mark_sweep.cc revision 8118781ebc9659f806716c451bdb3fe9b77ae32b
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#define ATRACE_TAG ATRACE_TAG_DALVIK 26#include "cutils/trace.h" 27 28#include "base/bounded_fifo.h" 29#include "base/logging.h" 30#include "base/macros.h" 31#include "base/mutex-inl.h" 32#include "base/time_utils.h" 33#include "base/timing_logger.h" 34#include "gc/accounting/card_table-inl.h" 35#include "gc/accounting/heap_bitmap-inl.h" 36#include "gc/accounting/mod_union_table.h" 37#include "gc/accounting/space_bitmap-inl.h" 38#include "gc/heap.h" 39#include "gc/reference_processor.h" 40#include "gc/space/large_object_space.h" 41#include "gc/space/space-inl.h" 42#include "mark_sweep-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 49namespace art { 50namespace gc { 51namespace collector { 52 53// Performance options. 54static constexpr bool kUseRecursiveMark = false; 55static constexpr bool kUseMarkStackPrefetch = true; 56static constexpr size_t kSweepArrayChunkFreeSize = 1024; 57static constexpr bool kPreCleanCards = true; 58 59// Parallelism options. 60static constexpr bool kParallelCardScan = true; 61static constexpr bool kParallelRecursiveMark = true; 62// Don't attempt to parallelize mark stack processing unless the mark stack is at least n 63// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not 64// having this can add overhead in ProcessReferences since we may end up doing many calls of 65// ProcessMarkStack with very small mark stacks. 66static constexpr size_t kMinimumParallelMarkStackSize = 128; 67static constexpr bool kParallelProcessMarkStack = true; 68 69// Profiling and information flags. 70static constexpr bool kProfileLargeObjects = false; 71static constexpr bool kMeasureOverhead = false; 72static constexpr bool kCountTasks = false; 73static constexpr bool kCountJavaLangRefs = false; 74static constexpr bool kCountMarkedObjects = false; 75 76// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%. 77static constexpr bool kCheckLocks = kDebugLocking; 78static constexpr bool kVerifyRootsMarked = kIsDebugBuild; 79 80// If true, revoke the rosalloc thread-local buffers at the 81// checkpoint, as opposed to during the pause. 82static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true; 83 84void MarkSweep::BindBitmaps() { 85 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 86 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 87 // Mark all of the spaces we never collect as immune. 88 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 89 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 90 CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space; 91 } 92 } 93} 94 95MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix) 96 : GarbageCollector(heap, 97 name_prefix + 98 (is_concurrent ? "concurrent mark sweep": "mark sweep")), 99 current_space_bitmap_(nullptr), mark_bitmap_(nullptr), mark_stack_(nullptr), 100 gc_barrier_(new Barrier(0)), 101 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 102 is_concurrent_(is_concurrent), live_stack_freeze_size_(0) { 103 std::string error_msg; 104 MemMap* mem_map = MemMap::MapAnonymous( 105 "mark sweep sweep array free buffer", nullptr, 106 RoundUp(kSweepArrayChunkFreeSize * sizeof(mirror::Object*), kPageSize), 107 PROT_READ | PROT_WRITE, false, false, &error_msg); 108 CHECK(mem_map != nullptr) << "Couldn't allocate sweep array free buffer: " << error_msg; 109 sweep_array_free_buffer_mem_map_.reset(mem_map); 110} 111 112void MarkSweep::InitializePhase() { 113 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 114 mark_stack_ = heap_->GetMarkStack(); 115 DCHECK(mark_stack_ != nullptr); 116 immune_region_.Reset(); 117 class_count_.StoreRelaxed(0); 118 array_count_.StoreRelaxed(0); 119 other_count_.StoreRelaxed(0); 120 large_object_test_.StoreRelaxed(0); 121 large_object_mark_.StoreRelaxed(0); 122 overhead_time_ .StoreRelaxed(0); 123 work_chunks_created_.StoreRelaxed(0); 124 work_chunks_deleted_.StoreRelaxed(0); 125 reference_count_.StoreRelaxed(0); 126 mark_null_count_.StoreRelaxed(0); 127 mark_immune_count_.StoreRelaxed(0); 128 mark_fastpath_count_.StoreRelaxed(0); 129 mark_slowpath_count_.StoreRelaxed(0); 130 { 131 // TODO: I don't think we should need heap bitmap lock to Get the mark bitmap. 132 ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 133 mark_bitmap_ = heap_->GetMarkBitmap(); 134 } 135 if (!GetCurrentIteration()->GetClearSoftReferences()) { 136 // Always clear soft references if a non-sticky collection. 137 GetCurrentIteration()->SetClearSoftReferences(GetGcType() != collector::kGcTypeSticky); 138 } 139} 140 141void MarkSweep::RunPhases() { 142 Thread* self = Thread::Current(); 143 InitializePhase(); 144 Locks::mutator_lock_->AssertNotHeld(self); 145 if (IsConcurrent()) { 146 GetHeap()->PreGcVerification(this); 147 { 148 ReaderMutexLock mu(self, *Locks::mutator_lock_); 149 MarkingPhase(); 150 } 151 ScopedPause pause(this); 152 GetHeap()->PrePauseRosAllocVerification(this); 153 PausePhase(); 154 RevokeAllThreadLocalBuffers(); 155 } else { 156 ScopedPause pause(this); 157 GetHeap()->PreGcVerificationPaused(this); 158 MarkingPhase(); 159 GetHeap()->PrePauseRosAllocVerification(this); 160 PausePhase(); 161 RevokeAllThreadLocalBuffers(); 162 } 163 { 164 // Sweeping always done concurrently, even for non concurrent mark sweep. 165 ReaderMutexLock mu(self, *Locks::mutator_lock_); 166 ReclaimPhase(); 167 } 168 GetHeap()->PostGcVerification(this); 169 FinishPhase(); 170} 171 172void MarkSweep::ProcessReferences(Thread* self) { 173 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 174 GetHeap()->GetReferenceProcessor()->ProcessReferences( 175 true, GetTimings(), GetCurrentIteration()->GetClearSoftReferences(), this); 176} 177 178void MarkSweep::PausePhase() { 179 TimingLogger::ScopedTiming t("(Paused)PausePhase", GetTimings()); 180 Thread* self = Thread::Current(); 181 Locks::mutator_lock_->AssertExclusiveHeld(self); 182 if (IsConcurrent()) { 183 // Handle the dirty objects if we are a concurrent GC. 184 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 185 // Re-mark root set. 186 ReMarkRoots(); 187 // Scan dirty objects, this is only required if we are not doing concurrent GC. 188 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 189 } 190 { 191 TimingLogger::ScopedTiming t2("SwapStacks", GetTimings()); 192 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 193 heap_->SwapStacks(self); 194 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 195 // Need to revoke all the thread local allocation stacks since we just swapped the allocation 196 // stacks and don't want anybody to allocate into the live stack. 