mark_sweep.cc revision 727b294b4091cf3cc2f8137cd654552f477fe46a
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#define ATRACE_TAG ATRACE_TAG_DALVIK 25#include "cutils/trace.h" 26 27#include "base/bounded_fifo.h" 28#include "base/logging.h" 29#include "base/macros.h" 30#include "base/mutex-inl.h" 31#include "base/timing_logger.h" 32#include "gc/accounting/card_table-inl.h" 33#include "gc/accounting/heap_bitmap-inl.h" 34#include "gc/accounting/mod_union_table.h" 35#include "gc/accounting/space_bitmap-inl.h" 36#include "gc/heap.h" 37#include "gc/reference_processor.h" 38#include "gc/space/image_space.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/art_field-inl.h" 43#include "mirror/object-inl.h" 44#include "runtime.h" 45#include "scoped_thread_state_change.h" 46#include "thread-inl.h" 47#include "thread_list.h" 48 49using ::art::mirror::Object; 50 51namespace art { 52namespace gc { 53namespace collector { 54 55// Performance options. 56static constexpr bool kUseRecursiveMark = false; 57static constexpr bool kUseMarkStackPrefetch = true; 58static constexpr size_t kSweepArrayChunkFreeSize = 1024; 59static constexpr bool kPreCleanCards = true; 60 61// Parallelism options. 62static constexpr bool kParallelCardScan = true; 63static constexpr bool kParallelRecursiveMark = true; 64// Don't attempt to parallelize mark stack processing unless the mark stack is at least n 65// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not 66// having this can add overhead in ProcessReferences since we may end up doing many calls of 67// ProcessMarkStack with very small mark stacks. 68static constexpr size_t kMinimumParallelMarkStackSize = 128; 69static constexpr bool kParallelProcessMarkStack = true; 70 71// Profiling and information flags. 72static constexpr bool kProfileLargeObjects = false; 73static constexpr bool kMeasureOverhead = false; 74static constexpr bool kCountTasks = false; 75static constexpr bool kCountJavaLangRefs = false; 76static constexpr bool kCountMarkedObjects = false; 77 78// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%. 79static constexpr bool kCheckLocks = kDebugLocking; 80static constexpr bool kVerifyRootsMarked = kIsDebugBuild; 81 82// If true, revoke the rosalloc thread-local buffers at the 83// checkpoint, as opposed to during the pause. 84static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true; 85 86void MarkSweep::BindBitmaps() { 87 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 88 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 89 // Mark all of the spaces we never collect as immune. 90 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 91 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 92 CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space; 93 } 94 } 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 current_space_bitmap_(nullptr), mark_bitmap_(nullptr), mark_stack_(nullptr), 102 gc_barrier_(new Barrier(0)), 103 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 104 is_concurrent_(is_concurrent), live_stack_freeze_size_(0) { 105 std::string error_msg; 106 MemMap* mem_map = MemMap::MapAnonymous( 107 "mark sweep sweep array free buffer", nullptr, 108 RoundUp(kSweepArrayChunkFreeSize * sizeof(mirror::Object*), kPageSize), 109 PROT_READ | PROT_WRITE, false, &error_msg); 110 CHECK(mem_map != nullptr) << "Couldn't allocate sweep array free buffer: " << error_msg; 111 sweep_array_free_buffer_mem_map_.reset(mem_map); 112} 113 114void MarkSweep::InitializePhase() { 115 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 116 mark_stack_ = heap_->GetMarkStack(); 117 DCHECK(mark_stack_ != nullptr); 118 immune_region_.Reset(); 119 class_count_.StoreRelaxed(0); 120 array_count_.StoreRelaxed(0); 121 other_count_.StoreRelaxed(0); 122 large_object_test_.StoreRelaxed(0); 123 large_object_mark_.StoreRelaxed(0); 124 overhead_time_ .StoreRelaxed(0); 125 work_chunks_created_.StoreRelaxed(0); 126 work_chunks_deleted_.StoreRelaxed(0); 127 reference_count_.StoreRelaxed(0); 128 mark_null_count_.StoreRelaxed(0); 129 mark_immune_count_.StoreRelaxed(0); 130 mark_fastpath_count_.StoreRelaxed(0); 131 mark_slowpath_count_.StoreRelaxed(0); 132 { 133 // TODO: I don't think we should need heap bitmap lock to Get the mark bitmap. 134 ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 135 mark_bitmap_ = heap_->GetMarkBitmap(); 136 } 137 if (!GetCurrentIteration()->GetClearSoftReferences()) { 138 // Always clear soft references if a non-sticky collection. 139 GetCurrentIteration()->SetClearSoftReferences(GetGcType() != collector::kGcTypeSticky); 140 } 141} 142 143void MarkSweep::RunPhases() { 144 Thread* self = Thread::Current(); 145 InitializePhase(); 146 Locks::mutator_lock_->AssertNotHeld(self); 147 if (IsConcurrent()) { 148 GetHeap()->PreGcVerification(this); 149 { 150 ReaderMutexLock mu(self, *Locks::mutator_lock_); 151 MarkingPhase(); 152 } 153 ScopedPause pause(this); 154 GetHeap()->PrePauseRosAllocVerification(this); 155 PausePhase(); 156 RevokeAllThreadLocalBuffers(); 157 } else { 158 ScopedPause pause(this); 159 GetHeap()->PreGcVerificationPaused(this); 160 MarkingPhase(); 161 GetHeap()->PrePauseRosAllocVerification(this); 162 PausePhase(); 163 RevokeAllThreadLocalBuffers(); 164 } 165 { 166 // Sweeping always done concurrently, even for non concurrent mark sweep. 