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 <atomic> 20#include <functional> 21#include <numeric> 22#include <climits> 23#include <vector> 24 25#include "base/bounded_fifo.h" 26#include "base/enums.h" 27#include "base/logging.h" 28#include "base/macros.h" 29#include "base/mutex-inl.h" 30#include "base/systrace.h" 31#include "base/time_utils.h" 32#include "base/timing_logger.h" 33#include "gc/accounting/card_table-inl.h" 34#include "gc/accounting/heap_bitmap-inl.h" 35#include "gc/accounting/mod_union_table.h" 36#include "gc/accounting/space_bitmap-inl.h" 37#include "gc/heap.h" 38#include "gc/reference_processor.h" 39#include "gc/space/large_object_space.h" 40#include "gc/space/space-inl.h" 41#include "mark_sweep-inl.h" 42#include "mirror/object-inl.h" 43#include "runtime.h" 44#include "scoped_thread_state_change-inl.h" 45#include "thread-inl.h" 46#include "thread_list.h" 47 48namespace art { 49namespace gc { 50namespace collector { 51 52// Performance options. 53static constexpr bool kUseRecursiveMark = false; 54static constexpr bool kUseMarkStackPrefetch = true; 55static constexpr size_t kSweepArrayChunkFreeSize = 1024; 56static constexpr bool kPreCleanCards = true; 57 58// Parallelism options. 59static constexpr bool kParallelCardScan = true; 60static constexpr bool kParallelRecursiveMark = true; 61// Don't attempt to parallelize mark stack processing unless the mark stack is at least n 62// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not 63// having this can add overhead in ProcessReferences since we may end up doing many calls of 64// ProcessMarkStack with very small mark stacks. 65static constexpr size_t kMinimumParallelMarkStackSize = 128; 66static constexpr bool kParallelProcessMarkStack = true; 67 68// Profiling and information flags. 69static constexpr bool kProfileLargeObjects = false; 70static constexpr bool kMeasureOverhead = false; 71static constexpr bool kCountTasks = false; 72static constexpr bool kCountMarkedObjects = false; 73 74// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%. 75static constexpr bool kCheckLocks = kDebugLocking; 76static constexpr bool kVerifyRootsMarked = kIsDebugBuild; 77 78// If true, revoke the rosalloc thread-local buffers at the 79// checkpoint, as opposed to during the pause. 80static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true; 81 82void MarkSweep::BindBitmaps() { 83 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 84 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 85 // Mark all of the spaces we never collect as immune. 86 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 87 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 88 immune_spaces_.AddSpace(space); 89 } 90 } 91} 92 93MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix) 94 : GarbageCollector(heap, 95 name_prefix + 96 (is_concurrent ? "concurrent mark sweep": "mark sweep")), 97 current_space_bitmap_(nullptr), 98 mark_bitmap_(nullptr), 99 mark_stack_(nullptr), 100 gc_barrier_(new Barrier(0)), 101 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 102 is_concurrent_(is_concurrent), 103 live_stack_freeze_size_(0) { 104 std::string error_msg; 105 MemMap* mem_map = MemMap::MapAnonymous( 106 "mark sweep sweep array free buffer", nullptr, 107 RoundUp(kSweepArrayChunkFreeSize * sizeof(mirror::Object*), kPageSize), 108 PROT_READ | PROT_WRITE, false, false, &error_msg); 109 CHECK(mem_map != nullptr) << "Couldn't allocate sweep array free buffer: " << error_msg; 110 sweep_array_free_buffer_mem_map_.reset(mem_map); 111} 112 113void MarkSweep::InitializePhase() { 114 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 115 mark_stack_ = heap_->GetMarkStack(); 116 DCHECK(mark_stack_ != nullptr); 117 immune_spaces_.Reset(); 118 no_reference_class_count_.StoreRelaxed(0); 119 normal_count_.StoreRelaxed(0); 120 class_count_.StoreRelaxed(0); 121 object_array_count_.StoreRelaxed(0); 122 other_count_.StoreRelaxed(0); 123 reference_count_.StoreRelaxed(0); 124 large_object_test_.StoreRelaxed(0); 125 large_object_mark_.StoreRelaxed(0); 126 overhead_time_ .StoreRelaxed(0); 127 work_chunks_created_.StoreRelaxed(0); 128 work_chunks_deleted_.StoreRelaxed(0); 129 mark_null_count_.StoreRelaxed(0); 130 mark_immune_count_.StoreRelaxed(0); 131 mark_fastpath_count_.StoreRelaxed(0); 132 mark_slowpath_count_.StoreRelaxed(0); 133 { 134 // TODO: I don't think we should need heap bitmap lock to Get the mark bitmap. 135 ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 136 mark_bitmap_ = heap_->GetMarkBitmap(); 137 } 138 if (!GetCurrentIteration()->GetClearSoftReferences()) { 139 // Always clear soft references if a non-sticky collection. 140 GetCurrentIteration()->SetClearSoftReferences(GetGcType() != collector::kGcTypeSticky); 141 } 142} 143 144void MarkSweep::RunPhases() { 145 Thread* self = Thread::Current(); 146 InitializePhase(); 147 Locks::mutator_lock_->AssertNotHeld(self); 148 if (IsConcurrent()) { 149 GetHeap()->PreGcVerification(this); 150 { 151 ReaderMutexLock mu(self, *Locks::mutator_lock_); 152 MarkingPhase(); 153 } 154 ScopedPause pause(this); 155 GetHeap()->PrePauseRosAllocVerification(this); 156 PausePhase(); 157 RevokeAllThreadLocalBuffers(); 158 } else { 159 ScopedPause pause(this); 160 GetHeap()->PreGcVerificationPaused(this); 161 MarkingPhase(); 162 GetHeap()->PrePauseRosAllocVerification(this); 163 PausePhase(); 164 RevokeAllThreadLocalBuffers(); 165 } 166 { 167 // Sweeping always done concurrently, even for non concurrent mark sweep. 168 ReaderMutexLock mu(self, *Locks::mutator_lock_); 169 ReclaimPhase(); 170 } 171 GetHeap()->PostGcVerification(this); 172 FinishPhase(); 173} 174 175void MarkSweep::ProcessReferences(Thread* self) { 176 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 177 GetHeap()->GetReferenceProcessor()->ProcessReferences( 178 true, 179 GetTimings(), 180 GetCurrentIteration()->GetClearSoftReferences(), 181 this); 182} 183 184void MarkSweep::PausePhase() { 185 TimingLogger::ScopedTiming t("(Paused)PausePhase", GetTimings()); 186 Thread* self = Thread::Current(); 187 Locks::mutator_lock_->AssertExclusiveHeld(self); 188 if (IsConcurrent()) { 189 // Handle the dirty objects if we are a concurrent GC. 190 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 191 // Re-mark root set. 192 ReMarkRoots(); 193 // Scan dirty objects, this is only required if we are not doing concurrent GC. 194 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 195 } 196 { 197 TimingLogger::ScopedTiming t2("SwapStacks", GetTimings()); 198 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 199 heap_->SwapStacks(); 200 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 201 // Need to revoke all the thread local allocation stacks since we just swapped the allocation 202 // stacks and don't want anybody to allocate into the live stack. 