197 RevokeAllThreadLocalAllocationStacks(self); 198 } 199 heap_->PreSweepingGcVerification(this); 200 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 201 // weak before we sweep them. Since this new system weak may not be marked, the GC may 202 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 203 // reference to a string that is about to be swept. 204 Runtime::Current()->DisallowNewSystemWeaks(); 205 // Enable the reference processing slow path, needs to be done with mutators paused since there 206 // is no lock in the GetReferent fast path. 207 GetHeap()->GetReferenceProcessor()->EnableSlowPath(); 208} 209 210void MarkSweep::PreCleanCards() { 211 // Don't do this for non concurrent GCs since they don't have any dirty cards. 212 if (kPreCleanCards && IsConcurrent()) { 213 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 214 Thread* self = Thread::Current(); 215 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 216 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 217 heap_->ProcessCards(GetTimings(), false, true, false); 218 // The checkpoint root marking is required to avoid a race condition which occurs if the 219 // following happens during a reference write: 220 // 1. mutator dirties the card (write barrier) 221 // 2. GC ages the card (the above ProcessCards call) 222 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 223 // 4. mutator writes the value (corresponding to the write barrier in 1.) 224 // This causes the GC to age the card but not necessarily mark the reference which the mutator 225 // wrote into the object stored in the card. 226 // Having the checkpoint fixes this issue since it ensures that the card mark and the 227 // reference write are visible to the GC before the card is scanned (this is due to locks being 228 // acquired / released in the checkpoint code). 229 // The other roots are also marked to help reduce the pause. 230 MarkRootsCheckpoint(self, false); 231 MarkNonThreadRoots(); 232 MarkConcurrentRoots( 233 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 234 // Process the newly aged cards. 235 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 236 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 237 // in the next GC. 238 } 239} 240 241void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 242 if (kUseThreadLocalAllocationStack) { 243 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 244 Locks::mutator_lock_->AssertExclusiveHeld(self); 245 heap_->RevokeAllThreadLocalAllocationStacks(self); 246 } 247} 248 249void MarkSweep::MarkingPhase() { 250 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 251 Thread* self = Thread::Current(); 252 BindBitmaps(); 253 FindDefaultSpaceBitmap(); 254 // Process dirty cards and add dirty cards to mod union tables. 255 // If the GC type is non sticky, then we just clear the cards instead of ageing them. 256 heap_->ProcessCards(GetTimings(), false, true, GetGcType() != kGcTypeSticky); 257 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 258 MarkRoots(self); 259 MarkReachableObjects(); 260 // Pre-clean dirtied cards to reduce pauses. 261 PreCleanCards(); 262} 263 264void MarkSweep::UpdateAndMarkModUnion() { 265 for (const auto& space : heap_->GetContinuousSpaces()) { 266 if (immune_region_.ContainsSpace(space)) { 267 const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" : 268 "UpdateAndMarkImageModUnionTable"; 269 TimingLogger::ScopedTiming t(name, GetTimings()); 270 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 271 CHECK(mod_union_table != nullptr); 272 mod_union_table->UpdateAndMarkReferences(this); 273 } 274 } 275} 276 277void MarkSweep::MarkReachableObjects() { 278 UpdateAndMarkModUnion(); 279 // Recursively mark all the non-image bits set in the mark bitmap. 280 RecursiveMark(); 281} 282 283void MarkSweep::ReclaimPhase() { 284 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 285 Thread* self = Thread::Current(); 286 // Process the references concurrently. 287 ProcessReferences(self); 288 SweepSystemWeaks(self); 289 Runtime::Current()->AllowNewSystemWeaks(); 290 { 291 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 292 GetHeap()->RecordFreeRevoke(); 293 // Reclaim unmarked objects. 294 Sweep(false); 295 // Swap the live and mark bitmaps for each space which we modified space. This is an 296 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 297 // bitmaps. 298 SwapBitmaps(); 299 // Unbind the live and mark bitmaps. 300 GetHeap()->UnBindBitmaps(); 301 } 302} 303 304void MarkSweep::FindDefaultSpaceBitmap() { 305 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 306 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 307 accounting::ContinuousSpaceBitmap* bitmap = space->GetMarkBitmap(); 308 // We want to have the main space instead of non moving if possible. 309 if (bitmap != nullptr && 310 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 311 current_space_bitmap_ = bitmap; 312 // If we are not the non moving space exit the loop early since this will be good enough. 313 if (space != heap_->GetNonMovingSpace()) { 314 break; 315 } 316 } 317 } 318 CHECK(current_space_bitmap_ != nullptr) << "Could not find a default mark bitmap\n" 319 << heap_->DumpSpaces(); 320} 321 322void MarkSweep::ExpandMarkStack() { 323 ResizeMarkStack(mark_stack_->Capacity() * 2); 324} 325 326void MarkSweep::ResizeMarkStack(size_t new_size) { 327 // Rare case, no need to have Thread::Current be a parameter. 328 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 329 // Someone else acquired the lock and expanded the mark stack before us. 330 return; 331 } 332 std::vector<StackReference<mirror::Object>> temp(mark_stack_->Begin(), mark_stack_->End()); 333 CHECK_LE(mark_stack_->Size(), new_size); 334 mark_stack_->Resize(new_size); 335 for (auto& obj : temp) { 336 mark_stack_->PushBack(obj.AsMirrorPtr()); 337 } 338} 339 340inline void MarkSweep::MarkObjectNonNullParallel(mirror::Object* obj) { 341 DCHECK(obj != nullptr); 342 if (MarkObjectParallel(obj)) { 343 MutexLock mu(Thread::Current(), mark_stack_lock_); 344 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 345 ExpandMarkStack(); 346 } 347 // The object must be pushed on to the mark stack. 