167 ReaderMutexLock mu(self, *Locks::mutator_lock_); 168 ReclaimPhase(); 169 } 170 GetHeap()->PostGcVerification(this); 171 FinishPhase(); 172} 173 174void MarkSweep::ProcessReferences(Thread* self) { 175 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 176 GetHeap()->GetReferenceProcessor()->ProcessReferences( 177 true, GetTimings(), GetCurrentIteration()->GetClearSoftReferences(), 178 &HeapReferenceMarkedCallback, &MarkObjectCallback, &ProcessMarkStackCallback, this); 179} 180 181void MarkSweep::PausePhase() { 182 TimingLogger::ScopedTiming t("(Paused)PausePhase", GetTimings()); 183 Thread* self = Thread::Current(); 184 Locks::mutator_lock_->AssertExclusiveHeld(self); 185 if (IsConcurrent()) { 186 // Handle the dirty objects if we are a concurrent GC. 187 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 188 // Re-mark root set. 189 ReMarkRoots(); 190 // Scan dirty objects, this is only required if we are not doing concurrent GC. 191 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 192 } 193 { 194 TimingLogger::ScopedTiming t2("SwapStacks", GetTimings()); 195 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 196 heap_->SwapStacks(self); 197 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 198 // Need to revoke all the thread local allocation stacks since we just swapped the allocation 199 // stacks and don't want anybody to allocate into the live stack. 200 RevokeAllThreadLocalAllocationStacks(self); 201 } 202 heap_->PreSweepingGcVerification(this); 203 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 204 // weak before we sweep them. Since this new system weak may not be marked, the GC may 205 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 206 // reference to a string that is about to be swept. 207 Runtime::Current()->DisallowNewSystemWeaks(); 208 // Enable the reference processing slow path, needs to be done with mutators paused since there 209 // is no lock in the GetReferent fast path. 210 GetHeap()->GetReferenceProcessor()->EnableSlowPath(); 211} 212 213void MarkSweep::PreCleanCards() { 214 // Don't do this for non concurrent GCs since they don't have any dirty cards. 215 if (kPreCleanCards && IsConcurrent()) { 216 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 217 Thread* self = Thread::Current(); 218 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 219 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 220 heap_->ProcessCards(GetTimings(), false); 221 // The checkpoint root marking is required to avoid a race condition which occurs if the 222 // following happens during a reference write: 223 // 1. mutator dirties the card (write barrier) 224 // 2. GC ages the card (the above ProcessCards call) 225 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 226 // 4. mutator writes the value (corresponding to the write barrier in 1.) 227 // This causes the GC to age the card but not necessarily mark the reference which the mutator 228 // wrote into the object stored in the card. 229 // Having the checkpoint fixes this issue since it ensures that the card mark and the 230 // reference write are visible to the GC before the card is scanned (this is due to locks being 231 // acquired / released in the checkpoint code). 232 // The other roots are also marked to help reduce the pause. 233 MarkRootsCheckpoint(self, false); 234 MarkNonThreadRoots(); 235 MarkConcurrentRoots( 236 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 237 // Process the newly aged cards. 238 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 239 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 240 // in the next GC. 241 } 242} 243 244void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 245 if (kUseThreadLocalAllocationStack) { 246 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 247 Locks::mutator_lock_->AssertExclusiveHeld(self); 248 heap_->RevokeAllThreadLocalAllocationStacks(self); 249 } 250} 251 252void MarkSweep::MarkingPhase() { 253 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 254 Thread* self = Thread::Current(); 255 BindBitmaps(); 256 FindDefaultSpaceBitmap(); 257 // Process dirty cards and add dirty cards to mod union tables. 258 heap_->ProcessCards(GetTimings(), false); 259 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 260 MarkRoots(self); 261 MarkReachableObjects(); 262 // Pre-clean dirtied cards to reduce pauses. 263 PreCleanCards(); 264} 265 266void MarkSweep::UpdateAndMarkModUnion() { 267 for (const auto& space : heap_->GetContinuousSpaces()) { 268 if (immune_region_.ContainsSpace(space)) { 269 const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" : 270 "UpdateAndMarkImageModUnionTable"; 271 TimingLogger::ScopedTiming t(name, GetTimings()); 272 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 273 CHECK(mod_union_table != nullptr); 274 mod_union_table->UpdateAndMarkReferences(MarkHeapReferenceCallback, this); 275 } 276 } 277} 278 279void MarkSweep::MarkReachableObjects() { 280 UpdateAndMarkModUnion(); 281 // Recursively mark all the non-image bits set in the mark bitmap. 282 RecursiveMark(); 283} 284 285void MarkSweep::ReclaimPhase() { 286 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 287 Thread* self = Thread::Current(); 288 // Process the references concurrently. 289 ProcessReferences(self); 290 SweepSystemWeaks(self); 291 Runtime::Current()->AllowNewSystemWeaks(); 292 { 293 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 294 // Reclaim unmarked objects. 295 Sweep(false); 296 // Swap the live and mark bitmaps for each space which we modified space. This is an 297 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 298 // bitmaps. 299 SwapBitmaps(); 300 // Unbind the live and mark bitmaps. 301 GetHeap()->UnBindBitmaps(); 302 } 303} 304 305void MarkSweep::FindDefaultSpaceBitmap() { 306 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 307 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 308 accounting::ContinuousSpaceBitmap* bitmap = space->GetMarkBitmap(); 309 // We want to have the main space instead of non moving if possible. 310 if (bitmap != nullptr && 311 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 312 current_space_bitmap_ = bitmap; 313 // If we are not the non moving space exit the loop early since this will be good enough. 314 if (space != heap_->GetNonMovingSpace()) { 315 break; 316 } 317 } 318 } 319 CHECK(current_space_bitmap_ != nullptr) << "Could not find a default mark bitmap\n" 320 << heap_->DumpSpaces(); 321} 322 323void MarkSweep::ExpandMarkStack() { 324 ResizeMarkStack(mark_stack_->Capacity() * 2); 325} 326 327void MarkSweep::ResizeMarkStack(size_t new_size) { 328 // Rare case, no need to have Thread::Current be a parameter. 