mark_sweep.cc revision df386c551405ce9668e827584f744c6f098761fa
1/* 2 * Copyright (C) 2011 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include "mark_sweep.h" 18 19#include <functional> 20#include <numeric> 21#include <climits> 22#include <vector> 23 24#include "base/bounded_fifo.h" 25#include "base/logging.h" 26#include "base/macros.h" 27#include "base/mutex-inl.h" 28#include "base/timing_logger.h" 29#include "gc/accounting/card_table-inl.h" 30#include "gc/accounting/heap_bitmap-inl.h" 31#include "gc/accounting/mod_union_table.h" 32#include "gc/accounting/space_bitmap-inl.h" 33#include "gc/heap.h" 34#include "gc/space/image_space.h" 35#include "gc/space/large_object_space.h" 36#include "gc/space/space-inl.h" 37#include "mark_sweep-inl.h" 38#include "mirror/art_field-inl.h" 39#include "mirror/object-inl.h" 40#include "runtime.h" 41#include "scoped_thread_state_change.h" 42#include "thread-inl.h" 43#include "thread_list.h" 44 45using ::art::mirror::ArtField; 46using ::art::mirror::Class; 47using ::art::mirror::Object; 48using ::art::mirror::ObjectArray; 49 50namespace art { 51namespace gc { 52namespace collector { 53 54// Performance options. 55static constexpr bool kUseRecursiveMark = false; 56static constexpr bool kUseMarkStackPrefetch = true; 57static constexpr size_t kSweepArrayChunkFreeSize = 1024; 58static constexpr bool kPreCleanCards = true; 59 60// Parallelism options. 61static constexpr bool kParallelCardScan = true; 62static constexpr bool kParallelRecursiveMark = true; 63// Don't attempt to parallelize mark stack processing unless the mark stack is at least n 64// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not 65// having this can add overhead in ProcessReferences since we may end up doing many calls of 66// ProcessMarkStack with very small mark stacks. 67static constexpr size_t kMinimumParallelMarkStackSize = 128; 68static constexpr bool kParallelProcessMarkStack = true; 69 70// Profiling and information flags. 71static constexpr bool kProfileLargeObjects = false; 72static constexpr bool kMeasureOverhead = false; 73static constexpr bool kCountTasks = false; 74static constexpr bool kCountJavaLangRefs = false; 75static constexpr bool kCountMarkedObjects = false; 76 77// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%. 78static constexpr bool kCheckLocks = kDebugLocking; 79static constexpr bool kVerifyRootsMarked = kIsDebugBuild; 80 81// If true, revoke the rosalloc thread-local buffers at the 82// checkpoint, as opposed to during the pause. 83static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true; 84 85void MarkSweep::BindBitmaps() { 86 timings_.StartSplit("BindBitmaps"); 87 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 88 // Mark all of the spaces we never collect as immune. 89 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 90 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 91 CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space; 92 } 93 } 94 timings_.EndSplit(); 95} 96 97MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix) 98 : GarbageCollector(heap, 99 name_prefix + 100 (is_concurrent ? "concurrent mark sweep": "mark sweep")), 101 gc_barrier_(new Barrier(0)), 102 large_object_lock_("mark sweep large object lock", kMarkSweepLargeObjectLock), 103 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 104 is_concurrent_(is_concurrent) { 105} 106 107void MarkSweep::InitializePhase() { 108 timings_.Reset(); 109 TimingLogger::ScopedSplit split("InitializePhase", &timings_); 110 mark_stack_ = heap_->mark_stack_.get(); 111 DCHECK(mark_stack_ != nullptr); 112 immune_region_.Reset(); 113 class_count_ = 0; 114 array_count_ = 0; 115 other_count_ = 0; 116 large_object_test_ = 0; 117 large_object_mark_ = 0; 118 overhead_time_ = 0; 119 work_chunks_created_ = 0; 120 work_chunks_deleted_ = 0; 121 reference_count_ = 0; 122 mark_null_count_ = 0; 123 mark_immune_count_ = 0; 124 mark_fastpath_count_ = 0; 125 mark_slowpath_count_ = 0; 126 FindDefaultSpaceBitmap(); 127 { 128 // TODO: I don't think we should need heap bitmap lock to get the mark bitmap. 129 ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 130 mark_bitmap_ = heap_->GetMarkBitmap(); 131 } 132 if (!clear_soft_references_) { 133 // Always clear soft references if a non-sticky collection. 134 clear_soft_references_ = GetGcType() != collector::kGcTypeSticky; 135 } 136 // Do any pre GC verification. 137 timings_.NewSplit("PreGcVerification"); 138 heap_->PreGcVerification(this); 139} 140 141void MarkSweep::ProcessReferences(Thread* self) { 142 TimingLogger::ScopedSplit split("ProcessReferences", &timings_); 143 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 144 GetHeap()->ProcessReferences(timings_, clear_soft_references_, &IsMarkedCallback, 145 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 146} 147 148void MarkSweep::PreProcessReferences() { 149 if (IsConcurrent()) { 150 // No reason to do this for non-concurrent GC since pre processing soft references only helps 151 // pauses. 152 timings_.NewSplit("PreProcessReferences"); 153 GetHeap()->ProcessSoftReferences(timings_, clear_soft_references_, &IsMarkedCallback, 154 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 155 } 156} 157 158void MarkSweep::PausePhase() { 159 TimingLogger::ScopedSplit split("(Paused)PausePhase", &timings_); 160 Thread* self = Thread::Current(); 161 Locks::mutator_lock_->AssertExclusiveHeld(self); 162 if (IsConcurrent()) { 163 // Handle the dirty objects if we are a concurrent GC. 164 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 165 // Re-mark root set. 166 ReMarkRoots(); 167 // Scan dirty objects, this is only required if we are not doing concurrent GC. 168 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 169 } 170 ProcessReferences(self); 171 { 172 timings_.NewSplit("SwapStacks"); 173 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 174 heap_->SwapStacks(self); 175 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 176 // Need to revoke all the thread local allocation stacks since we just swapped the allocation 177 // stacks and don't want anybody to allocate into the live stack. 