348 mark_stack_->PushBack(obj); 349 } 350} 351 352bool MarkSweep::IsMarkedHeapReference(mirror::HeapReference<mirror::Object>* ref) { 353 return IsMarked(ref->AsMirrorPtr()); 354} 355 356class MarkSweepMarkObjectSlowPath { 357 public: 358 explicit MarkSweepMarkObjectSlowPath(MarkSweep* mark_sweep, mirror::Object* holder = nullptr, 359 MemberOffset offset = MemberOffset(0)) 360 : mark_sweep_(mark_sweep), holder_(holder), offset_(offset) { 361 } 362 363 void operator()(const mirror::Object* obj) const ALWAYS_INLINE NO_THREAD_SAFETY_ANALYSIS { 364 if (kProfileLargeObjects) { 365 // TODO: Differentiate between marking and testing somehow. 366 ++mark_sweep_->large_object_test_; 367 ++mark_sweep_->large_object_mark_; 368 } 369 space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace(); 370 if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) || 371 (kIsDebugBuild && large_object_space != nullptr && 372 !large_object_space->Contains(obj)))) { 373 LOG(INTERNAL_FATAL) << "Tried to mark " << obj << " not contained by any spaces"; 374 if (holder_ != nullptr) { 375 size_t holder_size = holder_->SizeOf(); 376 ArtField* field = holder_->FindFieldByOffset(offset_); 377 LOG(INTERNAL_FATAL) << "Field info: " 378 << " holder=" << holder_ 379 << " holder is " 380 << (mark_sweep_->GetHeap()->IsLiveObjectLocked(holder_) 381 ? "alive" : "dead") 382 << " holder_size=" << holder_size 383 << " holder_type=" << PrettyTypeOf(holder_) 384 << " offset=" << offset_.Uint32Value() 385 << " field=" << (field != nullptr ? field->GetName() : "nullptr") 386 << " field_type=" 387 << (field != nullptr ? field->GetTypeDescriptor() : "") 388 << " first_ref_field_offset=" 389 << (holder_->IsClass() 390 ? holder_->AsClass()->GetFirstReferenceStaticFieldOffset( 391 sizeof(void*)) 392 : holder_->GetClass()->GetFirstReferenceInstanceFieldOffset()) 393 << " num_of_ref_fields=" 394 << (holder_->IsClass() 395 ? holder_->AsClass()->NumReferenceStaticFields() 396 : holder_->GetClass()->NumReferenceInstanceFields()) 397 << "\n"; 398 // Print the memory content of the holder. 399 for (size_t i = 0; i < holder_size / sizeof(uint32_t); ++i) { 400 uint32_t* p = reinterpret_cast<uint32_t*>(holder_); 401 LOG(INTERNAL_FATAL) << &p[i] << ": " << "holder+" << (i * sizeof(uint32_t)) << " = " 402 << std::hex << p[i]; 403 } 404 } 405 PrintFileToLog("/proc/self/maps", LogSeverity::INTERNAL_FATAL); 406 MemMap::DumpMaps(LOG(INTERNAL_FATAL), true); 407 { 408 LOG(INTERNAL_FATAL) << "Attempting see if it's a bad root"; 409 Thread* self = Thread::Current(); 410 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 411 mark_sweep_->VerifyRoots(); 412 } else { 413 const bool heap_bitmap_exclusive_locked = 414 Locks::heap_bitmap_lock_->IsExclusiveHeld(self); 415 if (heap_bitmap_exclusive_locked) { 416 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 417 } 418 Locks::mutator_lock_->SharedUnlock(self); 419 ThreadList* tl = Runtime::Current()->GetThreadList(); 420 tl->SuspendAll(__FUNCTION__); 421 mark_sweep_->VerifyRoots(); 422 tl->ResumeAll(); 423 Locks::mutator_lock_->SharedLock(self); 424 if (heap_bitmap_exclusive_locked) { 425 Locks::heap_bitmap_lock_->ExclusiveLock(self); 426 } 427 } 428 } 429 LOG(FATAL) << "Can't mark invalid object"; 430 } 431 } 432 433 private: 434 MarkSweep* const mark_sweep_; 435 mirror::Object* const holder_; 436 MemberOffset offset_; 437}; 438 439inline void MarkSweep::MarkObjectNonNull(mirror::Object* obj, mirror::Object* holder, 440 MemberOffset offset) { 441 DCHECK(obj != nullptr); 442 if (kUseBakerOrBrooksReadBarrier) { 443 // Verify all the objects have the correct pointer installed. 444 obj->AssertReadBarrierPointer(); 445 } 446 if (immune_region_.ContainsObject(obj)) { 447 if (kCountMarkedObjects) { 448 ++mark_immune_count_; 449 } 450 DCHECK(mark_bitmap_->Test(obj)); 451 } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) { 452 if (kCountMarkedObjects) { 453 ++mark_fastpath_count_; 454 } 455 if (UNLIKELY(!current_space_bitmap_->Set(obj))) { 456 PushOnMarkStack(obj); // This object was not previously marked. 457 } 458 } else { 459 if (kCountMarkedObjects) { 460 ++mark_slowpath_count_; 461 } 462 MarkSweepMarkObjectSlowPath visitor(this, holder, offset); 463 // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set 464 // will check again. 465 if (!mark_bitmap_->Set(obj, visitor)) { 466 PushOnMarkStack(obj); // Was not already marked, push. 467 } 468 } 469} 470 471inline void MarkSweep::PushOnMarkStack(mirror::Object* obj) { 472 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 473 // Lock is not needed but is here anyways to please annotalysis. 474 MutexLock mu(Thread::Current(), mark_stack_lock_); 475 ExpandMarkStack(); 476 } 477 // The object must be pushed on to the mark stack. 478 mark_stack_->PushBack(obj); 479} 480 481inline bool MarkSweep::MarkObjectParallel(mirror::Object* obj) { 482 DCHECK(obj != nullptr); 483 if (kUseBakerOrBrooksReadBarrier) { 484 // Verify all the objects have the correct pointer installed. 485 obj->AssertReadBarrierPointer(); 486 } 487 if (immune_region_.ContainsObject(obj)) { 488 DCHECK(IsMarked(obj) != nullptr); 489 return false; 490 } 491 // Try to take advantage of locality of references within a space, failing this find the space 492 // the hard way. 493 accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_; 494 if (LIKELY(object_bitmap->HasAddress(obj))) { 495 return !object_bitmap->AtomicTestAndSet(obj); 496 } 497 MarkSweepMarkObjectSlowPath visitor(this); 498 return !mark_bitmap_->AtomicTestAndSet(obj, visitor); 499} 500 501void MarkSweep::MarkHeapReference(mirror::HeapReference<mirror::Object>* ref) { 502 MarkObject(ref->AsMirrorPtr(), nullptr, MemberOffset(0)); 503} 504 505// Used to mark objects when processing the mark stack. If an object is null, it is not marked. 506inline void MarkSweep::MarkObject(mirror::Object* obj, mirror::Object* holder, 507 MemberOffset offset) { 508 if (obj != nullptr) { 509 MarkObjectNonNull(obj, holder, offset); 510 } else if (kCountMarkedObjects) { 511 ++mark_null_count_; 512 } 513} 514 515class VerifyRootMarkedVisitor : public SingleRootVisitor { 516 public: 517 explicit VerifyRootMarkedVisitor(MarkSweep* collector) : collector_(collector) { } 518 519 void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE 520 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 521 CHECK(collector_->IsMarked(root) != nullptr) << info.ToString(); 522 } 523 524 private: 525 MarkSweep* const collector_; 526}; 527 528void MarkSweep::VisitRoots(mirror::Object*** roots, size_t count, 529 const RootInfo& info ATTRIBUTE_UNUSED) { 530 for (size_t i = 0; i < count; ++i) { 531 MarkObjectNonNull(*roots[i]); 532 } 533} 534 535void MarkSweep::VisitRoots(mirror::CompressedReference<mirror::Object>** roots, size_t count, 536 const RootInfo& info ATTRIBUTE_UNUSED) { 537 for (size_t i = 0; i < count; ++i) { 538 MarkObjectNonNull(roots[i]->AsMirrorPtr()); 539 } 540} 541 542class VerifyRootVisitor : public SingleRootVisitor { 543 public: 544 void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE 545 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 546 // See if the root is on any space bitmap. 