329 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 330 // Someone else acquired the lock and expanded the mark stack before us. 331 return; 332 } 333 std::vector<Object*> temp(mark_stack_->Begin(), mark_stack_->End()); 334 CHECK_LE(mark_stack_->Size(), new_size); 335 mark_stack_->Resize(new_size); 336 for (const auto& obj : temp) { 337 mark_stack_->PushBack(obj); 338 } 339} 340 341inline void MarkSweep::MarkObjectNonNullParallel(Object* obj) { 342 DCHECK(obj != nullptr); 343 if (MarkObjectParallel(obj)) { 344 MutexLock mu(Thread::Current(), mark_stack_lock_); 345 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 346 ExpandMarkStack(); 347 } 348 // The object must be pushed on to the mark stack. 349 mark_stack_->PushBack(obj); 350 } 351} 352 353mirror::Object* MarkSweep::MarkObjectCallback(mirror::Object* obj, void* arg) { 354 MarkSweep* mark_sweep = reinterpret_cast<MarkSweep*>(arg); 355 mark_sweep->MarkObject(obj); 356 return obj; 357} 358 359void MarkSweep::MarkHeapReferenceCallback(mirror::HeapReference<mirror::Object>* ref, void* arg) { 360 reinterpret_cast<MarkSweep*>(arg)->MarkObject(ref->AsMirrorPtr()); 361} 362 363bool MarkSweep::HeapReferenceMarkedCallback(mirror::HeapReference<mirror::Object>* ref, void* arg) { 364 return reinterpret_cast<MarkSweep*>(arg)->IsMarked(ref->AsMirrorPtr()); 365} 366 367class MarkSweepMarkObjectSlowPath { 368 public: 369 explicit MarkSweepMarkObjectSlowPath(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) { 370 } 371 372 void operator()(const Object* obj) const ALWAYS_INLINE { 373 if (kProfileLargeObjects) { 374 // TODO: Differentiate between marking and testing somehow. 375 ++mark_sweep_->large_object_test_; 376 ++mark_sweep_->large_object_mark_; 377 } 378 space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace(); 379 if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) || 380 (kIsDebugBuild && large_object_space != nullptr && 381 !large_object_space->Contains(obj)))) { 382 LOG(ERROR) << "Tried to mark " << obj << " not contained by any spaces"; 383 LOG(ERROR) << "Attempting see if it's a bad root"; 384 mark_sweep_->VerifyRoots(); 385 LOG(FATAL) << "Can't mark invalid object"; 386 } 387 } 388 389 private: 390 MarkSweep* const mark_sweep_; 391}; 392 393inline void MarkSweep::MarkObjectNonNull(Object* obj) { 394 DCHECK(obj != nullptr); 395 if (kUseBakerOrBrooksReadBarrier) { 396 // Verify all the objects have the correct pointer installed. 397 obj->AssertReadBarrierPointer(); 398 } 399 if (immune_region_.ContainsObject(obj)) { 400 if (kCountMarkedObjects) { 401 ++mark_immune_count_; 402 } 403 DCHECK(mark_bitmap_->Test(obj)); 404 } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) { 405 if (kCountMarkedObjects) { 406 ++mark_fastpath_count_; 407 } 408 if (UNLIKELY(!current_space_bitmap_->Set(obj))) { 409 PushOnMarkStack(obj); // This object was not previously marked. 410 } 411 } else { 412 if (kCountMarkedObjects) { 413 ++mark_slowpath_count_; 414 } 415 MarkSweepMarkObjectSlowPath visitor(this); 416 // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set 417 // will check again. 418 if (!mark_bitmap_->Set(obj, visitor)) { 419 PushOnMarkStack(obj); // Was not already marked, push. 420 } 421 } 422} 423 424inline void MarkSweep::PushOnMarkStack(Object* obj) { 425 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 426 // Lock is not needed but is here anyways to please annotalysis. 427 MutexLock mu(Thread::Current(), mark_stack_lock_); 428 ExpandMarkStack(); 429 } 430 // The object must be pushed on to the mark stack. 431 mark_stack_->PushBack(obj); 432} 433 434inline bool MarkSweep::MarkObjectParallel(const Object* obj) { 435 DCHECK(obj != nullptr); 436 if (kUseBakerOrBrooksReadBarrier) { 437 // Verify all the objects have the correct pointer installed. 438 obj->AssertReadBarrierPointer(); 439 } 440 if (immune_region_.ContainsObject(obj)) { 441 DCHECK(IsMarked(obj)); 442 return false; 443 } 444 // Try to take advantage of locality of references within a space, failing this find the space 445 // the hard way. 446 accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_; 447 if (LIKELY(object_bitmap->HasAddress(obj))) { 448 return !object_bitmap->AtomicTestAndSet(obj); 449 } 450 MarkSweepMarkObjectSlowPath visitor(this); 451 return !mark_bitmap_->AtomicTestAndSet(obj, visitor); 452} 453 454// Used to mark objects when processing the mark stack. If an object is null, it is not marked. 455inline void MarkSweep::MarkObject(Object* obj) { 456 if (obj != nullptr) { 457 MarkObjectNonNull(obj); 458 } else if (kCountMarkedObjects) { 459 ++mark_null_count_; 460 } 461} 462 463void MarkSweep::MarkRootParallelCallback(Object** root, void* arg, const RootInfo& /*root_info*/) { 464 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNullParallel(*root); 465} 466 467void MarkSweep::VerifyRootMarked(Object** root, void* arg, const RootInfo& /*root_info*/) { 468 CHECK(reinterpret_cast<MarkSweep*>(arg)->IsMarked(*root)); 469} 470 471void MarkSweep::MarkRootCallback(Object** root, void* arg, const RootInfo& /*root_info*/) { 472 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNull(*root); 473} 474 475void MarkSweep::VerifyRootCallback(Object** root, void* arg, const RootInfo& root_info) { 476 reinterpret_cast<MarkSweep*>(arg)->VerifyRoot(*root, root_info); 477} 478 479void MarkSweep::VerifyRoot(const Object* root, const RootInfo& root_info) { 480 // See if the root is on any space bitmap. 481 if (heap_->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) { 482 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 483 if (large_object_space != nullptr && !large_object_space->Contains(root)) { 484 LOG(ERROR) << "Found invalid root: " << root << " "; 485 root_info.Describe(LOG(ERROR)); 486 } 487 } 488} 489 490void MarkSweep::VerifyRoots() { 491 Runtime::Current()->GetThreadList()->VisitRoots(VerifyRootCallback, this); 492} 493 494void MarkSweep::MarkRoots(Thread* self) { 495 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 496 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 497 // If we exclusively hold the mutator lock, all threads must be suspended. 