178 RevokeAllThreadLocalAllocationStacks(self); 179 } 180 timings_.StartSplit("PreSweepingGcVerification"); 181 heap_->PreSweepingGcVerification(this); 182 timings_.EndSplit(); 183 if (IsConcurrent()) { 184 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 185 // weak before we sweep them. Since this new system weak may not be marked, the GC may 186 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 187 // reference to a string that is about to be swept. 188 Runtime::Current()->DisallowNewSystemWeaks(); 189 } 190} 191 192void MarkSweep::PreCleanCards() { 193 // Don't do this for non concurrent GCs since they don't have any dirty cards. 194 if (kPreCleanCards && IsConcurrent()) { 195 Thread* self = Thread::Current(); 196 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 197 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 198 heap_->ProcessCards(timings_, false); 199 // The checkpoint root marking is required to avoid a race condition which occurs if the 200 // following happens during a reference write: 201 // 1. mutator dirties the card (write barrier) 202 // 2. GC ages the card (the above ProcessCards call) 203 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 204 // 4. mutator writes the value (corresponding to the write barrier in 1.) 205 // This causes the GC to age the card but not necessarily mark the reference which the mutator 206 // wrote into the object stored in the card. 207 // Having the checkpoint fixes this issue since it ensures that the card mark and the 208 // reference write are visible to the GC before the card is scanned (this is due to locks being 209 // acquired / released in the checkpoint code). 210 // The other roots are also marked to help reduce the pause. 211 MarkRootsCheckpoint(self, false); 212 MarkNonThreadRoots(); 213 MarkConcurrentRoots( 214 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 215 // Process the newly aged cards. 216 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 217 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 218 // in the next GC. 219 } 220} 221 222void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 223 if (kUseThreadLocalAllocationStack) { 224 timings_.NewSplit("RevokeAllThreadLocalAllocationStacks"); 225 Locks::mutator_lock_->AssertExclusiveHeld(self); 226 heap_->RevokeAllThreadLocalAllocationStacks(self); 227 } 228} 229 230void MarkSweep::MarkingPhase() { 231 TimingLogger::ScopedSplit split("MarkingPhase", &timings_); 232 Thread* self = Thread::Current(); 233 234 BindBitmaps(); 235 FindDefaultSpaceBitmap(); 236 237 // Process dirty cards and add dirty cards to mod union tables. 238 heap_->ProcessCards(timings_, false); 239 240 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 241 MarkRoots(self); 242 MarkReachableObjects(); 243 // Pre-clean dirtied cards to reduce pauses. 244 PreCleanCards(); 245 PreProcessReferences(); 246} 247 248void MarkSweep::UpdateAndMarkModUnion() { 249 for (const auto& space : heap_->GetContinuousSpaces()) { 250 if (immune_region_.ContainsSpace(space)) { 251 const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" : 252 "UpdateAndMarkImageModUnionTable"; 253 TimingLogger::ScopedSplit split(name, &timings_); 254 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 255 CHECK(mod_union_table != nullptr); 256 mod_union_table->UpdateAndMarkReferences(MarkHeapReferenceCallback, this); 257 } 258 } 259} 260 261void MarkSweep::MarkReachableObjects() { 262 UpdateAndMarkModUnion(); 263 // Recursively mark all the non-image bits set in the mark bitmap. 264 RecursiveMark(); 265} 266 267void MarkSweep::ReclaimPhase() { 268 TimingLogger::ScopedSplit split("ReclaimPhase", &timings_); 269 Thread* self = Thread::Current(); 270 SweepSystemWeaks(self); 271 if (IsConcurrent()) { 272 Runtime::Current()->AllowNewSystemWeaks(); 273 } 274 { 275 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 276 277 // Reclaim unmarked objects. 278 Sweep(false); 279 280 // Swap the live and mark bitmaps for each space which we modified space. This is an 281 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 282 // bitmaps. 283 timings_.StartSplit("SwapBitmaps"); 284 SwapBitmaps(); 285 timings_.EndSplit(); 286 287 // Unbind the live and mark bitmaps. 288 TimingLogger::ScopedSplit split("UnBindBitmaps", &timings_); 289 GetHeap()->UnBindBitmaps(); 290 } 291} 292 293void MarkSweep::FindDefaultSpaceBitmap() { 294 TimingLogger::ScopedSplit split("FindDefaultMarkBitmap", &timings_); 295 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 296 accounting::SpaceBitmap* bitmap = space->GetMarkBitmap(); 297 if (bitmap != nullptr && 298 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 299 current_space_bitmap_ = bitmap; 300 return; 301 } 302 } 303 GetHeap()->DumpSpaces(); 304 LOG(FATAL) << "Could not find a default mark bitmap"; 305} 306 307void MarkSweep::ExpandMarkStack() { 308 ResizeMarkStack(mark_stack_->Capacity() * 2); 309} 310 311void MarkSweep::ResizeMarkStack(size_t new_size) { 312 // Rare case, no need to have Thread::Current be a parameter. 313 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 314 // Someone else acquired the lock and expanded the mark stack before us. 315 return; 316 } 317 std::vector<Object*> temp(mark_stack_->Begin(), mark_stack_->End()); 318 CHECK_LE(mark_stack_->Size(), new_size); 319 mark_stack_->Resize(new_size); 320 for (const auto& obj : temp) { 321 mark_stack_->PushBack(obj); 322 } 323} 324 325inline void MarkSweep::MarkObjectNonNullParallel(Object* obj) { 326 DCHECK(obj != NULL); 327 if (MarkObjectParallel(obj)) { 328 MutexLock mu(Thread::Current(), mark_stack_lock_); 329 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 330 ExpandMarkStack(); 331 } 332 // The object must be pushed on to the mark stack. 