547 auto* heap = Runtime::Current()->GetHeap(); 548 if (heap->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) { 549 space::LargeObjectSpace* large_object_space = heap->GetLargeObjectsSpace(); 550 if (large_object_space != nullptr && !large_object_space->Contains(root)) { 551 LOG(INTERNAL_FATAL) << "Found invalid root: " << root << " " << info; 552 } 553 } 554 } 555}; 556 557void MarkSweep::VerifyRoots() { 558 VerifyRootVisitor visitor; 559 Runtime::Current()->GetThreadList()->VisitRoots(&visitor); 560} 561 562void MarkSweep::MarkRoots(Thread* self) { 563 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 564 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 565 // If we exclusively hold the mutator lock, all threads must be suspended. 566 Runtime::Current()->VisitRoots(this); 567 RevokeAllThreadLocalAllocationStacks(self); 568 } else { 569 MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint); 570 // At this point the live stack should no longer have any mutators which push into it. 571 MarkNonThreadRoots(); 572 MarkConcurrentRoots( 573 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 574 } 575} 576 577void MarkSweep::MarkNonThreadRoots() { 578 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 579 Runtime::Current()->VisitNonThreadRoots(this); 580} 581 582void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 583 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 584 // Visit all runtime roots and clear dirty flags. 585 Runtime::Current()->VisitConcurrentRoots( 586 this, static_cast<VisitRootFlags>(flags | kVisitRootFlagNonMoving)); 587} 588 589class ScanObjectVisitor { 590 public: 591 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 592 : mark_sweep_(mark_sweep) {} 593 594 void operator()(mirror::Object* obj) const ALWAYS_INLINE 595 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 596 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 597 if (kCheckLocks) { 598 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 599 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 600 } 601 mark_sweep_->ScanObject(obj); 602 } 603 604 private: 605 MarkSweep* const mark_sweep_; 606}; 607 608class DelayReferenceReferentVisitor { 609 public: 610 explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) { 611 } 612 613 void operator()(mirror::Class* klass, mirror::Reference* ref) const 614 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 615 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 616 collector_->DelayReferenceReferent(klass, ref); 617 } 618 619 private: 620 MarkSweep* const collector_; 621}; 622 623template <bool kUseFinger = false> 624class MarkStackTask : public Task { 625 public: 626 MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size, 627 StackReference<mirror::Object>* mark_stack) 628 : mark_sweep_(mark_sweep), 629 thread_pool_(thread_pool), 630 mark_stack_pos_(mark_stack_size) { 631 // We may have to copy part of an existing mark stack when another mark stack overflows. 632 if (mark_stack_size != 0) { 633 DCHECK(mark_stack != nullptr); 634 // TODO: Check performance? 635 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 636 } 637 if (kCountTasks) { 638 ++mark_sweep_->work_chunks_created_; 639 } 640 } 641 642 static const size_t kMaxSize = 1 * KB; 643 644 protected: 645 class MarkObjectParallelVisitor { 646 public: 647 explicit MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task, 648 MarkSweep* mark_sweep) ALWAYS_INLINE 649 : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {} 650 651 void operator()(mirror::Object* obj, MemberOffset offset, bool /* static */) const ALWAYS_INLINE 652 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 653 mirror::Object* ref = obj->GetFieldObject<mirror::Object>(offset); 654 if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) { 655 if (kUseFinger) { 656 std::atomic_thread_fence(std::memory_order_seq_cst); 657 if (reinterpret_cast<uintptr_t>(ref) >= 658 static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) { 659 return; 660 } 661 } 662 chunk_task_->MarkStackPush(ref); 663 } 664 } 665 666 private: 667 MarkStackTask<kUseFinger>* const chunk_task_; 668 MarkSweep* const mark_sweep_; 669 }; 670 671 class ScanObjectParallelVisitor { 672 public: 673 explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE 674 : chunk_task_(chunk_task) {} 675 676 // No thread safety analysis since multiple threads will use this visitor. 677 void operator()(mirror::Object* obj) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 678 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 679 MarkSweep* const mark_sweep = chunk_task_->mark_sweep_; 680 MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep); 681 DelayReferenceReferentVisitor ref_visitor(mark_sweep); 682 mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor); 683 } 684 685 private: 686 MarkStackTask<kUseFinger>* const chunk_task_; 687 }; 688 689 virtual ~MarkStackTask() { 690 // Make sure that we have cleared our mark stack. 691 DCHECK_EQ(mark_stack_pos_, 0U); 692 if (kCountTasks) { 693 ++mark_sweep_->work_chunks_deleted_; 694 } 695 } 696 697 MarkSweep* const mark_sweep_; 698 ThreadPool* const thread_pool_; 699 // Thread local mark stack for this task. 700 StackReference<mirror::Object> mark_stack_[kMaxSize]; 701 // Mark stack position. 