will be done (see ParGCArrayScanChunk). 1282 // When using compressed oops, this length field is kept in the 1283 // lower 32 bits of the erstwhile klass word and cannot be used 1284 // for the overflow chaining pointer (OCP below). As such the OCP 1285 // would itself need to be compressed into the top 32-bits in this 1286 // case. Unfortunately, see below, in the event that we have a 1287 // promotion failure, the node to be pushed on the list can be 1288 // outside of the Java heap, so the heap-based pointer compression 1289 // would not work (we would have potential aliasing between C-heap 1290 // and Java-heap pointers). For this reason, when using compressed 1291 // oops, we simply use a worker-thread-local, non-shared overflow 1292 // list in the form of a growable array, with a slightly different 1293 // overflow stack draining strategy. If/when we start using fat 1294 // stacks here, we can go back to using (fat) pointer chains 1295 // (although some performance comparisons would be useful since 1296 // single global lists have their own performance disadvantages 1297 // as we were made painfully aware not long ago, see 6786503). 1298 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 1299 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { 1300 assert(is_in_reserved(from_space_obj), "Should be from this generation"); 1301 if (ParGCUseLocalOverflow) { 1302 // In the case of compressed oops, we use a private, not-shared 1303 // overflow stack. 1304 par_scan_state->push_on_overflow_stack(from_space_obj); 1305 } else { 1306 assert(!UseCompressedOops, "Error"); 1307 // if the object has been forwarded to itself, then we cannot 1308 // use the klass pointer for the linked list. Instead we have 1309 // to allocate an oopDesc in the C-Heap and use that for the linked list. 1310 // XXX This is horribly inefficient when a promotion failure occurs 1311 // and should be fixed. XXX FIX ME !!! 1312 #ifndef PRODUCT 1313 Atomic::inc_ptr(&_num_par_pushes); 1314 assert(_num_par_pushes > 0, "Tautology"); 1315 #endif 1316 if (from_space_obj->forwardee() == from_space_obj) { 1317 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); 1318 listhead->forward_to(from_space_obj); 1319 from_space_obj = listhead; 1320 } 1321 oop observed_overflow_list = _overflow_list; 1322 oop cur_overflow_list; 1323 do { 1324 cur_overflow_list = observed_overflow_list; 1325 if (cur_overflow_list != BUSY) { 1326 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); 1327 } else { 1328 from_space_obj->set_klass_to_list_ptr(NULL); 1329 } 1330 observed_overflow_list = 1331 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); 1332 } while (cur_overflow_list != observed_overflow_list); 1333 } 1334 } 1335 1336 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { 1337 bool res; 1338 1339 if (ParGCUseLocalOverflow) { 1340 res = par_scan_state->take_from_overflow_stack(); 1341 } else { 1342 assert(!UseCompressedOops, "Error"); 1343 res = take_from_overflow_list_work(par_scan_state); 1344 } 1345 return res; 1346 } 1347 1348 1349 // *NOTE*: The overflow list manipulation code here and 1350 // in CMSCollector:: are very similar in shape, 1351 // except that in the CMS case we thread the objects 1352 // directly into the list via their mark word, and do 1353 // not need to deal with special cases below related 1354 // to chunking of object arrays and promotion failure 1355 // handling. 1356 // CR 6797058 has been filed to attempt consolidation of 1357 // the common code. 1358 // Because of the common code, if you make any changes in 1359 // the code below, please check the CMS version to see if 1360 // similar changes might be needed. 1361 // See CMSCollector::par_take_from_overflow_list() for 1362 // more extensive documentation comments. 1363 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { 1364 ObjToScanQueue* work_q = par_scan_state->work_queue(); 1365 // How many to take? 1366 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 1367 (size_t)ParGCDesiredObjsFromOverflowList); 1368 1369 assert(!UseCompressedOops, "Error"); 1370 assert(par_scan_state->overflow_stack() == NULL, "Error"); 1371 if (_overflow_list == NULL) return false; 1372 1373 // Otherwise, there was something there; try claiming the list. 1374 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1375 // Trim off a prefix of at most objsFromOverflow items 1376 Thread* tid = Thread::current(); 1377 size_t spin_count = (size_t)ParallelGCThreads; 1378 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); 1379 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { 1380 // someone grabbed it before we did ... 