203 RevokeAllThreadLocalAllocationStacks(self); 204 } 205 heap_->PreSweepingGcVerification(this); 206 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 207 // weak before we sweep them. Since this new system weak may not be marked, the GC may 208 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 209 // reference to a string that is about to be swept. 210 Runtime::Current()->DisallowNewSystemWeaks(); 211 // Enable the reference processing slow path, needs to be done with mutators paused since there 212 // is no lock in the GetReferent fast path. 213 GetHeap()->GetReferenceProcessor()->EnableSlowPath(); 214} 215 216void MarkSweep::PreCleanCards() { 217 // Don't do this for non concurrent GCs since they don't have any dirty cards. 218 if (kPreCleanCards && IsConcurrent()) { 219 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 220 Thread* self = Thread::Current(); 221 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 222 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 223 heap_->ProcessCards(GetTimings(), false, true, false); 224 // The checkpoint root marking is required to avoid a race condition which occurs if the 225 // following happens during a reference write: 226 // 1. mutator dirties the card (write barrier) 227 // 2. GC ages the card (the above ProcessCards call) 228 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 229 // 4. mutator writes the value (corresponding to the write barrier in 1.) 230 // This causes the GC to age the card but not necessarily mark the reference which the mutator 231 // wrote into the object stored in the card. 232 // Having the checkpoint fixes this issue since it ensures that the card mark and the 233 // reference write are visible to the GC before the card is scanned (this is due to locks being 234 // acquired / released in the checkpoint code). 235 // The other roots are also marked to help reduce the pause. 236 MarkRootsCheckpoint(self, false); 237 MarkNonThreadRoots(); 238 MarkConcurrentRoots( 239 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 240 // Process the newly aged cards. 241 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 242 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 243 // in the next GC. 244 } 245} 246 247void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 248 if (kUseThreadLocalAllocationStack) { 249 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 250 Locks::mutator_lock_->AssertExclusiveHeld(self); 251 heap_->RevokeAllThreadLocalAllocationStacks(self); 252 } 253} 254 255void MarkSweep::MarkingPhase() { 256 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 257 Thread* self = Thread::Current(); 258 BindBitmaps(); 259 FindDefaultSpaceBitmap(); 260 // Process dirty cards and add dirty cards to mod union tables. 261 // If the GC type is non sticky, then we just clear the cards instead of ageing them. 262 heap_->ProcessCards(GetTimings(), false, true, GetGcType() != kGcTypeSticky); 263 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 264 MarkRoots(self); 265 MarkReachableObjects(); 266 // Pre-clean dirtied cards to reduce pauses. 267 PreCleanCards(); 268} 269 270class MarkSweep::ScanObjectVisitor { 271 public: 272 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 273 : mark_sweep_(mark_sweep) {} 274 275 void operator()(ObjPtr<mirror::Object> obj) const 276 ALWAYS_INLINE 277 REQUIRES(Locks::heap_bitmap_lock_) 278 REQUIRES_SHARED(Locks::mutator_lock_) { 279 if (kCheckLocks) { 280 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 281 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 282 } 283 mark_sweep_->ScanObject(obj.Ptr()); 284 } 285 286 private: 287 MarkSweep* const mark_sweep_; 288}; 289 290void MarkSweep::UpdateAndMarkModUnion() { 291 for (const auto& space : immune_spaces_.GetSpaces()) { 292 const char* name = space->IsZygoteSpace() 293 ? "UpdateAndMarkZygoteModUnionTable" 294 : "UpdateAndMarkImageModUnionTable"; 295 DCHECK(space->IsZygoteSpace() || space->IsImageSpace()) << *space; 296 TimingLogger::ScopedTiming t(name, GetTimings()); 297 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 298 if (mod_union_table != nullptr) { 299 mod_union_table->UpdateAndMarkReferences(this); 300 } else { 301 // No mod-union table, scan all the live bits. This can only occur for app images. 302 space->GetLiveBitmap()->VisitMarkedRange(reinterpret_cast<uintptr_t>(space->Begin()), 303 reinterpret_cast<uintptr_t>(space->End()), 304 ScanObjectVisitor(this)); 305 } 306 } 307} 308 309void MarkSweep::MarkReachableObjects() { 310 UpdateAndMarkModUnion(); 311 // Recursively mark all the non-image bits set in the mark bitmap. 312 RecursiveMark(); 313} 314 315void MarkSweep::ReclaimPhase() { 316 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 317 Thread* const self = Thread::Current(); 318 // Process the references concurrently. 319 ProcessReferences(self); 320 SweepSystemWeaks(self); 321 Runtime* const runtime = Runtime::Current(); 322 runtime->AllowNewSystemWeaks(); 323 // Clean up class loaders after system weaks are swept since that is how we know if class 324 // unloading occurred. 325 runtime->GetClassLinker()->CleanupClassLoaders(); 326 { 327 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 328 GetHeap()->RecordFreeRevoke(); 329 // Reclaim unmarked objects. 330 Sweep(false); 331 // Swap the live and mark bitmaps for each space which we modified space. This is an 332 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 333 // bitmaps. 334 SwapBitmaps(); 335 // Unbind the live and mark bitmaps. 336 GetHeap()->UnBindBitmaps(); 337 } 338} 339 340void MarkSweep::FindDefaultSpaceBitmap() { 341 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 342 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 343 accounting::ContinuousSpaceBitmap* bitmap = space->GetMarkBitmap(); 344 // We want to have the main space instead of non moving if possible. 345 if (bitmap != nullptr && 346 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 347 current_space_bitmap_ = bitmap; 348 // If we are not the non moving space exit the loop early since this will be good enough. 349 if (space != heap_->GetNonMovingSpace()) { 350 break; 351 } 352 } 353 } 354 CHECK(current_space_bitmap_ != nullptr) << "Could not find a default mark bitmap\n" 355 << heap_->DumpSpaces(); 356} 357 358void MarkSweep::ExpandMarkStack() { 359 ResizeMarkStack(mark_stack_->Capacity() * 2); 360} 361 362void MarkSweep::ResizeMarkStack(size_t new_size) { 363 // Rare case, no need to have Thread::Current be a parameter. 