702 size_t mark_stack_pos_; 703 704 ALWAYS_INLINE void MarkStackPush(mirror::Object* obj) 705 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 706 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 707 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 708 mark_stack_pos_ /= 2; 709 auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_, 710 mark_stack_ + mark_stack_pos_); 711 thread_pool_->AddTask(Thread::Current(), task); 712 } 713 DCHECK(obj != nullptr); 714 DCHECK_LT(mark_stack_pos_, kMaxSize); 715 mark_stack_[mark_stack_pos_++].Assign(obj); 716 } 717 718 virtual void Finalize() { 719 delete this; 720 } 721 722 // Scans all of the objects 723 virtual void Run(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 724 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 725 UNUSED(self); 726 ScanObjectParallelVisitor visitor(this); 727 // TODO: Tune this. 728 static const size_t kFifoSize = 4; 729 BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo; 730 for (;;) { 731 mirror::Object* obj = nullptr; 732 if (kUseMarkStackPrefetch) { 733 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 734 mirror::Object* const mark_stack_obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr(); 735 DCHECK(mark_stack_obj != nullptr); 736 __builtin_prefetch(mark_stack_obj); 737 prefetch_fifo.push_back(mark_stack_obj); 738 } 739 if (UNLIKELY(prefetch_fifo.empty())) { 740 break; 741 } 742 obj = prefetch_fifo.front(); 743 prefetch_fifo.pop_front(); 744 } else { 745 if (UNLIKELY(mark_stack_pos_ == 0)) { 746 break; 747 } 748 obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr(); 749 } 750 DCHECK(obj != nullptr); 751 visitor(obj); 752 } 753 } 754}; 755 756class CardScanTask : public MarkStackTask<false> { 757 public: 758 CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 759 accounting::ContinuousSpaceBitmap* bitmap, 760 uint8_t* begin, uint8_t* end, uint8_t minimum_age, size_t mark_stack_size, 761 StackReference<mirror::Object>* mark_stack_obj, bool clear_card) 762 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 763 bitmap_(bitmap), 764 begin_(begin), 765 end_(end), 766 minimum_age_(minimum_age), clear_card_(clear_card) { 767 } 768 769 protected: 770 accounting::ContinuousSpaceBitmap* const bitmap_; 771 uint8_t* const begin_; 772 uint8_t* const end_; 773 const uint8_t minimum_age_; 774 const bool clear_card_; 775 776 virtual void Finalize() { 777 delete this; 778 } 779 780 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 781 ScanObjectParallelVisitor visitor(this); 782 accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable(); 783 size_t cards_scanned = clear_card_ ? 784 card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_) : 785 card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_); 786 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 787 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 788 // Finish by emptying our local mark stack. 789 MarkStackTask::Run(self); 790 } 791}; 792 793size_t MarkSweep::GetThreadCount(bool paused) const { 794 if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) { 795 return 1; 796 } 797 return (paused ? heap_->GetParallelGCThreadCount() : heap_->GetConcGCThreadCount()) + 1; 798} 799 800void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) { 801 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 802 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 803 size_t thread_count = GetThreadCount(paused); 804 // The parallel version with only one thread is faster for card scanning, TODO: fix. 805 if (kParallelCardScan && thread_count > 1) { 806 Thread* self = Thread::Current(); 807 // Can't have a different split for each space since multiple spaces can have their cards being 808 // scanned at the same time. 809 TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__, 810 GetTimings()); 811 // Try to take some of the mark stack since we can pass this off to the worker tasks. 812 StackReference<mirror::Object>* mark_stack_begin = mark_stack_->Begin(); 813 StackReference<mirror::Object>* mark_stack_end = mark_stack_->End(); 814 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 815 // Estimated number of work tasks we will create. 816 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 817 DCHECK_NE(mark_stack_tasks, 0U); 818 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 819 mark_stack_size / mark_stack_tasks + 1); 820 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 821 if (space->GetMarkBitmap() == nullptr) { 822 continue; 823 } 824 uint8_t* card_begin = space->Begin(); 825 uint8_t* card_end = space->End(); 826 // Align up the end address. For example, the image space's end 827 // may not be card-size-aligned. 828 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 829 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin)); 830 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end)); 831 // Calculate how many bytes of heap we will scan, 832 const size_t address_range = card_end - card_begin; 833 // Calculate how much address range each task gets. 834 const size_t card_delta = RoundUp(address_range / thread_count + 1, 835 accounting::CardTable::kCardSize); 836 // If paused and the space is neither zygote nor image space, we could clear the dirty 837 // cards to avoid accumulating them to increase card scanning load in the following GC 838 // cycles. We need to keep dirty cards of image space and zygote space in order to track 839 // references to the other spaces. 840 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 841 // Create the worker tasks for this space. 842 while (card_begin != card_end) { 843 // Add a range of cards. 844 size_t addr_remaining = card_end - card_begin; 845 size_t card_increment = std::min(card_delta, addr_remaining); 846 // Take from the back of the mark stack. 847 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 848 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 849 mark_stack_end -= mark_stack_increment; 850 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 851 DCHECK_EQ(mark_stack_end, mark_stack_->End()); 852 // Add the new task to the thread pool. 853 auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin, 854 card_begin + card_increment, minimum_age, 855 mark_stack_increment, mark_stack_end, clear_card); 856 thread_pool->AddTask(self, task); 857 card_begin += card_increment; 858 } 859 } 860 861 // Note: the card scan below may dirty new cards (and scan them) 862 // as a side effect when a Reference object is encountered and 863 // queued during the marking. See b/11465268. 