333 mark_stack_->PushBack(obj); 334 } 335} 336 337mirror::Object* MarkSweep::MarkObjectCallback(mirror::Object* obj, void* arg) { 338 MarkSweep* mark_sweep = reinterpret_cast<MarkSweep*>(arg); 339 mark_sweep->MarkObject(obj); 340 return obj; 341} 342 343void MarkSweep::MarkHeapReferenceCallback(mirror::HeapReference<mirror::Object>* ref, void* arg) { 344 reinterpret_cast<MarkSweep*>(arg)->MarkObject(ref->AsMirrorPtr()); 345} 346 347inline void MarkSweep::MarkObjectNonNull(Object* obj) { 348 DCHECK(obj != nullptr); 349 if (kUseBakerOrBrooksReadBarrier) { 350 // Verify all the objects have the correct pointer installed. 351 obj->AssertReadBarrierPointer(); 352 } 353 if (immune_region_.ContainsObject(obj)) { 354 if (kCountMarkedObjects) { 355 ++mark_immune_count_; 356 } 357 DCHECK(IsMarked(obj)); 358 return; 359 } 360 // Try to take advantage of locality of references within a space, failing this find the space 361 // the hard way. 362 accounting::SpaceBitmap* object_bitmap = current_space_bitmap_; 363 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 364 object_bitmap = mark_bitmap_->GetContinuousSpaceBitmap(obj); 365 if (kCountMarkedObjects) { 366 ++mark_slowpath_count_; 367 } 368 if (UNLIKELY(object_bitmap == nullptr)) { 369 MarkLargeObject(obj, true); 370 return; 371 } 372 } else if (kCountMarkedObjects) { 373 ++mark_fastpath_count_; 374 } 375 // This object was not previously marked. 376 if (!object_bitmap->Set(obj)) { 377 PushOnMarkStack(obj); 378 } 379} 380 381inline void MarkSweep::PushOnMarkStack(Object* obj) { 382 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 383 // Lock is not needed but is here anyways to please annotalysis. 384 MutexLock mu(Thread::Current(), mark_stack_lock_); 385 ExpandMarkStack(); 386 } 387 // The object must be pushed on to the mark stack. 388 mark_stack_->PushBack(obj); 389} 390 391// Rare case, probably not worth inlining since it will increase instruction cache miss rate. 392bool MarkSweep::MarkLargeObject(const Object* obj, bool set) { 393 // TODO: support >1 discontinuous space. 394 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 395 accounting::ObjectSet* large_objects = large_object_space->GetMarkObjects(); 396 if (kProfileLargeObjects) { 397 ++large_object_test_; 398 } 399 if (UNLIKELY(!large_objects->Test(obj))) { 400 if (!large_object_space->Contains(obj)) { 401 LOG(ERROR) << "Tried to mark " << obj << " not contained by any spaces"; 402 LOG(ERROR) << "Attempting see if it's a bad root"; 403 VerifyRoots(); 404 LOG(FATAL) << "Can't mark bad root"; 405 } 406 if (kProfileLargeObjects) { 407 ++large_object_mark_; 408 } 409 if (set) { 410 large_objects->Set(obj); 411 } else { 412 large_objects->Clear(obj); 413 } 414 return true; 415 } 416 return false; 417} 418 419inline bool MarkSweep::MarkObjectParallel(const Object* obj) { 420 DCHECK(obj != nullptr); 421 if (kUseBakerOrBrooksReadBarrier) { 422 // Verify all the objects have the correct pointer installed. 423 obj->AssertReadBarrierPointer(); 424 } 425 if (immune_region_.ContainsObject(obj)) { 426 DCHECK(IsMarked(obj)); 427 return false; 428 } 429 // Try to take advantage of locality of references within a space, failing this find the space 430 // the hard way. 431 accounting::SpaceBitmap* object_bitmap = current_space_bitmap_; 432 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 433 accounting::SpaceBitmap* new_bitmap = mark_bitmap_->GetContinuousSpaceBitmap(obj); 434 if (new_bitmap != NULL) { 435 object_bitmap = new_bitmap; 436 } else { 437 // TODO: Remove the Thread::Current here? 438 // TODO: Convert this to some kind of atomic marking? 439 MutexLock mu(Thread::Current(), large_object_lock_); 440 return MarkLargeObject(obj, true); 441 } 442 } 443 // Return true if the object was not previously marked. 444 return !object_bitmap->AtomicTestAndSet(obj); 445} 446 447// Used to mark objects when processing the mark stack. If an object is null, it is not marked. 448inline void MarkSweep::MarkObject(Object* obj) { 449 if (obj != nullptr) { 450 MarkObjectNonNull(obj); 451 } else if (kCountMarkedObjects) { 452 ++mark_null_count_; 453 } 454} 455 456void MarkSweep::MarkRootParallelCallback(Object** root, void* arg, uint32_t /*thread_id*/, 457 RootType /*root_type*/) { 458 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNullParallel(*root); 459} 460 461void MarkSweep::VerifyRootMarked(Object** root, void* arg, uint32_t /*thread_id*/, 462 RootType /*root_type*/) { 463 CHECK(reinterpret_cast<MarkSweep*>(arg)->IsMarked(*root)); 464} 465 466void MarkSweep::MarkRootCallback(Object** root, void* arg, uint32_t /*thread_id*/, 467 RootType /*root_type*/) { 468 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNull(*root); 469} 470 471void MarkSweep::VerifyRootCallback(const Object* root, void* arg, size_t vreg, 472 const StackVisitor* visitor, RootType root_type) { 473 reinterpret_cast<MarkSweep*>(arg)->VerifyRoot(root, vreg, visitor, root_type); 474} 475 476void MarkSweep::VerifyRoot(const Object* root, size_t vreg, const StackVisitor* visitor, 477 RootType root_type) { 478 // See if the root is on any space bitmap. 479 if (GetHeap()->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) { 480 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 481 if (!large_object_space->Contains(root)) { 482 LOG(ERROR) << "Found invalid root: " << root << " with type " << root_type; 483 if (visitor != NULL) { 484 LOG(ERROR) << visitor->DescribeLocation() << " in VReg: " << vreg; 485 } 486 } 487 } 488} 489 490void MarkSweep::VerifyRoots() { 491 Runtime::Current()->GetThreadList()->VerifyRoots(VerifyRootCallback, this); 492} 493 494void MarkSweep::MarkRoots(Thread* self) { 495 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 496 // If we exclusively hold the mutator lock, all threads must be suspended. 