364 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 365 // Someone else acquired the lock and expanded the mark stack before us. 366 return; 367 } 368 std::vector<StackReference<mirror::Object>> temp(mark_stack_->Begin(), mark_stack_->End()); 369 CHECK_LE(mark_stack_->Size(), new_size); 370 mark_stack_->Resize(new_size); 371 for (auto& obj : temp) { 372 mark_stack_->PushBack(obj.AsMirrorPtr()); 373 } 374} 375 376mirror::Object* MarkSweep::MarkObject(mirror::Object* obj) { 377 MarkObject(obj, nullptr, MemberOffset(0)); 378 return obj; 379} 380 381inline void MarkSweep::MarkObjectNonNullParallel(mirror::Object* obj) { 382 DCHECK(obj != nullptr); 383 if (MarkObjectParallel(obj)) { 384 MutexLock mu(Thread::Current(), mark_stack_lock_); 385 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 386 ExpandMarkStack(); 387 } 388 // The object must be pushed on to the mark stack. 389 mark_stack_->PushBack(obj); 390 } 391} 392 393bool MarkSweep::IsNullOrMarkedHeapReference(mirror::HeapReference<mirror::Object>* ref, 394 bool do_atomic_update ATTRIBUTE_UNUSED) { 395 mirror::Object* obj = ref->AsMirrorPtr(); 396 if (obj == nullptr) { 397 return true; 398 } 399 return IsMarked(obj); 400} 401 402class MarkSweep::MarkObjectSlowPath { 403 public: 404 explicit MarkObjectSlowPath(MarkSweep* mark_sweep, 405 mirror::Object* holder = nullptr, 406 MemberOffset offset = MemberOffset(0)) 407 : mark_sweep_(mark_sweep), 408 holder_(holder), 409 offset_(offset) {} 410 411 void operator()(const mirror::Object* obj) const NO_THREAD_SAFETY_ANALYSIS { 412 if (kProfileLargeObjects) { 413 // TODO: Differentiate between marking and testing somehow. 414 ++mark_sweep_->large_object_test_; 415 ++mark_sweep_->large_object_mark_; 416 } 417 space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace(); 418 if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) || 419 (kIsDebugBuild && large_object_space != nullptr && 420 !large_object_space->Contains(obj)))) { 421 // Lowest priority logging first: 422 PrintFileToLog("/proc/self/maps", LogSeverity::FATAL_WITHOUT_ABORT); 423 MemMap::DumpMaps(LOG_STREAM(FATAL_WITHOUT_ABORT), true); 424 // Buffer the output in the string stream since it is more important than the stack traces 425 // and we want it to have log priority. The stack traces are printed from Runtime::Abort 426 // which is called from LOG(FATAL) but before the abort message. 427 std::ostringstream oss; 428 oss << "Tried to mark " << obj << " not contained by any spaces" << std::endl; 429 if (holder_ != nullptr) { 430 size_t holder_size = holder_->SizeOf(); 431 ArtField* field = holder_->FindFieldByOffset(offset_); 432 oss << "Field info: " 433 << " holder=" << holder_ 434 << " holder is " 435 << (mark_sweep_->GetHeap()->IsLiveObjectLocked(holder_) 436 ? "alive" : "dead") 437 << " holder_size=" << holder_size 438 << " holder_type=" << holder_->PrettyTypeOf() 439 << " offset=" << offset_.Uint32Value() 440 << " field=" << (field != nullptr ? field->GetName() : "nullptr") 441 << " field_type=" 442 << (field != nullptr ? field->GetTypeDescriptor() : "") 443 << " first_ref_field_offset=" 444 << (holder_->IsClass() 445 ? holder_->AsClass()->GetFirstReferenceStaticFieldOffset( 446 kRuntimePointerSize) 447 : holder_->GetClass()->GetFirstReferenceInstanceFieldOffset()) 448 << " num_of_ref_fields=" 449 << (holder_->IsClass() 450 ? holder_->AsClass()->NumReferenceStaticFields() 451 : holder_->GetClass()->NumReferenceInstanceFields()) 452 << std::endl; 453 // Print the memory content of the holder. 454 for (size_t i = 0; i < holder_size / sizeof(uint32_t); ++i) { 455 uint32_t* p = reinterpret_cast<uint32_t*>(holder_); 456 oss << &p[i] << ": " << "holder+" << (i * sizeof(uint32_t)) << " = " << std::hex << p[i] 457 << std::endl; 458 } 459 } 460 oss << "Attempting see if it's a bad thread root" << std::endl; 461 mark_sweep_->VerifySuspendedThreadRoots(oss); 462 LOG(FATAL) << oss.str(); 463 } 464 } 465 466 private: 467 MarkSweep* const mark_sweep_; 468 mirror::Object* const holder_; 469 MemberOffset offset_; 470}; 471 472inline void MarkSweep::MarkObjectNonNull(mirror::Object* obj, 473 mirror::Object* holder, 474 MemberOffset offset) { 475 DCHECK(obj != nullptr); 476 if (kUseBakerReadBarrier) { 477 // Verify all the objects have the correct state installed. 478 obj->AssertReadBarrierState(); 479 } 480 if (immune_spaces_.IsInImmuneRegion(obj)) { 481 if (kCountMarkedObjects) { 482 ++mark_immune_count_; 483 } 484 DCHECK(mark_bitmap_->Test(obj)); 485 } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) { 486 if (kCountMarkedObjects) { 487 ++mark_fastpath_count_; 488 } 489 if (UNLIKELY(!current_space_bitmap_->Set(obj))) { 490 PushOnMarkStack(obj); // This object was not previously marked. 491 } 492 } else { 493 if (kCountMarkedObjects) { 494 ++mark_slowpath_count_; 495 } 496 MarkObjectSlowPath visitor(this, holder, offset); 497 // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set 498 // will check again. 499 if (!mark_bitmap_->Set(obj, visitor)) { 500 PushOnMarkStack(obj); // Was not already marked, push. 501 } 502 } 503} 504 505inline void MarkSweep::PushOnMarkStack(mirror::Object* obj) { 506 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 507 // Lock is not needed but is here anyways to please annotalysis. 508 MutexLock mu(Thread::Current(), mark_stack_lock_); 509 ExpandMarkStack(); 510 } 511 // The object must be pushed on to the mark stack. 512 mark_stack_->PushBack(obj); 513} 514 515inline bool MarkSweep::MarkObjectParallel(mirror::Object* obj) { 516 DCHECK(obj != nullptr); 517 if (kUseBakerReadBarrier) { 518 // Verify all the objects have the correct state installed. 519 obj->AssertReadBarrierState(); 520 } 521 if (immune_spaces_.IsInImmuneRegion(obj)) { 522 DCHECK(IsMarked(obj) != nullptr); 523 return false; 524 } 525 // Try to take advantage of locality of references within a space, failing this find the space 526 // the hard way. 527 accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_; 528 if (LIKELY(object_bitmap->HasAddress(obj))) { 529 return !object_bitmap->AtomicTestAndSet(obj); 530 } 531 MarkObjectSlowPath visitor(this); 532 return !mark_bitmap_->AtomicTestAndSet(obj, visitor); 533} 534 535void MarkSweep::MarkHeapReference(mirror::HeapReference<mirror::Object>* ref, 536 bool do_atomic_update ATTRIBUTE_UNUSED) { 537 MarkObject(ref->AsMirrorPtr(), nullptr, MemberOffset(0)); 538} 539 540// Used to mark objects when processing the mark stack. If an object is null, it is not marked. 