864 thread_pool->SetMaxActiveWorkers(thread_count - 1); 865 thread_pool->StartWorkers(self); 866 thread_pool->Wait(self, true, true); 867 thread_pool->StopWorkers(self); 868 } else { 869 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 870 if (space->GetMarkBitmap() != nullptr) { 871 // Image spaces are handled properly since live == marked for them. 872 const char* name = nullptr; 873 switch (space->GetGcRetentionPolicy()) { 874 case space::kGcRetentionPolicyNeverCollect: 875 name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects"; 876 break; 877 case space::kGcRetentionPolicyFullCollect: 878 name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects"; 879 break; 880 case space::kGcRetentionPolicyAlwaysCollect: 881 name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects"; 882 break; 883 default: 884 LOG(FATAL) << "Unreachable"; 885 UNREACHABLE(); 886 } 887 TimingLogger::ScopedTiming t(name, GetTimings()); 888 ScanObjectVisitor visitor(this); 889 bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace(); 890 if (clear_card) { 891 card_table->Scan<true>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, 892 minimum_age); 893 } else { 894 card_table->Scan<false>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, 895 minimum_age); 896 } 897 } 898 } 899 } 900} 901 902class RecursiveMarkTask : public MarkStackTask<false> { 903 public: 904 RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 905 accounting::ContinuousSpaceBitmap* bitmap, uintptr_t begin, uintptr_t end) 906 : MarkStackTask<false>(thread_pool, mark_sweep, 0, nullptr), bitmap_(bitmap), begin_(begin), 907 end_(end) { 908 } 909 910 protected: 911 accounting::ContinuousSpaceBitmap* const bitmap_; 912 const uintptr_t begin_; 913 const uintptr_t end_; 914 915 virtual void Finalize() { 916 delete this; 917 } 918 919 // Scans all of the objects 920 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 921 ScanObjectParallelVisitor visitor(this); 922 bitmap_->VisitMarkedRange(begin_, end_, visitor); 923 // Finish by emptying our local mark stack. 924 MarkStackTask::Run(self); 925 } 926}; 927 928// Populates the mark stack based on the set of marked objects and 929// recursively marks until the mark stack is emptied. 930void MarkSweep::RecursiveMark() { 931 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 932 // RecursiveMark will build the lists of known instances of the Reference classes. See 933 // DelayReferenceReferent for details. 934 if (kUseRecursiveMark) { 935 const bool partial = GetGcType() == kGcTypePartial; 936 ScanObjectVisitor scan_visitor(this); 937 auto* self = Thread::Current(); 938 ThreadPool* thread_pool = heap_->GetThreadPool(); 939 size_t thread_count = GetThreadCount(false); 940 const bool parallel = kParallelRecursiveMark && thread_count > 1; 941 mark_stack_->Reset(); 942 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 943 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 944 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 945 current_space_bitmap_ = space->GetMarkBitmap(); 946 if (current_space_bitmap_ == nullptr) { 947 continue; 948 } 949 if (parallel) { 950 // We will use the mark stack the future. 951 // CHECK(mark_stack_->IsEmpty()); 952 // This function does not handle heap end increasing, so we must use the space end. 953 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 954 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 955 atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue()); 956 957 // Create a few worker tasks. 958 const size_t n = thread_count * 2; 959 while (begin != end) { 960 uintptr_t start = begin; 961 uintptr_t delta = (end - begin) / n; 962 delta = RoundUp(delta, KB); 963 if (delta < 16 * KB) delta = end - begin; 964 begin += delta; 965 auto* task = new RecursiveMarkTask(thread_pool, this, current_space_bitmap_, start, 966 begin); 967 thread_pool->AddTask(self, task); 968 } 969 thread_pool->SetMaxActiveWorkers(thread_count - 1); 970 thread_pool->StartWorkers(self); 971 thread_pool->Wait(self, true, true); 972 thread_pool->StopWorkers(self); 973 } else { 974 // This function does not handle heap end increasing, so we must use the space end. 975 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 976 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 977 current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 978 } 979 } 980 } 981 } 982 ProcessMarkStack(false); 983} 984 985void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) { 986 ScanGrayObjects(paused, minimum_age); 987 ProcessMarkStack(paused); 988} 989 990void MarkSweep::ReMarkRoots() { 991 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 992 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 993 Runtime::Current()->VisitRoots(this, static_cast<VisitRootFlags>( 994 kVisitRootFlagNewRoots | kVisitRootFlagStopLoggingNewRoots | kVisitRootFlagClearRootLog)); 995 if (kVerifyRootsMarked) { 996 TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings()); 997 VerifyRootMarkedVisitor visitor(this); 998 Runtime::Current()->VisitRoots(&visitor); 999 } 1000} 1001 1002void MarkSweep::SweepSystemWeaks(Thread* self) { 1003 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1004 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 1005 Runtime::Current()->SweepSystemWeaks(this); 1006} 1007 1008class VerifySystemWeakVisitor : public IsMarkedVisitor { 1009 public: 1010 explicit VerifySystemWeakVisitor(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {} 1011 1012 virtual mirror::Object* IsMarked(mirror::Object* obj) OVERRIDE 1013 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) { 1014 mark_sweep_->VerifyIsLive(obj); 1015 return obj; 1016 } 1017 1018 MarkSweep* const mark_sweep_; 1019}; 1020 1021void MarkSweep::VerifyIsLive(const mirror::Object* obj) { 1022 if (!heap_->GetLiveBitmap()->Test(obj)) { 1023 // TODO: Consider live stack? Has this code bitrotted? 1024 CHECK(!heap_->allocation_stack_->Contains(obj)) 1025 << "Found dead object " << obj << "\n" << heap_->DumpSpaces(); 1026 } 1027} 1028 1029void MarkSweep::VerifySystemWeaks() { 1030 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1031 // Verify system weaks, uses a special object visitor which returns the input object. 1032 VerifySystemWeakVisitor visitor(this); 1033 Runtime::Current()->SweepSystemWeaks(&visitor); 1034} 1035 1036class CheckpointMarkThreadRoots : public Closure, public RootVisitor { 1037 public: 1038 explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep, 1039 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) 1040 : mark_sweep_(mark_sweep), 1041 revoke_ros_alloc_thread_local_buffers_at_checkpoint_( 1042 revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 1043 } 1044 1045 void VisitRoots(mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED) 1046 OVERRIDE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 1047 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1048 for (size_t i = 0; i < count; ++i) { 1049 mark_sweep_->MarkObjectNonNullParallel(*roots[i]); 1050 } 1051 } 1052 1053 void VisitRoots(mirror::CompressedReference<mirror::Object>** roots, size_t count, 1054 const RootInfo& info ATTRIBUTE_UNUSED) 1055 OVERRIDE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 1056 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1057 for (size_t i = 0; i < count; ++i) { 1058 mark_sweep_->MarkObjectNonNullParallel(roots[i]->AsMirrorPtr()); 1059 } 1060 } 1061 1062 virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS { 1063 ATRACE_BEGIN("Marking thread roots"); 1064 // Note: self is not necessarily equal to thread since thread may be suspended. 