497 timings_.StartSplit("MarkRoots"); 498 Runtime::Current()->VisitRoots(MarkRootCallback, this); 499 timings_.EndSplit(); 500 RevokeAllThreadLocalAllocationStacks(self); 501 } else { 502 MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint); 503 // At this point the live stack should no longer have any mutators which push into it. 504 MarkNonThreadRoots(); 505 MarkConcurrentRoots( 506 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 507 } 508} 509 510void MarkSweep::MarkNonThreadRoots() { 511 timings_.StartSplit("MarkNonThreadRoots"); 512 Runtime::Current()->VisitNonThreadRoots(MarkRootCallback, this); 513 timings_.EndSplit(); 514} 515 516void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 517 timings_.StartSplit("MarkConcurrentRoots"); 518 // Visit all runtime roots and clear dirty flags. 519 Runtime::Current()->VisitConcurrentRoots(MarkRootCallback, this, flags); 520 timings_.EndSplit(); 521} 522 523class ScanObjectVisitor { 524 public: 525 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 526 : mark_sweep_(mark_sweep) {} 527 528 void operator()(Object* obj) const ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 529 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 530 if (kCheckLocks) { 531 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 532 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 533 } 534 mark_sweep_->ScanObject(obj); 535 } 536 537 private: 538 MarkSweep* const mark_sweep_; 539}; 540 541class DelayReferenceReferentVisitor { 542 public: 543 explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) { 544 } 545 546 void operator()(mirror::Class* klass, mirror::Reference* ref) const 547 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 548 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 549 collector_->DelayReferenceReferent(klass, ref); 550 } 551 552 private: 553 MarkSweep* const collector_; 554}; 555 556template <bool kUseFinger = false> 557class MarkStackTask : public Task { 558 public: 559 MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size, 560 Object** mark_stack) 561 : mark_sweep_(mark_sweep), 562 thread_pool_(thread_pool), 563 mark_stack_pos_(mark_stack_size) { 564 // We may have to copy part of an existing mark stack when another mark stack overflows. 565 if (mark_stack_size != 0) { 566 DCHECK(mark_stack != NULL); 567 // TODO: Check performance? 568 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 569 } 570 if (kCountTasks) { 571 ++mark_sweep_->work_chunks_created_; 572 } 573 } 574 575 static const size_t kMaxSize = 1 * KB; 576 577 protected: 578 class MarkObjectParallelVisitor { 579 public: 580 explicit MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task, 581 MarkSweep* mark_sweep) ALWAYS_INLINE 582 : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {} 583 584 void operator()(Object* obj, MemberOffset offset, bool /* static */) const ALWAYS_INLINE 585 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 586 mirror::Object* ref = obj->GetFieldObject<mirror::Object>(offset, false); 587 if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) { 588 if (kUseFinger) { 589 android_memory_barrier(); 590 if (reinterpret_cast<uintptr_t>(ref) >= 591 static_cast<uintptr_t>(mark_sweep_->atomic_finger_)) { 592 return; 593 } 594 } 595 chunk_task_->MarkStackPush(ref); 596 } 597 } 598 599 private: 600 MarkStackTask<kUseFinger>* const chunk_task_; 601 MarkSweep* const mark_sweep_; 602 }; 603 604 class ScanObjectParallelVisitor { 605 public: 606 explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE 607 : chunk_task_(chunk_task) {} 608 609 // No thread safety analysis since multiple threads will use this visitor. 610 void operator()(Object* obj) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 611 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 612 MarkSweep* const mark_sweep = chunk_task_->mark_sweep_; 613 MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep); 614 DelayReferenceReferentVisitor ref_visitor(mark_sweep); 615 mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor); 616 } 617 618 private: 619 MarkStackTask<kUseFinger>* const chunk_task_; 620 }; 621 622 virtual ~MarkStackTask() { 623 // Make sure that we have cleared our mark stack. 624 DCHECK_EQ(mark_stack_pos_, 0U); 625 if (kCountTasks) { 626 ++mark_sweep_->work_chunks_deleted_; 627 } 628 } 629 630 MarkSweep* const mark_sweep_; 631 ThreadPool* const thread_pool_; 632 // Thread local mark stack for this task. 633 Object* mark_stack_[kMaxSize]; 634 // Mark stack position. 635 size_t mark_stack_pos_; 636 637 void MarkStackPush(Object* obj) ALWAYS_INLINE { 638 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 639 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 640 mark_stack_pos_ /= 2; 641 auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_, 642 mark_stack_ + mark_stack_pos_); 643 thread_pool_->AddTask(Thread::Current(), task); 644 } 645 DCHECK(obj != nullptr); 646 DCHECK_LT(mark_stack_pos_, kMaxSize); 647 mark_stack_[mark_stack_pos_++] = obj; 648 } 649 650 virtual void Finalize() { 651 delete this; 652 } 653 654 // Scans all of the objects 655 virtual void Run(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 656 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 657 ScanObjectParallelVisitor visitor(this); 658 // TODO: Tune this. 659 static const size_t kFifoSize = 4; 660 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 661 for (;;) { 662 Object* obj = nullptr; 663 if (kUseMarkStackPrefetch) { 664 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 665 Object* obj = mark_stack_[--mark_stack_pos_]; 666 DCHECK(obj != nullptr); 667 __builtin_prefetch(obj); 668 prefetch_fifo.push_back(obj); 669 } 670 if (UNLIKELY(prefetch_fifo.empty())) { 671 break; 672 } 673 obj = prefetch_fifo.front(); 674 prefetch_fifo.pop_front(); 675 } else { 676 if (UNLIKELY(mark_stack_pos_ == 0)) { 677 break; 678 } 679 obj = mark_stack_[--mark_stack_pos_]; 680 } 681 DCHECK(obj != nullptr); 682 visitor(obj); 683 } 684 } 685}; 686 687class CardScanTask : public MarkStackTask<false> { 688 public: 689 CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, accounting::SpaceBitmap* bitmap, 690 byte* begin, byte* end, byte minimum_age, size_t mark_stack_size, 691 Object** mark_stack_obj) 692 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 693 bitmap_(bitmap), 694 begin_(begin), 695 end_(end), 696 minimum_age_(minimum_age) { 697 } 698 699 protected: 700 accounting::SpaceBitmap* const bitmap_; 701 byte* const begin_; 702 byte* const end_; 703 const byte minimum_age_; 704 705 virtual void Finalize() { 706 delete this; 707 } 708 709 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 710 ScanObjectParallelVisitor visitor(this); 711 accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable(); 712 size_t cards_scanned = card_table->Scan(bitmap_, begin_, end_, visitor, minimum_age_); 713 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 714 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 715 // Finish by emptying our local mark stack. 