541inline void MarkSweep::MarkObject(mirror::Object* obj, 542 mirror::Object* holder, 543 MemberOffset offset) { 544 if (obj != nullptr) { 545 MarkObjectNonNull(obj, holder, offset); 546 } else if (kCountMarkedObjects) { 547 ++mark_null_count_; 548 } 549} 550 551class MarkSweep::VerifyRootMarkedVisitor : public SingleRootVisitor { 552 public: 553 explicit VerifyRootMarkedVisitor(MarkSweep* collector) : collector_(collector) { } 554 555 void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE 556 REQUIRES_SHARED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 557 CHECK(collector_->IsMarked(root) != nullptr) << info.ToString(); 558 } 559 560 private: 561 MarkSweep* const collector_; 562}; 563 564void MarkSweep::VisitRoots(mirror::Object*** roots, 565 size_t count, 566 const RootInfo& info ATTRIBUTE_UNUSED) { 567 for (size_t i = 0; i < count; ++i) { 568 MarkObjectNonNull(*roots[i]); 569 } 570} 571 572void MarkSweep::VisitRoots(mirror::CompressedReference<mirror::Object>** roots, 573 size_t count, 574 const RootInfo& info ATTRIBUTE_UNUSED) { 575 for (size_t i = 0; i < count; ++i) { 576 MarkObjectNonNull(roots[i]->AsMirrorPtr()); 577 } 578} 579 580class MarkSweep::VerifyRootVisitor : public SingleRootVisitor { 581 public: 582 explicit VerifyRootVisitor(std::ostream& os) : os_(os) {} 583 584 void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE 585 REQUIRES_SHARED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 586 // See if the root is on any space bitmap. 587 auto* heap = Runtime::Current()->GetHeap(); 588 if (heap->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) { 589 space::LargeObjectSpace* large_object_space = heap->GetLargeObjectsSpace(); 590 if (large_object_space != nullptr && !large_object_space->Contains(root)) { 591 os_ << "Found invalid root: " << root << " " << info << std::endl; 592 } 593 } 594 } 595 596 private: 597 std::ostream& os_; 598}; 599 600void MarkSweep::VerifySuspendedThreadRoots(std::ostream& os) { 601 VerifyRootVisitor visitor(os); 602 Runtime::Current()->GetThreadList()->VisitRootsForSuspendedThreads(&visitor); 603} 604 605void MarkSweep::MarkRoots(Thread* self) { 606 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 607 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 608 // If we exclusively hold the mutator lock, all threads must be suspended. 609 Runtime::Current()->VisitRoots(this); 610 RevokeAllThreadLocalAllocationStacks(self); 611 } else { 612 MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint); 613 // At this point the live stack should no longer have any mutators which push into it. 614 MarkNonThreadRoots(); 615 MarkConcurrentRoots( 616 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 617 } 618} 619 620void MarkSweep::MarkNonThreadRoots() { 621 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 622 Runtime::Current()->VisitNonThreadRoots(this); 623} 624 625void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 626 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 627 // Visit all runtime roots and clear dirty flags. 628 Runtime::Current()->VisitConcurrentRoots(this, flags); 629} 630 631class MarkSweep::DelayReferenceReferentVisitor { 632 public: 633 explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) {} 634 635 void operator()(ObjPtr<mirror::Class> klass, ObjPtr<mirror::Reference> ref) const 636 REQUIRES(Locks::heap_bitmap_lock_) 637 REQUIRES_SHARED(Locks::mutator_lock_) { 638 collector_->DelayReferenceReferent(klass, ref); 639 } 640 641 private: 642 MarkSweep* const collector_; 643}; 644 645template <bool kUseFinger = false> 646class MarkSweep::MarkStackTask : public Task { 647 public: 648 MarkStackTask(ThreadPool* thread_pool, 649 MarkSweep* mark_sweep, 650 size_t mark_stack_size, 651 StackReference<mirror::Object>* mark_stack) 652 : mark_sweep_(mark_sweep), 653 thread_pool_(thread_pool), 654 mark_stack_pos_(mark_stack_size) { 655 // We may have to copy part of an existing mark stack when another mark stack overflows. 656 if (mark_stack_size != 0) { 657 DCHECK(mark_stack != nullptr); 658 // TODO: Check performance? 659 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 660 } 661 if (kCountTasks) { 662 ++mark_sweep_->work_chunks_created_; 663 } 664 } 665 666 static const size_t kMaxSize = 1 * KB; 667 668 protected: 669 class MarkObjectParallelVisitor { 670 public: 671 ALWAYS_INLINE MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task, 672 MarkSweep* mark_sweep) 673 : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {} 674 675 ALWAYS_INLINE void operator()(mirror::Object* obj, 676 MemberOffset offset, 677 bool is_static ATTRIBUTE_UNUSED) const 678 REQUIRES_SHARED(Locks::mutator_lock_) { 679 Mark(obj->GetFieldObject<mirror::Object>(offset)); 680 } 681 682 void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const 683 REQUIRES_SHARED(Locks::mutator_lock_) { 684 if (!root->IsNull()) { 685 VisitRoot(root); 686 } 687 } 688 689 void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const 690 REQUIRES_SHARED(Locks::mutator_lock_) { 691 if (kCheckLocks) { 692 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 693 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 694 } 695 Mark(root->AsMirrorPtr()); 696 } 697 698 private: 699 ALWAYS_INLINE void Mark(mirror::Object* ref) const REQUIRES_SHARED(Locks::mutator_lock_) { 700 if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) { 701 if (kUseFinger) { 702 std::atomic_thread_fence(std::memory_order_seq_cst); 703 if (reinterpret_cast<uintptr_t>(ref) >= 704 static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) { 705 return; 706 } 707 } 708 chunk_task_->MarkStackPush(ref); 709 } 710 } 711 712 MarkStackTask<kUseFinger>* const chunk_task_; 713 MarkSweep* const mark_sweep_; 714 }; 715 716 class ScanObjectParallelVisitor { 717 public: 718 ALWAYS_INLINE explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) 719 : chunk_task_(chunk_task) {} 720 721 // No thread safety analysis since multiple threads will use this visitor. 722 void operator()(mirror::Object* obj) const 723 REQUIRES(Locks::heap_bitmap_lock_) 724 REQUIRES_SHARED(Locks::mutator_lock_) { 725 MarkSweep* const mark_sweep = chunk_task_->mark_sweep_; 726 MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep); 727 DelayReferenceReferentVisitor ref_visitor(mark_sweep); 728 mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor); 729 } 730 731 private: 732 MarkStackTask<kUseFinger>* const chunk_task_; 733 }; 734 735 virtual ~MarkStackTask() { 736 // Make sure that we have cleared our mark stack. 737 DCHECK_EQ(mark_stack_pos_, 0U); 738 if (kCountTasks) { 739 ++mark_sweep_->work_chunks_deleted_; 740 } 741 } 742 743 MarkSweep* const mark_sweep_; 744 ThreadPool* const thread_pool_; 745 // Thread local mark stack for this task. 746 StackReference<mirror::Object> mark_stack_[kMaxSize]; 747 // Mark stack position. 