1065 Thread* const self = Thread::Current(); 1066 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 1067 << thread->GetState() << " thread " << thread << " self " << self; 1068 thread->VisitRoots(this); 1069 ATRACE_END(); 1070 if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) { 1071 ATRACE_BEGIN("RevokeRosAllocThreadLocalBuffers"); 1072 mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread); 1073 ATRACE_END(); 1074 } 1075 // If thread is a running mutator, then act on behalf of the garbage collector. 1076 // See the code in ThreadList::RunCheckpoint. 1077 if (thread->GetState() == kRunnable) { 1078 mark_sweep_->GetBarrier().Pass(self); 1079 } 1080 } 1081 1082 private: 1083 MarkSweep* const mark_sweep_; 1084 const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_; 1085}; 1086 1087void MarkSweep::MarkRootsCheckpoint(Thread* self, 1088 bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) { 1089 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1090 CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint); 1091 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 1092 // Request the check point is run on all threads returning a count of the threads that must 1093 // run through the barrier including self. 1094 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1095 // Release locks then wait for all mutator threads to pass the barrier. 1096 // If there are no threads to wait which implys that all the checkpoint functions are finished, 1097 // then no need to release locks. 1098 if (barrier_count == 0) { 1099 return; 1100 } 1101 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1102 Locks::mutator_lock_->SharedUnlock(self); 1103 { 1104 ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun); 1105 gc_barrier_->Increment(self, barrier_count); 1106 } 1107 Locks::mutator_lock_->SharedLock(self); 1108 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1109} 1110 1111void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1112 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1113 Thread* self = Thread::Current(); 1114 mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>( 1115 sweep_array_free_buffer_mem_map_->BaseBegin()); 1116 size_t chunk_free_pos = 0; 1117 ObjectBytePair freed; 1118 ObjectBytePair freed_los; 1119 // How many objects are left in the array, modified after each space is swept. 1120 StackReference<mirror::Object>* objects = allocations->Begin(); 1121 size_t count = allocations->Size(); 1122 // Change the order to ensure that the non-moving space last swept as an optimization. 1123 std::vector<space::ContinuousSpace*> sweep_spaces; 1124 space::ContinuousSpace* non_moving_space = nullptr; 1125 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1126 if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) && 1127 space->GetLiveBitmap() != nullptr) { 1128 if (space == heap_->GetNonMovingSpace()) { 1129 non_moving_space = space; 1130 } else { 1131 sweep_spaces.push_back(space); 1132 } 1133 } 1134 } 1135 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1136 // the other alloc spaces as an optimization. 1137 if (non_moving_space != nullptr) { 1138 sweep_spaces.push_back(non_moving_space); 1139 } 1140 // Start by sweeping the continuous spaces. 1141 for (space::ContinuousSpace* space : sweep_spaces) { 1142 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1143 accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1144 accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1145 if (swap_bitmaps) { 1146 std::swap(live_bitmap, mark_bitmap); 1147 } 1148 StackReference<mirror::Object>* out = objects; 1149 for (size_t i = 0; i < count; ++i) { 1150 mirror::Object* const obj = objects[i].AsMirrorPtr(); 1151 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1152 continue; 1153 } 1154 if (space->HasAddress(obj)) { 1155 // This object is in the space, remove it from the array and add it to the sweep buffer 1156 // if needed. 1157 if (!mark_bitmap->Test(obj)) { 1158 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1159 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1160 freed.objects += chunk_free_pos; 1161 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1162 chunk_free_pos = 0; 1163 } 1164 chunk_free_buffer[chunk_free_pos++] = obj; 1165 } 1166 } else { 1167 (out++)->Assign(obj); 1168 } 1169 } 1170 if (chunk_free_pos > 0) { 1171 TimingLogger::ScopedTiming t2("FreeList", GetTimings()); 1172 freed.objects += chunk_free_pos; 1173 freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1174 chunk_free_pos = 0; 1175 } 1176 // All of the references which space contained are no longer in the allocation stack, update 1177 // the count. 1178 count = out - objects; 1179 } 1180 // Handle the large object space. 1181 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1182 if (large_object_space != nullptr) { 1183 accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap(); 1184 accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap(); 1185 if (swap_bitmaps) { 1186 std::swap(large_live_objects, large_mark_objects); 1187 } 1188 for (size_t i = 0; i < count; ++i) { 1189 mirror::Object* const obj = objects[i].AsMirrorPtr(); 1190 // Handle large objects. 