498 Runtime::Current()->VisitRoots(MarkRootCallback, this); 499 RevokeAllThreadLocalAllocationStacks(self); 500 } else { 501 MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint); 502 // At this point the live stack should no longer have any mutators which push into it. 503 MarkNonThreadRoots(); 504 MarkConcurrentRoots( 505 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 506 } 507} 508 509void MarkSweep::MarkNonThreadRoots() { 510 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 511 Runtime::Current()->VisitNonThreadRoots(MarkRootCallback, this); 512} 513 514void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 515 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 516 // Visit all runtime roots and clear dirty flags. 517 Runtime::Current()->VisitConcurrentRoots(MarkRootCallback, this, flags); 518} 519 520class ScanObjectVisitor { 521 public: 522 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 523 : mark_sweep_(mark_sweep) {} 524 525 void operator()(Object* obj) const ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 526 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 527 if (kCheckLocks) { 528 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 529 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 530 } 531 mark_sweep_->ScanObject(obj); 532 } 533 534 private: 535 MarkSweep* const mark_sweep_; 536}; 537 538class DelayReferenceReferentVisitor { 539 public: 540 explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) { 541 } 542 543 void operator()(mirror::Class* klass, mirror::Reference* ref) const 544 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 545 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 546 collector_->DelayReferenceReferent(klass, ref); 547 } 548 549 private: 550 MarkSweep* const collector_; 551}; 552 553template <bool kUseFinger = false> 554class MarkStackTask : public Task { 555 public: 556 MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size, 557 Object** mark_stack) 558 : mark_sweep_(mark_sweep), 559 thread_pool_(thread_pool), 560 mark_stack_pos_(mark_stack_size) { 561 // We may have to copy part of an existing mark stack when another mark stack overflows. 562 if (mark_stack_size != 0) { 563 DCHECK(mark_stack != NULL); 564 // TODO: Check performance? 565 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 566 } 567 if (kCountTasks) { 568 ++mark_sweep_->work_chunks_created_; 569 } 570 } 571 572 static const size_t kMaxSize = 1 * KB; 573 574 protected: 575 class MarkObjectParallelVisitor { 576 public: 577 explicit MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task, 578 MarkSweep* mark_sweep) ALWAYS_INLINE 579 : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {} 580 581 void operator()(Object* obj, MemberOffset offset, bool /* static */) const ALWAYS_INLINE 582 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 583 mirror::Object* ref = obj->GetFieldObject<mirror::Object>(offset); 584 if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) { 585 if (kUseFinger) { 586 android_memory_barrier(); 587 if (reinterpret_cast<uintptr_t>(ref) >= 588 static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) { 589 return; 590 } 591 } 592 chunk_task_->MarkStackPush(ref); 593 } 594 } 595 596 private: 597 MarkStackTask<kUseFinger>* const chunk_task_; 598 MarkSweep* const mark_sweep_; 599 }; 600 601 class ScanObjectParallelVisitor { 602 public: 603 explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE 604 : chunk_task_(chunk_task) {} 605 606 // No thread safety analysis since multiple threads will use this visitor. 607 void operator()(Object* obj) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 608 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 609 MarkSweep* const mark_sweep = chunk_task_->mark_sweep_; 610 MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep); 611 DelayReferenceReferentVisitor ref_visitor(mark_sweep); 612 mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor); 613 } 614 615 private: 616 MarkStackTask<kUseFinger>* const chunk_task_; 617 }; 618 619 virtual ~MarkStackTask() { 620 // Make sure that we have cleared our mark stack. 621 DCHECK_EQ(mark_stack_pos_, 0U); 622 if (kCountTasks) { 623 ++mark_sweep_->work_chunks_deleted_; 624 } 625 } 626 627 MarkSweep* const mark_sweep_; 628 ThreadPool* const thread_pool_; 629 // Thread local mark stack for this task. 630 Object* mark_stack_[kMaxSize]; 631 // Mark stack position. 632 size_t mark_stack_pos_; 633 634 void MarkStackPush(Object* obj) ALWAYS_INLINE { 635 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 636 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 637 mark_stack_pos_ /= 2; 638 auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_, 639 mark_stack_ + mark_stack_pos_); 640 thread_pool_->AddTask(Thread::Current(), task); 641 } 642 DCHECK(obj != nullptr); 643 DCHECK_LT(mark_stack_pos_, kMaxSize); 644 mark_stack_[mark_stack_pos_++] = obj; 645 } 646 647 virtual void Finalize() { 648 delete this; 649 } 650 651 // Scans all of the objects 652 virtual void Run(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 653 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 654 UNUSED(self); 655 ScanObjectParallelVisitor visitor(this); 656 // TODO: Tune this. 657 static const size_t kFifoSize = 4; 658 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 659 for (;;) { 660 Object* obj = nullptr; 661 if (kUseMarkStackPrefetch) { 662 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 663 Object* mark_stack_obj = mark_stack_[--mark_stack_pos_]; 664 DCHECK(mark_stack_obj != nullptr); 665 __builtin_prefetch(mark_stack_obj); 666 prefetch_fifo.push_back(mark_stack_obj); 667 } 668 if (UNLIKELY(prefetch_fifo.empty())) { 669 break; 670 } 671 obj = prefetch_fifo.front(); 672 prefetch_fifo.pop_front(); 673 } else { 674 if (UNLIKELY(mark_stack_pos_ == 0)) { 675 break; 676 } 677 obj = mark_stack_[--mark_stack_pos_]; 678 } 679 DCHECK(obj != nullptr); 680 visitor(obj); 681 } 682 } 683}; 684 685class CardScanTask : public MarkStackTask<false> { 686 public: 687 CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 688 accounting::ContinuousSpaceBitmap* bitmap, 689 uint8_t* begin, uint8_t* end, uint8_t minimum_age, size_t mark_stack_size, 690 Object** mark_stack_obj, bool clear_card) 691 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 692 bitmap_(bitmap), 693 begin_(begin), 694 end_(end), 695 minimum_age_(minimum_age), clear_card_(clear_card) { 696 } 697 698 protected: 699 accounting::ContinuousSpaceBitmap* const bitmap_; 700 uint8_t* const begin_; 701 uint8_t* const end_; 702 const uint8_t minimum_age_; 703 const bool clear_card_; 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 = clear_card_ ? 