748 size_t mark_stack_pos_; 749 750 ALWAYS_INLINE void MarkStackPush(mirror::Object* obj) 751 REQUIRES_SHARED(Locks::mutator_lock_) { 752 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 753 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 754 mark_stack_pos_ /= 2; 755 auto* task = new MarkStackTask(thread_pool_, 756 mark_sweep_, 757 kMaxSize - mark_stack_pos_, 758 mark_stack_ + mark_stack_pos_); 759 thread_pool_->AddTask(Thread::Current(), task); 760 } 761 DCHECK(obj != nullptr); 762 DCHECK_LT(mark_stack_pos_, kMaxSize); 763 mark_stack_[mark_stack_pos_++].Assign(obj); 764 } 765 766 virtual void Finalize() { 767 delete this; 768 } 769 770 // Scans all of the objects 771 virtual void Run(Thread* self ATTRIBUTE_UNUSED) 772 REQUIRES(Locks::heap_bitmap_lock_) 773 REQUIRES_SHARED(Locks::mutator_lock_) { 774 ScanObjectParallelVisitor visitor(this); 775 // TODO: Tune this. 776 static const size_t kFifoSize = 4; 777 BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo; 778 for (;;) { 779 mirror::Object* obj = nullptr; 780 if (kUseMarkStackPrefetch) { 781 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 782 mirror::Object* const mark_stack_obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr(); 783 DCHECK(mark_stack_obj != nullptr); 784 __builtin_prefetch(mark_stack_obj); 785 prefetch_fifo.push_back(mark_stack_obj); 786 } 787 if (UNLIKELY(prefetch_fifo.empty())) { 788 break; 789 } 790 obj = prefetch_fifo.front(); 791 prefetch_fifo.pop_front(); 792 } else { 793 if (UNLIKELY(mark_stack_pos_ == 0)) { 794 break; 795 } 796 obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr(); 797 } 798 DCHECK(obj != nullptr); 799 visitor(obj); 800 } 801 } 802}; 803 804class MarkSweep::CardScanTask : public MarkStackTask<false> { 805 public: 806 CardScanTask(ThreadPool* thread_pool, 807 MarkSweep* mark_sweep, 808 accounting::ContinuousSpaceBitmap* bitmap, 809 uint8_t* begin, 810 uint8_t* end, 811 uint8_t minimum_age, 812 size_t mark_stack_size, 813 StackReference<mirror::Object>* mark_stack_obj, 814 bool clear_card) 815 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 816 bitmap_(bitmap), 817 begin_(begin), 818 end_(end), 819 minimum_age_(minimum_age), 820 clear_card_(clear_card) {} 821 822 protected: 823 accounting::ContinuousSpaceBitmap* const bitmap_; 824 uint8_t* const begin_; 825 uint8_t* const end_; 826 const uint8_t minimum_age_; 827 const bool clear_card_; 828 829 virtual void Finalize() { 830 delete this; 831 } 832 833 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 834 ScanObjectParallelVisitor visitor(this); 835 accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable(); 836 size_t cards_scanned = clear_card_ 837 ? card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_) 838 : card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_); 839 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 840 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 841 // Finish by emptying our local mark stack. 842 MarkStackTask::Run(self); 843 } 844}; 845 846size_t MarkSweep::GetThreadCount(bool paused) const { 847 // Use less threads if we are in a background state (non jank perceptible) since we want to leave 848 // more CPU time for the foreground apps. 849 if (heap_->GetThreadPool() == nullptr || !Runtime::Current()->InJankPerceptibleProcessState()) { 850 return 1; 851 } 852 return (paused ? heap_->GetParallelGCThreadCount() : heap_->GetConcGCThreadCount()) + 1; 853} 854 855void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) { 856 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 857 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 858 size_t thread_count = GetThreadCount(paused); 859 // The parallel version with only one thread is faster for card scanning, TODO: fix. 860 if (kParallelCardScan && thread_count > 1) { 861 Thread* self = Thread::Current(); 862 // Can't have a different split for each space since multiple spaces can have their cards being 863 // scanned at the same time. 864 TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__, 865 GetTimings()); 866 // Try to take some of the mark stack since we can pass this off to the worker tasks. 867 StackReference<mirror::Object>* mark_stack_begin = mark_stack_->Begin(); 868 StackReference<mirror::Object>* mark_stack_end = mark_stack_->End(); 869 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 870 // Estimated number of work tasks we will create. 871 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 872 DCHECK_NE(mark_stack_tasks, 0U); 873 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 874 mark_stack_size / mark_stack_tasks + 1); 875 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 876 if (space->GetMarkBitmap() == nullptr) { 877 continue; 878 } 879 uint8_t* card_begin = space->Begin(); 880 uint8_t* card_end = space->End(); 881 // Align up the end address. For example, the image space's end 882 // may not be card-size-aligned. 883 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 884 DCHECK_ALIGNED(card_begin, accounting::CardTable::kCardSize); 885 DCHECK_ALIGNED(card_end, accounting::CardTable::kCardSize); 886 // Calculate how many bytes of heap we will scan, 887 const size_t address_range = card_end - card_begin; 888 // Calculate how much address range each task gets. 889 const size_t card_delta = RoundUp(address_range / thread_count + 1, 890 accounting::CardTable::kCardSize); 891 // If paused and the space is neither zygote nor image space, we could clear the dirty 892 // cards to avoid accumulating them to increase card scanning load in the following GC 893 // cycles. We need to keep dirty cards of image space and zygote space in order to track 894 // references to the other spaces. 895 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 896 // Create the worker tasks for this space. 897 while (card_begin != card_end) { 898 // Add a range of cards. 899 size_t addr_remaining = card_end - card_begin; 900 size_t card_increment = std::min(card_delta, addr_remaining); 901 // Take from the back of the mark stack. 902 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 903 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 904 mark_stack_end -= mark_stack_increment; 905 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 906 DCHECK_EQ(mark_stack_end, mark_stack_->End()); 907 // Add the new task to the thread pool. 908 auto* task = new CardScanTask(thread_pool, 909 this, 910 space->GetMarkBitmap(), 911 card_begin, 912 card_begin + card_increment, 913 minimum_age, 914 mark_stack_increment, 915 mark_stack_end, 916 clear_card); 917 thread_pool->AddTask(self, task); 918 card_begin += card_increment; 919 } 920 } 921 922 // Note: the card scan below may dirty new cards (and scan them) 923 // as a side effect when a Reference object is encountered and 924 // queued during the marking. See b/11465268. 