716 MarkStackTask::Run(self); 717 } 718}; 719 720size_t MarkSweep::GetThreadCount(bool paused) const { 721 if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) { 722 return 0; 723 } 724 if (paused) { 725 return heap_->GetParallelGCThreadCount() + 1; 726 } else { 727 return heap_->GetConcGCThreadCount() + 1; 728 } 729} 730 731void MarkSweep::ScanGrayObjects(bool paused, byte minimum_age) { 732 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 733 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 734 size_t thread_count = GetThreadCount(paused); 735 // The parallel version with only one thread is faster for card scanning, TODO: fix. 736 if (kParallelCardScan && thread_count > 0) { 737 Thread* self = Thread::Current(); 738 // Can't have a different split for each space since multiple spaces can have their cards being 739 // scanned at the same time. 740 timings_.StartSplit(paused ? "(Paused)ScanGrayObjects" : "ScanGrayObjects"); 741 // Try to take some of the mark stack since we can pass this off to the worker tasks. 742 Object** mark_stack_begin = mark_stack_->Begin(); 743 Object** mark_stack_end = mark_stack_->End(); 744 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 745 // Estimated number of work tasks we will create. 746 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 747 DCHECK_NE(mark_stack_tasks, 0U); 748 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 749 mark_stack_size / mark_stack_tasks + 1); 750 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 751 if (space->GetMarkBitmap() == nullptr) { 752 continue; 753 } 754 byte* card_begin = space->Begin(); 755 byte* card_end = space->End(); 756 // Align up the end address. For example, the image space's end 757 // may not be card-size-aligned. 758 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 759 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin)); 760 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end)); 761 // Calculate how many bytes of heap we will scan, 762 const size_t address_range = card_end - card_begin; 763 // Calculate how much address range each task gets. 764 const size_t card_delta = RoundUp(address_range / thread_count + 1, 765 accounting::CardTable::kCardSize); 766 // Create the worker tasks for this space. 767 while (card_begin != card_end) { 768 // Add a range of cards. 769 size_t addr_remaining = card_end - card_begin; 770 size_t card_increment = std::min(card_delta, addr_remaining); 771 // Take from the back of the mark stack. 772 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 773 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 774 mark_stack_end -= mark_stack_increment; 775 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 776 DCHECK_EQ(mark_stack_end, mark_stack_->End()); 777 // Add the new task to the thread pool. 778 auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin, 779 card_begin + card_increment, minimum_age, 780 mark_stack_increment, mark_stack_end); 781 thread_pool->AddTask(self, task); 782 card_begin += card_increment; 783 } 784 } 785 786 // Note: the card scan below may dirty new cards (and scan them) 787 // as a side effect when a Reference object is encountered and 788 // queued during the marking. See b/11465268. 789 thread_pool->SetMaxActiveWorkers(thread_count - 1); 790 thread_pool->StartWorkers(self); 791 thread_pool->Wait(self, true, true); 792 thread_pool->StopWorkers(self); 793 timings_.EndSplit(); 794 } else { 795 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 796 if (space->GetMarkBitmap() != nullptr) { 797 // Image spaces are handled properly since live == marked for them. 798 switch (space->GetGcRetentionPolicy()) { 799 case space::kGcRetentionPolicyNeverCollect: 800 timings_.StartSplit(paused ? "(Paused)ScanGrayImageSpaceObjects" : 801 "ScanGrayImageSpaceObjects"); 802 break; 803 case space::kGcRetentionPolicyFullCollect: 804 timings_.StartSplit(paused ? "(Paused)ScanGrayZygoteSpaceObjects" : 805 "ScanGrayZygoteSpaceObjects"); 806 break; 807 case space::kGcRetentionPolicyAlwaysCollect: 808 timings_.StartSplit(paused ? "(Paused)ScanGrayAllocSpaceObjects" : 809 "ScanGrayAllocSpaceObjects"); 810 break; 811 } 812 ScanObjectVisitor visitor(this); 813 card_table->Scan(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, minimum_age); 814 timings_.EndSplit(); 815 } 816 } 817 } 818} 819 820class RecursiveMarkTask : public MarkStackTask<false> { 821 public: 822 RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 823 accounting::SpaceBitmap* bitmap, uintptr_t begin, uintptr_t end) 824 : MarkStackTask<false>(thread_pool, mark_sweep, 0, NULL), 825 bitmap_(bitmap), 826 begin_(begin), 827 end_(end) { 828 } 829 830 protected: 831 accounting::SpaceBitmap* const bitmap_; 832 const uintptr_t begin_; 833 const uintptr_t end_; 834 835 virtual void Finalize() { 836 delete this; 837 } 838 839 // Scans all of the objects 840 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 841 ScanObjectParallelVisitor visitor(this); 842 bitmap_->VisitMarkedRange(begin_, end_, visitor); 843 // Finish by emptying our local mark stack. 844 MarkStackTask::Run(self); 845 } 846}; 847 848// Populates the mark stack based on the set of marked objects and 849// recursively marks until the mark stack is emptied. 