713 card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_) : 714 card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_); 715 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 716 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 717 // Finish by emptying our local mark stack. 718 MarkStackTask::Run(self); 719 } 720}; 721 722size_t MarkSweep::GetThreadCount(bool paused) const { 723 if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) { 724 return 1; 725 } 726 if (paused) { 727 return heap_->GetParallelGCThreadCount() + 1; 728 } else { 729 return heap_->GetConcGCThreadCount() + 1; 730 } 731} 732 733void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) { 734 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 735 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 736 size_t thread_count = GetThreadCount(paused); 737 // The parallel version with only one thread is faster for card scanning, TODO: fix. 738 if (kParallelCardScan && thread_count > 1) { 739 Thread* self = Thread::Current(); 740 // Can't have a different split for each space since multiple spaces can have their cards being 741 // scanned at the same time. 742 TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__, 743 GetTimings()); 744 // Try to take some of the mark stack since we can pass this off to the worker tasks. 745 Object** mark_stack_begin = mark_stack_->Begin(); 746 Object** mark_stack_end = mark_stack_->End(); 747 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 748 // Estimated number of work tasks we will create. 749 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 750 DCHECK_NE(mark_stack_tasks, 0U); 751 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 752 mark_stack_size / mark_stack_tasks + 1); 753 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 754 if (space->GetMarkBitmap() == nullptr) { 755 continue; 756 } 757 uint8_t* card_begin = space->Begin(); 758 uint8_t* card_end = space->End(); 759 // Align up the end address. For example, the image space's end 760 // may not be card-size-aligned. 761 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 762 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin)); 763 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end)); 764 // Calculate how many bytes of heap we will scan, 765 const size_t address_range = card_end - card_begin; 766 // Calculate how much address range each task gets. 767 const size_t card_delta = RoundUp(address_range / thread_count + 1, 768 accounting::CardTable::kCardSize); 769 // If paused and the space is neither zygote nor image space, we could clear the dirty 770 // cards to avoid accumulating them to increase card scanning load in the following GC 771 // cycles. We need to keep dirty cards of image space and zygote space in order to track 772 // references to the other spaces. 773 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 774 // Create the worker tasks for this space. 775 while (card_begin != card_end) { 776 // Add a range of cards. 777 size_t addr_remaining = card_end - card_begin; 778 size_t card_increment = std::min(card_delta, addr_remaining); 779 // Take from the back of the mark stack. 780 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 781 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 782 mark_stack_end -= mark_stack_increment; 783 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 784 DCHECK_EQ(mark_stack_end, mark_stack_->End()); 785 // Add the new task to the thread pool. 786 auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin, 787 card_begin + card_increment, minimum_age, 788 mark_stack_increment, mark_stack_end, clear_card); 789 thread_pool->AddTask(self, task); 790 card_begin += card_increment; 791 } 792 } 793 794 // Note: the card scan below may dirty new cards (and scan them) 795 // as a side effect when a Reference object is encountered and 796 // queued during the marking. See b/11465268. 797 thread_pool->SetMaxActiveWorkers(thread_count - 1); 798 thread_pool->StartWorkers(self); 799 thread_pool->Wait(self, true, true); 800 thread_pool->StopWorkers(self); 801 } else { 802 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 803 if (space->GetMarkBitmap() != nullptr) { 804 // Image spaces are handled properly since live == marked for them. 805 const char* name = nullptr; 806 switch (space->GetGcRetentionPolicy()) { 807 case space::kGcRetentionPolicyNeverCollect: 808 name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects"; 809 break; 810 case space::kGcRetentionPolicyFullCollect: 811 name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects"; 812 break; 813 case space::kGcRetentionPolicyAlwaysCollect: 814 name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects"; 815 break; 816 default: 817 LOG(FATAL) << "Unreachable"; 818 UNREACHABLE(); 819 } 820 TimingLogger::ScopedTiming t(name, GetTimings()); 821 ScanObjectVisitor visitor(this); 822 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 823 if (clear_card) { 824 card_table->Scan<true>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, 825 minimum_age); 826 } else { 827 card_table->Scan<false>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, 828 minimum_age); 829 } 830 } 831 } 832 } 833} 834 835class RecursiveMarkTask : public MarkStackTask<false> { 836 public: 837 RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 838 accounting::ContinuousSpaceBitmap* bitmap, uintptr_t begin, uintptr_t end) 839 : MarkStackTask<false>(thread_pool, mark_sweep, 0, NULL), bitmap_(bitmap), begin_(begin), 840 end_(end) { 841 } 842 843 protected: 844 accounting::ContinuousSpaceBitmap* const bitmap_; 845 const uintptr_t begin_; 846 const uintptr_t end_; 847 848 virtual void Finalize() { 849 delete this; 850 } 851 852 // Scans all of the objects 853 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 854 ScanObjectParallelVisitor visitor(this); 855 bitmap_->VisitMarkedRange(begin_, end_, visitor); 856 // Finish by emptying our local mark stack. 