1191 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1192 continue; 1193 } 1194 if (!large_mark_objects->Test(obj)) { 1195 ++freed_los.objects; 1196 freed_los.bytes += large_object_space->Free(self, obj); 1197 } 1198 } 1199 } 1200 { 1201 TimingLogger::ScopedTiming t2("RecordFree", GetTimings()); 1202 RecordFree(freed); 1203 RecordFreeLOS(freed_los); 1204 t2.NewTiming("ResetStack"); 1205 allocations->Reset(); 1206 } 1207 sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero(); 1208} 1209 1210void MarkSweep::Sweep(bool swap_bitmaps) { 1211 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1212 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 1213 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 1214 { 1215 TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings()); 1216 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 1217 // knowing that new allocations won't be marked as live. 1218 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 1219 heap_->MarkAllocStackAsLive(live_stack); 1220 live_stack->Reset(); 1221 DCHECK(mark_stack_->IsEmpty()); 1222 } 1223 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1224 if (space->IsContinuousMemMapAllocSpace()) { 1225 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1226 TimingLogger::ScopedTiming split( 1227 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", GetTimings()); 1228 RecordFree(alloc_space->Sweep(swap_bitmaps)); 1229 } 1230 } 1231 SweepLargeObjects(swap_bitmaps); 1232} 1233 1234void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1235 space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace(); 1236 if (los != nullptr) { 1237 TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings()); 1238 RecordFreeLOS(los->Sweep(swap_bitmaps)); 1239 } 1240} 1241 1242// Process the "referent" field in a java.lang.ref.Reference. If the referent has not yet been 1243// marked, put it on the appropriate list in the heap for later processing. 1244void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) { 1245 if (kCountJavaLangRefs) { 1246 ++reference_count_; 1247 } 1248 heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, this); 1249} 1250 1251class MarkVisitor { 1252 public: 1253 explicit MarkVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) { 1254 } 1255 1256 void operator()(mirror::Object* obj, MemberOffset offset, bool /* is_static */) const 1257 ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 1258 EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1259 if (kCheckLocks) { 1260 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1261 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1262 } 1263 mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset), obj, offset); 1264 } 1265 1266 private: 1267 MarkSweep* const mark_sweep_; 1268}; 1269 1270// Scans an object reference. Determines the type of the reference 1271// and dispatches to a specialized scanning routine. 1272void MarkSweep::ScanObject(mirror::Object* obj) { 1273 MarkVisitor mark_visitor(this); 1274 DelayReferenceReferentVisitor ref_visitor(this); 1275 ScanObjectVisit(obj, mark_visitor, ref_visitor); 1276} 1277 1278void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1279 Thread* self = Thread::Current(); 1280 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1281 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1282 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1283 CHECK_GT(chunk_size, 0U); 1284 // Split the current mark stack up into work tasks. 1285 for (auto* it = mark_stack_->Begin(), *end = mark_stack_->End(); it < end; ) { 1286 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1287 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it)); 1288 it += delta; 1289 } 1290 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1291 thread_pool->StartWorkers(self); 1292 thread_pool->Wait(self, true, true); 1293 thread_pool->StopWorkers(self); 1294 mark_stack_->Reset(); 1295 CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(), 1296 work_chunks_deleted_.LoadSequentiallyConsistent()) 1297 << " some of the work chunks were leaked"; 1298} 1299 1300// Scan anything that's on the mark stack. 1301void MarkSweep::ProcessMarkStack(bool paused) { 1302 TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings()); 1303 size_t thread_count = GetThreadCount(paused); 1304 if (kParallelProcessMarkStack && thread_count > 1 && 1305 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1306 ProcessMarkStackParallel(thread_count); 1307 } else { 1308 // TODO: Tune this. 1309 static const size_t kFifoSize = 4; 1310 BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo; 1311 for (;;) { 1312 mirror::Object* obj = nullptr; 1313 if (kUseMarkStackPrefetch) { 1314 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1315 mirror::Object* mark_stack_obj = mark_stack_->PopBack(); 1316 DCHECK(mark_stack_obj != nullptr); 1317 __builtin_prefetch(mark_stack_obj); 1318 prefetch_fifo.push_back(mark_stack_obj); 1319 } 1320 if (prefetch_fifo.empty()) { 1321 break; 1322 } 1323 obj = prefetch_fifo.front(); 1324 prefetch_fifo.pop_front(); 1325 } else { 1326 if (mark_stack_->IsEmpty()) { 1327 break; 1328 } 1329 obj = mark_stack_->PopBack(); 1330 } 1331 DCHECK(obj != nullptr); 1332 ScanObject(obj); 1333 } 1334 } 1335} 1336 1337inline mirror::Object* MarkSweep::IsMarked(mirror::Object* object) { 1338 if (immune_region_.ContainsObject(object)) { 1339 return object; 1340 } 1341 if (current_space_bitmap_->HasAddress(object)) { 1342 return current_space_bitmap_->Test(object) ? object : nullptr; 1343 } 1344 return mark_bitmap_->Test(object) ? object : nullptr; 1345} 1346 1347void MarkSweep::FinishPhase() { 1348 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1349 if (kCountScannedTypes) { 1350 VLOG(gc) << "MarkSweep scanned classes=" << class_count_.LoadRelaxed() 1351 << " arrays=" << array_count_.LoadRelaxed() << " other=" << other_count_.LoadRelaxed(); 1352 } 1353 if (kCountTasks) { 1354 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed(); 1355 } 1356 if (kMeasureOverhead) { 1357 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed()); 1358 } 1359 if (kProfileLargeObjects) { 1360 VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed() 1361 << " marked " << large_object_mark_.LoadRelaxed(); 1362 } 1363 if (kCountJavaLangRefs) { 1364 VLOG(gc) << "References scanned " << reference_count_.LoadRelaxed(); 1365 } 1366 if (kCountMarkedObjects) { 1367 VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed() 1368 << " immune=" << mark_immune_count_.LoadRelaxed() 1369 << " fastpath=" << mark_fastpath_count_.LoadRelaxed() 1370 << " slowpath=" << mark_slowpath_count_.LoadRelaxed(); 1371 } 1372 CHECK(mark_stack_->IsEmpty()); // Ensure that the mark stack is empty. 1373 mark_stack_->Reset(); 1374 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 1375 heap_->ClearMarkedObjects(); 1376} 1377 1378void MarkSweep::RevokeAllThreadLocalBuffers() { 1379 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) { 1380 // If concurrent, rosalloc thread-local buffers are revoked at the 1381 // thread checkpoint. Bump pointer space thread-local buffers must 1382 // not be in use. 1383 GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked(); 1384 } else { 1385 TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings()); 1386 GetHeap()->RevokeAllThreadLocalBuffers(); 1387 } 1388} 1389 1390} // namespace collector 1391} // namespace gc 1392} // namespace art 1393