850void MarkSweep::RecursiveMark() { 851 TimingLogger::ScopedSplit split("RecursiveMark", &timings_); 852 // RecursiveMark will build the lists of known instances of the Reference classes. See 853 // DelayReferenceReferent for details. 854 if (kUseRecursiveMark) { 855 const bool partial = GetGcType() == kGcTypePartial; 856 ScanObjectVisitor scan_visitor(this); 857 auto* self = Thread::Current(); 858 ThreadPool* thread_pool = heap_->GetThreadPool(); 859 size_t thread_count = GetThreadCount(false); 860 const bool parallel = kParallelRecursiveMark && thread_count > 1; 861 mark_stack_->Reset(); 862 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 863 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 864 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 865 current_space_bitmap_ = space->GetMarkBitmap(); 866 if (current_space_bitmap_ == nullptr) { 867 continue; 868 } 869 if (parallel) { 870 // We will use the mark stack the future. 871 // CHECK(mark_stack_->IsEmpty()); 872 // This function does not handle heap end increasing, so we must use the space end. 873 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 874 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 875 atomic_finger_ = static_cast<int32_t>(0xFFFFFFFF); 876 877 // Create a few worker tasks. 878 const size_t n = thread_count * 2; 879 while (begin != end) { 880 uintptr_t start = begin; 881 uintptr_t delta = (end - begin) / n; 882 delta = RoundUp(delta, KB); 883 if (delta < 16 * KB) delta = end - begin; 884 begin += delta; 885 auto* task = new RecursiveMarkTask(thread_pool, this, current_space_bitmap_, start, 886 begin); 887 thread_pool->AddTask(self, task); 888 } 889 thread_pool->SetMaxActiveWorkers(thread_count - 1); 890 thread_pool->StartWorkers(self); 891 thread_pool->Wait(self, true, true); 892 thread_pool->StopWorkers(self); 893 } else { 894 // This function does not handle heap end increasing, so we must use the space end. 895 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 896 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 897 current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 898 } 899 } 900 } 901 } 902 ProcessMarkStack(false); 903} 904 905mirror::Object* MarkSweep::IsMarkedCallback(mirror::Object* object, void* arg) { 906 if (reinterpret_cast<MarkSweep*>(arg)->IsMarked(object)) { 907 return object; 908 } 909 return nullptr; 910} 911 912void MarkSweep::RecursiveMarkDirtyObjects(bool paused, byte minimum_age) { 913 ScanGrayObjects(paused, minimum_age); 914 ProcessMarkStack(paused); 915} 916 917void MarkSweep::ReMarkRoots() { 918 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 919 timings_.StartSplit("(Paused)ReMarkRoots"); 920 Runtime::Current()->VisitRoots( 921 MarkRootCallback, this, static_cast<VisitRootFlags>(kVisitRootFlagNewRoots | 922 kVisitRootFlagStopLoggingNewRoots | 923 kVisitRootFlagClearRootLog)); 924 timings_.EndSplit(); 925 if (kVerifyRootsMarked) { 926 timings_.StartSplit("(Paused)VerifyRoots"); 927 Runtime::Current()->VisitRoots(VerifyRootMarked, this); 928 timings_.EndSplit(); 929 } 930} 931 932void MarkSweep::SweepSystemWeaks(Thread* self) { 933 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 934 timings_.StartSplit("SweepSystemWeaks"); 935 Runtime::Current()->SweepSystemWeaks(IsMarkedCallback, this); 936 timings_.EndSplit(); 937} 938 939mirror::Object* MarkSweep::VerifySystemWeakIsLiveCallback(Object* obj, void* arg) { 940 reinterpret_cast<MarkSweep*>(arg)->VerifyIsLive(obj); 941 // We don't actually want to sweep the object, so lets return "marked" 942 return obj; 943} 944 945void MarkSweep::VerifyIsLive(const Object* obj) { 946 if (!heap_->GetLiveBitmap()->Test(obj)) { 947 space::LargeObjectSpace* large_object_space = heap_->GetLargeObjectsSpace(); 948 if (!large_object_space->GetLiveObjects()->Test(obj)) { 949 if (std::find(heap_->allocation_stack_->Begin(), heap_->allocation_stack_->End(), obj) == 950 heap_->allocation_stack_->End()) { 951 // Object not found! 952 heap_->DumpSpaces(); 953 LOG(FATAL) << "Found dead object " << obj; 954 } 955 } 956 } 957} 958 959void MarkSweep::VerifySystemWeaks() { 960 // Verify system weaks, uses a special object visitor which returns the input object. 961 Runtime::Current()->SweepSystemWeaks(VerifySystemWeakIsLiveCallback, this); 962} 963 964class CheckpointMarkThreadRoots : public Closure { 965 public: 966 explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep, 967 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) 968 : mark_sweep_(mark_sweep), 969 revoke_ros_alloc_thread_local_buffers_at_checkpoint_( 970 revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 971 } 972 973 virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS { 974 ATRACE_BEGIN("Marking thread roots"); 975 // Note: self is not necessarily equal to thread since thread may be suspended. 976 Thread* self = Thread::Current(); 977 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 978 << thread->GetState() << " thread " << thread << " self " << self; 979 thread->VisitRoots(MarkSweep::MarkRootParallelCallback, mark_sweep_); 980 ATRACE_END(); 981 if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) { 982 ATRACE_BEGIN("RevokeRosAllocThreadLocalBuffers"); 983 mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread); 984 ATRACE_END(); 985 } 986 mark_sweep_->GetBarrier().Pass(self); 987 } 988 989 private: 990 MarkSweep* const mark_sweep_; 991 const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_; 992}; 993 994void MarkSweep::MarkRootsCheckpoint(Thread* self, 995 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 996 CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint); 997 timings_.StartSplit("MarkRootsCheckpoint"); 998 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 999 // Request the check point is run on all threads returning a count of the threads that must 1000 // run through the barrier including self. 1001 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1002 // Release locks then wait for all mutator threads to pass the barrier. 