925 thread_pool->SetMaxActiveWorkers(thread_count - 1); 926 thread_pool->StartWorkers(self); 927 thread_pool->Wait(self, true, true); 928 thread_pool->StopWorkers(self); 929 } else { 930 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 931 if (space->GetMarkBitmap() != nullptr) { 932 // Image spaces are handled properly since live == marked for them. 933 const char* name = nullptr; 934 switch (space->GetGcRetentionPolicy()) { 935 case space::kGcRetentionPolicyNeverCollect: 936 name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects"; 937 break; 938 case space::kGcRetentionPolicyFullCollect: 939 name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects"; 940 break; 941 case space::kGcRetentionPolicyAlwaysCollect: 942 name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects"; 943 break; 944 default: 945 LOG(FATAL) << "Unreachable"; 946 UNREACHABLE(); 947 } 948 TimingLogger::ScopedTiming t(name, GetTimings()); 949 ScanObjectVisitor visitor(this); 950 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 951 if (clear_card) { 952 card_table->Scan<true>(space->GetMarkBitmap(), 953 space->Begin(), 954 space->End(), 955 visitor, 956 minimum_age); 957 } else { 958 card_table->Scan<false>(space->GetMarkBitmap(), 959 space->Begin(), 960 space->End(), 961 visitor, 962 minimum_age); 963 } 964 } 965 } 966 } 967} 968 969class MarkSweep::RecursiveMarkTask : public MarkStackTask<false> { 970 public: 971 RecursiveMarkTask(ThreadPool* thread_pool, 972 MarkSweep* mark_sweep, 973 accounting::ContinuousSpaceBitmap* bitmap, 974 uintptr_t begin, 975 uintptr_t end) 976 : MarkStackTask<false>(thread_pool, mark_sweep, 0, nullptr), 977 bitmap_(bitmap), 978 begin_(begin), 979 end_(end) {} 980 981 protected: 982 accounting::ContinuousSpaceBitmap* const bitmap_; 983 const uintptr_t begin_; 984 const uintptr_t end_; 985 986 virtual void Finalize() { 987 delete this; 988 } 989 990 // Scans all of the objects 991 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 992 ScanObjectParallelVisitor visitor(this); 993 bitmap_->VisitMarkedRange(begin_, end_, visitor); 994 // Finish by emptying our local mark stack. 995 MarkStackTask::Run(self); 996 } 997}; 998 999// Populates the mark stack based on the set of marked objects and 1000// recursively marks until the mark stack is emptied. 1001void MarkSweep::RecursiveMark() { 1002 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1003 // RecursiveMark will build the lists of known instances of the Reference classes. See 1004 // DelayReferenceReferent for details. 1005 if (kUseRecursiveMark) { 1006 const bool partial = GetGcType() == kGcTypePartial; 1007 ScanObjectVisitor scan_visitor(this); 1008 auto* self = Thread::Current(); 1009 ThreadPool* thread_pool = heap_->GetThreadPool(); 1010 size_t thread_count = GetThreadCount(false); 1011 const bool parallel = kParallelRecursiveMark && thread_count > 1; 1012 mark_stack_->Reset(); 1013 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1014 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 1015 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 1016 current_space_bitmap_ = space->GetMarkBitmap(); 1017 if (current_space_bitmap_ == nullptr) { 1018 continue; 1019 } 1020 if (parallel) { 1021 // We will use the mark stack the future. 1022 // CHECK(mark_stack_->IsEmpty()); 1023 // This function does not handle heap end increasing, so we must use the space end. 1024 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 1025 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 1026 atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue()); 1027 1028 // Create a few worker tasks. 1029 const size_t n = thread_count * 2; 1030 while (begin != end) { 1031 uintptr_t start = begin; 1032 uintptr_t delta = (end - begin) / n; 1033 delta = RoundUp(delta, KB); 1034 if (delta < 16 * KB) delta = end - begin; 1035 begin += delta; 1036 auto* task = new RecursiveMarkTask(thread_pool, 1037 this, 1038 current_space_bitmap_, 1039 start, 1040 begin); 1041 thread_pool->AddTask(self, task); 1042 } 1043 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1044 thread_pool->StartWorkers(self); 1045 thread_pool->Wait(self, true, true); 1046 thread_pool->StopWorkers(self); 1047 } else { 1048 // This function does not handle heap end increasing, so we must use the space end. 1049 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 1050 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 1051 current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 1052 } 1053 } 1054 } 1055 } 1056 ProcessMarkStack(false); 1057} 1058 1059void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) { 1060 ScanGrayObjects(paused, minimum_age); 1061 ProcessMarkStack(paused); 1062} 1063 1064void MarkSweep::ReMarkRoots() { 1065 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1066 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 1067 Runtime::Current()->VisitRoots(this, static_cast<VisitRootFlags>( 1068 kVisitRootFlagNewRoots | kVisitRootFlagStopLoggingNewRoots | kVisitRootFlagClearRootLog)); 1069 if (kVerifyRootsMarked) { 1070 TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings()); 1071 VerifyRootMarkedVisitor visitor(this); 1072 Runtime::Current()->VisitRoots(&visitor); 1073 } 1074} 1075 1076void MarkSweep::SweepSystemWeaks(Thread* self) { 1077 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1078 ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_); 1079 Runtime::Current()->SweepSystemWeaks(this); 1080} 1081 1082class MarkSweep::VerifySystemWeakVisitor : public IsMarkedVisitor { 1083 public: 1084 explicit VerifySystemWeakVisitor(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {} 1085 1086 virtual mirror::Object* IsMarked(mirror::Object* obj) 1087 OVERRIDE 1088 REQUIRES_SHARED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 1089 mark_sweep_->VerifyIsLive(obj); 1090 return obj; 1091 } 1092 1093 MarkSweep* const mark_sweep_; 1094}; 1095 1096void MarkSweep::VerifyIsLive(const mirror::Object* obj) { 1097 if (!heap_->GetLiveBitmap()->Test(obj)) { 1098 // TODO: Consider live stack? Has this code bitrotted? 1099 CHECK(!heap_->allocation_stack_->Contains(obj)) 1100 << "Found dead object " << obj << "\n" << heap_->DumpSpaces(); 1101 } 1102} 1103 1104void MarkSweep::VerifySystemWeaks() { 1105 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1106 // Verify system weaks, uses a special object visitor which returns the input object. 1107 VerifySystemWeakVisitor visitor(this); 1108 Runtime::Current()->SweepSystemWeaks(&visitor); 1109} 1110 1111class MarkSweep::CheckpointMarkThreadRoots : public Closure, public RootVisitor { 1112 public: 1113 CheckpointMarkThreadRoots(MarkSweep* mark_sweep, 1114 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) 1115 : mark_sweep_(mark_sweep), 1116 revoke_ros_alloc_thread_local_buffers_at_checkpoint_( 1117 revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 1118 } 1119 1120 void VisitRoots(mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED) 1121 OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) 1122 REQUIRES(Locks::heap_bitmap_lock_) { 1123 for (size_t i = 0; i < count; ++i) { 1124 mark_sweep_->MarkObjectNonNullParallel(*roots[i]); 1125 } 1126 } 1127 1128 void VisitRoots(mirror::CompressedReference<mirror::Object>** roots, 1129 size_t count, 1130 const RootInfo& info ATTRIBUTE_UNUSED) 1131 OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) 1132 REQUIRES(Locks::heap_bitmap_lock_) { 1133 for (size_t i = 0; i < count; ++i) { 1134 mark_sweep_->MarkObjectNonNullParallel(roots[i]->AsMirrorPtr()); 1135 } 1136 } 1137 1138 virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS { 1139 ScopedTrace trace("Marking thread roots"); 1140 // Note: self is not necessarily equal to thread since thread may be suspended. 