1381 // ... we spin for a short while... 1382 os::sleep(tid, sleep_time_millis, false); 1383 if (_overflow_list == NULL) { 1384 // nothing left to take 1385 return false; 1386 } else if (_overflow_list != BUSY) { 1387 // try and grab the prefix 1388 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 1389 } 1390 } 1391 if (prefix == NULL || prefix == BUSY) { 1392 // Nothing to take or waited long enough 1393 if (prefix == NULL) { 1394 // Write back the NULL in case we overwrote it with BUSY above 1395 // and it is still the same value. 1396 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1397 } 1398 return false; 1399 } 1400 assert(prefix != NULL && prefix != BUSY, "Error"); 1401 size_t i = 1; 1402 oop cur = prefix; 1403 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { 1404 i++; cur = cur->list_ptr_from_klass(); 1405 } 1406 1407 // Reattach remaining (suffix) to overflow list 1408 if (cur->klass_or_null() == NULL) { 1409 // Write back the NULL in lieu of the BUSY we wrote 1410 // above and it is still the same value. 1411 if (_overflow_list == BUSY) { 1412 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 1413 } 1414 } else { 1415 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); 1416 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list 1417 cur->set_klass_to_list_ptr(NULL); // break off suffix 1418 // It's possible that the list is still in the empty(busy) state 1419 // we left it in a short while ago; in that case we may be 1420 // able to place back the suffix. 1421 oop observed_overflow_list = _overflow_list; 1422 oop cur_overflow_list = observed_overflow_list; 1423 bool attached = false; 1424 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 1425 observed_overflow_list = 1426 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1427 if (cur_overflow_list == observed_overflow_list) { 1428 attached = true; 1429 break; 1430 } else cur_overflow_list = observed_overflow_list; 1431 } 1432 if (!attached) { 1433 // Too bad, someone else got in in between; we'll need to do a splice. 1434 // Find the last item of suffix list 1435 oop last = suffix; 1436 while (last->klass_or_null() != NULL) { 1437 last = last->list_ptr_from_klass(); 1438 } 1439 // Atomically prepend suffix to current overflow list 1440 observed_overflow_list = _overflow_list; 1441 do { 1442 cur_overflow_list = observed_overflow_list; 1443 if (cur_overflow_list != BUSY) { 1444 // Do the splice ... 1445 last->set_klass_to_list_ptr(cur_overflow_list); 1446 } else { // cur_overflow_list == BUSY 1447 last->set_klass_to_list_ptr(NULL); 1448 } 1449 observed_overflow_list = 1450 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); 1451 } while (cur_overflow_list != observed_overflow_list); 1452 } 1453 } 1454 1455 // Push objects on prefix list onto this thread's work queue 1456 assert(prefix != NULL && prefix != BUSY, "program logic"); 1457 cur = prefix; 1458 ssize_t n = 0; 1459 while (cur != NULL) { 1460 oop obj_to_push = cur->forwardee(); 1461 oop next = cur->list_ptr_from_klass(); 1462 cur->set_klass(obj_to_push->klass()); 1463 // This may be an array object that is self-forwarded. In that case, the list pointer 1464 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. 1465 if (!is_in_reserved(cur)) { 1466 // This can become a scaling bottleneck when there is work queue overflow coincident 1467 // with promotion failure. 1468 oopDesc* f = cur; 1469 FREE_C_HEAP_ARRAY(oopDesc, f); 1470 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { 1471 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); 1472 obj_to_push = cur; 1473 } 1474 bool ok = work_q->push(obj_to_push); 1475 assert(ok, "Should have succeeded"); 1476 cur = next; 1477 n++; 1478 } 1479 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); 1480 #ifndef PRODUCT 1481 assert(_num_par_pushes >= n, "Too many pops?"); 1482 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 1483 #endif 1484 return true; 1485 } 1486 #undef BUSY 1487 1488 void ParNewGeneration::ref_processor_init() { 1489 if (_ref_processor == NULL) { 1490 // Allocate and initialize a reference processor 1491 _ref_processor = 1492 new ReferenceProcessor(_reserved, // span 1493 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 1494 (int) ParallelGCThreads, // mt processing degree 1495 refs_discovery_is_mt(), // mt discovery 1496 (int) ParallelGCThreads, // mt discovery degree 1497 refs_discovery_is_atomic(), // atomic_discovery 1498 NULL); // is_alive_non_header 1499 } 1500 } 1501 1502 const char* ParNewGeneration::name() const { 1503 return "par new generation"; 1504 }