857 MarkStackTask::Run(self); 858 } 859}; 860 861// Populates the mark stack based on the set of marked objects and 862// recursively marks until the mark stack is emptied. 863void MarkSweep::RecursiveMark() { 864 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 865 // RecursiveMark will build the lists of known instances of the Reference classes. See 866 // DelayReferenceReferent for details. 867 if (kUseRecursiveMark) { 868 const bool partial = GetGcType() == kGcTypePartial; 869 ScanObjectVisitor scan_visitor(this); 870 auto* self = Thread::Current(); 871 ThreadPool* thread_pool = heap_->GetThreadPool(); 872 size_t thread_count = GetThreadCount(false); 873 const bool parallel = kParallelRecursiveMark && thread_count > 1; 874 mark_stack_->Reset(); 875 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 876 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 877 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 878 current_space_bitmap_ = space->GetMarkBitmap(); 879 if (current_space_bitmap_ == nullptr) { 880 continue; 881 } 882 if (parallel) { 883 // We will use the mark stack the future. 884 // CHECK(mark_stack_->IsEmpty()); 885 // This function does not handle heap end increasing, so we must use the space end. 886 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 887 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 888 atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue()); 889 890 // Create a few worker tasks. 891 const size_t n = thread_count * 2; 892 while (begin != end) { 893 uintptr_t start = begin; 894 uintptr_t delta = (end - begin) / n; 895 delta = RoundUp(delta, KB); 896 if (delta < 16 * KB) delta = end - begin; 897 begin += delta; 898 auto* task = new RecursiveMarkTask(thread_pool, this, current_space_bitmap_, start, 899 begin); 900 thread_pool->AddTask(self, task); 901 } 902 thread_pool->SetMaxActiveWorkers(thread_count - 1); 903 thread_pool->StartWorkers(self); 904 thread_pool->Wait(self, true, true); 905 thread_pool->StopWorkers(self); 906 } else { 907 // This function does not handle heap end increasing, so we must use the space end. 908 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 909 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 910 current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 911 } 912 } 913 } 914 } 915 ProcessMarkStack(false); 916} 917 918mirror::Object* MarkSweep::IsMarkedCallback(mirror::Object* object, void* arg) { 919 if (reinterpret_cast<MarkSweep*>(arg)->IsMarked(object)) { 920 return object; 921 } 922 return nullptr; 923} 924 925void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) { 926 ScanGrayObjects(paused, minimum_age); 927 ProcessMarkStack(paused); 928} 929 930void MarkSweep::ReMarkRoots() { 931 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 932 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 933 Runtime::Current()->VisitRoots( 934 MarkRootCallback, this, static_cast<VisitRootFlags>(kVisitRootFlagNewRoots | 935 kVisitRootFlagStopLoggingNewRoots | 936 kVisitRootFlagClearRootLog)); 937 if (kVerifyRootsMarked) { 938 TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings()); 939 Runtime::Current()->VisitRoots(VerifyRootMarked, this); 940 } 941} 942 943void MarkSweep::SweepSystemWeaks(Thread* self) { 944 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 945 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 946 Runtime::Current()->SweepSystemWeaks(IsMarkedCallback, this); 947} 948 949mirror::Object* MarkSweep::VerifySystemWeakIsLiveCallback(Object* obj, void* arg) { 950 reinterpret_cast<MarkSweep*>(arg)->VerifyIsLive(obj); 951 // We don't actually want to sweep the object, so lets return "marked" 952 return obj; 953} 954 955void MarkSweep::VerifyIsLive(const Object* obj) { 956 if (!heap_->GetLiveBitmap()->Test(obj)) { 957 accounting::ObjectStack* allocation_stack = heap_->allocation_stack_.get(); 958 CHECK(std::find(allocation_stack->Begin(), allocation_stack->End(), obj) != 959 allocation_stack->End()) << "Found dead object " << obj << "\n" << heap_->DumpSpaces(); 960 } 961} 962 963void MarkSweep::VerifySystemWeaks() { 964 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 965 // Verify system weaks, uses a special object visitor which returns the input object. 966 Runtime::Current()->SweepSystemWeaks(VerifySystemWeakIsLiveCallback, this); 967} 968 969class CheckpointMarkThreadRoots : public Closure { 970 public: 971 explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep, 972 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) 973 : mark_sweep_(mark_sweep), 974 revoke_ros_alloc_thread_local_buffers_at_checkpoint_( 975 revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 976 } 977 978 virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS { 979 ATRACE_BEGIN("Marking thread roots"); 980 // Note: self is not necessarily equal to thread since thread may be suspended. 