1141 Thread* const self = Thread::Current(); 1142 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 1143 << thread->GetState() << " thread " << thread << " self " << self; 1144 thread->VisitRoots(this); 1145 if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) { 1146 ScopedTrace trace2("RevokeRosAllocThreadLocalBuffers"); 1147 mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread); 1148 } 1149 // If thread is a running mutator, then act on behalf of the garbage collector. 1150 // See the code in ThreadList::RunCheckpoint. 1151 mark_sweep_->GetBarrier().Pass(self); 1152 } 1153 1154 private: 1155 MarkSweep* const mark_sweep_; 1156 const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_; 1157}; 1158 1159void MarkSweep::MarkRootsCheckpoint(Thread* self, 1160 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 1161 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1162 CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint); 1163 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 1164 // Request the check point is run on all threads returning a count of the threads that must 1165 // run through the barrier including self. 1166 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1167 // Release locks then wait for all mutator threads to pass the barrier. 1168 // If there are no threads to wait which implys that all the checkpoint functions are finished, 1169 // then no need to release locks. 1170 if (barrier_count == 0) { 1171 return; 1172 } 1173 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1174 Locks::mutator_lock_->SharedUnlock(self); 1175 { 1176 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun); 1177 gc_barrier_->Increment(self, barrier_count); 1178 } 1179 Locks::mutator_lock_->SharedLock(self); 1180 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1181} 1182 1183void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1184 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1185 Thread* self = Thread::Current(); 1186 mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>( 1187 sweep_array_free_buffer_mem_map_->BaseBegin()); 1188 size_t chunk_free_pos = 0; 1189 ObjectBytePair freed; 1190 ObjectBytePair freed_los; 1191 // How many objects are left in the array, modified after each space is swept. 1192 StackReference<mirror::Object>* objects = allocations->Begin(); 1193 size_t count = allocations->Size(); 1194 // Change the order to ensure that the non-moving space last swept as an optimization. 1195 std::vector<space::ContinuousSpace*> sweep_spaces; 1196 space::ContinuousSpace* non_moving_space = nullptr; 1197 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1198 if (space->IsAllocSpace() && 1199 !immune_spaces_.ContainsSpace(space) && 1200 space->GetLiveBitmap() != nullptr) { 1201 if (space == heap_->GetNonMovingSpace()) { 1202 non_moving_space = space; 1203 } else { 1204 sweep_spaces.push_back(space); 1205 } 1206 } 1207 } 1208 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1209 // the other alloc spaces as an optimization. 1210 if (non_moving_space != nullptr) { 1211 sweep_spaces.push_back(non_moving_space); 1212 } 1213 // Start by sweeping the continuous spaces. 1214 for (space::ContinuousSpace* space : sweep_spaces) { 1215 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1216 accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1217 accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1218 if (swap_bitmaps) { 1219 std::swap(live_bitmap, mark_bitmap); 1220 } 1221 StackReference<mirror::Object>* out = objects; 1222 for (size_t i = 0; i < count; ++i) { 1223 mirror::Object* const obj = objects[i].AsMirrorPtr(); 1224 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1225 continue; 1226 } 1227 if (space->HasAddress(obj)) { 1228 // This object is in the space, remove it from the array and add it to the sweep buffer 1229 // if needed. 1230 if (!mark_bitmap->Test(obj)) { 1231 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1232 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1233 freed.objects += chunk_free_pos; 1234 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1235 chunk_free_pos = 0; 1236 } 1237 chunk_free_buffer[chunk_free_pos++] = obj; 1238 } 1239 } else { 1240 (out++)->Assign(obj); 1241 } 1242 } 1243 if (chunk_free_pos > 0) { 1244 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1245 freed.objects += chunk_free_pos; 1246 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1247 chunk_free_pos = 0; 1248 } 1249 // All of the references which space contained are no longer in the allocation stack, update 1250 // the count. 1251 count = out - objects; 1252 } 1253 // Handle the large object space. 1254 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1255 if (large_object_space != nullptr) { 1256 accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap(); 1257 accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap(); 1258 if (swap_bitmaps) { 1259 std::swap(large_live_objects, large_mark_objects); 1260 } 1261 for (size_t i = 0; i < count; ++i) { 1262 mirror::Object* const obj = objects[i].AsMirrorPtr(); 1263 // Handle large objects. 1264 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1265 continue; 1266 } 1267 if (!large_mark_objects->Test(obj)) { 1268 ++freed_los.objects; 1269 freed_los.bytes += large_object_space->Free(self, obj); 1270 } 1271 } 1272 } 1273 { 1274 TimingLogger::ScopedTiming t2("RecordFree", GetTimings()); 1275 RecordFree(freed); 1276 RecordFreeLOS(freed_los); 1277 t2.NewTiming("ResetStack"); 1278 allocations->Reset(); 1279 } 1280 sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero(); 1281} 1282 1283void MarkSweep::Sweep(bool swap_bitmaps) { 1284 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1285 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 1286 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 1287 { 1288 TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings()); 1289 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 1290 // knowing that new allocations won't be marked as live. 1291 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 1292 heap_->MarkAllocStackAsLive(live_stack); 1293 live_stack->Reset(); 1294 DCHECK(mark_stack_->IsEmpty()); 1295 } 1296 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1297 if (space->IsContinuousMemMapAllocSpace()) { 1298 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1299 TimingLogger::ScopedTiming split( 1300 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", 1301 GetTimings()); 1302 RecordFree(alloc_space->Sweep(swap_bitmaps)); 1303 } 1304 } 1305 SweepLargeObjects(swap_bitmaps); 1306} 1307 1308void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1309 space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace(); 1310 if (los != nullptr) { 1311 TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings()); 1312 RecordFreeLOS(los->Sweep(swap_bitmaps)); 1313 } 1314} 1315 1316// Process the "referent" field lin a java.lang.ref.Reference. If the referent has not yet been 1317// marked, put it on the appropriate list in the heap for later processing. 