1003 // TODO: optimize to not release locks when there are no threads to wait for. 1004 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1005 Locks::mutator_lock_->SharedUnlock(self); 1006 { 1007 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun); 1008 gc_barrier_->Increment(self, barrier_count); 1009 } 1010 Locks::mutator_lock_->SharedLock(self); 1011 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1012 timings_.EndSplit(); 1013} 1014 1015void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1016 timings_.StartSplit("SweepArray"); 1017 Thread* self = Thread::Current(); 1018 mirror::Object* chunk_free_buffer[kSweepArrayChunkFreeSize]; 1019 size_t chunk_free_pos = 0; 1020 size_t freed_bytes = 0; 1021 size_t freed_large_object_bytes = 0; 1022 size_t freed_objects = 0; 1023 size_t freed_large_objects = 0; 1024 // How many objects are left in the array, modified after each space is swept. 1025 Object** objects = allocations->Begin(); 1026 size_t count = allocations->Size(); 1027 // Change the order to ensure that the non-moving space last swept as an optimization. 1028 std::vector<space::ContinuousSpace*> sweep_spaces; 1029 space::ContinuousSpace* non_moving_space = nullptr; 1030 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1031 if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) && 1032 space->GetLiveBitmap() != nullptr) { 1033 if (space == heap_->GetNonMovingSpace()) { 1034 non_moving_space = space; 1035 } else { 1036 sweep_spaces.push_back(space); 1037 } 1038 } 1039 } 1040 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1041 // the other alloc spaces as an optimization. 1042 if (non_moving_space != nullptr) { 1043 sweep_spaces.push_back(non_moving_space); 1044 } 1045 // Start by sweeping the continuous spaces. 1046 for (space::ContinuousSpace* space : sweep_spaces) { 1047 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1048 accounting::SpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1049 accounting::SpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1050 if (swap_bitmaps) { 1051 std::swap(live_bitmap, mark_bitmap); 1052 } 1053 Object** out = objects; 1054 for (size_t i = 0; i < count; ++i) { 1055 Object* obj = objects[i]; 1056 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1057 continue; 1058 } 1059 if (space->HasAddress(obj)) { 1060 // This object is in the space, remove it from the array and add it to the sweep buffer 1061 // if needed. 1062 if (!mark_bitmap->Test(obj)) { 1063 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1064 timings_.StartSplit("FreeList"); 1065 freed_objects += chunk_free_pos; 1066 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1067 timings_.EndSplit(); 1068 chunk_free_pos = 0; 1069 } 1070 chunk_free_buffer[chunk_free_pos++] = obj; 1071 } 1072 } else { 1073 *(out++) = obj; 1074 } 1075 } 1076 if (chunk_free_pos > 0) { 1077 timings_.StartSplit("FreeList"); 1078 freed_objects += chunk_free_pos; 1079 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1080 timings_.EndSplit(); 1081 chunk_free_pos = 0; 1082 } 1083 // All of the references which space contained are no longer in the allocation stack, update 1084 // the count. 1085 count = out - objects; 1086 } 1087 // Handle the large object space. 1088 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1089 accounting::ObjectSet* large_live_objects = large_object_space->GetLiveObjects(); 1090 accounting::ObjectSet* large_mark_objects = large_object_space->GetMarkObjects(); 1091 if (swap_bitmaps) { 1092 std::swap(large_live_objects, large_mark_objects); 1093 } 1094 for (size_t i = 0; i < count; ++i) { 1095 Object* obj = objects[i]; 1096 // Handle large objects. 1097 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1098 continue; 1099 } 1100 if (!large_mark_objects->Test(obj)) { 1101 ++freed_large_objects; 1102 freed_large_object_bytes += large_object_space->Free(self, obj); 1103 } 1104 } 1105 timings_.EndSplit(); 1106 1107 timings_.StartSplit("RecordFree"); 1108 VLOG(heap) << "Freed " << freed_objects << "/" << count 1109 << " objects with size " << PrettySize(freed_bytes); 1110 heap_->RecordFree(freed_objects + freed_large_objects, freed_bytes + freed_large_object_bytes); 1111 freed_objects_.FetchAndAdd(freed_objects); 1112 freed_large_objects_.FetchAndAdd(freed_large_objects); 1113 freed_bytes_.FetchAndAdd(freed_bytes); 1114 freed_large_object_bytes_.FetchAndAdd(freed_large_object_bytes); 1115 timings_.EndSplit(); 1116 1117 timings_.StartSplit("ResetStack"); 1118 allocations->Reset(); 1119 timings_.EndSplit(); 1120} 1121 1122void MarkSweep::Sweep(bool swap_bitmaps) { 1123 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 1124 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 1125 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 1126 // knowing that new allocations won't be marked as live. 1127 timings_.StartSplit("MarkStackAsLive"); 1128 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 1129 heap_->MarkAllocStackAsLive(live_stack); 1130 live_stack->Reset(); 1131 timings_.EndSplit(); 1132 1133 DCHECK(mark_stack_->IsEmpty()); 1134 TimingLogger::ScopedSplit("Sweep", &timings_); 1135 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1136 if (space->IsContinuousMemMapAllocSpace()) { 1137 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1138 TimingLogger::ScopedSplit split( 1139 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", &timings_); 1140 size_t freed_objects = 0; 1141 size_t freed_bytes = 0; 1142 alloc_space->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1143 heap_->RecordFree(freed_objects, freed_bytes); 1144 freed_objects_.FetchAndAdd(freed_objects); 1145 freed_bytes_.FetchAndAdd(freed_bytes); 1146 } 1147 } 1148 SweepLargeObjects(swap_bitmaps); 1149} 1150 1151void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1152 TimingLogger::ScopedSplit("SweepLargeObjects", &timings_); 1153 size_t freed_objects = 0; 1154 size_t freed_bytes = 0; 1155 GetHeap()->GetLargeObjectsSpace()->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1156 freed_large_objects_.FetchAndAdd(freed_objects); 1157 freed_large_object_bytes_.FetchAndAdd(freed_bytes); 1158 GetHeap()->RecordFree(freed_objects, freed_bytes); 1159} 1160 1161// Process the "referent" field in a java.lang.ref.Reference. If the referent has not yet been 1162// marked, put it on the appropriate list in the heap for later processing. 