981 Thread* self = Thread::Current(); 982 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 983 << thread->GetState() << " thread " << thread << " self " << self; 984 thread->VisitRoots(MarkSweep::MarkRootParallelCallback, mark_sweep_); 985 ATRACE_END(); 986 if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) { 987 ATRACE_BEGIN("RevokeRosAllocThreadLocalBuffers"); 988 mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread); 989 ATRACE_END(); 990 } 991 mark_sweep_->GetBarrier().Pass(self); 992 } 993 994 private: 995 MarkSweep* const mark_sweep_; 996 const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_; 997}; 998 999void MarkSweep::MarkRootsCheckpoint(Thread* self, 1000 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 1001 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1002 CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint); 1003 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 1004 // Request the check point is run on all threads returning a count of the threads that must 1005 // run through the barrier including self. 1006 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1007 // Release locks then wait for all mutator threads to pass the barrier. 1008 // TODO: optimize to not release locks when there are no threads to wait for. 1009 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1010 Locks::mutator_lock_->SharedUnlock(self); 1011 { 1012 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun); 1013 gc_barrier_->Increment(self, barrier_count); 1014 } 1015 Locks::mutator_lock_->SharedLock(self); 1016 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1017} 1018 1019void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1020 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1021 Thread* self = Thread::Current(); 1022 mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>( 1023 sweep_array_free_buffer_mem_map_->BaseBegin()); 1024 size_t chunk_free_pos = 0; 1025 ObjectBytePair freed; 1026 ObjectBytePair freed_los; 1027 // How many objects are left in the array, modified after each space is swept. 1028 Object** objects = allocations->Begin(); 1029 size_t count = allocations->Size(); 1030 // Change the order to ensure that the non-moving space last swept as an optimization. 1031 std::vector<space::ContinuousSpace*> sweep_spaces; 1032 space::ContinuousSpace* non_moving_space = nullptr; 1033 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1034 if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) && 1035 space->GetLiveBitmap() != nullptr) { 1036 if (space == heap_->GetNonMovingSpace()) { 1037 non_moving_space = space; 1038 } else { 1039 sweep_spaces.push_back(space); 1040 } 1041 } 1042 } 1043 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1044 // the other alloc spaces as an optimization. 1045 if (non_moving_space != nullptr) { 1046 sweep_spaces.push_back(non_moving_space); 1047 } 1048 // Start by sweeping the continuous spaces. 1049 for (space::ContinuousSpace* space : sweep_spaces) { 1050 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1051 accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1052 accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1053 if (swap_bitmaps) { 1054 std::swap(live_bitmap, mark_bitmap); 1055 } 1056 Object** out = objects; 1057 for (size_t i = 0; i < count; ++i) { 1058 Object* obj = objects[i]; 1059 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1060 continue; 1061 } 1062 if (space->HasAddress(obj)) { 1063 // This object is in the space, remove it from the array and add it to the sweep buffer 1064 // if needed. 1065 if (!mark_bitmap->Test(obj)) { 1066 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1067 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1068 freed.objects += chunk_free_pos; 1069 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1070 chunk_free_pos = 0; 1071 } 1072 chunk_free_buffer[chunk_free_pos++] = obj; 1073 } 1074 } else { 1075 *(out++) = obj; 1076 } 1077 } 1078 if (chunk_free_pos > 0) { 1079 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1080 freed.objects += chunk_free_pos; 1081 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1082 chunk_free_pos = 0; 1083 } 1084 // All of the references which space contained are no longer in the allocation stack, update 1085 // the count. 1086 count = out - objects; 1087 } 1088 // Handle the large object space. 1089 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1090 if (large_object_space != nullptr) { 1091 accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap(); 1092 accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap(); 1093 if (swap_bitmaps) { 1094 std::swap(large_live_objects, large_mark_objects); 1095 } 1096 for (size_t i = 0; i < count; ++i) { 1097 Object* obj = objects[i]; 1098 // Handle large objects. 1099 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1100 continue; 1101 } 1102 if (!large_mark_objects->Test(obj)) { 1103 ++freed_los.objects; 1104 freed_los.bytes += large_object_space->Free(self, obj); 1105 } 1106 } 1107 } 1108 { 1109 TimingLogger::ScopedTiming t2("RecordFree", GetTimings()); 1110 RecordFree(freed); 1111 RecordFreeLOS(freed_los); 1112 t2.NewTiming("ResetStack"); 1113 allocations->Reset(); 1114 } 1115 sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero(); 1116} 1117 1118void MarkSweep::Sweep(bool swap_bitmaps) { 1119 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1120 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 1121 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 1122 { 1123 TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings()); 1124 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 1125 // knowing that new allocations won't be marked as live. 1126 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 1127 heap_->MarkAllocStackAsLive(live_stack); 1128 live_stack->Reset(); 1129 DCHECK(mark_stack_->IsEmpty()); 1130 } 1131 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1132 if (space->IsContinuousMemMapAllocSpace()) { 1133 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1134 TimingLogger::ScopedTiming split( 1135 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", GetTimings()); 1136 RecordFree(alloc_space->Sweep(swap_bitmaps)); 1137 } 1138 } 1139 SweepLargeObjects(swap_bitmaps); 1140} 1141 1142void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1143 space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace(); 1144 if (los != nullptr) { 1145 TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings()); 1146 RecordFreeLOS(los->Sweep(swap_bitmaps)); 1147 } 1148} 1149 1150// Process the "referent" field in a java.lang.ref.Reference. If the referent has not yet been 1151// marked, put it on the appropriate list in the heap for later processing. 