1318void MarkSweep::DelayReferenceReferent(ObjPtr<mirror::Class> klass, ObjPtr<mirror::Reference> ref) { 1319 heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, this); 1320} 1321 1322class MarkVisitor { 1323 public: 1324 ALWAYS_INLINE explicit MarkVisitor(MarkSweep* const mark_sweep) : mark_sweep_(mark_sweep) {} 1325 1326 ALWAYS_INLINE void operator()(mirror::Object* obj, 1327 MemberOffset offset, 1328 bool is_static ATTRIBUTE_UNUSED) const 1329 REQUIRES(Locks::heap_bitmap_lock_) 1330 REQUIRES_SHARED(Locks::mutator_lock_) { 1331 if (kCheckLocks) { 1332 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1333 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1334 } 1335 mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset), obj, offset); 1336 } 1337 1338 void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const 1339 REQUIRES(Locks::heap_bitmap_lock_) 1340 REQUIRES_SHARED(Locks::mutator_lock_) { 1341 if (!root->IsNull()) { 1342 VisitRoot(root); 1343 } 1344 } 1345 1346 void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const 1347 REQUIRES(Locks::heap_bitmap_lock_) 1348 REQUIRES_SHARED(Locks::mutator_lock_) { 1349 if (kCheckLocks) { 1350 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1351 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1352 } 1353 mark_sweep_->MarkObject(root->AsMirrorPtr()); 1354 } 1355 1356 private: 1357 MarkSweep* const mark_sweep_; 1358}; 1359 1360// Scans an object reference. Determines the type of the reference 1361// and dispatches to a specialized scanning routine. 1362void MarkSweep::ScanObject(mirror::Object* obj) { 1363 MarkVisitor mark_visitor(this); 1364 DelayReferenceReferentVisitor ref_visitor(this); 1365 ScanObjectVisit(obj, mark_visitor, ref_visitor); 1366} 1367 1368void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1369 Thread* self = Thread::Current(); 1370 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1371 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1372 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1373 CHECK_GT(chunk_size, 0U); 1374 // Split the current mark stack up into work tasks. 1375 for (auto* it = mark_stack_->Begin(), *end = mark_stack_->End(); it < end; ) { 1376 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1377 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it)); 1378 it += delta; 1379 } 1380 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1381 thread_pool->StartWorkers(self); 1382 thread_pool->Wait(self, true, true); 1383 thread_pool->StopWorkers(self); 1384 mark_stack_->Reset(); 1385 CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(), 1386 work_chunks_deleted_.LoadSequentiallyConsistent()) 1387 << " some of the work chunks were leaked"; 1388} 1389 1390// Scan anything that's on the mark stack. 1391void MarkSweep::ProcessMarkStack(bool paused) { 1392 TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings()); 1393 size_t thread_count = GetThreadCount(paused); 1394 if (kParallelProcessMarkStack && thread_count > 1 && 1395 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1396 ProcessMarkStackParallel(thread_count); 1397 } else { 1398 // TODO: Tune this. 1399 static const size_t kFifoSize = 4; 1400 BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo; 1401 for (;;) { 1402 mirror::Object* obj = nullptr; 1403 if (kUseMarkStackPrefetch) { 1404 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1405 mirror::Object* mark_stack_obj = mark_stack_->PopBack(); 1406 DCHECK(mark_stack_obj != nullptr); 1407 __builtin_prefetch(mark_stack_obj); 1408 prefetch_fifo.push_back(mark_stack_obj); 1409 } 1410 if (prefetch_fifo.empty()) { 1411 break; 1412 } 1413 obj = prefetch_fifo.front(); 1414 prefetch_fifo.pop_front(); 1415 } else { 1416 if (mark_stack_->IsEmpty()) { 1417 break; 1418 } 1419 obj = mark_stack_->PopBack(); 1420 } 1421 DCHECK(obj != nullptr); 1422 ScanObject(obj); 1423 } 1424 } 1425} 1426 1427inline mirror::Object* MarkSweep::IsMarked(mirror::Object* object) { 1428 if (immune_spaces_.IsInImmuneRegion(object)) { 1429 return object; 1430 } 1431 if (current_space_bitmap_->HasAddress(object)) { 1432 return current_space_bitmap_->Test(object) ? object : nullptr; 1433 } 1434 return mark_bitmap_->Test(object) ? object : nullptr; 1435} 1436 1437void MarkSweep::FinishPhase() { 1438 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1439 if (kCountScannedTypes) { 1440 VLOG(gc) 1441 << "MarkSweep scanned" 1442 << " no reference objects=" << no_reference_class_count_.LoadRelaxed() 1443 << " normal objects=" << normal_count_.LoadRelaxed() 1444 << " classes=" << class_count_.LoadRelaxed() 1445 << " object arrays=" << object_array_count_.LoadRelaxed() 1446 << " references=" << reference_count_.LoadRelaxed() 1447 << " other=" << other_count_.LoadRelaxed(); 1448 } 1449 if (kCountTasks) { 1450 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed(); 1451 } 1452 if (kMeasureOverhead) { 1453 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed()); 1454 } 1455 if (kProfileLargeObjects) { 1456 VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed() 1457 << " marked " << large_object_mark_.LoadRelaxed(); 1458 } 1459 if (kCountMarkedObjects) { 1460 VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed() 1461 << " immune=" << mark_immune_count_.LoadRelaxed() 1462 << " fastpath=" << mark_fastpath_count_.LoadRelaxed() 1463 << " slowpath=" << mark_slowpath_count_.LoadRelaxed(); 1464 } 1465 CHECK(mark_stack_->IsEmpty()); // Ensure that the mark stack is empty. 1466 mark_stack_->Reset(); 1467 Thread* const self = Thread::Current(); 1468 ReaderMutexLock mu(self, *Locks::mutator_lock_); 1469 WriterMutexLock mu2(self, *Locks::heap_bitmap_lock_); 1470 heap_->ClearMarkedObjects(); 1471} 1472 1473void MarkSweep::RevokeAllThreadLocalBuffers() { 1474 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) { 1475 // If concurrent, rosalloc thread-local buffers are revoked at the 1476 // thread checkpoint. Bump pointer space thread-local buffers must 1477 // not be in use. 1478 GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked(); 1479 } else { 1480 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1481 GetHeap()->RevokeAllThreadLocalBuffers(); 1482 } 1483} 1484 1485} // namespace collector 1486} // namespace gc 1487} // namespace art 1488