1163void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) { 1164 DCHECK(klass != nullptr); 1165 if (kCountJavaLangRefs) { 1166 ++reference_count_; 1167 } 1168 heap_->DelayReferenceReferent(klass, ref, IsMarkedCallback, this); 1169} 1170 1171class MarkObjectVisitor { 1172 public: 1173 explicit MarkObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) { 1174 } 1175 1176 void operator()(Object* obj, MemberOffset offset, bool /* is_static */) const 1177 ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 1178 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1179 if (kCheckLocks) { 1180 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1181 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1182 } 1183 mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset, false)); 1184 } 1185 1186 private: 1187 MarkSweep* const mark_sweep_; 1188}; 1189 1190// Scans an object reference. Determines the type of the reference 1191// and dispatches to a specialized scanning routine. 1192void MarkSweep::ScanObject(Object* obj) { 1193 MarkObjectVisitor mark_visitor(this); 1194 DelayReferenceReferentVisitor ref_visitor(this); 1195 ScanObjectVisit(obj, mark_visitor, ref_visitor); 1196} 1197 1198void MarkSweep::ProcessMarkStackPausedCallback(void* arg) { 1199 reinterpret_cast<MarkSweep*>(arg)->ProcessMarkStack(true); 1200} 1201 1202void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1203 Thread* self = Thread::Current(); 1204 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1205 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1206 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1207 CHECK_GT(chunk_size, 0U); 1208 // Split the current mark stack up into work tasks. 1209 for (mirror::Object **it = mark_stack_->Begin(), **end = mark_stack_->End(); it < end; ) { 1210 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1211 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it)); 1212 it += delta; 1213 } 1214 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1215 thread_pool->StartWorkers(self); 1216 thread_pool->Wait(self, true, true); 1217 thread_pool->StopWorkers(self); 1218 mark_stack_->Reset(); 1219 CHECK_EQ(work_chunks_created_, work_chunks_deleted_) << " some of the work chunks were leaked"; 1220} 1221 1222// Scan anything that's on the mark stack. 1223void MarkSweep::ProcessMarkStack(bool paused) { 1224 timings_.StartSplit(paused ? "(Paused)ProcessMarkStack" : "ProcessMarkStack"); 1225 size_t thread_count = GetThreadCount(paused); 1226 if (kParallelProcessMarkStack && thread_count > 1 && 1227 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1228 ProcessMarkStackParallel(thread_count); 1229 } else { 1230 // TODO: Tune this. 1231 static const size_t kFifoSize = 4; 1232 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 1233 for (;;) { 1234 Object* obj = NULL; 1235 if (kUseMarkStackPrefetch) { 1236 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1237 Object* obj = mark_stack_->PopBack(); 1238 DCHECK(obj != NULL); 1239 __builtin_prefetch(obj); 1240 prefetch_fifo.push_back(obj); 1241 } 1242 if (prefetch_fifo.empty()) { 1243 break; 1244 } 1245 obj = prefetch_fifo.front(); 1246 prefetch_fifo.pop_front(); 1247 } else { 1248 if (mark_stack_->IsEmpty()) { 1249 break; 1250 } 1251 obj = mark_stack_->PopBack(); 1252 } 1253 DCHECK(obj != nullptr); 1254 ScanObject(obj); 1255 } 1256 } 1257 timings_.EndSplit(); 1258} 1259 1260inline bool MarkSweep::IsMarked(const Object* object) const 1261 SHARED_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1262 if (immune_region_.ContainsObject(object)) { 1263 return true; 1264 } 1265 if (current_space_bitmap_->HasAddress(object)) { 1266 return current_space_bitmap_->Test(object); 1267 } 1268 return mark_bitmap_->Test(object); 1269} 1270 1271void MarkSweep::FinishPhase() { 1272 TimingLogger::ScopedSplit split("FinishPhase", &timings_); 1273 // Can't enqueue references if we hold the mutator lock. 1274 timings_.NewSplit("PostGcVerification"); 1275 heap_->PostGcVerification(this); 1276 if (kCountScannedTypes) { 1277 VLOG(gc) << "MarkSweep scanned classes=" << class_count_ << " arrays=" << array_count_ 1278 << " other=" << other_count_; 1279 } 1280 if (kCountTasks) { 1281 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_; 1282 } 1283 if (kMeasureOverhead) { 1284 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_); 1285 } 1286 if (kProfileLargeObjects) { 1287 VLOG(gc) << "Large objects tested " << large_object_test_ << " marked " << large_object_mark_; 1288 } 1289 if (kCountJavaLangRefs) { 1290 VLOG(gc) << "References scanned " << reference_count_; 1291 } 1292 if (kCountMarkedObjects) { 1293 VLOG(gc) << "Marked: null=" << mark_null_count_ << " immune=" << mark_immune_count_ 1294 << " fastpath=" << mark_fastpath_count_ << " slowpath=" << mark_slowpath_count_; 1295 } 1296 CHECK(mark_stack_->IsEmpty()); // Ensure that the mark stack is empty. 1297 mark_stack_->Reset(); 1298 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 1299 heap_->ClearMarkedObjects(); 1300} 1301 1302void MarkSweep::RevokeAllThreadLocalBuffers() { 1303 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) { 1304 // If concurrent, rosalloc thread-local buffers are revoked at the 1305 // thread checkpoint. Bump pointer space thread-local buffers must 1306 // not be in use. 1307 GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked(); 1308 } else { 1309 timings_.StartSplit("(Paused)RevokeAllThreadLocalBuffers"); 1310 GetHeap()->RevokeAllThreadLocalBuffers(); 1311 timings_.EndSplit(); 1312 } 1313} 1314 1315} // namespace collector 1316} // namespace gc 1317} // namespace art 1318