1152void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) { 1153 if (kCountJavaLangRefs) { 1154 ++reference_count_; 1155 } 1156 heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, &HeapReferenceMarkedCallback, 1157 this); 1158} 1159 1160class MarkObjectVisitor { 1161 public: 1162 explicit MarkObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) { 1163 } 1164 1165 void operator()(Object* obj, MemberOffset offset, bool /* is_static */) const 1166 ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 1167 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1168 if (kCheckLocks) { 1169 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1170 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1171 } 1172 mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset)); 1173 } 1174 1175 private: 1176 MarkSweep* const mark_sweep_; 1177}; 1178 1179// Scans an object reference. Determines the type of the reference 1180// and dispatches to a specialized scanning routine. 1181void MarkSweep::ScanObject(Object* obj) { 1182 MarkObjectVisitor mark_visitor(this); 1183 DelayReferenceReferentVisitor ref_visitor(this); 1184 ScanObjectVisit(obj, mark_visitor, ref_visitor); 1185} 1186 1187void MarkSweep::ProcessMarkStackCallback(void* arg) { 1188 reinterpret_cast<MarkSweep*>(arg)->ProcessMarkStack(false); 1189} 1190 1191void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1192 Thread* self = Thread::Current(); 1193 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1194 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1195 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1196 CHECK_GT(chunk_size, 0U); 1197 // Split the current mark stack up into work tasks. 1198 for (mirror::Object **it = mark_stack_->Begin(), **end = mark_stack_->End(); it < end; ) { 1199 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1200 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it)); 1201 it += delta; 1202 } 1203 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1204 thread_pool->StartWorkers(self); 1205 thread_pool->Wait(self, true, true); 1206 thread_pool->StopWorkers(self); 1207 mark_stack_->Reset(); 1208 CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(), 1209 work_chunks_deleted_.LoadSequentiallyConsistent()) 1210 << " some of the work chunks were leaked"; 1211} 1212 1213// Scan anything that's on the mark stack. 1214void MarkSweep::ProcessMarkStack(bool paused) { 1215 TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings()); 1216 size_t thread_count = GetThreadCount(paused); 1217 if (kParallelProcessMarkStack && thread_count > 1 && 1218 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1219 ProcessMarkStackParallel(thread_count); 1220 } else { 1221 // TODO: Tune this. 1222 static const size_t kFifoSize = 4; 1223 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 1224 for (;;) { 1225 Object* obj = NULL; 1226 if (kUseMarkStackPrefetch) { 1227 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1228 Object* mark_stack_obj = mark_stack_->PopBack(); 1229 DCHECK(mark_stack_obj != NULL); 1230 __builtin_prefetch(mark_stack_obj); 1231 prefetch_fifo.push_back(mark_stack_obj); 1232 } 1233 if (prefetch_fifo.empty()) { 1234 break; 1235 } 1236 obj = prefetch_fifo.front(); 1237 prefetch_fifo.pop_front(); 1238 } else { 1239 if (mark_stack_->IsEmpty()) { 1240 break; 1241 } 1242 obj = mark_stack_->PopBack(); 1243 } 1244 DCHECK(obj != nullptr); 1245 ScanObject(obj); 1246 } 1247 } 1248} 1249 1250inline bool MarkSweep::IsMarked(const Object* object) const { 1251 if (immune_region_.ContainsObject(object)) { 1252 return true; 1253 } 1254 if (current_space_bitmap_->HasAddress(object)) { 1255 return current_space_bitmap_->Test(object); 1256 } 1257 return mark_bitmap_->Test(object); 1258} 1259 1260void MarkSweep::FinishPhase() { 1261 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1262 if (kCountScannedTypes) { 1263 VLOG(gc) << "MarkSweep scanned classes=" << class_count_.LoadRelaxed() 1264 << " arrays=" << array_count_.LoadRelaxed() << " other=" << other_count_.LoadRelaxed(); 1265 } 1266 if (kCountTasks) { 1267 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed(); 1268 } 1269 if (kMeasureOverhead) { 1270 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed()); 1271 } 1272 if (kProfileLargeObjects) { 1273 VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed() 1274 << " marked " << large_object_mark_.LoadRelaxed(); 1275 } 1276 if (kCountJavaLangRefs) { 1277 VLOG(gc) << "References scanned " << reference_count_.LoadRelaxed(); 1278 } 1279 if (kCountMarkedObjects) { 1280 VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed() 1281 << " immune=" << mark_immune_count_.LoadRelaxed() 1282 << " fastpath=" << mark_fastpath_count_.LoadRelaxed() 1283 << " slowpath=" << mark_slowpath_count_.LoadRelaxed(); 1284 } 1285 CHECK(mark_stack_->IsEmpty()); // Ensure that the mark stack is empty. 1286 mark_stack_->Reset(); 1287 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 1288 heap_->ClearMarkedObjects(); 1289} 1290 1291void MarkSweep::RevokeAllThreadLocalBuffers() { 1292 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) { 1293 // If concurrent, rosalloc thread-local buffers are revoked at the 1294 // thread checkpoint. Bump pointer space thread-local buffers must 1295 // not be in use. 1296 GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked(); 1297 } else { 1298 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1299 GetHeap()->RevokeAllThreadLocalBuffers(); 1300 } 1301} 1302 1303} // namespace collector 1304} // namespace gc 1305} // namespace art 1306